Fault Diagnosis Method of Distribution Equipment Based on Hybrid Model of Robot and Deep Learning

Shan Rongrong; Ma Zhenyu; Ye Hong; Lin Zhenxing; Qiu Gongming; Ge Chengyu; Lu Yang; Yu Kun

doi:10.1155/2022/9742815

Fault Diagnosis Method of Distribution Equipment Based on Hybrid Model of Robot and Deep Learning

Rongrong, Shan;Zhenyu, Ma;Hong, Ye;Zhenxing, Lin;Gongming, Qiu;Chengyu, Ge;Yang, Lu;Kun, Yu 2022-04-14 00:00:00 Hindawi Journal of Robotics Volume 2022, Article ID 9742815, 11 pages https://doi.org/10.1155/2022/9742815 Research Article Fault Diagnosis Method of Distribution Equipment Based on Hybrid Model of Robot and Deep Learning 1 2 3 3 1 1 Shan Rongrong , Ma Zhenyu, Ye Hong, Lin Zhenxing, Qiu Gongming, Ge Chengyu, 1 1 Lu Yang, and Yu Kun NARI Group Co., Ltd, State Grid Electric Power Research Institute, Nanjing 210000, China Zhejiang Electric Power Corporation, Hangzhou 310013, China State Grid Wenzhou Power Supply Company Ouhai Power Supply Branch, Wenzhou 325000, China Correspondence should be addressed to Shan Rongrong; shanrongrong@sgepri.sgcc.com.cn Received 8 February 2022; Revised 14 March 2022; Accepted 19 March 2022; Published 14 April 2022 Academic Editor: Shan Zhong Copyright © 2022 Shan Rongrong et al. 'is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. In view of the poor eﬀect of most fault diagnosis methods on the intelligent recognition of equipment images, a fault diagnosis methodofdistributionequipmentbasedonthehybridmodelofrobotanddeeplearningisproposedtoreducethedependenceon manpower and realize eﬃcient intelligent diagnosis. Firstly, the robot is used to collect the on-site state images of distribution equipmenttobuildtheimageinformationdatabaseofdistribution equipment.At thesametime, therobot backgroundisused as the comprehensive database data analysis platform to optimize the sample quality of the database. 'en, the massive infrared images are segmented based on chroma saturation brightness space to distinguish the defective equipment images, and the defective equipment areas are extracted from the images by OTSU method. Finally, the residual network is used to improve the region-based fully convolutional networks (R-FCN) algorithm, and the improved R-FCN algorithm trained by the online hard example mining method is used for fault feature learning. 'e fault type, grade, and location of distribution equipment are obtained through fault criterion analysis. 'e experimental analysis of the proposed method based on PyTorch platform shows that the fault diagnosis time and accuracy are about 5.5s and 92.06%, respectively, which are better than other comparison methods and provide a certain theoretical basis for the automatic diagnosis of power grid equipment. be “allocated and used” [3]. 'e distribution equipment will 1. Introduction have some faults under long-term operation, resulting in With the continuous improvement of social economy and abnormal temperature. 'erefore, by detecting the tem- people’s living standards, the power demand is increasing perature of the distribution equipment, the thermal fault day by day, and the scale of power system is growing day by diagnosis of the distribution equipment can be carried out day. It includes transmission and transformation networks quickly, which plays a great role in the safe operation of the power grid. of various voltage levels. 'erefore, ensuring the safe and stable operation of complex power grid is an inevitable 'e infrared image of equipment is used for fault di- requirement to ensure the economic and social develop- agnosis with high eﬃciency, accurate judgment, safety, and ment. At the same time, the economic and social devel- reliability. At the same time, it is free from electromagnetic opmenthasagreatimpactonthesecurityandeconomy,and interference, fast detection speed, and no power failure of higher reliability is required [1, 2]. As the terminal of the live equipment. 'erefore, infrared diagnosis is widely used wholepowergridoperation,distributionnetwork isthepart in the ﬁeld of equipment fault monitoring and diagnosis with the widest coverage and the largest scale in China’s technology [4]. However, due to the characteristics of large powersystem, andit is thekey link toensure that powercan quantity and complex types of distribution equipment, if we 2 Journal of Robotics only rely on manual work in the process of data acquisition, breaker fault identiﬁcation. 'e image analysis process is analysis, and processing, the workload is relatively large, the complex, resulting in the reduction of fault identiﬁcation eﬃciency. Reference [15] proposes a method based on eﬃciencyislow,andtheaccuracy isrelativelylowdue tothe high dependence on manual experience. 'erefore, auto- feature model for single-phase grounding fault in active matic image acquisition and analysis of distribution distribution network system, which transforms the solution equipment are of great signiﬁcance to ensure the safety and of nonlinear feature model into single-objective optimiza- stability of distribution network [5]. tion of feature entropy, which can well identify single-phase In recent years, the automatic inspection technology of fault, but the identiﬁcation eﬀect of equipment with feature power distribution room has been popularized, and various type is not ideal. automaticrobotsandUAVshavemadegreatprogressinthe With the continuous development of computer original data acquisition stage. However, the accurate and technology and the rapid development of 5G communi- eﬃcient processing of collected image data is still in its cation technology, machine learning algorithm has been infancy.Howtoextractthefeaturesofinterestfrominfrared widely used this year, especially the deep learning algo- rithm has certain advantages in the ﬁeld of fault identi- images for power distribution equipment recognition is a problem to be solved [6]. Among them, the deep learning ﬁcation and classiﬁcation. Reference [16] proposes algorithmhasmadegreatachievementsinimageprocessing, artiﬁcial neural network algorithm to identify the insu- speechrecognition,andtextanalysis.Byestablishingadeep- lator stateand uses single-layerand multilayerperceptron seatedneuralnetwork,high-levelfeaturesareextractedfrom artiﬁcial intelligence algorithm to classify the conditions low-level features layer by layer, so as to achieve the eﬀect of of distribution insulators. 'is technology can make the target classiﬁcation and recognition [7]. Compared with the automatic inspection of electrical system more accurate manually designed feature extraction method, the distrib- and eﬃcient, but it lacks high reliable database for sup- uted features obtained by deep learning network model can port. Reference [17] proposed a Mask R convolution better express the essence of data [8, 9]. 'erefore, in order neural network method and used transfer learning and to improve the eﬃciency of thermal fault detection of dis- dynamic learning rate algorithm to realize eﬃcient rec- ognition of annotated image data sets, but it relied too tributionequipment,improvetheintelligenceofpowergrid, reduce the labor cost of detection, and reduce the false much on graphics annotation and lacked practical ap- detection rate, a fault diagnosis method of distribution plication value. In [18], appropriate traveling wave time- equipment based on the hybrid model of robot and deep frequency characteristic parameters of fault current are learning is proposed, which eﬀectively ensures the safe and selected as the input of adaptive depth belief network reliable operation of distribution equipment. model to obtain the fault type, but only considering the fault current characteristics as the basis, the reliability needs to be further improved. 2. Related Research Basedontheaboveanalysis,aimingattheproblemssuch as the complexity and diversity of smart grid distribution At present, there are many researches on fault diagnosis of distribution equipment at home and abroad, which can be equipment and the unsatisfactory eﬀect of most existing image recognition methods, a distribution equipment fault divided into traditional fault identiﬁcation and classiﬁcation methods and machine learning based identiﬁcation and diagnosis method based on robot and deep learning hybrid model is proposed. Its innovations are summarized as classiﬁcation methods [10]. Among them, the traditional fault identiﬁcation and classiﬁcation methods mainly in- follows: clude fuzzy clustering, discrete wavelet transform, and (1) In order to obtain the image information of distri- chaotic algorithm. For example, [11] proposed an infrared bution equipment more comprehensively, the pro- image segmentation algorithm based on intuitionistic fuzzy posed method introduces the robot to construct the clustering algorithm based on spatial distribution infor- corresponding image knowledge database, which mation, which is suitable for power equipment. It can well provides the basis for fault classiﬁcation and fault suppress the strong interference of nontarget objects in location. infrared image to image segmentation, but the method is (2) In order to locate the equipment defect area in the more traditional and has poor segmentation eﬀect for infrared image of distribution equipment, the pro- complex intelligent power grid equipment. Reference [12] posed method performs threshold segmentation on proposed an anomaly detection method based on spatial the infrared image in hue saturation value (HSV) clustering applied by auxiliary feature vector and density space and uses OTSU method to extract the noise. 'e auxiliary feature vector of each conditional equipmentdefectarea,soas toimprovetheaccuracy variable is constructed for clustering to identify normal data of subsequent fault diagnosis. patterns and diﬀerent types of anomalies. Reference [13] (3) Aiming at the problem that the deep learning al- proposed a data mining driving scheme based on discrete gorithm is prone to gradient disappearance and wavelet transform to realize high impedance fault detection gradient explosion, the proposed method uses the in active distribution network, but the universality of the residual network to improve the region-based fully method is not high. Reference [14] proposed a method to convolutional networks (R-FCN) algorithm and obtain the vibration characteristics of circuit breaker based appliesittothelearningoffaultyequipment,soasto on time-frequency and chaos analysis to realize circuit Journal of Robotics 3 obtainthe fault type and location with high accuracy Robot and further improve the safety of equipment. Defect backstage system 3. Proposed Method 3.1. Construction of Image Information Database of Distri- System data Fault information Model base analysis center bution Equipment Based on Robot Inspection. 'etraditional knowledge base equipment status is usually determined by manual analysis. 'e workload is huge and error-prone, which aﬀects the judgment of system status, resulting in potential safety Knowledge database hazards. 'erefore, the robot is used for patrol inspection to Database obtain the status image of distribution equipment and build the corresponding information base for the analysis of Production Robot data system equipment status, so as to ﬁnd the faulty equipment in time Online and ensure the reliable operation of power grid [19]. 'e monitoring construction process of distribution equipment image in- formation base based on robot inspection is shown in Figure 1: Construction process of image information database for Figure 1. distribution equipment. 'e basic data sources of the database mainly include production system, online monitoring system, and robot (1) 'e data of the three-party platform includes the backgroundinspectionsystem.'erelevant data ofthestate information required in the database structure table. quantity of power equipment mainly comes from the power After eliminating the redundant information, the production management system (PMS), which can provide integrated data in a uniﬁed format can be obtained, the real-time operation condition, historical operation state, and the defect alarm data can be located and re- historical maintenance record, historical test data, equip- trieved quickly. ment account, equipment parameters, and other informa- tionoftheequipment.'eonlinemonitoringsystemmainly (2) 'e fault information base is mainly taken from the defectsystemrecordsandcontainsalargenumberof relies on various sensors on each power equipment for real- time monitoring. 'e robot background inspection system relevant equipment fault cases, including fault characteristics, solutions, expert opinions, and can not only provide the observation of some state quan- manufacturer records. At the same time, the tities, but also carry out corresponding state evaluation and maintenance record database and equipment ac- analysis for diﬀerent equipment states according to the count database are used to build a comprehensive automatic state evaluation system. In addition, the data databaseoffaultinformation,soastoscreenthefault composition of the system includes infrared temperature inspection points. measurement, visible light reading, and telemetry reading. 'e robot inspection cycle generally refers to the in- (3) 'eknowledgeinformationbaseistheengineforthe spection plan formulated by the distribution network op- system to evaluate the equipment status and judge eration inspection center, and two inspection robots the fault. 'e internal rules at all levels provide the complete the tasks of infrared temperature measurement logicalbasisforthesystemtojudgethefault.'ekey and data transcription of equipment in the area [20]. At the is knowledge acquisition, that is, collecting and same time, the robot background uses the threshold out of mining the knowledge at all levels to enrich the limit judgment method to automatically evaluate the knowledge base. equipmentstatus.Inordertoensuresuﬃcientchargingtime of the robot and avoid the daily patrol and infrared tem- perature measurement period, the special patrol at night is 3.2. Defect Feature Extraction of Distribution Equipment. set in the nonbusy working period of the robot every day, When extracting the defect features of distribution equip- with the upper limit of one time. 'e data reports collected ment, it is necessary to perform threshold segmentation on by the special patrol at night and infrared temperature theinfraredimageinHSVspace,separatetheinfraredimage measurement are included in the database for screening and background from irrelevant equipment and defective preprocessing. equipment, and then extract the equipment defect area [21]. In addition, the background of the inspection robot is equipped with a system server, which includes data analysis software terminal, data exchange server, data storage server, 3.2.1. OTSU 4reshold Segmentation. OTSUisconsideredto dataoperationserver,andothermodules.'edataexchange be one of the best algorithms in image threshold segmen- server is responsible for collecting and classifying the pro- tation. 'e threshold segmentation process of OTSU algo- rithm is as follows: ﬁrstly, the image is processed in gray duction system, online monitoring system, and robot patrol data into the storage server. 'ere are three-party databases, level, the number of pixels in the whole image is counted, fault information base and knowledge information base in and the probability distribution of each pixel in the whole the data storage server. image is calculated; then, the gray level is traversed and 4 Journal of Robotics searchedinthewholeimage,andtheinterclassprobabilityof classiﬁcation [22]. 'e ﬂow of HSV based defect region the image foreground and background at the current gray extraction algorithm is shown in Figure 2. leveliscalculated;ﬁnally,thethresholdcorrespondingtothe When processing the infrared image of defective variance between classes and within classes is calculated by equipment, ﬁrst merge the similar pixels corresponding to the given objective function. the area with the same temperature, and segment the image Suppose there are D gray levels in the image, in which according to the threshold of the three components of the the number of pixels with gray value of i is N and the total defective areaintheHSV colorspace toextract thedefective number of pixels in the image is N. 'en, the average gray area.'en,thediscretedefectregionsareconnectedthrough value of the whole image is the closed operation in mathematical morphology, and the threshold segmentation of the original image is carried out D−1 by OTSU method to separate the power equipment and the μ � 􏽘 i . (1) background region. Finally, the defect area is found in the i�0 binary image separated by OTSU method; that is, the de- According to the gray characteristics of the image, the fective power equipment is separated from other areas, so as image is divided into foreground B and background B . 0 1 to achieve the purpose of extracting defective power 'en, p (T) and p (T) represent the probability of oc- 0 1 equipment and facilitate the identiﬁcation and diagnosis of currence of foreground B and background B when the 0 1 power equipment types and fault types. threshold is T, respectively. 'e calculation is as follows: p (T) � 􏽘 􏼒 􏼓, 3.3. Fault Diagnosis of Distribution Equipment Based on Deep (2) i�0 Learning Hybrid Model p (T) � 1 − p (T). 1 0 3.3.1. Defect Training Based on Deep Learning Hybrid Model. R-FCN algorithm architecture mainly includes backbone 'en,themeanvaluesofforeground B andbackground network, region proposal network (RPN), and region of B are interest (ROI) subnet [23]. When fault diagnosis of power ⎧ ⎪ 􏽐 i N /N 􏼁 distribution equipment is carried out, ﬁrst input the col- ⎪ i�0 μ (T) � , ⎪ 0 ⎪ lected infrared image of power equipment into convolution p (T) neural network and extract the convolution feature map of (3) ⎪ infrared image. In this process, deeper and more abstract ⎪ T μ − 􏽐 i N /N􏼁 i�0 ⎪ 􏽐 image features can be extracted by using a larger backbone ⎩ μ (T) � . network (ResNet 101) to improve the recognition accuracy p (T) [24]. 'en, the feature map is sent to the RPN network to 'e interclass variance with threshold T in the gray generate anchors, which are marked with foreground and histogram is calculated as follows: background, and the foreground area with high score is 2 2 selectedastherecommendedareaROIs.'eseROIsaresent to the ROI subnet for further training, and 300 recom- σ (T) � p (T)􏼢μ (T) − μ 􏼣 + p (T)􏼢μ (T) − μ 􏼣 . 0 1 B 0 1 􏽘 􏽘 mended windows are generated for each infrared image of (4) power equipment. At the same time, the characteristic map of the full convolution layer is calculated with the multilayer 'e optimal threshold is deﬁned as the T value corre- convolution kernel to generate a position sensitive score sponding to the maximum variance between classes, which map. 'e ROI and Score Maps are input into the later is calculated as Softmax layer for vote. 'rough the Softmax layer for 2 2 classiﬁcation, the ROI with the highest score is ﬁnally ob- σ (T) � max 􏽮σ (T)􏽯. (5) B B 0≤T≤D−1 tained,thatis,thelocationandtypeoftheobjectlocatedand recognized. 'e architecture of R-FCN algorithm is shown in Figure 3. 3.2.2. Defect Region Extraction Based on HSV. In order to (1) Residual Network. When the depth of the deep learning improve the accuracy of equipment fault image classiﬁca- network reaches a certain degree, the problems of gradient tion,thedefectregionandbackgroundintheinfraredimage disappearance and gradient explosion often appear during of fault power equipment are separated by using the defect training.Inordertosolvethisproblem,theresidualnetwork region segmentation algorithm based on HSV. Since it is (ResNet) is used to improve the R-FCN algorithm; that is, impossible to determine the defect type only by analyzing the residual network is selected as the backbone network. thefaultarea,itisnecessarytosegmentthedefectareabased 'e residual element is essentially the mapping residual on mathematical morphology according to the location of required for ﬁtting through these stacked layers. Suppose the defect area. 'rough this method, the defective power thatthenetworkmappingis H(x)andtheresidualmapping equipment and the background area in the infrared image function of the network is F(x), F(x) � H(x) − x. 'e so- are separated, so as to reduce the interference of the called residual is the diﬀerence between the observed value background area in the infrared image on the defect type Journal of Robotics 5 it only needs to make F(x) � 0 to get H(x) � x, so as to Start avoid the disappearance and explosion of gradient. 'e input x and output x of the m-th residual unit m m+1 Initialize the image clustering center point so that the distance S of are expressed as follows: each clustering center point is evenly distributed in the image. x � f h x + δ x , ω , 􏼁 􏼁 􏼁 m+1 m m m M−1 In each 3 × 3 select the pixel with the smallest gradient in the (6) neighborhood as the clustering center in the neighborhood. x � x + 􏽘 δ x , ω􏼁 , M m i i i�1 The distance metric D is calculated by calculating the distance where δ is the ReLu activation function, ω is the weight metric from each pixel to the cluster center in the super-pixel between each unit, m, M respectively represent the shallow segmentation algorithm, and the smallest distance metric is selected to assign pixels to the corresponding cluster center. residual unit and deep residual unit, h(·) represents the identity mapping, and x is the ﬁnal output of the residual unit. Normalize the processed image and update the H value. Inordertolearnmoreandmoreabstractimagefeatures, the proposed method selects ResNet 101 network, and its conﬁguration is shown in Table 1. Set the threshold. If the three components of HSV meet 10 < H < 150, 10 < S < 100, 200 < V < 255, then H = 0, S = 0, V = 0. The equipment defect area in the infrared image is separated through the threshold. (2) RPN Network.'einputofRPNnetworkisimagefeature graph.Accordingtoanchormechanism,9rectangularboxes with diﬀerent sizes are generated for each point. When OTSU threshold segmentation algorithm is used to segment training the RPN network, compare the anchor with the the original image to obtain the binary image I . manually calibrated true value area in the data set, mark the anchor frame with the largest overlap ratio as the fore- ground, and mark the anchor frame with an overlap ratio Because there may be one or several defect points in the segmented binary image, each isolated defect point is connected according to the greater than 0.7 as the foreground sample. Mark the anchor closed operation of mathematical morphology to determine the defect box whose overlap ratio is less than 0.3 as the background region R . sample.Selectthepositiveandnegativesamplesofanchorin proportion, use the maximum suppression method (NMS) Find each connected region R in the binary image I . If the defective 0 b and other methods to screen the top 250 ROIs with the region is R ∩R > 0.8R , this region is the defective power equipment region 0 s s to be extracted. highest score, and send these preliminarily screened pre- selected frames to the ROI subnet. In addition, RPN network adopts anchor mechanism, End which not only solves the problem of translation invariance, but also enables R-FCN algorithm to identify and locate Figure 2: Defect region extraction process based on HSV. targets with diﬀerent overall dimensions. In the actual process of infrared image recognition of power distribution equipment, due to diﬀerent equipment with diﬀerent shape and structure, diﬀerent sizes and variable aspect ratio, in Input image Position sensitive convolution order to ensure that there are targets in the receptive ﬁeld k (c+1) ROI subnet corresponding to each sliding window on the feature map, conv multiscale anchor is required to ensure that the candidate ROI frame is as complete as possible to select the target [25]. In Softmax pool vote conv Result the implementation of RPN network anchor, multiscale output anchor can be obtained by setting the area of reference c+1 k window (base_size), diﬀerent area multiples and anchor aspect ratio, so that RPN can give more accurate foreground Convolution characteristic recommendation area. RPN graph ROIs (3) ROI Subnet.'eroleofROIsubnetistocorrecttheROIs conv locationobtainedfrom RPNnetwork,soastoobtainamore accuratetargetlocationofpowerequipment,soastoidentify ROI.Apositionsensitiveconvolutionlayerisaddedafterthe Figure 3: Architecture of R-FCN algorithm. last layer of the full convolution network, which can realize the translation variability of the algorithm and output the H(x) and the estimated value x. 'e advantage of ResNet (c + 1)-dimensional position sensitive fractional graph. networkisthatitusesthestackinglayertoﬁt H(x)togetthe 'e RPN network ﬁlters out the characteristic map of ROI mapping H(x) � F(x) + x. 'e advantage of this repre- with size a × b, which is divided into k × k parts (bin) and sentation is that if the model has been ﬁtted to the beststate, convoluted with k (c + 1) convolution cores; that is, each 6 Journal of Robotics Table 1: Structure table of ResNet 101. Layer name 101-Layer conv1 7 7, stride 2 conv2_x 3 3max pool, stride 2 1 ×1 64 ⎡ ⎢ ⎤ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎣3 ×3 64 ⎥ ⎦ × 3 1 ×1 256 1 ×1 128 ⎢ ⎥ ⎡ ⎢ ⎤ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ conv3_x ⎣3 ×3 128 ⎦ × 4 1 ×1 512 1 ×1 256 ⎡ ⎢ ⎤ ⎥ ⎢ ⎢ ⎥ ⎥ ⎢ ⎥ ⎢ ⎥ conv4_x ⎣3 ×3 256 ⎦ × 23 1 ×1 1024 1 ×1 512 ⎡ ⎢ ⎤ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ conv5_x ⎣ 3 ×3 512 ⎦ × 3 1 ×1 2048 Average pool, 1000-D FC, softmax part is mapped to a score map, in which (c +1) is the Classification number of categories plus the background to obtain the 2 loss L cls k (c + 1)-layer position sensitive score map. 'e number of channels of bin in diﬀerent positions is a × b × (c +1), and c+1 vote ROI the score of class (c + 1) in this position is stored. After the pool poolingprocessiscompleted,voteontheROI.Sumthe k × k Regression parts bin to get the output of (c + 1) dimension, that is, the loss L reg probability of category (score), and classify it through the softmax layer. At this time, the output result is the posi- ROI cls pool tioning coordinates and type of the object. vote c+1 (4) OHEM. In RPN network, a large number of rectangular reg boxes are generated, and hundreds or thousands of regions will participate in the training of predicting target categories Hard RoI and locations. 'e proportion of power equipment is small, Sampler so the ratio of equipment background area to target area is Figure 4: Training framework of OHEM method. too large, resulting in sample imbalance, which makes it diﬃcult to identify distribution equipment. 'erefore, online hard example mining (OHEM) method is used to backpropagated to the network and trained again to update trainthenetworkmodel[26].Duringtraining,whenthereis theweightofthewholenetwork,andthepenaltyforlowloss a preselected area with large loss, the hard example can be samples is ignored. Among them, the weight update is re- trained and classiﬁed again, which can solve the problem of alized based on the random gradient descent method. imbalance between positive and negative sample categories and improve the accuracy of infrared image recognition model of power equipment. 'e training framework of 3.3.2. Fault Diagnosis of Distribution Equipment. 'ere are OHEM method is shown in Figure 4. many kinds of equipment in the distribution network, such WhenOHEMcarriesoutspeciﬁctraining,ﬁrstly,ResNet as circuit breaker, potential transformer (PT), current 101 is used to extract the image features of training samples, transformer(CT),lightningarrester,andtransformer.'ese train the image classiﬁcation and positioning branches, and equipment can be classiﬁed into current heating type, calculatetheclassiﬁcationloss L andregressionloss L of voltage heating type, and comprehensive heating type cls reg each target. 'e loss function is calculated as follows: according to the heating factors. Diﬀerent diagnostic methods are selected for diﬀerent types of equipment, and ∗ ∗ L􏼐s, d 􏼑 � L s 􏼁 + ζ c >0 􏼁 L d, d 􏼁 , (7) x,y,a,b cls c reg diﬀerent diagnostic methods have diﬀerent diagnostic cri- teria. 'e defects of the equipment, such as general defects, where d istheupperleftcoordinateofthetargetarea, d is x,y a serious defects, and critical defects, are determined through thewidthofthetargetarea, d istheheightofthetargetarea, various diagnostic methods. ζ is the balance coeﬃcient of classiﬁcation loss and re- gression loss, and s is the image type. (1) 'e heating of current heating equipment is mainly 'en, sort according to the loss value from high to low, due to the current thermal eﬀect. Generally, the and use the NMS method to select the ﬁrst Ω samples with surface temperature judgment method and relative the largest loss value to screen out the diﬃcult cases of this temperature diﬀerence judgment method are used round of samples. Finally, these diﬃcult samples are for fault diagnosis. 'e relative temperature Position sensitive convolution Journal of Robotics 7 diﬀerence ∇ andthemaximumsurfacetemperature PyTorch v1.3 version, the programming language is Python T of diﬀerent defect degrees of each distribution 3.6.0, and third-party dependent libraries such as Open CV max equipment are shown in Table 2. 4.0 and NumPy 1.3 are used for batch processing of data. At the same time, the network model is trained on the GPU of (2) Voltage heating equipment is mainly due to voltage dual card Tesla P100, with a total video memory of 8GB. In eﬀect. 'e main equipment categories include zinc this way, large batch data can be set during training to oxide arrester, high-voltage bushing, and coupling improve the convergence speed of the model. Other hard- capacitor. 'is kind of equipment generally adopts ware environments are as follows: 512GB of memory re- imagefeaturejudgmentmethod,similarcomparison sources and 1TB of hard disk. When processing data and I/ judgment method, and comprehensive analysis O operations on a large scale, this can realize parallel judgment method. Using the thermal image char- processing and high speed and ensure the training re- acteristics of the equipment and the image feature quirements of network model. judgment method, the fault can be found quickly. If the similar comparison discrimination method and comprehensive analysis judgment method are used, 4.1. Experimental Data Set. 'e research scenario is the the temperature diﬀerence ΔT shall be taken as the power equipment in the distribution network, so it is fault diagnosis index, in which the ΔT of zinc oxide necessarytocollectthepicturematerialsofequipmentfaults arrester is 0.5∼1K, and the ΔT of high-voltage in the distribution system and then construct a ﬁxed format bushing, coupling capacitor, oil immersed PT and data set for test training. 'rough safety training and pro- CT are 2∼3K. fessional leadership, use mobile phones, cameras, and other (3) Comprehensive heating equipment needs to be di- equipment to shoot at the site of power distribution equipment. 'e scene of abnormal power grid equipment is agnosed in combination with the diagnosis methods of voltage heating equipment and current heating selected,andalargenumberofpositivesamplesarecollected in multiple directions according to the shooting angle of equipment, mainly including insulators, generators, and transformers. In the actual thermal fault diag- video monitoring. In order to ensure the rationalization of data distribution, all types and forms of equipment anomaly nosis, the fault diagnosis indexes of various methods should be combined to improve the eﬃciency and types are covered in the acquisition process. After that, the accuracy of fault diagnosis [27]. abnormal categories and areas of power grid equipment are marked through the open-source and free wizard marking According to diﬀerent types of distribution equipment, software. the corresponding thermal fault diagnosis and judgment 'e data set contains 2580 on-site abnormal images of methods can be selected, the corresponding parameters can RGBpowergridequipmentduringtheday,6300markboxes be calculated, and the thermal fault diagnosis of distribution intotal,and2769on-siteabnormalimagesofinfraredpower equipment can be carried out according to the diagnosis grid equipment at night, 7000 mark boxes in total, all from criteria. 'e fault diagnosis process of power distribution the real power distribution room, power equipment plant, equipment is shown in Figure 5. etc. After the data annotation is completed, the annotation Firstly, the infrared image of distribution equipment is ﬁle in XML format is generated, which corresponds to the input into the detection model for image preprocessing, and real image. the defect area is extracted based on HSV to divide the structure of distribution equipment. 'en,the trained depth 4.2. Comparative Analysis of Training Speed. 'e proposed learninghybridmodelisusedtoobtainthetypeandlocation ofthetargetequipment,andthetemperatureinformationof method combines the hybrid model of robot and deep eachstructuralareaontheinfraredimageofthedistribution learning for equipment fault diagnosis. ResNet network is equipmentis readat thesametime.Finally,accordingtothe usedtooptimizetheR-FCNalgorithm,andthedefectareais selected fault diagnosis method, the thermal fault state, extracted based on HSV to improve the diagnosis eﬀect. In thermalfault level,andthermalfault locationof distribution order to demonstrate the improvement eﬀect of the pro- equipment are determined by using the diagnosis criterion. posed method, it is compared with the diagnosis methods of ResNet network, Otsu threshold segmentation, and OHEM Diﬀerent from the existing fault diagnosis methods of ar- tiﬁcial equipment, the proposed method can use the deep training.'ediagnosisaccuracyandtrainingtimeareshown in Table 3. learning hybrid model to realize the intelligent classiﬁcation and fault type diagnosis of infrared images of distribution It can be seen from Table 3 that extracting deeper equipment fault features using ResNet network can greatly equipment, greatly reduce the workload of inspectors, and improve the automation level of fault diagnosis of distri- improve the diagnosis accuracy, which is 9.88% higher than bution equipment. thatofthemodel.However,duetothedeepeningofnetwork layers, the training time is also increased, more than 10s. At thesametime,byintegratingOHEMtrainingdepthlearning 4. Experiment and Analysis model, the diagnostic accuracy continued to improve by 'e experiment is carried out in the 64 bit operating system 4.41%. Due to the simple and easy implementation of the training process, the training time is only increased by 0.7s. environment of Ubuntu 6.04.4 LTS, in which the deep learning hybrid model uses the deep learning framework It can be seen that the diagnostic accuracy of the proposed 8 Journal of Robotics Table 2: Fault diagnosis criterion of current heating equipment. General defect Serious defect Emergency defect Circuit breaker 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 55%≤ T ≤80% T >80% max max Disconnecting switch 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 90%≤ T ≤130% T >130% max max CT 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 55%≤ T ≤80% T >80% max max Capacitor 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 55%≤ T ≤80% T >80% max max High voltage bushing 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 55%≤ T ≤80% T >80% max max Gets the type and Infrared image input location of the target device Extract the original temperature of the target Image preprocessing device Select diagnostic Extracting defect method region based on HSV Calculate the diagnosis index of Structural region each structural area and division (image determine the fault level. segmentation) Trained deep learning Output results model Figure 5: Fault diagnosis process of distribution equipment. Table 3: Fault diagnosis criterion of current heating equipment. Method Accuracy/% Training time/s Original model 82.49 4.6 Original model+OTSU threshold segmentation 85.82 5.9 Original model+OTSU threshold segmentation+ResNet network 92.37 10.5 Original model+OTSU threshold segmentation+ResNet network+OHEM 96.78 11.2 method is higher than that of the basic method, but the methodsarecomplexandtakealongtimetocalculate,sothe training time increases, and the training speed decreases. training time is about 10s. By using the improved R-FCN algorithm for fault diagnosis, the proposed method uses In order to demonstrate the performance of the pro- posed method in training speed, it is compared with ref- OHEM method to train it, which can simplify the data processing process, the training speed is fast, and the erence[11],reference[13],andreference[16].'eresultsare shown in Figure 6. training time is about 5.5s. At the same time, the robot As can be seen from Figure 6, the training time of background is used for data analysis, which can reduce the reference [13] is the shortest, only about 5S. Because its transmission time of image information. intuitionistic fuzzy clustering algorithm based on spatial distributioninformationforimagerecognitionissimpleand easy to implement, the training speed is fast. Reference [13] 4.3. Comparative Analysis of Fault Diagnosis Accuracy. 'e accuracy of fault diagnosis is a key judgment index. combines discrete wavelet transform and support vector machinealgorithmtocompletefaultdiagnosis,and[16]uses 'e accuracy of the proposed method and the methods in reference [11], reference [13], and reference [16] for artiﬁcial neural network algorithm to classify faults. Both Journal of Robotics 9 Ref. [11] Ref. [13] Ref. [16] Proposed method Figure 6: Training time of diﬀerent methods. 0 1000 2000 3000 4000 Epoch Proposed method Ref. [13] Ref. [16] Ref. [11] Figure 7: Diagnostic accuracy of diﬀerent methods. fault diagnosis of distribution equipment is shown in diagnosis accuracy is about 80%. Reference [11] uses the Figure 7. traditional intuitionistic fuzzy clustering algorithm for As can be seen from Figure 7, compared with other graphic classiﬁcation. 'e traditional method is diﬃcult to methods, with the iteration of epoch, the fault diagnosis apply to a large number of distribution equipment, so the accuracy of the proposed method tends to be stable, about diagnosis accuracy is low. 92.06%. Due to its combination of robot and deep learning For the three fault types, the diagnostic accuracy of hybridmodel,itdeeplyextractsthecharacteristicsofvarious diﬀerent methods is shown in Table 4. typesoffaultequipmentfordiagnosis,whichfurtherensures It can be seen from Table 4 that the diagnostic accuracy the reliability of diagnosis results. Similarly, [16] uses arti- ofcomprehensiveheatingequipmentisgenerallylowerthan ﬁcial neural network algorithm for state recognition, but that of current heating equipment and voltage heating there is no eﬃcient way to obtain the equipment state, and equipment. Taking the proposed method as an example, the there is no complete database to support it. 'erefore, the diagnostic accuracy of comprehensive heating equipment is diagnosis accuracy is reduced by about 6% compared with 89.31%, and the other two types are higher than 90%. Be- cause the diagnostic criteria of comprehensive heating the proposed method. Reference [13] adopts the improved support vector machine algorithm of genetic algorithm for equipment are complex and easy to be confused, they aﬀect fault detection, which has a good eﬀect on high impedance the fault diagnosis. 'e recognition accuracy of current fault diagnosis, but its universality is not high, so the heating type defects is slightly higher, which may be due to Diagnostic accuracy (%) Training time (s) 10 Journal of Robotics Table 4: Comparison results of diagnostic accuracy of each fault type. Method Current heating type (%) Voltage heating type (%) Comprehensive heating type (%) Reference [11] 73.23 69.18 65.75 Reference [13] 82.84 80.36 79.04 Reference [16] 86.69 87.05 84.27 Proposed method 93.52 91.88 89.31 the obvious characteristics and large amount of data of defective power equipment can be segmented in a more infrared images of current heating type defects in the data complex background to further improve the recognition accuracy of the subsequent model. set. However, the diagnosis accuracy of the proposed method is higher than that of other comparison methods. Taking the current heating equipment as an example, its Data Availability diagnosis accuracy is as high as 93.52%, because it can well 'e data of the paper can be obtained from the corre- distinguish all kinds of faulty equipment by using the im- sponding author. proved R-FCN algorithm to learn the equipment image features and evaluate the fault level according to the fault Conflicts of Interest judgment. Other comparison methods only diagnose whether the equipment is faulty or not, but the diagnosis 'e authors declare no conﬂicts of interest. eﬀect is poor for various speciﬁc fault types. References 5. Conclusion [1] S. Gangolu, P. Raja, M. P. Selvan, and V. K. Murali, “Eﬀective algorithm for fault discrimination and estimation of fault Nowadays, the construction of smart grid in China has location in transmission lines,” IET Generation, Transmission entered a new stage of comprehensive and rapid develop- & Distribution, vol. 13, no. 13, pp. 2789–2798, 2019. ment. 'e traditional manual detection methods have been [2] N. Hu, H. Du, S. Liu, and Q. Lin, “Power equipment status diﬃculttodealwithalargeamountofinfraredimagedataof information parallel fault diagnosis of based on MapReduce,” power equipment. 'erefore, a fault diagnosis method of Journal of Computational Methods in Science and Engineering, distribution equipment based on the hybrid model of robot vol. 19, no. 88, pp. 1–6, 2019. and deep learning is proposed. 'e image information [3] E. Bashar, Q. Han, R. Wu, L. Ran, O. Alatise, and S. Jupe, database of distribution equipment based on robot in- “AnalysisofDCoﬀsetinfaultcurrentcausedbymachinesina spection is constructed, and the OTSU method is used to medium voltage distribution network,” Journal of Engineer- ing, vol. 2019, no. 17, pp. 3494–3499, 2019. extract the defect features of distribution equipment from [4] H. Tian, P. Liu, S. Zhou et al., “Research on the deterioration the binary image of equipment based on HSV space. 'en, process of electrical contact structure inside the±500kV theequipmentdefectcharacteristicsaresenttotheimproved converter transformer RIP bushings and its prediction R-FCN algorithm for learning and analysis to obtain the strategy,” IET Generation, Transmission & Distribution, fault type and location, and the fault level is obtained vol. 13, no. 12, pp. 2391–2400, 2019. through the calculation of fault criterion. 'e experimental [5] W.Luo,H.Wang,L.Wang,Z.Zhu,andH.Gao,“Faultedline results based on PyTorch platform show that: location method for distribution systems based on the equipment’s information exchange,” Dianli Xitong Baohu yu (1) Using the robot platform to build the image infor- Kongzhi/Power System Protection and Control, vol. 47, no. 4, mation database of distribution equipment can pp. 73–82, 2019. improve the accuracy of fault diagnosis of various [6] N.Narasimhulu,D.Kumar,andM.V.Kumar,“Detectionand equipment. 'e diagnostic accuracy of the proposed classiﬁcation of high impedance fault in power distribution method for current heating equipment, voltage system using hybrid technique,” Journal of Circuits, Systems, heating equipment, and comprehensive heating and Computers, vol. 29, no. 08, pp. 67–976, 2020. equipment is 93.52%, 91.88%, and 89.31%, [7] K. Chen, J. Hu, Y. Zhang, Z. Yu, and J. He, “Fault location in respectively. power distribution systems via deep graph convolutional networks,” IEEE Journal on Selected Areas in Communica- (2) Using OHEM method to train the improved R-FCN tions, vol. 38, no. 1, pp. 119–131, 2020. algorithm can shorten the model training time, [8] F. Wan, P. Madhika, J. Chwa, M. Mozumdar, and A. Ameri, improve the fault diagnosis eﬃciency, and further “Automatic optimal synthesis of aircraft electric power dis- improvethediagnosiseﬀect.'efaultdiagnosistime tribution system,” International Journal of Computing and and accuracy are 5.5s and 92.06%, respectively. Digital Systems, vol. 9, no. 3, pp. 363–375, 2020. [9] K. Jia, Q. Zhao, T. Feng, and T. Bi, “Distance protection Since the extraction degree of the defective region will scheme for DC distribution systems based on the high-fre- aﬀect the recognition accuracy of the subsequent model, quency characteristics of faults,” IEEE Transactions on Power considering the subsequent semantic segmentation of the Delivery, vol. 35, no. 1, pp. 234–243, 2020. overall infrared image by using the convolution neural [10] A. S. Alayande, I. K. Okakwu, O. E. Olabode, and network optimized by conditional random ﬁeld, the O. K. Nwankwoh, “Analysis of unsymmetricalfaults based on Journal of Robotics 11 artiﬁcial neural network using 11kV distribution network of [25] S.YamaneandK.Matsuo,“Adaptivecontrolbyconvolutional neural network in plasma arc welding system,” ISIJ Inter- University of lagos as case study,” Journal of Advances in Science and Engineering, vol. 4, no. 1, pp. 53–64, 2021. national, vol. 60, no. 5, pp. 998–1005, 2020. [26] M. Nabati, H. Navidan, R. Shahbazian, S. A. Ghorashi, and [11] F. Hu,H.Chen,and X.Wang, “An intuitionistickernel-based fuzzy C-means clustering algorithm with local information D. Windridge, “Using synthetic data to enhance the accuracy of ﬁngerprint-based localization: a deep learning approach,” for power equipment image segmentation,” IEEE Access, IEEE Sensors Letters, vol. 4, no. 4, pp. 1–4, 2020. vol. 8, no. 6, pp. 4500–4514, 2020. [27] M. S. Erbnescu, N. C. Manea, L. Streba et al., “Automated [12] H. Liu, Y. Wang, and W. Chen, “Anomaly detection for gleason grading of prostate cancer using transfer learning condition monitoring data using auxiliary feature vector and from general-purpose deep-learning networks,” Romanian density-based clustering,” IET Generation, Transmission & Journal of Morphology and Embryology, vol. 61, no. 1, Distribution, vol. 14, no. 1, pp. 108–118, 2020. pp. 149–155, 2020. [13] Youness, Mohammadnian, A. Turaj, and A. Soroudi, “Fault detection in distribution networks in presence of distributed generations using a data mining–driven wavelet transform,” IET Smart Grid, vol. 2, no. 2, pp. 163–171, 2019. [14] Q. Yang, J. Ruan, and Z. Zhuang, “Fault diagnosis of circuit breakers based on time-frequency and chaotic vibration analysis,” IET Generation, Transmission & Distribution, vol. 14, no. 7, pp. 1214–1221, 2020. [15] T. Zhang, H. Yu, P. Zeng, L. Sun, C. Song, and J. Liu, “Single phase fault diagnosis and location in active distribution network using synchronized voltage measurement,” Inter- national Journal of Electrical Power and Energy Systems, vol. 117, no. 5, 2020. [16] S.Frizzo-Stefenon,M.C.Silva,D.W.Bertol,L.H.Meyer,and A. Nied, “Fault diagnosis of insulators from ultrasound de- tection using neural networks,” Journal of Intelligent and Fuzzy Systems, vol. 37, no. 5, pp. 6655–6664, 2019. [17] B. Wang, M. Dong, M. Ren et al., “Automatic fault diagnosis of infrared insulator images based on image instance seg- mentation and temperature analysis,” IEEE Transactions on Instrumentation and Measurement, vol. 69, no. 8, pp. 5345– 5355, 2020. [18] H. Liang, Y. Liu, G. Sheng, and X. Jiang, “Fault-cause iden- tiﬁcation method based on adaptive deep belief network and time-frequency characteristics of travelling wave,” IET Gen- eration, Transmission & Distribution, vol. 13, no. 5, pp. 724– 732, 2019. [19] M. Gholami, A. Abbaspour, M. Moeini-Aghtaie, M. Fotuhi- Firuzabad,andM.Lehtonen,“Detectingthelocationofshort- circuit faults in active distribution network using PMU-based state estimation,” IEEE Transactions on Smart Grid, vol. 11, no. 2, pp. 1396–1406, 2020. [20] W.Hu,C.Ruan,H.Nian,andD.Sun,“Simpliﬁedmodulation scheme for open-end winding PMSM system with common DC bus under open-phase fault based on circulating current suppression,” IEEE Transactions on Power Electronics, vol. 35, no. 1, pp. 10–14, 2020. [21] K. Zhu and P. W. T. Pong, “Fault classiﬁcation of power distribution cables by detecting decaying DC components with magnetic sensing,” IEEE Transactions on Instrumenta- tion and Measurement, vol. 69, no. 5, pp. 2016–2027, 2020. [22] L. Romero, J. Blesa, V. Puig, G. Cembrano, and C. Trapiello, “First results in leak localization in water distribution net- worksusinggraph-basedclusteringanddeeplearning,” IFAC- PapersOnLine, vol. 53, no. 2, pp. 16691–16696, 2020. [23] W. Wang, N. Yu, Y. Gao, and J. Shi, “Safe oﬀ-policy deep reinforcement learning algorithm for volt-VAR control in power distribution systems,” IEEE Transactions on Smart Grid, vol. 11, no. 4, pp. 3008–3018, 2020. [24] D. A. Leo´n-Vargas, V. A. Bucheli-Guerrero, and H. A. Ordoez, “Solarradiationpredictiononphotovoltaicsystemsusingmachine learningtechniques,” Revista Facultad de Ingenier´ıa,vol.29,no.10, pp. 1–20, 2020. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Robotics Hindawi Publishing Corporation http://www.deepdyve.com/lp/hindawi-publishing-corporation/fault-diagnosis-method-of-distribution-equipment-based-on-hybrid-model-FgW5cq1bg6

Loading next page...

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

Publisher: Hindawi Publishing Corporation
Copyright: Copyright © 2022 Shan Rongrong et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN: 1687-9600
eISSN: 1687-9619
DOI: 10.1155/2022/9742815
Publisher site: See Article on Publisher Site

Abstract

Hindawi Journal of Robotics Volume 2022, Article ID 9742815, 11 pages https://doi.org/10.1155/2022/9742815 Research Article Fault Diagnosis Method of Distribution Equipment Based on Hybrid Model of Robot and Deep Learning 1 2 3 3 1 1 Shan Rongrong , Ma Zhenyu, Ye Hong, Lin Zhenxing, Qiu Gongming, Ge Chengyu, 1 1 Lu Yang, and Yu Kun NARI Group Co., Ltd, State Grid Electric Power Research Institute, Nanjing 210000, China Zhejiang Electric Power Corporation, Hangzhou 310013, China State Grid Wenzhou Power Supply Company Ouhai Power Supply Branch, Wenzhou 325000, China Correspondence should be addressed to Shan Rongrong; shanrongrong@sgepri.sgcc.com.cn Received 8 February 2022; Revised 14 March 2022; Accepted 19 March 2022; Published 14 April 2022 Academic Editor: Shan Zhong Copyright © 2022 Shan Rongrong et al. 'is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. In view of the poor eﬀect of most fault diagnosis methods on the intelligent recognition of equipment images, a fault diagnosis methodofdistributionequipmentbasedonthehybridmodelofrobotanddeeplearningisproposedtoreducethedependenceon manpower and realize eﬃcient intelligent diagnosis. Firstly, the robot is used to collect the on-site state images of distribution equipmenttobuildtheimageinformationdatabaseofdistribution equipment.At thesametime, therobot backgroundisused as the comprehensive database data analysis platform to optimize the sample quality of the database. 'en, the massive infrared images are segmented based on chroma saturation brightness space to distinguish the defective equipment images, and the defective equipment areas are extracted from the images by OTSU method. Finally, the residual network is used to improve the region-based fully convolutional networks (R-FCN) algorithm, and the improved R-FCN algorithm trained by the online hard example mining method is used for fault feature learning. 'e fault type, grade, and location of distribution equipment are obtained through fault criterion analysis. 'e experimental analysis of the proposed method based on PyTorch platform shows that the fault diagnosis time and accuracy are about 5.5s and 92.06%, respectively, which are better than other comparison methods and provide a certain theoretical basis for the automatic diagnosis of power grid equipment. be “allocated and used” [3]. 'e distribution equipment will 1. Introduction have some faults under long-term operation, resulting in With the continuous improvement of social economy and abnormal temperature. 'erefore, by detecting the tem- people’s living standards, the power demand is increasing perature of the distribution equipment, the thermal fault day by day, and the scale of power system is growing day by diagnosis of the distribution equipment can be carried out day. It includes transmission and transformation networks quickly, which plays a great role in the safe operation of the power grid. of various voltage levels. 'erefore, ensuring the safe and stable operation of complex power grid is an inevitable 'e infrared image of equipment is used for fault di- requirement to ensure the economic and social develop- agnosis with high eﬃciency, accurate judgment, safety, and ment. At the same time, the economic and social devel- reliability. At the same time, it is free from electromagnetic opmenthasagreatimpactonthesecurityandeconomy,and interference, fast detection speed, and no power failure of higher reliability is required [1, 2]. As the terminal of the live equipment. 'erefore, infrared diagnosis is widely used wholepowergridoperation,distributionnetwork isthepart in the ﬁeld of equipment fault monitoring and diagnosis with the widest coverage and the largest scale in China’s technology [4]. However, due to the characteristics of large powersystem, andit is thekey link toensure that powercan quantity and complex types of distribution equipment, if we 2 Journal of Robotics only rely on manual work in the process of data acquisition, breaker fault identiﬁcation. 'e image analysis process is analysis, and processing, the workload is relatively large, the complex, resulting in the reduction of fault identiﬁcation eﬃciency. Reference [15] proposes a method based on eﬃciencyislow,andtheaccuracy isrelativelylowdue tothe high dependence on manual experience. 'erefore, auto- feature model for single-phase grounding fault in active matic image acquisition and analysis of distribution distribution network system, which transforms the solution equipment are of great signiﬁcance to ensure the safety and of nonlinear feature model into single-objective optimiza- stability of distribution network [5]. tion of feature entropy, which can well identify single-phase In recent years, the automatic inspection technology of fault, but the identiﬁcation eﬀect of equipment with feature power distribution room has been popularized, and various type is not ideal. automaticrobotsandUAVshavemadegreatprogressinthe With the continuous development of computer original data acquisition stage. However, the accurate and technology and the rapid development of 5G communi- eﬃcient processing of collected image data is still in its cation technology, machine learning algorithm has been infancy.Howtoextractthefeaturesofinterestfrominfrared widely used this year, especially the deep learning algo- rithm has certain advantages in the ﬁeld of fault identi- images for power distribution equipment recognition is a problem to be solved [6]. Among them, the deep learning ﬁcation and classiﬁcation. Reference [16] proposes algorithmhasmadegreatachievementsinimageprocessing, artiﬁcial neural network algorithm to identify the insu- speechrecognition,andtextanalysis.Byestablishingadeep- lator stateand uses single-layerand multilayerperceptron seatedneuralnetwork,high-levelfeaturesareextractedfrom artiﬁcial intelligence algorithm to classify the conditions low-level features layer by layer, so as to achieve the eﬀect of of distribution insulators. 'is technology can make the target classiﬁcation and recognition [7]. Compared with the automatic inspection of electrical system more accurate manually designed feature extraction method, the distrib- and eﬃcient, but it lacks high reliable database for sup- uted features obtained by deep learning network model can port. Reference [17] proposed a Mask R convolution better express the essence of data [8, 9]. 'erefore, in order neural network method and used transfer learning and to improve the eﬃciency of thermal fault detection of dis- dynamic learning rate algorithm to realize eﬃcient rec- ognition of annotated image data sets, but it relied too tributionequipment,improvetheintelligenceofpowergrid, reduce the labor cost of detection, and reduce the false much on graphics annotation and lacked practical ap- detection rate, a fault diagnosis method of distribution plication value. In [18], appropriate traveling wave time- equipment based on the hybrid model of robot and deep frequency characteristic parameters of fault current are learning is proposed, which eﬀectively ensures the safe and selected as the input of adaptive depth belief network reliable operation of distribution equipment. model to obtain the fault type, but only considering the fault current characteristics as the basis, the reliability needs to be further improved. 2. Related Research Basedontheaboveanalysis,aimingattheproblemssuch as the complexity and diversity of smart grid distribution At present, there are many researches on fault diagnosis of distribution equipment at home and abroad, which can be equipment and the unsatisfactory eﬀect of most existing image recognition methods, a distribution equipment fault divided into traditional fault identiﬁcation and classiﬁcation methods and machine learning based identiﬁcation and diagnosis method based on robot and deep learning hybrid model is proposed. Its innovations are summarized as classiﬁcation methods [10]. Among them, the traditional fault identiﬁcation and classiﬁcation methods mainly in- follows: clude fuzzy clustering, discrete wavelet transform, and (1) In order to obtain the image information of distri- chaotic algorithm. For example, [11] proposed an infrared bution equipment more comprehensively, the pro- image segmentation algorithm based on intuitionistic fuzzy posed method introduces the robot to construct the clustering algorithm based on spatial distribution infor- corresponding image knowledge database, which mation, which is suitable for power equipment. It can well provides the basis for fault classiﬁcation and fault suppress the strong interference of nontarget objects in location. infrared image to image segmentation, but the method is (2) In order to locate the equipment defect area in the more traditional and has poor segmentation eﬀect for infrared image of distribution equipment, the pro- complex intelligent power grid equipment. Reference [12] posed method performs threshold segmentation on proposed an anomaly detection method based on spatial the infrared image in hue saturation value (HSV) clustering applied by auxiliary feature vector and density space and uses OTSU method to extract the noise. 'e auxiliary feature vector of each conditional equipmentdefectarea,soas toimprovetheaccuracy variable is constructed for clustering to identify normal data of subsequent fault diagnosis. patterns and diﬀerent types of anomalies. Reference [13] (3) Aiming at the problem that the deep learning al- proposed a data mining driving scheme based on discrete gorithm is prone to gradient disappearance and wavelet transform to realize high impedance fault detection gradient explosion, the proposed method uses the in active distribution network, but the universality of the residual network to improve the region-based fully method is not high. Reference [14] proposed a method to convolutional networks (R-FCN) algorithm and obtain the vibration characteristics of circuit breaker based appliesittothelearningoffaultyequipment,soasto on time-frequency and chaos analysis to realize circuit Journal of Robotics 3 obtainthe fault type and location with high accuracy Robot and further improve the safety of equipment. Defect backstage system 3. Proposed Method 3.1. Construction of Image Information Database of Distri- System data Fault information Model base analysis center bution Equipment Based on Robot Inspection. 'etraditional knowledge base equipment status is usually determined by manual analysis. 'e workload is huge and error-prone, which aﬀects the judgment of system status, resulting in potential safety Knowledge database hazards. 'erefore, the robot is used for patrol inspection to Database obtain the status image of distribution equipment and build the corresponding information base for the analysis of Production Robot data system equipment status, so as to ﬁnd the faulty equipment in time Online and ensure the reliable operation of power grid [19]. 'e monitoring construction process of distribution equipment image in- formation base based on robot inspection is shown in Figure 1: Construction process of image information database for Figure 1. distribution equipment. 'e basic data sources of the database mainly include production system, online monitoring system, and robot (1) 'e data of the three-party platform includes the backgroundinspectionsystem.'erelevant data ofthestate information required in the database structure table. quantity of power equipment mainly comes from the power After eliminating the redundant information, the production management system (PMS), which can provide integrated data in a uniﬁed format can be obtained, the real-time operation condition, historical operation state, and the defect alarm data can be located and re- historical maintenance record, historical test data, equip- trieved quickly. ment account, equipment parameters, and other informa- tionoftheequipment.'eonlinemonitoringsystemmainly (2) 'e fault information base is mainly taken from the defectsystemrecordsandcontainsalargenumberof relies on various sensors on each power equipment for real- time monitoring. 'e robot background inspection system relevant equipment fault cases, including fault characteristics, solutions, expert opinions, and can not only provide the observation of some state quan- manufacturer records. At the same time, the tities, but also carry out corresponding state evaluation and maintenance record database and equipment ac- analysis for diﬀerent equipment states according to the count database are used to build a comprehensive automatic state evaluation system. In addition, the data databaseoffaultinformation,soastoscreenthefault composition of the system includes infrared temperature inspection points. measurement, visible light reading, and telemetry reading. 'e robot inspection cycle generally refers to the in- (3) 'eknowledgeinformationbaseistheengineforthe spection plan formulated by the distribution network op- system to evaluate the equipment status and judge eration inspection center, and two inspection robots the fault. 'e internal rules at all levels provide the complete the tasks of infrared temperature measurement logicalbasisforthesystemtojudgethefault.'ekey and data transcription of equipment in the area [20]. At the is knowledge acquisition, that is, collecting and same time, the robot background uses the threshold out of mining the knowledge at all levels to enrich the limit judgment method to automatically evaluate the knowledge base. equipmentstatus.Inordertoensuresuﬃcientchargingtime of the robot and avoid the daily patrol and infrared tem- perature measurement period, the special patrol at night is 3.2. Defect Feature Extraction of Distribution Equipment. set in the nonbusy working period of the robot every day, When extracting the defect features of distribution equip- with the upper limit of one time. 'e data reports collected ment, it is necessary to perform threshold segmentation on by the special patrol at night and infrared temperature theinfraredimageinHSVspace,separatetheinfraredimage measurement are included in the database for screening and background from irrelevant equipment and defective preprocessing. equipment, and then extract the equipment defect area [21]. In addition, the background of the inspection robot is equipped with a system server, which includes data analysis software terminal, data exchange server, data storage server, 3.2.1. OTSU 4reshold Segmentation. OTSUisconsideredto dataoperationserver,andothermodules.'edataexchange be one of the best algorithms in image threshold segmen- server is responsible for collecting and classifying the pro- tation. 'e threshold segmentation process of OTSU algo- rithm is as follows: ﬁrstly, the image is processed in gray duction system, online monitoring system, and robot patrol data into the storage server. 'ere are three-party databases, level, the number of pixels in the whole image is counted, fault information base and knowledge information base in and the probability distribution of each pixel in the whole the data storage server. image is calculated; then, the gray level is traversed and 4 Journal of Robotics searchedinthewholeimage,andtheinterclassprobabilityof classiﬁcation [22]. 'e ﬂow of HSV based defect region the image foreground and background at the current gray extraction algorithm is shown in Figure 2. leveliscalculated;ﬁnally,thethresholdcorrespondingtothe When processing the infrared image of defective variance between classes and within classes is calculated by equipment, ﬁrst merge the similar pixels corresponding to the given objective function. the area with the same temperature, and segment the image Suppose there are D gray levels in the image, in which according to the threshold of the three components of the the number of pixels with gray value of i is N and the total defective areaintheHSV colorspace toextract thedefective number of pixels in the image is N. 'en, the average gray area.'en,thediscretedefectregionsareconnectedthrough value of the whole image is the closed operation in mathematical morphology, and the threshold segmentation of the original image is carried out D−1 by OTSU method to separate the power equipment and the μ � 􏽘 i . (1) background region. Finally, the defect area is found in the i�0 binary image separated by OTSU method; that is, the de- According to the gray characteristics of the image, the fective power equipment is separated from other areas, so as image is divided into foreground B and background B . 0 1 to achieve the purpose of extracting defective power 'en, p (T) and p (T) represent the probability of oc- 0 1 equipment and facilitate the identiﬁcation and diagnosis of currence of foreground B and background B when the 0 1 power equipment types and fault types. threshold is T, respectively. 'e calculation is as follows: p (T) � 􏽘 􏼒 􏼓, 3.3. Fault Diagnosis of Distribution Equipment Based on Deep (2) i�0 Learning Hybrid Model p (T) � 1 − p (T). 1 0 3.3.1. Defect Training Based on Deep Learning Hybrid Model. R-FCN algorithm architecture mainly includes backbone 'en,themeanvaluesofforeground B andbackground network, region proposal network (RPN), and region of B are interest (ROI) subnet [23]. When fault diagnosis of power ⎧ ⎪ 􏽐 i N /N 􏼁 distribution equipment is carried out, ﬁrst input the col- ⎪ i�0 μ (T) � , ⎪ 0 ⎪ lected infrared image of power equipment into convolution p (T) neural network and extract the convolution feature map of (3) ⎪ infrared image. In this process, deeper and more abstract ⎪ T μ − 􏽐 i N /N􏼁 i�0 ⎪ 􏽐 image features can be extracted by using a larger backbone ⎩ μ (T) � . network (ResNet 101) to improve the recognition accuracy p (T) [24]. 'en, the feature map is sent to the RPN network to 'e interclass variance with threshold T in the gray generate anchors, which are marked with foreground and histogram is calculated as follows: background, and the foreground area with high score is 2 2 selectedastherecommendedareaROIs.'eseROIsaresent to the ROI subnet for further training, and 300 recom- σ (T) � p (T)􏼢μ (T) − μ 􏼣 + p (T)􏼢μ (T) − μ 􏼣 . 0 1 B 0 1 􏽘 􏽘 mended windows are generated for each infrared image of (4) power equipment. At the same time, the characteristic map of the full convolution layer is calculated with the multilayer 'e optimal threshold is deﬁned as the T value corre- convolution kernel to generate a position sensitive score sponding to the maximum variance between classes, which map. 'e ROI and Score Maps are input into the later is calculated as Softmax layer for vote. 'rough the Softmax layer for 2 2 classiﬁcation, the ROI with the highest score is ﬁnally ob- σ (T) � max 􏽮σ (T)􏽯. (5) B B 0≤T≤D−1 tained,thatis,thelocationandtypeoftheobjectlocatedand recognized. 'e architecture of R-FCN algorithm is shown in Figure 3. 3.2.2. Defect Region Extraction Based on HSV. In order to (1) Residual Network. When the depth of the deep learning improve the accuracy of equipment fault image classiﬁca- network reaches a certain degree, the problems of gradient tion,thedefectregionandbackgroundintheinfraredimage disappearance and gradient explosion often appear during of fault power equipment are separated by using the defect training.Inordertosolvethisproblem,theresidualnetwork region segmentation algorithm based on HSV. Since it is (ResNet) is used to improve the R-FCN algorithm; that is, impossible to determine the defect type only by analyzing the residual network is selected as the backbone network. thefaultarea,itisnecessarytosegmentthedefectareabased 'e residual element is essentially the mapping residual on mathematical morphology according to the location of required for ﬁtting through these stacked layers. Suppose the defect area. 'rough this method, the defective power thatthenetworkmappingis H(x)andtheresidualmapping equipment and the background area in the infrared image function of the network is F(x), F(x) � H(x) − x. 'e so- are separated, so as to reduce the interference of the called residual is the diﬀerence between the observed value background area in the infrared image on the defect type Journal of Robotics 5 it only needs to make F(x) � 0 to get H(x) � x, so as to Start avoid the disappearance and explosion of gradient. 'e input x and output x of the m-th residual unit m m+1 Initialize the image clustering center point so that the distance S of are expressed as follows: each clustering center point is evenly distributed in the image. x � f h x + δ x , ω , 􏼁 􏼁 􏼁 m+1 m m m M−1 In each 3 × 3 select the pixel with the smallest gradient in the (6) neighborhood as the clustering center in the neighborhood. x � x + 􏽘 δ x , ω􏼁 , M m i i i�1 The distance metric D is calculated by calculating the distance where δ is the ReLu activation function, ω is the weight metric from each pixel to the cluster center in the super-pixel between each unit, m, M respectively represent the shallow segmentation algorithm, and the smallest distance metric is selected to assign pixels to the corresponding cluster center. residual unit and deep residual unit, h(·) represents the identity mapping, and x is the ﬁnal output of the residual unit. Normalize the processed image and update the H value. Inordertolearnmoreandmoreabstractimagefeatures, the proposed method selects ResNet 101 network, and its conﬁguration is shown in Table 1. Set the threshold. If the three components of HSV meet 10 < H < 150, 10 < S < 100, 200 < V < 255, then H = 0, S = 0, V = 0. The equipment defect area in the infrared image is separated through the threshold. (2) RPN Network.'einputofRPNnetworkisimagefeature graph.Accordingtoanchormechanism,9rectangularboxes with diﬀerent sizes are generated for each point. When OTSU threshold segmentation algorithm is used to segment training the RPN network, compare the anchor with the the original image to obtain the binary image I . manually calibrated true value area in the data set, mark the anchor frame with the largest overlap ratio as the fore- ground, and mark the anchor frame with an overlap ratio Because there may be one or several defect points in the segmented binary image, each isolated defect point is connected according to the greater than 0.7 as the foreground sample. Mark the anchor closed operation of mathematical morphology to determine the defect box whose overlap ratio is less than 0.3 as the background region R . sample.Selectthepositiveandnegativesamplesofanchorin proportion, use the maximum suppression method (NMS) Find each connected region R in the binary image I . If the defective 0 b and other methods to screen the top 250 ROIs with the region is R ∩R > 0.8R , this region is the defective power equipment region 0 s s to be extracted. highest score, and send these preliminarily screened pre- selected frames to the ROI subnet. In addition, RPN network adopts anchor mechanism, End which not only solves the problem of translation invariance, but also enables R-FCN algorithm to identify and locate Figure 2: Defect region extraction process based on HSV. targets with diﬀerent overall dimensions. In the actual process of infrared image recognition of power distribution equipment, due to diﬀerent equipment with diﬀerent shape and structure, diﬀerent sizes and variable aspect ratio, in Input image Position sensitive convolution order to ensure that there are targets in the receptive ﬁeld k (c+1) ROI subnet corresponding to each sliding window on the feature map, conv multiscale anchor is required to ensure that the candidate ROI frame is as complete as possible to select the target [25]. In Softmax pool vote conv Result the implementation of RPN network anchor, multiscale output anchor can be obtained by setting the area of reference c+1 k window (base_size), diﬀerent area multiples and anchor aspect ratio, so that RPN can give more accurate foreground Convolution characteristic recommendation area. RPN graph ROIs (3) ROI Subnet.'eroleofROIsubnetistocorrecttheROIs conv locationobtainedfrom RPNnetwork,soastoobtainamore accuratetargetlocationofpowerequipment,soastoidentify ROI.Apositionsensitiveconvolutionlayerisaddedafterthe Figure 3: Architecture of R-FCN algorithm. last layer of the full convolution network, which can realize the translation variability of the algorithm and output the H(x) and the estimated value x. 'e advantage of ResNet (c + 1)-dimensional position sensitive fractional graph. networkisthatitusesthestackinglayertoﬁt H(x)togetthe 'e RPN network ﬁlters out the characteristic map of ROI mapping H(x) � F(x) + x. 'e advantage of this repre- with size a × b, which is divided into k × k parts (bin) and sentation is that if the model has been ﬁtted to the beststate, convoluted with k (c + 1) convolution cores; that is, each 6 Journal of Robotics Table 1: Structure table of ResNet 101. Layer name 101-Layer conv1 7 7, stride 2 conv2_x 3 3max pool, stride 2 1 ×1 64 ⎡ ⎢ ⎤ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎣3 ×3 64 ⎥ ⎦ × 3 1 ×1 256 1 ×1 128 ⎢ ⎥ ⎡ ⎢ ⎤ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ conv3_x ⎣3 ×3 128 ⎦ × 4 1 ×1 512 1 ×1 256 ⎡ ⎢ ⎤ ⎥ ⎢ ⎢ ⎥ ⎥ ⎢ ⎥ ⎢ ⎥ conv4_x ⎣3 ×3 256 ⎦ × 23 1 ×1 1024 1 ×1 512 ⎡ ⎢ ⎤ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ conv5_x ⎣ 3 ×3 512 ⎦ × 3 1 ×1 2048 Average pool, 1000-D FC, softmax part is mapped to a score map, in which (c +1) is the Classification number of categories plus the background to obtain the 2 loss L cls k (c + 1)-layer position sensitive score map. 'e number of channels of bin in diﬀerent positions is a × b × (c +1), and c+1 vote ROI the score of class (c + 1) in this position is stored. After the pool poolingprocessiscompleted,voteontheROI.Sumthe k × k Regression parts bin to get the output of (c + 1) dimension, that is, the loss L reg probability of category (score), and classify it through the softmax layer. At this time, the output result is the posi- ROI cls pool tioning coordinates and type of the object. vote c+1 (4) OHEM. In RPN network, a large number of rectangular reg boxes are generated, and hundreds or thousands of regions will participate in the training of predicting target categories Hard RoI and locations. 'e proportion of power equipment is small, Sampler so the ratio of equipment background area to target area is Figure 4: Training framework of OHEM method. too large, resulting in sample imbalance, which makes it diﬃcult to identify distribution equipment. 'erefore, online hard example mining (OHEM) method is used to backpropagated to the network and trained again to update trainthenetworkmodel[26].Duringtraining,whenthereis theweightofthewholenetwork,andthepenaltyforlowloss a preselected area with large loss, the hard example can be samples is ignored. Among them, the weight update is re- trained and classiﬁed again, which can solve the problem of alized based on the random gradient descent method. imbalance between positive and negative sample categories and improve the accuracy of infrared image recognition model of power equipment. 'e training framework of 3.3.2. Fault Diagnosis of Distribution Equipment. 'ere are OHEM method is shown in Figure 4. many kinds of equipment in the distribution network, such WhenOHEMcarriesoutspeciﬁctraining,ﬁrstly,ResNet as circuit breaker, potential transformer (PT), current 101 is used to extract the image features of training samples, transformer(CT),lightningarrester,andtransformer.'ese train the image classiﬁcation and positioning branches, and equipment can be classiﬁed into current heating type, calculatetheclassiﬁcationloss L andregressionloss L of voltage heating type, and comprehensive heating type cls reg each target. 'e loss function is calculated as follows: according to the heating factors. Diﬀerent diagnostic methods are selected for diﬀerent types of equipment, and ∗ ∗ L􏼐s, d 􏼑 � L s 􏼁 + ζ c >0 􏼁 L d, d 􏼁 , (7) x,y,a,b cls c reg diﬀerent diagnostic methods have diﬀerent diagnostic cri- teria. 'e defects of the equipment, such as general defects, where d istheupperleftcoordinateofthetargetarea, d is x,y a serious defects, and critical defects, are determined through thewidthofthetargetarea, d istheheightofthetargetarea, various diagnostic methods. ζ is the balance coeﬃcient of classiﬁcation loss and re- gression loss, and s is the image type. (1) 'e heating of current heating equipment is mainly 'en, sort according to the loss value from high to low, due to the current thermal eﬀect. Generally, the and use the NMS method to select the ﬁrst Ω samples with surface temperature judgment method and relative the largest loss value to screen out the diﬃcult cases of this temperature diﬀerence judgment method are used round of samples. Finally, these diﬃcult samples are for fault diagnosis. 'e relative temperature Position sensitive convolution Journal of Robotics 7 diﬀerence ∇ andthemaximumsurfacetemperature PyTorch v1.3 version, the programming language is Python T of diﬀerent defect degrees of each distribution 3.6.0, and third-party dependent libraries such as Open CV max equipment are shown in Table 2. 4.0 and NumPy 1.3 are used for batch processing of data. At the same time, the network model is trained on the GPU of (2) Voltage heating equipment is mainly due to voltage dual card Tesla P100, with a total video memory of 8GB. In eﬀect. 'e main equipment categories include zinc this way, large batch data can be set during training to oxide arrester, high-voltage bushing, and coupling improve the convergence speed of the model. Other hard- capacitor. 'is kind of equipment generally adopts ware environments are as follows: 512GB of memory re- imagefeaturejudgmentmethod,similarcomparison sources and 1TB of hard disk. When processing data and I/ judgment method, and comprehensive analysis O operations on a large scale, this can realize parallel judgment method. Using the thermal image char- processing and high speed and ensure the training re- acteristics of the equipment and the image feature quirements of network model. judgment method, the fault can be found quickly. If the similar comparison discrimination method and comprehensive analysis judgment method are used, 4.1. Experimental Data Set. 'e research scenario is the the temperature diﬀerence ΔT shall be taken as the power equipment in the distribution network, so it is fault diagnosis index, in which the ΔT of zinc oxide necessarytocollectthepicturematerialsofequipmentfaults arrester is 0.5∼1K, and the ΔT of high-voltage in the distribution system and then construct a ﬁxed format bushing, coupling capacitor, oil immersed PT and data set for test training. 'rough safety training and pro- CT are 2∼3K. fessional leadership, use mobile phones, cameras, and other (3) Comprehensive heating equipment needs to be di- equipment to shoot at the site of power distribution equipment. 'e scene of abnormal power grid equipment is agnosed in combination with the diagnosis methods of voltage heating equipment and current heating selected,andalargenumberofpositivesamplesarecollected in multiple directions according to the shooting angle of equipment, mainly including insulators, generators, and transformers. In the actual thermal fault diag- video monitoring. In order to ensure the rationalization of data distribution, all types and forms of equipment anomaly nosis, the fault diagnosis indexes of various methods should be combined to improve the eﬃciency and types are covered in the acquisition process. After that, the accuracy of fault diagnosis [27]. abnormal categories and areas of power grid equipment are marked through the open-source and free wizard marking According to diﬀerent types of distribution equipment, software. the corresponding thermal fault diagnosis and judgment 'e data set contains 2580 on-site abnormal images of methods can be selected, the corresponding parameters can RGBpowergridequipmentduringtheday,6300markboxes be calculated, and the thermal fault diagnosis of distribution intotal,and2769on-siteabnormalimagesofinfraredpower equipment can be carried out according to the diagnosis grid equipment at night, 7000 mark boxes in total, all from criteria. 'e fault diagnosis process of power distribution the real power distribution room, power equipment plant, equipment is shown in Figure 5. etc. After the data annotation is completed, the annotation Firstly, the infrared image of distribution equipment is ﬁle in XML format is generated, which corresponds to the input into the detection model for image preprocessing, and real image. the defect area is extracted based on HSV to divide the structure of distribution equipment. 'en,the trained depth 4.2. Comparative Analysis of Training Speed. 'e proposed learninghybridmodelisusedtoobtainthetypeandlocation ofthetargetequipment,andthetemperatureinformationof method combines the hybrid model of robot and deep eachstructuralareaontheinfraredimageofthedistribution learning for equipment fault diagnosis. ResNet network is equipmentis readat thesametime.Finally,accordingtothe usedtooptimizetheR-FCNalgorithm,andthedefectareais selected fault diagnosis method, the thermal fault state, extracted based on HSV to improve the diagnosis eﬀect. In thermalfault level,andthermalfault locationof distribution order to demonstrate the improvement eﬀect of the pro- equipment are determined by using the diagnosis criterion. posed method, it is compared with the diagnosis methods of ResNet network, Otsu threshold segmentation, and OHEM Diﬀerent from the existing fault diagnosis methods of ar- tiﬁcial equipment, the proposed method can use the deep training.'ediagnosisaccuracyandtrainingtimeareshown in Table 3. learning hybrid model to realize the intelligent classiﬁcation and fault type diagnosis of infrared images of distribution It can be seen from Table 3 that extracting deeper equipment fault features using ResNet network can greatly equipment, greatly reduce the workload of inspectors, and improve the automation level of fault diagnosis of distri- improve the diagnosis accuracy, which is 9.88% higher than bution equipment. thatofthemodel.However,duetothedeepeningofnetwork layers, the training time is also increased, more than 10s. At thesametime,byintegratingOHEMtrainingdepthlearning 4. Experiment and Analysis model, the diagnostic accuracy continued to improve by 'e experiment is carried out in the 64 bit operating system 4.41%. Due to the simple and easy implementation of the training process, the training time is only increased by 0.7s. environment of Ubuntu 6.04.4 LTS, in which the deep learning hybrid model uses the deep learning framework It can be seen that the diagnostic accuracy of the proposed 8 Journal of Robotics Table 2: Fault diagnosis criterion of current heating equipment. General defect Serious defect Emergency defect Circuit breaker 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 55%≤ T ≤80% T >80% max max Disconnecting switch 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 90%≤ T ≤130% T >130% max max CT 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 55%≤ T ≤80% T >80% max max Capacitor 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 55%≤ T ≤80% T >80% max max High voltage bushing 35%≤ ∇ <80% 80%≤∇ <95% ∇ ≥95% T T T — 55%≤ T ≤80% T >80% max max Gets the type and Infrared image input location of the target device Extract the original temperature of the target Image preprocessing device Select diagnostic Extracting defect method region based on HSV Calculate the diagnosis index of Structural region each structural area and division (image determine the fault level. segmentation) Trained deep learning Output results model Figure 5: Fault diagnosis process of distribution equipment. Table 3: Fault diagnosis criterion of current heating equipment. Method Accuracy/% Training time/s Original model 82.49 4.6 Original model+OTSU threshold segmentation 85.82 5.9 Original model+OTSU threshold segmentation+ResNet network 92.37 10.5 Original model+OTSU threshold segmentation+ResNet network+OHEM 96.78 11.2 method is higher than that of the basic method, but the methodsarecomplexandtakealongtimetocalculate,sothe training time increases, and the training speed decreases. training time is about 10s. By using the improved R-FCN algorithm for fault diagnosis, the proposed method uses In order to demonstrate the performance of the pro- posed method in training speed, it is compared with ref- OHEM method to train it, which can simplify the data processing process, the training speed is fast, and the erence[11],reference[13],andreference[16].'eresultsare shown in Figure 6. training time is about 5.5s. At the same time, the robot As can be seen from Figure 6, the training time of background is used for data analysis, which can reduce the reference [13] is the shortest, only about 5S. Because its transmission time of image information. intuitionistic fuzzy clustering algorithm based on spatial distributioninformationforimagerecognitionissimpleand easy to implement, the training speed is fast. Reference [13] 4.3. Comparative Analysis of Fault Diagnosis Accuracy. 'e accuracy of fault diagnosis is a key judgment index. combines discrete wavelet transform and support vector machinealgorithmtocompletefaultdiagnosis,and[16]uses 'e accuracy of the proposed method and the methods in reference [11], reference [13], and reference [16] for artiﬁcial neural network algorithm to classify faults. Both Journal of Robotics 9 Ref. [11] Ref. [13] Ref. [16] Proposed method Figure 6: Training time of diﬀerent methods. 0 1000 2000 3000 4000 Epoch Proposed method Ref. [13] Ref. [16] Ref. [11] Figure 7: Diagnostic accuracy of diﬀerent methods. fault diagnosis of distribution equipment is shown in diagnosis accuracy is about 80%. Reference [11] uses the Figure 7. traditional intuitionistic fuzzy clustering algorithm for As can be seen from Figure 7, compared with other graphic classiﬁcation. 'e traditional method is diﬃcult to methods, with the iteration of epoch, the fault diagnosis apply to a large number of distribution equipment, so the accuracy of the proposed method tends to be stable, about diagnosis accuracy is low. 92.06%. Due to its combination of robot and deep learning For the three fault types, the diagnostic accuracy of hybridmodel,itdeeplyextractsthecharacteristicsofvarious diﬀerent methods is shown in Table 4. typesoffaultequipmentfordiagnosis,whichfurtherensures It can be seen from Table 4 that the diagnostic accuracy the reliability of diagnosis results. Similarly, [16] uses arti- ofcomprehensiveheatingequipmentisgenerallylowerthan ﬁcial neural network algorithm for state recognition, but that of current heating equipment and voltage heating there is no eﬃcient way to obtain the equipment state, and equipment. Taking the proposed method as an example, the there is no complete database to support it. 'erefore, the diagnostic accuracy of comprehensive heating equipment is diagnosis accuracy is reduced by about 6% compared with 89.31%, and the other two types are higher than 90%. Be- cause the diagnostic criteria of comprehensive heating the proposed method. Reference [13] adopts the improved support vector machine algorithm of genetic algorithm for equipment are complex and easy to be confused, they aﬀect fault detection, which has a good eﬀect on high impedance the fault diagnosis. 'e recognition accuracy of current fault diagnosis, but its universality is not high, so the heating type defects is slightly higher, which may be due to Diagnostic accuracy (%) Training time (s) 10 Journal of Robotics Table 4: Comparison results of diagnostic accuracy of each fault type. Method Current heating type (%) Voltage heating type (%) Comprehensive heating type (%) Reference [11] 73.23 69.18 65.75 Reference [13] 82.84 80.36 79.04 Reference [16] 86.69 87.05 84.27 Proposed method 93.52 91.88 89.31 the obvious characteristics and large amount of data of defective power equipment can be segmented in a more infrared images of current heating type defects in the data complex background to further improve the recognition accuracy of the subsequent model. set. However, the diagnosis accuracy of the proposed method is higher than that of other comparison methods. Taking the current heating equipment as an example, its Data Availability diagnosis accuracy is as high as 93.52%, because it can well 'e data of the paper can be obtained from the corre- distinguish all kinds of faulty equipment by using the im- sponding author. proved R-FCN algorithm to learn the equipment image features and evaluate the fault level according to the fault Conflicts of Interest judgment. Other comparison methods only diagnose whether the equipment is faulty or not, but the diagnosis 'e authors declare no conﬂicts of interest. eﬀect is poor for various speciﬁc fault types. References 5. Conclusion [1] S. Gangolu, P. Raja, M. P. Selvan, and V. K. Murali, “Eﬀective algorithm for fault discrimination and estimation of fault Nowadays, the construction of smart grid in China has location in transmission lines,” IET Generation, Transmission entered a new stage of comprehensive and rapid develop- & Distribution, vol. 13, no. 13, pp. 2789–2798, 2019. ment. 'e traditional manual detection methods have been [2] N. Hu, H. Du, S. Liu, and Q. Lin, “Power equipment status diﬃculttodealwithalargeamountofinfraredimagedataof information parallel fault diagnosis of based on MapReduce,” power equipment. 'erefore, a fault diagnosis method of Journal of Computational Methods in Science and Engineering, distribution equipment based on the hybrid model of robot vol. 19, no. 88, pp. 1–6, 2019. and deep learning is proposed. 'e image information [3] E. Bashar, Q. Han, R. Wu, L. Ran, O. Alatise, and S. Jupe, database of distribution equipment based on robot in- “AnalysisofDCoﬀsetinfaultcurrentcausedbymachinesina spection is constructed, and the OTSU method is used to medium voltage distribution network,” Journal of Engineer- ing, vol. 2019, no. 17, pp. 3494–3499, 2019. extract the defect features of distribution equipment from [4] H. Tian, P. Liu, S. Zhou et al., “Research on the deterioration the binary image of equipment based on HSV space. 'en, process of electrical contact structure inside the±500kV theequipmentdefectcharacteristicsaresenttotheimproved converter transformer RIP bushings and its prediction R-FCN algorithm for learning and analysis to obtain the strategy,” IET Generation, Transmission & Distribution, fault type and location, and the fault level is obtained vol. 13, no. 12, pp. 2391–2400, 2019. through the calculation of fault criterion. 'e experimental [5] W.Luo,H.Wang,L.Wang,Z.Zhu,andH.Gao,“Faultedline results based on PyTorch platform show that: location method for distribution systems based on the equipment’s information exchange,” Dianli Xitong Baohu yu (1) Using the robot platform to build the image infor- Kongzhi/Power System Protection and Control, vol. 47, no. 4, mation database of distribution equipment can pp. 73–82, 2019. improve the accuracy of fault diagnosis of various [6] N.Narasimhulu,D.Kumar,andM.V.Kumar,“Detectionand equipment. 'e diagnostic accuracy of the proposed classiﬁcation of high impedance fault in power distribution method for current heating equipment, voltage system using hybrid technique,” Journal of Circuits, Systems, heating equipment, and comprehensive heating and Computers, vol. 29, no. 08, pp. 67–976, 2020. equipment is 93.52%, 91.88%, and 89.31%, [7] K. Chen, J. Hu, Y. Zhang, Z. Yu, and J. He, “Fault location in respectively. power distribution systems via deep graph convolutional networks,” IEEE Journal on Selected Areas in Communica- (2) Using OHEM method to train the improved R-FCN tions, vol. 38, no. 1, pp. 119–131, 2020. algorithm can shorten the model training time, [8] F. Wan, P. Madhika, J. Chwa, M. Mozumdar, and A. Ameri, improve the fault diagnosis eﬃciency, and further “Automatic optimal synthesis of aircraft electric power dis- improvethediagnosiseﬀect.'efaultdiagnosistime tribution system,” International Journal of Computing and and accuracy are 5.5s and 92.06%, respectively. Digital Systems, vol. 9, no. 3, pp. 363–375, 2020. [9] K. Jia, Q. Zhao, T. Feng, and T. Bi, “Distance protection Since the extraction degree of the defective region will scheme for DC distribution systems based on the high-fre- aﬀect the recognition accuracy of the subsequent model, quency characteristics of faults,” IEEE Transactions on Power considering the subsequent semantic segmentation of the Delivery, vol. 35, no. 1, pp. 234–243, 2020. overall infrared image by using the convolution neural [10] A. S. Alayande, I. K. Okakwu, O. E. Olabode, and network optimized by conditional random ﬁeld, the O. K. Nwankwoh, “Analysis of unsymmetricalfaults based on Journal of Robotics 11 artiﬁcial neural network using 11kV distribution network of [25] S.YamaneandK.Matsuo,“Adaptivecontrolbyconvolutional neural network in plasma arc welding system,” ISIJ Inter- University of lagos as case study,” Journal of Advances in Science and Engineering, vol. 4, no. 1, pp. 53–64, 2021. national, vol. 60, no. 5, pp. 998–1005, 2020. [26] M. Nabati, H. Navidan, R. Shahbazian, S. A. Ghorashi, and [11] F. Hu,H.Chen,and X.Wang, “An intuitionistickernel-based fuzzy C-means clustering algorithm with local information D. Windridge, “Using synthetic data to enhance the accuracy of ﬁngerprint-based localization: a deep learning approach,” for power equipment image segmentation,” IEEE Access, IEEE Sensors Letters, vol. 4, no. 4, pp. 1–4, 2020. vol. 8, no. 6, pp. 4500–4514, 2020. [27] M. S. Erbnescu, N. C. Manea, L. Streba et al., “Automated [12] H. Liu, Y. Wang, and W. Chen, “Anomaly detection for gleason grading of prostate cancer using transfer learning condition monitoring data using auxiliary feature vector and from general-purpose deep-learning networks,” Romanian density-based clustering,” IET Generation, Transmission & Journal of Morphology and Embryology, vol. 61, no. 1, Distribution, vol. 14, no. 1, pp. 108–118, 2020. pp. 149–155, 2020. [13] Youness, Mohammadnian, A. Turaj, and A. Soroudi, “Fault detection in distribution networks in presence of distributed generations using a data mining–driven wavelet transform,” IET Smart Grid, vol. 2, no. 2, pp. 163–171, 2019. [14] Q. Yang, J. Ruan, and Z. Zhuang, “Fault diagnosis of circuit breakers based on time-frequency and chaotic vibration analysis,” IET Generation, Transmission & Distribution, vol. 14, no. 7, pp. 1214–1221, 2020. [15] T. Zhang, H. Yu, P. Zeng, L. Sun, C. Song, and J. Liu, “Single phase fault diagnosis and location in active distribution network using synchronized voltage measurement,” Inter- national Journal of Electrical Power and Energy Systems, vol. 117, no. 5, 2020. [16] S.Frizzo-Stefenon,M.C.Silva,D.W.Bertol,L.H.Meyer,and A. Nied, “Fault diagnosis of insulators from ultrasound de- tection using neural networks,” Journal of Intelligent and Fuzzy Systems, vol. 37, no. 5, pp. 6655–6664, 2019. [17] B. Wang, M. Dong, M. Ren et al., “Automatic fault diagnosis of infrared insulator images based on image instance seg- mentation and temperature analysis,” IEEE Transactions on Instrumentation and Measurement, vol. 69, no. 8, pp. 5345– 5355, 2020. [18] H. Liang, Y. Liu, G. Sheng, and X. Jiang, “Fault-cause iden- tiﬁcation method based on adaptive deep belief network and time-frequency characteristics of travelling wave,” IET Gen- eration, Transmission & Distribution, vol. 13, no. 5, pp. 724– 732, 2019. [19] M. Gholami, A. Abbaspour, M. Moeini-Aghtaie, M. Fotuhi- Firuzabad,andM.Lehtonen,“Detectingthelocationofshort- circuit faults in active distribution network using PMU-based state estimation,” IEEE Transactions on Smart Grid, vol. 11, no. 2, pp. 1396–1406, 2020. [20] W.Hu,C.Ruan,H.Nian,andD.Sun,“Simpliﬁedmodulation scheme for open-end winding PMSM system with common DC bus under open-phase fault based on circulating current suppression,” IEEE Transactions on Power Electronics, vol. 35, no. 1, pp. 10–14, 2020. [21] K. Zhu and P. W. T. Pong, “Fault classiﬁcation of power distribution cables by detecting decaying DC components with magnetic sensing,” IEEE Transactions on Instrumenta- tion and Measurement, vol. 69, no. 5, pp. 2016–2027, 2020. [22] L. Romero, J. Blesa, V. Puig, G. Cembrano, and C. Trapiello, “First results in leak localization in water distribution net- worksusinggraph-basedclusteringanddeeplearning,” IFAC- PapersOnLine, vol. 53, no. 2, pp. 16691–16696, 2020. [23] W. Wang, N. Yu, Y. Gao, and J. Shi, “Safe oﬀ-policy deep reinforcement learning algorithm for volt-VAR control in power distribution systems,” IEEE Transactions on Smart Grid, vol. 11, no. 4, pp. 3008–3018, 2020. [24] D. A. Leo´n-Vargas, V. A. Bucheli-Guerrero, and H. A. Ordoez, “Solarradiationpredictiononphotovoltaicsystemsusingmachine learningtechniques,” Revista Facultad de Ingenier´ıa,vol.29,no.10, pp. 1–20, 2020.

Journal

Journal of Robotics – Hindawi Publishing Corporation

Published: Apr 14, 2022

References

Effective algorithm for fault discrimination and estimation of fault location in transmission lines,

S. Gangolu, P. Raja, M. P. Selvan, and V. K. Murali
Power equipment status information parallel fault diagnosis of based on MapReduce,

N. Hu, H. Du, S. Liu, and Q. Lin
Analysis of DC offset in fault current caused by machines in a medium voltage distribution network,

E. Bashar, Q. Han, R. Wu, L. Ran, O. Alatise, and S. Jupe
Research on the deterioration process of electrical contact structure inside the ± 500 kV converter transformer RIP bushings and its prediction strategy,

H. Tian, P. Liu, S. Zhou et al.
Faulted line location method for distribution systems based on the equipment’s information exchange,

W. Luo, H. Wang, L. Wang, Z. Zhu, and H. Gao
Detection and classification of high impedance fault in power distribution system using hybrid technique,

N. Narasimhulu, D. Kumar, and M. V. Kumar
Fault location in power distribution systems via deep graph convolutional networks,

K. Chen, J. Hu, Y. Zhang, Z. Yu, and J. He
Automatic optimal synthesis of aircraft electric power distribution system,

F. Wan, P. Madhika, J. Chwa, M. Mozumdar, and A. Ameri
Distance protection scheme for DC distribution systems based on the high-frequency characteristics of faults,

K. Jia, Q. Zhao, T. Feng, and T. Bi
Analysis of unsymmetrical faults based on artificial neural network using 11 kV distribution network of University of lagos as case study,

A. S. Alayande, I. K. Okakwu, O. E. Olabode, and O. K. Nwankwoh
An intuitionistic kernel-based fuzzy C-means clustering algorithm with local information for power equipment image segmentation,

F. Hu, H. Chen, and X. Wang
Anomaly detection for condition monitoring data using auxiliary feature vector and density-based clustering,

H. Liu, Y. Wang, and W. Chen
Fault detection in distribution networks in presence of distributed generations using a data mining–driven wavelet transform,

Youness, Mohammadnian, A. Turaj, and A. Soroudi
Fault diagnosis of circuit breakers based on time-frequency and chaotic vibration analysis,

Q. Yang, J. Ruan, and Z. Zhuang
Single phase fault diagnosis and location in active distribution network using synchronized voltage measurement,

T. Zhang, H. Yu, P. Zeng, L. Sun, C. Song, and J. Liu
Fault diagnosis of insulators from ultrasound detection using neural networks,

S. Frizzo-Stefenon, M. C. Silva, D. W. Bertol, L. H. Meyer, and A. Nied
Automatic fault diagnosis of infrared insulator images based on image instance segmentation and temperature analysis,

B. Wang, M. Dong, M. Ren et al.
Fault-cause identification method based on adaptive deep belief network and time-frequency characteristics of travelling wave,

H. Liang, Y. Liu, G. Sheng, and X. Jiang
Detecting the location of short-circuit faults in active distribution network using PMU-based state estimation,

M. Gholami, A. Abbaspour, M. Moeini-Aghtaie, M. Fotuhi-Firuzabad, and M. Lehtonen
Simplified modulation scheme for open-end winding PMSM system with common DC bus under open-phase fault based on circulating current suppression,

W. Hu, C. Ruan, H. Nian, and D. Sun
Fault classification of power distribution cables by detecting decaying DC components with magnetic sensing,

K. Zhu and P. W. T. Pong
First results in leak localization in water distribution networks using graph-based clustering and deep learning,

L. Romero, J. Blesa, V. Puig, G. Cembrano, and C. Trapiello
Safe off-policy deep reinforcement learning algorithm for volt-VAR control in power distribution systems,

W. Wang, N. Yu, Y. Gao, and J. Shi
Solar radiation prediction on photovoltaic systems using machine learning techniques,

D. A. León-Vargas, V. A. Bucheli-Guerrero, and H. A. Ordoez
Adaptive control by convolutional neural network in plasma arc welding system,

S. Yamane and K. Matsuo
Using synthetic data to enhance the accuracy of fingerprint-based localization: a deep learning approach,

M. Nabati, H. Navidan, R. Shahbazian, S. A. Ghorashi, and D. Windridge
Automated gleason grading of prostate cancer using transfer learning from general-purpose deep-learning networks,

M. S. Erbnescu, N. C. Manea, L. Streba et al.

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Fault Diagnosis Method of Distribution Equipment Based on Hybrid Model of Robot and Deep Learning

Fault Diagnosis Method of Distribution Equipment Based on Hybrid Model of Robot and Deep Learning

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Fault Diagnosis Method of Distribution Equipment Based on Hybrid Model of Robot and Deep Learning

Fault Diagnosis Method of Distribution Equipment Based on Hybrid Model of Robot and Deep Learning

References

Abstract

Journal

Recommended Articles

References

Our policy towards the use of cookies