CN114092838A - Photovoltaic defect detection method based on deep learning target detection - Google Patents
Photovoltaic defect detection method based on deep learning target detection Download PDFInfo
- Publication number
- CN114092838A CN114092838A CN202111337200.2A CN202111337200A CN114092838A CN 114092838 A CN114092838 A CN 114092838A CN 202111337200 A CN202111337200 A CN 202111337200A CN 114092838 A CN114092838 A CN 114092838A
- Authority
- CN
- China
- Prior art keywords
- frame
- detection
- training
- photovoltaic
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 55
- 238000013135 deep learning Methods 0.000 title claims abstract description 13
- 230000007547 defect Effects 0.000 title claims abstract description 11
- 238000000034 method Methods 0.000 claims abstract description 42
- 238000012549 training Methods 0.000 claims abstract description 32
- 238000000605 extraction Methods 0.000 claims abstract description 8
- 238000013528 artificial neural network Methods 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 15
- 238000011176 pooling Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 5
- 238000002329 infrared spectrum Methods 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 5
- 238000002372 labelling Methods 0.000 claims description 4
- 238000005457 optimization Methods 0.000 claims description 4
- 238000011897 real-time detection Methods 0.000 claims description 4
- 230000002159 abnormal effect Effects 0.000 claims description 3
- 230000000694 effects Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 2
- 230000004927 fusion Effects 0.000 claims description 2
- 238000011478 gradient descent method Methods 0.000 claims description 2
- 238000012423 maintenance Methods 0.000 claims 1
- 238000005516 engineering process Methods 0.000 abstract description 2
- 238000003331 infrared imaging Methods 0.000 abstract description 2
- 239000000463 material Substances 0.000 abstract description 2
- 238000010248 power generation Methods 0.000 description 12
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000005855 radiation Effects 0.000 description 2
- MARUHZGHZWCEQU-UHFFFAOYSA-N 5-phenyl-2h-tetrazole Chemical compound C1=CC=CC=C1C1=NNN=N1 MARUHZGHZWCEQU-UHFFFAOYSA-N 0.000 description 1
- KTSFMFGEAAANTF-UHFFFAOYSA-N [Cu].[Se].[Se].[In] Chemical compound [Cu].[Se].[Se].[In] KTSFMFGEAAANTF-UHFFFAOYSA-N 0.000 description 1
- 229910021417 amorphous silicon Inorganic materials 0.000 description 1
- 229910021419 crystalline silicon Inorganic materials 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000013021 overheating Methods 0.000 description 1
- 230000000258 photobiological effect Effects 0.000 description 1
- 230000001699 photocatalysis Effects 0.000 description 1
- 229910021420 polycrystalline silicon Inorganic materials 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention discloses a photovoltaic defect detection method based on deep learning target detection, and belongs to the technical field of deep learning, neural networks and photovoltaic power station defect detection. The specific method comprises the following steps: firstly, an unmanned aerial vehicle is used for collecting original data, and the collected data are transmitted back to a background system for frame-by-frame extraction and pretreatment; then, making an infrared image data set; then putting the picture into a network for training, and outputting a characteristic picture; and finally, integrating the prediction network into a photovoltaic detection background system, and detecting the infrared video shot and returned by the unmanned aerial vehicle in real time. The method realizes infrared photovoltaic fault detection by combining deep learning with an unmanned aerial vehicle infrared imaging technology, has simple operation process, greatly saves manpower and material resources, and has higher recognition speed, strong robustness and high detection efficiency compared with the traditional photovoltaic hot spot detection method.
Description
Technical Field
The invention relates to the technical field of photovoltaic power station fault detection, in particular to a photovoltaic defect detection method based on deep learning target detection.
Background
Solar photovoltaic power generation refers to a power generation mode of directly converting light energy into electric energy without a thermal process. The method comprises photovoltaic power generation, photochemical power generation, photoinduction power generation and photobiological power generation. Photovoltaic power generation is a direct power generation method that utilizes a solar-grade semiconductor electronic device to effectively absorb solar radiation energy and convert the solar radiation energy into electric energy, and is the mainstream of current solar power generation. Electrochemical photovoltaic cells, photoelectrolytic cells and photocatalytic cells are used in photochemical power generation, and photovoltaic cells are currently used in practice.
The photovoltaic power generation system mainly comprises a solar cell, a storage battery, a controller and an inverter, wherein the solar cell is a key part of the photovoltaic power generation system, and the quality and the cost of a solar cell panel directly determine the quality and the cost of the whole system. The solar cell mainly comprises a crystalline silicon cell and a polycrystalline silicon cell, and the solar cell mainly comprises an amorphous silicon solar cell, a copper indium gallium selenide solar cell and a cadmium telluride solar cell.
The photovoltaic module works for a long time, a plurality of faults are easy to generate, the most serious fault is a hot spot fault, and the photovoltaic defect detected in the invention is a hot spot. When the photovoltaic module works, the current of the single battery in the module is reduced due to shading or self reasons, when the working current exceeds the current of the single battery, the partial battery is in a reverse bias state, the function in a circuit is changed from a power supply to a load, energy is consumed, and therefore a local overheating phenomenon is formed in the module, and hot spots are generated.
The traditional hot spot detection method comprises an I-V curve method, a manual handheld thermal imager inspection method and the like, and is low in detection speed, power loss and inaccurate in detection, complex in detection process, time-consuming and labor-consuming. With the development of science and technology, the method for carrying the infrared spectrum camera by the unmanned aerial vehicle can carry out preliminary detection on infrared photovoltaic faults.
According to the invention, the rotor unmanned aerial vehicle is adopted to carry an infrared spectrum camera to acquire images, the images are quickly returned through a 5G network, real-time detection is carried out at the background, and a target detection model is deeply learned while the timeliness is ensured, so that higher detection precision is obtained.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to provide a photovoltaic defect detection method based on deep learning target detection, which ensures detection speed and detection accuracy.
The technical scheme is as follows: the invention provides an infrared photovoltaic defect detection method based on a deep learning target detection algorithm, which comprises the following steps of:
and S1, collecting the original data. In the sunny weather with better illuminance, the rotor unmanned aerial vehicle and the infrared spectrum camera carried by the rotor unmanned aerial vehicle are used for shooting the photovoltaic module, path planning is carried out in advance before shooting, and shooting is carried out in a time period from 11 am to 16 pm in the early autumn. The shot photovoltaic panel infrared video is transmitted back in real time and is processed in the next step;
and S2, creating an infrared image data set. Since deep learning target detection algorithms require a large amount of labeling data, infrared images need to be labeled prior to training. First, all images are annotated using an open source tool, Labelimage, and an xml file including annotation information is generated. Storing the xml file in a VOCdevkit structure directory prepared in advance, wherein the xml file comprises an options file for storing the xml file, a JPEGImages file for storing an original picture and an ImageSets-Main for storing the txt file;
and S3, performing data enhancement operation on the original data. Because the number of samples is small, the number of original data sets needs to be increased by methods such as horizontal turning, rotation, inversion, cutting and splicing of the marked infrared pictures so as to ensure the accuracy of detection results;
and S4, carrying out prior frame clustering on the marked infrared images. And (4) clustering the prior frames by using a K-means + + algorithm. The K-Means + + algorithm is an unsupervised clustering algorithm, is relatively simple to realize, has good clustering effect, is widely applied, and meets the requirements on rapidness and accuracy when being applied to the method;
and S5, training a network. The invention trains data using an improved yolo network. And putting the processed data set into a network, changing the corresponding category information, the category name, the training iteration number and the corresponding file path, and running a train. The essence of the neural network is to obtain target weight information and calculate loss in the forward propagation process, and add a loss function to adjust the weight in the backward propagation process. When the loss value obtained through hundreds of iterations is very small, the weight file can be saved for target detection. The backbone network in the invention uses Darknet53 network structure, which is composed of a series of residual error structures, the original residual error block is continuously stacked on the basic backbone part, the branch part is equivalent to a residual error edge, and is directly connected to the last through a small amount of processing, and each convolution layer is operated through Batch Normalization Batch Normalization and Mish activation function. The backbone network divides the feature mapping of the basic layer into two parts, and then combines the two parts through a cross-stage hierarchical structure, thereby reducing the calculated amount and ensuring the accuracy;
and S6, outputting the feature map. After the image input model is processed, feature maps of three scales are output, the feature maps of different sizes correspond to different prior frames, and the three feature maps of different levels are fused. The shape of the 3 output tensors of the YOLO algorithm are: (19, 19, 225), (38, 38, 255), (76, 76, 225). Each grid corresponds to 3 Anchor boxes, and each bounding box to be predicted corresponds to 5 values;
s7, the prediction algorithm is integrated into a photovoltaic detection big data processing platform, after the infrared video shot and returned by the unmanned aerial vehicle is extracted at intervals of frames, real-time detection can be performed, abnormal information is accurately analyzed, a detection report is sent, and the unmanned aerial vehicle can be maintained and processed conveniently in time. For the extraction of the infrared video key frame, an algorithm based on interframe difference is adopted, and the principle is that difference is carried out on two frame images to obtain the average pixel intensity of the images, so that the change size of the two frame images can be measured. Therefore, based on the average intensity of the inter-frame difference, whenever a certain frame in the video has a large change from the picture content of the previous frame, it is considered as a key frame and extracted.
Further, in step S5, in order to enhance the detection of the target, the CutMix algorithm is used for data enhancement. The CutMix used by the invention is spliced by using two pictures and spliced in the modes of random zooming, random cutting and random arrangement, so that the data set is further increased and the robustness of the system is enhanced. The infrared photovoltaic hot spot target is generally small, the AP of the small target is generally much lower than that of the medium target and the large target, in order to enable the identification to be more accurate, 2 pictures are randomly used, randomly scaled and then randomly distributed for splicing, the detection data set is greatly enriched, particularly, the random scaling increases many small targets, and the robustness of the network is better. In addition, the CutMix data enhancement can directly calculate the data of 2 pictures during training, and the training speed is ensured while the data set is expanded;
the photovoltaic defect detection method based on deep learning target detection provided by the invention has the following beneficial effects:
1. the unmanned aerial vehicle is adopted to automatically acquire the infrared video of the photovoltaic panel, so that manpower and material resources are greatly saved, the real-time performance is ensured, and the system has stronger robustness;
2. the adopted Yolo model can better adapt to the identification of the infrared small target, has excellent performance on detection speed and detection precision, adopts a data enhancement method to more effectively increase the original data volume, and is more beneficial to identifying fault characteristics by a new backbone network.
Drawings
FIG. 1 is a flow chart of an unmanned aerial vehicle infrared imaging photovoltaic detection system;
FIG. 2 is an overall flow chart of the present invention;
Detailed Description
The technical solution of the method of the present invention is further described below with reference to the accompanying drawings of the method of the present invention.
A specific work flow of a photovoltaic defect detection method based on Yolov4 and a thermal infrared image is as follows:
and S1, collecting the original data. In the sunny weather with better illuminance, the rotor unmanned aerial vehicle and the infrared spectrum camera carried by the rotor unmanned aerial vehicle are used for shooting the photovoltaic module, path planning is carried out in advance before shooting, and shooting is carried out in a time period from 11 am to 16 pm in the early autumn. The shot photovoltaic panel infrared video is transmitted back in real time and is processed in the next step;
and S2, creating an infrared image data set. Since deep learning target detection algorithms require a large amount of labeling data, infrared images need to be labeled prior to training. First, all images are annotated using an open source tool, Labelimage, and an xml file including annotation information is generated. Storing the xml file in a VOCdevkit structure directory prepared in advance, wherein the xml file comprises an options file for storing the xml file, a JPEGImages file for storing an original picture and an ImageSets-Main for storing the txt file;
and S3, performing data enhancement operation on the original data. Because the number of samples is small, the number of original data sets needs to be increased by methods such as horizontal turning, rotation, inversion, cutting and splicing of the marked infrared pictures so as to ensure the accuracy of detection results;
and S4, carrying out prior frame clustering on the marked infrared images. And (4) clustering the prior frames by using a K-means + + algorithm. The K-Means + + algorithm is an unsupervised clustering algorithm, is relatively simple to realize, has good clustering effect, is widely applied, and meets the requirements on rapidness and accuracy when being applied to the method:
the location selection of k initialized centroids has a great influence on the final clustering result and the running time, so that the proper k centroids need to be selected. If it is simply a completely random selection, it may result in slow algorithm convergence. The K-Means + + algorithm is the optimization of the method for initializing the centroid randomly by K-Means.
The K-Means + + optimization strategy for initializing centroids is also simple, as follows:
a) randomly selecting a point from the input data point set as a first cluster center mu1；
b) For each point X in the datasetiCalculating its distance to the nearest cluster center among the selected cluster centers
c) Selecting a new data point as a new cluster center according to the following selection principles: d (x) larger points, with a higher probability of being selected as cluster centers;
d) repeating b and c until k clustered centroids are selected;
e) the K centroids are used as initialization centroids to run the standard K-Means algorithm.
And S5, training a network. And putting the processed data set into a network, and changing the corresponding class file category information, the class name, the training iteration times and the corresponding file path, wherein the total training time is set to be 5000 epochs in the initial training, and because the shallow layer characteristics have higher similarity after the iteration times are increased to a certain number and the generalization capability is better when the number of layers is higher, a part of Darknet53 training network is frozen when the 4000 th epoch is set, and in addition, the training speed can be accelerated to a certain extent. Then adding a CutMix online data enhancement method, and optimizing an objective function by using a gradient descent method in the network training process.
Then run a train. The essence of the neural network is to obtain target weight information and calculate loss in the forward propagation process, and add a loss function to adjust the weight in the backward propagation process. When the loss value obtained through hundreds of iterations is very small, the weight file can be saved for target detection. The backbone network in the invention uses Darknet53 network structure, which is composed of a series of residual error structures, the original residual error block is continuously stacked on the basic backbone part, the branch part is equivalent to a residual error edge, and is directly connected to the last through a small amount of processing, and each convolution layer is operated through Batch Normalization Batch Normalization and Mish activation function. The backbone network divides the feature mapping of the basic layer into two parts, and then combines the two parts through a cross-stage hierarchical structure, thereby reducing the calculated amount and simultaneously ensuring the accuracy, and the Mish activation function can further improve the training precision, and the principle is as follows:
y＝x*tanh(ln(1+exp(x)))
during training, the CutMix algorithm is used for data enhancement. The CutMix used by the invention is spliced by using two pictures and spliced in the modes of random zooming, random cutting and random arrangement, so that the data set is further increased and the robustness of the system is enhanced. The infrared photovoltaic hot spot target is generally small, the AP of the small target is generally much lower than that of the medium target and the large target, in order to enable the identification to be more accurate, 2 pictures are randomly used, randomly scaled and then randomly distributed for splicing, the detection data set is greatly enriched, particularly, the random scaling increases many small targets, and the robustness of the network is better. In addition, the CutMix data enhancement can directly calculate the data of 2 pictures during training, and the training speed is ensured while the data set is expanded;
in the process of training the network, the SPP module is inserted between the backhaul and the output layer, so that fusion characteristics can be better extracted. The SPP structure is applied to convolution of the last feature layer of Darknet53, after the last feature layer of PDarknet53 is convoluted for three times, the SPP structure is processed by using the maximum pooling of four different scales respectively, and the pooling kernel sizes of the maximum pooling downsampling are respectively 13 × 13, 9 × 9, 5 × 5 and 1 × 1;
and S6, outputting the feature map. After the image input model is processed, feature maps of three scales are output, the feature maps of different sizes correspond to different prior frames, and the three feature maps of different levels are fused. The shape of the 3 output tensors of the YOLO algorithm are: (19, 19, 225), (38, 38, 255), (76, 76, 225). Each grid corresponds to 3 Anchor boxes, and each bounding box to be predicted corresponds to 5 values. The loss function of the network includes three aspects: category loss, confidence loss, location loss:
the category loss function adopts multi-category cross entropy loss, and the formula is as follows:
in the above formula, a represents a prediction box, b represents an object class, and XabE {0,1}, X when the prediction box contains the real targetab1, X when the prediction box does not contain a true target valueab＝0。
The Sigmoid probability of the target category existing in the network prediction target boundary box is represented;
the target confidence coefficient adopts binary cross entropy loss, and the formula is as follows:
the position loss adopts DIOU loss, and the formula is as follows:
LDIOU＝1-IOU(A,B)+ρ2(Actr,Bctr)/c2
wherein, A is a prediction frame, and B is a real frame; a. thectrTo predict the frame center point coordinates, BctrAnd coordinates of the center point of the real frame. ρ (g) is the Euclidean distance calculation; c is diagonal line of A, B smallest enclosing frameA length;
s7, the prediction algorithm is integrated into a photovoltaic detection big data processing platform, after the infrared video shot and returned by the unmanned aerial vehicle is extracted at intervals of frames, real-time detection can be performed, abnormal information is accurately analyzed, a detection report is sent, and the unmanned aerial vehicle can be maintained and processed conveniently in time. For the extraction of the infrared video key frame, an algorithm based on interframe difference is adopted, and the principle is that difference is carried out on two frame images to obtain the average pixel intensity of the images, so that the change size of the two frame images can be measured. Therefore, based on the average intensity of the inter-frame difference, whenever a certain frame in the video has a large change from the picture content of the previous frame, it is considered as a key frame and extracted.
The flow of the algorithm is briefly described as follows:
firstly, reading a video, and sequentially calculating the interframe difference between every two frames to further obtain the average interframe difference strength;
secondly, we can choose one of the following three methods to extract the key frame, which are all based on the difference between frames;
and finally, selecting the frame with the local maximum of the average inter-frame differential strength as a key frame of the video by using the local maximum, wherein the extraction result of the method is better in richness and is uniformly dispersed in the video. It should be noted that smoothing the average inter-frame differential intensity time series is a significant skill when using this method. The method can effectively remove noise to avoid simultaneously extracting a plurality of frames in similar scenes as key frames.
Claims (1)
1. A photovoltaic defect detection method based on deep learning target detection is mainly characterized by comprising the following steps:
s1, collecting original data: in sunny weather with good illuminance, the rotor unmanned aerial vehicle and the carried infrared spectrum camera are used for shooting the photovoltaic module, path planning is carried out in advance before shooting, and the shot photovoltaic panel infrared video is transmitted back in real time and is processed in the next step;
s2, preparing an infrared image data set: since a deep learning target detection algorithm requires a large amount of labeled data, the infrared image needs to be labeled before training; firstly, labeling all images by using an open source tool Labelimage to generate an xml file containing labeling information; storing the xml file in a VOCdevkit structure directory prepared in advance, wherein the xml file comprises an options file for storing the xml file, a JPEGImages file for storing an original picture and an ImageSets-Main for storing the txt file;
s3, performing data enhancement operation on the original data: because the number of samples is small, the number of original data sets needs to be increased by methods such as horizontal turning, rotation, inversion, cutting and splicing of the marked infrared pictures so as to ensure the accuracy of detection results;
s4, carrying out prior frame clustering on the marked infrared images: clustering the prior frames by using a K-means + + algorithm; the K-Means + + algorithm is an unsupervised clustering algorithm, is relatively simple to realize, has good clustering effect, is widely applied, and meets the requirements on rapidness and accuracy when being applied to the method:
the position selection of k initialized centroids has great influence on the final clustering result and the running time, so that proper k centroids need to be selected; if it is only a completely random choice, it may result in slow algorithm convergence; the K-Means + + algorithm is the optimization of the method for initializing the centroid randomly by the K-Means;
the K-Means + + optimization strategy for initializing centroids is also simple, as follows:
a) randomly selecting a point from the input data point set as a first cluster center mu1；
b) For each point X in the datasetiCalculating its distance to the nearest cluster center among the selected cluster centers
c) Selecting a new data point as a new cluster center according to the following selection principles: d (x) larger points, with a higher probability of being selected as cluster centers;
d) repeating b and c until k clustered centroids are selected;
e) the K centroids are used as initialized centroids to run a standard K-Means algorithm;
s5, training the network: putting the processed data set into a network, and changing corresponding class file category information, category name, training iteration times and corresponding file paths, wherein the total training time is set to 5000 epochs in the primary training, and because the similarity of shallow layer characteristics is higher after the iteration times are increased to a certain number, and the generalization capability is better when the number of layers is higher, a part of Darknet53 training network is frozen when the number of layers is 4000 epochs, and in addition, the training speed can be accelerated to a certain extent; adding a CutMix online data enhancement method, and optimizing an objective function by using a gradient descent method in the network training process;
then running a train. The essence of the neural network is that target weight information is obtained and loss is calculated in the forward propagation process, and a loss function is added to adjust the weight in the backward propagation process; when the loss value obtained through hundreds of iterations is very small, the weight file can be saved for target detection; the backbone network in the invention uses Darknet53 network structure, which is composed of a series of residual error structures, the original residual error block is continuously stacked on the basic backbone part, the branch part is equivalent to a residual error edge, and is directly connected to the last through a small amount of processing, and each convolution layer is operated through Batch Normalization and Mish activation function; the backbone network divides the feature mapping of the basic layer into two parts, and then combines the two parts through a cross-stage hierarchical structure, thereby reducing the calculated amount and simultaneously ensuring the accuracy, and the Mish activation function can further improve the training precision, and the principle is as follows:
y＝x*tanh(ln(1+exp(x)))
in the training process, a CutMix algorithm is used for data enhancement; the CutMix used by the invention is spliced by using two pictures and spliced in the modes of random zooming, random cutting and random arrangement, so that the data set is further increased, and the robustness of the system is enhanced; the infrared photovoltaic hot spot target is generally small, the AP of the small target is generally much lower than that of the medium target and the large target, in order to ensure that the identification is more accurate, 2 pictures are randomly used, randomly scaled and then randomly distributed for splicing, so that a detection data set is greatly enriched, particularly, the random scaling increases many small targets, and the robustness of the network is better; in addition, the CutMix data enhancement can directly calculate the data of 2 pictures during training, and the training speed is ensured while the data set is expanded;
in the process of training the network, an SPP module is inserted between the backhaul and the output layer to better extract fusion characteristics; the SPP structure is applied to convolution of the last feature layer of Darknet53, after the last feature layer of PDarknet53 is convoluted for three times, the SPP structure is processed by using the maximum pooling of four different scales respectively, and the pooling kernel sizes of the maximum pooling downsampling are respectively 13 × 13, 9 × 9, 5 × 5 and 1 × 1;
s6, outputting of the feature map: after the picture input model is processed, feature maps of three scales are output, the feature maps of different sizes correspond to different prior frames, and the feature maps of three different levels are fused; the shape of the 3 output tensors of the YOLO algorithm are: (19, 19, 225), (38, 38, 255), (76, 76, 225); each grid corresponds to 3 Anchor boxes, and each bounding box to be predicted corresponds to 5 values; the loss function of the network includes three aspects: category loss, confidence loss, location loss:
the category loss function adopts multi-category cross entropy loss, and the formula is as follows:
in the above formula, a represents a prediction box, b represents an object class, and XabE {0,1}, when the prediction box contains trueReal object time Xab1, X when the prediction box does not contain a true target valueab＝0；
The Sigmoid probability of the target category existing in the network prediction target boundary box is represented;
the target confidence coefficient adopts binary cross entropy loss, and the formula is as follows:
the position loss adopts DIOU loss, and the formula is as follows:
LDIOU＝1-IOU(A,B)+ρ2(Actr,Bctr)/c2
wherein, A is a prediction frame, and B is a real frame; a. thectrTo predict the frame center point coordinates, BctrCoordinates of the center point of the real frame; ρ (g) is the Euclidean distance calculation; c is the diagonal length of A, B minimum enclosing frame;
s7, integrating a prediction algorithm into a photovoltaic detection big data processing platform, performing frame-by-frame extraction on an infrared video shot and returned by an unmanned aerial vehicle, performing real-time detection, accurately analyzing abnormal information and sending a detection report, wherein the detection report comprises fault information and positioning information, and is convenient for timely maintenance; for the extraction of the infrared video key frame, an algorithm based on interframe difference is adopted, and the principle is that difference is carried out on two frames of images to obtain the average pixel intensity of the images, so that the change size of the two frames of images can be measured; therefore, based on the average intensity of the inter-frame difference, whenever a certain frame in the video and the content of the previous frame generate a large change, the certain frame is considered as a key frame and is extracted;
the flow of the algorithm is briefly described as follows:
firstly, reading a video, and sequentially calculating the interframe difference between every two frames to further obtain the average interframe difference strength;
secondly, we can choose one of the following three methods to extract the key frame, which are all based on the difference between frames;
finally, the local maximum is used, the frame with the average inter-frame difference intensity local maximum is selected as the key frame of the video, the extraction result of the method is better in richness, and the extraction result is uniformly dispersed in the video; it should be noted that when this method is used, it is an effective skill to smooth the time series of the average inter-frame differential intensity; the method can effectively remove noise to avoid simultaneously extracting a plurality of frames in similar scenes as key frames.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111337200.2A CN114092838A (en) | 2021-11-12 | 2021-11-12 | Photovoltaic defect detection method based on deep learning target detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111337200.2A CN114092838A (en) | 2021-11-12 | 2021-11-12 | Photovoltaic defect detection method based on deep learning target detection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114092838A true CN114092838A (en) | 2022-02-25 |
Family
ID=80300067
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111337200.2A Pending CN114092838A (en) | 2021-11-12 | 2021-11-12 | Photovoltaic defect detection method based on deep learning target detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114092838A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117788462A (en) * | 2024-02-26 | 2024-03-29 | 湖南铁华精斧汽车集团股份有限公司 | Nondestructive testing method for breakage defect of self-unloading semitrailer axle |
| CN117788462B (en) * | 2024-02-26 | 2024-05-10 | 湖南铁华精斧汽车集团股份有限公司 | Nondestructive testing method for breakage defect of self-unloading semitrailer axle |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113538391A (en) * | 2021-07-25 | 2021-10-22 | 吉林大学 | Photovoltaic defect detection method based on Yolov4 and thermal infrared image |
-
2021
- 2021-11-12 CN CN202111337200.2A patent/CN114092838A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113538391A (en) * | 2021-07-25 | 2021-10-22 | 吉林大学 | Photovoltaic defect detection method based on Yolov4 and thermal infrared image |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117788462A (en) * | 2024-02-26 | 2024-03-29 | 湖南铁华精斧汽车集团股份有限公司 | Nondestructive testing method for breakage defect of self-unloading semitrailer axle |
| CN117788462B (en) * | 2024-02-26 | 2024-05-10 | 湖南铁华精斧汽车集团股份有限公司 | Nondestructive testing method for breakage defect of self-unloading semitrailer axle |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113538391B (en) | Photovoltaic defect detection method based on Yolov4 and thermal infrared image | |
| CN113284124B (en) | Photovoltaic panel defect detection method based on unmanned aerial vehicle vision | |
| CN114529817A (en) | Unmanned aerial vehicle photovoltaic fault diagnosis and positioning method based on attention neural network | |
| CN114037918B (en) | Photovoltaic hot spot detection method based on unmanned aerial vehicle inspection and image processing | |
| CN114283137A (en) | Photovoltaic module hot spot defect detection method based on multi-scale characteristic diagram inference network | |
| CN114021741A (en) | Photovoltaic cell panel inspection method based on deep learning | |
| Sun et al. | Defect detection of photovoltaic modules based on convolutional neural network | |
| CN114972225B (en) | Deep learning-based two-stage photovoltaic panel defect detection method | |
| Ren et al. | Design of photovoltaic hot spot detection system based on deep learning | |
| CN113781466B (en) | Solar panel defect identification method based on multi-scale attention-dense network | |
| CN114897857A (en) | Solar cell defect detection method based on light neural network | |
| CN108008633A (en) | Irradiation level comprising a variety of Changes in weather and photovoltaic module coordinate incidence relation method for building up | |
| Huang et al. | Solar panel defect detection design based on YOLO v5 algorithm | |
| CN113902792A (en) | Building height detection method and system based on improved RetinaNet network and electronic equipment | |
| Choulli et al. | Hybrid optimization based on the analytical approach and the particle swarm optimization algorithm (Ana-PSO) for the extraction of single and double diode models parameters | |
| CN114092838A (en) | Photovoltaic defect detection method based on deep learning target detection | |
| CN109388845B (en) | Photovoltaic array parameter extraction method based on reverse learning and enhanced complex evolution | |
| Mehta et al. | Assessment of solar energy generation potential on roof-tops using image processing | |
| CN114120059A (en) | Method for real-time multi-cloud judgment based on Resnet18 model | |
| CN114387466A (en) | Solar cell color difference detection method and system | |
| Otamendi et al. | A scalable framework for annotating photovoltaic cell defects in electroluminescence images | |
| Wu et al. | Ghost-RetinaNet: Fast Shadow DetectionMethod for Photovoltaic Panels Based on Improved RetinaNet. | |
| CN112132835A (en) | SeFa and artificial intelligence-based jelly effect analysis method for photovoltaic track camera | |
| Özer et al. | An approach based on deep learning methods to detect the condition of solar panels in solar power plants | |
| Verma et al. | Automatic Defect Classification of Electro-Luminescence Images of Photovoltaic Modules Based on Deep Learning CNN |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |