Insulator damage identification method based on uncertainty estimation
By combining the MC-dropout algorithm with YOLOv5s and DenseNet201 networks, efficient identification and uncertainty estimation of insulator damage in distribution networks are achieved, solving the problems of insufficient detection accuracy and efficiency in existing technologies, and providing accurate identification results and uncertainty assessment.
Patent Information
- Application Number
- CN202111674765.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-31
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-12-31
AI Technical Summary
Existing technologies lack sufficient uncertainty analysis for insulator defect detection in power distribution networks, and the computational complexity of convolutional neural networks limits their generalization ability to detect morphologies that have not yet appeared, resulting in insufficient detection accuracy and efficiency.
The traditional convolutional neural network is adjusted using the MC-dropout algorithm, and approximate Bayesian inference is performed using the Monte Carlo dropout algorithm. Combined with YOLOv5s and DenseNet201 networks, quantitative identification and uncertainty estimation of insulator damage are achieved.
It improves the accuracy and efficiency of insulator damage detection, and can quantitatively describe the uncertainty of the identification results, assisting maintenance personnel in maintenance.
Smart Images

Figure CN114359695B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of insulator defect identification, and in particular to an insulator damage identification method based on uncertainty estimation. Background Art
[0002] Most of the distribution network lines are outdoor overhead lines. Insulators are prone to dirtiness and damage under the influence of complex weather conditions in the field for a long time. Therefore, regular inspection of distribution network insulators is very important. With the widespread application of drone inspection in distribution lines, the workload of manual line inspection has been effectively reduced. At the same time, the technology of automatic recognition of visible light inspection image data by machine vision has greatly improved the efficiency of power inspection.
[0003] Automatic inspection image recognition based on deep learning has performed well in recent years. At present, certain results have been achieved in defect detection of transmission line inspection images, but it is less used in distribution networks. The on-site environment and meteorological conditions of distribution network inspection insulator images are complex, the shooting angles are variable, and the insulator defects are diverse in image manifestation. It is difficult for the training set to cover all defect forms. The structure of the neural network limits the generalization ability of defect forms that do not appear in the training set. Research in recent years has mainly focused on improving detection accuracy, and there is currently almost no uncertainty analysis of power inspection results.
[0004] Bayesian neural networks can solve the problem of uncertainty quantification very well. By combining prior information with the likelihood information under the observed data to obtain the posterior probability distribution of the parameters, it can improve the generalization ability of unknown samples and provide uncertainty in the prediction results. The commonly used inference methods in Bayesian neural networks, Markov chain Monte Carlo method (MCMC) and variational inference method (VI), have achieved good results in feedforward networks and recurrent neural networks, but are limited by the amount of computation in convolutional neural networks.
[0005] Therefore, the present invention proposes an insulator damage identification method based on uncertainty estimation, and introduces cognitive uncertainty by using the MC-dropout algorithm. Only a small adjustment is needed to the traditional convolutional neural network to obtain prediction results similar to those of the Bayesian neural network, which greatly reduces the amount of calculation. Summary of the invention
[0006] In view of all the above problems, the present invention provides an insulator damage identification method based on uncertainty estimation, which adopts Monte Carlo dropout (MC-dropout) algorithm to perform approximate Bayesian reasoning, and can quantitatively describe the uncertainty of the identification result to assist maintenance personnel in maintenance.
[0007] The technical solution adopted by the present invention is:
[0008] The insulator damage identification method based on uncertainty estimation includes the following steps:
[0009] Step 1: Create a data set for insulator damage detection in distribution networks and divide it into training data and test data in a ratio of 9:1;
[0010] Step 2: Use the insulator damage detection dataset in step 1 to train the YOLOv5s network;
[0011] Step 3: According to the positioning frame information, the rectangular area of the insulator is intercepted from the dataset image, and classified and stored according to whether there is a damage fault. The intercepted insulator image data is used for DenseNet201 network training;
[0012] Step 4: Use the trained YOLOv5s and DenseNet201 models for real-time detection of insulator damage in the distribution network. First, use the YOLOv5s model to locate the insulator area. Then, based on the positioning frame, extract the insulator area from the inspection image, and use the DenseNet201 model to further classify whether the insulator is damaged.
[0013] Step 5: The DenseNet201 model is trained according to the MC-dropout method. During testing, the dropout operation is still retained in the forward propagation process. Under the effect of dropout, multiple different network structures can be obtained by randomly closing neurons, thereby making multiple predictions for the same sample.
[0014] Step 6: Calculate the mean of multiple prediction results to represent the classification result of the DenseNet201 model, and calculate the variance of multiple prediction results to represent the uncertainty of the classification result.
[0015] In step 1, the experimental data includes 2,000 aerial images taken by drones for inspection of a distribution network in a certain province, and the insulator area is manually labeled with a rectangular frame using the Labelimg labeling tool and saved in the YOLO format.
[0016] In step 2, a lightweight YOLOv5s model is used, and the weight file size is only 27MB. While ensuring the accuracy, the detection speed is also faster, which is suitable for embedded devices for UAV power inspection. Set the YOLOv5s network training parameters: batch size (batchsize) is set to 64, the number of iterations (epoch) is set to 300, the initial learning rate is set to 0.001, and the decay coefficient is set to 0.0005.
[0017] The step 3 specifically includes:
[0018] Step 3.1: The captured insulator image data includes 2000 normal insulators and 63 damaged insulators. Due to the small amount of image data of damaged insulators in the distribution network and the imbalance of the number of image samples, the damaged insulator images are synthesized by image flipping, brightness adjustment and adding Gaussian noise to the data enhancement method, and the damaged insulator images are expanded to 800.
[0019] Step 3.2: Set the DenseNet201 network training parameters: batch size is set to 64, number of iterations (epoch) is set to 100, initial learning rate is set to 0.001, decay coefficient is set to 0.005, and dropout probability is 0.5.
[0020] In step 5, the MC-dropout method is used to obtain the uncertainty of the test results of the DenseNet201 model. The specific scheme is as follows:
[0021] During training, there is no need to change the structure of the DenseNet201 model. During testing, dropout is turned on, and multiple different network structures are formed by randomly shutting down neurons. Specifically, the activation function value of each neuron in the fully connected layer of the DenseNet201 model is set to 0 with a probability of p=0.5. After dropout, the number of neurons in each layer becomes about half of the original. Dropout is turned on during testing. The neural network structure will be different each time the test is performed. Through these multiple different network structures, multiple predictions are made for the same sample, and the average of the multiple prediction results is calculated as the classification result. The variance of the multiple predictions is calculated as the uncertainty of the detection result.
[0022] In step 6, the mean of multiple prediction results calculated by formula (1) represents the classification result of the DenseNet201 model, and the classification result is a normal insulator or a damaged insulator. The variance of multiple prediction results is calculated to represent the uncertainty of the classification result. The uncertainty represented by the variance is specifically calculated by formula (2). Since the ability of the variance to describe uncertainty is low when the data distribution is multi-peaked, the information entropy is further used to calculate the uncertainty as shown in formula (3). The specific scheme is as follows:
[0023] The traditional classification network dropout layer improves the model's generalization ability and reduces overfitting by randomly shutting down some neurons with a certain probability during the training process, and shuts down the dropout layer during prediction. The MC-dropout algorithm does not need to change the model training process. By turning on the dropout layer during model prediction, the mean of T predictions is used as the classification result, and the variance is used as the uncertainty. T samples can be processed in parallel, which is equivalent to the time it takes to run the model once. The prediction result of the DenseNet201 model is shown in formula (1), and the variance is shown in formula (2).
[0024]
[0025] In formula (1), E(y) represents the mean of T prediction results, T is the number of sampling times, X is the input sample, The output of the network at the tth prediction given the input X.
[0026] In formula (2), Var(y) represents the variance of T prediction results.
[0027] Since the ability of variance to describe uncertainty is low when the data distribution is multi-peak, information entropy is further used to represent uncertainty, as shown in formula (3).
[0028]
[0029] In formula (3), H(p) represents the information entropy of the result, which is used to describe the uncertainty measurement, I is the number of categories, and p is i is the probability distribution when the sampling result is i, i=0 means the sample is a normal insulator, i=1 means the sample is a damaged insulator, and the defect classification result is the category with the highest probability.
[0030] The present invention provides an insulator damage identification method based on uncertainty estimation, and the technical effects are as follows:
[0031] 1) Compared with the traditional image recognition method based on color and shape features, which is easily affected by uncertain factors such as lighting, weather, and complex background of the picture, the method of the present invention can accurately identify the damaged insulators photographed by drone inspection of distribution lines.
[0032] 2) The present invention adopts the MC-dropout method to capture the cognitive uncertainty of the DenseNet201 model. Compared with the traditional variational inference method, it simplifies the approximate calculation of the posterior probability distribution. In the modeling process, only a small adjustment is needed to the traditional deep neural network to obtain a classification result similar to that of the Bayesian neural network, taking into account the timeliness of the classification and the measurement of uncertainty, and obtaining better classification accuracy.
[0033] 3) The present invention uses a deep learning method to detect insulator damage defects based on the different insulator shape and color features in the image data collected by drone inspections. It can quickly and effectively make accurate judgments on the degree of damage to the insulator, and give a quantitative assessment of the uncertainty of the recognition results, assisting users in establishing confidence standards, giving priority to high-confidence recognition results, and facilitating maintenance personnel to inspect. In terms of evaluating the uncertainty of recognition results, the method of the present invention is simpler and more real-time than the traditional Bayesian method. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is an overall flow chart of the method of the present invention.
[0035] Figure 2(1) is the second example of Mosaic data enhancement;
[0036] Figure 2(2) is the third example of Mosaic data enhancement;
[0037] Figure 3 (1) is an example of damaged insulator data enhancement (original image);
[0038] Figure 3(2) is an example of enhanced data of a damaged insulator (color jitter);
[0039] Figure 3 (3) is an example of damaged insulator data enhancement (random flip);
[0040] Figure 3(4) is an example of damaged insulator data enhancement (random occlusion);
[0041] Figure 3(5) is an example of enhanced damaged insulator data (Gaussian noise).
[0042] Figure 4 It is a specific flow chart of real-time detection of insulators.
[0043] Figure 5(1) is a conventional neural network diagram;
[0044] Figure 5(2) is the principle diagram of the MC-dropout method.
[0045] Figure 6 (1) shows the insulator positioning effect Figure 1 ;
[0046] Figure 6(2) is the second insulator positioning effect diagram.
[0047] Figure 7(1) shows the parameter changes during the YOLOv5s model training process Figure 1 ;
[0048] Figure 7(2) is the second diagram of parameter changes during the training of the YOLOv5s model;
[0049] Figure 7(3) is the third diagram of parameter changes during the training of the YOLOv5s model;
[0050] Figure 7(4) shows the parameter changes during the YOLOv5s model training process Figure 4 .
[0051] Figure 8 (1) is a graph showing the change in loss value during DenseNet201 model training;
[0052] Figure 8(2) is a graph showing the accuracy changes during the training of the DenseNet201 model. DETAILED DESCRIPTION
[0053] Insulator damage identification method based on uncertainty estimation, the specific process is as follows Figure 1 As shown, the following steps are included:
[0054] Step 1: Create a data set for insulator damage detection in distribution networks and divide it into training data and test data in a ratio of 9:1;
[0055] Step 2: Use the insulator damage detection dataset in step 1 to train the YOLOv5s network;
[0056] Step 3: According to the positioning frame information, the rectangular area of the insulator is intercepted from the dataset image, and classified and stored according to whether there is a damage fault. The intercepted insulator image data is used for DenseNet201 network training;
[0057] Step 4: Use the trained YOLOv5s and DenseNet201 models for real-time detection of insulator damage in the distribution network. First, use the YOLOv5s model to locate the insulator area. Then, based on the positioning frame, extract the insulator area from the inspection image, and use the DenseNet201 model to further classify whether the insulator is damaged.
[0058] Step 5: The DenseNet201 model is trained according to the MC-dropout method. During testing, the dropout operation is still retained in the forward propagation process. Under the effect of dropout, multiple different network structures can be obtained by randomly closing neurons, thereby making multiple predictions for the same sample.
[0059] Step 6: Calculate the mean of multiple prediction results to represent the classification result of the DenseNet201 model, and calculate the variance of multiple prediction results to represent the uncertainty of the classification result.
[0060] In step 1, the experimental data includes 2,000 aerial images taken by drones for inspection of a distribution network in a certain province, and the insulator area is manually annotated with a rectangular frame using the Labelimg annotation tool and saved in YOLO format. The annotation interface selects all insulators and saves them as insulators.
[0061] In step 2, the YOLO series model transforms the target detection problem into a regression problem, and directly regresses the target prediction box and classification category at multiple positions in the image, which greatly improves the detection speed. Compared with the traditional model that uses sliding windows or extracts proposals, YOLO directly uses the entire image to train the model, which can reduce the false detection rate in the background area and effectively solve the positioning problem of distribution network insulators in the complex environment.
[0062] YOLOv5 is an open source target detection model released in June 2020. It is divided into four models according to size: YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x. The present invention adopts a lightweight YOLOv5s model with a weight file size of only 27MB. While ensuring accuracy, the detection speed is also faster, which is suitable for embedded devices for drone power inspection. At the same time, since the insulator targets of the distribution network are small and difficult to detect, the YOLOv5 model uses Mosaic data enhancement to randomly scale and crop 4 pictures for splicing to increase the detection data set. Mosaic data enhancement examples are as follows: Figure 2(1) to Figure 2(2) As shown in the figure, a large number of small targets in the detection data set are added, the effect of small target detection is improved, and the problem of difficulty in detecting small targets in distribution network insulators can be effectively solved.
[0063] Set the YOLOv5s network training parameters: batch size is set to 64, the number of iterations (epoch) is set to 300, the initial learning rate is set to 0.001, and the decay coefficient is set to 0.0005.
[0064] In step 3, according to the positioning frame information, the rectangular area of the insulator is intercepted from the data set image, and classified and stored according to whether there is a damage fault. The specific process is as follows: Figure 1 As shown, in the offline training stage, the insulator area is cropped into separate images according to the rectangular frame marking the insulator area in step 1, and the cropped images are manually classified into normal insulators and damaged insulators as the training data set of the DenseNet201 model.
[0065] The step 3 specifically includes:
[0066] Step 3.1: The captured insulator image data includes 2,000 normal insulators and 63 damaged insulators. Since the image data of damaged insulators in the distribution network is small and the number of image samples is unbalanced, the damaged insulator images are increased by color jitter, random flipping, random cropping and adding Gaussian noise to the data enhancement method, and the damaged insulator images are expanded to 800. Data enhancement examples are as follows: Figure 3(1) to Figure 3(5)As shown in the figure, color jitter is to randomly adjust the saturation, brightness and contrast of the original image, which can simulate the image captured during inspection is easily affected by light and weather changes; random flip is to randomly flip the image horizontally and vertically; random occlusion is to cover random areas of the image with black, which can simulate the situation that the insulators of the distribution network are easily blocked by lines and tower materials during shooting; Gaussian noise is to add noise whose probability density function obeys Gaussian distribution to the original image, which can better simulate the unknown real noise.
[0067] Step 3.2: The DenseNet model is a classic neural network proposed in 2017. It achieves better results and fewer parameters by repeatedly utilizing shallow features through a unique densely connected module, and has a good effect in preventing overfitting. It is suitable for power inspection target detection where defect training data is relatively scarce. There are four network structures: DenseNet121, DenseNet169, DenseNet201, and DenseNet264. The present invention uses the DenseNet201 network structure. Set the DenseNet201 network training parameters: batch size (batchsize) is set to 64, number of iterations (epoch) is set to 100, initial learning rate is set to 0.001, attenuation coefficient is set to 0.005, and dropout probability is 0.5.
[0068] In step 4, the YOLOv5s model trained in step 2 and the DenseNet201 model trained in step 3 are used in the real-time detection of insulator damage in the distribution network. The insulator area is first located by the YOLOv5s model; then, the insulator area is intercepted from the inspection image according to the positioning frame, and the DenseNet201 model is used to further classify whether the insulator is damaged. The specific process of real-time detection is as follows: Figure 4 As shown in the figure, the insulator area located by the YOLOv5s model is cropped into separate images, and then the DenseNet201 model performs fault detection and classifies the cropped images into normal insulators and damaged insulators.
[0069] In step 5, the MC-dropout method is used to obtain the uncertainty of the test results of the DenseNet201 model. The specific scheme is as follows:
[0070] The traditional method uses a neural network based on a Bayesian probability model to estimate the uncertainty of the output, adding probability distribution to the model parameters and model output to consider the impact of uncertainty factors. The Bayesian inference formula is:
[0071]
[0072] In formula (4), X is the input data, Y is the category label, and W is the model parameter. When evaluating the posterior probability, the required p(Y|X) is difficult to calculate.
[0073] Commonly used inference methods in Bayesian neural networks, Markov chain Monte Carlo method (MCMC) and variational inference method (VI) have achieved good results in feedforward networks and recurrent neural networks, but are limited by the amount of computation in convolutional neural networks. Research on the MC-dropout method has shown that this method simplifies the approximate calculation of the posterior probability distribution. Assuming that each parameter obeys a Bernoulli distribution, the optimization process of the neural network after the dropout operation is equivalent to the Bayesian approximation of the Gaussian process.
[0074] Therefore, the present invention proposes an insulator damage identification method based on MC-dropout approximate Bayesian reasoning. During training, there is no need to change the DenseNet201 model structure. During testing, dropout is turned on and multiple different network structures are formed by randomly shutting down neurons. The specific principle is shown in Figures 5(1) and 5(2). By setting the activation function value of each neuron in the fully connected layer of the DenseNet201 model to 0 with a probability of p=0.5, the number of neurons in each layer becomes about half of the original after dropout. During testing, dropout is turned on. The neural network structure will be different each time the test is performed. Multiple predictions are made to the same sample through these multiple different network structures, and the average of the multiple prediction results is calculated as the classification result. The variance of the multiple predictions is calculated as the uncertainty of the detection result.
[0075] In step 6, the mean of multiple prediction results calculated by formula (1) represents the classification result of the DenseNet201 model, and the classification result is a normal insulator or a damaged insulator. The variance of multiple prediction results is calculated to represent the uncertainty of the classification result. The uncertainty represented by the variance is specifically calculated by formula (2). Since the ability of the variance to describe uncertainty is low when the data distribution is multi-peaked, the uncertainty is further calculated using information entropy as shown in formula (3).
[0076] The traditional classification network dropout layer improves the model's generalization ability and reduces overfitting by randomly shutting down some neurons with a certain probability during the training process, and shuts down the dropout layer during prediction. The MC-dropout algorithm does not need to change the model training process. By turning on the dropout layer during model prediction, the mean of T predictions is used as the classification result, and the variance is used as the uncertainty. T samples can be processed in parallel, which is equivalent to the time it takes to run the model once. The prediction result of the DenseNet201 model is shown in formula (1), and the variance is shown in formula (2).
[0077]
[0078]
[0079] In formula (1), E(y) represents the mean of T prediction results, T is the number of sampling times, X is the input sample, The output of the network at the tth prediction given the input X.
[0080] In formula (2), Var(y) represents the variance of T prediction results.
[0081] Since the ability of variance to describe uncertainty is low when the data distribution is multi-peak, information entropy is further used to represent uncertainty, as shown in formula (3).
[0082]
[0083] In formula (3), H(p) represents the information entropy of the result, which is used to describe the uncertainty measurement, I is the number of categories, and p is i is the probability distribution when the sampling result is i, i=0 means the sample is a normal insulator, i=1 means the sample is a damaged insulator, and the defect classification result is the category with the highest probability.
[0084] Example:
[0085] The simulation example of the present invention uses 90% of the distribution line insulator damage image data as the training set and 10% of the transmission and distribution line insulator damage image data as the verification set. The insulator positioning effect of the YOLOv5s model is shown in Figure 2. Figure 6(1) to Figure 6(2) As shown by Figure 6(1) to Figure 6(2) It can be seen that the method proposed in the present invention can accurately locate the insulator of the distribution line and further realize fault detection.
[0086] The changes in accuracy (Precision), recall (Recall) and average precision (mAP) during YOLOv5s model training are as follows Figure 7(1) to Figure 7(4) As shown in the figure, when epoch 150 or more, mAP tends to be stable, and the average accuracy tested on the test set is 92.75%, which ensures that all insulator areas can be effectively selected, providing a basis for further defect detection.
[0087] During the training process of the DenseNet201 model, the change in loss value is shown in Figure 8(1). When the number of epochs exceeds 80, the loss value tends to be stable and the entire network tends to converge. The change in accuracy is shown in Figure 8(2). When the number of epochs exceeds 80, the accuracy tends to be stable. The accuracy tested on the test set is 93.14%, which can well realize the detection of broken insulator faults.
[0088] In summary, the present invention provides an insulator damage identification method based on uncertainty estimation, and studies the effect of applying the MC-dropout method to estimate uncertainty in UAV distribution network inspection. Compared with traditional image processing technology, it can quantitatively describe the uncertainty of the recognition results to assist users in establishing confidence standards, give priority to high-confidence recognition results, and give priority to detecting and correcting low-confidence results.
[0089] Compared with the traditional variational inference method, the approximate calculation of the posterior probability distribution is simplified. During the modeling process, only a small adjustment is needed to the traditional deep neural network to obtain classification results similar to those of the Bayesian neural network, taking into account the timeliness of classification and the measurement of uncertainty, and obtaining better classification accuracy, which can better assist maintenance personnel in inspections.
Claims
1. An insulator damage identification method based on uncertainty estimation is characterized by The following steps are involved: Step 1: Create a data set for insulator damage detection in distribution networks and divide it into training data and test data; Step 2: Use the insulator damage detection dataset in step 1 to train the YOLOv5s network; Step 3: According to the positioning frame information, the rectangular area of the insulator is intercepted from the dataset image, and classified and stored according to whether there is a damage fault. The intercepted insulator image data is used for DenseNet201 network training; Step 4: Use the trained YOLOv5s and DenseNet201 models for real-time detection of insulator damage in the distribution network. First, use the YOLOv5s model to locate the insulator area. Then, based on the positioning frame, extract the insulator area from the inspection image, and use the DenseNet201 model to further classify whether the insulator is damaged. Step 5: The DenseNet201 model trained according to the MC-dropout method is retained during the test. The dropout operation is retained during the forward propagation process. Under the action of dropout, multiple different network structures can be obtained by randomly closing neurons, thereby making multiple predictions for the same sample. Step 6: Calculate the mean of multiple prediction results to represent the classification result of the DenseNet201 model, and calculate the variance of multiple prediction results to represent the uncertainty of the classification result.
2. The insulator damage identification method based on uncertainty estimation according to claim 1 is characterized in that : In step 1, the experimental data contains 2000 distribution network inspection drone aerial images, and the Labelimg annotation tool is used to manually annotate the insulator area with a rectangular box and save it in YOLO format.
3. The insulator damage identification method based on uncertainty estimation according to claim 1, characterized in that: In step 2, a lightweight YOLOv5s model is used with a weight file size of 27MB. The YOLOv5s network training parameters are set: the batch size is set to 64, the number of iterations is set to 300, the initial learning rate is set to 0.001, and the decay coefficient is set to 0.0005.
4. The insulator damage identification method based on uncertainty estimation according to claim 1, characterized in that: Step 3 specifically includes: Step 3.1: The captured insulator image data includes 2000 normal insulators and 63 damaged insulators. The damaged insulator images are added by using the data enhancement method of image flipping, brightness adjustment and Gaussian noise addition to synthesize the images, and the damaged insulator images are expanded to 800; Step 3.2: Set the DenseNet201 network training parameters: batch size is set to 64, number of iterations is set to 100, initial learning rate is set to 0.001, decay coefficient is set to 0.005, and dropout probability is 0.
5.
5. The insulator damage identification method based on uncertainty estimation according to claim 1, characterized in that: In step 5, the MC-dropout method is used to obtain the uncertainty of the test results of the DenseNet201 model. The specific scheme is as follows: During training, there is no need to change the structure of the DenseNet201 model. During testing, dropout is turned on and multiple different network structures are formed by randomly shutting down neurons. Specifically, the activation function value of each neuron in the fully connected layer of the DenseNet201 model is set to 0 with a probability of p=0.
5. After dropout, the number of neurons in each layer is reduced to half of the original. During testing, dropout is turned on and the neural network structure is different each time. Through these multiple different network structures, multiple predictions are made for the same sample, and the average of the multiple prediction results is calculated as the classification result. The variance of the multiple predictions is calculated as the uncertainty of the detection result.
6. The insulator damage identification method based on uncertainty estimation according to claim 1, characterized in that: In step 6, the mean of multiple prediction results calculated by formula (1) represents the classification result of the DenseNet201 model, and the classification result is a normal insulator or a damaged insulator; the variance of multiple prediction results is calculated to represent the uncertainty of the classification result; the uncertainty represented by the variance is specifically calculated by formula (2). Since the ability of the variance to describe uncertainty is low when the data distribution is multi-peaked, the information entropy is further used to calculate the uncertainty as shown in formula (3). The specific scheme is as follows: The prediction results of the DenseNet201 model are shown in formula (1), and the variance is shown in formula (2); In formula (1), E(y) represents the mean of T prediction results, T is the number of sampling times, X is the input sample, The output of the network at the tth prediction given input X; In formula (2), Var(y) represents the variance of T prediction results; Since the ability of variance to describe uncertainty is low when the data distribution is multi-peak, information entropy is further used to represent uncertainty, as shown in formula (3); In formula (3), H(p) represents the information entropy of the result, which is used to describe the uncertainty measurement, I is the number of categories, and p is i is the probability distribution when the sampling result is i, i=0 means the sample is a normal insulator, i=1 means the sample is a damaged insulator, and the defect classification result is the category with the highest probability.
