CN114359695A - Insulator breakage identification method based on uncertainty estimation - Google Patents
Insulator breakage identification method based on uncertainty estimation Download PDFInfo
- Publication number
- CN114359695A CN114359695A CN202111674765.XA CN202111674765A CN114359695A CN 114359695 A CN114359695 A CN 114359695A CN 202111674765 A CN202111674765 A CN 202111674765A CN 114359695 A CN114359695 A CN 114359695A
- Authority
- CN
- China
- Prior art keywords
- insulator
- uncertainty
- model
- dropout
- data
- 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
Images
Landscapes
- Insulators (AREA)
Abstract
The method for identifying the damage of the insulator based on uncertainty estimation extends the collected damage image of the insulator as training data and test data through data enhancement; firstly, positioning an insulator region by a YOLOv5s model; then, according to the information of the positioning frame, an insulator region is intercepted from the inspection image, and a DenseNet201 model is adopted to further classify the intercepted insulator region; according to the MC-dropout method, during the test of the DenseNet201 model, dropout operation is still reserved in the forward propagation process, a plurality of outputs of different network structures are obtained, and the mean value and the variance of the outputs represent the classification result and uncertainty of the DenseNet201 model. Compared with the traditional Bayesian method, the method provided by the invention is simpler in evaluation of the uncertainty of the recognition result and high in real-time performance.
Description
Technical Field
The invention relates to the field of insulator defect identification, in particular to an insulator damage identification method based on uncertainty estimation.
Background
Most of power distribution network lines are field overhead lines, and the insulator is prone to causing the problems of dirt, damage and the like under the influence of a field complex meteorological environment for a long time, so that the insulator is of great importance for regular inspection of the power distribution network insulator. Along with unmanned aerial vehicle patrols and examines extensive the development application in the distribution lines, the effectual work load that reduces artifical patrolling line, patrols and examines the technique of image data by machine vision automatic identification visible light simultaneously, very big improvement the efficiency that electric power was patrolled and examined.
The inspection image automatic identification based on deep learning is excellent in research in recent years, and at present, certain achievements are obtained for defect detection of the inspection image of the power transmission line, but the application of the inspection image automatic identification in the power distribution network is less. The field environment and meteorological conditions of the power distribution network inspection insulator image are complex, the shooting angle is variable, the insulator defects are diversified in image expression form, a training set is difficult to cover all defect forms, and the generalization capability of the defect forms which do not appear in the training set is limited by the structure of the neural network. Recent research mainly focuses on improving detection accuracy, and there is almost no uncertainty analysis of the power patrol detection result at present.
The Bayesian neural network can well solve the problem of uncertainty quantification, posterior probability distribution of parameters is obtained by combining prior information and likelihood information under observation data, generalization capability of unknown samples can be improved, and uncertainty of prediction results can be provided. The inference methods commonly used in bayesian neural networks, the markov chain Monte Carlo Method (MCMC) and the variational inference method (VI), have achieved better results in feed-forward networks and recurrent neural networks at present, but are limited by the amount of computation in convolutional neural networks.
Therefore, the invention provides an insulator damage identification method based on uncertainty estimation, cognitive uncertainty is introduced by adopting the MC-dropout algorithm, a prediction result similar to a Bayesian neural network can be obtained only by carrying out small adjustment on the traditional convolutional neural network, and the calculated amount is greatly reduced.
Disclosure of Invention
Aiming at all the problems, the invention provides an insulator damage identification method based on uncertainty estimation, which adopts Monte Carlo drop algorithm (MC-drop) to carry out approximate Bayesian inference, and can quantitatively describe the uncertainty of the identification result to assist maintenance personnel to maintain.
The technical scheme adopted by the invention is as follows:
the method for identifying the damage of the insulator based on uncertainty estimation comprises the following steps:
step 1: making a data set for detecting the damage of the insulator of the power distribution network, and pressing 9: 1 into training data and test data;
step 2: using the data set of insulator breakage detection in the step 1 for YOLOv5s network training;
and step 3: intercepting a rectangular region of the insulator from the data set image according to the information of the positioning frame, performing classified storage according to whether a damage fault exists or not, and using the intercepted image data of the insulator in a DenseNet201 network for training;
and 4, step 4: the trained YOLOv5s and DenseNet201 models are used for detecting the damage of the insulator of the power distribution network in real time, and the insulator region is positioned through the YOLOv5s model; then according to the positioning frame, an insulator region is cut out from the inspection image, and a DenseNet201 model is adopted to further classify whether the insulator is damaged or not.
And 5: the trained DenseNet201 model is completed according to the MC-dropout method, dropout operation is still reserved in the forward propagation process during testing, and a plurality of different network structures can be obtained by randomly closing neurons under the action of dropout, so that the same sample is predicted for multiple times.
Step 6: calculating the mean value of the multiple prediction results represents the classification result of the DenseNet201 model, and calculating the variance of the multiple prediction results represents the uncertainty of the classification result.
In the step 1, the experimental data comprise 2000 aerial images of the power distribution network inspection unmanned aerial vehicle in a certain province, and the rectangular frame for the insulator region is manually marked by using a Labelimg marking tool and stored in a YOLO format.
In the step 2, a lightweight YOLOv5s model is adopted, the size of the weight file is only 27MB, the detection speed is higher while the accuracy is ensured, and the method is suitable for embedded equipment of unmanned aerial vehicle power inspection. Setting the YOLOv5s network training parameters: the batch size (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 attenuation coefficient is set to 0.0005.
The step 3 specifically includes:
step 3.1: the intercepted insulator image data comprises 2000 normal insulators and 63 damaged insulators, and due to the fact that the number of damaged fault insulator image data of the power distribution network is small and the number of picture samples is unbalanced, picture turning, brightness adjustment and Gaussian noise adding data enhancement modes are adopted to synthesize pictures to increase damaged insulator images, and the damaged insulator images are expanded to 800.
Step 3.2: setting the DenseNet201 network training parameters: the batch size (batch size) is set to 64, the number of iterations (epoch) is set to 100, the initial learning rate is set to 0.001, the attenuation factor is set to 0.005, and the dropout probability is 0.5.
In the step 5, an MC-dropout method is adopted to obtain the uncertainty of the test result of the DenseNet201 model, and the specific scheme is as follows:
during training, the structure of the DenseNet201 model does not need to be changed, dropout is started during testing, and a plurality of different network structures are formed by randomly closing neurons. Specifically, an activation function value of each neuron is set to be 0 with the probability that p is 0.5 at a full connection layer of a DenseNet201 model, the number of the neurons of each layer after the neurons pass through the dropout is about half of the original number, the dropout is started during testing, the neural network structures are different during each testing, multiple predictions are performed on the same sample through multiple different network structures, the average value of multiple prediction results is calculated as a classification result, and the variance of the multiple predictions is calculated as the uncertainty of the detection result.
In the step 6, the average value of the multiple prediction results calculated by the 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 the multiple predictions is calculated to represent the uncertainty of the classification result. The uncertainty represented by the variance is specifically calculated by formula (2), and since the variance has a low capability of describing the uncertainty when the data distribution is multimodal, the uncertainty is further calculated by using the information entropy as shown in formula (3). The specific scheme is as follows:
the traditional classification network dropout layer improves the generalization capability of the model and reduces overfitting by randomly closing part of neurons with certain probability in the training process, and closes dropout in the prediction process. The MC-dropout algorithm does not need to change a model training process, a dropout layer is started during model prediction, the mean value of T times of prediction is used as a classification result, the variance is used as uncertainty, and T times of sampling can be processed in parallel, which is equivalent to the time of one-time model operation. The prediction result of the DenseNet201 model is shown in formula (1), and the variance is shown in formula (2).
In the formula (1), E (y) represents the average value of the prediction results of T times, T is the sampling times, X is the input sample,
the network is given the output of input X at the time of the t-th prediction.
In the formula (2), var (y) represents the variance of the T prediction results.
Since the data distribution is multimodal, the variance has low capability of describing uncertainty, and further information entropy is used to represent uncertainty, as shown in formula (3).
In the formula (3), H (p) represents the information entropy of the result and is used for describing the uncertaintyAmount, I is the number of classes, piThe probability distribution when the sampling result is i is shown, i-0 indicates that the sample is a normal insulator, i-1 indicates that the sample is a broken insulator, and the defect classification result is the type with the highest probability.
The invention discloses an insulator damage identification method based on uncertainty estimation, which has the following technical effects:
1) compared with the traditional image identification method which is based on color and shape characteristics and is easily influenced by uncertain factors such as illumination, weather and complex picture background, the method disclosed by the invention can be used for accurately identifying the damage of the insulator shot by the unmanned aerial vehicle for the distribution line in the inspection process.
2) The cognitive uncertainty of the DenseNet201 model is captured by adopting an MC-dropout method, and compared with a traditional variational inference method, the method simplifies the approximate calculation of posterior probability distribution; in the modeling process, the classification result similar to a Bayesian neural network can be obtained only by slightly adjusting the traditional deep neural network, the timeliness and uncertainty measurement of classification are considered, and better classification accuracy is obtained.
3) According to the method, the defect of insulator breakage is detected by using a deep learning method according to the difference of the shape and color characteristics of the insulator in the image data acquired by the unmanned aerial vehicle inspection. The method can quickly and effectively make accurate judgment on the damage degree of the insulator, give quantitative evaluation on the uncertainty of the identification result, assist a user in establishing a confidence standard, preferentially adopt the identification result with high confidence level, and facilitate the maintenance of maintenance personnel. Compared with the traditional Bayesian method, the method provided by the invention is simpler in evaluation of the uncertainty of the recognition result and high in real-time performance.
Drawings
FIG. 1 is an overall flow diagram of the method of the present invention.
Fig. 2 is an illustration of an insulator labeling example.
FIG. 3(1) is a diagram of a Mosaic data enhancement example;
FIG. 3(2) is a second example of the Mosaic data enhancement;
FIG. 3(3) is a third exemplary graph of Mosaic data enhancement;
fig. 3(4) is a diagram of a Mosaic data enhancement example.
FIG. 4(1) is an exemplary graph (original image) of the damaged insulation sub-data enhancement;
FIG. 4(2) is a diagram of an example of damaged insulator data enhancement (color dithering);
fig. 4(3) is an exemplary graph of data enhancement of a broken insulator (random inversion);
FIG. 4(4) is an exemplary graph of damaged insulator data enhancement (random occlusion);
fig. 4(5) is a diagram illustrating an example of the broken insulator data enhancement (gaussian noise).
Fig. 5 is a detailed flowchart of insulator real-time detection.
FIG. 6(1) is a diagram of a conventional neural network;
FIG. 6(2) is a schematic diagram of the MC-dropout method.
Fig. 7(1) is a first diagram of the positioning effect of the insulator;
FIG. 7(2) is a second diagram illustrating the positioning effect of the insulator;
fig. 7(3) is a third diagram of the positioning effect of the insulator;
fig. 7(4) shows the positioning effect of the insulator.
FIG. 8(1) is a first parameter variation diagram in the training process of the YOLOv5s model;
FIG. 8(2) is a second parameter variation diagram in the process of training the YOLOv5s model;
FIG. 8(3) is a third parameter variation diagram in the training process of the YOLOv5s model;
fig. 8(4) is a graph of parameter variation in the YOLOv5s model training process. (ii) a
Fig. 9(1) is a graph of the change in the loss value during the training of the DenseNet201 model;
fig. 9(2) is a graph of the accuracy rate change in the course of the DenseNet201 model training.
Detailed Description
The method for identifying the damage of the insulator based on uncertainty estimation comprises the following steps of:
step 1: making a data set for detecting the damage of the insulator of the power distribution network, and pressing 9: 1 into training data and test data;
step 2: using the data set of insulator breakage detection in the step 1 for YOLOv5s network training;
and step 3: intercepting a rectangular region of the insulator from the data set image according to the information of the positioning frame, performing classified storage according to whether a damage fault exists or not, and using the intercepted image data of the insulator in a DenseNet201 network for training;
and 4, step 4: the trained YOLOv5s and DenseNet201 models are used for detecting the damage of the insulator of the power distribution network in real time, and the insulator region is positioned through the YOLOv5s model; then according to the positioning frame, an insulator region is cut out from the inspection image, and a DenseNet201 model is adopted to further classify whether the insulator is damaged or not.
And 5: the trained DenseNet201 model is completed according to the MC-dropout method, dropout operation is still reserved in the forward propagation process during testing, and a plurality of different network structures can be obtained by randomly closing neurons under the action of dropout, so that the same sample is predicted for multiple times.
Step 6: calculating the mean value of the multiple prediction results represents the classification result of the DenseNet201 model, and calculating the variance of the multiple prediction results represents the uncertainty of the classification result.
In the step 1, the experimental data include 2000 aerial images of the power distribution network inspection unmanned aerial vehicle in a certain province, the insulator regions are manually marked by rectangular frames by using a Labelimg marking tool and stored in a YOLO format, the specific marking interface is shown in FIG. 2, and all insulator frames are selected and stored and marked as insulators.
In the step 2, the YoLO series model converts the target detection problem into a regression problem, and directly regresses the target prediction frame and classification category at a plurality of positions of the image, thereby greatly improving the detection speed. Compared with the traditional model adopting a sliding window or extracting proposal, the YOLO directly selects the whole picture to train the model, so that the false detection rate of a background area can be reduced, and the problem of positioning of the power distribution network insulator under the complex environment background is effectively solved.
YOLOv5 is an open source target detection model released in 6 months of 2020, and is divided into four models of YOLOv5s, YOLOv5m, YOLOv5l and YOLOv5x according to size. The invention adopts a lightweight YOLOv5s model, the size of the weight file is only 27MB, the detection speed is higher while the accuracy is ensured, and the invention is suitable for embedded equipment of unmanned aerial vehicle power inspection. Meanwhile, as the insulator target of the power distribution network is small and the detection difficulty is large, the YOLOv5 model randomly scales and cuts 4 pictures for splicing through Mosaic data enhancement to increase the detection data set. As shown in fig. 3(1) to fig. 3(4), the Mosaic data enhancement method greatly increases small targets in the detection data set, improves the effect of detecting the small targets, and can effectively solve the problem of difficulty in detecting the small targets of the power distribution network insulator.
Setting the YOLOv5s network training parameters: the batch size (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 attenuation coefficient is set to 0.0005.
In the step 3, according to the information of the positioning frame, a rectangular region of the insulator is cut out from the data set image, and classified storage is performed according to whether a damage fault exists, the specific process is shown in fig. 1, in the off-line training stage, according to the rectangular frame marked with the insulator region in the step 1, the insulator region is cut into separate pictures, and then the cut pictures are manually classified into a normal insulator and a damaged insulator as a training data set of a densnet 201 model.
The step 3 specifically includes:
step 3.1: the intercepted insulator image data comprises 2000 normal insulators and 63 damaged insulators, and due to the fact that the number of the damaged and failed insulator image data of the power distribution network is small, the number of picture samples is unbalanced, damaged insulator images are increased in a data enhancement mode of color dithering, random overturning, random cutting and Gaussian noise adding, and the damaged insulator images are expanded to 800. The data enhancement is shown in fig. 4(1) -4 (5), the color dithering randomly adjusts the saturation, brightness and contrast of the original picture, and can simulate that the shot image is easily affected by illumination and weather change during patrol; the random turning is to turn the picture in the horizontal and vertical directions randomly; the random shielding is to shield the random area of the picture with black, so that the condition that the insulator of the power distribution network is easily shielded by circuits and tower materials during shooting can be simulated; the Gaussian noise is noise which is obtained by adding a probability density function in an original picture and follows Gaussian distribution, and unknown real noise can be better simulated.
Step 3.2: the DenseNet model is a classical neural network proposed in 2017, achieves better effect and fewer parameters by recycling shallow features through a unique dense connection module, has a good over-fitting prevention effect, and is suitable for power inspection target detection with relatively-deficient defect training data. The invention has four network structures of DenseNet121, DenseNet169, DenseNet201 and DenseNet264, wherein the DenseNet201 network structure is selected. Setting the DenseNet201 network training parameters: the batch size (batch size) is set to 64, the number of iterations (epoch) is set to 100, the initial learning rate is set to 0.001, the attenuation factor is set to 0.005, and the dropout probability is 0.5.
In the step 4: applying the YOLOv5s model trained in the step 2 and the DenseNet201 model trained in the step 3 to real-time detection of damage of the power distribution network insulator, and positioning an insulator region through the YOLOv5s model; then according to the positioning frame, an insulator region is cut out from the inspection image, and a DenseNet201 model is adopted to further classify whether the insulator is damaged or not. The real-time detection process is as shown in fig. 5, the insulator region located by the YOLOv5s model is cut into separate pictures, and then fault detection is performed by the DenseNet201 model, so that the cut pictures are classified into normal insulators and damaged insulators.
In the step 5, an MC-dropout method is adopted to obtain the uncertainty of the test result of the DenseNet201 model, and the specific scheme is as follows:
the traditional method uses a neural network based on a Bayesian probability model to estimate the uncertainty of the output, adds probability distribution to the model parameters and the model output to consider the influence of uncertainty factors, and adopts a Bayesian inference formula:
in the formula (4), X is input data, Y is a class label, and W is a model parameter. The required p (Y | X) computation difficulty is large when evaluating the posterior probability.
The inference methods commonly used in bayesian neural networks, the markov chain Monte Carlo Method (MCMC) and the variational inference method (VI), have achieved better results in feed-forward networks and recurrent neural networks at present, but are limited by the amount of computation in convolutional neural networks. The MC-dropout method is researched to prove that the 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 operated by dropout is equivalent to the Bayesian approximation of the Gaussian process.
Therefore, the invention provides an insulator damage identification method based on MC-dropout approximate Bayesian inference, the structure of a DenseNet201 model does not need to be changed during training, dropout is started during testing, and a plurality of different network structures are formed by randomly closing neurons. As shown in fig. 6(1) and 6(2), the activation function value of each neuron is set at the full connection layer of the DenseNet201 model, the probability that p is 0.5 becomes 0, the number of neurons in each layer after dropout becomes about half of the original number, the dropout is opened during testing, the neural network structure is different during each testing, multiple predictions are performed on the same sample through multiple different network structures, the average value of the multiple prediction results is calculated as a classification result, and the variance of the multiple predictions is calculated as the uncertainty of the detection result.
In the step 6, the average value of the multiple prediction results calculated by the 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 the multiple predictions is calculated to represent the uncertainty of the classification result. The uncertainty represented by the variance is specifically calculated by formula (2), and since the variance has a low capability of describing the uncertainty when the data distribution is multimodal, the uncertainty is further calculated by using the information entropy as shown in formula (3).
The traditional classification network dropout layer improves the generalization capability of the model and reduces overfitting by randomly closing part of neurons with certain probability in the training process, and closes dropout in the prediction process. The MC-dropout algorithm does not need to change a model training process, a dropout layer is started during model prediction, the mean value of T times of prediction is used as a classification result, the variance is used as uncertainty, and T times of sampling can be processed in parallel, which is equivalent to the time of one-time model operation. The prediction result of the DenseNet201 model is shown in formula (1), and the variance is shown in formula (2).
In the formula (1), E (y) represents the average value of the prediction results of T times, T is the sampling times, X is the input sample,
the network is given the output of input X at the time of the t-th prediction.
In the formula (2), var (y) represents the variance of the T prediction results.
Since the data distribution is multimodal, the variance has low capability of describing uncertainty, and further information entropy is used to represent uncertainty, as shown in formula (3).
In the formula (3), H (p) represents the information entropy of the result and is used for describing uncertainty measurement, I is the number of categories, piThe probability distribution when the sampling result is i is shown, i-0 indicates that the sample is a normal insulator, i-1 indicates that the sample is a broken insulator, and the defect classification result is the type with the highest probability.
Example (b):
the simulation example of the invention takes 90% distribution line insulator damage image data as a training set and takes 10% transmission and distribution line insulator damage image data as a verification set. The insulator positioning effect of the YOLOv5s model is shown in fig. 7(1) to fig. 7(4), and it can be seen from fig. 7(1) to fig. 7(4) that the distribution line insulator can be accurately positioned and fault detection can be further realized by the method provided by the invention.
Precision (Precision), Recall (Recall) and average accuracy (mAP) changes in the YOLOv5s model training process are shown in fig. 8(1) -8 (4), when the epoch is more than 150 times, the mAP tends to be stable, the average accuracy of the test on the test set is 92.75%, all insulator regions can be effectively selected, and a foundation is provided for further defect detection.
In the process of training the DenseNet201 model, the loss value changes as shown in (1) in FIG. 9, when the frequency of epoch is more than 80, the loss value tends to be stable, the whole network tends to be convergent, the accuracy rate changes as shown in (2) in FIG. 9, when the frequency of epoch is more than 80, the accuracy rate tends to be stable, the accuracy rate of the test on the test set is 93.14%, and the fault detection of the damaged insulator can be better realized.
In conclusion, the invention provides an insulator damage identification method based on uncertainty estimation, and researches the effect of applying the estimated uncertainty of the MC-dropout method to the unmanned aerial vehicle power distribution network inspection. Compared with the traditional image processing technology, the method can be used for quantitatively describing the uncertainty of the recognition result to assist the user in establishing a confidence standard, preferentially adopting the recognition result with high confidence degree, and preferentially detecting and correcting the result with low confidence degree.
Compared with the traditional variational inference method, the method simplifies the approximate calculation of posterior probability distribution, can obtain the classification result similar to a Bayesian neural network only by carrying out small adjustment on the traditional deep neural network in the modeling process, gives consideration to the timeliness and uncertainty measurement of classification, obtains better classification accuracy and can better assist maintenance personnel to overhaul.
Claims (6)
1. The method for identifying the damage of the insulator based on uncertainty estimation is characterized by comprising the following steps of:
step 1: manufacturing a data set for detecting the damage of the power distribution network insulator, and dividing the data set into training data and test data;
step 2: using the data set of insulator breakage detection in the step 1 for YOLOv5s network training;
and step 3: intercepting a rectangular region of the insulator from the data set image according to the information of the positioning frame, performing classified storage according to whether a damage fault exists or not, and using the intercepted image data of the insulator in a DenseNet201 network for training;
and 4, step 4: the trained YOLOv5s and DenseNet201 models are used for detecting the damage of the insulator of the power distribution network in real time, and the insulator region is positioned through the YOLOv5s model; then according to the positioning frame, intercepting an insulator region from the inspection image, and further classifying whether the insulator is damaged or not by adopting a DenseNet201 model;
and 5: the method comprises the following steps that a trained DenseNet201 model is completed according to an MC-dropout method, dropout operation is reserved in the forward propagation process during testing, and a plurality of different network structures can be obtained by closing neurons randomly under the action of dropout, so that the same sample is predicted for multiple times;
step 6: calculating the mean value of the multiple prediction results represents the classification result of the DenseNet201 model, and calculating the variance of the multiple prediction results represents the uncertainty of the classification result.
2. The method for identifying insulator damage based on uncertainty estimation according to claim 1, wherein in step 1, the experimental data comprises 2000 aerial images of the power distribution network inspection unmanned aerial vehicle, and the insulator region is manually marked with a rectangular frame by using a Labelimg marking tool and stored in a YOLO format.
3. The uncertainty estimation based insulator breakage identification method according to claim 1, characterized in that: in step 2, a lightweight YOLOv5s model is adopted, the weight file size is 27MB, and YOLOv5s network training parameters are set: the batch size was set to 64, the number of iterations was set to 300, the initial learning rate was set to 0.001, and the attenuation factor was set to 0.0005.
4. The uncertainty estimation based insulator breakage identification method according to claim 1, characterized in that: the step 3 specifically comprises the following steps:
step 3.1: the intercepted insulator image data comprises 2000 normal insulators and 63 damaged insulators, and the damaged insulator images are increased by synthesizing pictures in a data enhancement mode of picture turning, brightness adjustment and Gaussian noise addition, and are expanded to 800 images;
step 3.2: setting the DenseNet201 network training parameters: the batch size is set to 64, the number of iterations is set to 100, the initial learning rate is set to 0.001, the attenuation factor is set to 0.005, and the dropout probability is 0.5.
5. The uncertainty estimation based insulator breakage identification method according to claim 1, characterized in that: in step 5, obtaining the uncertainty of the test result of the DenseNet201 model by adopting an MC-dropout method, wherein the specific scheme is as follows:
during training, the structure of a DenseNet201 model does not need to be changed, dropout is started during testing, and a plurality of different network structures are formed by randomly closing neurons; specifically, an activation function value of each neuron is set to be 0 with the probability that p is 0.5 at a full connection layer of a DenseNet201 model, the number of the neurons of each layer after the neurons pass through the dropout is about half of the original number, the dropout is started during testing, the neural network structures are different during each testing, multiple predictions are performed on the same sample through multiple different network structures, the average value of multiple prediction results is calculated as a classification result, and the variance of the multiple predictions is calculated as the uncertainty of the detection result.
6. The uncertainty estimation based insulator breakage identification method according to claim 1, characterized in that: in step 6, the mean value of the multiple prediction results calculated by the formula (1) represents the classification result of the DenseNet201 model, and the classification result is a normal insulator or a damaged insulator; calculating the variance of the multiple prediction results to represent the uncertainty of the classification result; the uncertainty represented by the variance is specifically calculated by formula (2), and since the variance has a low uncertainty describing capability when the data distribution is multimodal, the uncertainty is further calculated by using the information entropy as shown in formula (3), and the specific scheme is as follows:
the prediction result of the DenseNet201 model is shown in formula (1), and the variance is shown in formula (2);
in the formula (1), E (y) represents the average value of the prediction results of T times, T is the sampling times, X is the input sample,
giving the output of the input X in the t prediction for the network;
in the formula (2), var (y) represents the variance of the prediction results of T times;
when the data distribution is multimodal, the uncertainty describing capability of the variance is low, and further, the uncertainty is represented by using the information entropy as shown in formula (3);
in the formula (3), H (p) represents the information entropy of the result and is used for describing uncertainty measurement, I is the number of categories, piThe probability distribution when the sampling result is i is shown, i-0 indicates that the sample is a normal insulator, i-1 indicates that the sample is a broken insulator, and the defect classification result is the type with the highest probability.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111674765.XA CN114359695A (en) | 2021-12-31 | 2021-12-31 | Insulator breakage identification method based on uncertainty estimation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111674765.XA CN114359695A (en) | 2021-12-31 | 2021-12-31 | Insulator breakage identification method based on uncertainty estimation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114359695A true CN114359695A (en) | 2022-04-15 |
Family
ID=81104979
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111674765.XA Pending CN114359695A (en) | 2021-12-31 | 2021-12-31 | Insulator breakage identification method based on uncertainty estimation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114359695A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114648683A (en) * | 2022-05-23 | 2022-06-21 | 天津所托瑞安汽车科技有限公司 | Neural network performance improving method and device based on uncertainty analysis |
| CN115035108A (en) * | 2022-08-10 | 2022-09-09 | 四川中电启明星信息技术有限公司 | Insulator defect detection method based on deep learning |
| CN116087198A (en) * | 2022-12-02 | 2023-05-09 | 河南交通发展研究院有限公司 | Highway road surface situation data acquisition equipment and automatic rapid detection system thereof |
| CN116228778A (en) * | 2023-05-10 | 2023-06-06 | 国网山东省电力公司菏泽供电公司 | Insulator rupture detection method and system based on multi-mode information fusion |
-
2021
- 2021-12-31 CN CN202111674765.XA patent/CN114359695A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114648683A (en) * | 2022-05-23 | 2022-06-21 | 天津所托瑞安汽车科技有限公司 | Neural network performance improving method and device based on uncertainty analysis |
| CN115035108A (en) * | 2022-08-10 | 2022-09-09 | 四川中电启明星信息技术有限公司 | Insulator defect detection method based on deep learning |
| CN116087198A (en) * | 2022-12-02 | 2023-05-09 | 河南交通发展研究院有限公司 | Highway road surface situation data acquisition equipment and automatic rapid detection system thereof |
| CN116087198B (en) * | 2022-12-02 | 2024-02-23 | 河南交通发展研究院有限公司 | Highway road surface situation data acquisition equipment and automatic rapid detection system thereof |
| CN116228778A (en) * | 2023-05-10 | 2023-06-06 | 国网山东省电力公司菏泽供电公司 | Insulator rupture detection method and system based on multi-mode information fusion |
| CN116228778B (en) * | 2023-05-10 | 2023-09-08 | 国网山东省电力公司菏泽供电公司 | Insulator rupture detection method and system based on multi-mode information fusion |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111353413B (en) | Low-missing-report-rate defect identification method for power transmission equipment | |
| CN114359695A (en) | Insulator breakage identification method based on uncertainty estimation | |
| CN110335270B (en) | Power transmission line defect detection method based on hierarchical regional feature fusion learning | |
| US10546242B2 (en) | Image analysis neural network systems | |
| CN107609525B (en) | Remote sensing image target detection method for constructing convolutional neural network based on pruning strategy | |
| CN107437245B (en) | High-speed railway contact net fault diagnosis method based on deep convolutional neural network | |
| CN108229550B (en) | Cloud picture classification method based on multi-granularity cascade forest network | |
| CN112434643A (en) | Classification and identification method for low-slow small targets | |
| CN111860106B (en) | Unsupervised bridge crack identification method | |
| CN111832607B (en) | Bridge disease real-time detection method based on model pruning | |
| CN111223087B (en) | Automatic bridge crack detection method based on generation countermeasure network | |
| CN112686833A (en) | Industrial product surface defect detecting and classifying device based on convolutional neural network | |
| CN115909093A (en) | Power equipment fault detection method based on unmanned aerial vehicle inspection and infrared image semantic segmentation | |
| CN116229380A (en) | Method for identifying bird species related to bird-related faults of transformer substation | |
| CN113469950A (en) | Method for diagnosing abnormal heating defect of composite insulator based on deep learning | |
| CN106650823A (en) | Probability extreme learning machine integration-based foam nickel surface defect classification method | |
| CN111178438A (en) | ResNet 101-based weather type identification method | |
| CN112633257A (en) | Potato disease identification method based on improved convolutional neural network | |
| CN112258490A (en) | Low-emissivity coating intelligent damage detection method based on optical and infrared image fusion | |
| CN115984543A (en) | Target detection algorithm based on infrared and visible light images | |
| CN111539302B (en) | Glass insulator self-explosion identification method based on multi-scale deep disturbance neural network | |
| CN114049305A (en) | Distribution line pin defect detection method based on improved ALI and fast-RCNN | |
| CN115761537A (en) | Power transmission line foreign matter intrusion identification method oriented to dynamic characteristic supplement mechanism | |
| CN114219763A (en) | Infrared picture detection method for abnormal heating point of power distribution equipment based on fast RCNN algorithm | |
| CN117333948A (en) | End-to-end multi-target broiler behavior identification method integrating space-time attention mechanism |
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 |