Markov conversion and improved ResNet18 model-based insulator contamination identification method
The Markov transition and improved ResNet18 model encode leakage current dynamics into images for enhanced feature extraction, addressing the limitations of linear feature-based methods and enabling automated insulator pollution state recognition.
Patent Information
- Application Number
- CN202510383150.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-07-15
AI Technical Summary
The prior art is difficult to accurately and automatically identify the dirty state of the composite insulator surface. It is greatly affected by environmental factors and relies on manual experience to ensure the safety and reliability of the power system quickly and effectively.
The combination of Markov conversion and improved ResNet18 model is adopted to convert leakage current data into Markov conversion field images, and combined with the convolutional block attention module, an insulator filth state recognition model is constructed to realize automated recognition.
It realizes rapid and accurate identification of the filthy state of composite insulators, reduces manual intervention, improves the reliability and speed of identification, and ensures the safety and reliability of the power system.
Smart Images
Figure CN120316613A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of detecting and identifying the surface pollution state of insulators, and particularly to an insulator pollution identification method based on Markov transition and improved ResNet18 model. Background Art
[0002] In modern power systems, composite insulators have been widely used in high-voltage and extra-high-voltage power grids due to their superior electrical performance, lightweight design, and good anti-aging characteristics. These composite insulators not only play an important role in improving the safety and reliability of transmission lines, but also promote the sustainable development of the power industry due to their low maintenance costs. However, composite insulators face serious pollution problems during actual operation. Insulator pollution mainly consists of components such as dust, salts, and industrial pollutants. These pollutants deposit on the surface of the insulator, and over time, these impurities gradually accumulate, resulting in a decline in the electrical performance of the composite insulator and shortening its service life. Corona discharge or flashover phenomena can cause serious electrical faults and even lead to the outage of the entire line. Therefore, in order to ensure power supply stability and reliability, it is of great practical significance to timely identify the pollution state of composite insulators.
[0003] Analysis of the leakage current on the insulator surface is one of the effective methods for on-line detecting and identifying the pollution state of insulators. It does not require time-consuming and power-off operations like the equivalent salt deposit density (ESDD) estimation and non-soluble deposit density (NSDD) estimation, so it has received extensive attention from researchers at home and abroad. Currently, scholars mainly use certain characteristics of the leakage current as monitoring indicators to evaluate the pollution state of insulators. These indicators include time-domain statistical indicators of the leakage current, the ratio of the third harmonic to the fifth harmonic content, the phase difference between voltage and current, and the high-order harmonic content, etc. However, most of these characteristic indicators are based on the linear characteristics of the leakage current and are analyzed in the time domain or frequency domain. Since the leakage current signal of the insulator is a non-linear time series with complex dynamic change characteristics and is affected by various factors such as input voltage, climate conditions, and pollution degree, it is difficult to accurately describe the non-linear characteristics of the leakage current signal by only analyzing its time-domain or frequency-domain information. In addition, the insulator pollution state identification method based on characteristic indicators often requires professionals to compare the change rules of the indicators, which is greatly affected by manual experience and cannot automatically obtain the pollution state of the composite insulator.
[0004] Therefore, it is necessary to design an insulator pollution identification method based on Markov transition and improved ResNet18 model to overcome the above problems. Summary of the Invention
[0005] To avoid the above problems, a method for identifying the contamination of insulators based on Markov transformation and improved ResNet18 model is provided, which fully excavates the spatial hidden information such as the microscopic fluctuations and macroscopic patterns of the leakage current signal and combines deep learning methods to achieve a more comprehensive, accurate and automated identification of the contamination state, without being affected by external factors such as environmental temperature and air humidity, which is conducive to the staff to timely grasp the contamination state of composite insulators, thus ensuring the safety and reliability of the power system.
[0006] A method for identifying the contamination of insulators based on Markov transformation and improved ResNet18 model provided by the present invention includes the following steps:
[0007] S1. Obtain the leakage current data of the composite insulator in different contamination states;
[0008] S2. Preprocess the leakage current data of the composite insulator;
[0009] S3. Obtain the Markov transformation field images of the composite insulator in different contamination states;
[0010] S4. Train a contamination state identification model based on the improved ResNet18 model;
[0011] S5. Test the contamination state identification model of the composite insulator;
[0012] S6. Identify the contamination state of the composite insulator to be identified.
[0013] Preferably, step S1 specifically includes: measuring the leakage current of the composite insulator in a controlled laboratory to obtain the leakage current data of the composite insulator in different contamination states; simultaneously measuring the equivalent salt deposit density value of the insulator, and classifying the contamination state of the insulator into five contamination levels: very light, light, moderate, severe and very severe according to the equivalent salt deposit density value; by changing the environmental temperature and air humidity in the controlled laboratory, measuring the contaminated samples of the insulator multiple times until the required amount of leakage current data with known contamination states is obtained.
[0014] Preferably, step S2 specifically includes: using the ensemble empirical mode decomposition algorithm to adaptively decompose the leakage current data to obtain multiple intrinsic mode components, then calculating the correlation coefficients between each mode component and the original current data, and selecting the mode components with correlation coefficients greater than 0.4 for signal reconstruction to obtain the denoised leakage current data; assuming that the number of observed data points of the denoised leakage current in one cycle is N, then normalizing the leakage current observation values Y = {y1, y2,... y N} according to the following formula:
[0015]
[0016] Wherein, m is 1, …, N; x m is the m-th data point of the leakage current after normalization; y m is the m-th data point in the sequence Y; max(Y) and min(Y) are the maximum value and the minimum value in the sequence Y, respectively.
[0017] Preferably, step S3 specifically includes:
[0018] 3.1 Divide the normalized leakage current observation values X = {x1, x2, … x N} into Q quantile bins according to the "quantile binning" strategy, and the corresponding quantile region for each data point is q j , j ∈ [1, Q], and construct a transition matrix W with dimensions Q × Q. The construction formula is as follows:
[0019]
[0020] w ij = p i,j (x t ∈ q i |x t-1 ∈ q j ),
[0021] Wherein, t is the moment corresponding to a certain data point in the leakage current observation value X; w ij is the value of the i-th row and j-th column of the transition matrix W, where i, j ∈ [1, Q], and i - j > 1, indicating that the data point located in the q j quantile region at the moment t - 1 jumps to the q i quantile region at the next moment t; x t-1 and x t represent the data points of the leakage current at the moments t - 1 and t, respectively;
[0022] 3.2 Expand the transition matrix W into a Markov transition domain M with dimensions N × N to form a Markov transition field image. The expansion process is as follows:
[0023]
[0024] Wherein, P i,j represents the transition probability from the quantile q i to the quantile q j .
[0025] Preferably, step S4 specifically includes: dividing the Markov transition field images under different contamination states into a training set, a validation set, and a test set; embedding a convolutional block attention module after the first layer of convolution and the last layer of convolution of the original ResNet18 model to improve the feature extraction ability, thereby constructing an improved ResNet18 model based on the convolutional attention module. Then, use the training set with known contamination levels to input into the improved ResNet18 model for training to obtain a contamination state recognition model for composite insulators.
[0026] Preferably, step S5 specifically includes: using the test set to test the obtained contamination state recognition model to obtain the recognition accuracy of the contamination state of the composite insulator.
[0027] Preferably, step S6 specifically includes: for composite insulators of the same model with unknown contamination states, convert the measured leakage current into a Markov transition field image according to steps S2 and S3, and input the Markov transition field image into the trained contamination state recognition model to obtain the contamination state of the composite insulator.
[0028] Compared with the prior art, the present invention has the following beneficial effects: The present invention encodes the dynamic evolution relationship of the leakage current of the composite insulator into a two-dimensional image through the Markov transition field technology to intuitively display the microscopic fluctuations and macroscopic patterns of the leakage current and fully excavate the hidden information of the leakage current. On this basis, combined with the improved ResNet18 model based on the convolutional block attention module to identify the contamination state; compared with other traditional methods of converting one-dimensional data into two-dimensional images such as short-time Fourier transform, Gram angle field, recurrence plot, etc., the Markov transition field technology can transform the statistical evolution characteristics of the time series into a spatial probability distribution, and has stronger anti-noise ability while retaining the time dependence. The method of the present invention can quickly and accurately perform automatic identification of the contamination state of the composite insulator, is not affected by external factors such as environmental temperature and air humidity, is conducive to the staff to timely master the contamination state of the composite insulator, and thus ensures the safety and reliability of the power system. Description of the Drawings
[0029] Figure 1 It is a flowchart of an insulator contamination identification method based on Markov transition and improved ResNet18 model according to a preferred embodiment of the present invention;
[0030] Figure 2 It is a comparison diagram of the original leakage current and the current waveform after denoising by the ensemble empirical mode decomposition algorithm in a preferred embodiment of the present invention;
[0031] Figure 3 It is a typical Markov transition field diagram under different contamination levels in a preferred embodiment of the present invention;
[0032] Figure 4 In a preferred embodiment of the present invention, it is a comparison chart of the loss values between the improved ResNet18 model and the original ResNet18 model.
[0033] Figure 5 In a preferred embodiment of the present invention, it is the confusion matrix result of the test set for the pollution state recognition model. Detailed implementation manners
[0034] As Figures 1 to 3 shown, an insulator pollution recognition method based on Markov transition and improved ResNet18 model provided in this embodiment encodes the dynamic evolution relationship of the leakage current signal of the composite insulator into a Markov transition field image to intuitively display the microscopic fluctuations and macroscopic patterns of the leakage current and fully exploit the hidden information of the leakage current, and then realizes the automatic recognition of the pollution state of the composite insulator through the following steps, specifically including:
[0035] S1. Obtain the leakage current data of the composite insulator under different pollution states
[0036] Measure the leakage current of the composite insulator in a controlled laboratory to obtain the leakage current data of the composite insulator under different pollution states. At the same time, measure the equivalent salt deposit density value of the insulator according to the regulations in GB / T 16434-1996, and divide the pollution state of the insulator into five pollution levels: very light, light, medium, severe, and very severe according to the equivalent salt deposit density value; by changing the environmental temperature and air humidity in the controlled laboratory, measure the polluted insulator samples multiple times until the required amount of leakage current data with pollution state marks is obtained.
[0037] S2. Preprocess the leakage current data of the composite insulator
[0038] Use the ensemble empirical mode decomposition algorithm to adaptively decompose the leakage current data to obtain multiple intrinsic mode components, and then calculate the correlation coefficients of each mode component and the original current data, and select the mode components with correlation coefficients greater than 0.4 for signal reconstruction to obtain the denoised leakage current data; assume that the number of observed data points of the denoised leakage current in one cycle is N, and then normalize the observed values Y = {y1, y2,... y N} according to formula (1):
[0039]
[0040] In the formula, m is 1,..., N; x m is the m-th data point of the leakage current after normalization; y mis the m-th data point in sequence Y; max(Y) and min(Y) are the maximum and minimum values in sequence Y respectively.
[0041] S3. Obtain the Markov transition field images of composite insulators under different pollution states
[0042] The image coding method based on the Markov transition matrix can map one-dimensional leakage current data to the Markov transform domain. The specific steps are as follows:
[0043] (1) Divide the normalized leakage current observations X = {x1, x2,... x N} into Q quantile bins according to the "quantile binning" strategy. The corresponding quantile region for each data point is q j , j ∈ [1, Q], and construct a transition matrix W with dimension Q×Q. The construction formula is as follows:
[0044]
[0045] w ij = P i,j (x t ∈ q i | x t-1 ∈ q j ) (3)
[0046] In the formula, t is the moment corresponding to a certain data point in the leakage current observations X; w ij is the value of the i-th row and j-th column of the transition matrix W, where i, j ∈ [1, Q], and i - j > 1, indicating the probability that the data point located in the q j quantile region at time t - 1 jumps to the q i quantile region at the next time t; x t-1 and x t represent the data points of the leakage current at time t - 1 and time t respectively.
[0047] (2) Expand the transition matrix W into a Markov transform domain M with dimension N×N to form a Markov transition field image. The expansion process is as follows:
[0048]
[0049] In the formula, P i,j represents the transition probability from quantile q i to quantile q j .
[0050] S4. Train a pollution state recognition model based on the improved ResNet18 model
[0051] Obtain the Markov transition field images of all leakage current data through step S3, and divide the Markov transition field images under different pollution states into a training set, a validation set, and a test set according to a ratio of 6:2:2; embed the convolutional block attention module after the first layer of convolution and the last layer of convolution of the original ResNet18 model to improve the feature extraction ability, and thus construct an improved ResNet18 model based on the convolutional attention module. Then, use the training set marked with pollution levels to input into the improved ResNet18 model for training to obtain a pollution state recognition model for composite insulators.
[0052] S5. Test and apply the pollution state recognition model for composite insulators
[0053] Use the test set to test the obtained pollution state recognition model to obtain the recognition accuracy of the pollution state of the composite insulator; for a composite insulator of the same model with an unknown pollution state, convert the measured leakage current into a Markov transition field image according to steps two and three, and input the Markov transition field image into the trained pollution state recognition model to automatically output the pollution state of the composite insulator.
[0054] To further illustrate the effectiveness and practicality of the method of the present invention, the following conducts a detailed analysis in combination with an example of the pollution state recognition method of a certain type of silicone rubber composite insulator:
[0055] First, artificially pollute the insulator samples in the laboratory by the "solid layer method". In this method, a slurry made of NaCl, kaolin, and distilled water is evenly coated on the surface of the insulator skirt with a brush. The equivalent salt deposit density value of the insulator can be controlled by changing the content of NaCl and kaolin, and according to the regulations in Standard GB / T 16434-1996, the equivalent salt deposit density value ESDD is respectively set to five intervals of 0 < ESDD ≤ 0.03, 0.03 < ESDD ≤ 0.06, 0.06 < ESDD ≤ 0.10, 0.10 < ESDD ≤ 0.25, and ESDD > 0.25, corresponding to five insulator pollution levels of very light, light, moderate, severe, and very severe respectively.
[0056] Then, place the composite insulators under different pollution levels in a controlled laboratory, continuously change the environmental temperature and air humidity, and use a digital oscilloscope to record the leakage current data of the composite insulators under the rated voltage. The sampling frequency of the digital oscilloscope is 500 kHz, that is, the number of data points N in each cycle = 10000. By changing the environmental parameters, the polluted samples of the insulators are measured multiple times. For each pollution state, the leakage current data of 600 cycles are measured, that is, there are a total of 3000 groups of leakage current data.
[0057] Next, the ensemble empirical mode decomposition algorithm is used to adaptively decompose the leakage current data to obtain multiple intrinsic mode components. The correlation coefficient between each mode classification and the original current data is calculated, and the mode components with a correlation coefficient greater than 0.4 are selected for signal reconstruction to obtain the denoised leakage current data. Attachment Figure 2 Figure shows the comparison diagram of a certain original leakage current and the current waveform after denoising by the ensemble empirical mode decomposition algorithm under the moderate pollution level. From the attachment Figure 2 It can be seen that the original leakage current exhibits obvious fluctuations and irregularities, accompanied by frequent spikes and mutations. These burrs not only make the overall contour of the signal blurred, resulting in a significant decrease in the smoothness and recognizability of the signal. In contrast, the leakage current waveform after denoising by the ensemble empirical mode decomposition algorithm is more regular, and the irregular burrs caused by noise are significantly reduced, so it can better reflect the pollution characteristics of the insulator itself.
[0058] After normalizing the leakage current according to formula (1), the one-dimensional leakage current data is mapped into a Markov transition field image according to step S3. In our research, the "quantile binning" strategy is adopted, and the number of quantile bins q = 3. The generated image pixels are fixed at 300×300 dpi. In this way, the Markov transition field images of all leakage currents can be obtained. As an example, attachment Figure 3 shows typical Markov transition field images under different pollution states. It can be seen that the Markov transition field images under each pollution state show their specific shapes and contours, which essentially reflect the macroscopic mode and microscopic fluctuation characteristics of the leakage current signal. By extracting these image features, they can be used to identify the pollution state of the insulator.
[0059] The Markov transition field images of all leakage current data are obtained through step S3. There are 600 Markov transition field images under each pollution state, and a total of 3000 images. Then, they are divided into a training set, a validation set, and a test set according to the ratio of 6:2:2, that is, the number of Markov transition field images in the training set, validation set, and test set under each pollution state is 360, 120, and 120. In the Matlab software, the convolutional block attention module is embedded after the first convolution and the last convolution of the original ResNet18 model to improve the feature extraction ability, and the improved ResNet18 model based on the convolutional attention module is constructed. On this basis, the training set marked with the pollution level is input into the improved ResNet18 model for training, and the validation set is used for verification to obtain the pollution state recognition model of the composite insulator. During the training process of the model, the Adam optimizer with adaptive learning is used to update the network parameters, the initial learning rate is set to 0.004, and the batch size is 16. Attachment Figure 4It shows a comparison graph of the loss value curves between the improved ResNet18 model and the original ResNet18 model. It can be seen that the loss value of the improved ResNet18 model converges faster and is lower after reaching stability, indicating that the feature extraction ability of the original ResNet18 model is improved by introducing the convolutional attention module. After the model reaches stability, the fitting of the data distribution is more accurate, making the model perform better.
[0060] To evaluate the performance of the insulator contamination status recognition model based on the improved ResNet18 model, a confusion matrix is used to present the results of the test set and evaluate the accuracy of the model. The obtained results are as follows Figure 5 shown. The abscissa in the confusion matrix is the actual target label, and the ordinate represents the network output label. The numbers 1-5 in the figure correspond to the above 5 contamination levels respectively. This figure shows that there are only 16 misclassifications in total for the improved ResNet18 model, and the average recognition accuracy is 97.3%. This proves the effectiveness and strong performance of the method of the present invention in evaluating the contamination status of composite insulators. When this recognition model is applied to composite insulators of the same type with unknown contamination status, the measured leakage current can be converted into a Markov transition field image according to steps S2 and S3 and input into the trained contamination status recognition model, and then the model can automatically output the contamination status of the composite insulator. The method of the present invention simplifies the process of identifying the contamination status of composite insulators, thus significantly improving the judgment speed, enabling on-site workers to quickly obtain the status information of the insulators, reducing the time cost of manual inspection, improving the reliability of recognition, and being beneficial to optimizing the management and maintenance strategies of power equipment.
[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. An insulator pollution identification method based on Markov transformation and improved ResNet18 model, characterized in that, It includes the following steps: S1. Obtain the leakage current data of the composite insulator under different pollution states; S2. Preprocess the leakage current data of the composite insulator; S3. Obtain the Markov transition field images of the composite insulator under different pollution states; S4. Train a pollution state recognition model based on the improved ResNet18 model; S5. Test the pollution state recognition model of the composite insulator; S6. Recognize the pollution state of the composite insulator to be recognized.
2. The insulator pollution identification method based on Markov transition and improved ResNet18 model according to claim 1, characterized in that: Step S1 specifically includes: Measuring the leakage current of the composite insulator in a controlled laboratory to obtain the leakage current data of the composite insulator under different pollution states; At the same time, measuring the equivalent salt deposit density value of the insulator, and classifying the pollution state of the insulator into five pollution levels: very light, light, moderate, severe, and very severe according to the equivalent salt deposit density value; By changing the environmental temperature and air humidity in the controlled laboratory, measuring the contaminated samples of the insulator multiple times until the required amount of leakage current data with known pollution states is obtained.
3. The insulator contamination identification method based on Markov transition and improved ResNet18 model as described in claim 1, characterized in that: Step S2 specifically includes: adaptively decomposing the leakage current data by using the ensemble empirical mode decomposition algorithm to obtain multiple intrinsic mode components, then calculating the correlation coefficients between each mode component and the original current data, and selecting the mode components with correlation coefficients greater than 0.4 for signal reconstruction to obtain the denoised leakage current data; assuming that the number of observed data points of the denoised leakage current in one cycle is N, and then normalizing the leakage current observation values Y = {y1, y2, … y N} according to the following formula: where m is 1, …, N; x m is the m-th data point of the leakage current after normalization; y m is the m-th data point in the sequence Y; max(Y) and min(Y) are the maximum and minimum values in the sequence Y, respectively.
4. The insulator pollution identification method based on Markov transformation and improved ResNet18 model as described in claim 3, characterized in that: Step S3 specifically includes: 3.1 Divide the normalized leakage current observations X = {x1, x2, … x N} into Q quantile bins according to the "quantile binning" strategy, and the corresponding quantile region for each data point is q j , j ∈ [1, Q], and construct a transition matrix W with dimensions Q×Q. The construction formula is as follows: w ij = p i,j (x t ∈ q i | x t-1 ∈ q j ), where t is the moment corresponding to a certain data point in the leakage current observation value X; w ij is the value of the i-th row and j-th column of the transition matrix W, where i, j ∈ [1, Q], and i - j > 1, indicating that the data point located in the q j quantile region at the moment t - 1 jumps to the q i quantile region at the next moment t; x t-1 and x t respectively represent the data points of the leakage current at the moments t - 1 and t; 3.2 Expand the transition matrix W into a Markov transition domain M with a dimension of N×N to form a Markov transition field image. The expansion process is as follows: where P i,j represents the quantile q i transition to the quantile q j is the transition probability.
5. The insulator pollution identification method based on Markov transition and improved ResNet18 model according to claim 1, characterized in that: Step S4 specifically includes: Divide the Markov transition field images under different pollution states into a training set, a validation set, and a test set; Embed the convolutional block attention module after the first convolution and the last convolution of the original ResNet18 model to improve the feature extraction ability, and thus construct an improved ResNet18 model based on the convolutional attention module. Then, use the training set with known pollution levels to input into the improved ResNet18 model for training to obtain the pollution state recognition model of the composite insulator.
6. The insulator pollution identification method based on Markov transformation and improved ResNet18 model as described in claim 5, characterized in that: Step S5 specifically includes: Using the test set to test the obtained pollution state recognition model to obtain the recognition accuracy of the pollution state of the composite insulator.
7. The insulator pollution identification method based on Markov transition and improved ResNet18 model as described in claim 1, wherein: Step S6 specifically includes: For composite insulators of the same model with unknown pollution states, convert the measured leakage current into a Markov transition field image according to steps S2 and S3, and input the Markov transition field image into the trained pollution state recognition model to obtain the pollution state of the composite insulator.