Method for predicting aging degree of anti-corona layer of stator bar

Abstract

The invention discloses a method for predicting the aging degree of an anti-corona layer of a stator bar, and relates to the field of prediction of the aging degree of the anti-corona layer of the stator bar, and the method comprises the following steps: S1, data sampling; image data and real-time temperature information are obtained; s2, data processing; preprocessing the acquired image data to obtain a data set, and performing evaluation weight division and state division on the data set; s3, constructing a model; constructing a model by using a convolutional neural network; s4, training the model; s5, testing the model; s6, practical application; and predicting the aging degree of the anti-corona layer of the stator bar by using the anti-corona layer life prediction model. According to the method, the machine learning algorithm based on learning is adopted, the infrared thermal image, the ultraviolet sub-image and the temperature data in the aging process of the anti-corona layer of the stator bar are collected, the aging state of the stator bar is classified by applying multi-modal feature fusion, and the method has a good guiding effect on evaluation of the aging degree of the anti-corona layer of the stator bar.

Description

Technical Field

[0001] The invention relates to the field of prediction of aging degree of stator wire rod anti-corona layer, in particular to a method for predicting aging degree of stator wire rod anti-corona layer. Background Art

[0002] As an important part of the generator, the stator bar plays a vital role in the stable operation of the entire unit. The state of its anti-corona layer largely determines the service life of the generator. The stator coil anti-corona layer is located on the surface of the stator coil to prevent corona discharge; when there is a high-voltage electric field in the air, corona discharge will be generated on the coil surface, which will damage the insulation, shorten the life of the motor, and even cause motor failure. The stator coil anti-corona layer is covered on the coil surface in the form of conductive paint or metal foil to disperse the electric field and prevent corona discharge.

[0003] The material of the anti-corona layer is usually an organic substance, which is susceptible to moisture, pollutants and ultraviolet rays, and is prone to aging and deterioration. Over time, the anti-corona layer may crack, fall off, discolor, etc., which may cause the electrical performance of the motor to deteriorate or even cause motor failure. Thermal aging is a major form of aging of the anti-corona layer material; that is, under long-term high temperature conditions, the molecular activity increases, the molecular structure of the material may break, degrade or decompose, and some materials may even undergo oxidation and corrosion reactions. These phenomena may cause the insulation resistance to decrease and the dielectric constant to increase, resulting in a decrease in the performance of the anti-corona layer.

[0004] In the field of stator wire rod aging degree prediction, the technical personnel in this field are currently focusing on the prediction of the aging degree of the insulation layer of the stator wire rod. For example, the Chinese patent with publication number "CN111553432A" discloses "A method for predicting the aging degree of the insulation of a stator wire rod based on image feature support vector machine", the Chinese patent with publication number "CN111563622A" discloses "A method for predicting the aging degree of the insulation of a stator wire rod based on gray level co-occurrence matrix and deep learning", and the Chinese patent with publication number "CN113569507A" discloses "A method for predicting the aging degree of the insulation of a stator wire rod based on gray level co-occurrence matrix and deep learning". A composite prediction method for the insulation aging status of stator wire bars based on machine learning"; The insulation layer of the stator wire bar is used to isolate the parts between different voltage levels in the motor to prevent current short circuit and breakdown. Usually, special materials are used to cover the coils and other electrical components to prevent dust, moisture and pollutants in the external environment from invading the interior of the motor, while also providing dielectric properties to isolate the parts between different voltage levels; that is, the insulation layer and the anti-corona layer are different structures on the stator wire bar, so those skilled in the art cannot apply the prediction method for the aging degree of the insulation layer to the prediction of the aging degree of the anti-corona layer.

[0005] Therefore, in the prediction of the aging degree of the anti-corona layer, a new prediction method is needed in this field. Summary of the invention

[0006] The purpose of the invention is to provide a method for predicting the aging degree of the anti-corona layer of a stator wire bar in response to the above-mentioned problems. A learning-based machine learning algorithm is used to collect infrared thermal images, ultraviolet photon images and temperature data during the aging process of the anti-corona layer of the stator wire bar, and multimodal feature fusion is used to classify the aging state of the stator wire bar. The method has accurate prediction results and plays a good guiding role in evaluating the aging degree of the anti-corona layer of the stator wire bar.

[0007] The technical solution adopted by the present invention is as follows: A method for predicting the aging degree of the anti-corona layer of a stator wire bar comprises the following steps:

[0008] S1: Data sampling; the stator wire bar is continuously powered on for heating, and the infrared thermal image and ultraviolet photon image of the anti-corona layer of the stator wire bar are respectively captured by an infrared camera and an ultraviolet camera to obtain image data; the temperature distribution cloud map of the stator wire bar is monitored in real time, and the temperature information of the wire bar is obtained by a sensor;

[0009] S2: Data processing: preprocess the acquired image data to obtain a data set, and divide the data set into evaluation weights and states;

[0010] S3: construct a model; divide the data set obtained in step S2 into a training set and a test set, extract features from the training set using the pre-trained model to obtain a feature vector for each image, and use a convolutional neural network to fuse the feature vector with its corresponding temperature information;

[0011] S4: training model; taking the fused training set as input, through the multi-layer perception layer, classifying all images in the training set according to the stages corresponding to the images, adjusting the model parameters to maximize the classification accuracy, and obtaining a trained anti-corona layer life prediction model;

[0012] S5: Test model; input the test set into the trained anti-corona layer life prediction model for test verification, compare the predicted life result with the real label, and calculate the recognition accuracy of the test set; if the accuracy meets expectations, enter the actual application; if it does not meet expectations, continue to improve the model recognition accuracy by increasing the batch size and adjusting the function coefficient until it meets expectations;

[0013] S6: Practical application: Use the anti-corona layer life prediction model to predict the aging degree of the stator wire bar anti-corona layer.

[0014] Furthermore, in step S2, the preprocessing process includes image size unification processing, elimination of data that does not match the time, and mask processing.

[0015] Furthermore, the image size unification process is to adjust the size and proportion of the images so that all images have the same size in order to solve the problem that the sizes of the infrared thermal image and the ultraviolet photon image are inconsistent.

[0016] Furthermore, when removing data with mismatched time, a time alignment operation should be performed.

[0017] Furthermore, during mask processing, for the infrared thermal image, the area with the highest temperature of the wire rod is masked; for the ultraviolet photon image, mask processing is performed to remove irrelevant text parts.

[0018] Further, in step S2, the weight division process includes steps S21 to S24; wherein:

[0019] S21: Calculate the status score value of each evaluation parameter;

[0020] S22: the health status of each evaluation parameter is divided into four health levels, namely, serious failure, failure, sub-health and health;

[0021] S23: By substituting the state score value into the membership function calculation formula, the weight of the state score value of each evaluation parameter belonging to the four health levels can be calculated, thereby obtaining the fuzzy relationship matrix of the evaluation parameter set;

[0022] S24: For the classification of the aging status of the anti-corona layer, a comprehensive weighted method is adopted. First, the health status of multiple characteristic parameters of the stator wire bar anti-corona layer is evaluated, and then the variable weight coefficient matrix and single factor evaluation matrix of each factor are obtained by using the variable weight theory. Then, these two matrices are multiplied to obtain the fuzzy relationship matrix of the system health status.

[0023] Further, in step S3, the fusion process includes steps S31 and S32; wherein:

[0024] S31: perform dimensionality reduction processing on the infrared thermal map in the preprocessed data set and convert it into a one-dimensional feature vector; extract the GLCM features and HOG features of the ultraviolet photon map of the infrared thermal map in the preprocessed data set and splice them;

[0025] S32: Use the convolutional neural network VGG-16 to stitch the obtained image features and temperature information.

[0026] Furthermore, in step S3, the preprocessed data set is divided into a training set and a test set by a random partitioning method.

[0027] Further, in step S4, the classification includes steps S41 and S42; wherein:

[0028] S41: The aging state of the stator bar anti-corona layer is used as a label to divide it into four stage categories;

[0029] S42: Input the fused features into a multi-layer perceptron for training to obtain a trained anti-corona layer life prediction model.

[0030] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0031] The present invention uses feature fusion to analyze infrared thermal images, ultraviolet photon images and temperature information. The method can use a classification model to predict the aging degree of the stator wire rod anti-corona layer. Compared with traditional prediction technology, the present invention uses a multi-image feature fusion method for modeling. By fusing multiple images together, richer and more diverse feature expressions can be obtained, reducing noise and interference in a single image, thereby improving the robustness of image classification, offsetting errors or anomalies in individual images, and improving classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The present invention is a flowchart of the aging degree prediction method disclosed in the present invention. DETAILED DESCRIPTION

[0033] A method for predicting the aging degree of the anti-corona layer of a stator wire rod is based on infrared thermal images, ultraviolet photon images and temperature information, and a one-dimensional feature vector is obtained by dimensionality reduction, and a convolutional neural network is used to fuse image data and text data, and then a multi-layer perception layer is used to classify the image in stages; the method includes the following steps (the flowchart is shown in FIG. Figure 1 shown):

[0034] Step S1: data sampling;

[0035] The stator wire rod is continuously powered on for heating, and the infrared thermal image and ultraviolet photon image of the anti-corona layer of the stator wire rod are captured by infrared camera and ultraviolet camera to obtain image data; the temperature distribution cloud map of the stator wire rod is monitored in real time and the data is recorded in the form of text, and the temperature information of the wire rod is obtained by sensor.

[0036] Step S2: data processing;

[0037] The acquired image data is preprocessed to obtain a data set, and the data set is divided into evaluation weights and states; specifically, the preprocessing process includes image size unification, elimination of data that does not match time, and mask processing.

[0038] Image size unification processing is used to unify the sizes of infrared thermal images and ultraviolet photon images. By adjusting the size and proportion of the images, all images have the same size for accurate comparison and analysis.

[0039] Eliminate data that does not match the time, perform time alignment operations, and eliminate data that does not match the time in the data.

[0040] Mask processing: for infrared thermal images, mask the places with the highest wire rod temperature; perform mask processing on ultraviolet photon images to remove irrelevant text parts; specifically, select the places with the highest wire rod temperature in each infrared thermal image for masking, and mask its size to 100*60; select each ultraviolet photon image for masking to remove irrelevant text parts; remove features that do not match the other two items in infrared thermal images, ultraviolet photon images, and temperature text information.

[0041] The weight division process is as follows, including steps S21 to S24; wherein:

[0042] S21: Calculate the status score value of each evaluation parameter;

[0043] S22: the health status of each evaluation parameter is divided into four health levels, namely, serious failure, failure, sub-health and health;

[0044] S23: By substituting the state score value into the membership function calculation formula, the weight of the state score value of each evaluation parameter belonging to the four health levels can be calculated, thereby obtaining the fuzzy relationship matrix of the evaluation parameter set;

[0045] S24: For the classification of the aging status of the anti-corona layer, a comprehensive weighted method is adopted. First, the health status of multiple characteristic parameters of the stator wire bar anti-corona layer is evaluated, and then the variable weight coefficient matrix and single factor evaluation matrix of each factor are obtained by using the variable weight theory. Then, these two matrices are multiplied to obtain the fuzzy relationship matrix of the system health status.

[0046] In step S24, the fuzzy relationship matrix S of the system health status i The formula is as follows:

[0047]

[0048] Where: S i1 , S i2 , S i3 , S i4 It indicates the weight of the system health status in the i-th state for the four status levels of severe fault, fault, sub-health, and health. The value range is 0-1. The larger the value, the better the operating status.

[0049] Step S3: constructing a model;

[0050] The data set obtained in step S2 is divided into a training set and a test set. The preprocessed data set is divided into a training set and a test set by a random partitioning method, in which the training set accounts for 80% and the test set accounts for 20%. The training set is feature extracted by the pre-trained model to obtain a feature vector for each image, and the feature vector is fused with its corresponding temperature information using a convolutional neural network.

[0051] Specifically, the fusion process includes steps S31 and S32, wherein:

[0052] S31: Perform dimensionality reduction processing on the infrared thermal image in the preprocessed data set and convert it into a one-dimensional feature vector; extract the GLCM features and HOG features of the infrared thermal image ultraviolet photon image in the preprocessed data set and splice them; specifically, perform feature extraction on the acquired multiple infrared thermal images and ultraviolet photon image training sets through a pre-trained model; perform dimensionality reduction on the infrared thermal image obtained by masking and convert it into a one-dimensional feature vector; extract the GLCM features and HOG features of the ultraviolet photon image obtained by masking; thus, obtain a one-dimensional feature vector for each image.

[0053] S32: Use the convolutional neural network VGG-16 to splice the obtained image features and temperature information; specifically, use the convolutional neural network to fuse the feature vector one-dimensional infrared features, ultraviolet GLCM features, ultraviolet HOG features and temperature text features to construct an anti-corona layer life prediction model.

[0054] Step S4: training model;

[0055] The fused training set is used as input, and all images in the training set are classified according to the stages corresponding to the images through a multi-layer perception layer. The model parameters are adjusted to maximize the classification accuracy, and a trained anti-corona layer life prediction model is obtained.

[0056] Wherein, the classification includes step S41-step S42;

[0057] S41: The aging state of the anti-corona layer of the stator wire bar is used as a label to divide it into four stage categories; specifically, according to the size of the training data set and the data fitting results, the hyperparameters and bias values ​​are set and adjusted to make the trained model perform best on the validation set, and the images are input into the model to obtain the stage division results of each image, and the loss is calculated by comparing with the real label and back propagation optimization is performed, and then the correct division ratio is statistically calculated to obtain the changes in the loss value and accuracy of the training set stage division;

[0058] S42: Input the fused features into a multi-layer perceptron for training to obtain a trained anti-corona layer life prediction model.

[0059] Step S5: testing the model;

[0060] The test set is input into the trained anti-corona layer life prediction model for test verification. The model loads the optimal model parameters, inputs the test set data, compares the predicted life results with the real labels, and calculates the test set recognition accuracy; if the accuracy meets expectations, it will enter the actual application; if it does not meet expectations, the model recognition accuracy will be further improved by increasing the batch size and adjusting the function coefficients until it meets expectations. The test set results are shown in Table 1:

[0061] Table 1 Test set results

[0062] Accuracy (%) Multi-image fusion method 100

[0063] From the results in Table 1, it can be seen that the multi-image fusion method has a very high accuracy in predicting the aging state of given data.

[0064] Step S6: practical application;

[0065] The anti-corona layer life prediction model is used to predict the aging degree of the stator wire bar anti-corona layer. Specifically, in actual use, the application mode can be single-machine operation or "client-server" operation. After the front-end interface design is completed and the back-end model loads the optimal model parameters, it can continuously recognize multiple sets of data passed in by each user. The more input image data and its corresponding text data, the more beneficial it is to accurate recognition results.

[0066] In summary, the present invention uses feature fusion to analyze infrared thermal images, ultraviolet photon images and temperature information. This method can use a classification model to predict the aging degree of the stator wire rod anti-corona layer. Compared with traditional prediction technology (traditional prediction technology has slow training speed, is prone to fall into local optimal solutions, is sensitive to initial weights and biases, requires a large amount of training data and is prone to overfitting), the present invention uses a multi-image feature fusion method for modeling. By fusing multiple images together, richer and more diverse feature expressions can be obtained, reducing noise and interference in a single image, thereby improving the robustness of image classification, offsetting errors or anomalies in individual images, and improving classification accuracy.

[0067] The present invention is not limited to the above-mentioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A method for predicting the aging degree of a stator bar anti-corona layer, characterized in that: The following steps are involved: S1: Data sampling; the stator wire bar is continuously powered on for heating, and the infrared thermal image and ultraviolet photon image of the anti-corona layer of the stator wire bar are respectively captured by an infrared camera and an ultraviolet camera to obtain image data; the temperature distribution cloud map of the stator wire bar is monitored in real time, and the temperature information of the wire bar is obtained by a sensor; S2: Data processing: preprocess the acquired image data to obtain a data set, and divide the data set into evaluation weights and states; S3: construct a model; divide the data set obtained in step S2 into a training set and a test set, extract features from the training set using the pre-trained model to obtain a feature vector for each image, and use a convolutional neural network to fuse the feature vector with its corresponding temperature information; S4: training model; taking the fused training set as input, through the multi-layer perception layer, classifying all images in the training set according to the stages corresponding to the images, adjusting the model parameters to maximize the classification accuracy, and obtaining a trained anti-corona layer life prediction model; S5: Test model; input the test set into the trained anti-corona layer life prediction model for test verification, compare the predicted life result with the real label, and calculate the recognition accuracy of the test set; if the accuracy meets expectations, enter the actual application; if it does not meet expectations, continue to improve the model recognition accuracy by increasing the batch size and adjusting the function coefficient until it meets expectations; S6: Practical application: Use the anti-corona layer life prediction model to predict the aging degree of the stator wire bar anti-corona layer.

2. The aging degree prediction method according to claim 1, characterized in that: In step S2, the preprocessing process includes image size unification, elimination of data that does not match the time, and mask processing.

3. The aging degree prediction method according to claim 2, characterized in that: Image size unification is used to adjust the size and proportion of the images so that all images have the same size.

4. The aging degree prediction method according to claim 2, characterized in that: When removing data with mismatched time, time alignment should be performed.

5. The aging degree prediction method according to claim 2, characterized in that: During mask processing, for the infrared thermal map, the area with the highest temperature of the wire rod is masked; for the ultraviolet photon map, mask processing is performed to remove irrelevant text parts.

6. The aging degree prediction method according to claim 1, characterized in that: In step S2, the weight division process includes steps S21 to S24; wherein: S21: Calculate the status score value of each evaluation parameter; S22: the health status of each evaluation parameter is divided into four health levels, namely, serious failure, failure, sub-health and health; S23: By substituting the state score value into the membership function calculation formula, the weight of the state score value of each evaluation parameter belonging to the four health levels can be calculated, thereby obtaining the fuzzy relationship matrix of the evaluation parameter set; S24: For the classification of the aging status of the anti-corona layer, a comprehensive weighted method is adopted. First, the health status of multiple characteristic parameters of the stator wire bar anti-corona layer is evaluated, and then the variable weight coefficient matrix and single factor evaluation matrix of each factor are obtained by using the variable weight theory. Then, these two matrices are multiplied to obtain the fuzzy relationship matrix of the system health status.

7. The aging degree prediction method according to claim 1, characterized in that: In step S3, the fusion process includes steps S31 and S32; wherein: S31: perform dimensionality reduction processing on the infrared thermal map in the preprocessed data set and convert it into a one-dimensional feature vector; extract the GLCM features and HOG features of the ultraviolet photon map of the infrared thermal map in the preprocessed data set and splice them; S32: Use the convolutional neural network VGG-16 to stitch the obtained image features and temperature information.

8. The aging degree prediction method according to claim 1, characterized in that: In step S3, the preprocessed data set is divided into a training set and a test set by a random partitioning method.

9. The aging degree prediction method according to claim 1, characterized in that: In step S4, the classification includes steps S41 and S42; wherein: S41: The aging state of the stator bar anti-corona layer is used as a label to divide it into four stage categories; S42: Input the fused features into a multi-layer perceptron for training to obtain a trained anti-corona layer life prediction model.

Images