Partial discharge identification method for generators based on feature fusion bag-of-words and parallel twin networks

Abstract

The present invention discloses a method for identifying partial discharges in generators based on feature fusion word bags and parallel twin networks. The method collects partial discharge data and phase signal data of generator stator bars in the laboratory, obtains corresponding phase maps, constructs a bag-of-words dictionary for SIFT, LBP, HOG, and Haar-like multi-descriptor feature detection, and retrieves similar partial discharge maps. AlexNet, GoogLeNet, VGG-16, and ResNet neural networks are trained based on the partial discharge samples collected in the laboratory. The trained neural network twins then share weights, calculate the loss using the cross-entropy function of the fully connected layer, and select the label with the smallest loss as the recognition result for each channel. During testing, the multi-channel recognition results based on the segmentation of the spectrum in different phase quadrants are integrated using an adaptive fusion strategy to obtain the final partial discharge pattern recognition result. The present invention overcomes the shortcomings of traditional deep learning classification methods in data imbalance and the inability to train with small sample data, and solves the data imbalance and small sample recognition problems in partial discharge pattern recognition.

Description

Technical Field

[0001] The present invention belongs to the field of online monitoring of partial discharge of generator stator bars, and relates to a generator partial discharge identification method based on feature fusion word bag and parallel twin network. Background Art

[0002] One of the primary threats to the safe operation of generators comes from the insulation system. Partial discharge (PD) is both a cause and a symptom of insulation degradation, and is a key parameter for generator condition monitoring. Therefore, effectively identifying PD types is crucial for assessing and diagnosing generator insulation conditions. Generator insulation defects vary in type, with varying damage mechanisms and destructive effects. Generator PD patterns often suffer from a shortage of samples or high acquisition costs. Therefore, pattern recognition is crucial for effectively identifying and determining PDs in generator stator PDs, even with limited sample sizes. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for identifying partial discharges in generators based on a feature fusion bag-of-words and parallel twin networks. The method performs partial discharge similarity PRPD retrieval, partial discharge metric calculation based on a multi-phase parallel twin network, partial discharge PRPD spectrum similarity measurement, and multi-channel fusion partial discharge pattern recognition based on phase quadrant segmentation, and adaptively weights the multi-channel phase recognition results in different quadrants to obtain the generator partial discharge pattern recognition result. The present invention creatively applies a multi-descriptor feature fusion bag-of-words model, a multi-phase parallel twin network, and multi-channel fusion based on phase quadrant segmentation to the generator stator bar pattern recognition, achieving high-precision identification of a small number of generator stator partial discharge types. The application of this method can provide timely evaluation of the insulation status of the stator of an operating generator and maintenance advice, which has high economic value and social benefits.

[0004] The purpose of the present invention is achieved by the following technical measures:

[0005] A method for identifying partial discharge in a generator based on feature fusion bag-of-words and parallel twin networks is characterized by the following steps:

[0006] S1: Constructing a bag-of-words retrieval model based on the fusion of multiple descriptor features for typical defects in generator stator bars: Discharge data and phase signal data for typical defects in generator stator bars at different voltage levels were collected in a laboratory environment. Phase maps corresponding to known bar insulation defects were obtained as samples. Multi-descriptor features were constructed using SIFT, LBP, HOG, and Haar-like features. K-Means was used to adaptively select the number of word-book clusters to obtain a bag-of-words word book. Similar partial discharge maps were retrieved by fusing different descriptor features with the bag-of-words word book.

[0007] S2: Partial discharge metric calculation based on multi-phase parallel twin networks: AlexNet, GoogLeNet, VGG-16, and ResNet neural networks are trained using typical partial discharge samples collected in the laboratory. These neural networks are then twinned to share weights. PRPD images in different quadrants and phases are segmented and mapped into vectors using various SIFT, LBP, HOG, and Haar-like multi-descriptor features. Loss is then calculated using the cross-entropy function, and the label with the smallest distance is selected as the test sample class, completing classification.

[0008] S3: Multi-channel fusion partial discharge pattern recognition based on phase quadrant segmentation: During the test, pattern recognition is performed based on different phases and quadrants of the PRPD spectrum. The multi-phase parallel twin network of each channel outputs the pattern recognition results of the typical defect partial discharge type of the corresponding generator stator wire bar. The multi-channel results determine the weight of the judgment results of each channel through an adaptive fusion strategy, and integrate the pattern recognition results of twin networks at different levels, which is the typical defect partial discharge type of the generator stator wire bar.

[0009] The construction of the bag-of-words retrieval model based on the fusion of multiple descriptor features of typical defects of generator stator bars described in S1 is specifically carried out in the following steps:

[0010] S1.1: Collect discharge data and phase signal data at different voltage levels for typical defects in generator stator bars in a laboratory environment. Obtain phase spectra corresponding to known bar insulation defects as samples. Generator stator partial discharge types include internal discharge, slot discharge, end discharge, and vibration spark discharge.

[0011] S1.2: Based on the partial discharge sample data of four different types of generator stator defects collected in the laboratory, we extracted different bag-of-words features of the PRPD graph based on the multi-descriptor features of SIFT (Scale-invariant feature transform), LBP (Local Binary Pattern), HOG (Histogram of oriented gradient), and Haar-like features to form a PRPD graph bag-of-words sample library. We then used K-Means to adaptively select the number of word book clusters to obtain the bag-of-words word book.

[0012] S1.3: Count the number of times each word in the SIFT, LBP, HOG, and Haar-like multi-descriptor feature word lists appears in the graph, thereby representing the PRPD graph as a K-dimensional numerical vector. Then, use the different multi-descriptor features to fuse the numerical vectors to retrieve similar PD graphs, and obtain a corresponding retrieval sample to add to the PRPD graph bag-of-words sample library, forming a sample pair with the original phase graph sample.

[0013] S1.4: Repeat steps S1.2-S1.3. Each time a new generator stator partial discharge signal is collected at the industrial site, update the bag-of-words sample library of typical stator wire bar partial discharge PRPD graphs and the corresponding SIFT, LBP, HOG, and Haar-like multi-descriptor features, thereby enriching the diversity of the generator stator wire bar bag-of-words samples.

[0014] The partial discharge measurement calculation based on the multi-phase parallel twin network described in S2 is specifically carried out according to the following steps:

[0015] S2.1: The segmentation results of different phases (0°~90°, 90°~180°, 180°~270°, and 270°~360°) and the segmentation results of different quadrants (first quadrant, second quadrant, third quadrant, and fourth quadrant) in the PRPD atlas word bag sample library are represented using SIFT, LBP, HOG, and Haar-like multi-descriptor features;

[0016] S2.2: The samples in the PRPD atlas bag-of-words sample library after multi-descriptor feature representation are divided into training and test sets. 70% of the data in the bag-of-words sample library is divided into the training set for training each neural network model, and 30% of the data is divided into the test set for testing the sub-recognition model of each neural network model. The AlexNet, GoogLeNet, VGG-16, and ResNet neural networks are trained using the samples in the PRPD atlas bag-of-words sample library after multi-descriptor feature representation to obtain the optimal parameter sets for different neural networks.

[0017] S2.3: Based on the trained AlexNet, GoogLeNet, VGG-16, and ResNet neural networks, each neural network is twinned and weights are shared. The twin network uses the same embedding network to extract the graph-level features of the graph, maps the graph into a vector, and uses the absolute difference between the two vectors to represent the similarity between the two graphs. If the two input graphs are from the same class, the label is 1, and the labels of different classes are 0. The loss is then calculated using the cross entropy function of the fully connected layer. During testing, the sample pairs in the test set are sequentially input into the twin network to obtain the distance between each pair of samples. The test set label with the smallest distance is selected as the category of the test set sample.

[0018] S2.4: By parallelizing the fully connected layer parameters of the AlexNet, GoogLeNet, VGG-16, and ResNet twin networks, we share parameters, calculate the loss using the cross-entropy function, and select the test set label with the smallest distance as the category of the test set sample to complete the classification.

[0019] The multi-channel fusion partial discharge pattern recognition based on phase quadrant segmentation described in S3 is specifically performed according to the following steps:

[0020] S3.1: First, the pattern recognition results of each channel of the eight-channel parallel multi-phase twin network based on different quadrants and phases of the PRPD spectrum are assigned the same result weight, that is, W1=W2=…=W10=0.1;

[0021] S3.2: Compare the pattern discrimination results of each channel of the multi-phase parallel twin network obtained in the steps described in S2.3 with the actual partial discharge type of the generator stator bar. Use an adaptive fusion strategy to determine the weight of each channel discrimination result, and obtain the weight value of each channel result of the eight-channel multi-phase parallel twin network.

[0022] S3.3: Based on the weight values ​​of the results of each channel of the eight-channel multi-phase parallel twin network obtained in S3.2, the pattern recognition results of twin networks at different levels are integrated to calculate the pattern recognition weighted result, which is the final pattern recognition result of the partial discharge defect type of the generator stator bar.

[0023] The beneficial effects achieved by the present invention are as follows: the present invention improves the accuracy of partial discharge identification of generator stator wire rods by constructing a bag-of-words retrieval model based on the fusion of multiple descriptor features of typical defects of generator stator wire rods, performs partial discharge similarity PRPD retrieval, calculates partial discharge metrics and performs similarity measurement based on a multi-phase parallel twin network, and performs partial discharge pattern recognition through multi-channel fusion based on phase quadrant segmentation, thereby realizing pattern recognition of a small number of samples or a single sample of generator stator partial discharge, which plays an important role in the operational stability of the generator. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a structural block diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0025] The following is a detailed description of the embodiments of the present invention with reference to the accompanying drawings. The purpose of the present invention is to provide a generator partial discharge identification method based on feature fusion word bag and parallel twin network, such as Figure 1 As shown, follow these steps:

[0026] S1: Constructing a bag-of-words retrieval model based on the fusion of multiple descriptor features for typical defects in generator stator bars: Discharge data and phase signal data for typical defects in generator stator bars at different voltage levels are collected in a laboratory environment. Phase maps corresponding to known bar insulation defects are obtained as samples. SIFT, LBP, HOG, and Haar-like multi-descriptor features are constructed, and the number of word-book clusters is adaptively selected using K-Means to obtain a bag-of-words word book. The bag-of-words word book is then fused with different descriptor features to retrieve similar partial discharge maps. The specific steps include:

[0027] The construction of the bag-of-words retrieval model based on the fusion of multiple descriptor features of typical defects of generator stator bars described in S1 is specifically carried out in the following steps:

[0028] S1.1: Collect discharge data and phase signal data at different voltage levels for typical defects in generator stator bars in a laboratory environment. Obtain phase spectra corresponding to known bar insulation defects as samples. Generator stator partial discharge types include internal discharge, slot discharge, end discharge, and vibration spark discharge.

[0029] S1.2: Based on the partial discharge sample data of four different types of generator stator defects collected in the laboratory, we extracted different bag-of-words features of the PRPD graph based on the multi-descriptor features of SIFT (Scale-invariant feature transform), LBP (Local Binary Pattern), HOG (Histogram of oriented gradient), and Haar-like features to form a PRPD graph bag-of-words sample library. We then used K-Means to adaptively select the number of word book clusters to obtain the bag-of-words word book.

[0030] S1.3: Count the number of times each word in the SIFT, LBP, HOG, and Haar-like multi-descriptor feature word lists appears in the graph, thereby representing the PRPD graph as a K-dimensional numerical vector. Then, use the different multi-descriptor features to fuse the numerical vectors to retrieve similar PD graphs, and obtain a corresponding retrieval sample to add to the PRPD graph bag-of-words sample library, forming a sample pair with the original phase graph sample.

[0031] S1.4: Repeat steps S1.2-S1.3. Each time a new generator stator partial discharge signal is collected at the industrial site, update the bag-of-words sample library of typical stator wire bar partial discharge PRPD graphs and the corresponding SIFT, LBP, HOG, and Haar-like multi-descriptor features, thereby enriching the diversity of the generator stator wire bar bag-of-words samples.

[0032] S2: PD metric calculation based on multi-phase parallel twin neural networks: AlexNet, GoogLeNet, VGG-16, and ResNet neural networks are trained using typical PD samples collected in the laboratory. These neural networks are then twinned to share weights. PRPD images in different quadrants and phases are segmented and mapped into vectors using SIFT, LBP, HOG, and Haar-like multi-descriptor features. Loss is then calculated using the cross-entropy function, and the label with the smallest distance is selected as the test sample category to complete classification. The specific steps include:

[0033] S2.1: The segmentation results of different phases (0°~90°, 90°~180°, 180°~270°, and 270°~360°) and the segmentation results of different quadrants (first quadrant, second quadrant, third quadrant, and fourth quadrant) in the PRPD atlas word bag sample library are represented using SIFT, LBP, HOG, and Haar-like multi-descriptor features;

[0034] S2.2: The samples in the PRPD atlas bag-of-words sample library after multi-descriptor feature representation are divided into training and test sets. 70% of the data in the bag-of-words sample library is divided into the training set for training each neural network model, and 30% of the data is divided into the test set for testing the sub-recognition model of each neural network model. The AlexNet, GoogLeNet, VGG-16, and ResNet neural networks are trained using the samples in the PRPD atlas bag-of-words sample library after multi-descriptor feature representation to obtain the optimal parameter sets for different neural networks.

[0035] S2.3: Based on the trained AlexNet, GoogLeNet, VGG-16, and ResNet neural networks, each neural network is twinned and weights are shared. The twin network uses the same embedding network to extract the graph-level features of the graph, maps the graph into a vector, and uses the absolute difference between the two vectors to represent the similarity between the two graphs. If the two input graphs are from the same class, the label is 1, and the labels of different classes are 0. The loss is then calculated using the cross entropy function of the fully connected layer. During testing, the sample pairs in the test set are sequentially input into the twin network to obtain the distance between each pair of samples. The test set label with the smallest distance is selected as the category of the test set sample.

[0036] S2.4: By parallelizing the fully connected layer parameters of the AlexNet, GoogLeNet, VGG-16, and ResNet twin networks, we share parameters, calculate the loss using the cross-entropy function, and select the test set label with the smallest distance as the category of the test set sample to complete the classification.

[0037] S3: Multi-channel fusion partial discharge pattern recognition based on phase quadrant segmentation: During the test, pattern recognition is performed based on different phases and quadrants of the PRPD spectrum. The multi-phase parallel twin network of each channel outputs the pattern recognition results of the corresponding typical partial discharge type of the generator stator bar defect. The multi-channel results are weighted by determining the discrimination results of each channel through an adaptive fusion strategy. The pattern recognition results of the twin networks at different levels are integrated to obtain the typical partial discharge type of the generator stator bar defect. The specific steps include:

[0038] S3.1: First, the pattern recognition results of each channel of the eight-channel parallel multi-phase twin network based on different quadrants and phases of the PRPD spectrum are assigned the same result weight, that is, W1=W2=…=W10=0.1;

[0039] S3.2: Compare the pattern discrimination results of each channel of the multi-phase parallel twin network obtained in the steps described in S2.3 with the actual partial discharge type of the generator stator bar. Use an adaptive fusion strategy to determine the weight of each channel discrimination result, and obtain the weight value of each channel result of the eight-channel multi-phase parallel twin network.

[0040] S3.3: Based on the weight values ​​of the results of each channel of the eight-channel multi-phase parallel twin network obtained in S3.2, the pattern recognition results of twin networks at different levels are integrated to calculate the pattern recognition weighted result, which is the final pattern recognition result of the partial discharge defect type of the generator stator bar.

[0041] The beneficial effects achieved by the present invention are as follows: through S1, a multi-descriptor feature fusion bag-of-words retrieval model based on typical defects of generator stator wire bars is obtained, and through S2, a partial discharge metric calculation based on a multi-phase parallel twin network is obtained, and S3 is constructed for multi-channel fusion partial discharge pattern recognition based on phase quadrant segmentation, which improves the recognition accuracy of partial discharges in generator stator wire bars and realizes pattern recognition of a small number of samples or a single sample of generator stator partial discharges, which plays an important role in the operation stability of the generator.

Claims

1. A generator partial discharge identification method based on feature fusion bag-of-words and parallel twin network, characterized by Follow these steps: S1: Constructing a bag-of-words retrieval model based on the fusion of multiple descriptor features for typical defects in generator stator bars: Discharge data and phase signal data for typical defects in generator stator bars at different voltage levels were collected in a laboratory environment. Phase maps corresponding to known bar insulation defects were obtained as samples. Multi-descriptor features were constructed using SIFT, LBP, HOG, and Haar-like features. K-Means was used to adaptively select the number of word-book clusters to obtain a bag-of-words word book. Similar partial discharge maps were retrieved by fusing different descriptor features with the bag-of-words word book. S2: Partial discharge metric calculation based on multi-phase parallel twin networks: AlexNet, GoogLeNet, VGG-16, and ResNet neural networks are trained using typical partial discharge samples collected in the laboratory. These neural networks are then twinned to share weights. PRPD images in different quadrants and phases are segmented and mapped into vectors using various SIFT, LBP, HOG, and Haar-like multi-descriptor features. Loss is then calculated using the cross-entropy function, and the label with the smallest distance is selected as the test sample class, completing classification. S3: Multi-channel fusion partial discharge pattern recognition based on phase quadrant segmentation: During the test, pattern recognition is performed based on different phases and quadrants of the PRPD spectrum. The multi-phase parallel twin network of each channel outputs the pattern recognition results of the typical defect partial discharge type of the corresponding generator stator wire bar. The multi-channel results determine the weight of the judgment results of each channel through an adaptive fusion strategy, and integrate the pattern recognition results of twin networks at different levels, which is the typical defect partial discharge type of the generator stator wire bar.

2. The generator partial discharge identification method based on feature fusion bag-of-words and parallel twin network according to claim 1 is characterized by: The construction of the bag-of-words retrieval model based on the fusion of multiple descriptor features of typical defects of generator stator bars described in S1 is specifically carried out in the following steps: S1.1: Collect discharge data and phase signal data at different voltage levels for typical defects in generator stator bars in a laboratory environment. Obtain phase spectra corresponding to known bar insulation defects as samples. Generator stator partial discharge types include internal discharge, slot discharge, end discharge, and vibration spark discharge. S1.2: Based on partial discharge sample data of four different types of generator stator defects collected in the laboratory, we extracted different bag-of-words features of the PRPD graph using multiple descriptor features based on SIFT (Scale-invariant feature transform), LBP (Local Binary Pattern), HOG (Histogram of oriented gradient), and Haar-like features to form a PRPD graph bag-of-words sample library. We then used K-Means to adaptively select the number of word book clusters to obtain the bag-of-words word book. S1.3: Count the number of times each word in the SIFT, LBP, HOG, and Haar-like multi-descriptor feature word lists appears in the graph, thereby representing the PRPD graph as a K-dimensional numerical vector. Then, use the different multi-descriptor features to fuse the numerical vectors to retrieve similar PD graphs, and obtain a corresponding retrieval sample to add to the PRPD graph bag-of-words sample library, forming a sample pair with the original phase graph sample. S1.4: Repeat steps S1.2-S1.

3. Each time a new generator stator partial discharge signal is collected at the industrial site, update the bag-of-words sample library of typical stator wire bar partial discharge PRPD graphs and the corresponding SIFT, LBP, HOG, and Haar-like multi-descriptor features, thereby enriching the diversity of the generator stator wire bar bag-of-words samples.

3. The generator partial discharge identification method based on feature fusion bag-of-words and parallel twin network according to claim 2 is characterized by: The partial discharge measurement calculation based on the multi-phase parallel twin network described in S2 is specifically carried out according to the following steps: S2.1: The segmentation results of different phases (0°~90°, 90°~180°, 180°~270°, and 270°~360°) and the segmentation results of different quadrants (first quadrant, second quadrant, third quadrant, and fourth quadrant) in the PRPD atlas word bag sample library are represented using SIFT, LBP, HOG, and Haar-like multi-descriptor features; S2.2: The samples in the PRPD atlas bag-of-words sample library after multi-descriptor feature representation are divided into training and test sets. 70% of the data in the bag-of-words sample library is divided into the training set for training each neural network model, and 30% of the data is divided into the test set for testing the sub-recognition model of each neural network model. The AlexNet, GoogLeNet, VGG-16, and ResNet neural networks are trained using the samples in the PRPD atlas bag-of-words sample library after multi-descriptor feature representation to obtain the optimal parameter sets for different neural networks. S2.3: Based on the trained AlexNet, GoogLeNet, VGG-16, and ResNet neural networks, each neural network is twinned and weights are shared. The twin network uses the same embedding network to extract the graph-level features of the graph, maps the graph into a vector, and uses the absolute difference between the two vectors to represent the similarity between the two graphs. If the two input graphs are from the same class, the label is 1, and the labels of different classes are 0. The loss is then calculated using the cross entropy function of the fully connected layer. During testing, the sample pairs in the test set are sequentially input into the twin network to obtain the distance between each pair of samples. The test set label with the smallest distance is selected as the category of the test set sample. S2.4: By parallelizing the fully connected layer parameters of the AlexNet, GoogLeNet, VGG-16, and ResNet twin networks, we share parameters, calculate the loss using the cross-entropy function, and select the test set label with the smallest distance as the category of the test set sample to complete the classification.

4. The method for identifying partial discharge in a generator based on feature fusion bag-of-words and parallel twin networks according to claim 3 is characterized by: The multi-channel fusion partial discharge pattern recognition based on phase quadrant segmentation described in S3 is specifically performed according to the following steps: S3.1: First, the pattern recognition results of each channel of the eight-channel parallel multi-phase twin network based on different quadrants and phases of the PRPD spectrum are assigned the same result weight, that is, W1=W2=…=W10=0.1; S3.2: Compare the pattern discrimination results of each channel of the multi-phase parallel twin network obtained in the steps described in S2.3 with the actual partial discharge type of the generator stator bar. Use an adaptive fusion strategy to determine the weight of each channel discrimination result, and obtain the weight value of each channel result of the eight-channel multi-phase parallel twin network. S3.3: Based on the weight values ​​of the results of each channel of the eight-channel multi-phase parallel twin network obtained in S3.2, the pattern recognition results of twin networks at different levels are integrated to calculate the pattern recognition weighted result, which is the final pattern recognition result of the partial discharge defect type of the generator stator bar.

Images