A generator partial discharge pattern recognition method based on parallel CNN-BiLSTM
By combining parallel CNN-BiLSTM neural networks with laboratory and industrial field data, partial discharge modes of generator stator bars are identified, solving the problem of untimely identification in existing technologies and achieving efficient and accurate partial discharge type discrimination.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-14
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies are insufficient to effectively identify the type of partial discharge in generator stator bars in real time, leading to delayed fault detection and potentially serious consequences such as fires.
By employing a parallel CNN-BiLSTM neural network and combining laboratory and industrial field data, multi-channel feature extraction and weighting are used to identify the partial discharge mode of generator stator bars, thereby reducing identification bias and improving accuracy.
It achieves efficient identification of partial discharge types in generator stator bars, improves identification accuracy, shortens identification time, and has high economic value and social benefits.
Smart Images
Figure CN115586407B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a generator partial discharge pattern recognition method based on parallel CNN-BiLSTM, belonging to the field of generator partial discharge online monitoring. Background Technology
[0002] As a crucial piece of equipment in the power system, the safe and stable operation of generators directly ensures the reliability of the power grid. When a generator malfunctions, it not only affects the generator itself but also causes losses to industrial production and people's lives. Partial discharge (PD), a type of generator fault, can even cause fires and result in significant economic losses in severe cases. Therefore, online monitoring of the generator stator bars and real-time identification of PD types can promptly detect problems and propose corresponding modification solutions, thus ensuring the stable operation of the generator to a certain extent. Summary of the Invention
[0003] The purpose of this invention is to provide a generator partial discharge pattern recognition method based on parallel CNN-BiLSTM to solve the problems existing in the background art.
[0004] The objective of this invention is achieved by the following technical measures:
[0005] A generator partial discharge pattern recognition method based on parallel CNN-BiLSTM includes the following steps:
[0006] S1: Build a partial discharge test platform for typical defects of generator stator bars, conduct partial discharge tests under different generator stator bar defect types and different pressure levels, collect various signals during the partial discharge process of generator stator bars, and construct ten different types of input parameters for training a ten-channel parallel CNN neural network;
[0007] S2: Ten input parameters are fed into the input layer of a ten-channel parallel CNN neural network. First, the input parameters are extracted by the CNN neural network of each parallel channel. Then, the features are fed into the BiLSTM neural network through the output layer. The BiLSTM neural network is trained and predicted to identify the typical partial discharge type of generator stator bars.
[0008] S3: Each channel of the BiLSTM neural network outputs the pattern recognition results of the typical partial discharge type of the generator stator bar. Combining the partial discharge signals of different defect types collected on site and the accuracy of the discrimination results of each channel, the weight of the discrimination results of each channel is determined, and the pattern recognition result after weighting of ten channels is calculated, which is the typical partial discharge type of the generator stator bar.
[0009] Furthermore, the specific steps of S1 include:
[0010] S1.1: Construct a partial discharge test platform for typical defects in generator stator bars under laboratory conditions, and conduct partial discharge tests on generator stator bars under different defect types and different voltage levels.
[0011] S1.2: During the process of inducing partial discharge under pressure, partial discharge signals are collected by high voltage coupling capacitor, A / B phase voltage signals are collected by two PT sensors respectively, and ground current signals are collected by CT sensor. The partial discharge signals, voltage and current signal samples of generator stator bar power supply and spark discharge defects, internal discharge defects, end discharge curves and slot discharge defects are obtained respectively.
[0012] S1.3: Based on the collected partial discharge signal samples, in order to retain different proportions of partial discharge signal samples, normalized phase maps of partial discharge with denoising coefficients n1=0, n2=0.25, n3=0.5, and n4=0.75 are plotted respectively.
[0013] S1.4: Based on the partial discharge phase maps plotted with denoising coefficients n1=0, n2=0.25, n3=0.5, and n4=0.75, the discharge density distribution and maximum discharge quantity features of the corresponding phase maps are extracted respectively; based on the collected partial discharge signal samples, the skewness, kurtosis, skewness, peak value, mean value, and variance parameters of the partial discharge waveform are extracted respectively.
[0014] S1.5: The normalized phase spectrum of partial discharge with denoising coefficients n1=0, n2=0.25, n3=0.5, and n4=0.75 is used as four input parameters of the CNN neural network; the A / B two-phase voltage signal acquired by the PT sensor is used as two input parameters of the CNN neural network; the partial discharge signal acquired by the high-voltage coupling capacitor is used as one input parameter of the CNN neural network; the grounding current signal acquired by the CT sensor is used as one input parameter of the CNN neural network; the extracted phase spectrum feature parameter is used as one input parameter of the CNN neural network; and the extracted partial discharge waveform feature parameter is used as one input parameter of the CNN neural network, for a total of ten feature parameters, which are used as input parameters of the ten-channel parallel CNN neural network.
[0015] Furthermore, the specific steps of S2 include:
[0016] S2.1: The ten different feature parameters extracted in S1.5 are fed into the input layer of a ten-channel parallel CNN neural network. The feature parameters are then subjected to convolution and pooling operations through each channel of the CNN neural network, and the feature parameters are then extracted and reconstructed.
[0017] S2.2: The reconstructed feature parameters are fed into the input gates of the forward LSTM and the backward LSTM of the BiLSTM, respectively. The reconstructed feature parameters are then processed by the forget gate to determine the information retention state. The input retained by the forget gate is combined with the input of the input gate to determine the state of the current output information, namely the generator stator bar partial discharge pattern recognition result. The pattern recognition result is output through the output gate. The BiLSTM output gate outputs the generator stator bar partial discharge pattern recognition result of each channel of the ten-channel parallel CNN-BiLSTM neural network.
[0018] Furthermore, the specific steps of S3 include:
[0019] S3.1: First, assign the same result weight to the pattern recognition result of each channel of the ten-channel parallel CNN-BiLSTM neural network output by the BiLSTM output gate, i.e., W1=W2=…=W10=0.1;
[0020] S3.2: Based on the steps described in S2.2, the mode discrimination results of each channel of the ten-channel parallel CNN neural network are compared with the actual partial discharge type of the generator stator bar to obtain the accuracy of the judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network.
[0021] S3.3: Collect different types of partial discharge signals of generator stator bars in industrial sites, perform pattern recognition of discharge types based on the steps described in S1 and S2, compare with the actual situation, and obtain the accuracy of the judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network.
[0022] S3.4: Based on the accuracy of the discharge type judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network under the laboratory conditions of S2.2 and the accuracy of the discharge type judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network under the industrial field conditions of S3.3, the weight of each channel output result of the ten-channel parallel CNN-BiLSTM neural network is determined by combining the two. This weight is the final weight value of each channel result of the ten-channel parallel CNN-BiLSTM neural network.
[0023] S3.5: Based on the weight values of each channel of the final ten-channel parallel CNN-BiLSTM neural network obtained from S3.4, calculate the weighted result of pattern recognition, which is the final pattern recognition result of the partial discharge defect type of the generator stator bar.
[0024] This invention develops a generator partial discharge pattern recognition method based on parallel CNN. It creatively combines typical defect partial discharge samples of laboratory generator stator bars to construct multi-type, multi-dimensional partial discharge feature parameters. These parameters are then processed through a ten-channel parallel CNN-BiLSTM neural network to obtain pattern recognition results. This reduces the bias of single neural network pattern recognition, improves the accuracy of pattern recognition, and shortens the pattern recognition time. Furthermore, by incorporating industrial field data for weight adjustment, this method can be effectively applied to the identification of generator partial discharge types in actual industrial settings, demonstrating high economic value and social benefits. Attached Figure Description
[0025] Figure 1 This is a flowchart illustrating the overall process structure of the present invention. Detailed Implementation
[0026] 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 pattern recognition method based on parallel CNN-BiLSTM.
[0027] like Figure 1 As shown, a generator partial discharge pattern recognition method based on parallel CNN-BiLSTM includes the following steps:
[0028] S1: Construct a partial discharge test platform for typical defects in generator stator bars. Conduct partial discharge tests under different generator stator bar defect types and different pressure levels. Collect various signals during the partial discharge process of the generator stator bars to construct ten different types of input parameters for training a ten-channel parallel CNN neural network. The specific steps include:
[0029] S1.1: Construct a partial discharge test platform for typical defects in generator stator bars under laboratory conditions, and conduct partial discharge tests on generator stator bars under different defect types and different voltage levels.
[0030] S1.2: During the process of inducing partial discharge under pressure, partial discharge signals are collected by high voltage coupling capacitor, A / B phase voltage signals are collected by two PT sensors respectively, and ground current signals are collected by CT sensor. The partial discharge signals, voltage and current signal samples of generator stator bar power supply and spark discharge defects, internal discharge defects, end discharge curves and slot discharge defects are obtained respectively.
[0031] S1.3: Based on the collected partial discharge signal samples, in order to retain different proportions of partial discharge signal samples, normalized phase maps of partial discharge with denoising coefficients n1=0, n2=0.25, n3=0.5, and n4=0.75 are plotted respectively.
[0032] S1.4: Based on the partial discharge phase maps plotted with denoising coefficients n1=0, n2=0.25, n3=0.5, and n4=0.75, the discharge density distribution and maximum discharge quantity features of the corresponding phase maps are extracted respectively; based on the collected partial discharge signal samples, the skewness, kurtosis, skewness, peak value, mean value, and variance parameters of the partial discharge waveform are extracted respectively.
[0033] S1.5: The normalized phase spectrum of partial discharge with denoising coefficients n1=0, n2=0.25, n3=0.5, and n4=0.75 is used as four input parameters of the CNN neural network; the A / B two-phase voltage signal acquired by the PT sensor is used as two input parameters of the CNN neural network; the partial discharge signal acquired by the high-voltage coupling capacitor is used as one input parameter of the CNN neural network; the grounding current signal acquired by the CT sensor is used as one input parameter of the CNN neural network; the extracted phase spectrum feature parameter is used as one input parameter of the CNN neural network; and the extracted partial discharge waveform feature parameter is used as one input parameter of the CNN neural network, for a total of ten feature parameters, which are used as input parameters of the ten-channel parallel CNN neural network.
[0034] S2: Ten input parameters are fed into the input layer of a ten-channel parallel CNN neural network. First, the input parameters are used to extract features through the CNN neural network of each parallel channel. Then, the features are fed into a BiLSTM neural network through the output layer. The BiLSTM neural network is trained and used to predict and identify the typical partial discharge type of generator stator bars. The specific steps include:
[0035] S2.1: The ten different feature parameters extracted in S1.5 are fed into the input layer of a ten-channel parallel CNN neural network. The feature parameters are then subjected to convolution and pooling operations through each channel of the CNN neural network, and the feature parameters are then extracted and reconstructed.
[0036] S2.2: The reconstructed feature parameters are fed into the input gates of the forward LSTM and the backward LSTM of the BiLSTM, respectively. The reconstructed feature parameters are then processed by the forget gate to determine the information retention state. The input retained by the forget gate is combined with the input of the input gate to determine the state of the current output information, namely the generator stator bar partial discharge pattern recognition result. The pattern recognition result is output through the output gate. The BiLSTM output gate outputs the generator stator bar partial discharge pattern recognition result of each channel of the ten-channel parallel CNN-BiLSTM neural network.
[0037] S3: Each channel of the BiLSTM neural network outputs the pattern recognition result of the typical partial discharge type of the generator stator bar. Combining the partial discharge signals of different defect types collected on site and the accuracy of the discrimination results of each channel, the weight of the discrimination results of each channel is determined, and the ten-channel weighted pattern recognition result is calculated, which is the typical partial discharge type of the generator stator bar. The specific steps include:
[0038] S3.1: First, assign the same result weight to the pattern recognition result of each channel of the ten-channel parallel CNN-BiLSTM neural network output by the BiLSTM output gate, i.e., W1=W2=…=W10=0.1;
[0039] S3.2: Based on the steps described in S2.2, the mode discrimination results of each channel of the ten-channel parallel CNN neural network are compared with the actual partial discharge type of the generator stator bar to obtain the accuracy of the judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network.
[0040] S3.3: Collect different types of partial discharge signals of generator stator bars in industrial sites, perform pattern recognition of discharge types based on the steps described in S1 and S2, compare with the actual situation, and obtain the accuracy of the judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network.
[0041] S3.4: Based on the accuracy of the discharge type judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network under the laboratory conditions of S2.2 and the accuracy of the discharge type judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network under the industrial field conditions of S3.3, the weight of each channel output result of the ten-channel parallel CNN-BiLSTM neural network is determined by combining the two. This weight is the final weight value of each channel result of the ten-channel parallel CNN-BiLSTM neural network.
[0042] S3.5: Based on the weight values of each channel of the final ten-channel parallel CNN-BiLSTM neural network obtained from S3.4, calculate the weighted result of pattern recognition, which is the final pattern recognition result of the partial discharge defect type of the generator stator bar.
Claims
1. A generator partial discharge pattern recognition method based on parallel CNN-BiLSTM, characterized in that... Follow these steps: S1: Build a partial discharge test platform for typical defects of generator stator bars, conduct partial discharge tests under different generator stator bar defect types and different pressure levels, collect various signals during the partial discharge process of generator stator bars, and construct ten different types of input parameters for training a ten-channel parallel CNN neural network; Specifically, the following steps are performed: S1.1: Under laboratory conditions, a partial discharge test platform for typical defects of generator stator bars is built, and partial discharge tests of generator stator bars are carried out under different defect types and different voltage levels. S1.2: During the process of inducing partial discharge under pressure, partial discharge signals are collected by high voltage coupling capacitor, A / B phase voltage signals are collected by two PT sensors respectively, and ground current signals are collected by CT sensor. The partial discharge signals, voltage and current signal samples of generator stator bar power supply and spark discharge defects, internal discharge defects, end discharge curves and slot discharge defects are obtained respectively. S1.3: Based on the collected partial discharge signal samples, in order to retain different proportions of partial discharge signal samples, normalized phase maps of partial discharge with denoising coefficients n1=0, n2=0.25, n3=0.5, and n4=0.75 are plotted respectively. S1.4: Based on the partial discharge phase maps plotted with denoising coefficients n1=0, n2=0.25, n3=0.5, and n4=0.75, the characteristics of the discharge density distribution and maximum discharge amount of the corresponding phase maps are extracted respectively. Based on the collected partial discharge signal samples, the skewness, kurtosis, skewness, peak value, mean value and variance parameters of the partial discharge waveform are extracted respectively. S1.5: The normalized phase spectrum of partial discharge with denoising coefficients n1=0, n2=0.25, n3=0.5, and n4=0.75 is used as four input parameters of the CNN neural network; the A / B two-phase voltage signal acquired by the PT sensor is used as two input parameters of the CNN neural network; the partial discharge signal acquired by the high-voltage coupling capacitor is used as one input parameter of the CNN neural network; the grounding current signal acquired by the CT sensor is used as one input parameter of the CNN neural network; the extracted phase spectrum feature parameter is used as one input parameter of the CNN neural network; and the extracted partial discharge waveform feature parameter is used as one input parameter of the CNN neural network, for a total of ten feature parameters, which are used as input parameters of the ten-channel parallel CNN neural network. S2: Ten input parameters are fed into the input layer of a ten-channel parallel CNN neural network. First, the input parameters are extracted by the CNN neural network of each parallel channel. Then, the features are fed into the BiLSTM neural network through the output layer. The BiLSTM neural network is trained and predicted to identify the typical partial discharge type of generator stator bars. S3: Each channel of the BiLSTM neural network outputs the pattern recognition results of the typical partial discharge type of the generator stator bar. Combining the partial discharge signals of different defect types collected on site and the accuracy of the discrimination results of each channel, the weight of the discrimination results of each channel is determined, and the pattern recognition result after weighting of ten channels is calculated, which is the typical partial discharge type of the generator stator bar.
2. The generator partial discharge pattern recognition method based on parallel CNN-BiLSTM according to claim 1, characterized in that: S2 is performed according to the following steps: S2.1: The ten different feature parameters extracted in S1.5 are fed into the input layer of a ten-channel parallel CNN neural network. The feature parameters are then subjected to convolution and pooling operations through each channel of the CNN neural network, and the feature parameters are then extracted and reconstructed. S2.2: The reconstructed feature parameters are fed into the input gates of the forward LSTM and the backward LSTM of the BiLSTM, respectively. The reconstructed feature parameters are then processed by the forget gate to determine the information retention state. The input retained by the forget gate is combined with the input of the input gate to determine the state of the current output information, namely the generator stator bar partial discharge pattern recognition result. The pattern recognition result is output through the output gate. The BiLSTM output gate outputs the generator stator bar partial discharge pattern recognition result of each channel of the ten-channel parallel CNN-BiLSTM neural network.
3. The generator partial discharge pattern recognition method based on parallel CNN-BiLSTM according to claim 2, characterized in that: The specific steps of S3 include: S3.1: First, assign the same result weight to the pattern recognition result of each channel of the ten-channel parallel CNN-BiLSTM neural network output by the BiLSTM output gate, i.e., W1=W2=…=W10=0.1; S3.2: Based on the steps described in S2.2, the mode discrimination results of each channel of the ten-channel parallel CNN neural network are compared with the actual partial discharge type of the generator stator bar to obtain the accuracy of the judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network. S3.3: Collect different types of partial discharge signals of generator stator bars in industrial sites, perform pattern recognition of discharge types based on the steps described in S1 and S2, compare with the actual situation, and obtain the accuracy of the judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network. S3.4: Based on the accuracy of the discharge type judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network under the laboratory conditions of S2.2 and the accuracy of the discharge type judgment results of each channel of the ten-channel parallel CNN-BiLSTM neural network under the industrial field conditions of S3.3, the weight of each channel output result of the ten-channel parallel CNN-BiLSTM neural network is determined by combining the two. This weight is the final weight value of each channel result of the ten-channel parallel CNN-BiLSTM neural network. S3.5: Based on the weight values of each channel of the final ten-channel parallel CNN-BiLSTM neural network obtained from S3.4, calculate the weighted result of pattern recognition, which is the final pattern recognition result of the partial discharge defect type of the generator stator bar.