A generator partial discharge pattern recognition method based on threshold spectrogram and light neural network

Abstract

The application discloses a generator partial discharge pattern recognition method based on a threshold spectrum and a light neural network, collects generator stator bar discharge data and phase signal data of different defect types in a laboratory environment, obtains a known phase spectrum corresponding to bar insulation defects, selects optimal threshold parameters of different types of typical defects and optimizes threshold spectrum features based on an adaptive threshold binarization method, constructs HOG, SHIF, SURF and MSER different types of image features, and carries out multi-dimensional feature fusion, selects four light neural networks of SqueezeNet, MobileNet, ShuffleNet and Xception based on different industrial application scenes for integrated learning, trains the extracted multi-dimensional fusion features on different light neural network models, judges the pattern recognition result through weight adjustment, and outputs a judgment result of the generator stator bar defect type. The method is based on the threshold spectrum and the light neural network, and improves the accuracy and real-time performance of the pattern recognition.

Description

TECHNICAL FIELD

[0001] The application relates to a generator partial discharge pattern recognition method based on a threshold spectrum and a light neural network and belongs to the field of generator stator bar partial discharge online monitoring. BACKGROUND

[0002] One of the main threats to the safe operation of a generator comes from the insulation system. According to the statistics of CIGRE Working Group WG A3.06, insulation faults account for about 56% of the faults of a hydro-generator, and stator insulation faults account for nearly two-thirds of the insulation faults. Therefore, the stator insulation performance is directly related to the operation reliability and service life of the unit. Partial discharge (PRPD, referred to as partial discharge) is both a cause and a manifestation of insulation degradation and is one of the important parameters of generator state monitoring. Different types of generator partial discharge have different insulation damage mechanisms and destructive effects. Therefore, it is of great significance to effectively identify and determine the partial discharge. With the increasing demand for urban electricity, the proportion of high-voltage cables used in urban power transmission and distribution is increasing year by year, and the operation state of the high-voltage cable directly affects the stability of the entire power transmission and distribution network. SUMMARY

[0003] The application aims to provide a generator partial discharge pattern recognition method based on a threshold spectrum and a light neural network. The application creatively applies adaptive threshold interference suppression, multi-dimensional fusion image features and integrated light neural network to generator stator bar pattern recognition, realizes high-precision and low-power consumption recognition of the type of generator stator partial discharge, and can give timely state evaluation and maintenance opinions for the running cable, thereby having high economic value and social benefits.

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

[0005] A generator partial discharge pattern recognition method based on a threshold spectrum and a light neural network is characterized by the following steps:

[0006] S1: Feature optimization method based on adaptive threshold binarization method: collect typical defect discharge data and phase signal data of generator stator bars of different defect types in a laboratory environment, obtain phase spectrum data sets corresponding to known bar insulation defects, select optimal threshold parameters of different types of typical defects, and select optimal threshold interference suppression parameters based on an adaptive threshold binarization method, and obtain feature spectrum after adaptive threshold binarization interference suppression processing;

[0007] S2: based on HOG, SHIF, SURF, MSER different types of image features, extract the fusion feature parameters of different dimensions of atlas and carry out multi-dimensional feature fusion: based on the adaptive threshold binarization, obtain the picture sample of the generator stator bar partial discharge PRPD feature atlas after interference suppression, extract the HOG, SHIF, SURF, MSER image features of different types of picture samples, and obtain the multi-dimensional fusion feature atlas sample by fusing different dimensions of image features;

[0008] S3: based on different industrial application scenarios, select SqueezeNet, MobileNet, ShuffleNet and Xception four kinds of lightweight neural network models for integrated learning: the extracted multi-dimensional fusion feature atlas sample is trained on the above four different lightweight neural network models, the neural network integrated weight is adjusted by selecting different industrial application scenarios, and the pattern recognition result is judged, and the discrimination result of the generator stator bar defect type is output.

[0009] After S1, the feature atlas based on adaptive threshold binarization and interference suppression is obtained, and after S2, the multi-dimensional fusion feature atlas sample of the generator stator bar typical defect partial discharge based on multi-dimensional fusion image features is obtained, and the local discharge pattern recognition method based on integrated lightweight neural network of S3 is constructed.

[0010] Further, the feature optimization method based on adaptive threshold binarization method in S1 is carried out according to the following steps:

[0011] S1.1: based on laboratory cable partial discharge experiment, collecting partial discharge data of different types of defects, all the partial discharge data collected based on different defect types of cables are classified into four types of internal discharge, slot discharge, end discharge and vibration spark discharge;

[0012] S1.2: based on the partial discharge sample data of four different types of cable defect types collected in the laboratory, by adjusting the threshold parameters to 0, 0.25, 0.5 and 0.75, the PRPD atlas of the generator stator bar typical defect under different noise environment is obtained under different threshold values;

[0013] S1.3: based on the adaptive threshold binarization method, the optimal threshold interference suppression parameter is selected, the PRPD feature atlas after adaptive threshold segmentation is obtained, and the PRPD feature atlas obtained by adaptive thresholding is compared with the PRPD atlas with adjusted threshold parameters, by comparing the suppression degree of interference signals under different industrial application scenarios of the two kinds of atlas, the appropriate threshold range of different types of defects is selected, the reliability of threshold selection and the accuracy of interference suppression are improved;

[0014] S1.4: Repeat the steps of S1.2-S1.3, and update the stator bar different type of typical defect partial discharge PRPD pattern sample library and the corresponding threshold selection pattern sample library after each new high-voltage cable partial discharge signal is collected in the industrial field, and then adjust and select the interference suppression threshold based on different industrial application scenarios, and continuously improve the reliability of threshold selection and the accuracy of interference suppression.

[0015] Further, the method for constructing multi-dimensional fusion features based on different types of image features of HOG, SHIF, SURF, and MSER described in S2 is specifically performed according to the following steps:

[0016] S2.1: Extract different types of image features based on PRPD feature pattern extraction of adaptive thresholding processing, including HOG (histogram of oriented gradient), SHIF (Scale-invariant feature transform), SURF (Speeded Up Robust Features), and MSER (Maximally Stable Extremal Regions);

[0017] S2.2: Compare the four types of feature patterns, HOG feature obtains edge features, SHIF feature obtains point distribution features, SURF feature obtains phase features, and MSER feature obtains region distribution features; superimpose and fuse the above four different dimensional image features to obtain a multi-dimensional fusion feature pattern sample;

[0018] S2.3: Construct an internal discharge, slot discharge, end discharge, and vibration spark different type of fingerprint feature library based on the multi-dimensional fusion feature pattern sample, divide the picture samples of the fingerprint feature library into a training set and a test set and input them into an integrated lightweight neural network; divide 70% of the data in the picture samples of the fingerprint feature library into a training set, which is used for training the model of each lightweight neural network, and divide 30% of the data into a test set, which is used for testing the respective sub-recognition models of the four integrated lightweight neural network models.

[0019] Further, the method for constructing multi-dimensional fusion features based on different types of image features of HOG, SHIF, SURF, and MSER described in S2 is specifically performed according to the following steps:

[0020] S3.1: Construct four lightweight neural network models, namely SqueezeNet, MobileNet, ShuffleNet, and Xception, to build their respective sub-recognition models, and pre-train them using image samples from the training set in S2.3; select appropriate optimizers, learning rates, and gradient decay threshold parameters for each sub-recognition model.

[0021] S3.2: Based on ensemble learning, the sub-recognition models of the four pre-trained lightweight neural network models are integrated. According to the different accuracy levels of the sub-recognition models of the four lightweight neural network models and the industrial application scenarios, weights are assigned to each sub-recognition result of the sub-recognition models of the four lightweight neural network models pre-trained in S3.1.

[0022] S3.3: Using recognition accuracy and time as the performance evaluation criteria for each sub-recognition model of the four integrated lightweight neural network models, the most suitable sub-recognition model of the four integrated lightweight neural network models is selected. The image sample test data of the S2.3 test set is input into the selected sub-recognition model of the four integrated lightweight neural network models and then enters the recognition test, and the results are output.

[0023] The beneficial effects achieved by this invention are as follows: This invention optimizes the features of the acquired phase spectrum through an adaptive thresholding method, constructs features through multi-dimensional fusion based on four image features (HOG, SHIF, SURF, and MSER), and performs partial discharge pattern recognition of generator stator bars by integrating lightweight neural networks based on SqueezeNet, MobileNet, ShuffleNet, and Xception. This improves the recognition accuracy of generator stator bar partial discharge under noisy conditions, enabling the pattern recognition method to better balance recognition accuracy and real-time performance, which plays an important role in the stability of generator operation. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall process structure of the present invention. Detailed Implementation

[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 design method for a high-voltage cable intelligent monitoring central system.

[0026] like Figure 1 As shown, a generator partial discharge pattern recognition method based on threshold spectrum and lightweight neural network includes the following steps:

[0027] S1: Feature optimization method based on adaptive threshold binarization method: Collect typical defect discharge data and phase signal data of generator stator bars of different defect types in a laboratory environment, obtain the phase spectrum data set corresponding to the known bar insulation defect, select the optimal threshold parameter of different types of typical defects, and select the optimal thresholding interference suppression parameter based on the adaptive threshold binarization method, and obtain the feature spectrum after adaptive threshold binarization interference suppression processing. The specific steps include:

[0028] S1.1: Based on the laboratory cable partial discharge experiment, collect partial discharge data of different types of defects, and classify all the partial discharge data collected based on different types of defects into four types: 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 cable defects collected in the laboratory, by adjusting the threshold parameters to 0, 0.25, 0.5 and 0.75, obtain the PRPD spectrum of the generator stator bar under different noise environments of different threshold values;

[0030] S1.3: Based on the adaptive threshold binarization method, select the optimal thresholding interference suppression parameter, obtain the PRPD feature spectrum after adaptive threshold segmentation, and compare the PRPD feature spectrum obtained by adaptive thresholding processing with the PRPD spectrum with adjusted threshold parameters. By comparing the suppression degree of interference signals in different industrial application scenarios, select the appropriate threshold range for different types of defects, and improve the reliability of threshold selection and the accuracy of interference suppression;

[0031] S1.4: Repeat the steps of S1.2-S1.3. After collecting new high-voltage cable partial discharge signals in the industrial field each time, update the stator bar partial discharge PRPD spectrum sample library of different types of typical defects and the corresponding threshold selection spectrum sample library, and then adjust and select the interference suppression threshold based on different industrial application scenarios, and continuously improve the reliability of threshold selection and the accuracy of interference suppression.

[0032] S2: Based on HOG, SHIF, SURF, MSER different types of image features, extract different dimensional fusion feature parameters of the spectrum and perform multi-dimensional feature fusion: Based on the adaptive threshold binarization, obtain the picture sample of the generator stator bar partial discharge PRPD feature spectrum after interference suppression, extract the HOG, SHIF, SURF, MSER image features of different types of picture samples, and obtain the multi-dimensional fusion feature spectrum sample by fusing different dimensional image features. The specific steps include:

[0033] S2.1: Different types of image features are extracted based on the PRPD feature map extraction of adaptive thresholding processing, including HOG (histogram of oriented gradient), SHIF (Scale-invariant feature transform), SURF (Speeded Up Robust Features), and MSER (Maximally Stable Extremal Regions);

[0034] S2.2: The four types of feature maps are compared, and the HOG feature obtains edge features, the SHIF feature obtains point distribution features, the SURF feature obtains phase features, and the MSER feature obtains region distribution features. The multi-dimensional fusion feature map samples are obtained by superimposing and fusing the above four types of image features with different dimensions;

[0035] S2.3: Based on the multi-dimensional fusion feature map samples, internal discharge, slot discharge, end discharge, and vibration spark fingerprint feature libraries of different types are constructed. The picture samples of the fingerprint feature library are divided into a training set and a test set and input into an integrated lightweight neural network. 70% of the data in the picture samples of the fingerprint feature library are divided into a training set, which is used for training the models of each lightweight neural network. 30% of the data are divided into a test set, which is used for testing the respective sub-recognition models of the four integrated lightweight neural network models.

[0036] S3: Based on different industrial application scenarios, SqueezeNet, MobileNet, ShuffleNet, and Xception four types of lightweight neural network models are selected for integrated learning. The extracted multi-dimensional fusion feature map samples are trained on the above four different lightweight neural network models. By selecting different industrial application scenarios, the neural network integrated weights are adjusted and the pattern recognition results are judged, and the discrimination results of the generator stator bar defect type are output. The specific steps include:

[0037] S3.1: SqueezeNet, MobileNet, ShuffleNet, and Xception four types of lightweight neural network models are constructed to construct their respective sub-recognition models, and the picture samples of the S2.3 training set are used to pre-train the respective sub-recognition models. The appropriate optimizer, learning rate, and gradient decay threshold parameters of the respective sub-recognition models are selected;

[0038] S3.2: Based on ensemble learning, the pre-trained sub-identification models of the four lightweight neural network models are integrated. According to the accuracy of the sub-identification models of the four lightweight neural network models and the industrial application scenarios, the weight of each sub-identification result of the pre-trained sub-identification models of the four lightweight neural network models in S3.1 is allocated;

[0039] S3.3: The recognition accuracy and time are taken as the performance evaluation criteria for integrating the sub-identification models of the four lightweight neural network models, and the most suitable integrated sub-identification models of the four lightweight neural network models are selected. The image sample test data of the test set in S2.3 are input into the selected integrated sub-identification models of the four lightweight neural network models, and then the recognition test is performed to output the results.

[0040] After S1, the interference suppression map based on the adaptive thresholding method is obtained, and after S2, the local discharge characteristic map of the typical defect of the generator stator bar is obtained based on the multi-dimensional fusion image feature. The local discharge pattern recognition method based on the integrated lightweight neural network established in S3 is constructed.

Claims

1. A generator partial discharge pattern recognition method based on threshold spectrum and lightweight neural network, characterized in that... Follow these steps: S1: Feature optimization method based on adaptive threshold binarization: Discharge data and phase signal data of typical defects in generator stator bars of different defect types are collected in a laboratory environment. A phase spectrum dataset corresponding to known bar insulation defects is obtained. Optimal threshold parameters for different types of typical defects are selected, and optimal thresholding interference suppression parameters are selected based on the adaptive threshold binarization method. Feature spectra after interference suppression processing based on adaptive threshold binarization are obtained. Specifically, the following steps are performed: S1.1: Based on laboratory cable partial discharge experiments, partial discharge data of cables with different types of defects were collected, and all partial discharge data collected based on cables with different defect types were classified into four types: internal discharge, slot discharge, end discharge and vibration spark discharge. S1.2: Based on partial discharge sample data of four different types of cable defects collected in the laboratory, by adjusting the threshold parameters to 0, 0.25, 0.5 and 0.75, PRPD maps of typical defects of generator stator bars under different noise environments were obtained at different thresholds. S1.3: Based on the adaptive threshold binarization method, the optimal thresholding interference suppression parameter is selected, the PRPD feature map after adaptive thresholding is obtained, and the obtained adaptive thresholding PRPD feature map is compared with the PRPD map with the adjusted threshold parameter. By comparing the two maps for the degree of suppression of interference signals under different industrial application scenarios, the appropriate threshold range for different types of defects is selected, thereby improving the reliability of threshold selection and the accuracy of interference suppression. S1.4: Repeat steps S1.2-S1.

3. After each new high-voltage cable partial discharge signal is collected from the industrial site, update the stator bar partial discharge PRPD spectrum sample library for different types of typical defects and the corresponding threshold selection spectrum sample library. Then, adjust and select the interference suppression threshold based on different industrial application scenarios to continuously improve the reliability of threshold selection and the accuracy of interference suppression. S2: Based on different types of image features such as HOG, SHIF, SURF, and MSER, extract fusion feature parameters of different dimensions of the map and perform multi-dimensional feature fusion: Based on the image samples of the generator stator bar partial discharge PRPD feature map obtained by adaptive threshold binarization after interference suppression, extract HOG, SHIF, SURF, and MSER image features of different types of image samples, and obtain multi-dimensional fusion feature map samples by fusing image features of different dimensions; S3: Based on different industrial application scenarios, four lightweight neural network models, namely SqueezeNet, MobileNet, ShuffleNet, and Xception, are selected for integrated learning: Based on the multi-dimensional fusion feature map samples, fingerprint feature libraries of different types, such as internal discharge, slot discharge, end discharge, and vibration spark, are constructed. The image samples of the fingerprint feature libraries are used to train the above four different lightweight neural network models respectively. By selecting different industrial application scenarios, the integrated weights of the neural networks are adjusted and the pattern recognition results are judged to output the discrimination results of generator stator bar defect types.

2. The generator partial discharge pattern recognition method based on threshold spectrum and lightweight neural network according to claim 1, characterized in that: S2 describes the extraction of fusion feature parameters from different dimensions of the image atlas based on different types of image features such as HOG, SHIF, SURF, and MSER, and the subsequent multi-dimensional feature fusion, which is carried out according to the following steps: S2.1: Based on adaptive thresholding, PRPD feature maps extract different types of image features, including four types: HOG histogram of oriented gradients, SHIF scale-invariant feature transform, SURF accelerated robust features, and MSER maximum stable external region features. S2.2: Compare the four types of feature maps: HOG feature to obtain edge features, SHIF feature to obtain point distribution features, SURF feature to obtain phase features, and MSER feature to obtain region distribution features; superimpose and fuse the above four different dimensions of image features to obtain multidimensional fused feature map samples; S2.3: Construct fingerprint feature libraries of different types, including internal discharge, slot discharge, end discharge, and vibration spark, based on multi-dimensional fusion feature map samples. Divide the image samples of the fingerprint feature library into training set and test set and input them into the integrated lightweight neural network respectively. Divide 70% of the data in the image samples of the fingerprint feature library into training set for training the models of each lightweight neural network, and divide 30% of the data into test set for testing the sub-recognition models of the four integrated lightweight neural network models.

3. The generator partial discharge pattern recognition method based on threshold spectrum and lightweight neural network according to claim 2, characterized in that: S3 describes the selection of four lightweight neural network models—SqueezeNet, MobileNet, ShuffleNet, and Xception—for ensemble learning based on different industrial application scenarios. The specific steps are as follows: S3.1: Construct four lightweight neural network models, namely SqueezeNet, MobileNet, ShuffleNet, and Xception, to build their respective sub-recognition models, and pre-train them using image samples from the training set in S2.3; select appropriate optimizers, learning rates, and gradient decay threshold parameters for each sub-recognition model. S3.2: Based on ensemble learning, the sub-recognition models of the four pre-trained lightweight neural network models are integrated. According to the different accuracy levels of the sub-recognition models of the four lightweight neural network models and the industrial application scenarios, weights are assigned to each sub-recognition result of the sub-recognition models of the four lightweight neural network models pre-trained in S3.

1. S3.3: Using recognition accuracy and time as the performance evaluation criteria for each sub-recognition model of the four integrated lightweight neural network models, the most suitable sub-recognition model of the four integrated lightweight neural network models is selected. The image sample test data of the S2.3 test set is input into the selected sub-recognition model of the four integrated lightweight neural network models and then enters the recognition test, and the results are output.

Images