CN-122020435-A - Cable partial discharge identification method for multi-layer feature fusion and layered mRMR

CN122020435ACN 122020435 ACN122020435 ACN 122020435ACN-122020435-A

Abstract

The invention belongs to the technical field of high-voltage cable detection, and particularly relates to a cable partial discharge identification method of multi-layer feature fusion and layered mRMR, which constructs a three-layer feature system comprising physical features, image texture features and shape features, extracts 39-dimensional features altogether, and characterizes partial discharge characteristics from multiple dimensions such as energy distribution, texture modes, geometric forms and the like. On the basis, a layering mRMR method is adopted to select the characteristics, and a random forest classifier is combined to realize cable defect identification. In order to improve the interpretability of the model, a SHAP method is adopted to analyze the contribution degree of different feature layers and key features in classification decisions. Experimental results show that the model still maintains higher recognition accuracy under the condition that the feature dimension is compressed from 39 dimensions to 15 dimensions. The method can effectively reduce the feature redundancy, reveals the roles of different feature layers in discharge identification, and provides a feasible scheme for interpretation analysis and engineering application of cable partial discharge signals.

Inventors

Song guangdong
FENG JINGJIE
WANG JIQIANG
HOU MOYU
ZHANG HUA
HAN HAIYAN
Wang Zunting

Assignees

山东省科学院激光研究所

Dates

Publication Date: 20260512
Application Date: 20260416

Claims (7)

1. A cable partial discharge identification method of multilayer feature fusion and layering mRMR is characterized by comprising the following steps: Step one, collecting cable partial discharge signals to obtain a time-frequency diagram matrix of each section of cable partial discharge signals; The method comprises the steps of constructing a physical layer, constructing a pattern layer, converting a time-frequency image matrix into a gray image matrix, extracting 14-dimensional features from the gray image matrix through a gray symbiotic matrix, extracting 10-dimensional features from the gray image matrix through a local binary mode, constructing a shape layer, performing binarization processing on the gray image matrix to obtain a binarized gray image matrix, extracting 7-dimensional Hu invariant moment from the binarized gray image matrix, and obtaining 39-dimensional features; Step three, extracting 39-dimensional characteristics of partial discharge signals of each section of cable according to the step two, wherein characteristic vectors of all samples and labels form a data set together, and the data set is divided according to the proportion of 70% training set and 30% testing set; Dividing 39-dimensional features into three layers according to sources, namely physical layer features, pattern layer features and shape layer features, executing hierarchical mRMR feature sequencing on each layer, respectively selecting features with the most discriminating capability from all layers according to a preset proportion, constructing a low redundancy feature subset on a training set, calculating the mean value and standard deviation of each feature of the low redundancy feature subset on the training set, storing the mean value and standard deviation, completing fitting of a normalizer, and then respectively carrying out centering and scaling on each feature of the low redundancy feature subset of the training set by utilizing the stored mean value and standard deviation to obtain the normalized low redundancy feature subset on the training set; training a random forest classification model based on the standardized low-redundancy feature subsets on the training set, and identifying different types of partial discharge to obtain a trained random forest classification model; And step six, extracting the corresponding low-redundancy feature subset from the test set according to the index of the low-redundancy feature subset on the training set, carrying out centering and scaling treatment on the test set by using a normalizer fitted by the training set, then calling a trained random forest classification model to complete prediction on the test set sample, obtaining a prediction label and classification probability, calculating classification accuracy, generating a confusion matrix and storing.
2. The method for identifying partial discharge of a cable by multi-layer feature fusion and layered mRMR according to claim 1, further comprising a step seven of calculating marginal contribution of each feature of the standardized low-redundancy feature subset in the training set in random forest classification model prediction by SHAP, and normalizing the feature contribution of each category to quantify the relative importance degree of each feature and each feature layer in classification decision.
3. The method for identifying the partial discharge of the cable by the multi-layer feature fusion and layering mRMR according to claim 1 or 2, wherein in the first step, the partial discharge signals of each section of the cable are subjected to normalization processing after DC removal, and then the partial discharge signals of the cable are subjected to time-frequency representation by adopting short-time Fourier transformation, so that a time-frequency diagram matrix of the partial discharge signals of each section of the cable is obtained.
4. The method for identifying partial discharge of a cable by multi-layer feature fusion and layering mRMR according to claim 1 or 2, wherein in the fourth step, 4-dimensional features are selected by a physical layer, 8-dimensional features are selected by a pattern layer, 3-dimensional features are selected by a shape layer, and a 15-dimensional low redundancy feature subset is constructed.
5. The method for identifying partial discharge of a cable by multi-layer feature fusion and layering mRMR according to claim 1 or 2, wherein in the fourth step, the Pearson correlation coefficient is used for redundancy verification of the low redundancy feature subset.
6. The method for identifying partial discharge of a cable with multi-layer feature fusion and layering mRMR as set forth in claim 5, wherein in step four, the number of redundant feature pairs is counted by setting a threshold value, and the overall redundancy degree is evaluated by assisting with an average absolute correlation coefficient and a maximum absolute correlation coefficient.
7. The method for identifying partial discharge of a cable by multi-layer feature fusion and layering mRMR according to claim 1 or 2, wherein in the fifth step, the number of decision trees in the random forest classification model is set to 200, and the fixed random seeds are set to 42.

Description

Cable partial discharge identification method for multi-layer feature fusion and layered mRMR Technical Field The invention belongs to the technical field of high-voltage cable detection, and particularly relates to a cable partial discharge identification method of multilayer feature fusion and layering mRMR. Background Long-term safe operation of high voltage cables has a significance for power system reliability, where partial discharge is an important precursor to aging of cable insulation and potential failure. The AE signals (acoustic emission signals) generated by the partial discharge of the cable are identified and classified, so that early diagnosis of different defect types is facilitated, and an important basis is provided for the operation and maintenance of the cable. At present, aiming at the problem of partial discharge signal identification, scholars at home and abroad have proposed various signal characterization and mode identification methods, and are widely applied to cable insulation state evaluation and fault diagnosis. In general, existing studies can be largely classified into two types, namely, an automatic feature learning method based on deep learning and a conventional machine learning method based on artificial feature extraction. The method realizes automatic feature extraction and classification by constructing a deep network model, for example Xu Chenhang and the like propose a PRPD spectrogram identification method based on a deep residual network, rajat Srivastava and the like propose a partial discharge identification method based on an optimized convolutional neural network, and the identification capability of a discharge mode is improved through model optimization. On the other hand, the traditional method still has better engineering applicability by constructing features with physical significance and combining with a classifier for identification. For example, from the point of view of image processing, the multi-dimensional statistical features are extracted from time domain signals by using the gray level CO-occurrence matrix (GRAY LEVEL CO-Occurrence Matrix, GLCM) and the local binary pattern (Local binary patterns, LBP) to extract the local discharge image texture features by combining a support vector machine (Support Vector Machine, SVM) to realize cable insulation state identification, and from the point of view of image processing, zhao Lei is used to effectively improve identification performance. In general, existing studies model and analyze partial discharge signals from different angles and have made some progress. However, the above method has a certain limitation while achieving a high recognition accuracy. Firstly, although the deep learning-based method has strong feature expression capability, the model structure is complex, the decision process lacks clear physical explanation, and the actual contribution of different features to the classification result is difficult to reveal. Second, artificial feature extraction-based methods typically focus on single-dimensional information characterization, e.g., using only temporal statistical features or image texture features, lacking systematic analysis of complementary relationships between multi-source features. In addition, the existing research is mainly carried out by modeling from the whole feature set, the structural description of the discharge signal is less from the feature level angle, and the comprehensive characteristics of partial discharge in the aspects of energy distribution, texture mode, morphological structure and the like are difficult to be simultaneously described, so that the interpretability and generalization capability of the model are limited to a certain extent. Disclosure of Invention The invention aims to overcome the defects of the background technology and provide a cable partial discharge identification method with multi-layer feature fusion and layered mRMR. The invention is realized by the following technical scheme: a cable partial discharge identification method of multilayer feature fusion and layering mRMR comprises the following steps: Step one, collecting cable partial discharge signals to obtain a time-frequency diagram matrix of each section of cable partial discharge signals. The method comprises the steps of constructing a physical layer, extracting 8-dimensional features from a time-frequency image matrix, constructing a pattern layer, converting the time-frequency image matrix into a gray image matrix, extracting 14-dimensional features from the gray image matrix through a gray level co-occurrence matrix, extracting 10-dimensional features from the gray image matrix through a local binary pattern, constructing a shape layer, performing binarization processing on the gray image matrix to obtain a binarized gray image matrix, extracting 7-dimensional Hu invariant moment from the binarized gray image matrix, and obtaining 39-dimensional features. And thirdly,