GIS ultrahigh frequency partial discharge signal separation method based on pattern recognition

Abstract

The invention relates to the technical field of pattern recognition, in particular to a GIS ultrahigh frequency partial discharge signal separation method based on pattern recognition, which comprises the following steps: synchronously acquiring original data by using various sensors, and preliminarily eliminating interference data to obtain data to be analyzed; generating a first characteristic spectrogram and a second characteristic spectrogram based on the to-be-analyzed data, and extracting spectrogram features; according to the spectrogram characteristics, judging judgment conditions which are met by the to-be-analyzed data, wherein the judgment conditions comprise a first mode, a second mode and a third mode; if the to-be-analyzed data meets the corresponding discrimination condition, classifying the to-be-analyzed data through a corresponding data classification path; and fusing classification results obtained through the data classification paths based on confidence, outputting a classification result of the to-be-analyzed data, and separating identified real data from interference data.

Description

Technical Field

[0001] The present invention relates to the technical field of pattern recognition, and specifically to a GIS ultra-high frequency partial discharge signal separation method based on pattern recognition. Background Art

[0002] Most of the existing partial discharge signal recognition methods rely on manual feature construction. For example, the energy distribution of the signal in the time-frequency domain, pulse amplitude statistics, envelope shape features, etc. are extracted, and after constructing the feature vector, it is input into traditional machine learning models (such as support vector machines, random forests, etc.) for classification. However, in the case where the pattern boundaries are not obvious between multi-source partial discharge signal samples, it is difficult for such methods to form feature expressions with sufficient discrimination, resulting in insufficient generalization ability of the classifier. In addition, in practical applications, partial discharge signals often have high-similarity feature maps (such as PRPD, PRPS diagrams, etc.), and it is difficult for artificial experience to accurately label and discriminate, further exacerbating the uncertainty of pattern recognition.

[0003] With the development of deep learning, some studies have tried to construct recognition frameworks based on convolutional neural networks (CNNs) or recurrent neural networks (RNNs), and use their self-learning ability to automatically extract features from the original signals. However, most of these methods lack structural design and fail to construct targeted feature extraction channels for the cross-modal and multi-resolution characteristics of partial discharge signals themselves, resulting in unsatisfactory identification effects of the model for different discharge types. At the same time, existing recognition models often extract features from a single perspective, lack a multi-dimensional cross-modeling mechanism, and also fail to fully exploit the collaborative discrimination information between the time-series structure, spectral aggregation, and phase distribution of discharge signals.

[0004] Therefore, aiming at the recognition problems such as the ambiguity in class discrimination of partial discharge signals, small distribution differences in high-dimensional spaces, and serious feature drift under noise interference, it is urgent to construct a high-discrimination feature extraction structure for high-similarity category signals, improve the robustness, accuracy, and adaptive ability of pattern recognition, and solve the problem of highly reliable intelligent classification of multi-class discharge signals under weak signal conditions. Summary of the Invention

[0005] The object of the present invention is to provide a method for separating GIS ultra-high frequency partial discharge signals based on pattern recognition, so as to solve the recognition problems such as the ambiguity in category distinction of partial discharge signals, the small distribution difference in high-dimensional space, and the serious feature drift under noise interference, and improve the accuracy of distinguishing real partial discharge signals from external interference signals. The method includes: synchronously collecting raw data using multiple sensors, and performing preliminary elimination processing on the interference data to obtain data to be analyzed; generating a first feature spectrogram and a second feature spectrogram based on the data to be analyzed, and extracting the spectrogram features; judging the discrimination conditions that the data to be analyzed conforms to according to the spectrogram features, including a first mode, a second mode, and a third mode; if the data to be analyzed conforms to the corresponding discrimination conditions, classifying the data to be analyzed through the corresponding data classification path; fusing the classification results obtained through each data classification path based on the confidence level, outputting the classification result of the data to be analyzed, and separating the identified real data from the interference data.

[0006] To achieve the above object, the present invention provides the following technical solutions: A method for separating GIS ultra-high frequency partial discharge signals based on pattern recognition, including: Synchronously collecting raw data using multiple sensors, and performing preliminary elimination processing on the interference data to obtain data to be analyzed; Generating a first feature spectrogram and a second feature spectrogram based on the data to be analyzed, and extracting the spectrogram features of the first feature spectrogram and the second feature spectrogram; Judging the discrimination conditions that the data to be analyzed conforms to according to the spectrogram features; the discrimination conditions include a first mode, a second mode, and a third mode; If the data to be analyzed conforms to the first mode, it is determined as the first data type, and the data to be analyzed is classified through the first data classification path; If the data to be analyzed conforms to the second mode, it is determined as the second data type, and the data to be analyzed is classified through the second data classification path; If the data to be analyzed conforms to the third mode, it is determined as the third data type, and the data to be analyzed is classified through the third data classification path; Fusing the classification results of the first data classification path, the second data classification path, and the third data classification path based on the confidence level, outputting the classification result of the data to be analyzed, and separating the identified real data from the interference data.

[0007] Preferably, the multiple sensors include: an ultra-high frequency internal sensor, an ultra-high frequency external sensor, a noise sensor, and an ultrasonic sensor; The process of performing preliminary elimination processing on the interference data is: Perform amplitude threshold filtering on the original data to eliminate signals below the background noise baseline threshold and above the device saturation response threshold; Perform short-time energy detection on the original data based on the sampling time window to identify and exclude burst interference pulses with poor energy distribution continuity and extremely short signal duration; Combine the wavelet transform method to perform band-pass filtering on the original data, and only retain the effective frequency band components within the preset working bandwidth range to obtain the data to be analyzed.

[0008] Preferably, generate a first feature spectrogram and a second feature spectrogram based on the data to be analyzed, and extract the spectrogram features of the first feature spectrogram and the second feature spectrogram, specifically including: Perform phase alignment and full-cycle segmentation processing on the data to be analyzed to generate the first feature spectrogram; Perform time-domain pulse sequence tracking processing on the data to be analyzed, and generate the second feature spectrogram based on the amplitude and time interval; Extract the spectrogram features based on the first feature spectrogram and the second feature spectrogram, including spectrogram structure features, trend perturbation features, and spatial delay features; the spectrogram structure features include graphic texture, symmetry, and clustering degree; the trend perturbation features include peak drift and abnormal period; the spatial delay features include the signal arrival time delay difference of multiple sensors.

[0009] Preferably, determine whether the data to be analyzed conforms to the first mode, and the specific steps are as follows: Calculate the point cloud distribution concentration of the spectrogram structure features; evaluate the symmetry index of the spectrogram features around the predetermined symmetry center; obtain the periodic intensity index through autocorrelation analysis of the data to be analyzed; If the point cloud distribution concentration is higher than the preset concentration threshold, the symmetry index is higher than the preset symmetry threshold, and the periodic intensity index is higher than the preset periodic threshold, it is determined as the first data type, and the data to be analyzed is classified through the first data classification path; The first data classification path uses a CNN model based on the spectrogram texture features to classify the data to be analyzed, and outputs the probability label and the first confidence level of the first data type.

[0010] Preferably, determine whether the data to be analyzed conforms to the second mode, and the specific steps are as follows: Calculate the statistical dispersion parameter of the signal trend perturbation features; detect the amplitude sequence of the data to be analyzed, and calculate the adjacent pulse amplitude difference; analyze the pulse repetition time interval and the instantaneous frequency of the data to be analyzed, and calculate the instability index; If the statistical dispersion parameter is higher than a preset perturbation threshold, the difference between adjacent pulse amplitudes exceeds a preset amplitude change threshold, and the instability index is higher than a preset sudden increase frequency threshold, it is determined as the second data type, and the data to be analyzed is classified through the second data classification path; The second data classification path uses a time series feature learning model to analyze the data to be analyzed, and obtains a classification label and a second confidence level of the second data type.

[0011] Preferably, to determine whether the data to be analyzed conforms to the second mode, the specific steps are as follows: Calculate the structural order parameter of the spectrogram structure feature; obtain the spatial positioning result according to the spatial delay feature, and calculate the confidence level of the spatial positioning result; If the structural order parameter is lower than a preset order threshold and the confidence level is lower than a preset delay confidence threshold, it is determined as the third data type, and the data to be analyzed is classified through the third data classification path; The third data classification path uses an interference feature matching and pattern exclusion algorithm to distinguish the data to be analyzed, and obtains a classification label and a third confidence level of the third data type.

[0012] Preferably, the classification results of the first data classification path, the second data classification path, and the third data classification path are fused based on the confidence level, and the classification result of the data to be analyzed is output. The specific steps are as follows: Use a weighted voting mechanism to fuse the classification results of the first data classification path, the second data classification path, and the third data classification path; the weights of the weighted voting mechanism are determined based on the first confidence level, the second confidence level, and the third confidence level; Output the classification result of the fused data to be analyzed, and at the same time generate a confidence level label of the classification result; separate the identified real data and interference data according to the classification result and the confidence level label.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In the feature extraction stage, the present invention generates a first feature spectrogram (discharge phase distribution diagram) and a second feature spectrogram (pulse sequence phase distribution diagram) from the data to be analyzed respectively, further extracts the spectrogram structure feature, trend perturbation feature, and spatial delay feature, and constructs a multi - dimensional high - dimensional feature set. This feature not only enhances the distinguishability between the subsequent three discrimination modes (periodic symmetry type, trend perturbation type, spatial interference type), but also significantly improves the mapping accuracy between data samples and classification paths, avoiding the problem of misclassification into the wrong path, thereby improving the overall classification efficiency and accuracy of the subsequent three - path model.

[0014] 2. The present invention constructs a first data classification path (a CNN model based on spectral texture features), a second data classification path (a temporal feature learning model), and a third data classification path (an interference matching and pattern exclusion algorithm), which respectively correspond to different types of UHF partial discharge signals. By combining the original data collected synchronously by multiple sensors to form spectral features and trend perturbation features, the high-precision intelligent classification of partial discharge signals in GIS equipment is realized. The present invention significantly reduces the probability of misjudgment and missed judgment, and improves the overall recognition accuracy and automation degree of the system.

[0015] 3. In the classification result output stage, the present invention introduces a fusion mechanism based on confidence weighting to integrally and consistently combine the independent classification results from the three data classification paths. This fusion strategy makes full use of the advantages of each path for different data types, and enhances the robustness and stability of the overall recognition and judgment. This mechanism not only improves the consistency of the final classification result in complex scenarios, but also provides a reliable classification basis for the accurate separation of interference signals and real signals, significantly reducing the misclassification rate, and improving the signal restoration and abnormal alarm capabilities of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic flowchart of a method for separating UHF partial discharge signals of GIS based on pattern recognition provided by an embodiment of the present invention; Figure 2 It is a schematic flowchart of determining that the data to be analyzed conforms to the first pattern provided by an embodiment of the present invention; Figure 3 It is a schematic flowchart of determining that the data to be analyzed conforms to the second pattern provided by an embodiment of the present invention; Figure 4 It is a schematic flowchart of determining that the data to be analyzed conforms to the third pattern provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] Gas Insulated Switchgear (GIS for short), as a key high-voltage device in the power system, its operating state has an important impact on the safety and stability of the power transmission and transformation system. In GIS equipment, due to factors such as manufacturing, installation, or operating environment, insulation defects may occur, leading to the phenomenon of partial discharge (PD). Partial discharge is a typical manifestation of early insulation deterioration in GIS and can provide early warning signals before a fault occurs. Therefore, the early identification and interference stripping of PD signals in GIS have become one of the key technical issues in power system condition detection and intelligent diagnosis.

[0019] The present invention proposes a method for separating UHF partial discharge signals in GIS based on pattern recognition, which can solve the identification problems such as the ambiguity in class discrimination of partial discharge signals, the small distribution difference in high-dimensional space, and the serious feature drift under noise interference, and improve the accuracy of distinguishing real partial discharge signals from external interference signals. In order to illustrate that the method of the present invention can play a role in improving the accuracy of distinguishing real partial discharge signals from external interference signals, the effectiveness of the present invention will be described below from two embodiments.

[0020] Embodiment 1 In the embodiment of the present application, the method proposed by the present invention is used to elaborate on the process of solving the identification problems such as the ambiguity in class discrimination of partial discharge signals, the small distribution difference in high-dimensional space, and the serious feature drift under noise interference, and improving the accuracy of distinguishing real partial discharge signals from external interference signals. Figure 1 It is a specific flowchart of the method proposed by the present invention, including: synchronously collecting raw data using multiple sensors, and preliminarily eliminating interference data to obtain data to be analyzed; generating a first feature spectrogram and a second feature spectrogram based on the data to be analyzed, and extracting spectrogram features; judging the discriminant conditions that the data to be analyzed conforms to according to the spectrogram features, including the first mode, the second mode, and the third mode; if the data to be analyzed conforms to the corresponding discriminant conditions, classifying the data to be analyzed through the corresponding data classification paths; fusing the classification results obtained through each data classification path based on the confidence level, outputting the classification result of the data to be analyzed, and separating the identified real data from the interference data. The following is based on Figure 1 the content of Synchronously collecting raw data using multiple sensors, and preliminarily eliminating interference data to obtain data to be analyzed; the multiple sensors include: an ultra-high frequency built-in sensor, an ultra-high frequency external sensor, a noise sensor, and an ultrasonic sensor; The process of preliminarily eliminating interference data is: Perform amplitude threshold filtering on the original data to eliminate signals below the background noise baseline threshold and above the device saturation response threshold; Perform short-time energy detection on the original data based on the sampling time window, and identify and exclude burst interference pulses with poor energy distribution continuity and extremely short signal duration; Combine the wavelet transform method to perform band-pass filtering on the original data, and only retain the effective frequency band components within the preset working bandwidth range to obtain the data to be analyzed.

[0021] Specifically, the original data is the original partial discharge signal; The background noise baseline threshold is dynamically set according to the on-site environmental noise level, for example, set to be 3 to 6 dB higher than the average noise level; the device saturation response threshold is set according to the maximum linear response range of the used sensor and data acquisition system to avoid signal clipping distortion.

[0022] In the short-time energy detection, the sampling time window is set to several power frequency cycles, for example, 5 power frequency cycles. By calculating the signal energy within the window and comparing it with the preset energy threshold, and combining the pulse duration (for example, pulses below 10 ns are regarded as burst interference) for judgment.

[0023] In the wavelet transform method, the maximum overlap discrete wavelet transform is adopted, and wavelet bases such as db4 or sym8 are selected. The original signal is decomposed into 5 to 8 layers. According to the UHF frequency band characteristics of typical partial discharge signals in GIS equipment (within the range of 0.3 GHz to 3 GHz), the wavelet coefficients corresponding to the corresponding frequency band are selected for reconstruction to achieve band-pass filtering; amplitude normalization adopts the maximum-minimum normalization method to map the signal amplitude to the interval [0,1].

[0024] Through the collaborative acquisition of the above-mentioned multiple sensors and the targeted preliminary elimination of interference data, it is possible to effectively suppress electromagnetic interference from space, pulse interference caused by poor contact, and some narrow-band interference of non-partial discharge. At the same time, the effective frequency band components of the partial discharge signal are retained, and standardized processing is carried out, significantly improving the signal-to-noise ratio and normality of the original data, laying a solid data foundation for subsequent accurate feature extraction and pattern recognition, and reducing the complexity of subsequent processing.

[0025] Preferably, generate a first feature spectrogram and a second feature spectrogram based on the data to be analyzed, and extract the spectrogram features of the first feature spectrogram and the second feature spectrogram; specifically include: Perform phase alignment and full-cycle segmentation processing on the data to be analyzed to generate the first feature spectrogram; Perform time-domain pulse sequence tracking processing on the data to be analyzed, and generate the second feature spectrogram based on the amplitude and time interval; Extracting the spectrogram features based on the first characteristic spectrogram and the second characteristic spectrogram includes spectrogram structure features, trend perturbation features, and spatial delay features; the spectrogram structure features include graphic texture, symmetry, and clustering degree; the trend perturbation features include peak drift and period anomaly; the spatial delay features include the signal arrival time delay difference of multiple sensors.

[0026] Specifically, when generating the first characteristic spectrogram (discharge phase distribution diagram), first obtain the phase reference through the synchronously collected power frequency voltage signal, count each UHF pulse signal according to its occurring power frequency phase (0 - 360°) and amplitude, and accumulate data for multiple power frequency cycles (for example, 100 - 500 cycles) to form a two-dimensional statistical histogram. When generating the second characteristic spectrogram (pulse sequence phase distribution diagram), record the occurrence time of consecutive pulses and the pulse amplitude , and display the time correlation of the pulse sequence in a two-dimensional plane in the form of (representing the time interval between the pulse charge amount and the previous pulse), where is the modulo operation, and is the power frequency cycle.

[0027] Extract the spectrogram structure features: For graphic texture, calculate texture parameters such as energy, contrast, and correlation using the gray-level co-occurrence matrix; for symmetry, calculate the symmetry index of the first characteristic spectrogram (discharge phase distribution diagram) about the 90° and 270° phase axes; for clustering degree, identify the number and density of the main discharge clusters through the DBSCAN clustering algorithm. Extract the trend perturbation features: For peak drift, monitor the change in the amplitude mean of the main discharge clusters in the first characteristic spectrogram (discharge phase distribution diagram) at different time intervals (such as per minute or per hour); for period anomaly, analyze the stability of the pulse repetition frequency and detect whether there are abnormal pulsations with non-power frequency periods. Extract the spatial delay features: Use the same discharge pulse signal received by at least three UHF sensors, and calculate the precise time difference of the signal arriving at different sensors through the generalized cross-correlation algorithm for subsequent positioning and interference elimination.

[0028] By generating two key characteristic spectrograms, namely the discharge phase distribution diagram and the pulse sequence phase distribution diagram, and extracting multi-dimensional features such as structure, trend, and spatial delay from them, the essential attributes of the signal can be comprehensively characterized from multiple angles such as discharge phase distribution, pulse time series correlation, signal amplitude stability, periodicity, and spatial propagation characteristics. This combination of multi-dimensional features significantly enhances the ability to distinguish different types of partial discharge and interference signals, compensates for the deficiency of the discrimination ability of single features under complex working conditions, and provides a basis for subsequent pattern determination of the data to be analyzed through the spectrogram features.

[0029] Table 1 is a comparison table of classification accuracies of introducing bispectrum diagrams and feature extraction with other methods. Among them, the classification accuracy is obtained by obtained.

[0030] Table 1 Comparison Table of Classification Accuracies of Introducing Bispectrum Diagrams and Feature Extraction

[0031] Preferably, determine the discriminant conditions that the data to be analyzed conforms to according to the spectral diagram features; the discriminant conditions include a first mode, a second mode, and a third mode; If the data to be analyzed conforms to the first mode, it is determined as the first data type, and the data to be analyzed is classified through the first data classification path; Judge that the data to be analyzed conforms to the first mode, referring to Figure 2 , and the specific steps are: Calculate the concentration degree of the point cloud distribution of the spectral diagram structure features; evaluate the symmetry index of the spectral diagram features around a predetermined symmetry axis; obtain the periodic intensity index through autocorrelation analysis of the data to be analyzed; If the concentration degree of the point cloud distribution is higher than the preset concentration threshold, the symmetry index is higher than the preset symmetry threshold, and the periodic intensity index is higher than the preset periodic threshold, it is determined as the first data type, and the data to be analyzed is classified through the first data classification path; The first data classification path uses a CNN model based on spectral texture features to classify the data to be analyzed, and outputs the probability label and the first confidence level of the first data type.

[0032] Specifically, the concentration degree of the point cloud distribution is quantified as the energy ratio of the high-density area in the discharge phase distribution diagram, and the preset concentration threshold is 0.6; the symmetry index is the reciprocal of the difference in pixel values on both sides of the specific symmetry axis of the spectral diagram, and the preset symmetry threshold is 0.7; the periodic intensity index is obtained by performing FFT transformation on the pulse sequence and observing whether there are significant peaks at the power frequency and its multiples. The preset periodic threshold is set according to the statistical characteristics of typical metal particle discharges.

[0033] The first data type is the metal particle type discharge signal; in the first data classification path, the CNN model includes 3 convolutional layers and 2 fully connected layers, the convolutional kernel size is 3×3, the activation function uses ReLU, and the pooling layer uses max pooling. The input is the discharge phase distribution diagram or the pulse sequence phase distribution diagram, and the output layer outputs the probability value including the metal particle discharge signal through the Softmax function. This model is pre-trained on a large number of spectral diagram samples including known metal particle discharges and other types of discharges / interferences. The first confidence level is the probability value of the corresponding category output by the Softmax function.

[0034] By quantitatively judging the concentration, symmetry and periodicity of the spectral structure features, and combining with a CNN model optimized specifically for the spectral texture features for fine classification, it is possible to efficiently and accurately identify metal particle type discharge signals with typical "rabbit ear shape" or symmetric aggregation morphology. This two-stage judgment and classification mechanism first quickly screens out highly suspected signals and then confirms them through a deep learning model, significantly improving the recognition accuracy and robustness of metal particle type discharges and effectively avoiding confusion with other types of signals.

[0035] Preferably, if the data to be analyzed conforms to the second pattern, it is determined as the second data type, and the data to be analyzed is classified through the second data classification path; Judge that the data to be analyzed conforms to the second pattern, referring to Figure 3 , and the specific steps are as follows: Calculate the statistical dispersion parameter of the signal trend perturbation feature; detect the amplitude sequence of the data to be analyzed and calculate the adjacent pulse amplitude difference; analyze the pulse repetition time interval and instantaneous frequency of the data to be analyzed and calculate the instability index; If the statistical dispersion parameter is higher than the preset perturbation threshold, the adjacent pulse amplitude difference exceeds the preset amplitude change threshold, and the instability index is higher than the preset sudden increase frequency threshold, it is determined as the second data type, and the data to be analyzed is classified through the second data classification path; The second data classification path uses a time series feature learning model to analyze the data to be analyzed, and obtains the classification label and the second confidence level of the second data type.

[0036] Specifically, the statistical dispersion parameter is obtained by calculating the standard deviation of the pulse amplitude sequence, and the preset perturbation threshold is set according to the statistical value of the reference stable signal (such as exceeding 3 times the standard deviation of the normal fluctuation); the adjacent pulse amplitude difference is obtained by calculating and the preset amplitude change threshold is 20% of the maximum amplitude; the instability index is obtained by calculating the coefficient of variation (such as the standard deviation) of the pulse time interval sequence, and the preset sudden increase frequency threshold is 0.4.

[0037] The second data type is internal corona and / or insulation breakdown type signal; in the second data classification path, the time series feature learning model uses an LSTM model including 2 layers of LSTM (128 units in each layer) and an output fully connected layer, and the input is the extracted pulse amplitude sequence, time interval sequence, etc.; learn the sequence data of 100 consecutive pulses, and the output is the classification label of the internal corona and / or insulation breakdown type signal and the corresponding Softmax probability as the second confidence level.

[0038] By comprehensively evaluating the perturbation of the signal trend, the severity of the amplitude change, and the stability of the pulse frequency, and using a time-series feature learning model to deeply explore the dynamic evolution law of the signal in the time dimension, this method can effectively capture the continuous cluster-like characteristics of internal corona discharge and / or the sudden and high-energy characteristics of insulation breakdown. This improves the recognition sensitivity and accuracy of such rapidly developing and harmful internal discharges, providing a reliable basis for timely early warning.

[0039] Preferably, if the data to be analyzed conforms to the third mode, it is determined as the third data type, and the data to be analyzed is classified through the third data classification path; Judge that the data to be analyzed conforms to the third mode, refer to Figure 4 , and the specific steps are as follows: Calculate the structural order parameter of the spectrogram structure feature; obtain the spatial positioning result according to the spatial delay feature, and calculate the confidence of the spatial positioning result; If the structural order parameter is lower than the preset order threshold and the confidence is lower than the preset delay confidence threshold, it is determined as the third data type, and the data to be analyzed is classified through the third data classification path; The third data classification path uses an interference feature matching and pattern exclusion algorithm to distinguish the data to be analyzed, and obtains the classification label and the third confidence of the third data type.

[0040] Specifically, the structural order parameter is obtained by calculating the entropy value of the discharge phase distribution diagram. The higher the entropy value, the lower the order. The preset order threshold is set according to expert experience; the confidence of the spatial positioning result is evaluated by comparing the consistency of the positioning results of multiple sensor combinations after TDOA positioning. The preset delay confidence threshold is 0.5; The third data type is interference and / or floating discharge type signal; in the third data classification path, interference feature matching first compares the features of the data to be analyzed (such as specific peaks in the spectrum, morphological parameters of the discharge phase distribution diagram) with the pre-stored interference feature database (such as the feature signatures of switch noise, radio interference); the pattern exclusion algorithm determines it as an interference and / or floating discharge type signal when the signal does not conform to the strong features of the first and second modes, and the matching degree with the known interference feature library is not high, but the spectrum is disordered and the positioning is fuzzy, and outputs the classification label and the third confidence based on the exclusion rule.

[0041] Table 2 is a comparison data table of the partial discharge classification accuracy and stability of different methods. Among them, the classification accuracy is obtained through ; the misjudgment rate is obtained through .

[0042] Table 2 Comparison data table of different methods in partial discharge classification accuracy and stability

[0043] By evaluating the disorder of the spectrogram and the ambiguity of the multi-sensor positioning results, and combining the interference feature library matching and pattern exclusion strategies, this method can effectively identify and separate interference signals with complex structures and uncertain sources, as well as some floating discharge signals. This mechanism avoids misjudging interference as real discharge, and also provides a reasonable way out for signals that are difficult to classify, improving the anti-interference ability of the overall system and the comprehensiveness of diagnosis.

[0044] Preferably, the classification results of the first data classification path, the second data classification path, and the third data classification path are fused based on confidence, and the classification result of the data to be analyzed is output, and the identified real data and interference data are separated; the specific steps are as follows: Use the weighted voting mechanism to fuse the classification results of the first data classification path, the second data classification path, and the third data classification path; the weights of the weighted voting mechanism are determined based on the first confidence, the second confidence, and the third confidence; Output the classification result of the data to be analyzed after fusion, and at the same time generate a confidence level label for the classification result; separate the identified real data and interference data according to the classification result and the confidence level label.

[0045] Specifically, in the weighted voting mechanism, the classification result output by each path and its confidence . If a certain path is activated (that is, the discrimination condition is satisfied), then its weight ( is the index of all activated paths), and the weight of the unactivated path is 0. The final classification result is the category with the highest weighted confidence. For example, if the first path determines it as an effective partial discharge (confidence 0.82), the second path determines it as an interference signal (confidence 0.65), and the third path determines it as an effective partial discharge (confidence 0.72), then it is finally determined as an effective partial discharge (confidence 0.78) through the weighted voting mechanism. If only one path is activated, its result is directly adopted. Set the confidence level label >0.85 as high, 0.6 - 0.85 as medium, and <0.6 as low. When separating real data and interference data, those classified as metal particles, internal corona, insulation breakdown and with high or medium confidence are determined as real partial discharge signals, and the rest can be marked as interference signals.

[0046] Through the confidence-weighted fusion mechanism of multi-path classification results, the advantages of different recognition strategies are integrated, effectively improving the accuracy and reliability of the final decision on complex and ambiguous signals. The introduction of confidence level labels provides a decision-making reference for operation and maintenance personnel, facilitating the distinction between real discharges with high risks and general interferences, realizing effective hierarchical management and precise separation of monitoring data, and improving the practical value and automation level of the monitoring system.

[0047] The method for separating UHF partial discharge signals of GIS based on pattern recognition proposed by the present invention systematically solves the problems of inaccurate recognition of partial discharge signals and susceptibility to interference in complex electromagnetic environments through multi-sensor data acquisition and refined preprocessing, multi-dimensional feature extraction and spectrogram generation, multi-mode recognition path selection based on signal characteristics, and final confidence fusion decision-making. This method can not only accurately distinguish different types of partial discharges (such as metal particles, internal corona / breakdown, floating discharge), but also effectively identify and eliminate various interference signals, significantly improving the accuracy and intelligent level of GIS equipment condition monitoring, and providing strong technical support for ensuring the safe and stable operation of power systems.

[0048] Embodiment 2 In Embodiment 1, the method proposed by the present invention successfully solves the recognition problems such as the ambiguity in class discrimination of partial discharge signals, the small distribution difference in high-dimensional space, and the serious feature drift under noise interference, improving the accuracy of distinguishing real partial discharge signals from external interference signals. To further verify the effectiveness of the present invention, the partial discharge signals generated in another GIS device are also distinguished and recognized in the embodiments of this application.

[0049] Synchronously collect raw data using multiple sensors, and perform preliminary elimination processing on interference data to obtain data to be analyzed; the multiple sensors include: an ultra-high frequency internal sensor, an ultra-high frequency external sensor, a noise sensor, and an ultrasonic sensor; The process of performing preliminary elimination processing on the interference data is as follows: Perform amplitude threshold filtering processing on the raw data to eliminate signals below the background noise baseline threshold and above the device saturation response threshold; Based on the sampling time window, perform short-time energy detection on the raw data to identify and exclude burst interference pulses with poor energy distribution continuity and abnormally short signal duration; Combined with the wavelet transform method, perform band-pass filtering on the raw data, and only retain the effective frequency band components within the preset working bandwidth range to obtain the data to be analyzed.

[0050] Preferably, generate a first feature spectrogram and a second feature spectrogram based on the data to be analyzed, and extract the spectrogram features of the first feature spectrogram and the second feature spectrogram; specifically include: Perform phase alignment and integer-cycle segmentation processing on the data to be analyzed to generate the first feature spectrogram; Perform time-domain pulse sequence tracking processing on the data to be analyzed, and generate the second feature spectrogram based on the amplitude and time interval; Extract spectrogram features based on the first feature spectrogram and the second feature spectrogram, including spectrogram structure features, trend perturbation features, and spatial delay features; the spectrogram structure features include graphic texture, symmetry, and clustering degree; the trend perturbation features include peak drift and period anomaly; the spatial delay features include the signal arrival time delay difference of multiple sensors.

[0051] Preferably, judge the discrimination conditions that the data to be analyzed conforms to according to the spectrogram features; the discrimination conditions include the first mode, the second mode, and the third mode; If the data to be analyzed conforms to the first mode, it is determined as the first data type, and the data to be analyzed is classified through the first data classification path; the specific steps for judging that the data to be analyzed conforms to the first mode are as follows: Calculate the point cloud distribution concentration of the spectrogram structure features; evaluate the symmetry index of the spectrogram features around a predetermined symmetry center; obtain the periodic intensity index through autocorrelation analysis of the data to be analyzed; If the point cloud distribution concentration is higher than the preset concentration threshold, the symmetry index is higher than the preset symmetry threshold, and the periodic intensity index is higher than the preset periodic threshold, it is determined as the first data type, and the data to be analyzed is classified through the first data classification path; The first data classification path uses a CNN model based on spectrogram texture features to classify the data to be analyzed, and outputs the probability label and the first confidence level of the first data type.

[0052] Preferably, if the data to be analyzed conforms to the second mode, it is determined as the second data type, and the data to be analyzed is classified through the second data classification path; The specific steps for judging that the data to be analyzed conforms to the second mode are as follows: Calculate the statistical dispersion parameter of the signal trend perturbation features; detect the amplitude sequence of the data to be analyzed, and calculate the adjacent pulse amplitude difference; analyze the pulse repetition time interval and the instantaneous frequency of the data to be analyzed, and calculate the instability index; If the statistical dispersion parameter is higher than the preset perturbation threshold, the adjacent pulse amplitude difference exceeds the preset amplitude change threshold, and the instability index is higher than the preset sudden increase frequency threshold, it is determined as the second data type, and the data to be analyzed is classified through the second data classification path; The second data classification path uses a time series feature learning model to analyze the data to be analyzed, obtaining a classification label and a second confidence level for the second data type.

[0053] Preferably, if the data to be analyzed conforms to the third pattern, it is determined as the third data type, and the data to be analyzed is classified through the third data classification path; The steps for determining whether the data to be analyzed conforms to the third pattern are as follows: Calculate the structural order parameter of the spectrogram structure feature; obtain the spatial positioning result based on the spatial delay feature, and calculate the confidence level of the spatial positioning result; If the structural order parameter is lower than a preset order threshold and the confidence level is lower than a preset delay confidence threshold, it is determined as the third data type, and the data to be analyzed is classified through the third data classification path; The third data classification path uses an interference feature matching and pattern exclusion algorithm to distinguish the data to be analyzed, obtaining a classification label and a third confidence level for the third data type.

[0054] Preferably, the classification results of the first data classification path, the second data classification path, and the third data classification path are fused based on the confidence levels, outputting the classification result of the data to be analyzed, and separating the identified real data from the interference data; the specific steps are as follows: Use a weighted voting mechanism to fuse the classification results of the first data classification path, the second data classification path, and the third data classification path; the weights of the weighted voting mechanism are determined based on the first confidence level, the second confidence level, and the third confidence level; Output the classification result of the fused data to be analyzed, and at the same time generate a confidence level label for the classification result; separate the identified real data from the interference data according to the classification result and the confidence level label.

[0055] Table 3 shows the influence of the fusion mechanism on the final classification result and the interference separation effect, where , represents the proportion of different classification results output by the three classification paths under the same input; , represents the ability of the system to completely retain the real signal during processing.

[0056] Table 3 Influence of the Fusion Mechanism on the Final Classification Result and the Interference Separation Effect

[0057] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for separating GIS UHF partial discharge signals based on pattern recognition, characterized in that, Including: Collecting raw data and preprocessing it to obtain the data to be analyzed; Generating a first feature spectrogram and a second feature spectrogram based on the data to be analyzed, and extracting spectrogram features; Judging the discrimination conditions that the data to be analyzed conforms to according to the spectrogram features; the discrimination conditions include a first mode, a second mode, and a third mode; If the data to be analyzed conforms to the first mode, it is determined as the first data type, and the data to be analyzed is classified through the first data classification path; If the data to be analyzed conforms to the second mode, it is determined as the second data type, and the data to be analyzed is classified through the second data classification path; If the data to be analyzed conforms to the third mode, it is determined as the third data type, and the data to be analyzed is classified through the third data classification path; According to the classification results of the first data classification path, the second data classification path, and the third data classification path, output the classification result of the data to be analyzed.

2. The method for separating GIS ultra-high frequency partial discharge signals based on pattern recognition according to claim 1, wherein The raw data is synchronously collected by a variety of sensors, and the variety of sensors include: an ultra-high frequency built-in sensor, an ultra-high frequency external sensor, a noise sensor, and an ultrasonic sensor; The preprocessing is to initially eliminate interference data, and the specific process is: performing amplitude threshold filtering on the raw data to eliminate signals below the background noise baseline threshold and above the device saturation response threshold; performing short-time energy detection on the raw data based on the sampling time window to identify and exclude burst interference pulses with poor energy distribution continuity and abnormally short signal duration; combining the wavelet transform method to perform band-pass filtering on the raw data, and only retaining the effective frequency band components within the preset working bandwidth range to obtain the data to be analyzed.

3. The method for separating GIS ultra-high frequency partial discharge signals based on pattern recognition according to claim 1, wherein Generating a first feature spectrogram and a second feature spectrogram based on the data to be analyzed, and extracting spectrogram features, specifically including: Performing phase alignment and integer period segmentation processing on the data to be analyzed to generate the first feature spectrogram; performing time-domain pulse sequence tracking processing on the data to be analyzed, and generating the second feature spectrogram based on the amplitude and time interval; Extracting the spectrogram features based on the first feature spectrogram and the second feature spectrogram, including spectrogram structure features, trend perturbation features, and spatial delay features; the spectrogram structure features include graphic texture, symmetry, and clustering degree; the trend perturbation features include peak drift and abnormal period; the spatial delay features include the signal arrival time delay difference of multiple sensors.

4. The method for separating GIS ultra-high frequency partial discharge signals based on pattern recognition according to claim 1, wherein Judging that the data to be analyzed conforms to the first mode, the specific steps are: Calculating the point cloud distribution concentration of the spectrogram structure features; evaluating the symmetry index of the spectrogram features around a predetermined symmetry center; obtaining the periodic intensity index through autocorrelation analysis of the data to be analyzed; If the point cloud distribution concentration is higher than the preset concentration threshold, the symmetry index is higher than the preset symmetry threshold, and the periodic intensity index is higher than the preset periodic threshold, it is determined as the first data type, and the data to be analyzed is classified through the first data classification path; The first data classification path uses a CNN model based on the spectrogram texture features to classify the data to be analyzed, and outputs the probability label and the first confidence level of the first data type.

5. The method for separating GIS UHF partial discharge signals based on pattern recognition according to claim 1, characterized in that, To determine whether the data to be analyzed conforms to the second pattern, the specific steps are as follows: Calculate the statistical dispersion parameter of the trend perturbation feature; detect the amplitude sequence of the data to be analyzed and calculate the difference in adjacent pulse amplitudes; Analyze the pulse repetition time interval and instantaneous frequency of the data to be analyzed and calculate the instability index; If the statistical dispersion parameter is higher than the preset perturbation threshold, the difference in adjacent pulse amplitudes exceeds the preset amplitude change threshold, and the instability index is higher than the preset sudden increase frequency threshold, it is determined as the second data type, and the data to be analyzed is classified through the second data classification path; The second data classification path uses a time series feature learning model to analyze the data to be analyzed to obtain the classification label and the second confidence level of the second data type.

6. The method for separating GIS ultra-high frequency partial discharge signals based on pattern recognition according to claim 1, wherein To determine whether the data to be analyzed conforms to the third pattern, the specific steps are as follows: Calculate the structural order parameter of the spectrogram structure feature; obtain the spatial positioning result based on the spatial delay feature and calculate the confidence level of the spatial positioning result; If the structural order parameter is lower than the preset order threshold and the confidence level is lower than the preset delay confidence threshold, it is determined as the third data type, and the data to be analyzed is classified through the third data classification path; The third data classification path uses an interference feature matching and pattern exclusion algorithm to distinguish the data to be analyzed to obtain the classification label and the third confidence level of the third data type.

7. The method for separating GIS UHF partial discharge signals based on pattern recognition according to claim 1, characterized in that According to the classification results of the first data classification path, the second data classification path, and the third data classification path, output the classification result of the data to be analyzed. The specific steps are as follows: Use a weighted voting mechanism to fuse the classification results of the first data classification path, the second data classification path, and the third data classification path; the weights of the weighted voting mechanism are determined based on the first confidence level, the second confidence level, and the third confidence level; Output the classification result of the fused data to be analyzed, and at the same time generate the confidence level label of the classification result; Separate the identified real data and interference data according to the classification result and the confidence level label.

Images