Separation method of GIS UHF partial discharge signal based on pattern recognition

Abstract

The present invention relates to the field of pattern recognition technology, and specifically to a GIS ultra-high frequency partial discharge signal separation method based on pattern recognition. The method comprises the following steps: using a plurality of sensors to synchronously collect raw data, and preliminarily eliminating interference data to obtain data to be analyzed; generating a first characteristic spectrum and a second characteristic spectrum based on the data to be analyzed, and extracting spectrum features; judging, based on the spectrum features, whether the data to be analyzed meets a discrimination condition, including a first mode, a second mode, and a third mode; if the data to be analyzed meets the corresponding discrimination condition, classifying the data to be analyzed through a corresponding data classification path; fusing the classification results obtained through each data classification path based on confidence, outputting the classification result of the data to be analyzed, and separating the identified real data from the interference data.

Description

Technical Field

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

[0002] Most existing PD signal recognition methods rely on manual feature construction. For example, they extract the signal's energy distribution, pulse amplitude statistics, and envelope morphology in the time-frequency domain. These construct feature vectors and then input them into traditional machine learning models (such as support vector machines and random forests) for classification. However, when the pattern boundaries between multi-source PD signal samples are unclear, these methods struggle to generate sufficiently discriminative feature representations, resulting in insufficient classifier generalization. Furthermore, in practical applications, PD signals often exhibit highly similar feature maps (such as PRPD and PRPS maps), making accurate labeling and identification difficult through manual experience, further exacerbating the uncertainty of pattern recognition.

[0003] With the development of deep learning, some studies have attempted to construct recognition frameworks based on convolutional neural networks (CNNs) or recurrent neural networks (RNNs), leveraging their self-learning capabilities to automatically extract features from raw signals. However, most of these methods lack structural design and fail to construct targeted feature extraction channels tailored to the cross-modal and multi-resolution characteristics of partial discharge signals. As a result, the models' ability to distinguish different discharge types remains suboptimal. Furthermore, existing recognition models often extract features from a single perspective, lacking multidimensional cross-modeling mechanisms and failing to fully exploit the collaborative discriminative information between the discharge signal's temporal structure, spectral clustering, and phase distribution.

[0004] Therefore, in order to address the recognition difficulties of partial discharge signals, such as the ambiguity in category distinction, small distribution differences in high-dimensional space, and severe feature drift under noise interference, it is urgent to construct a high-discrimination feature extraction structure for highly similar category signals, improve the robustness, accuracy and adaptability of pattern recognition, and solve the problem of highly reliable intelligent classification of multiple types of discharge signals under weak signal conditions. Summary of the Invention

[0005] The present invention aims to provide a GIS ultra-high frequency partial discharge signal separation method based on pattern recognition to address identification challenges such as ambiguity in the classification of partial discharge signals, small distribution differences in high-dimensional space, and severe feature drift under noise interference, thereby improving the accuracy of distinguishing true partial discharge signals from external interference signals. The method comprises: using multiple sensors to synchronously collect raw data and preliminarily eliminating interference data to obtain data to be analyzed; generating a first characteristic spectrum and a second characteristic spectrum based on the data to be analyzed, and extracting spectrum features; judging whether the data to be analyzed meets the discrimination conditions, including a first mode, a second mode, and a third mode, based on the spectrum features; if the data to be analyzed meets 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 confidence, outputting the classification results of the data to be analyzed, and separating the identified true data from the interference data.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The GIS UHF partial discharge signal separation method based on pattern recognition includes:

[0008] Use multiple sensors to synchronously collect raw data, and preliminarily eliminate interference data to obtain data to be analyzed;

[0009] generating a first characteristic spectrum and a second characteristic spectrum based on the data to be analyzed, and extracting spectrum features of the first characteristic spectrum and the second characteristic spectrum;

[0010] Determine the discrimination condition that the data to be analyzed meets based on the spectral characteristics; the discrimination condition includes a first mode, a second mode, and a third mode;

[0011] If the data to be analyzed conforms to the first pattern, the data is determined to be of the first data type, and the data to be analyzed is classified through the first data classification path;

[0012] If the data to be analyzed conforms to the second mode, the data is determined to be of the second data type, and the data to be analyzed is classified through a second data classification path;

[0013] If the data to be analyzed conforms to the third pattern, the data is determined to be of a third data type, and the data to be analyzed is classified through a third data classification path;

[0014] 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, the classification results of the data to be analyzed are output, and the identified real data is separated from the interference data.

[0015] Preferably, the multiple sensors include: ultra-high frequency built-in sensors, ultra-high frequency external sensors, noise sensors and ultrasonic sensors;

[0016] The process of performing preliminary elimination processing on the interference data is as follows:

[0017] Performing amplitude threshold filtering on the raw data to eliminate signals below a background noise baseline threshold and above a device saturation response threshold;

[0018] Perform short-term energy detection on the raw data based on the sampling time window to identify and eliminate sudden interference pulses with poor energy distribution continuity and abnormally short signal duration;

[0019] In combination with the wavelet transform method, the original data is band-pass filtered to retain only the effective frequency band components within the preset working bandwidth range to obtain the data to be analyzed.

[0020] Preferably, generating a first characteristic spectrum and a second characteristic spectrum based on the data to be analyzed, and extracting spectrum features of the first characteristic spectrum and the second characteristic spectrum, specifically includes:

[0021] Performing phase alignment and full-cycle segmentation processing on the data to be analyzed to generate the first characteristic spectrum;

[0022] Performing time-domain pulse sequence tracking processing on the data to be analyzed, and generating the second characteristic spectrum based on amplitude and time interval;

[0023] The spectrum features are extracted based on the first characteristic spectrum and the second characteristic spectrum, including spectrum structure features, trend disturbance features and spatial delay features; the spectrum structure features include graphic texture, symmetry and clustering degree; the trend disturbance features include peak drift and periodic anomaly; the spatial delay features include signal arrival delay differences of multiple sensors.

[0024] Preferably, the specific steps of determining whether the data to be analyzed conforms to the first mode are:

[0025] Calculating the point cloud distribution concentration of the spectral structural features; evaluating the symmetry index of the spectral features around a predetermined symmetry center; and obtaining a periodicity intensity index by performing autocorrelation analysis on the data to be analyzed;

[0026] If the point cloud distribution concentration is higher than a preset concentration threshold, the symmetry index is higher than a preset symmetry threshold, and the periodicity intensity index is higher than a preset periodicity threshold, it is determined to be a first data type, and the data to be analyzed is classified through a first data classification path;

[0027] The first data classification path uses a CNN model based on atlas texture features to classify the data to be analyzed, and outputs a probability label and a first confidence level of the first data type.

[0028] Preferably, the specific steps of determining whether the data to be analyzed conforms to the second mode are:

[0029] Calculating the statistical dispersion parameter of the signal trend disturbance feature; detecting the amplitude sequence of the data to be analyzed and calculating the amplitude difference of adjacent pulses; analyzing the pulse repetition time interval and instantaneous frequency of the data to be analyzed and calculating the instability index;

[0030] If the statistical dispersion parameter is higher than a preset disturbance 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 to be a second data type, and the data to be analyzed is classified through a second data classification path;

[0031] The second data classification path uses a time series feature learning model to analyze the data to be analyzed to obtain a classification label and a second confidence level of the second data type.

[0032] Preferably, the specific steps of determining whether the data to be analyzed conforms to the second mode are:

[0033] Calculating the structural order parameter of the spectrum structure feature; obtaining the spatial positioning result according to the spatial delay feature, and calculating the confidence of the spatial positioning result;

[0034] If the structural orderliness parameter is lower than a preset orderliness threshold and the confidence is lower than a preset delay confidence threshold, it is determined to be a third data type, and the data to be analyzed is classified through a third data classification path;

[0035] The third data classification path uses interference feature matching and pattern exclusion algorithms to distinguish the data to be analyzed, and obtains a classification label and a third confidence level of the third data type.

[0036] 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:

[0037] A weighted voting mechanism is used 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;

[0038] Output the fused classification results of the data to be analyzed, and generate confidence level labels for the classification results; separate the identified real data from the interference data based on the classification results and the confidence level labels.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. During the feature extraction phase, the present invention generates a first characteristic spectrum (discharge phase distribution diagram) and a second characteristic spectrum (pulse sequence phase distribution diagram) from the data to be analyzed. It then further extracts the spectrum structure features, trend perturbation features, and spatial delay features, constructing a multivariate high-dimensional feature set. This feature not only enhances the distinguishability between the three subsequent discrimination modes (periodic symmetry, trend perturbation, and spatial interference), but also significantly improves the accuracy of the mapping 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-pathway model.

[0041] 2. This invention achieves high-precision intelligent classification of partial discharge signals in GIS equipment by constructing a first data classification path (a CNN model based on spectral texture features), a second data classification path (a time series feature learning model), and a third data classification path (an interference matching and pattern elimination algorithm), each corresponding to different types of ultra-high frequency partial discharge signals. Combining the raw data collected synchronously by multiple sensors to form spectral features and trend disturbance characteristics, this method significantly reduces the probability of false positives and missed positives, improving the system's overall recognition accuracy and automation level.

[0042] 3. During the classification result output phase, the present invention integrates the independent classification results from the three data classification paths by introducing a confidence-weighted fusion mechanism. This fusion strategy leverages the strengths of each path for different data types, enhancing the robustness and stability of the overall recognition judgment. This mechanism not only improves the consistency of the final classification results in complex scenarios, but also provides a reliable classification foundation for the precise separation of interference signals from true signals, significantly reducing the misclassification rate and improving the system's signal restoration and anomaly alarm capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A schematic flow chart of a method for separating ultra-high frequency partial discharge signals from a GIS based on pattern recognition according to an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of a process for determining whether the data to be analyzed conforms to the first mode provided by an embodiment of the present invention;

[0045] Figure 3 A schematic diagram of a process for determining whether the data to be analyzed conforms to the second mode provided by an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of a process for determining whether the data to be analyzed conforms to the third mode provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] Gas-insulated switchgear (GIS) is a critical high-voltage component in power systems, and its operating status significantly impacts the safety and stability of the transmission and distribution system. Insulation defects can occur in GIS equipment due to factors such as manufacturing, installation, and the operating environment, leading to partial discharge (PD). PD is a typical manifestation of early insulation degradation in GIS and can provide early warning signals before a fault occurs. Therefore, early identification and interference removal of GIS PD signals have become key technical challenges in power system status monitoring and intelligent diagnosis.

[0049] This paper proposes a pattern recognition-based method for separating GIS ultra-high frequency partial discharge (PD) signals. This method addresses identification challenges such as ambiguity in PD signal classification, minimal distribution differences in high-dimensional space, and severe feature drift under noise interference, thereby improving the accuracy of distinguishing true PD signals from external interference signals. To demonstrate the effectiveness of this method in improving the accuracy of distinguishing true PD signals from external interference signals, the following two examples will illustrate its effectiveness.

[0050] Example 1

[0051] In the embodiments of the present application, the method proposed in the present invention is used to solve the identification problems such as the ambiguity in the category distinction of partial discharge signals, the small distribution difference in high-dimensional space, and the severe feature drift under noise interference, and to improve the accuracy of distinguishing between real partial discharge signals and external interference signals. The process is described in detail. Figure 1The specific flow chart of the method proposed in the present invention includes: using multiple sensors to synchronously collect raw data, and preliminarily eliminating interference data to obtain data to be analyzed; generating a first characteristic spectrum and a second characteristic spectrum based on the data to be analyzed, and extracting spectrum features; judging whether the data to be analyzed meets the discrimination conditions according to the spectrum features, including the first mode, the second mode and the third mode; if the data to be analyzed meets the corresponding discrimination conditions, the data to be analyzed is classified through the corresponding data classification path; the classification results obtained through each data classification path are integrated based on the confidence level, the classification results of the data to be analyzed are output, and the identified real data is separated from the interference data. The following is based on Figure 1 The following content is described:

[0052] Using multiple sensors to synchronously collect raw data and preliminarily eliminate interference data to obtain data to be analyzed; the multiple sensors include: ultra-high frequency built-in sensors, ultra-high frequency external sensors, noise sensors and ultrasonic sensors;

[0053] The process of performing preliminary elimination processing on the interference data is as follows:

[0054] Performing amplitude threshold filtering on the raw data to eliminate signals below a background noise baseline threshold and above a device saturation response threshold;

[0055] Perform short-term energy detection on the raw data based on the sampling time window to identify and eliminate sudden interference pulses with poor energy distribution continuity and abnormally short signal duration;

[0056] In combination with the wavelet transform method, the original data is band-pass filtered to retain only the effective frequency band components within the preset working bandwidth range to obtain the data to be analyzed.

[0057] Specifically, the original data is an original partial discharge signal;

[0058] The background noise baseline threshold is dynamically set according to the on-site environmental noise level, for example, it is set to 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 sensor and data acquisition system used to avoid signal clipping and distortion.

[0059] In the short-time energy detection, the sampling time window is set to several power frequency cycles, for example, 5 power frequency cycles. The signal energy in the window is calculated and compared with the preset energy threshold, and the judgment is made in combination with the pulse duration (for example, a pulse less than 10ns is regarded as a burst interference).

[0060] In the wavelet transform method, the maximum overlap discrete wavelet transform is adopted, and a wavelet basis such as db4 or sym8 is selected to decompose the original signal into 5 to 8 layers. According to the UHF frequency band characteristics (within the range of 0.3 GHz to 3 GHz) of typical partial discharge signals in GIS equipment, the wavelet coefficients of the corresponding frequency band are selected for reconstruction to achieve bandpass filtering; the amplitude normalization adopts the maximum and minimum normalization method to map the signal amplitude to the interval [0, 1].

[0061] Through the coordinated collection of the above-mentioned multiple sensors and targeted preliminary interference data elimination processing, it is possible to effectively suppress electromagnetic interference from space, pulse interference caused by poor contact, and narrowband interference from some non-local discharges. At the same time, the effective local discharge signal frequency band components are retained and standardized, which significantly improves the signal-to-noise ratio and standardization of the original data, lays a solid data foundation for subsequent accurate feature extraction and pattern recognition, and reduces the complexity of subsequent processing.

[0062] Preferably, generating a first characteristic spectrum and a second characteristic spectrum based on the data to be analyzed, and extracting spectrum features of the first characteristic spectrum and the second characteristic spectrum; specifically comprising:

[0063] Performing phase alignment and full-cycle segmentation processing on the data to be analyzed to generate the first characteristic spectrum;

[0064] Performing time-domain pulse sequence tracking processing on the data to be analyzed, and generating the second characteristic spectrum based on amplitude and time interval;

[0065] The spectrum features are extracted based on the first characteristic spectrum and the second characteristic spectrum, including spectrum structure features, trend disturbance features and spatial delay features; the spectrum structure features include graphic texture, symmetry and clustering degree; the trend disturbance features include peak drift and periodic anomaly; the spatial delay features include signal arrival delay differences of multiple sensors.

[0066] Specifically, when generating the first characteristic spectrum (discharge phase distribution diagram), a phase reference is first obtained through the synchronously collected power frequency voltage signal. Each UHF pulse signal is statistically analyzed based on its power frequency phase (0-360 degrees) and amplitude. The data of multiple power frequency cycles (for example, 100-500 cycles) are accumulated to form a two-dimensional statistical histogram.

[0067] When generating the second characteristic spectrum (pulse sequence phase distribution diagram), record the time when the continuous pulses occur and pulse amplitude , in a two-dimensional plane The temporal correlation of the pulse train is shown in the form of (representing the time interval between the pulse charge and the previous pulse), where is a modular operation, is the power frequency period.

[0068] Extracting spectral structural features: The gray-level co-occurrence matrix is ​​used to calculate texture parameters such as energy, contrast, and correlation. Symmetry is determined by calculating the symmetry index of the first characteristic spectrum (discharge phase distribution diagram) about the 90° and 270° phase axes. Clustering is determined by using the DBSCAN clustering algorithm to identify the number and density of major discharge clusters.

[0069] Extract trend disturbance features: Peak drift monitoring measures the change in the mean amplitude of the main discharge clusters in different time periods (such as every minute or every hour) in the first characteristic spectrum (discharge phase distribution diagram); periodic anomaly analysis measures the stability of the pulse repetition frequency and detects whether there are abnormal pulsations in non-power frequency periods;

[0070] Extract spatial delay features: Using the same discharge pulse signal received by at least three UHF sensors, the generalized cross-correlation algorithm is used to calculate the precise time difference between the signals reaching different sensors for subsequent positioning and interference elimination.

[0071] By generating two key characteristic spectra—the discharge phase distribution diagram and the pulse sequence phase distribution diagram—and extracting multi-dimensional features such as structure, trend, and spatial delay, the system can comprehensively characterize the essential properties of the signal from multiple perspectives, including discharge phase distribution, pulse time series correlation, signal amplitude stability, periodicity, and spatial propagation characteristics. This multi-dimensional feature combination significantly enhances the ability to distinguish different types of partial discharges and interference signals, compensating for the lack of distinguishing power of a single feature under complex working conditions and providing a foundation for subsequent pattern determination of the analyzed data through spectral features.

[0072] Table 1 is a comparison table of the classification accuracy of the introduction of bispectral graph and feature extraction and other methods, among which the classification accuracy is get.

[0073] Table 1 Comparison of classification accuracy using bispectral graphs and feature extraction

[0074]

[0075] Preferably, the discrimination conditions that the data to be analyzed meet are determined according to the spectral characteristics; the discrimination conditions include a first mode, a second mode, and a third mode;

[0076] If the data to be analyzed conforms to the first pattern, the data is determined to be of the first data type, and the data to be analyzed is classified through the first data classification path;

[0077] Determine whether the data to be analyzed conforms to the first mode, referring to Figure 2 , the specific steps are:

[0078] Calculating the point cloud distribution concentration of the spectral structural feature; evaluating the symmetry index of the spectral feature around a predetermined symmetry axis; and obtaining a periodicity strength index by performing autocorrelation analysis on the data to be analyzed;

[0079] If the point cloud distribution concentration is higher than a preset concentration threshold, the symmetry index is higher than a preset symmetry threshold, and the periodicity intensity index is higher than a preset periodicity threshold, it is determined to be a first data type, and the data to be analyzed is classified through a first data classification path;

[0080] The first data classification path uses a CNN model based on atlas texture features to classify the data to be analyzed, and outputs a probability label and a first confidence level of the first data type.

[0081] Specifically, the concentration of the point cloud distribution is quantified as the energy proportion of the high-density area in the discharge phase distribution map, and the preset concentration threshold is 0.6; the symmetry index is the inverse of the difference in pixel values ​​on both sides of the specific symmetry axis of the spectrum, 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 is a significant peak at the power frequency and its multiples. The preset periodic threshold is set according to the statistical characteristics of typical metal particle discharge.

[0082] The first data type is a metal particle discharge signal. In the first data classification path, the CNN model includes three convolutional layers and two fully connected layers, with a 3×3 convolution kernel size, a Reinforced Lu (ReLU) activation function, and max pooling in the pooling layer. The input is a discharge phase distribution map or a pulse train phase distribution map, and the output layer uses a Softmax function to output a probability value containing a metal particle discharge signal. The model is pre-trained on a large number of atlas samples containing known metal particle discharges and other types of discharges / interference. The first confidence level is the probability value of the corresponding category output by the Softmax function.

[0083] By quantifying the concentration, symmetry, and periodicity of the spectral structural features, and combining them with a CNN model optimized specifically for spectral texture features for refined classification, the system can efficiently and accurately identify metallic particle-type discharge signals with typical "rabbit-ear" or symmetrically clustered morphologies. This two-stage judgment and classification mechanism first quickly screens out highly suspected signals, then confirms them through a deep learning model. This significantly improves the accuracy and robustness of metallic particle-type discharge identification, effectively avoiding confusion with other types of signals.

[0084] Preferably, if the data to be analyzed conforms to the second mode, it is determined to be a second data type, and the data to be analyzed is classified through a second data classification path;

[0085] Determine whether the data to be analyzed conforms to the second mode, referring to Figure 3 , the specific steps are:

[0086] Calculating the statistical dispersion parameter of the signal trend disturbance feature; detecting the amplitude sequence of the data to be analyzed and calculating the amplitude difference of adjacent pulses; analyzing the pulse repetition time interval and instantaneous frequency of the data to be analyzed and calculating the instability index;

[0087] If the statistical dispersion parameter is higher than a preset disturbance 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 to be a second data type, and the data to be analyzed is classified through a second data classification path;

[0088] The second data classification path uses a time series feature learning model to analyze the data to be analyzed to obtain a classification label and a second confidence level of the second data type.

[0089] Specifically, the statistical dispersion parameter is obtained by calculating the standard deviation of the pulse amplitude sequence, and the preset disturbance 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 amplitude difference of adjacent pulses is calculated It is found that the preset amplitude change threshold is 20% of the maximum amplitude; the instability index is obtained by calculating the coefficient of variation (such as standard deviation) of the pulse time interval sequence, and the preset sudden increase frequency threshold is 0.4.

[0090] 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 adopts an LSTM model including 2 layers of LSTM (128 units per layer) and an output fully connected layer, and the input is the extracted pulse amplitude sequence, time interval sequence, etc.; the sequence data with a length of 100 continuous pulses is learned, 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.

[0091] By comprehensively evaluating the perturbations of signal trends, the severity of amplitude changes, and the stability of pulse frequency, and leveraging a time series feature learning model to deeply explore the dynamic evolution of the signal over time, this method effectively captures the continuous clustering characteristics of internal corona discharges and / or the sudden, high-energy characteristics of insulation breakdown. This improves the sensitivity and accuracy of identifying these rapidly developing and highly hazardous internal discharges, providing a reliable basis for timely early warning.

[0092] Preferably, if the data to be analyzed conforms to the third pattern, it is determined to be a third data type, and the data to be analyzed is classified through a third data classification path;

[0093] Determine whether the data to be analyzed conforms to the third mode, refer to Figure 4 , the specific steps are:

[0094] Calculating the structural order parameter of the spectrum structure feature; obtaining the spatial positioning result according to the spatial delay feature, and calculating the confidence of the spatial positioning result;

[0095] If the structural orderliness parameter is lower than a preset orderliness threshold and the confidence is lower than a preset delay confidence threshold, it is determined to be a third data type, and the data to be analyzed is classified through a third data classification path;

[0096] The third data classification path uses interference feature matching and pattern exclusion algorithms to distinguish the data to be analyzed, and obtains a classification label and a third confidence level of the third data type.

[0097] Specifically, the structural order parameter is obtained by calculating the entropy of the discharge phase distribution map. The higher the entropy value, the lower the order. The preset order threshold is set based on 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.

[0098] The third data type is interference and / or suspended discharge type signal; in the third data classification path, the interference feature matching first compares the features of the data to be analyzed (such as specific peaks of the spectrum, morphological parameters of the discharge phase distribution diagram) with a pre-stored interference feature database (such as switching noise, characteristic signatures of radio interference); the pattern exclusion algorithm judges the signal as interference and / or suspended discharge type signal when it does not meet the strong features of the first and second patterns and the matching degree with the known interference feature library is not high, but the spectrum is disordered and the positioning is vague, and outputs the classification label and the third confidence based on the exclusion rule.

[0099] Table 2 is a comparative data table of different methods in terms of partial discharge classification accuracy and stability, among which the classification accuracy is Get; the false positive rate is obtained by get.

[0100] Table 2 Comparison of different methods in partial discharge classification accuracy and stability

[0101]

[0102] By evaluating the disorder of the spectra and the ambiguity of multi-sensor positioning results, combined with interference signature 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 suspended discharge signals. This mechanism avoids misidentifying interference as real discharges and provides a reasonable outlet for difficult-to-classify signals, improving the overall system's anti-interference capabilities and comprehensive diagnostic capabilities.

[0103] 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, the classification results of the data to be analyzed are output, and the identified real data is separated from the interference data; the specific steps are:

[0104] A weighted voting mechanism is used 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;

[0105] Output the fused classification results of the data to be analyzed, and generate confidence level labels for the classification results; separate the identified real data from the interference data based on the classification results and the confidence level labels.

[0106] Specifically, in the weighted voting mechanism, the classification result output by each path is and its confidence If a path is activated (i.e. the judgment condition is met), its weight ( (where is the index of all activated paths), and the weight of inactivated paths is 0. The final classification result is the category with the highest weighted confidence. For example, if the first path is judged as a valid PD (confidence 0.82), the second path is judged as an interference signal (confidence 0.65), and the third path is judged as a valid PD (confidence 0.72), then a weighted voting mechanism will determine the final category as a valid PD (confidence 0.78). If only one path is activated, its result is used directly. Confidence levels are set as high for >0.85, medium for 0.6-0.85, and low for <0.6. When separating real data from interference data, those classified as metal particles, internal corona, or insulation breakdown with high or medium confidence are considered real PD signals, while the rest can be labeled as interference signals.

[0107] The confidence-weighted fusion mechanism of multi-path classification results combines the advantages of different identification strategies, effectively improving the accuracy and reliability of final judgments on complex and ambiguous signals. The introduction of confidence level labels provides decision-making references for operators, facilitating the distinction between high-risk real discharges and general interference. This enables effective hierarchical management and precise separation of monitoring data, enhancing the practical value and automation level of the monitoring system.

[0108] The proposed pattern recognition-based method for separating ultra-high frequency partial discharge (PD) signals from GIS (GIS) is a systematic approach that addresses the inaccurate identification and susceptibility to interference of traditional methods in complex electromagnetic environments. This method not only accurately distinguishes different types of partial discharge (such as metal particles, internal corona / breakdown, and suspended discharge), but also effectively identifies and eliminates various interference signals. This significantly improves the accuracy and intelligence of GIS equipment condition monitoring, providing strong technical support for ensuring the safe and stable operation of power systems.

[0109] Example 2

[0110] In Example 1, the method proposed by the present invention successfully addressed identification challenges such as ambiguity in PD signal classification, minimal distribution differences in high-dimensional space, and severe feature drift under noise interference, thereby improving the accuracy of distinguishing true PD signals from external interference signals. To further verify the effectiveness of the present invention, the present example also distinguished and identified PD signals generated by another GIS device.

[0111] Using multiple sensors to synchronously collect raw data and preliminarily eliminate interference data to obtain data to be analyzed; the multiple sensors include: ultra-high frequency built-in sensors, ultra-high frequency external sensors, noise sensors and ultrasonic sensors;

[0112] The process of performing preliminary elimination processing on the interference data is as follows:

[0113] Performing amplitude threshold filtering on the raw data to eliminate signals below a background noise baseline threshold and above a device saturation response threshold;

[0114] Perform short-term energy detection on the raw data based on the sampling time window to identify and eliminate sudden interference pulses with poor energy distribution continuity and abnormally short signal duration;

[0115] In combination with the wavelet transform method, the original data is band-pass filtered to retain only the effective frequency band components within the preset working bandwidth range to obtain the data to be analyzed.

[0116] Preferably, generating a first characteristic spectrum and a second characteristic spectrum based on the data to be analyzed, and extracting spectrum features of the first characteristic spectrum and the second characteristic spectrum; specifically comprising:

[0117] Performing phase alignment and full-cycle segmentation processing on the data to be analyzed to generate the first characteristic spectrum;

[0118] Performing time-domain pulse sequence tracking processing on the data to be analyzed, and generating the second characteristic spectrum based on amplitude and time interval;

[0119] The spectrum features are extracted based on the first characteristic spectrum and the second characteristic spectrum, including spectrum structure features, trend disturbance features and spatial delay features; the spectrum structure features include graphic texture, symmetry and clustering degree; the trend disturbance features include peak drift and periodic anomaly; the spatial delay features include signal arrival delay differences of multiple sensors.

[0120] Preferably, the discrimination conditions that the data to be analyzed meet are determined according to the spectral characteristics; the discrimination conditions include a first mode, a second mode, and a third mode;

[0121] If the data to be analyzed conforms to the first pattern, it is determined to be a first data type, and the data to be analyzed is classified through a first data classification path; determining that the data to be analyzed conforms to the first pattern, specifically the steps are:

[0122] Calculating the point cloud distribution concentration of the spectral structural feature; evaluating the symmetry index of the spectral feature around a predetermined symmetry center; and obtaining a periodicity intensity index by performing autocorrelation analysis on the data to be analyzed;

[0123] If the point cloud distribution concentration is higher than a preset concentration threshold, the symmetry index is higher than a preset symmetry threshold, and the periodicity intensity index is higher than a preset periodicity threshold, it is determined to be a first data type, and the data to be analyzed is classified through a first data classification path;

[0124] The first data classification path uses a CNN model based on atlas texture features to classify the data to be analyzed, and outputs a probability label and a first confidence level of the first data type.

[0125] Preferably, if the data to be analyzed conforms to the second mode, it is determined to be a second data type, and the data to be analyzed is classified through a second data classification path;

[0126] Determining whether the data to be analyzed conforms to the second mode includes the following specific steps:

[0127] Calculating the statistical dispersion parameter of the signal trend disturbance feature; detecting the amplitude sequence of the data to be analyzed and calculating the amplitude difference of adjacent pulses; analyzing the pulse repetition time interval and instantaneous frequency of the data to be analyzed and calculating the instability index;

[0128] If the statistical dispersion parameter is higher than a preset disturbance 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 to be a second data type, and the data to be analyzed is classified through a second data classification path;

[0129] The second data classification path uses a time series feature learning model to analyze the data to be analyzed to obtain a classification label and a second confidence level of the second data type.

[0130] Preferably, if the data to be analyzed conforms to the third pattern, it is determined to be a third data type, and the data to be analyzed is classified through a third data classification path;

[0131] Determining whether the data to be analyzed conforms to the third mode, specifically the steps are:

[0132] Calculating the structural order parameter of the spectrum structure feature; obtaining the spatial positioning result according to the spatial delay feature, and calculating the confidence of the spatial positioning result;

[0133] If the structural orderliness parameter is lower than a preset orderliness threshold and the confidence is lower than a preset delay confidence threshold, it is determined to be a third data type, and the data to be analyzed is classified through a third data classification path;

[0134] The third data classification path uses interference feature matching and pattern exclusion algorithms to distinguish the data to be analyzed, and obtains a classification label and a third confidence level of the third data type.

[0135] 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, the classification results of the data to be analyzed are output, and the identified real data is separated from the interference data; the specific steps are:

[0136] A weighted voting mechanism is used 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;

[0137] Output the fused classification results of the data to be analyzed, and generate confidence level labels for the classification results; separate the identified real data from the interference data based on the classification results and the confidence level labels.

[0138] Table 3 shows the impact of using the fusion mechanism on the final classification results and interference separation effect, among which, , represents the ratio of the three classification paths outputting different classification results under the same input; , which indicates the system's ability to fully preserve the real signal during processing.

[0139] Table 3 The impact of fusion mechanism on the final classification results and interference separation effect

[0140]

[0141] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A GIS UHF partial discharge signal separation method based on pattern recognition is characterized by: include: Collect and preprocess raw data to obtain data to be analyzed; The preprocessing is to preliminarily remove interference data; Generate a first characteristic spectrum and a second characteristic spectrum based on the data to be analyzed, and extract spectrum features; perform phase alignment and full cycle segmentation processing on the data to be analyzed to generate a first characteristic spectrum; Performing time domain pulse sequence tracking processing on the data to be analyzed, and generating a second characteristic spectrum based on the amplitude and time interval; Spectral features include spectral structure features, trend disturbance features and spatial delay features; Determine whether the data to be analyzed meets the discrimination conditions according to the spectral characteristics, including the first mode, the second mode and the third mode; If the data to be analyzed conforms to the first pattern, it is determined to be a first data type and classified through the first data classification path; the first pattern includes that the point cloud distribution concentration is higher than a preset concentration threshold, the symmetry index is higher than a preset symmetry threshold, and the periodicity intensity index is higher than a preset periodicity threshold; If the data to be analyzed conforms to the second pattern, it is determined to be of the second data type and classified through the second data classification path; the second pattern includes the statistical dispersion parameter being higher than the preset disturbance threshold, the difference in amplitudes of adjacent pulses being higher than the preset amplitude change threshold, and the instability index being higher than the preset sudden increase frequency threshold; If the data to be analyzed conforms to the third pattern, it is determined to be a third data type and the data to be analyzed is classified through a third data classification path; the third pattern includes a structural orderliness parameter being lower than a preset orderliness threshold and a confidence level being lower than a preset delay confidence level threshold; According to the classification results of the first data classification path, the second data classification path and the third data classification path, the classification results of the data to be analyzed are output, and the identified real data is separated from the interference data.

2. The GIS UHF partial discharge signal separation method based on pattern recognition according to claim 1 is characterized in that: The raw data is collected synchronously by multiple sensors, including: an ultra-high frequency built-in sensor, an ultra-high frequency external sensor, a noise sensor, and an ultrasonic sensor; The specific process of the preprocessing is as follows: 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 eliminate burst interference pulses with poor energy distribution continuity and abnormally short signal duration; combining the wavelet transform method to bandpass filter the raw data to retain only the effective frequency band components within the preset working bandwidth to obtain the data to be analyzed.

3. The GIS UHF partial discharge signal separation method based on pattern recognition according to claim 1 is characterized in that: Generating a first characteristic spectrum and a second characteristic spectrum based on the data to be analyzed, and extracting spectrum features, specifically includes: The spectrum features are extracted based on the first characteristic spectrum and the second characteristic spectrum, including spectrum structure features, trend disturbance features and spatial delay features; the spectrum structure features include graphic texture, symmetry and clustering degree; the trend disturbance features include peak drift and periodic anomaly; the spatial delay features include signal arrival delay differences of multiple sensors.

4. The GIS UHF partial discharge signal separation method based on pattern recognition according to claim 1 is characterized in that: Determining whether the data to be analyzed conforms to the first mode includes the following specific steps: Calculating the point cloud distribution concentration of the spectral structural features; evaluating the symmetry index of the spectral features around a predetermined symmetry center; and obtaining a periodicity intensity index by performing autocorrelation analysis on the data to be analyzed; If it is determined to be the first data type, classifying the data to be analyzed through a first data classification path; The first data classification path uses a CNN model based on atlas texture features to classify the data to be analyzed, and outputs a probability label and a first confidence level of the first data type.

5. The GIS UHF partial discharge signal separation method based on pattern recognition according to claim 1 is characterized in that: Determining whether the data to be analyzed conforms to the second mode includes the following specific steps: Calculating the statistical dispersion parameter of the trend disturbance feature; detecting the amplitude sequence of the data to be analyzed and calculating the amplitude difference of adjacent pulses; Analyzing the pulse repetition time interval and instantaneous frequency of the data to be analyzed, and calculating an instability index; If it is determined to be the second data type, classifying the data to be analyzed through a 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 a classification label and a second confidence level of the second data type.

6. The GIS UHF partial discharge signal separation method based on pattern recognition according to claim 1 is characterized in that: Determining whether the data to be analyzed conforms to the third mode, specifically the steps are: Calculating the structural order parameter of the spectrum structure feature; obtaining the spatial positioning result according to the spatial delay feature, and calculating the confidence of the spatial positioning result; If it is determined to be a third data type, classifying the data to be analyzed through a third data classification path; The third data classification path uses interference feature matching and pattern exclusion algorithms to distinguish the data to be analyzed, and obtains a classification label and a third confidence level of the third data type.

7. The GIS UHF partial discharge signal separation method based on pattern recognition according to claim 1 is characterized in that: Outputting the classification results of the data to be analyzed according to the classification results of the first data classification path, the second data classification path, and the third data classification path. The specific steps are: A weighted voting mechanism is used 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; Outputting the fused classification results of the data to be analyzed, and generating a confidence level label for the classification results; The identified real data is separated from the interference data based on the classification results and confidence level labels.

Images