Search

CN-121049664-B - Partial discharge identification and positioning method based on multi-mode calibration and blind source separation

CN121049664BCN 121049664 BCN121049664 BCN 121049664BCN-121049664-B

Abstract

The invention discloses a partial discharge identification and positioning method based on multi-mode calibration and blind source separation, which comprises the following steps of 1, distributing a high-frequency current sensor along the grounding flat iron of a transformer iron core or a clamp, 2, collecting a mixed pulse signal, performing frequency conversion, outputting multi-mode characteristics, 3, using the blind source separation to output a plurality of independent pulse source signals and signal parameters for the mixed pulse signal, 4, periodically injecting a reference pulse signal into the flat iron, estimating and adaptively updating the signal propagation speed along the flat iron based on the arrival time difference, 5, obtaining the multi-mode characteristics and the signal parameters, inputting the multi-mode characteristics and the signal parameters into a deep learning classifier, outputting classification confidence, comprehensively judging pulse signal interference sources, and realizing internal and external discharge identification and positioning along the flat iron.

Inventors

  • ZHAO HONGYI
  • SHEN DAOYI
  • HU YONG
  • GAN YUANFENG
  • TIAN GUANGLIANG
  • Qian Dazhao

Assignees

  • 上海格鲁布科技有限公司

Dates

Publication Date
20260508
Application Date
20250829

Claims (10)

  1. 1. The partial discharge identification and positioning method based on multi-mode calibration and blind source separation is characterized by comprising the following steps of: Step 1, arranging a high-frequency current sensor, connecting the output end of the high-frequency current sensor to a high-speed data acquisition device, and setting sampling channel clocks for synchronization, wherein the high-frequency current sensor is connected to the grounding flat iron of a transformer core or the grounding flat iron of a clamping piece along the line; Step 2, collecting a mixed pulse signal, wherein the mixed pulse signal comprises a high-frequency sensor signal and at least one signal of other sensors, performing time-frequency conversion on the mixed pulse signal, and outputting multi-mode characteristics, wherein the multi-mode characteristics comprise a time sequence original waveform, a time-frequency spectrogram, an ultrasonic envelope signal, a PRPD image, an energy ratio, a frequency spectrum centroid and instantaneous phase information; step 3, a convolution mixed model of blind source separation is customized, a plurality of independent pulse source signals are output for mixed pulse signals by using blind source separation, and signal parameters of the independent pulse source signals are calculated, wherein the signal parameters comprise waveform similarity, pulse polarity, arrival time difference, energy and frequency spectrum characteristics; Step 4, periodically injecting a reference pulse signal into the flat iron, and estimating and adaptively updating the signal propagation speed along the flat iron based on the arrival time difference of the reference pulse; step 5, acquiring the multi-modal characteristics and the signal parameters, inputting the multi-modal characteristics and the signal parameters into a deep learning classifier, outputting classification confidence, and comprehensively judging pulse signal interference sources; The multi-mode characteristics are input into a pre-trained CNN-transducer mixed model, the CNN-transducer mixed model comprises a CNN module, a transducer module and a fusion module, the CNN module extracts image characteristics, the transducer module processes time sequence waveforms, the pre-training is based on screened actual and simulated partial discharge high frequency, ultrasonic and other data, the pre-training is input into the CNN-transducer mixed model, the recognition accuracy is output, the model parameters are adjusted based on the recognition preparation rate, and the training is completed after 98 percent of the parameters are reached; The CNN-transducer mixed model outputs three results of internal discharge, external interference and unknown event; The specific calculation process of the deep learning result is that when abnormal pulses on the grounding flat iron of the transformer are monitored, the deep learning classifier starts to work, firstly, an original pulse signal captured by a high-frequency current sensor is subjected to wavelet transformation, a time-frequency spectrogram is output, the phase-amplitude relation of the original pulse in a power frequency period is synchronously extracted, a PRPD image is output, an envelope waveform is extracted from a signal acquired by an ultrasonic sensor through Hilbert transformation, and interpolation is aligned to the same time axis; The energy ratio and spectrum centroid of the pulse are quantized into numerical values, normalized and converted into gray values, integrated into 256 multiplied by 256 pixel tensors of a four-channel, wherein the first channel is a time-frequency spectrogram, records frequency domain characteristics, the second channel is a PRPD image, records discharge phase rules, the third channel is an ultrasonic envelope, captures mechanical vibration association, the fourth channel is integrated into quantization indexes of energy and spectrum, and is used for completely describing physical characteristics of the pulse, The pixel tensor is input into a pre-trained CNN-transducer mixed model, a CNN module analyzes a time-frequency spectrogram and a space mode in a PRPD image through a three-layer convolution network, a small kernel of 3 multiplied by 3 is used for sliding scanning on the image by first-layer convolution, local details are captured, a residual error connection structure ensures that a deep network can continuously optimize feature extraction capacity to avoid detail loss; The method comprises the steps of outputting a CNN module and a transducer module to a fusion module, flattening and splicing 64X 128 dimensional space features extracted by the CNN module and 16X 64 dimensional time sequence features generated by the transducer module to form a joint feature vector exceeding 52 ten thousand dimensions, processing the joint feature vector through two layers of fully-connected neural networks, screening significant features through a ReLU activation function at the first layer, outputting three original score values at the second layer, and converting the scores into probability distribution by a Softmax function.
  2. 2. The method for identifying and positioning partial discharge based on multi-modal calibration and blind source separation according to claim 1, wherein the distance between the high-frequency current sensors is set to be 2-3 meters, the sampling rate of the high-speed data acquisition device is not less than 1GS/s, and the clock synchronization error between sampling channels is less than 1 nanosecond.
  3. 3. The method for identifying and positioning partial discharge based on multi-mode calibration and blind source separation according to claim 1 is characterized in that in the step 5, comprehensive judgment is carried out by carrying out evidence fusion on classification confidence and evidence based on polarity, time difference and similarity to judge that pulse signals come from external interference or internal discharge of a transformer, wherein the evidence fusion is carried out by detecting that pulses detected by 3 high-frequency current sensors are negative and accord with internal discharge characteristics, setting weight to be 0.3, matching arrival time difference, substituting adjacent sensor time difference DeltaT 12 =3.2ns,ΔT 23 =3.0 ns into a positioning equation error <5%, setting weight to be 0.4, waveform similarity, and setting separated pulse cross correlation coefficient ρ=0.91 >0.85, and weight to be 0.3; And judging that the internal discharge is generated, positioning along the flat iron vector based on the signal propagation speed and the arrival time difference, recording position information, and storing a judging result, an interpretable output and a safety time stamp.
  4. 4. The method for identifying and positioning partial discharge based on multi-modal calibration and blind source separation according to claim 1, wherein in the step 3, the blind source separation uses independent component analysis, sparse component analysis or decomposition method based on sparse representation, and the blind source separation is performed after the short-time energy segment is segmented preferentially.
  5. 5. The method for identifying and locating partial discharge based on multi-modal calibration and blind source separation according to claim 1, wherein the deep learning classifier comprises an anomaly detection module, an online incremental learning module and a pseudo tag self-learning module.
  6. 6. The method for identifying and locating partial discharge based on multi-modal calibration and blind source separation as claimed in claim 5, further comprising an anomaly detection module, wherein the anomaly detection module uploads an anomaly event identified based on historical event detection, and performs manual review and training set expansion, and the historical event is specifically an unsupervised detection model constructed based on historical characteristics and an event which is significantly different from a historical sample.
  7. 7. A positioning system based on the multi-modal calibration and blind source separation based partial discharge identification and positioning method as claimed in any one of claims 1-6, comprising: The sensor module is used for collecting and outputting mixed pulse signals and comprises a plurality of high-frequency current sensors and at least one other type of sensor; The high-speed data acquisition module receives the mixed pulse signals and performs analog-to-digital conversion, the sampling rate is more than or equal to 1GS/s, and the clock synchronization error is less than or equal to 1ns; the processing judgment module is used for processing the mixed pulse signals to generate classification confidence, wherein the processing comprises blind source separation, multi-mode feature extraction, self-adaptive propagation speed estimation, deep learning classification and evidence fusion judgment; The self-checking calibration module is used for injecting a reference pulse signal into the flat iron, calibrating the time delay and the gain of the channel and outputting a calibration result; The anomaly detection module is used for uploading the anomaly event detected and identified based on the historical event, and carrying out manual review and training set expansion; the storage module is used for storing process data and output results, wherein the process data comprises mixed pulse signals, independent pulse source signals and multi-mode characteristics, and the output results comprise judgment results, interpretable output and safety time stamps.
  8. 8. The positioning system of claim 7 wherein the number of high frequency current sensors is greater than or equal to 3 and is disposed along the flat iron array.
  9. 9. The positioning system of claim 7 wherein the processing determination module processes the mixed pulse signal by beamforming to enhance the signal and suppress out-of-space interference.
  10. 10. A computer readable storage medium, wherein at least one program is stored in the computer readable storage medium, the at least one program being loaded and executed by a processor to implement the partial discharge identification and localization method of any one of claims 1 to 6.

Description

Partial discharge identification and positioning method based on multi-mode calibration and blind source separation Technical Field The invention relates to the technical field of partial discharge signal identification, in particular to a partial discharge identification and positioning method based on multi-mode calibration and blind source separation. Background Partial discharge is an important early indicator of transformer insulation defects. For 500kV and above, common online detection means include high frequency current sensors (HFCT) and ultrasonic sensors. However, in actual operation, the iron core grounding flat iron or the clamp grounding flat iron of the transformer is connected with the power grid grounding grid, and discharge or switching impact generated by other devices in the grid can be coupled to the grounding point, so that a large amount of external interference and coupling noise exist in the detection signal. The conventional rule judging method based on the high-frequency current sensors at two ends can be used for basically distinguishing the inside and the outside and positioning along the flat iron under simple working conditions, for example, based on the judgment of arrival time difference, polarity and waveform similarity, but the performance is obviously reduced and false alarm or missing alarm is easy to occur under the conditions of complex coupling, multi-pulse superposition, low signal to noise ratio, influence of on-site structure and temperature change on propagation speed and the like. Furthermore, conventional schemes typically rely on fixed signal propagation velocity estimates, lack in-situ adaptive calibration means, and are poorly adapted to overlapping pulses or new sources of interference. Disclosure of Invention The invention aims to provide a partial discharge identification and positioning method based on multi-mode calibration and blind source separation, which is used for solving the problems that the traditional scheme generally depends on fixed signal propagation speed estimation, lacks a field self-adaptive calibration means and has insufficient adaptability to overlapping pulses or novel interference sources. The invention solves the technical problems by adopting the following technical scheme: a partial discharge identification and positioning method based on multi-mode calibration and blind source separation, Step 1, arranging a high-frequency current sensor, connecting the output end of the high-frequency current sensor to a high-speed data acquisition device, and setting sampling channel clocks for synchronization, wherein the high-frequency current sensor is connected to the grounding flat iron of a transformer core or the grounding flat iron of a clamping piece along the line; Step 2, collecting a mixed pulse signal, wherein the mixed pulse signal comprises a high-frequency sensor signal and at least one signal of other sensors, performing time-frequency conversion on the mixed pulse signal, and outputting multi-mode characteristics, wherein the multi-mode characteristics comprise a time sequence original waveform, a time-frequency spectrogram, an ultrasonic envelope signal, a PRPD image, an energy ratio, a frequency spectrum centroid and instantaneous phase information; step 3, a convolution mixed model of blind source separation is customized, a plurality of independent pulse source signals are output for mixed pulse signals by using blind source separation, and signal parameters of the independent pulse source signals are calculated, wherein the signal parameters comprise waveform similarity, pulse polarity, arrival time difference, energy and frequency spectrum characteristics; Step 4, periodically injecting a reference pulse signal into the flat iron, and estimating and adaptively updating the signal propagation speed along the flat iron based on the arrival time difference of the reference pulse; And step 5, acquiring the multi-modal characteristics and the signal parameters, inputting the multi-modal characteristics and the signal parameters into a deep learning classifier, outputting classification confidence, and comprehensively judging the source of the pulse signal interference. Preferably, the distance between the high-frequency current sensors is set to be 2-3 meters, the sampling rate of the high-speed data acquisition device is not less than 1GS/s, and the clock synchronization error between sampling channels is less than 1 nanosecond. Preferably, in the step 5, the comprehensive judgment is specifically that the classification confidence coefficient is evidently fused with evidence based on polarity, time difference and similarity, and the pulse signal is judged to be from external interference or internal discharge of the transformer; And judging that the internal discharge is generated, positioning along the flat iron vector based on the signal propagation speed and the arrival time difference, recording position information,