An acoustic signal partial discharge detection system for distribution network transformers

Abstract

The application discloses an acoustic signal partial discharge detection system of a distribution network transformer, and has the technical scheme as follows: a multi-sensor acquisition unit, a synchronous clock control unit, a signal preprocessing unit, an edge computing processing unit and a communication and storage unit; the edge computing processing unit is integrated with a cross-frequency band denoising module, a multi-scale feature extraction module, a lightweight neural network diagnosis module and a diagnosis result fusion module, wherein the cross-frequency band denoising module is used for jointly denoising audible sound and ultrasonic band signals. The application establishes a partial discharge detection scheme for the distribution network transformer scene, and the scheme process includes denoising, feature extraction, neural network training and diagnosis; the scheme covers audible sound band data and ultrasonic band data, and removes the interference generated by the mechanical vibration of the transformer in the ultrasonic band through a soft threshold value of wavelet transform and a pulse detection denoising algorithm.

Description

Technical Field

[0001] This invention relates to the field of signal detection technology, and more specifically to an acoustic signal partial discharge detection system for distribution network transformers. Background Technology

[0002] Distribution transformers are critical equipment in power systems, connecting the high-voltage grid to the user end. Their failures can lead to regional power outages, impacting industrial production and residential lives. If the potential for partial discharge in distribution transformers is not detected in time, it can eventually escalate into serious accidents such as insulation breakdown or even equipment explosion, posing a severe threat to the safe operation of the distribution network.

[0003] When conventional ultrasonic testing methods are used for distribution network transformer testing, the transformer itself exhibits mechanical vibrations, primarily displaying a 100Hz periodic characteristic. This makes it highly susceptible to generating interference signals similar to the partial discharge mechanism, affecting the test results. Furthermore, the partial discharge signals generated inside the distribution network transformer undergo significant attenuation as they propagate outward through the insulating oil and casing, further increasing the difficulty of detection.

[0004] In the existing technology, the ultrasonic array detection scheme proposed in patent CN202110345678.9 does not consider the harmonic interference of mechanical vibration in the ultrasonic frequency band; the joint detection system in patent CN202210123456.7 relies on synchronous acquisition of electrical signals, which is not suitable for the field environment of pole-mounted transformers. Therefore, there is an urgent need for a partial discharge detection scheme that takes into account anti-interference ability, signal penetration (ability to identify weak signals), and timeliness (early detection of partial discharge phenomena). Summary of the Invention

[0005] In view of the problems mentioned in the background art, the purpose of this invention is to provide an acoustic signal partial discharge detection system for distribution network transformers to solve the problems existing in the prior art.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: An acoustic signal partial discharge detection system for a distribution network transformer includes a multi-sensor acquisition unit, a synchronous clock control unit, a signal preprocessing unit, an edge computing processing unit, and a communication and storage unit. The edge computing processing unit integrates a cross-band denoising module, a multi-scale feature extraction module, a lightweight neural network diagnostic module, and a diagnostic result fusion module. The cross-band denoising module is used to jointly denoise audible and ultrasonic frequency signals. The multi-scale feature extraction module is used to generate multi-dimensional features containing time-frequency domain and power frequency correlation. The lightweight neural network diagnostic module is used for partial discharge signal identification. The diagnostic result fusion module is used for time-dimensional weighted fusion of multi-frame prediction results.

[0007] Ideally, the following processes should also be included: Step 1: Denoise the collected voiceprint signal by removing environmental interference through wavelet transform combined with adaptive soft thresholding, and then remove homologous pulse interference from the audible sound band within the ultrasonic band through pulse detection. Step 2: Perform multi-scale decomposition on the denoised data using wavelet transform, then perform frame-by-frame processing on the multi-scale data, calculate the correlation coefficient of each frame of data under different power frequency harmonics, and construct a feature vector involving multi-scale, time domain, and multiple power frequency harmonic correlation coefficients. Step 3: Model training. The feature vectors are used for model training. The lightweight MobileNetV3 network architecture is adopted, and the accuracy and number of parameters are balanced through depthwise separable convolution and channel attention mechanism. Step 4: Model prediction. Input the feature vector into the trained model to obtain the partial discharge detection results.

[0008] Preferably, when the multi-sensor acquisition unit is used with a square transformer, the sensors are placed at the midpoint of the four corners; when the multi-sensor acquisition unit is used with a triangular three-dimensional transformer, the sensors are placed at the midpoint of the three areas without heat sinks; the sensor positions in the multi-sensor acquisition unit avoid the heat sinks; the sensor frequency response range of the multi-sensor acquisition unit covers 15k-60kHz, the sampling rate is not less than 128kHz, and the synchronization clock control unit adopts GPS clock synchronization technology.

[0009] Preferably, the cross-band denoising module adopts a cascaded denoising architecture: the first stage performs multi-level decomposition of the signal through wavelet transform, calculates an adaptive threshold based on the noise standard deviation and signal characteristics of the coefficients of each level, and uses a soft threshold function to denoise the wavelet coefficients; the second stage uses pulse detection technology to identify and suppress cross-band co-source pulse interference.

[0010] Preferably, the wavelet transform uses db1 wavelet for 7-level decomposition. When thresholding the coefficients of each decomposition level, a nonlinear adjustment factor is used to optimize the threshold based on the coefficient distribution characteristics. After denoising, the wavelet coefficients are inversely transformed to reconstruct the signal.

[0011] Preferably, the multi-scale feature extraction module includes: a wavelet decomposition unit performing 7-level wavelet decomposition on the denoised signal to generate 8 sub-bands; a framing unit performing framing processing on the sub-band signals with 5 power frequency cycles as time windows; and a power frequency correlation calculation unit calculating the correlation coefficient of 1-10 times the power frequency to construct a multi-dimensional feature tensor.

[0012] Preferably, the lightweight neural network diagnostic module is based on the MobileNetV3 architecture and embeds a cross-band attention mechanism in the bottleneck module. This mechanism includes: performing band-wide global pooling on the input features, calculating the band attention weights through a two-layer fully connected network, and combining time-frequency cross-weighting operations to enhance the key band features.

[0013] Preferably, the lightweight neural network diagnostic module performs dimensionality transformation on the input features, expands them using bilinear interpolation to adapt to the network input requirements, and outputs multi-frame prediction results. Preferably, the diagnostic result fusion module adopts a time-weighted voting mechanism to ultimately output a partial discharge diagnostic probability value.

[0014] In summary, the present invention has the following main beneficial effects: This invention establishes a partial discharge (PD) detection scheme for distribution network transformer scenarios. The scheme process includes denoising, feature extraction, neural network training, and diagnosis. This scheme covers audible and ultrasonic frequency bands, using wavelet transform soft thresholding and pulse detection denoising algorithms to remove interference caused by transformer mechanical vibration in the ultrasonic band. In the feature extraction step, this scheme uses multi-scale, frame-based correlation coefficients from multiple power frequency harmonics to form a feature vector, containing the characteristic patterns of correlation coefficients changing with the frequency domain at different power frequency harmonics. This scheme adopts a lightweight deep neural network architecture, and preprocessing and post-processing optimizations have been performed on the model architecture to adapt to the current scenario. This invention designs an adaptive time-frequency feature extraction algorithm to enhance the power frequency correlation representation of PD signals. This invention optimizes the lightweight neural network architecture to achieve real-time diagnosis in edge computing scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart of the present invention; Figure 2 These are comparison images of wavelet transform and adaptive soft thresholding denoising before and after the process. Figure 3 This is the MobileNetV3 architecture diagram; Figure 4 This is a flowchart of feature vector dimension transformation and model prediction. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0017] refer to Figures 1 to 4 An acoustic signal partial discharge detection system for distribution network transformers is described, which performs partial discharge diagnosis of transformers through the following steps.

[0018] (1) The collected voiceprint signal is denoised. Environmental interference is removed by wavelet transform combined with adaptive soft threshold, and then the homogeneous pulse interference caused by the audible sound band in the ultrasonic band is removed by pulse detection.

[0019] (2) The denoised data is decomposed into multiple scales by wavelet transform, and then the multi-scale data is processed into frames to improve the temporal resolution of partial discharge detection and the ability to identify intermittent partial discharge. Next, the correlation coefficient of each frame of data under different power frequency harmonics is calculated to construct a feature vector involving multiple scales, time domain, and multiple power frequency harmonic correlation coefficients.

[0020] (3) Model training: The feature vectors are used for model training. The lightweight MobileNetV3 network architecture is adopted, which achieves a balance between accuracy and parameter quantity through depthwise separable convolution and channel attention mechanism.

[0021] (4) Model prediction: Input the feature vector into the trained model to obtain the partial discharge detection result.

[0022] The specific implementation is as follows: (1) Wavelet transform combined with adaptive soft thresholding and impulse detection for denoising 1) Use the db1 wavelet to perform a 7-level wavelet decomposition on the signal, calculate the variance of each decomposition level, and thus obtain the threshold of each frequency band: , The standard deviation of noise. The signal length; 2) Next, perform soft thresholding denoising: ,in, For the threshold, The sign function is used. Soft thresholding denoising can suppress noise-dominant coefficients while retaining signal-dominant coefficients; finally, an inverse transform is performed on the wavelet coefficients to reconstruct the denoised signal.

[0023] 3) Pulse detection is performed on the denoised reconstructed signal. First, the amplitude threshold, pulse width, and significance parameters are applied to detect and locate potential pulse segments. Here, the amplitude threshold is set to the mean of the absolute value of the signal + 0.5 * the standard deviation of the signal, the pulse width is greater than 30us, and the peak significance is greater than 0.1. 4) Perform time-domain and frequency-domain parameter checks on the extracted pulse signals, retaining those with kurtosis greater than 0.5 and peak amplitude exceeding 10% of the signal median; locate the pulse signals that do not meet the requirements in the corresponding wavelet coefficients and set the corresponding coefficients to zero; reconstruct the signal by inverse wavelet transform of the corrected coefficients to obtain the final denoising result.

[0024] like Figure 2 As shown, by using an adaptive soft threshold and pulse detection denoising algorithm, random interference and abnormal pulse interference in the environment can be effectively removed.

[0025] (2) For the denoised data, 1) First, a 7-level wavelet decomposition is performed to obtain multi-scale data in 8 frequency bands; among which... The original signal, and These are the approximation and detail coefficients for the j-th layer, respectively. Approximation coefficients for the j-th layer (low frequency): ; J-th level detail coefficients (high frequency): ; 2) Next, the signal is divided into frames. Each frame contains 5 power frequency cycles (power frequency is 50Hz, each frame is 0.1s data). For 10s of data, it can be divided into 100 frames. 3) Calculate the correlation coefficient of 1-10 times the power frequency for each frame of data, and obtain the correlation coefficients of different power frequency multiples for multi-scale and frame-by-frame as shown in Table 1. Thus, the dimension of the feature vector is (8, 100, 10).

[0026] Table 1. Correlation coefficients of different power frequency harmonics in multi-scale and framing scenarios.

[0027] Next, the feature vectors are input into the model for partial discharge detection. MobileNetV3 is chosen as the model architecture, and the neural network is trained as follows: Figure 3 As shown, to adapt to the network architecture, the feature vector is decomposed from (8, 100, 10) into (8, 33, 3, 10), and then the feature tensor is transformed into (33, 3, 8, 10). Then, through bilinear interpolation, it becomes (33, 3, 224, 224), thus dividing a sample into 33 frames for independent prediction. Therefore, the results of the 33 frames need to be weighted and fused to give the final prediction result.

[0028] (1) This invention establishes a partial discharge detection scheme for distribution network transformer scenarios. The scheme process includes denoising, feature extraction, neural network training and diagnosis. (2) This scheme covers audible and ultrasonic frequency band data. It removes interference caused by transformer mechanical vibration in the ultrasonic frequency band by using wavelet transform soft threshold and pulse detection denoising algorithm. (3) In the feature extraction step, this scheme adopts multi-scale, frame-by-frame, and multi-frequency harmonic correlation coefficients to form a feature vector, which contains the feature law of the correlation coefficients of different frequency harmonics changing with the frequency domain. (4) This solution adopts a lightweight deep neural network architecture. In order to adapt to the current scenario, the model architecture has been preprocessed and postprocessed to optimize it.

[0029] Example 2 refer to Figure 1 and Figure 2 An acoustic signal partial discharge detection system for a distribution network transformer includes a multi-sensor acquisition unit, a synchronous clock control unit, a signal preprocessing unit, an edge computing processing unit, and a communication and storage unit. The edge computing processing unit integrates a cross-band denoising module, a multi-scale feature extraction module, a lightweight neural network diagnostic module, and a diagnostic result fusion module. The cross-band denoising module is used to jointly denoise audible and ultrasonic frequency signals. The multi-scale feature extraction module is used to generate multi-dimensional features containing time-frequency domain and power frequency correlation. The lightweight neural network diagnostic module is used for partial discharge signal identification. The diagnostic result fusion module is used for time-dimensional weighted fusion of multi-frame prediction results.

[0030] This also includes the following processes: Step 1: Denoise the collected voiceprint signal by removing environmental interference through wavelet transform combined with adaptive soft thresholding, and then remove homologous pulse interference from the audible sound band within the ultrasonic band through pulse detection. Step 2: Perform multi-scale decomposition on the denoised data using wavelet transform, then perform frame-by-frame processing on the multi-scale data, calculate the correlation coefficient of each frame of data under different power frequency harmonics, and construct a feature vector involving multi-scale, time domain, and multiple power frequency harmonic correlation coefficients. Step 3: Model training. The feature vectors are used for model training. The lightweight MobileNetV3 network architecture is adopted, and the accuracy and number of parameters are balanced through depthwise separable convolution and channel attention mechanism. Step 4: Model prediction. Input the feature vector into the trained model to obtain the partial discharge detection results.

[0031] In the application of the multi-sensor acquisition unit to a square transformer, sensors are placed at the midpoint of the four corners; in the application to a triangular three-dimensional transformer, sensors are placed in the middle of the three areas without heat sinks; the sensor positions in the multi-sensor acquisition unit avoid the heat sinks; the sensor frequency response range of the multi-sensor acquisition unit covers 15kHz-60kHz, the sampling rate is not less than 128kHz, and the synchronization clock control unit adopts GPS clock synchronization technology. The cross-band denoising module adopts a cascaded denoising architecture: the first stage performs multi-level decomposition of the signal through wavelet transform, calculates adaptive thresholds based on the noise standard deviation and signal characteristics of each layer coefficient, and uses a soft thresholding function to denoise the wavelet coefficients; the second stage uses pulse detection technology to identify and suppress cross-band co-source pulse interference.

[0032] The lightweight neural network diagnostic module is based on the MobileNetV3 architecture and embeds a cross-band attention mechanism in the bottleneck module. This mechanism includes: performing frequency band global pooling on the input features, calculating the frequency band attention weights through a two-layer fully connected network, and combining time-frequency cross-weighting operations to enhance the key frequency band features.

[0033] The lightweight neural network diagnostic module performs dimensionality transformation on the input features: it converts the multidimensional feature tensor into a feature map of a specific size through three-dimensional convolution, expands it through bilinear interpolation to adapt to the network input requirements, and then performs weighted fusion of the output multi-frame prediction results through a time sliding window mechanism. The diagnostic result fusion module employs a time-weighted voting mechanism to weight and fuse the prediction results, ultimately outputting a partial discharge diagnostic probability value.

[0034] Among them, the present invention designs an adaptive time-frequency feature extraction algorithm to enhance the power frequency correlation characterization of partial discharge signals; the present invention optimizes a lightweight neural network architecture to realize real-time diagnosis in edge computing scenarios.

[0035] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An acoustic signal partial discharge detection system for a distribution network transformer, characterized in that, It includes a multi-sensor acquisition unit, a synchronous clock control unit, a signal preprocessing unit, an edge computing processing unit, and a communication and storage unit. The edge computing processing unit integrates a cross-band denoising module, a multi-scale feature extraction module, a lightweight neural network diagnostic module, and a diagnostic result fusion module. The cross-band denoising module is used to jointly denoise audible and ultrasonic frequency band signals. The multi-scale feature extraction module is used to generate multi-dimensional features containing time-frequency domain and power frequency correlation. The lightweight neural network diagnostic module is used for partial discharge signal identification. The diagnostic result fusion module is used for time-dimensional weighted fusion of multi-frame prediction results. It also includes the following processes: Step 1: Denoise the collected voiceprint signal by removing environmental interference through wavelet transform combined with adaptive soft thresholding, and then remove homologous pulse interference from the audible sound band within the ultrasonic band through pulse detection. Step 2: Perform multi-scale decomposition on the denoised data using wavelet transform, then perform frame-by-frame processing on the multi-scale data, calculate the correlation coefficient of each frame of data under different power frequency harmonics, and construct a feature vector involving multi-scale, time domain, and multiple power frequency harmonic correlation coefficients. Step 3: Model training. The feature vectors are used for model training. The lightweight MobileNetV3 network architecture is adopted, and the accuracy and number of parameters are balanced through depthwise separable convolution and channel attention mechanism. Step 4: Model prediction. Input the feature vector into the trained model to obtain the partial discharge detection results. The multi-scale feature extraction module includes: a wavelet decomposition unit to decompose the denoised signal and generate multiple sub-frequency bands; a framing unit to process the sub-frequency band signals into frames; and a power frequency correlation calculation unit to extract the power frequency correlation coefficient.

2. The acoustic signal partial discharge detection system for a distribution network transformer according to claim 1, characterized in that: When the multi-sensor acquisition unit is used with a square transformer, the sensors are placed at the midpoint of the four corners. When the multi-sensor acquisition unit is used with a triangular three-dimensional transformer, the sensors are placed in the middle of the three areas without heat sinks. The sensor positions in the multi-sensor acquisition unit avoid the heat sinks. The sensor frequency response range of the multi-sensor acquisition unit covers 15k-60kHz, the sampling rate is not less than 128kHz, and the synchronization clock control unit adopts GPS clock synchronization technology.

3. The acoustic signal partial discharge detection system for a distribution network transformer according to claim 2, characterized in that: The cross-band denoising module adopts a cascaded denoising architecture: the first stage performs multi-level decomposition of the signal through wavelet transform, calculates an adaptive threshold based on the noise standard deviation and signal characteristics of the coefficients of each level, and uses a soft threshold function to denoise the wavelet coefficients; The second stage uses pulse detection technology to identify and suppress cross-frequency band co-source pulse interference.

4. The acoustic signal partial discharge detection system for a distribution network transformer according to claim 3, characterized in that: The wavelet transform uses db1 wavelet for 7-level decomposition. When thresholding the coefficients of each decomposition level, a nonlinear adjustment factor is used to optimize the threshold based on the coefficient distribution characteristics. After denoising, the wavelet coefficients are inversely transformed to reconstruct the signal.

5. The acoustic signal partial discharge detection system for a distribution network transformer according to claim 1, characterized in that: The lightweight neural network diagnostic module is based on the MobileNetV3 architecture and embeds a cross-band attention mechanism in the bottleneck module. This mechanism includes: performing band-wide pooling on the input features, calculating the band attention weights through a two-layer fully connected network, and combining time-frequency cross-weighting operations to enhance key band features.

6. The acoustic signal partial discharge detection system for a distribution network transformer according to claim 5, characterized in that: The lightweight neural network diagnostic module performs dimensionality transformation on the input features, expands them using bilinear interpolation to adapt to the network input requirements, and then weights and fuses the multi-frame prediction results through a time sliding window mechanism.

Images