Partial discharge pattern recognition method based on multi-channel acoustic signal time-frequency space spectrum fusion

Abstract

This application provides a partial discharge pattern recognition method based on the fusion of time-frequency spatial spectra of multi-channel acoustic signals, belonging to the field of partial discharge pattern recognition technology. It solves the problem that existing multi-channel acoustic signals have information redundancy and differences, making it difficult to effectively fuse and extract discriminative features. The method includes the following steps: collecting and summarizing multi-channel acoustic signal data corresponding to different partial discharge models as a dataset; normalizing the original multi-channel partial discharge acoustic signals and mapping the time-domain signals to a time-frequency joint distribution to obtain a grayscale time-frequency spectrum; performing multi-scale and multi-directional decomposition on the grayscale time-frequency spectrum to obtain sub-maps for each channel; fusing the decomposed sub-maps using different fusion strategies; reconstructing the time-frequency spatial fusion spectrum from the fused sub-maps through inverse transformation; feature extraction; partial discharge pattern recognition; this application is applied to partial discharge detection.

Description

Technical Field

[0001] This application relates to the field of partial discharge pattern recognition technology, and in particular to a partial discharge pattern recognition method based on the fusion of time-frequency spatial spectrum of multi-channel acoustic signals. Background Technology

[0002] During long-term operation, high-voltage electrical equipment is prone to partial discharge due to factors such as insulation aging, manufacturing defects, or mechanical stress. The continued development of partial discharge accelerates insulation degradation, potentially leading to equipment failure. Therefore, effective detection and pattern recognition of partial discharge are crucial for power equipment condition assessment and fault early warning.

[0003] Acoustic detection of partial discharge has attracted widespread attention due to its strong resistance to electromagnetic interference and flexible sensor deployment. However, traditional acoustic detection methods often rely on single-channel signal analysis, making them susceptible to interference from complex ambient noise and unable to effectively capture the spatial distribution characteristics of discharge signals. Existing identification methods often directly extract features from time-domain or frequency-domain signals, resulting in limited characterization capabilities and requiring improvement in accuracy when dealing with various discharge types.

[0004] In recent years, the introduction of multi-sensor arrays has significantly improved the robustness of partial discharge signal acquisition. Combined with the continuous development of time-frequency analysis technology, significant progress has been made in the field of partial discharge detection. However, information redundancy and differences exist between multi-channel signals, and how to effectively fuse multi-channel acoustic signals and extract discriminative discharge features remains a challenge. Summary of the Invention

[0005] To address the aforementioned technical issues, this application proposes a partial discharge pattern recognition method based on the fusion of time-frequency spatial spectrum of multi-channel acoustic signals, which can achieve effective fusion of multi-channel partial discharge acoustic signals and partial discharge pattern recognition.

[0006] The technical solution adopted in this application is: a partial discharge pattern recognition method based on the fusion of time-frequency spatial spectrum of multi-channel acoustic signals, comprising the following steps:

[0007] Step 1: Construct a multi-channel partial discharge acoustic signal dataset: Build a multi-channel partial discharge acoustic signal acquisition circuit corresponding to different partial discharge models to acquire partial discharge acoustic signals. The acquisition circuit uses an acoustic array to acquire multi-channel partial discharge acoustic signals. The multi-channel partial discharge acoustic signal data corresponding to different partial discharge models are summarized as a dataset.

[0008] Step 2: Drawing the time-frequency distribution spectrum of partial discharge: After normalizing the original multi-channel partial discharge acoustic signal, the time-domain signal is mapped to a time-frequency joint distribution to obtain a grayscale time-frequency spectrum.

[0009] Step 3: Fusion of partial discharge time-frequency spatial spectrum, including:

[0010] Step 3.1: Perform multi-scale, multi-directional decomposition on the grayscale time-spectrum to obtain sub-images for each channel;

[0011] Step 3.2: Use different fusion strategies to fuse the sub-maps obtained from the decomposition;

[0012] Step 3.3: Reconstruct the time-frequency spatial fusion spectrum from the fused channel sub-graphs using inverse transformation;

[0013] Step 4: Feature Extraction: Extract feature information from the time-frequency spatial fusion spectrogram. To construct a multidimensional representation vector, where, This refers to the spatial structural scale characteristics. The pulse intensity difference is a characteristic. This represents the directional distribution preference characteristic;

[0014] Step 5: Partial discharge mode identification.

[0015] Furthermore, the partial discharge model in step 1 includes at least the corona discharge model, the surface discharge model, the internal discharge model, and the suspended particle discharge model;

[0016] The multi-channel partial discharge acoustic signal acquisition circuit includes a high-voltage power supply, a protective resistor, a partial discharge model, an acoustic array, a synchronous data acquisition card, and a PC. The high-voltage power supply is connected to the partial discharge model through the protective resistor, and the acoustic array is connected to the PC through the synchronous data acquisition card.

[0017] The acoustic array consists of N partial discharge acoustic acquisition elements evenly distributed along the circumference. Each element serves as an acquisition channel, and all elements employ the same circuit structure to synchronously acquire multi-channel partial discharge acoustic signals. Each element is equipped with a MEMS microphone chip.

[0018] Furthermore, the number of array elements N in the acoustic array is configured to be a power of 2, and the circular array radius r is defined as the distance from the center of each MEMS microphone chip to the geometric center of the acoustic array. The array radius r is determined by analyzing the 3dB main lobe bandwidth and the maximum side lobe level of the array through a beamforming algorithm.

[0019] Furthermore, during the acquisition of partial discharge acoustic signals, experiments were conducted using a continuous voltage ramp method for different typical partial discharge models. The voltage applied to the partial discharge model was gradually increased in a stepwise manner until a stable partial discharge phenomenon was observed.

[0020] Under stable discharge conditions, multi-channel partial discharge acoustic signals are continuously acquired for a duration of t for each partial discharge model. The acquired continuous signals are then divided into signal samples of length Δt, where Δt is set as an integer multiple of the power frequency period. Finally, each partial discharge model obtains n0 independent samples, which together constitute a partial discharge multi-channel acoustic signal dataset with a total sample size of n.

[0021] Furthermore, the specific steps in step 2 to map the time-domain signal into a joint time-frequency distribution and obtain the grayscale time-frequency spectrum are as follows:

[0022] The normalized continuous time-domain signal is divided into a series of short time frames in chronological order, and the length of each frame is set to L sampling points;

[0023] Overlap processing is performed between adjacent frames, and the frame shift is set to S sampling points. The frame shift S is constrained to a fraction of the frame length L, i.e., S=L / k, where k is an integer greater than 1.

[0024] Multiply each segmented frame signal by a window function. For each frame of signal after windowing Perform a discrete Fourier transform and calculate its amplitude spectrum. The amplitude spectra are stacked along the time axis to form a time-frequency matrix. Further analysis of the time-frequency matrix Perform grayscale value mapping processing, using linear normalization, to map the amplitude spectrum to the [0,G] interval, where G is the preset upper limit of grayscale levels, thus generating an A×B grayscale time-frequency spectrum:

[0025] ;

[0026] In the formula: The partial discharge acoustic signal acquired for the i-th channel in the time-frequency domain grayscale value at that location For the duration of data collection, For frequency variables, For time frame indexing, For frequency indexing.

[0027] Furthermore, step 3.1 involves multi-scale, multi-directional decomposition of the grayscale time-spectrum image to obtain the sub-images of each channel, and the specific steps are as follows:

[0028] Step 3.1.1: Construct a two-level decomposition pyramid. First-level decomposition: Apply a low-pass filter to smooth the original time-spectrum image, downsample the smoothed image by a factor of 2 to obtain the first-level approximate sub-image AI, then upsample the first-level approximate sub-image AI to restore it to its original size, and subtract it from the original time-spectrum image to obtain the first-scale detail sub-image DI. Second-level decomposition: Use the first-level approximate sub-image AI as input, repeat the above smoothing and downsampling process to obtain the second-level approximate sub-image AII, and obtain the second-scale detail sub-image DII by calculating the residual.

[0029] Step 3.1.2: Perform directional decomposition of different levels on the detail sub-images of the two scales using a directional filter bank, and perform primary directional projection decomposition on the first-scale detail sub-image DI to generate 2 k1 A directional submap is generated, and a high-level directional projection decomposition is performed on the second-scale detail submap DII to generate 2 k2 There are 1 directional sub-maps, and the directional coverage range of each directional sub-map is evenly distributed from 0° to 180°.

[0030] The above decomposition is performed on the grayscale spectrogram of each channel, and the coefficient matrix of the approximate sub-image for each channel is output. and direction detail subgraph coefficient matrix Where m=(1,2,…,M) is the number of scale decomposition layers, p is the directional subband coefficient at scale m, and i is the channel index.

[0031] Furthermore, step 3.2 involves fusing the decomposed channel sub-images using different fusion strategies, and the specific steps are as follows:

[0032] Step 3.2.1: Introduce multi-scale contrast analysis for each pixel. Set an inner window to extract local texture features and an outer window as a regional background reference. Quantify the information contribution of each pixel in the overall image by comparing the contrast between the features of the inner and outer windows, thereby obtaining the local information contribution of each element at each pixel.

[0033] ;

[0034] In the formula: For the i-th channel at pixel point The contribution of local information at a given location; For The mean of the coefficients of the approximate subgraph within the inner window centered on the graph; For The mean of the outer ring region centered on the window is the outer window minus the mean of the inner window region. It is a very small constant to prevent the denominator from being zero;

[0035] Step 3.2.2: Normalize the contribution of local information to obtain the fusion weights:

[0036] ;

[0037] In the formula: For the i-th channel at pixel point Approximate subgraph fusion weights at the location; N is the number of array elements;

[0038] Step 3.2.3: Approximate the coefficient matrix of each channel's sub-graph With corresponding weights Weighted fusion is performed to obtain an approximate fused subgraph;

[0039] Step 3.2.4: For the directional detail submap, firstly, a fixed-size sliding window is used to traverse the high-frequency submap, and local difference statistics are calculated within the window region:

[0040] ;

[0041] In the formula: It is a local difference statistic; For the pixel of the p-th directional submap of the i-th channel at the m-th scale The directional detail subgraph coefficients at the location; k s The window radius controls the size of the local area;

[0042] Step 3.2.5: Generate fusion weights for the directional detail submap based on the above local difference statistics:

[0043] ;

[0044] In the formula: For the pixel of the p-th directional submap of the i-th channel at the m-th scale The fusion weights at the location can adaptively highlight regions containing strong texture and transient impulse features;

[0045] Step 3.2.6: Convert the coefficient matrix of the detail submap for each channel direction. With the corresponding fusion weights Perform weighted fusion to obtain the fused detailed sub-graphs in each direction.

[0046] Furthermore, step 3.3, which involves reconstructing the time-frequency spatial fusion spectrum of each channel sub-graph after fusion through inverse transformation, is to reconstruct the approximate sub-graph and the directional detail sub-graph after fusion according to the reverse steps of the decomposition process, thereby reconstructing the partial discharge time-frequency spatial fusion spectrum G(x,y).

[0047] Furthermore, The formula used to characterize the local energy distribution and spatial scale differences of discharge pulses in the time-frequency spatial fusion spectrum of partial discharge is as follows:

[0048] ;

[0049] In the formula: The number of discrete sampling points in the frequency domain. The number of discrete sampling points in the time domain. The optimal scale that best represents the global structural features;

[0050] The formula for quantifying the brightness difference between the discharge pulse and the background in the time-frequency spatial fusion spectrum of partial discharge is as follows:

[0051] ;

[0052] Where: V is the standard deviation of the time-frequency spatial fusion spectrum of partial discharge, and K is the fourth central moment of the time-frequency spatial fusion spectrum of partial discharge;

[0053] The formula used to reflect the main directional trend of texture in the time-frequency spatial fusion spectrum of partial discharge is as follows:

[0054] ;

[0055] In the formula: Let the center angle of the b-th interval be . dominant interval of the histogram The average center angle, H b Let be the cumulative magnitude of the b-th interval.

[0056] Furthermore, in step 5, the sample features are divided into training set, validation set, and test set in a specific ratio and fed into the classifier to complete the partial discharge pattern recognition.

[0057] The advantages of this application over the prior art are as follows:

[0058] (1) This application effectively overcomes the limitation of single-channel detection being susceptible to noise interference by array-type acoustic signal acquisition and multi-channel signal fusion, and significantly improves the reliability and stability of signal acquisition.

[0059] (2) This application constructs a time-frequency distribution spectrum, mapping the partial discharge acoustic signal into a time-frequency joint distribution form, which fully preserves the time-frequency local features of the signal. Furthermore, a multi-scale, multi-directional decomposition and adaptive fusion strategy is adopted to achieve deep integration of multi-channel time-frequency information in the spatial dimension, thereby enhancing the ability to characterize the discharge pulse features.

[0060] (3) Based on the above time-frequency spatial fusion spectrum, this application realizes the pattern recognition of partial discharge. This method breaks through the dependence of traditional methods on single signal modes and limited feature dimensions by enhancing spatial dimensions, extracting multi-scale directional features and adaptive fusion mechanism, and significantly improves the ability to distinguish various types of partial discharge and the accuracy of recognition. Attached Figure Description

[0061] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0062] Figure 1 A flowchart illustrating the method provided in this application embodiment;

[0063] Figure 2 This is a schematic diagram of a common partial discharge model in electrical equipment provided in the embodiments of this application;

[0064] Figure 3 A circuit diagram for acquiring multi-channel partial discharge acoustic signals provided in an embodiment of this application;

[0065] Figure 4 This is a flowchart of partial discharge time-frequency spatial spectrum fusion provided in an embodiment of this application;

[0066] Figure 5 A diagram showing the correspondence between array radius and maximum sidelobe level provided for embodiments of this application;

[0067] Figure 6 The time-frequency distribution spectrum of this application is given as an example using a corona discharge model;

[0068] Figure 7 These are detailed sub-images of the corona discharge time-frequency distribution spectrum after multi-scale and multi-directional decomposition, as provided in the embodiments of this application.

[0069] Figure 8 These are detailed sub-graphs in each direction after fusion using the fusion strategy of this application;

[0070] Figure 9 The final reconstructed time-frequency spatial fusion spectrum of corona discharge;

[0071] Figure 10 A graph showing the accuracy of partial discharge identification on the same test set using different methods, provided for embodiments of this application.

[0072] In the diagram: 1 is the high-voltage power supply, 2 is the protective resistor, 3 is the partial discharge model, 4 is the acoustic array, 5 is the synchronous data acquisition card, and 6 is the PC terminal. Detailed Implementation

[0073] like Figures 1 to 10As shown, this application provides a partial discharge pattern recognition method based on the fusion of time-frequency spatial spectrum of multi-channel acoustic signals, which mainly includes the following steps:

[0074] Step 1: Construct a multi-channel partial discharge acoustic signal dataset;

[0075] Step 2: Plot the time-frequency distribution spectrum of partial discharge;

[0076] Step 3: Fusion of partial discharge time-frequency spatial spectrum;

[0077] Step 4: Feature extraction;

[0078] Step 5: Partial discharge mode identification.

[0079] The construction of the multi-channel acoustic signal dataset in step 1 is achieved by designing a partial discharge model 3 based on common partial discharge types found in electrical equipment and building a corresponding multi-channel partial discharge acoustic signal acquisition circuit. The partial discharge model 3 includes corona discharge model, surface discharge model, internal discharge model, and suspended particle discharge model, such as... Figure 2 As shown. (2a) is a corona discharge model, consisting of a copper metal needle, an epoxy resin plate and a stainless steel plate; (2b) is a surface discharge model, in which an epoxy resin plate with a thickness of 5 mm is sandwiched between two cylindrical electrodes to simulate surface discharge; (2c) is an internal discharge model, using a cylindrical pore with a diameter of 8 mm as an air gap; (2d) is a suspended particle discharge model, using a copper wire with a length of 3 mm as metal particles.

[0080] The acquisition circuit for multi-channel partial discharge acoustic signals is as follows: Figure 3 As shown, the system includes a high-voltage power supply 1, a protective resistor 2, a partial discharge model 3, an acoustic array 4, a synchronous data acquisition card 5, and a PC terminal 6. The high-voltage power supply 1 is connected to the partial discharge model 3 through the protective resistor 2, and the acoustic array 4 is connected to the PC terminal 6 through the synchronous data acquisition card 5. In the figure, the high-voltage power supply 1 is a partial discharge-free test transformer that outputs an adjustable 50Hz AC voltage in the range of 0~100kV. The protective resistor 2 is a 20kΩ water resistor. The synchronous data acquisition card 5 is a PXIE-5323 data acquisition card, which can realize synchronous acquisition of up to 32 channels with a sampling rate of 100kS / s.

[0081] Acoustic array 4 is an array-type partial discharge acoustic signal acquisition device, which consists of N partial discharge acoustic acquisition elements evenly distributed along the circumference. Each element is equipped with a MEMS microphone chip. Each element serves as an acquisition channel, and all elements employ the same circuit structure to synchronously acquire multi-channel partial discharge acoustic signals, ensuring no time delay between the multi-channel output signals.

[0082] The number of array elements N is configured as a power of 2, including 8, 16, 32, 64, etc. The array elements are arranged in a uniform circular pattern to form a highly symmetrical array layout, thereby forming a uniformly symmetrical radiation pattern within a 360° range to maintain consistent detection performance and suppress sidelobe levels in specific directions.

[0083] The circular array radius *r* is defined as the distance from the center of each MEMS microphone chip to the geometric center of the acoustic array. When determining the array radius *r*, a beamforming algorithm is used to analyze the array's 3dB main lobe bandwidth and maximum sidelobe level. The 3dB main lobe bandwidth is used to evaluate the array's ability to distinguish nearby sound sources, and the maximum sidelobe level is used to evaluate the array's anti-interference performance. For the selected number of array elements *N* and the highest frequency *F* of the partial discharge acoustic signal received by the MEMS microphone, the radius value that optimizes the combined performance of the 3dB main lobe bandwidth and maximum sidelobe level is selected as the working radius. In actual deployment, a suitable number of array elements *N* and the corresponding optimal array radius *r* are selected based on the site installation conditions.

[0084] During the acquisition of partial discharge acoustic signals, experiments were conducted using a continuous voltage ramp method for four different typical partial discharge models 3. The voltage applied to the partial discharge model 3 was gradually increased in a stepwise manner until a stable partial discharge phenomenon was observed. Under stable discharge conditions, multi-channel partial discharge acoustic signals were continuously acquired for each partial discharge model 3 for a duration of t. The acquired continuous signals were then divided into signal samples of length Δt, where Δt was set as an integer multiple of the power frequency period (20ms) to ensure that each signal sample could reflect both local pulse characteristics and global periodic characteristics. Finally, n0 independent samples were obtained for each partial discharge model 3, which together constituted a partial discharge multi-channel acoustic signal dataset with a total sample size of n.

[0085] The process of drawing the time-frequency distribution spectrum of partial discharge in step 2 is as follows:

[0086] To avoid signal amplitude deviations and environmental noise interference caused by differences in sensitivity and installation location of partial discharge acoustic acquisition array elements, the original multi-channel partial discharge acoustic signals were processed. Normalization process is performed to obtain ,Right now:

[0087] (1);

[0088] In the formula: Let be the arithmetic mean of the i-th channel signal; Let be the standard deviation of the signal in the i-th channel, where i is the channel index, which is a positive integer 1, 2, 3, ..., N.

[0089] After normalization, the signals of each channel follow a zero-mean, unit-variance distribution, eliminating the interference of dimensional differences on subsequent time-frequency analysis.

[0090] Mapping the time-domain signal to a joint time-frequency distribution: First, the normalized continuous time-domain signal is divided into a series of short frames in chronological order, with each frame having a length of L sampling points. To ensure smooth transitions between frames and avoid information loss, adjacent frames are overlapped, with a frame shift of S sampling points. To ensure the rationality of framing, the frame shift S is constrained to a fraction of the frame length L, i.e., S = L / o, where o is an integer greater than 1. Then, each segmented frame signal is multiplied by a window function. To suppress spectral leakage caused by signal truncation, each frame of the windowed signal... Perform a discrete Fourier transform and calculate its amplitude spectrum. The amplitude spectra are stacked along the time axis to form a time-frequency matrix. Further analysis of the time-frequency matrix Perform grayscale value mapping processing, using linear normalization, to map the amplitude spectrum to the [0,G] interval, where G is the preset upper limit of grayscale levels, thus generating an A×B grayscale time-frequency spectrum:

[0091] (2);

[0092] In the formula: The partial discharge acoustic signal acquired for the i-th channel in the time-frequency domain grayscale value at that location For the duration of data collection, For frequency variables, For time frame indexing, For frequency indexing.

[0093] In this grayscale time-frequency spectrum, the horizontal axis corresponds to time [0, Δt], and the vertical axis corresponds to frequency [0, F]. All spectra are taken from the same coordinate system. To eliminate the interference of non-data elements such as coordinate axes and tick marks in the original image on subsequent computer vision algorithms, all time-frequency spectra are uniformly cropped before feature extraction, retaining only the image regions corresponding to pure time-frequency data.

[0094] Step 3, the partial discharge time-frequency spatial spectrum fusion, mainly consists of three steps: Step 3.1: Multi-scale, multi-directional decomposition of the grayscale time-frequency spectrum to obtain sub-images for each channel; Step 3.2: Fusing the decomposed sub-images for each channel using different fusion strategies; Step 3.3: Reconstructing the time-frequency spatial fused spectrum from the fused sub-images for each channel through inverse transformation. The flowchart is as follows. Figure 4 As shown.

[0095] Step 3.1 specifically includes:

[0096] Step 3.1.1: Construct a two-level decomposition pyramid. First-level decomposition: Apply a low-pass filter to smooth the original time-spectrum image. Downsample the smoothed image by a factor of 2 to obtain the first-level approximate sub-image (AI). Then upsample the AI ​​to restore its original size and subtract it from the original time-spectrum image to obtain the first-scale detail sub-image (DI). Second-level decomposition: Using AI as input, repeat the above smoothing and downsampling process to obtain the second-level approximate sub-image (AII). Calculate the residual to obtain the second-scale detail sub-image (DII).

[0097] Step 3.1.2: Perform directional decomposition of different levels on the detail sub-images of the two scales using a directional filter bank. Perform primary directional projection decomposition on the first-scale detail sub-image (DI) to generate 2 k1 The second-scale detail submap (DII) is decomposed using high-level directional projection to generate 2 directional submaps. k2 Each directional submap can focus on discharge texture features with a specific orientation, and the directional coverage range is uniformly distributed from 0° to 180°.

[0098] The above decomposition is performed on the grayscale spectrogram of each channel, and the coefficient matrix of the approximate sub-image for each channel is output. and direction detail subgraph coefficient matrix Where m=(1,2,…,M) is the number of scale decomposition layers, and p is the directional sub-band coefficient at scale m.

[0099] Step 3.2: The second-level approximation sub-image (AII) and the two-scale directional detail sub-images (DI and DII directional sub-images) are fused using different fusion strategies. The approximation sub-images contain global structural information of the image. Discharge pulses exhibit short-term energy accumulation characteristics in the time-frequency domain, and traditional averaging methods easily lead to feature diffusion. Therefore, this application proposes a fusion strategy based on local information contribution, the specific steps of which are as follows:

[0100] Step 3.2.1: Introduce multi-scale contrast analysis for each pixel, and set the inner window (window radius k). l1 ) is used to extract local texture features, with an outer window (window radius k) l2 As a regional background reference, the information contribution of a pixel in the overall image is quantified by the contrast between the inner and outer window features, thereby obtaining the local information contribution of each element at each pixel:

[0101] (3);

[0102] In the formula: For the i-th channel at pixel point The contribution of local information at a given location; For The mean of the coefficients of the approximate subgraph within the inner window centered on the graph; For The mean of the outer ring region centered on the window (outer window excluding inner window region); For a local constant, take 10. -6 To prevent the denominator from being zero.

[0103] Step 3.2.2: Further normalize the contribution of local information to obtain the fusion weights:

[0104] (4);

[0105] In the formula: For the i-th channel at pixel point The approximate subgraph fusion weight at the location; N is the number of array elements.

[0106] Step 3.2.3: Approximate the coefficient matrix of each channel's sub-graph With corresponding weights Weighted fusion is performed to obtain an approximate fused subgraph.

[0107] For the directional detail sub-image, in order to highlight the pulse transient components in the discharge signal and prevent weak pulse features from being masked during the fusion process, a fusion strategy based on local detail sensitivity is used to effectively preserve edge clarity while enhancing key textures and pulse features.

[0108] Step 3.2.4: First, a fixed-size sliding window is used to traverse the high-frequency subgraph, and local difference statistics are calculated within the window region:

[0109] (5);

[0110] In the formula: It is a local difference statistic; For the pixel of the p-th directional submap of the i-th channel at the m-th scale The directional detail subgraph coefficients at the location; k s The window radius controls the size of the local area.

[0111] Step 3.2.5: Further generate fusion weights for the directional detail submap based on the above local difference statistics:

[0112] (6);

[0113] In the formula: For the pixel of the p-th directional submap of the i-th channel at the m-th scale The fusion weights at the location can adaptively highlight regions containing strong texture and transient pulse features.

[0114] Step 3.2.6: Convert the coefficient matrix of the detail submap for each channel direction. With the corresponding fusion weights Perform weighted fusion to obtain the fused detailed sub-graphs in each direction.

[0115] In step 3.3, the fused approximate subgraph and the directional detail subgraph are reconstructed in the reverse order of the decomposition process, so as to reconstruct the partial discharge time-frequency spatial fusion spectrum G(x,y).

[0116] Step 4, feature extraction, involves extracting feature information from the temporal-frequency spatial fusion spectrum of partial discharge. To construct a multidimensional representation vector, where This refers to the spatial structural scale characteristics. The pulse intensity difference is a characteristic. This represents the directional distribution preference feature.

[0117] The calculation principle for characterizing the local energy distribution and spatial scale differences of discharge pulses in the time-frequency spatial fusion spectrum of partial discharge is as follows:

[0118] Calculate the value of each pixel (x, y) in 2. k ×2 k Average intensity within the window (k=0,1,…,5):

[0119] (7);

[0120] Calculate the local intensity changes at each scale:

[0121] (8);

[0122] Determine the optimal scale that best represents the global structural features:

[0123] (9);

[0124] The spatial structural scale characteristics of the entire image are:

[0125] (10);

[0126] In the formula: The number of discrete sampling points in the frequency domain. The number of discrete sampling points in the time domain. It is the optimal scale that best represents the global structural features.

[0127] The principle behind the calculation of the brightness difference between the discharge pulse and the background in the time-frequency spatial fusion spectrum of partial discharge is as follows:

[0128] Calculate the average intensity of the time-frequency spatial fusion spectrum of partial discharge:

[0129] (11);

[0130] Calculate the standard deviation of the time-frequency spatial fusion spectrum of partial discharge:

[0131] (12);

[0132] Calculate the fourth central moment of the time-frequency spatial fusion spectrum of partial discharge:

[0133] (13);

[0134] The pulse intensity difference characteristic is defined as:

[0135] (14);

[0136] The calculation principle for reflecting the main directional trend of texture in the time-frequency spatial fusion spectrum of partial discharge is as follows:

[0137] Calculate the horizontal gradient of the time-frequency spatial fusion spectrum of partial discharge. and vertical gradient Calculate the gradient magnitude With gradient direction angle :

[0138] (15);

[0139] (16);

[0140] (17);

[0141] (18).

[0142] Divide the area [0°, 180°] into 18 equal intervals. The angle range of the b-th interval is: For satisfying The pixel, based on its gradient direction angle The interval to which it belongs Perform counting and accumulating gradient magnitudes, where The critical value used to define "drastic changes in grayscale". The cumulative amplitude H in the b-th interval. b for:

[0143] (19);

[0144] The dominant interval of the histogram is: Then the directional distribution preference characteristics for:

[0145] (20);

[0146] In the formula: Let the center angle of the b-th interval be . dominant interval of the histogram The average center angle.

[0147] The feature vector is obtained through the above calculations. In step 5, the sample features are divided into training set, validation set and test set according to a specific ratio and fed into the classifier to complete the partial discharge pattern recognition.

[0148] The present application will be further described below with reference to a specific embodiment.

[0149] according to Figure 1 The experiment was conducted according to the procedure shown. Figure 3 The multi-channel partial discharge acoustic signal acquisition circuit shown uses four typical discharge models—corona discharge, surface discharge, internal discharge, and suspended particle discharge—as the detection targets.

[0150] Parameters of the array-type partial discharge acoustic signal acquisition device: Due to the constraints of the acquisition device, the number of array elements N is set to 32, and the highest frequency F of the partial discharge acoustic signal received by the MEMS microphone is 20kHz. Based on the above conditions, a beamforming algorithm is used to obtain the 3dB main lobe width and the maximum side lobe level. Figure 5 The optimal radius is 17cm. A 32-element acoustic array was placed 1m away from the partial discharge-free test transformer, and acoustic signal data for four discharge types were collected using a continuous voltage boosting method.

[0151] Dataset parameters: The continuous acquisition time t for each discharge type is 50s, the signal sample length Δt is 100ms, ensuring that each sample contains 5 power frequency cycles of signal, the number of samples n0 for each discharge type is 500, and the total number of samples n is 2000.

[0152] Time-frequency joint distribution parameters: The length L of each time frame is set to 1024 points, k=4, then the frame shift S=256, and the upper limit of grayscale level G is set to 255, thus generating a 256×256 grayscale time-frequency matrix. Taking corona discharge as an example, its time-frequency distribution spectrum is as follows... Figure 6 As shown.

[0153] Multi-scale multi-directional decomposition parameters: The number of scale decomposition layers M is set to 2; primary directional decomposition is performed on the first-scale detail sub-map (DI), generating 21 directional detail sub-maps (0°, 90°); advanced directional decomposition is performed on the second-scale detail sub-map (DII), generating 22 directional detail sub-maps (0°, 45°, 90°, 135°); the scale decomposition low-pass filter uses a maximum flatness filter; the directional filter bank uses an even-numbered Gabor filter. Multi-scale multi-directional decomposition is performed on the corona discharge time-frequency distribution spectrum, such as... Figure 7 As shown.

[0154] Fusion strategy parameters: The approximate subgraph adopts a fusion strategy based on the contribution of local information, and the inner window radius k l1 Set to 1, i.e., a 3×3 window, with an outer window radius of k. l2 =2, i.e., a 5×5 window; the directional detail submap adopts a fusion strategy based on local detail sensitivity, with a sliding window radius of k. s Also set to 2. The fused detail sub-images in each direction are as follows: Figure 8 As shown, the final reconstructed time-frequency spatial fusion spectrum of corona discharge is as follows: Figure 9 As shown.

[0155] Feature extraction and pattern recognition: Extracting feature information from the time-frequency spatial fusion spectrum of partial discharge Construct multidimensional representation vectors. 2000 samples are divided into training, validation, and test sets in an 8:1:1 ratio and fed into the classifier for pattern recognition.

[0156] Finally, to verify the superiority of this method, a comparative experiment was conducted with the following three methods: directly extracting features from the time-domain signal, directly extracting features from the frequency-domain signal, and extracting features from a single-channel time-spectrum graph. The recognition accuracy of each method on the same test set is shown below. Figure 10 As shown.

[0157] Experimental results show that the partial discharge pattern recognition method based on multi-channel acoustic signal time-frequency spatial spectrum fusion provided in this application can effectively integrate multi-channel information and highlight discharge characteristics. Ultimately, the method of this application significantly improves the accuracy and reliability of partial discharge pattern recognition, and its overall performance is significantly better than traditional methods.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for partial discharge pattern recognition based on time-frequency spatial spectrum fusion of multi-channel acoustic signals, characterized in that: Includes the following steps: Step 1: Construct a multi-channel partial discharge acoustic signal dataset: Build a multi-channel partial discharge acoustic signal acquisition circuit corresponding to different partial discharge models to acquire partial discharge acoustic signals. The acquisition circuit uses an acoustic array to acquire multi-channel partial discharge acoustic signals. The multi-channel partial discharge acoustic signal data corresponding to different partial discharge models are summarized as a dataset. Step 2: Drawing the time-frequency distribution spectrum of partial discharge: After normalizing the original multi-channel partial discharge acoustic signal, the time-domain signal is mapped to a time-frequency joint distribution to obtain a grayscale time-frequency spectrum. Step 3: Fusion of partial discharge time-frequency spatial spectrum, including: Step 3.1: Perform multi-scale, multi-directional decomposition on the grayscale time-spectrum to obtain sub-images for each channel. The specific steps are as follows: Step 3.1.1: Construct a two-level decomposition pyramid. First-level decomposition: Apply a low-pass filter to smooth the original time-spectrum image, downsample the smoothed image by a factor of 2 to obtain the first-level approximate sub-image AI, then upsample the first-level approximate sub-image AI to restore it to its original size, and subtract it from the original time-spectrum image to obtain the first-scale detail sub-image DI. Second-level decomposition: Use the first-level approximate sub-image AI as input, repeat the above smoothing and downsampling process to obtain the second-level approximate sub-image AII, and obtain the second-scale detail sub-image DII by calculating the residual. Step 3.1.2: Perform directional decomposition of different levels on the detail sub-images of the two scales using a directional filter bank, and perform primary directional projection decomposition on the first-scale detail sub-image DI to generate 2 k1 A directional submap is generated, and a high-level directional projection decomposition is performed on the second-scale detail submap DII to generate 2 k2 There are 1 directional sub-maps, and the directional coverage range of each directional sub-map is evenly distributed from 0° to 180°. The above decomposition is performed on the grayscale spectrogram of each channel, and the coefficient matrix of the approximate sub-image for each channel is output. and direction detail subgraph coefficient matrix Where m=(1,2,…,M) is the number of scale decomposition layers, p is the directional subband coefficient at scale m, and i is the channel index; Step 3.2: Use different fusion strategies to fuse the sub-maps obtained from the decomposition; Step 3.3: Reconstructing the time-frequency spatial fusion spectrum of each channel sub-graph after fusion by inverse transformation involves reconstructing the approximate sub-graph and directional detail sub-graph after fusion according to the reverse steps of the decomposition process, so as to reconstruct the partial discharge time-frequency spatial fusion spectrum G(x,y); Step 4: Feature Extraction: Extract feature information from the time-frequency spatial fusion spectrogram. To construct a multidimensional representation vector, where, This refers to the spatial structural scale characteristics. The pulse intensity difference is a characteristic. This represents the directional distribution preference characteristic; Step 5: Partial discharge mode identification.

2. The partial discharge pattern recognition method based on multi-channel acoustic signal time-frequency spatial spectrum fusion according to claim 1, characterized in that: The partial discharge model in step 1 includes at least the corona discharge model, the surface discharge model, the internal discharge model, and the suspended particle discharge model; The multi-channel partial discharge acoustic signal acquisition circuit includes a high-voltage power supply, a protective resistor, a partial discharge model, an acoustic array, a synchronous data acquisition card, and a PC. The high-voltage power supply is connected to the partial discharge model through the protective resistor, and the acoustic array is connected to the PC through the synchronous data acquisition card. The acoustic array consists of N partial discharge acoustic acquisition elements evenly distributed along the circumference. Each element serves as an acquisition channel, and all elements employ the same circuit structure to synchronously acquire multi-channel partial discharge acoustic signals. Each element is equipped with a MEMS microphone chip.

3. The partial discharge pattern recognition method based on multi-channel acoustic signal time-frequency spatial spectrum fusion according to claim 2, characterized in that: The number of array elements N in the acoustic array is configured to be a power of 2, and the circular array radius r is defined as the distance from the center of each MEMS microphone chip to the geometric center of the acoustic array. The array radius r is determined by analyzing the 3dB main lobe bandwidth and the maximum side lobe level of the array through a beamforming algorithm.

4. The partial discharge pattern recognition method based on multi-channel acoustic signal time-frequency spatial spectrum fusion according to claim 2, characterized in that: During the acquisition of partial discharge acoustic signals, experiments were conducted using the continuous voltage ramp method for different typical partial discharge models. The voltage applied to the partial discharge model was gradually increased in a stepwise manner until a stable partial discharge phenomenon was observed. Under stable discharge conditions, multi-channel partial discharge acoustic signals are continuously acquired for a duration of t for each partial discharge model. The acquired continuous signals are then divided into signal samples of length Δt, where Δt is set as an integer multiple of the power frequency period. Finally, each partial discharge model obtains n0 independent samples, which together constitute a partial discharge multi-channel acoustic signal dataset with a total sample size of n.

5. The partial discharge pattern recognition method based on multi-channel acoustic signal time-frequency spatial spectrum fusion according to claim 4, characterized in that: The specific steps in step 2 to map the time-domain signal to a joint time-frequency distribution and obtain the grayscale time-frequency spectrum are as follows: The normalized continuous time-domain signal is divided into a series of short time frames in chronological order, and the length of each frame is set to L sampling points; Overlap processing is performed between adjacent frames, and the frame shift is set to S sampling points. The frame shift S is constrained to a fraction of the frame length L, i.e., S=L / k, where k is an integer greater than 1. Multiply each segmented frame signal by a window function. For each frame of signal after windowing Perform a discrete Fourier transform and calculate its amplitude spectrum. The amplitude spectra are stacked along the time axis to form a time-frequency matrix. Further analysis of the time-frequency matrix Perform grayscale value mapping processing, using linear normalization, to map the amplitude spectrum to the [0,G] interval, where G is the preset upper limit of grayscale levels, thus generating an A×B grayscale time-frequency spectrum: ； In the formula: The partial discharge acoustic signal acquired for the i-th channel in the time-frequency domain grayscale value at that location For the duration of data collection, For frequency variables, For time frame indexing, For frequency indexing.

6. The partial discharge pattern recognition method based on time-frequency spatial spectrum fusion of multi-channel acoustic signals according to claim 5, characterized in that: Step 3.2 The specific steps for fusing the decomposed channel sub-images using different fusion strategies are as follows: Step 3.2.1: Introduce multi-scale contrast analysis for each pixel. Set an inner window to extract local texture features and an outer window as a regional background reference. Quantify the information contribution of each pixel in the overall image by comparing the contrast between the features of the inner and outer windows, thereby obtaining the local information contribution of each element at each pixel. ； In the formula: For the i-th channel at pixel point The contribution of local information at a given location; For The mean of the coefficients of the approximate subgraph within the inner window centered on the graph; For The mean of the outer ring region centered on the window is the outer window minus the mean of the inner window region. It is a very small constant to prevent the denominator from being zero; Step 3.2.2: Normalize the contribution of local information to obtain the fusion weights: ； In the formula: For the i-th channel at pixel point Approximate subgraph fusion weights at the location; N is the number of array elements; Step 3.2.3: Approximate the coefficient matrix of each channel's sub-graph With corresponding weights Weighted fusion is performed to obtain an approximate fused subgraph; Step 3.2.4: For the directional detail submap, firstly, a fixed-size sliding window is used to traverse the high-frequency submap, and local difference statistics are calculated within the window region: ； In the formula: This is a local difference statistic; For the pixel of the p-th directional submap of the i-th channel at the m-th scale The directional detail subgraph coefficients at the location; k s The window radius controls the size of the local area; Step 3.2.5: Generate fusion weights for the directional detail submap based on the above local difference statistics: ； In the formula: For the pixel of the p-th directional submap of the i-th channel at the m-th scale The fusion weights at the location can adaptively highlight regions containing strong texture and transient impulse features; Step 3.2.6: Convert the coefficient matrix of the detail submap for each channel direction. With the corresponding fusion weights Perform weighted fusion to obtain the fused detailed sub-graphs in each direction.

7. The partial discharge pattern recognition method based on time-frequency spatial spectrum fusion of multi-channel acoustic signals according to claim 6, characterized in that: The formula used to characterize the local energy distribution and spatial scale differences of discharge pulses in the time-frequency spatial fusion spectrum of partial discharge is as follows: ； In the formula: The number of discrete sampling points in the frequency domain. The number of discrete sampling points in the time domain. The optimal scale that best represents the global structural features; The formula for quantifying the brightness difference between the discharge pulse and the background in the time-frequency spatial fusion spectrum of partial discharge is as follows: ； Where: V is the standard deviation of the time-frequency spatial fusion spectrum of partial discharge, and K is the fourth central moment of the time-frequency spatial fusion spectrum of partial discharge; The formula used to reflect the main directional trend of texture in the time-frequency spatial fusion spectrum of partial discharge is as follows: ； In the formula: Let the center angle of the b-th interval be . dominant interval of the histogram The average center angle, H b Let be the cumulative magnitude of the b-th interval.

8. The partial discharge pattern recognition method based on time-frequency spatial spectrum fusion of multi-channel acoustic signals according to claim 7, characterized in that: In step 5, the sample features are divided into training set, validation set, and test set according to a specific ratio and fed into the classifier to complete the partial discharge pattern recognition.

Images