Power equipment abnormal sound positioning monitoring method and device based on high-throughput acoustic perception

Abstract

The application discloses a power equipment abnormal sound positioning monitoring method and device based on high-throughput acoustic perception. The method comprises the following steps: collecting multi-channel acoustic signals and performing clock synchronization calibration to construct aligned acoustic data sets; using the time difference of arrival mode to establish a propagation ray bundle, positioning a candidate sound source area through the intersection of the ray bundle, and forming an enhanced positioning area by beam focusing and strengthening on the candidate sound source area; converting the multi-channel acoustic signals to the frequency domain to obtain a frequency spectrum energy atlas, searching for an abnormal sound characteristic frequency band, establishing time-frequency positioning parameters, and generating a sound source position probability distribution according to the time-frequency positioning parameters; combining the phase distribution map and the sound source position probability distribution to eliminate the reflection path; performing sound pressure gradient vector field reconstruction on the spatial positioning data to form a three-dimensional sound field structure, searching for an energy peak value and matching a spectrum characteristic to determine the spatial position of the abnormal sound source and the abnormal sound type, and providing a monitoring means for intelligent operation and maintenance and fault early warning of power equipment.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, and in particular to a method and device for locating and monitoring abnormal noises in power equipment based on high-throughput acoustic sensing. Background Technology

[0002] Transformers, switchgear, busbars, and other power equipment are core components of the power grid system, and their operating status directly affects the safety and reliability of power supply. During long-term operation, these devices often generate various abnormal noises due to factors such as insulation aging, mechanical loosening, and partial discharge. Abnormal noises are an important indicator of internal equipment faults; timely and accurate location of the source and identification of the type of abnormal noise are crucial for preventing equipment failures and avoiding power grid accidents.

[0003] Traditional methods for monitoring abnormal noises in power equipment primarily rely on manual inspections and single-point acoustic sensor monitoring. Manual inspections are inefficient, heavily influenced by personnel experience, and struggle to achieve continuous monitoring and precise location. While single-point acoustic sensors can detect abnormal noise events, their lack of spatial localization capabilities makes it impossible to pinpoint the exact location and type of the noise source, or distinguish between direct and reflected sound interference. In recent years, although acoustic localization technologies based on sensor arrays have emerged, existing methods still have shortcomings in clock synchronization accuracy, multipath reflection suppression, and spectral feature extraction, resulting in low localization accuracy and high false alarm rates. Especially under complex electromagnetic environments and reflection path interference, localization performance deteriorates significantly. Therefore, a method is urgently needed to address at least one of the aforementioned problems. Summary of the Invention

[0004] This invention discloses a method and device for locating and monitoring abnormal noise sources in power equipment based on high-throughput acoustic sensing. Through techniques such as clock synchronization calibration of multi-channel acoustic signals, extraction of time difference of arrival patterns, location of propagating beam intersection, beam focusing enhancement, establishment of time-frequency positioning parameters, elimination of reflection paths, reconstruction of sound pressure gradient vector field, and search for energy peaks, it achieves accurate spatial location and type identification of abnormal noise sources in power equipment, providing a high-precision and high-reliability monitoring solution for intelligent operation and maintenance and fault early warning of power equipment.

[0005] The first aspect of this invention proposes a method for locating and monitoring abnormal noises in power equipment based on high-throughput acoustic sensing, comprising the following steps: Collect multi-channel acoustic signals and sensor array position information during the operation of power equipment, and perform clock synchronization calibration on the multi-channel acoustic signals to construct an aligned acoustic dataset; The time difference of arrival (TDOA) pattern is extracted from the aligned acoustic dataset. The TDOA pattern is used to establish a spatial propagation path to form a propagation ray beam. The intersection point of the propagation ray beam is used to generate candidate sound source regions. The candidate sound source regions are then enhanced to form an enhanced localization region. The multi-channel acoustic signal is converted to the frequency domain to obtain a spectral energy map. The abnormal noise characteristic frequency band is retrieved in the spectral energy map. The abnormal noise characteristic frequency band is combined with the sensor array position information to establish time-frequency positioning parameters. The sound source position probability distribution is generated based on the time-frequency positioning parameters. Sound pressure directionality analysis is performed on the enhanced positioning area to obtain the direction of the main sound beam. Array beam synthesis is performed around the direction of the main sound beam to form a superimposed sound field. Phase difference extraction is performed on the superimposed sound field to establish a phase distribution map. Reflection path elimination is performed by combining the phase distribution map with the probability distribution of the sound source position to generate spatial positioning data. Sound pressure change rate analysis is performed on the spatial positioning data to establish a sound pressure gradient vector field. Based on the sound pressure gradient vector field, the sound field around the sound source is reconstructed to form a three-dimensional sound field structure. The energy peak search is performed using the three-dimensional sound field structure to determine the spatial location and type of abnormal noise source.

[0006] A second aspect of the present invention provides a power equipment abnormal noise location and monitoring device based on high-throughput acoustic sensing, comprising: The data acquisition module is used to collect multi-channel acoustic signals and sensor array position information during the operation of power equipment, and to perform clock synchronization calibration on the multi-channel acoustic signals to construct an aligned acoustic dataset; The preliminary localization module is used to extract the time difference of arrival pattern from the aligned acoustic dataset, establish a spatial propagation path using the time difference of arrival pattern to form a propagation ray beam, perform intersection point localization on the propagation ray beam to generate candidate sound source regions, and enhance the candidate sound source regions to form an enhanced localization region. The frequency domain analysis module is used to convert the multi-channel acoustic signal to the frequency domain to obtain a spectral energy map, retrieve abnormal noise characteristic frequency bands in the spectral energy map, establish time-frequency positioning parameters by combining the abnormal noise characteristic frequency bands with the sensor array position information, and generate a sound source location probability distribution based on the time-frequency positioning parameters. The interference cancellation module is used to perform sound pressure directionality analysis to obtain the direction of the main sound beam in the enhanced positioning area, perform array beam synthesis around the direction of the main sound beam to form a superimposed sound field, extract the phase difference of the superimposed sound field to establish a phase distribution map, and combine the phase distribution map with the probability distribution of the sound source position to perform reflection path cancellation to generate spatial positioning data. The precise positioning module is used to perform sound pressure change rate analysis on the spatial positioning data to establish a sound pressure gradient vector field, reconstruct the sound field around the sound source based on the sound pressure gradient vector field to form a three-dimensional sound field structure, and use the three-dimensional sound field structure to perform energy peak search to determine the spatial location and type of abnormal noise source.

[0007] The beneficial effects of this invention are reflected in the following points: First, by identifying clock drift intervals from multi-channel acoustic signals and establishing synchronization anchors, time offset correction is implemented on multi-channel acoustic signals, solving the beamforming error problem caused by clock asynchrony in traditional multi-channel acquisition; furthermore, by identifying low-energy regions around candidate sound source areas, using low-energy regions to define the focusing boundary range, and generating spatial weighting coefficients based on beam focusing parameters, the limitations of traditional fixed beam methods in adapting to complex sound field environments are overcome, achieving adaptive beam focusing based on sound field characteristics, and avoiding positioning deviations caused by blind focusing. Secondly, frequency gradient analysis is performed on the frequency bands of abnormal noise to identify frequency jump locations. Spatial correlation is established using the frequency jump locations and sensor array location information, avoiding the limitation of traditional methods that only locate in a single dimension of time or frequency domain. A spatial search grid is established based on time-frequency positioning parameters and time-frequency response evaluation is performed. High-response clusters are identified from the response intensity sequence and probability mapping is performed based on confidence boundaries. A joint positioning mechanism combining time-domain arrival time difference and frequency-domain frequency jump is established, which solves the problem of insufficient robustness of single-dimensional positioning methods in complex abnormal noise scenarios. Finally, by performing phase gradient analysis on the phase distribution map to identify phase jump regions, extracting low-probability intervals from the probability distribution of sound source locations and tracing interference paths in reverse, and performing probability suppression processing on the interference source location set, the technical difficulty of traditional methods in distinguishing between direct and reflected paths is overcome. Furthermore, a three-dimensional sound field structure is formed by reconstructing the sound pressure gradient vector field, and energy peak search and spectral feature matching are performed using the three-dimensional sound field structure. A complete technical chain from spatial positioning to type identification is established, solving the deficiency of traditional methods that can only locate but not identify the type of abnormal noise. This provides equipment maintenance personnel with a comprehensive diagnostic basis that includes spatial location, equipment components, and abnormal noise type.

[0008] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0009] The accompanying drawings illustrate specific examples of the technical solutions described in this invention and, together with the detailed embodiments, form part of the specification, serving to explain the technical solutions, principles, and effects of this invention.

[0010] Unless otherwise specified, the same reference numerals in different figures represent the same or similar technical features, and different reference numerals may be used to represent the same or similar technical features.

[0011] Figure 1 This is a flowchart illustrating the method for locating and monitoring abnormal noises in power equipment based on high-throughput acoustic sensing, as described in this invention.

[0012] Figure 2 This is a structural block diagram of the power equipment abnormal noise location and monitoring device based on high-throughput acoustic sensing according to the present invention. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0014] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0015] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0016] The technical solutions of the embodiments of this application will be described below.

[0017] like Figure 1 As shown, this embodiment of the invention provides a method for locating and monitoring abnormal noises in power equipment based on high-throughput acoustic sensing, including the following steps S110-S150: Step S110: Collect multi-channel acoustic signals and sensor array position information during the operation of power equipment, and perform clock synchronization calibration on the multi-channel acoustic signals to construct an aligned acoustic dataset.

[0018] Specifically, multi-channel acoustic signals and sensor array location information are collected during the operation of power equipment. Acoustic sensor arrays are deployed around key equipment such as transformers, switchgear, and busbars in substations. The arrays adopt a distributed layout, with 16-32 sensor nodes deployed in each monitoring area. The sensor nodes use MEMS condenser microphones with a frequency response range of 50Hz-18kHz, sensitivity of -42dB±3dB, and a signal-to-noise ratio greater than 65dB. Each sensor node is connected to the data acquisition unit wirelessly or via wired connection. The acquisition unit synchronously records the acoustic signals of each channel at a sampling rate of 44.1kHz with a quantization accuracy of 24 bits. The spatial location information of the sensor array is recorded, including the three-dimensional coordinates of each sensor node, the distance between nodes, and the relative azimuth angle. For transformer monitoring scenarios, the sensor array is arranged in a circle centered on the transformer body, with an array radius of 2-3 meters and an adjacent sensor spacing of 0.4-0.6 meters, covering the main radiation surface of the transformer. For busbar monitoring scenarios, the sensor array is arranged linearly along the busbar direction, with a node spacing of 0.5-0.8 meters, and the array length covers the entire length of the monitored busbar segment. Acoustic signals generated during equipment operation are collected, including normal operating sounds, intermittent abnormal noises, and sudden discharge sounds. Each sensor node independently collects sound signals, forming multiple parallel acoustic data streams, which constitute a multi-channel acoustic signal. Simultaneously, the position coordinates of each sensor node are recorded, represented using the equipment coordinate system. The origin is set as the geometric center of the monitoring equipment, with the X-axis pointing east, the Y-axis pointing north, and the Z-axis vertically upward, with a coordinate accuracy of 0.01 meters. The sensor position coordinates, node numbers, and corresponding acoustic channel numbers are integrated into sensor array position information.

[0019] In some embodiments, the step of performing clock synchronization calibration on the multi-channel acoustic signals to construct an aligned acoustic dataset includes: identifying clock drift intervals from the multi-channel acoustic signals; establishing clock correction references for each channel using the clock drift intervals to generate synchronization anchors; performing time offset correction on the multi-channel acoustic signals based on the synchronization anchors to generate calibration sampling data; and constructing an aligned acoustic dataset based on the calibration sampling data.

[0020] Clock drift intervals are identified from multi-channel acoustic signals. Timestamp information for each channel is extracted, recording the acquisition time for each sampling point with millisecond precision. The increasing pattern of the timestamp sequence for each channel is analyzed. Theoretically, the timestamp interval between adjacent sampling points should equal the sampling period; for a 44.1kHz sampling rate, the theoretical interval is 0.0227 milliseconds. The deviation between the actual and theoretical timestamp intervals for each channel is calculated; the deviation equals the actual interval minus the theoretical interval. The cumulative value of the deviation sequence is statistically analyzed; the cumulative deviation reflects the degree of clock drift. When the cumulative deviation of a channel exceeds one sampling period (0.0227 milliseconds), it indicates clock drift in that channel. The point at which the cumulative deviation begins to exceed a threshold is identified as the drift start point, and the point at which the cumulative deviation stops increasing or begins to decrease is identified as the drift end point. The time interval between the drift start point and the drift end point is defined as the clock drift interval. For the 16-channel array, during a continuous 10-second acquisition process, the cumulative deviation of channel 3 increased from 0 to 0.035 milliseconds between seconds 2.8 and 5.6; this period was identified as the clock drift interval. The cumulative deviation of channel 7 reached 0.028 milliseconds between seconds 6.2 and 8.9, also marked as a drift interval. The start and end times of the clock drift interval and the peak value of the cumulative deviation for each channel were recorded.

[0021] A clock correction benchmark for each channel is established using clock drift intervals to generate synchronization anchor points. Based on the identified clock drift intervals, the density of time markers is increased within the drift intervals of channels exhibiting drift to improve correction accuracy. The channel with the best clock stability is selected as the reference channel from the multi-channel acoustic signal. The selection criteria for the reference channel are the channel with the smallest cumulative deviation and no clock drift intervals. For the 16-channel array, channel 1, whose cumulative deviation remains within 0.005 milliseconds and has no drift intervals, is selected as the reference channel. Several time markers are set on the time axis of the reference channel, with a marker interval of 1 second. For channels with clock drift intervals, the marker interval is shortened to 0.5 seconds within their drift intervals. For a 10-second acquisition duration, basic markers are set at positions 0, 1, 2, and 10 seconds. The sampling index and corresponding timestamp value of the reference channel at each time marker position are extracted. For channels with clock drift intervals, the difference between the timestamp at each time marker position and the timestamp of the reference channel is analyzed. Channel 3's timestamp at the 2-second marker is 2.0002 seconds, the reference channel's is 2.0000 seconds, and the difference is 0.0002 seconds; at the 5-second marker, the timestamp is 5.0032 seconds, the reference channel's is 5.0000 seconds, and the difference is 0.0032 seconds. The positions of each timestamp and their corresponding timestamp differences are defined as the clock calibration reference for that channel. A clock calibration reference is established for each channel, and each channel's calibration reference contains multiple timestamps and their timestamp deviation sequences. Synchronization anchors are set at each timestamp position, and each synchronization anchor contains three elements: a reference timestamp, a channel timestamp, and a deviation amount. The 5-second synchronization anchor for channel 3 has a reference timestamp of 5.0000 seconds, a channel timestamp of 5.0032 seconds, and a deviation amount of 0.0032 seconds.

[0022] Calibration sampling data is generated by performing time offset correction on multi-channel acoustic signals based on synchronization anchor points. Time offset interpolation correction is performed on the data segments of each channel's multi-channel acoustic signals between adjacent synchronization anchor points, based on the synchronization anchor points. For the data segment of channel 3 between the 4th and 5th seconds, the initial synchronization anchor point deviation is 0.0018 seconds, and the final synchronization anchor point deviation is 0.0032 seconds, with the deviation increasing linearly. The deviation of each sampling point within this data segment is calculated; the deviation of sampling point t is obtained by linear interpolation of the initial and final deviations. The corrected timestamp is obtained by subtracting the corresponding deviation from the timestamp of each sampling point. For the sampling point of channel 3 at the 4.5-second position, the original timestamp is 4.5025 seconds, the interpolation deviation is 0.0025 seconds, and the corrected timestamp is 4.5000 seconds, aligned with the reference channel. The corrected timestamp sequence is resampled to ensure that each channel has corresponding sampling data at the same timestamp position. A cubic spline interpolation algorithm is used to resample the acoustic signal waveform, generating interpolated sampling points at the corrected timestamp. For channel 3, at the corrected timestamp of 4.5000 seconds, the sampling value at that position is generated by interpolating the waveform data from the original sampling points of 4.5018 seconds and 4.5041 seconds. The same time offset correction and resampling operation is performed on the multi-channel acoustic signals of all channels to generate time-axis aligned sampling data. The corrected sampling data of each channel has synchronized sampling points at the same timestamp; these corrected sampling data are defined as calibration sampling data.

[0023] An aligned acoustic dataset was constructed based on calibration sampling data. The calibration sampling data for each channel was organized into a matrix according to channel number and time order. Rows of the matrix correspond to different time sampling points, and columns correspond to different channels. Matrix element values ​​are the acoustic signal amplitudes at the corresponding time and channel position. For a 16-channel, 10-second duration, 44.1kHz sampling rate acquisition, the matrix size is 441,000 rows × 16 columns. Sensor array position information was added to the matrix, and spatial coordinate labels for sensor nodes were added to each column. Channel 1 corresponds to sensor node coordinates X=0.50m, Y=0.00m, Z=1.20m, and this coordinate information serves as the metadata label for the Channel 1 column. The time synchronization accuracy of the calibration sampling data was verified by calculating the phase difference between sampling points at the same timestamp for each channel. The signal waveforms of each channel at the same moment were extracted, and the cross-correlation function between each pair of channels was calculated. The time delay corresponding to the peak position of the cross-correlation function reflects the synchronization error. The time delay error between calibrated channels should be less than 0.5 sampling periods, i.e., 0.0114 milliseconds. For the 16-channel array, the time delay error between any two channels is within the range of 0.008-0.011 milliseconds, meeting the synchronization accuracy requirements. The validated calibration sampling data matrix and its sensor location metadata are integrated into an aligned acoustic dataset, which includes time-aligned multi-channel acoustic signals, sensor spatial location information, and timestamp indexes.

[0024] Step S120: Extract the time difference of arrival (TDOA) pattern from the aligned acoustic dataset, establish a spatial propagation path using the TDOA pattern to form a propagation ray beam, perform intersection point localization on the propagation ray beam to generate candidate sound source regions, and enhance the candidate sound source regions to form an enhanced localization region.

[0025] Arrival time difference patterns are extracted from the aligned acoustic dataset. The acoustic signals of any two channels are selected from the aligned acoustic dataset, and the cross-correlation function between the two channels is calculated. The cross-correlation function reflects the similarity of two signals at different time offsets; the time offset corresponding to the peak of the function is the time difference between the two channels receiving the same acoustic event. For a 16-channel array, the signals from channel 1 and channel 5 are selected for cross-correlation calculation. The cross-correlation function peaks at a time delay of 0.85 milliseconds, indicating that the sound signal received by channel 5 is 0.85 milliseconds later than that received by channel 1. Cross-correlation calculations are performed on all channel pair combinations; the 16-channel array has 120 channel pair combinations, yielding 120 time delay values. The time delay values ​​of each channel pair are integrated with the corresponding channel pair identifiers into a time delay matrix. This matrix is ​​a 16×16 symmetric matrix; diagonal elements of 0 indicate that the time delay of the same channel is zero, and off-diagonal elements represent the time delay of the corresponding channel pair. To identify the consistency pattern in the time delay matrix, for acoustic events from the same sound source, the time delay of each channel pair should satisfy the spatial geometric consistency constraint. The time delay of channels 1-5 is 0.85 ms, the time delay of channels 1-9 is 1.32 ms, and the time delay of channels 5-9 is 0.47 ms. The three time delays satisfy the geometric consistency relationship. These combinations of time delays that satisfy geometric consistency are the time difference of arrival patterns.

[0026] A spatial propagation path is established using the time-of-arrival (TOA) pattern to form a propagation ray beam. The time delay values ​​and corresponding sensor position information for each channel pair in the TOA pattern are extracted. For channels 1 and 5, the position coordinates of channel 1 are X=0.50 m, Y=0.00 m, Z=1.20 m, and the position coordinates of channel 5 are X=-0.30 m, Y=0.70 m, Z=1.20 m, with a time delay of 0.85 milliseconds. Based on the time delay and the speed of sound (340 m / s), the distance difference between the two sensors and the sound source is calculated to be 0.289 m. In three-dimensional space, the trajectory of points with a fixed distance difference to the two sensors forms a hyperboloid, and the two foci of the hyperboloid represent the positions of the two sensors. For the time delays of channels 1-5, the sound source position lies on the hyperboloid with channels 1 and 5 as foci and a distance difference of 0.289 meters. Corresponding hyperboloid constraints are generated for all channel pairs; a total of 120 hyperboloids are generated for the 16-channel array. A spatial propagation path is established based on hyperboloid constraints. Spatial rays are emitted from each sensor position towards the corresponding hyperboloid. The ray direction is perpendicular to the hyperboloid and points in the direction where the distance difference decreases. The ray emitted by channel 1 towards its corresponding hyperboloid is in the direction from channel 1 towards the possible location of the sound source. The rays emitted by each sensor represent the spatial propagation path from the sound source to each sensor. The set of all spatial propagation paths emitted by all sensors constitutes the propagation ray beam. For transformer noise, the propagation ray beam shows a clear convergence trend in the area above the transformer body, while for busbar noise, the ray beam converges at a certain joint location along the busbar's direction.

[0027] Candidate sound source regions are generated by locating the intersection points of propagating ray beams. Intersection point location is achieved by statistically analyzing ray intersection density. A three-dimensional grid is divided with a resolution of 0.1 m × 0.1 m × 0.1 m, and the monitoring area is divided into several grid cells. The number of rays passing through each grid cell is counted; the number of rays reflects the probability of that location being a sound source. For a given grid cell, if 12 rays pass through, it indicates that the location matches the time delay constraints of the 12 sensors, and the probability of it being a sound source is relatively high. A ray density threshold of 30% of the total number of rays is set. For a 16-channel array with a total of 16 rays, the threshold is 5 rays. Grid cells with ray densities exceeding the threshold are considered candidate sound source locations. For locating the intersection point of transformer noise, the grid cell ray density near X=0.20m, Y=0.15m, Z=2.50m above the transformer body reached 11 ray lines, exceeding the threshold of 5 ray lines; the grid cell ray density near the bus joint location reached 8 ray lines. All grid cells with ray densities exceeding the threshold were aggregated into connected regions. For transformer noise, the connected region formed by the dense convergence of rays has a coverage area of ​​approximately 0.3m × 0.3m × 0.2m, and this connected region is the candidate sound source region. The center coordinates, coverage area, and peak ray density of each connected region are extracted as the results of intersection point location.

[0028] In some embodiments, enhancing the candidate sound source region to form an enhanced positioning region includes: identifying low-energy regions around the candidate sound source region; using the low-energy regions to define the focusing boundary range and determine beam focusing parameters; generating spatial weighting coefficients based on the beam focusing parameters; and using the spatial weighting coefficients to focus energy in the candidate sound source region to form an enhanced positioning region.

[0029] Low-sound-energy regions are identified around the candidate sound source area. A search range is defined around the candidate sound source area, extending 0.5-1.0 meters outward from the candidate area boundary. Within the search range, the sound energy intensity is calculated for each grid cell, and the sound pressure level at that grid cell location is estimated using a beamforming algorithm. Beamforming calculations are performed on all grid cells within the search range to obtain the sound pressure level values ​​for each cell. For transformer noise localization, the sound pressure level at the center of the candidate sound source area (X=0.20m, Y=0.15m, Z=2.50m) is 78dB, and the sound pressure levels of the surrounding grid cells within the search range range from 45-75dB. Grid cells with sound pressure levels below 50% of the average sound pressure level of the candidate sound source area are identified. The average sound pressure level of the candidate area is 75dB, and the 50% threshold is 37.5dB. Cells within the search range with sound pressure levels below 37.5dB are considered low-sound-energy cells. To the north and east of the candidate sound source area, the sound pressure level of the grid cells within a 0.6-0.9 meter radius of the center drops to 30-35 dB, forming a low-sound-energy accumulation zone, which is the low-sound-energy region. The appearance of low-sound-energy regions is usually caused by factors such as equipment obstruction, sound wave reflection cancellation, or sound source directivity. The low-sound-energy region to the north may correspond to the obstruction effect of the transformer cooler.

[0030] The beam focusing parameters are determined by defining the focusing boundary range using the low-energy region. The spatial orientation of the low-energy region relative to the candidate sound source region is analyzed, and the azimuth and elevation angles from the center of the candidate region to the center of each low-energy region are calculated. For transformer noise, the azimuth angle from the center of the candidate region to the north of the low-energy region is 15 degrees, and the elevation angle is -8 degrees; the azimuth angle to the east of the low-energy region is 85 degrees, and the elevation angle is -5 degrees. These azimuth and elevation angles are marked as suppression directions. The angular range of all suppression directions is statistically analyzed, and the blank areas between the suppression directions are identified. The angular range of 15-85 degrees between the north suppression direction of 15 degrees and the east suppression direction of 85 degrees is the non-suppressed region, which is the focusing boundary range. The focusing boundary range covers an azimuth angle of 15-85 degrees with an angular width of 70 degrees, and an elevation angle of -20 to 10 degrees with an angular width of 30 degrees. Based on the focusing boundary range and the spatial layout of the sensor array, the parameters required for beamforming are calculated. The main lobe width is set to 80% of the focusing boundary range angle width, i.e., 56 degrees. The side lobe suppression depth is set to -25dB. The null positions are set at 15 degrees and 85 degrees in the suppression direction. These parameters constitute the beam focusing parameters. The reasonable setting of parameters can enhance the response in the dominant direction and reduce interference in the suppression direction.

[0031] For example, generating spatial weighting coefficients based on the beam focusing parameters includes: identifying array response null regions from the beam focusing parameters; defining an effective focusing angle range using the array response null regions; performing weight optimization on the effective focusing angle range to generate an initial weighting vector; and performing energy normalization on the initial weighting vector to generate spatial weighting coefficients.

[0032] Identify null regions in the array response from beam focusing parameters. Extract null location information from the beam focusing parameters. For transformer noise localization, the beam focusing parameters specify null locations at azimuth angles of 15 degrees and 85 degrees. These null locations correspond to spatial angles where the response in the beamforming pattern is zero or close to zero. Based on the null locations and sensor array geometry, calculate the theoretical response characteristics of the array at these angles. At the null locations, the signals received by each sensor, after weighted summation, should cancel each other out, resulting in a total response close to zero. Analyze the angular range around the null locations to identify angular intervals where the response is below -30dB of the main lobe peak response. For a 15-degree azimuth null, the angular range with a response below -30dB is 10-20 degrees; this interval is the null region. For an 85-degree azimuth null, the null region ranges are 80-90 degrees; these low-response angular intervals are the array response null regions. The width of the array response null region is affected by the sensor array aperture and frequency. For high-frequency noise, the null region is typically narrower and has a deeper suppression depth.

[0033] The effective focusing angle range is defined using the null region of the array response. Analyzing the spatial distribution of the null region, the 20-80 degree range (between the 10-20 degree azimuth null region and the 80-90 degree null region) is considered a non-null region. This non-null region constitutes the effective focusing angle range. Within this range, the beam response is not constrained by nulls and can be freely adjusted to achieve optimal focusing. For transformer noise, the effective focusing angle range is 20-80 degrees azimuth and -15 to 5 degrees elevation. Verifying whether the effective focusing angle range covers the main radiation direction of the candidate sound source area, the candidate area's center azimuth of 45 degrees and elevation of -3 degrees fall within the effective focusing range of 20-80 degrees and -15 to 5 degrees, respectively, meeting the coverage requirements. The angular width of the effective focusing angle range and the covered spatial solid angle are calculated. The azimuth range width is 60 degrees, and the elevation range width is 20 degrees, corresponding to a spatial solid angle of approximately 0.36 spheroids. Analyzing the beam control capability of the sensor array within its effective focusing angle range, the 16-channel array can achieve a minimum beamwidth of approximately 12-15 degrees within this range, supporting precise focusing and localization of abnormal noises. For busbar noise localization scenarios, the effective focusing angle range is typically narrower, with an azimuth width of approximately 40 degrees and an elevation width of approximately 15 degrees, reflecting the stronger directional radiation characteristics of busbar noises.

[0034] An initial weighted vector is generated by weight optimization within the effective focusing angle range. A beamforming weight optimization model is established, with the optimization objective being to obtain the maximum response within the effective focusing angle range and the minimum response within the null region of the array response. Optimization constraints are set: the main lobe points to the center of the candidate sound source region at an azimuth angle of 45 degrees and an elevation angle of -3 degrees; the main lobe 3dB beamwidth does not exceed 15 degrees; and the null region response depth does not exceed -30dB. The minimum variance distortionless response (MVDR) algorithm is used to solve for the optimal weighted vector. This algorithm minimizes the response in other directions while maintaining a constant pointing direction response. For a 16-channel array, a 16-dimensional weighted vector is obtained: channel 1 has a weight amplitude of 0.85 and a phase of 0 degrees; channel 5 has a weight amplitude of 0.92 and a phase of -25 degrees; and channel 9 has a weight amplitude of 0.78 and a phase of -48 degrees. This 16-dimensional weighted vector is the initial weighted vector. The beam pattern corresponding to the weight vector was verified. The main lobe points to a 45-degree azimuth angle and a -3-degree elevation angle. The 3dB width of the main lobe at 12 degrees meets the requirement of not exceeding 15 degrees. The position responses at 15 degrees and 85 degrees in the null region are -32dB and -35dB respectively, which meet the depth requirements. The weight optimization successfully achieved the dual objectives of focusing and suppression.

[0035] Energy normalization is applied to the initial weighting vector to generate spatial weighting coefficients. The energy sum of each element in the initial weighting vector is calculated; the energy sum is equal to the sum of the squares of the amplitudes of each element. For a 16-channel array, the energy sum is 12.8. Each element of the initial weighting vector is divided by the square root of the energy sum to make the energy sum of the normalized weight vector equal to 1. The original weight amplitude of channel 1 is 0.85, and after normalization, the amplitude is 0.238, while the phase remains unchanged at 0 degrees. The original weight amplitude of channel 5 is 0.92, and after normalization, the amplitude is 0.257, while the phase remains at -25 degrees. The original weight amplitude of channel 9 is 0.78, and after normalization, the amplitude is 0.218, while the phase remains at -48 degrees. The normalized 16-dimensional weight vector is the spatial weighting coefficient. The beam pattern of the normalized weight vector is verified; the main lobe pointing and null position are consistent with those before normalization, with only the overall response amplitude scaling. The normalization operation ensures energy consistency when applying the weighting coefficients at different frequencies or at different times.

[0036] Spatial weighting coefficients are used to concentrate energy in candidate sound source regions to form enhanced localization regions. For each grid cell within the candidate sound source region, the 16-channel acoustic signals are weighted and summed using spatial weighting coefficients. The weighting summation formula is as follows: Where y(r) is the focused response intensity at position r, Wn is the spatial weighting coefficient of the nth channel, Xn(r) is the acoustic signal amplitude of the nth channel at position r, and N is the number of channels; in this embodiment, N=16. The focused response intensity at the location of this grid cell is calculated. For transformer noise, the center positions of the candidate sound source region are X=0.20m, Y=0.15m, and Z=2.50m. After applying the spatial weighting coefficient, the focused response intensity increases from the original value of 68dB to 82dB, an energy enhancement of 14dB. The same weighting process is performed on all grid cells within the candidate sound source region to generate a response distribution map after energy aggregation. Thresholding is applied to the response distribution map, and grid cells with focused response intensities exceeding the maximum value of -6dB are extracted to form enhanced localization regions. For transformer noise, the maximum response is 82dB, the -6dB threshold is 76dB, and grid cells with responses exceeding 76dB constitute enhanced localization regions. The enhanced positioning area covers approximately 0.15m × 0.15m × 0.10m, which is about 75% smaller than the candidate sound source area (0.3m × 0.3m × 0.2m), resulting in a significant improvement in positioning accuracy. The enhanced positioning area achieves finer focusing on the candidate sound source area through spatial weighting coefficients, improving the accuracy and resolution of sound source localization.

[0037] Step S130: Convert the multi-channel acoustic signal to the frequency domain to obtain the spectral energy map, search for the abnormal noise characteristic frequency band in the spectral energy map, combine the abnormal noise characteristic frequency band with the sensor array position information to establish time-frequency positioning parameters, and generate the sound source position probability distribution based on the time-frequency positioning parameters.

[0038] Specifically, the multi-channel acoustic signals are converted to the frequency domain to obtain spectral energy maps. Time-domain waveforms of the multi-channel acoustic signals are extracted from the aligned acoustic dataset, and a Short-Time Fourier Transform (STFT) is performed on each channel. The STFT time window length is set to 4096 sampling points, corresponding to a time resolution of 92.9 milliseconds at a sampling rate of 44.1 kHz. A Hamming window is used to reduce spectral leakage. The overlap rate between adjacent windows is set to 75%, and 1024 sampling points are slid in each step to ensure the continuity of time-frequency analysis. A Fast Fourier Transform (FFT) is performed on the signal within each time window to obtain the spectral distribution of that window, with a spectral resolution of 10.8 Hz. The energy density of each frequency component is calculated; the energy density is equal to the square of the spectral amplitude. For a 16-channel array and 10-second acquisition time, each channel generates approximately 108 time frames and a time-frequency matrix of 2048 frequency channels. Arrange the time-frequency matrices of each channel according to the channel number to construct a three-dimensional time-frequency energy cube. The first dimension of the cube corresponds to 108 time frames, the second dimension corresponds to 2048 frequency channels, and the third dimension corresponds to 16 acoustic channels. This three-dimensional cube is the spectral energy map.

[0039] The abnormal noise characteristic frequency bands were retrieved from the spectral energy map. The energy time series of each frequency channel in the spectral energy map were analyzed, and the time average and variance of the energy of each frequency channel were calculated. For frequency channels occupied by normal operation noise, the energy time series was relatively stable with a small variance. For frequency channels occupied by abnormal noise events, the energy time series showed intermittent peaks with a large variance. An energy variance threshold of 3 times the average variance was set to identify frequency channels whose energy variance exceeded the threshold. The energy variance of transformer partial discharge in the frequency range of 6-12kHz exceeded the threshold, and the energy variance of switchgear mechanical loosening in the frequency range of 800-1500Hz was significant. The frequency of energy peak occurrences for each frequency channel was statistically analyzed. The peak value was defined as the moment when the energy exceeded the average value of the frequency channel plus twice the standard deviation. Within a 10-second acquisition time, the peak frequency of partial discharge in the 6-12kHz frequency band was 15-25 times, while the peak frequency of normal noise was only 3-5 times. Frequency channels with peak frequencies exceeding 10 were identified. Continuous frequency channels with energy variance exceeding a threshold and peak frequencies exceeding 10 were aggregated into frequency bands. The bandwidth of these bands was required to be greater than 200Hz to exclude narrowband interference. These aggregated frequency bands constituted the characteristic frequency bands for abnormal noises. The characteristic frequency band for partial discharge abnormal noises covered 5.8-12.3kHz with a bandwidth of 6.5kHz, while the characteristic frequency band for mechanical loosening abnormal noises covered 750-1620Hz with a bandwidth of 870Hz. Parameters such as the start frequency, end frequency, center frequency, and bandwidth were recorded for each frequency band.

[0040] In some embodiments, establishing time-frequency positioning parameters by combining the abnormal noise characteristic frequency band with the sensor array position information includes: performing frequency gradient analysis on the abnormal noise characteristic frequency band to identify frequency jump positions; using the frequency jump positions and the sensor array position information to spatially correlate to form a frequency-position mapping table; extracting spatial response features from the frequency-position mapping table to generate a positioning reference vector; and establishing time-frequency positioning parameters based on the positioning reference vector.

[0041] Frequency gradient analysis was performed on the characteristic frequency band of the abnormal noise to identify frequency jump locations. Frequency-energy distribution curves were extracted for the characteristic frequency band of the abnormal noise, with frequency on the horizontal axis and average energy of the frequency channel on the vertical axis. For the characteristic frequency band of partial discharge abnormal noise (5.8-12.3kHz), average energy curves for each frequency channel within this band were plotted. The first derivative of the frequency-energy curve was calculated; the derivative reflects the rate of change of energy with frequency, i.e., the frequency gradient. A positive frequency gradient indicates that energy increases with increasing frequency, while a negative gradient indicates that energy decreases. In the partial discharge abnormal noise, the energy suddenly increases from -25dB to -15dB at frequency 8.2kHz, with the frequency gradient reaching a peak of 120dB / kHz at this location; at frequency 10.5kHz, the energy drops sharply from -12dB to -28dB, with a negative peak frequency gradient of -150dB / kHz. A frequency gradient threshold of ±80dB / kHz was set; frequency jump locations were defined as those where the absolute value of the gradient exceeded this threshold. For partial discharge noises, frequency jumps occur at 8.2kHz and 10.5kHz; for mechanical loosening noises, frequency jumps occur at 950Hz and 1350Hz. Analyzing the energy jump direction and amplitude at each frequency jump location, the energy jumps up by 10dB at 8.2kHz and down by 16dB at 10.5kHz. These jump characteristics mark key turning points in the spectral structure of the noises.

[0042] A frequency-position mapping table is formed by spatially associating frequency jump locations with sensor array position information. This spatial association establishes the correspondence between frequency jump locations and sensor positions. The energy response differences of each sensor channel at the frequency jump location are extracted. For the 8.2kHz frequency jump location, the energy values ​​of 16 sensor channels at that frequency are analyzed. The deviation of each channel's energy at that frequency from the average energy of all channels is calculated; channel 9 has an energy deviation of +6dB, and channel 13 has a deviation of -7dB. The spatial coordinates of each channel are extracted from the sensor array position information, and the spatial coordinates of each channel are paired with the energy deviation of that channel at the frequency jump location to achieve spatial association. For the 8.2kHz frequency jump location, 16 frequency-position-energy triplets are generated, each containing frequency, sensor spatial coordinates, and energy deviation. The same spatial association operation is performed on all frequency jump locations. For partial discharge abnormalities, there are 2 jump locations and 16 sensors, generating a total of 32 triplets. These triplets are arranged according to frequency and sensor number to form a two-dimensional frequency-position mapping table. The rows of the mapping table correspond to different frequency transition positions, the columns correspond to different sensor channels, and the table elements are the energy deviation values ​​of the corresponding frequencies and channels. Through this table, the response distribution patterns of each frequency component in space can be observed intuitively.

[0043] Spatial response features are extracted from the frequency-position mapping table to generate a positioning reference vector. The spatial response patterns at each frequency jump location in the frequency-position mapping table are analyzed. For the 8.2kHz frequency jump location, this row contains energy deviation values ​​for 16 sensors. Sensor locations with positive energy deviations are identified: channel 9 has an energy deviation of +6dB, channel 11 +4dB, and channel 14 +3dB, indicating strong responses from these three channels. The spatial coordinates of the strong-response sensors are extracted: channel 9 position X=-0.80m, Y=-0.50m, Z=1.20m; channel 11 position X=-0.50m, Y=-0.80m, Z=1.20m; channel 14 position X=0.30m, Y=-0.90m, Z=1.20m. The geometric center coordinates of these strong-response sensor locations are calculated as X=-0.33m, Y=-0.73m, Z=1.20m. Simultaneously, weak-response sensors with negative energy deviations were identified: channel 1 deviation -3dB, channel 5 deviation -1dB, and channel 13 deviation -7dB. The geometric center coordinates of the weak-response sensors were calculated as X = 0.07 m, Y = 0.23 m, and Z = 1.20 m. The direction of the line connecting the strong and weak response centers was calculated. The unit direction vector from the weak response region to the strong response region was X = -0.38, Y = -0.92, and Z = 0. This direction vector indicates the main propagation or radiation direction of this frequency component. Spatial response features were extracted from all frequency jump locations. The strong response center coordinates and direction vectors of each frequency were arranged in frequency order. Partial discharge noise has two frequency jump locations, forming a 12-dimensional vector containing two 3-dimensional center coordinates and two 3-dimensional direction vectors. This vector serves as the location reference vector, containing location clues for the noise in both spatial and frequency dimensions.

[0044] Time-frequency positioning parameters are established based on the positioning reference vector. The strong response center coordinates of each frequency transition location are extracted from the positioning reference vector. For partial discharge noise, the center coordinates at 8.2kHz are X=-0.33m, Y=-0.73m, Z=1.20m, and at 10.5kHz are X=-0.25m, Y=-0.68m, Z=1.20m. The average value of the center coordinates at each frequency is calculated to obtain the comprehensive spatial positioning reference point X=-0.29m, Y=-0.71m, Z=1.20m. Direction vectors for each frequency are extracted: at 8.2kHz, the direction vectors are X=-0.38, Y=-0.92, Z=0, and at 10.5kHz, the direction vectors are X=-0.42, Y=-0.91, Z=0. The average direction vector is calculated to be X=-0.40, Y=-0.92, Z=0. The energy jump amplitude at each frequency transition location was analyzed. The jump amplitude was 10dB at 8.2kHz and 16dB at 10.5kHz. The jump amplitude was used as the weighting factor for the frequency in the localization. The center frequency of 9.05kHz and the bandwidth of 6.5kHz were extracted from the abnormal noise characteristic frequency band. The spatial localization reference point coordinates, average direction vector, the frequency transition locations and their weighting factors, the abnormal noise center frequency and bandwidth, and other information were collected into time-frequency localization parameters. These parameters integrate the spatial dimension of position and direction, the frequency dimension of frequency and bandwidth, and the weighting dimension of jump amplitude, forming a complete input basis for subsequent sound source probability distribution calculations.

[0045] In some embodiments, generating a sound source location probability distribution based on the time-frequency positioning parameters includes: establishing a spatial search grid based on the time-frequency positioning parameters to divide candidate location points; performing time-frequency response evaluation on the candidate location points to generate a response intensity sequence; identifying high-response clustering regions from the response intensity sequence to determine confidence boundaries; and performing probability mapping on the response intensity sequence based on the confidence boundaries to generate a sound source location probability distribution.

[0046] Candidate location points are divided into a spatial search grid based on time-frequency positioning parameters. The coordinates of a spatial positioning reference point (X=-0.29m, Y=-0.71m, Z=1.20m) are extracted from the time-frequency positioning parameters. The spatial search area is defined with this reference point as the center. On the XY plane, the search area is a circular region with a radius of 1.5m centered on the reference point. On the Z direction, it extends 0.5m above and below the reference point, covering approximately 7 square meters in the XY plane and 1 meter in the Z direction. A three-dimensional grid is established within the search area, with a grid resolution of 0.05m × 0.05m × 0.1m. For the circular search area with a radius of 1.5m, the XY plane is divided into approximately 60×60 grids, and the Z direction is divided into 10 layers, resulting in a total of approximately 36,000 grid cells. The center coordinates of each grid cell constitute the set of candidate location points. The average direction vectors X=-0.40, Y=-0.92, and Z=0 are extracted from the time-frequency positioning parameters. The position vectors of each candidate point relative to the reference point are calculated. For example, the position vectors of a candidate point with X=-0.30 meters and Y=-0.70 meters relative to the reference point are X=-0.01 meters, Y=0.01 meters, and Z=0.05 meters. The angle between the position vector and the average direction vector is calculated. Candidate points with an angle less than 45 degrees are located in the dominant direction and are marked as priority search points. The set of candidate points covers the entire spatial range where the abnormal noise may exist.

[0047] A time-frequency response evaluation is performed on candidate locations to generate a response intensity sequence. Each candidate location is assumed to be a sound source location, and the distance and time delay from that location to each sensor are calculated. Frequency jump locations and their weighting factors are extracted from the time-frequency localization parameters; for partial discharge noise, the weight is 10 for 8.2kHz and 16 for 10.5kHz. For each frequency jump location, the energy values ​​of each sensor at that frequency are extracted from the spectral energy map. Based on the time delay vector from the candidate point to each sensor, the energy at that frequency is time-delay compensated and accumulated, ensuring that the signal from the candidate point is in phase and superimposed at each sensor. After time-delay compensation and accumulation for all 16 sensors, the response intensity of the candidate point at each frequency is calculated. The response intensity at each frequency is multiplied by its corresponding weighting factor and then summed to obtain the comprehensive response intensity of the candidate point. The same time-frequency response evaluation is performed on all candidate locations. Each candidate point generates a comprehensive response intensity value. The spatial coordinates of each candidate point are paired with their corresponding response intensities and arranged according to the candidate point index to form a response intensity sequence. The magnitude of the values ​​in the sequence reflects the strength of the probability that the corresponding spatial location is a sound source.

[0048] Confidence boundaries were determined by identifying high-response clusters from the response intensity sequence. The distribution characteristics of all response intensity values ​​in the sequence were statistically analyzed. For partial discharge anomalies, the maximum response intensity was -5dB, the minimum was -45dB, the average was -28dB, and the standard deviation was 8dB. A high-response threshold was set to the average plus one standard deviation, i.e., -20dB. Approximately 1800 candidate locations with response intensities exceeding -20dB were identified, accounting for 5% of the total candidate points. The spatial coordinates of these high-response candidate points were extracted. Spatially adjacent high-response candidate points were identified, defined as a distance of less than 0.15 meters in the XY plane and less than 0.2 meters in the Z direction. Adjacent high-response candidate points were then aggregated into connected regions. High-response candidate points of partial discharge abnormal noise aggregated into three connected regions. The largest connected region contained 1200 candidate points, with an average response intensity of -12dB and a spatial range of approximately 0.4m × 0.4m × 0.3m. This region is the high-response clustering area, with its center at X = -0.32m, Y = -0.74m, and Z = 1.25m. The boundary coordinates of the high-response clustering area were calculated, and the minimum response intensity within the boundary, -18dB, was extracted. This value is the confidence boundary threshold, marking the edge of the high-confidence region of the sound source location.

[0049] A probability distribution of sound source locations is generated by probabilistic mapping of the response intensity sequence based on confidence boundaries. The response intensities of each candidate location point in the response intensity sequence are normalized and mapped, converting the response intensities into probability values ​​to construct a preliminary probability distribution of sound source locations. Mapping rules are set: candidate points with a response intensity equal to the confidence boundary -18dB have a probability of 0.5; candidate points with a response intensity higher than the confidence boundary have a probability greater than 0.5; and candidate points with a response intensity lower than the confidence boundary have a probability less than 0.5. The Sigmoid function is used for probability mapping, with its center set at the confidence boundary -18dB. After mapping, candidate points with higher response intensities have a probability close to 1, while candidate points with lower response intensities have a probability close to 0. After performing probability mapping on all candidate location points, the spatial coordinates of each candidate point are paired with its probability value to construct a three-dimensional probability field. In the probability distribution of sound source locations, contour plots of the probability distribution are drawn at the Z-coordinate positions of each layer in the XY plane, and the probability peak positions of each Z-layer are extracted. For partial discharge noises, the probability peak location is concentrated in the range of Z=1.15 m to 1.35 m, and approximately in the region at the center of the XY plane at X=-0.31 m and Y=-0.73 m. This three-dimensional spatial probability field is the probability distribution of the sound source location, which quantifies the probability of the sound source in each spatial location in probabilistic form, providing a reliability assessment for noise localization.

[0050] Step S140: Perform sound pressure directionality analysis on the enhanced positioning area to obtain the direction of the main sound beam, perform array beam synthesis around the direction of the main sound beam to form a superimposed sound field, extract the phase difference of the superimposed sound field to establish a phase distribution map, and combine the phase distribution map with the probability distribution of the sound source position to perform reflection path elimination to generate spatial positioning data.

[0051] Specifically, sound pressure directionality analysis is performed on the enhanced positioning area to obtain the direction of the main sound beam. The spatial range and center coordinates of the enhanced positioning area are extracted. For transformer partial discharge noise, the center position of the enhanced positioning area is X=-0.32m, Y=-0.74m, Z=1.25m, with a spatial range of approximately 0.15m×0.15m×0.10m. Several test points are set within the enhanced positioning area, evenly distributed at 0.05m intervals, covering the three-dimensional space of the area. Beamforming calculations are performed on each test point to evaluate the sound pressure response intensity in different directions when each test point is used as a sound source. The center point of the enhanced positioning area is selected as the reference point, and the sound pressure response of the reference point to the surrounding 360-degree azimuth angle and ±90-degree elevation angle is calculated. The azimuth angle is evaluated in 10-degree increments and the elevation angle in 5-degree increments, for a total of 36×37=1332 directions. A beamforming weight vector is applied to each direction to calculate the beam output energy in that direction. For transformer noise, the beam output energy reaches a peak of -8dB in the direction of 225 degrees azimuth and -12 degrees elevation, and the sound pressure response in this direction is significantly higher than in other directions. The direction with the highest beam output energy is identified as the main beam direction. For transformer noise, the main beam direction is 225 degrees azimuth and -12 degrees elevation. Analyzing the energy distribution within ±15 degrees around the main beam direction, the energy within this range remains above the peak value of -3dB, indicating that the sound source has strong directionality in this direction.

[0052] A superimposed sound field is formed by array beamforming around the main sound beam. Multiple sub-beam directions are set within a ±30 degree angular range around the main sound beam's direction as the central axis. These sub-beam directions are fan-shaped in the azimuth and elevation planes, covering 195-255 degrees in azimuth and -42 to 18 degrees in elevation, for a total of 25 sub-beam directions. A corresponding beamforming weight vector is calculated for each sub-beam direction, and the weight vector for each direction is optimized for that direction. The weight vectors of the main beam and each sub-beam are applied to the acoustic signals of each channel in the aligned acoustic dataset to generate the main beam output signal and the 25 sub-beam output signals. The main beam output signal and the sub-beam output signals are then summed using amplitude weighting. The main beam weight coefficient is set to 0.50, and the weight coefficients of each sub-beam are determined based on their angular deviation from the main beam; the smaller the angular deviation, the greater the weight. The output signals of the main beam and 25 sub-beams are weighted and accumulated according to their respective weighting coefficients, and the accumulated signal is the superimposed sound field. The superimposed sound field maintains the same sampling rate of 44.1kHz and duration of 10 seconds as the original acoustic signal in the time domain, and spatially integrates the acoustic information of the main beam direction and the surrounding sub-beam directions. Compared with the single beam output, the superimposed sound field has a more comprehensive spatial coverage of the sound source.

[0053] A phase distribution map is established by extracting the phase difference of the superimposed sound field. A Short-Time Fourier Transform (STFT) is performed on the superimposed sound field to convert the time-domain signal into a complex representation in the time-frequency domain. The STFT time window length is set to 2048 sampling points, with a window overlap rate of 75%. Phase information is calculated for the spectrum of each time window, obtained from the amplitude of the complex spectrum, with values ​​ranging from -π to π. Phase information within the characteristic frequency band of the abnormal noise is extracted; for partial discharge abnormal noise, the phase in the 5.8-12.3kHz frequency band is extracted. The phase difference between the superimposed sound field and the original signal of the reference channel is calculated. Channel 1 is selected as the reference channel, and the phase difference between the superimposed sound field and the signal of Channel 1 at each frequency and time point is calculated. The phase difference is calculated for all time frames and frequency channels, constructing a three-dimensional matrix of time-frequency-phase difference. The three-dimensional matrix is ​​averaged along the frequency dimension to obtain a two-dimensional matrix of time-phase difference, which is the phase distribution map. The horizontal axis of the phase distribution diagram represents time (0-10 seconds), and the vertical axis represents phase difference (-180 to 180 degrees). The shades of color in the diagram indicate the phase difference values. The abnormal noise from the transformer maintains a phase difference of -30 to -40 degrees during the t=2-3 second period, but abruptly changes to 80-100 degrees during the t=6-7 second period, reflecting the phase characteristic changes of the abnormal noise event.

[0054] In some embodiments, the step of combining the phase distribution map with the sound source location probability distribution to generate spatial positioning data by performing reflection path elimination includes: performing phase gradient analysis on the phase distribution map to identify phase jump regions; extracting low probability intervals corresponding to the phase jump regions from the sound source location probability distribution; using the low probability intervals to perform interference path reverse tracing to form an interference source location set; and performing probability suppression processing on the interference source location set to generate spatial positioning data.

[0055] Phase gradient analysis is performed on the phase distribution map to identify phase jump regions. The phase gradient along the time dimension is calculated from the phase distribution map; the phase gradient is equal to the rate of change of phase difference between adjacent time frames. At time t=2.5 seconds, the phase difference is -35 degrees; at time t=2.55 seconds, the phase difference is -32 degrees, resulting in a phase gradient of 6 degrees / second. At time t=5.8 seconds, the phase difference is -25 degrees; at time t=5.85 seconds, the phase difference is 75 degrees, resulting in a phase gradient of 2000 degrees / second. A phase gradient threshold of 500 degrees / second is set, and the time locations where the absolute value of the phase gradient exceeds the threshold are identified from the phase distribution map. For transformer abnormal noise, the phase gradient exceeds the threshold at times t=5.8 seconds, t=6.2 seconds, and t=7.5 seconds, reaching 2000 degrees / second, 1800 degrees / second, and 1500 degrees / second, respectively. The time locations where the phase gradient exceeds the threshold and the surrounding ±0.2-second time range are extracted; these time ranges constitute the phase jump regions. For abnormal transformer noise, the phase jump region comprises three time periods: 5.6-6.0 seconds, 6.0-6.4 seconds, and 7.3-7.7 seconds, each lasting 0.4 seconds. Analyzing the phase change characteristics within each phase jump region, the phase jumps from -25 degrees to 75 degrees in the 5.6-6.0 second region, a span of 100 degrees; and the phase jumps from 70 degrees to -80 degrees in the 6.0-6.4 second region, a span of 150 degrees. These abrupt changes typically correspond to multipath propagation or reflection path interference of sound waves.

[0056] For example, extracting the low-probability interval corresponding to the phase transition region from the probability distribution of the sound source location includes: obtaining a probability distribution sequence by performing probability value retrieval from the probability distribution of the sound source location with the phase transition region as the center and moving outward; performing a rate of change analysis on the probability distribution sequence to generate a probability gradient curve; continuously shrinking the retrieval radius based on the probability gradient curve to obtain the positioning accuracy; and determining the current retrieval area as a low-probability interval when the positioning accuracy reaches a set requirement.

[0057] The probability distribution sequence is obtained by searching for probability values ​​outward from the phase transition region in the probability distribution of the sound source location. The time period corresponding to the phase transition region is extracted; for transformer noise, the first phase transition region is 5.6-6.0 seconds. Within this time period, the instantaneous distribution of the spatial probability field is extracted from the probability distribution of the sound source location. The spatial location with the highest probability within this time period is identified. For the 5.6-6.0 second time period, the probability peak locations are X=-0.28 meters, Y=-0.70 meters, and Z=1.22 meters, with a probability value of 0.65. A concentric circle search range is set outward from this probability peak location, starting with a search radius of 0.1 meters and gradually increasing to 1.0 meter in increments of 0.05 meters. For each search radius location, the probability values ​​of all spatial points on the sphere of that radius are extracted, and the average spherical probability is calculated. Each search radius is paired with its corresponding average spherical probability, arranged in ascending order of radius to form a radius-probability sequence, which is the probability distribution sequence. The probability distribution sequence shows that the probability value decreases as the distance from the center point increases. For transformer noise, the probability value decreases from 0.65 at the center to 0.08 at a radius of 1.0 meter.

[0058] A rate of change analysis is performed on the probability distribution sequence to generate a probability gradient curve. The rate of change of probability for adjacent radius positions in the probability distribution sequence is calculated; the rate of change of probability equals the probability difference divided by the radius difference. For a radius of 0.1 meters with a probability of 0.58 and a radius of 0.15 meters with a probability of 0.52, the rate of change of probability is -1.2 / meter. For a radius of 0.3 meters with a probability of 0.42 and a radius of 0.35 meters with a probability of 0.38, the rate of change of probability is -0.8 / meter. The rate of change of probability is calculated for all adjacent radius positions, and a curve is plotted by pairing each radius position with its corresponding rate of change of probability. The horizontal axis represents the search radius, and the vertical axis represents the rate of change of probability; this curve is the probability gradient curve. The probability gradient curve reflects the rate of probability decay as a function of radius. For transformer noise, within a radius of 0.1-0.3 meters, the probability gradient curve value is relatively large, remaining between -0.8 and -1.2 / meter, indicating rapid probability decay. Within a radius of 0.6-1.0 meters, the curve value is relatively small, remaining between -0.2 and -0.4 / meter, indicating slow probability decay. The inflection point of the curve usually corresponds to the boundary of the main radiation area of ​​the sound source.

[0059] The search radius is continuously reduced based on the probability gradient curve to obtain positioning accuracy. When the positioning accuracy reaches the set requirement, the current search area is determined to be a low-probability interval. The trend of the probability gradient curve is analyzed to identify the radius positions where the absolute value of the probability gradient is below a threshold. A probability gradient threshold of -0.5 / m is set, and radius positions where the absolute value of the probability gradient is below 0.5 / m are identified. For transformer noise, the absolute value of the probability gradient at locations with a radius greater than 0.5 meters is below 0.5 / m. Starting from the maximum search radius of 1.0 meter, the search is reduced inwards in 0.05-meter increments, and the ratio of the probability variance to the boundary probability within the reduced search area is calculated. The probability variance within the search area reflects the uniformity of the probability distribution within that area, while the ratio of the boundary probability to the central probability reflects the positioning accuracy of the search area; a smaller ratio indicates more accurate positioning. For a search area with a radius of 0.5 meters, the boundary probability is 0.28, the central probability is 0.65, and the ratio is 0.43, which represents the positioning accuracy. The positioning accuracy requirement is set to a ratio less than 0.50. When the search radius shrinks to 0.5 meters, a positioning accuracy of 0.43 meets the requirement, and the shrinking stops. The outer edge of the current search area is now considered a low-probability zone. For transformer noise, the low-probability zone is a ring-shaped area with a radius of 0.5-1.0 meters. The probability value within this area is 0.08-0.28, significantly lower than the 0.65 in the central area, indicating that this area is more likely to correspond to reflection paths or multipath interference.

[0060] Interference source location set is formed by reverse tracing the interference path using low-probability intervals. Local probability peak locations are extracted from these intervals; a local probability peak is defined as a location whose probability value is higher than that of all locations within a 0.1-meter radius. For transformer noise, two local probability peak locations are identified in the low-probability interval; these locations may correspond to sound wave reflection or scattering points. The azimuth and elevation angles of each local probability peak location relative to the center of the enhanced localization area are calculated, and local probability peak locations with angle deviations exceeding 45 degrees are identified, as these locations are more likely to correspond to reflection paths. A reverse tracing ray is emitted from the reflection point location into the surrounding space, with the ray direction pointing from the reflection point towards the normal of the possible reflecting surface. Based on the locations of reflecting surfaces such as walls and equipment casings in the substation environment, the intersection points of the reverse tracing ray with each reflecting surface are calculated. Reverse tracing is performed on all locations with angle deviations exceeding 45 degrees; a total of three reflecting surface locations are identified for transformer noise, and these three locations constitute the interference source location set. This interference source location set is used for subsequent probability suppression processing to eliminate the interference of reflection paths on sound source localization.

[0061] Spatial positioning data is generated by performing probabilistic suppression processing on the set of interference source locations. For each reflector in the set of interference source locations, a reflection path is established from the reflector to each sensor. The reflection path includes two segments: from the sound source to the reflector, and then from the reflector to the sensor. The arrival time delay of the sound wave corresponding to the reflection path is calculated and compared with the delay of the direct path to identify spatial regions with a delay difference in the range of 0.5-2.0 milliseconds. For the spatial regions affected by the reflection path, a suppression coefficient is applied to attenuate their probability values. The suppression coefficient η is obtained through the exponential decay function η=exp(-α·Δt), where η is the suppression coefficient (dimensionless, ranging from 0 to 1), α is the attenuation factor with a value of 1.2 / millisecond, and Δt is the delay difference between the reflection path and the direct path (in milliseconds). The original probability value of the affected region is multiplied by the suppression coefficient η to obtain the suppressed probability value. For the reflective surface of the north wall of the transformer, the reflection path delay in the area with a radius of 0.6 meters and an azimuth angle of 120 degrees is 1.2 milliseconds longer than that of the direct path. The suppression coefficient η = exp(-1.2 × 1.2) = 0.22, and the original probability value of 0.22 in this area is reduced to 0.05 after suppression. Probability suppression is performed on all reflective surface locations within the interference source location set. The three reflective surface locations of the transformer noise correspond to annular areas with radii of 0.5-0.7 meters, 0.7-0.9 meters, and 0.8-1.0 meters, respectively. The probability values ​​in these areas are all suppressed to below 0.10. The suppressed probability distribution is merged with the original sound source location probability distribution. The suppressed areas adopt the lower probability value after suppression, while the non-suppressed areas retain the original probability value. The merged three-dimensional spatial probability field is the spatial positioning data. The spatial positioning data provides more accurate sound source location information by eliminating reflection path interference.

[0062] Step S150: Perform sound pressure change rate analysis on the spatial positioning data to establish a sound pressure gradient vector field. Based on the sound pressure gradient vector field, reconstruct the sound field around the sound source to form a three-dimensional sound field structure. Use the three-dimensional sound field structure to perform energy peak search to determine the spatial location and type of abnormal noise source.

[0063] Specifically, a sound pressure gradient vector field is established by performing sound pressure change rate analysis on the spatial positioning data. Through sound pressure change rate analysis, the probability distribution values ​​in three-dimensional space are extracted from the spatial positioning data. For transformer noise, the spatial grid resolution is 0.05 m × 0.05 m × 0.1 m, covering a monitoring area of ​​approximately 2 m × 2 m × 1 m. The probability distribution in the spatial positioning data is mathematically equivalent to the normalized sound pressure energy distribution; therefore, the probability distribution field is considered equivalent to a sound pressure field for analysis. Spatial gradient calculations are performed on the probability values ​​of each grid cell. The spatial gradient includes three components: X-direction gradient, Y-direction gradient, and Z-direction gradient. For a grid cell with positions X = -0.32 m, Y = -0.74 m, and Z = 1.25 m, the probability value is 0.88. By calculating the probability difference between this cell and its adjacent cells, the X-direction gradient is -3.2 / m, the Y-direction gradient is -2.6 / m, and the Z-direction gradient is -0.8 / m. The three directional gradients are combined into a gradient vector (-3.2, -2.6, -0.8), with the vector direction pointing from the high-probability region to the low-probability region. Gradient calculations are performed on all grid cells, generating a three-dimensional gradient vector for each cell. The spatial distribution of these gradient vectors constitutes the sound pressure gradient vector field. Near the center of the enhanced localization region, the sound pressure gradient vector field exhibits a radiating pattern from the center outwards, with gradient amplitudes ranging from 2 to 4 / meter. In the outer low-probability region, the gradient vector direction is more chaotic, and the gradient amplitude drops to 0.5-1.0 / meter. This radial distribution pattern highly matches the typical radiation characteristics of a point sound source.

[0064] A three-dimensional sound field structure is formed by reconstructing the sound field around the sound source based on the sound pressure gradient vector field. The location of the grid cell with the largest gradient amplitude in the sound pressure gradient vector field is extracted. For transformer noise, the peak gradient amplitude locations are X = -0.30 m, Y = -0.72 m, and Z = 1.23 m, with a gradient amplitude of 4.2 / m. A reconstruction range is defined outward from the peak gradient amplitude location, covering a spatial volume of approximately 1.6 m × 1.6 m × 0.8 m. A refined grid is established within the reconstruction range, and sound field reconstruction around the sound source is performed. The grid resolution is set to 0.02 m × 0.02 m × 0.05 m, with a total of approximately 256,000 refined grid cells. For each refined grid cell, the gradient vector at that location is obtained by interpolation from the sound pressure gradient vector field using a trilinear interpolation algorithm, based on the gradient vectors of the eight coarse grid cells surrounding the refined grid cell. Divergence is calculated on the gradient vector of each refined grid cell. The divergence is obtained by summing the spatial derivatives of the three components of the gradient vector. Where div represents the divergence operator, This represents the gradient vector of the probability distribution p (equivalent to the sound pressure gradient vector). , , Here, represents the second-order partial derivatives of the probability distribution p with respect to spatial coordinates x, y, and z (unit: 1 / m²), where x, y, and z are spatial coordinates (unit: meters). A positive divergence value indicates that the location is a sound source, while a negative divergence value indicates that the location is a sound sink. The divergence of the refined unit near X = -0.32 m, Y = -0.74 m, and Z = 1.25 m is positive, 5-8 / m², indicating that this location is a sound source region. The Poisson equation is solved on the divergence field to reconstruct the sound pressure distribution. The resulting sound pressure field is then fused with the probability distribution of the original spatial positioning data using a weighted average method. The weight of the reconstructed sound pressure field is 0.60, and the weight of the original probability distribution is 0.40. The fused three-dimensional spatial sound pressure distribution constitutes the three-dimensional sound field structure. At the Z=1.25m layer, the sound pressure contour lines of the three-dimensional sound field structure form closed circles at X=-0.32m and Y=-0.74m, with the highest sound pressure at the center and gradually decreasing outwards. The sound pressure contour lines of the abnormal noises on the busbars are elliptical or irregular in shape, reflecting the differences in the spatial radiation characteristics of different abnormal noise sources.

[0065] In some embodiments, the step of using the three-dimensional sound field structure to perform energy peak search to determine the spatial location and type of abnormal noise source includes: extracting energy distribution features from the three-dimensional sound field structure; performing peak search on the energy distribution features to generate energy peak positions; extracting spectral features from the energy peak positions and performing type matching to generate abnormal noise type identifiers; and combining the energy peak positions with the abnormal noise type identifiers to determine the spatial location and type of abnormal noise source.

[0066] Energy distribution characteristics are extracted from the three-dimensional sound field structure. The sound pressure level (SPL) of each grid cell in the three-dimensional sound field structure is converted into energy, with the energy value equal to the square of the SPL. For cells at positions X = -0.32 m, Y = -0.74 m, and Z = 1.25 m, the SPL is 0.88 and the energy is 0.77. Energy conversion is performed on all grid cells to obtain the energy distribution field in three-dimensional space. Statistical characteristics of the energy distribution field are analyzed, including the maximum, minimum, average, and standard deviation of energy, as well as the spatial concentration of energy distribution. For transformer noise, the maximum energy is 0.77, the minimum is 0.00, the average is 0.12, and the standard deviation is 0.18. These statistics constitute the statistical dimensions of the energy distribution characteristics. Spatial concentration is represented by the percentage of grid cells with energy above the average; a concentration of 8% indicates that energy is concentrated in a few spatial regions. The spatial morphological characteristics of the energy distribution were analyzed to identify the principal axis direction, which was obtained through the inertial tensor eigenvector of the energy distribution. For transformer noise, the principal axis direction was azimuth 225 degrees and pitch -12 degrees, consistent with the direction of the main sound beam. The profile characteristics of the energy distribution along the principal axis were extracted. The profile showed that the energy attenuated from the peak position to both sides along the principal axis, and the attenuation pattern conformed to the energy attenuation characteristics of spherical waves in a free field. These spatial morphological characteristics further refined the description of the energy distribution characteristics.

[0067] Peak search is performed on the energy distribution characteristics to generate energy peak locations. Through peak search, the maximum energy value of 0.77 and its corresponding spatial coordinates X = -0.32 m, Y = -0.74 m, and Z = 1.25 m are extracted from the energy distribution characteristics as preliminary energy peak locations. A search window of 0.2 m × 0.2 m × 0.2 m cube size is set around the preliminary peak locations, and a fine peak search is performed within the window with a step size of 0.01 m. For each search point within the window, the sound pressure level is interpolated from the three-dimensional sound field structure and converted into an energy value. All search points within the window are traversed to identify the search point with the largest energy value. The precise energy peak location is X = -0.315 m, Y = -0.738 m, Z = 1.252 m, with an energy value of 0.79. To verify the stability of the energy peak position, 16 verification points were sampled on a spherical surface with a radius of 0.05 meters around the peak position. The average ratio of the energy at the verification points to the peak energy was 0.85, and the standard deviation was 0.06. The peak position stability of partial discharge-type abnormal noises is generally better than that of mechanical loosening-type abnormal noises.

[0068] Spectral features are extracted from the energy peak location and type matching is performed to generate abnormal noise type identifiers. At the energy peak location (X=-0.315m, Y=-0.738m, Z=1.252m), the beamforming output signal at this location is analyzed spectrally. The beam is directed towards this energy peak location, and the beamforming weight vector is applied to the weighted sum of the signals from each channel of the aligned acoustic dataset to obtain the focused signal at this location. A Fourier transform is performed on the focused signal to obtain the spectral distribution with a spectral resolution of 10.8Hz. The top 10 frequency components with the highest energy in the spectrum are extracted as the dominant frequency features. For transformer partial discharge abnormal noise, the dominant frequency range is 6.5-15.8kHz, and the bandwidth is 9.3kHz. An abnormal noise type feature library is established, containing spectral feature templates for eight typical abnormal noise types, including partial discharge, core loosening, winding deformation, mechanical loosening, and bearing wear. The spectral features extracted from the energy peak location are compared with the templates in the feature library for similarity calculation. The similarity is comprehensively evaluated by the overlap of the dominant frequency range and the bandwidth deviation. With a similarity of 0.88 to the partial discharge template, the partial discharge type with the highest similarity was selected as the matching result, and this type is the abnormal noise type identifier. The main frequency range of abnormal noises caused by loose iron cores is usually 200-1500Hz, which is significantly different from the 6-16kHz range of partial discharge. The difference in spectral characteristics ensures the accuracy of type identification.

[0069] The spatial location and type of abnormal noise were determined by combining the energy peak location and the abnormal noise type identifier. The three-dimensional coordinates of the energy peak location were extracted: X = -0.315 m, Y = -0.738 m, Z = 1.252 m. Using the origin of the equipment coordinate system, i.e., the geometric center of the monitoring equipment, as the reference, this location is located southwest of the monitoring equipment, at a horizontal distance of 0.81 m and a vertical height of 1.25 m. These three-dimensional coordinates represent the spatial location of the abnormal noise source. Based on the spatial layout of the transformer equipment, the spatial location of the abnormal noise source was mapped onto the equipment components. X = -0.315 m and Y = -0.738 m correspond to the A-phase winding area of ​​the transformer body, and Z = 1.252 m corresponds to the height of the middle of the winding. The specific component where the abnormal noise occurred was determined to be the middle area of ​​the A-phase winding. The abnormal noise type identifier was extracted as partial discharge. Combined with the component location information, the abnormal noise source was determined to be a partial discharge defect in the middle of the A-phase winding. The system outputs a report locating the source of the abnormal noise. The report includes the spatial coordinates of the source: X = -0.315 meters, Y = -0.738 meters, Z = 1.252 meters; location: center of phase A winding of the equipment component; noise type: partial discharge; similarity: 0.88; main frequency range: 6.5-15.8 kHz; peak energy: 0.79. Maintenance personnel can use the location and type of the abnormal noise from the report to develop targeted repair plans and prepare appropriate repair tools.

[0070] To implement the high-throughput acoustic sensing-based abnormal noise location and monitoring method for power equipment corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects. See also Figure 2 , Figure 2 A structural block diagram of a power equipment abnormal noise location and monitoring device 200 based on high-throughput acoustic sensing, provided in an embodiment of this application, is shown. For ease of explanation, only the parts relevant to this embodiment are shown. The power equipment abnormal noise location and monitoring device 200 based on high-throughput acoustic sensing provided in this embodiment includes: Data acquisition module 201 is used to acquire multi-channel acoustic signals and sensor array position information during the operation of power equipment, and to perform clock synchronization calibration on the multi-channel acoustic signals to construct an aligned acoustic dataset; The preliminary positioning module 202 is used to extract the time difference of arrival pattern from the aligned acoustic dataset, establish a spatial propagation path using the time difference of arrival pattern to form a propagation ray beam, perform intersection point positioning on the propagation ray beam to generate candidate sound source regions, and enhance the candidate sound source regions to form an enhanced positioning region. The frequency domain analysis module 203 is used to convert the multi-channel acoustic signal to the frequency domain to obtain a spectral energy map, retrieve abnormal noise characteristic frequency bands in the spectral energy map, establish time-frequency positioning parameters by combining the abnormal noise characteristic frequency bands with the sensor array position information, and generate a sound source position probability distribution based on the time-frequency positioning parameters. Interference cancellation module 204 is used to perform sound pressure directionality analysis to obtain the direction of the main sound beam in the enhanced positioning area, perform array beam synthesis around the direction of the main sound beam to form a superimposed sound field, extract the phase difference of the superimposed sound field to establish a phase distribution map, and combine the phase distribution map with the probability distribution of the sound source position to perform reflection path cancellation to generate spatial positioning data. The precise positioning module 205 is used to perform sound pressure change rate analysis on the spatial positioning data to establish a sound pressure gradient vector field, reconstruct the sound field around the sound source based on the sound pressure gradient vector field to form a three-dimensional sound field structure, and use the three-dimensional sound field structure to perform energy peak search to determine the spatial location and type of abnormal noise source.

[0071] The aforementioned high-throughput acoustic sensing-based power equipment abnormal noise location and monitoring device 200 can implement the high-throughput acoustic sensing-based power equipment abnormal noise location and monitoring method described in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment and will not be detailed here. The remaining content of this application embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0072] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

[0073] The above embodiments are not an exhaustive list based on the present invention, and there may be many other embodiments not listed. Any substitutions and improvements made without departing from the concept of the present invention are within the protection scope of the present invention.

Claims

1. A method for locating and monitoring abnormal noises in power equipment based on high-throughput acoustic sensing, characterized in that, include: Collect multi-channel acoustic signals and sensor array position information during the operation of power equipment, and perform clock synchronization calibration on the multi-channel acoustic signals to construct an aligned acoustic dataset; The time difference of arrival (TDOA) pattern is extracted from the aligned acoustic dataset. The TDOA pattern is used to establish a spatial propagation path to form a propagation ray beam. The intersection point of the propagation ray beam is used to generate candidate sound source regions. The candidate sound source regions are then enhanced to form an enhanced localization region. The multi-channel acoustic signal is converted to the frequency domain to obtain a spectral energy map. The abnormal noise characteristic frequency band is retrieved in the spectral energy map. The abnormal noise characteristic frequency band is combined with the sensor array position information to establish time-frequency positioning parameters. The sound source position probability distribution is generated based on the time-frequency positioning parameters. The method involves performing sound pressure directionality analysis to obtain the direction of the main sound beam in the enhanced positioning area, performing array beamforming around the direction of the main sound beam to form a superimposed sound field, extracting the phase difference of the superimposed sound field to establish a phase distribution map, and combining the phase distribution map with the probability distribution of the sound source location to perform reflection path elimination to generate spatial positioning data. This includes: performing phase gradient analysis on the phase distribution map to identify phase jump regions; extracting low-probability intervals corresponding to the phase jump regions from the probability distribution of the sound source location; using the low-probability intervals to trace interference paths in reverse to form a set of interference source locations; and performing probability suppression processing on the set of interference source locations to generate spatial positioning data. Specifically, extracting the low-probability intervals corresponding to the phase jump regions from the probability distribution of the sound source location includes: performing probability value retrieval from the probability distribution of the sound source location outwards from the phase jump region to obtain a probability distribution sequence; performing rate of change analysis on the probability distribution sequence to generate a probability gradient curve; continuously shrinking the retrieval radius based on the probability gradient curve to obtain positioning accuracy; and determining the current retrieval area as a low-probability interval when the positioning accuracy reaches a set requirement. Sound pressure change rate analysis is performed on the spatial positioning data to establish a sound pressure gradient vector field. Based on the sound pressure gradient vector field, the sound field around the sound source is reconstructed to form a three-dimensional sound field structure. The energy peak search is performed using the three-dimensional sound field structure to determine the spatial location and type of abnormal noise source.

2. The method according to claim 1, characterized in that, The step of performing clock synchronization calibration on the multi-channel acoustic signals to construct an aligned acoustic dataset includes: Identify clock drift intervals from the multi-channel acoustic signals; Establish a clock correction reference for each channel using the aforementioned clock drift interval to generate synchronization anchor points; Based on the synchronization anchor point, time offset correction is applied to the multi-channel acoustic signal to generate calibration sampling data; An aligned acoustic dataset is constructed based on the calibration sampling data.

3. The method according to claim 1, characterized in that, The step of enhancing the candidate sound source region to form an enhanced localization region includes: Identify low-sound-energy regions around the candidate sound source regions; The beam focusing parameters are determined by defining the focusing boundary range using the low acoustic energy region. Spatial weighting coefficients are generated based on the beam focusing parameters; The spatial weighting coefficients are used to concentrate energy in the candidate sound source region to form an enhanced localization region.

4. The method according to claim 1, characterized in that, The step of establishing time-frequency positioning parameters by combining the abnormal noise characteristic frequency band with the sensor array position information includes: Frequency gradient analysis is performed on the frequency bands of the abnormal noise to identify the frequency jump locations; A frequency-position mapping table is formed by spatially associating the frequency jump positions with the sensor array position information; Spatial response features are extracted from the frequency-location mapping table to generate a positioning reference vector; Time-frequency positioning parameters are established based on the positioning reference vector.

5. The method according to claim 1, characterized in that, The step of generating the sound source location probability distribution based on the time-frequency positioning parameters includes: Based on the time-frequency positioning parameters, a spatial search grid is established to divide candidate location points; Perform time-frequency response evaluation on the candidate location points to generate a response intensity sequence; Confidence boundaries are determined by identifying high-response clusters from the response intensity sequence; Based on the confidence boundary, a probability mapping is performed on the response intensity sequence to generate a probability distribution of the sound source location.

6. The method according to claim 1, characterized in that, The process of using the three-dimensional sound field structure to perform energy peak search to determine the spatial location and type of abnormal noise source includes: Extract energy distribution features from the three-dimensional sound field structure; Perform a peak search on the energy distribution characteristics to generate the energy peak locations; Extract spectral features from the energy peak location and perform type matching to generate anomaly type identifiers; The spatial location and type of abnormal noise source are determined by combining the energy peak location with the abnormal noise type identifier.

7. The method according to claim 3, characterized in that, The generation of spatial weighting coefficients based on the beam focusing parameters includes: Identify array response null regions from the beam focusing parameters; The effective focusing angle range is defined by the array response null region; An initial weighted vector is generated by performing weight optimization on the effective focusing angle range; Energy normalization is applied to the initial weighting vector to generate spatial weighting coefficients.

8. A power equipment abnormal noise location and monitoring device based on high-throughput acoustic sensing, characterized in that, include: The data acquisition module is used to collect multi-channel acoustic signals and sensor array position information during the operation of power equipment, and to perform clock synchronization calibration on the multi-channel acoustic signals to construct an aligned acoustic dataset; The preliminary localization module is used to extract the time difference of arrival pattern from the aligned acoustic dataset, establish a spatial propagation path using the time difference of arrival pattern to form a propagation ray beam, perform intersection point localization on the propagation ray beam to generate candidate sound source regions, and enhance the candidate sound source regions to form an enhanced localization region. The frequency domain analysis module is used to convert the multi-channel acoustic signal to the frequency domain to obtain a spectral energy map, retrieve abnormal noise characteristic frequency bands in the spectral energy map, establish time-frequency positioning parameters by combining the abnormal noise characteristic frequency bands with the sensor array position information, and generate a sound source location probability distribution based on the time-frequency positioning parameters. An interference cancellation module is used to perform sound pressure directionality analysis to obtain the direction of the main sound beam in the enhanced positioning area, perform array beamforming around the direction of the main sound beam to form a superimposed sound field, extract the phase difference of the superimposed sound field to establish a phase distribution map, and combine the phase distribution map with the probability distribution of the sound source location to perform reflection path elimination to generate spatial positioning data. This includes: performing phase gradient analysis on the phase distribution map to identify phase jump regions; extracting low-probability intervals corresponding to the phase jump regions from the probability distribution of the sound source location; using the low-probability intervals to trace interference paths in reverse to form an interference source location set; and performing probability suppression processing on the interference source location set to generate spatial positioning data. The step of extracting the low-probability intervals corresponding to the phase jump regions from the probability distribution of the sound source location includes: performing probability value retrieval from the probability distribution of the sound source location outwards from the phase jump region to obtain a probability distribution sequence; performing rate of change analysis on the probability distribution sequence to generate a probability gradient curve; continuously shrinking the retrieval radius based on the probability gradient curve to obtain positioning accuracy; and determining the current retrieval area as a low-probability interval when the positioning accuracy reaches a set requirement. The precise positioning module is used to perform sound pressure change rate analysis on the spatial positioning data to establish a sound pressure gradient vector field, reconstruct the sound field around the sound source based on the sound pressure gradient vector field to form a three-dimensional sound field structure, and use the three-dimensional sound field structure to perform energy peak search to determine the spatial location and type of abnormal noise source.

Images