A Transformer Fault Diagnosis Method Integrating Phase Localization and Adaptive Acoustic Mapping

Abstract

This invention discloses a transformer fault diagnosis method integrating phase localization and adaptive acoustic signature mapping, belonging to the field of power equipment condition monitoring technology. The method includes: acquiring multi-channel acoustic signals using acoustic sensors to estimate the spatial location of the fault source; weighting and fusing the multi-channel acoustic signals using localization information to generate a master signal; processing the master signal using adaptive window-length short-time Fourier transform and a trainable Mel filter bank to generate a two-dimensional acoustic signature map; processing the map in conjunction with spatial location information, and inputting the enhanced acoustic signature map into a probability distribution mapping module to achieve deep fusion and mutual verification of fault type identification and spatial localization, and continuously improving system performance through a closed-loop optimization mechanism. This invention effectively solves the problems of insufficient signal fusion, weak adaptive feature extraction, and separation of diagnosis and localization in traditional transformer fault diagnosis methods, improving the accuracy, reliability, and adaptive capability of transformer fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, specifically to a transformer fault diagnosis method that integrates phase localization and adaptive acoustic signature mapping. Background Technology

[0002] In the operation and maintenance of power systems, transformers are critical equipment, and their health status directly affects the safety and stability of the power grid. Voiceprint detection technology, due to its advantages such as non-contact operation, low cost, and rich information, has been introduced into the field of transformer fault diagnosis.

[0003] Existing technologies typically employ single or multiple microphones to collect sound signals and generate acoustic signatures through time-frequency analysis, then utilize pattern recognition or shallow machine learning models for fault classification. However, such methods have significant limitations. Multi-channel acoustic signal processing often employs simple fixed-weight fusion or averaging fusion, failing to fully utilize spatial information of the sound source to optimize signal quality, resulting in fault features being submerged in strong background noise. The acoustic signature generation process often uses fixed time-frequency analysis parameters, lacking adaptability to non-stationary signals, affecting the accuracy of feature extraction. Furthermore, diagnostic models typically separate fault type identification from fault location, failing to achieve information complementarity and collaborative optimization, thus limiting further improvements in diagnostic accuracy. Summary of the Invention

[0004] This invention addresses the technical problems in existing technologies, such as insufficient adaptive capability of acoustic signature feature extraction and limited diagnostic accuracy due to the separation of fault identification and localization processes. It provides a transformer fault diagnosis method that integrates phase localization and adaptive acoustic signature.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] This invention provides a transformer fault diagnosis method that integrates phase localization and adaptive acoustic signature mapping, including:

[0007] Multiple acoustic wave sensors arranged in space are used to synchronously collect and acquire multi-channel acoustic wave signals.

[0008] Phase deviation information is calculated based on the multi-channel acoustic signals, and the spatial location information of the internal fault source of the transformer is estimated by combining the sound source localization algorithm. The multi-channel acoustic signals are then fused according to the spatial location information to obtain the main signal.

[0009] An adaptive window-length short-time Fourier transform is performed on the main signal to obtain the time-frequency distribution, and the time-frequency distribution is converted into a two-dimensional acoustic signature using a trainable Mel filter bank.

[0010] The two-dimensional acoustic signature is subjected to adaptive denoising and key frequency band enhancement processing to generate an enhanced acoustic signature, which is then input into a pre-trained probability distribution mapping module to obtain fault type information. This information is then combined with the spatial location information and output as a fault diagnosis result.

[0011] The beneficial effects of this invention are:

[0012] Compared to existing technologies, this invention first achieves precise spatial localization of internal transformer fault sources through phase difference positioning technology. Based on this positioning information, it intelligently weights and fuses multi-channel acoustic signals, effectively improving the signal-to-noise ratio and fault feature abundance of the main signal used for diagnosis. Secondly, it employs adaptive window length time-frequency analysis technology and a trainable Mel filter bank to dynamically optimize the time-frequency resolution of the acoustic signature spectrum according to signal characteristics, focusing on the frequency bands most sensitive to faults and enhancing feature representation capabilities. Thirdly, it combines spatial location information to perform guided adaptive denoising and key frequency band enhancement on the acoustic signature spectrum, further suppressing background interference and highlighting fault components. Finally, by designing a probability distribution mapping module with a dual-branch discrimination mechanism, it achieves deep fusion and mutual verification of fault type identification and spatial location information, using positioning information to constrain and correct classification results, thereby comprehensively improving the accuracy and reliability of fault diagnosis. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the transformer fault diagnosis method that integrates phase localization and adaptive acoustic signature mapping provided by the present invention.

[0014] Figure 2 This is a schematic diagram of the probability distribution mapping module provided by the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0017] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0018] Example 1, as Figure 1 As shown, this embodiment of the invention provides a transformer fault diagnosis method that integrates phase localization and adaptive acoustic signature mapping, including:

[0019] S10: Acquires multi-channel acoustic signals by synchronously collecting data from multiple acoustic sensors arranged in space;

[0020] First, multiple acoustic wave sensors are arranged in a three-dimensional spatial distribution on the surface of the transformer casing or in the surrounding space. All acoustic wave sensors achieve strictly synchronized data acquisition through a synchronous triggering mechanism to ensure that the acquired acoustic wave signals are completely aligned on the time axis.

[0021] Specifically, the acoustic wave sensor continuously acquires the acoustic wave signals radiated by the transformer during operation at a preset sampling frequency, obtaining multi-channel acoustic wave signals. The preset sampling frequency is set based on the highest effective frequency component of the transformer fault acoustic signal, for example, 100 kHz to 200 kHz per second. Multi-channel refers to independent signal acquisition paths, each corresponding to an acoustic wave sensor at a specific spatial location. The acoustic wave signal acquired by each channel includes various acoustic components such as transformer operating sound, potential fault sound, and environmental background noise.

[0022] The acquisition of this multi-channel acoustic signal can provide raw data input for subsequent spatial location of fault sources and signal fusion processing.

[0023] S20: Calculate phase deviation information based on the multi-channel acoustic signal, estimate the spatial location information of the fault source inside the transformer by combining the sound source localization algorithm, and perform multi-channel fusion of the multi-channel acoustic signal according to the spatial location information to obtain the main signal;

[0024] Specifically, based on the multi-channel acoustic signal, phase deviation information is calculated, and combined with the sound source localization algorithm, the spatial location information of the fault source inside the transformer is estimated, including:

[0025] The multi-channel acoustic signal is preprocessed, including filtering and synchronization;

[0026] Using any two acoustic sensor channels as a channel pair, the signal is converted to the frequency domain to calculate the cross power spectrum and perform phase normalization processing. The cross-correlation function is obtained by combining the inverse transform, and the time delay corresponding to the peak value is taken as the phase deviation information.

[0027] The phase deviation information of multiple channel pairs is obtained iteratively, forming a phase difference matrix. Combined with the spatial arrangement of multiple acoustic sensors, the position coordinates of the sound source in three-dimensional space are estimated through a geometric inversion model, and the output is the spatial position information.

[0028] First, the acquired multi-channel acoustic signals undergo preprocessing. Specifically, this preprocessing includes two stages: filtering and synchronization. The filtering operation uses a bandpass filter to process the raw signal, with a passband range typically set to 100 Hz to 20 kHz, aiming to retain the effective frequency band related to transformer fault characteristics while suppressing out-of-band noise interference. The synchronization operation uses a hardware-triggered clock to perform timestamp alignment and interpolation correction on the data acquired from each channel, further ensuring accurate alignment of the signals on the time axis and laying the foundation for subsequent time delay analysis.

[0029] Secondly, within any channel pair consisting of two acoustic sensor channels, the time-domain signal is converted to the frequency domain, and the cross-power spectrum is calculated. Specifically, the time-domain signal is converted to the frequency domain using a Fourier transform. In the frequency domain, the complex spectrum of the first channel signal is multiplied point-by-point by the conjugate of the complex spectrum of the second channel signal to obtain the cross-power spectrum between the two channel signals. Then, the cross-power spectrum is phase-normalized by dividing each frequency component of the cross-power spectrum by its own magnitude to eliminate the influence of amplitude information and highlight the phase relationship. The processed result is then inversely transformed to obtain the generalized cross-correlation function. The generalized cross-correlation function is a time-domain function obtained by inverse Fourier transforming the phase-normalized cross-power spectrum, characterizing the correlation between the two channel signals at different time offsets. Analyzing the waveform of this generalized cross-correlation function, the peak point is identified. The time offset corresponding to this peak point is the phase deviation information between the channel pairs. This phase deviation information characterizes the arrival time difference of the acoustic wave propagating from the fault source to the two acoustic sensors located at different spatial positions.

[0030] Furthermore, by iterating through all possible channel pair combinations, the above calculation process is performed to obtain complete phase deviation information and construct a phase difference matrix. This phase difference matrix is ​​a symmetric matrix containing the arrival time differences between all pairs of sensors, with its dimension equal to the number of acoustic wave sensors. Each element in the matrix corresponds to the phase deviation information between a specific pair of sensors. This phase difference matrix systematically describes the time difference relationship of sound waves propagating from the fault source to each sensor in the spatial array, and completely records the relative time delay of sound waves arriving at sensors at different spatial locations.

[0031] Finally, by combining the known spatial arrangement of multiple acoustic sensors in three-dimensional space, the phase difference matrix is ​​input into the geometric inversion model to estimate the position coordinates of the sound source, i.e. the fault point, in the three-dimensional space inside the transformer.

[0032] Specifically, the position coordinates of the sound source in three-dimensional space are estimated using a geometric inversion model, and the output is the spatial position information, including:

[0033] Obtain prior fault records and, based on statistical analysis, obtain a probability distribution heatmap of the fault sources;

[0034] The spatial location information is determined by superimposing the probability distribution heatmap and the location coordinates output by the geometric inversion model.

[0035] First, prior fault records from the transformer's historical operation period are acquired. The historical operation period refers to the duration of at least one complete maintenance cycle since the transformer was commissioned, and is set according to the equipment type and operating environment characteristics, for example, a three- to five-year operating cycle. The acquired prior fault records contain statistical information on the spatial location of different types of faults within the transformer. Then, statistical analysis is performed based on these prior fault records to generate a probability distribution heatmap of the fault sources within the transformer's three-dimensional space. This probability distribution heatmap is a three-dimensional probability distribution model generated based on historical fault location data using a kernel density estimation method. It represents the statistical probability of faults occurring at different spatial locations within the transformer, visually representing the relative likelihood of faults occurring in different areas within the transformer using different colors or numerical intensities.

[0036] Then, the obtained probability distribution heatmap is superimposed and analyzed with the position coordinates output by the geometric inversion model. The geometric inversion model is a mathematical calculation model based on the physical principles of sound wave propagation and spatial geometric relationships. It uses a spherical wave propagation model to construct the propagation path of sound waves from the sound source to each sensor. By establishing a mapping relationship between the sound wave arrival time difference and the sound source position, the phase difference matrix and the spatial arrangement coordinates of multiple sound wave sensors are input into this geometric inversion model. This geometric inversion model calculates the sound source spatial coordinates that best match the observation data by solving the least squares solution of the spherical wave propagation equations or by using the maximum likelihood estimation algorithm, and outputs the three-dimensional spatial position coordinates, that is, the initial spatial position estimate of the fault source inside the transformer.

[0037] The location coordinates output by the geometric inversion model are overlaid with a probability distribution heatmap. This overlay analysis essentially fuses the positioning results based on real-time acoustic signals with prior knowledge based on historical statistical data. Specifically, a Bayesian estimation framework is used, employing real-time positioning results as the observation likelihood and historical probability distribution as the prior probability. The posterior probability distribution is obtained through probability product operations. The final determined spatial location information is the best estimate that integrates the real-time measurement results of acoustic positioning with the statistical patterns of historical fault distribution.

[0038] The final spatial location information of the internal fault source of the transformer takes into account both the real-time positioning results of the current acoustic features and the statistical patterns of historical fault data, thus improving the accuracy and reliability of the positioning.

[0039] Further, the multi-channel acoustic signal is fused according to the spatial location information to obtain the main signal, including:

[0040] Based on the spatial location information and the spatial arrangement of the multiple acoustic sensors, multiple spatial distance information is calculated and obtained;

[0041] Based on multiple spatial distance information, the signal weight of each channel in the multi-channel acoustic signal is determined, wherein the signal weight is negatively correlated with the spatial distance information;

[0042] The main signal is obtained by weighting the multi-channel acoustic signals based on the signal weights.

[0043] First, based on the determined spatial location information and the spatial arrangement of multiple acoustic sensors, the straight-line distances from the fault source location to each acoustic sensor are calculated, obtaining multiple spatial distance information. Specifically, this distance calculation is based on the Euclidean distance formula in three-dimensional space.

[0044] For example, suppose the spatial coordinates of the fault source inside the transformer, determined by a location algorithm, are (1.2, 0.8, 1.5), in meters. Four acoustic sensors are arranged on the surface of the transformer casing, with spatial coordinates of sensor A (0, 0, 0), sensor B (2, 0, 0), sensor C (2, 1.5, 0), and sensor D (0, 1.5, 0). The distance from the fault source to each sensor is calculated using the three-dimensional Euclidean distance formula:

[0045] The distance from the fault source to sensor A = √[(1.2-0)²+(0.8-0)²+(1.5-0)²] = √[1.44+0.64+2.25] ≈ 2.08 meters; the distance from the fault source to sensor B = √[(1.2-2)²+(0.8-0)²+(1.5-0)²] = √[0.64+0.64+2.25] ≈ 1.88 meters; The distance from the source to sensor C = √[(1.2-2)²+(0.8-1.5)²+(1.5-0)²] = √[0.64+0.49+2.25] ≈ 1.84 meters; the distance from the fault source to sensor D = √[(1.2-0)²+(0.8-1.5)²+(1.5-0)²] = √[1.44+0.49+2.25] ≈ 2.04 meters. Thus, four spatial distance values ​​are obtained: 2.08 meters, 1.88 meters, 1.84 meters, and 2.04 meters, which are used for subsequent signal weight allocation.

[0046] Furthermore, based on the calculated spatial distance information, the signal weight corresponding to each channel in the multi-channel acoustic signal is determined. Specifically, the weight allocation follows a distance-first principle, meaning that acoustic sensor channels closer to the fault source are assigned higher signal weights, while channels farther from the fault source are assigned lower signal weights. The theoretical basis for this allocation principle is that acoustic signals acquired by sensors closer to the fault source have a higher signal-to-noise ratio, and contain more complete and clearer fault characteristic information. Specifically, the weight allocation can employ mathematical relationships such as inverse proportional functions or Gaussian decay functions.

[0047] For example, the inverse proportional function directly reflects the distance-first principle through the inverse relationship between weight and distance. Let sensor A be 2.08 meters away, sensor B 1.88 meters away, sensor C 1.84 meters away, and sensor D 2.04 meters away. Calculate the reciprocal of the weights of each sensor: sensor A ≈ 1 / 2.08 ≈ 0.481, sensor B ≈ 1 / 1.88 ≈ 0.532, sensor C ≈ 1 / 1.84 ≈ 0.543, and sensor D ≈ 1 / 2.04 ≈ 0.490. Summing these four reciprocals gives a total of approximately 2.046. The final weight of each sensor is the ratio of its reciprocal to the sum of the reciprocals of the total weights: Sensor A's weight is approximately 0.481 / 2.046 ≈ 0.235, Sensor B's weight is approximately 0.532 / 2.046 ≈ 0.260, Sensor C's weight is approximately 0.543 / 2.046 ≈ 0.265, and Sensor D's weight is approximately 0.490 / 2.046 ≈ 0.240. Sensor C, which is closest to the fault source, received the highest weight of 0.265, while Sensor A, which is furthest away, received the lowest weight of 0.235.

[0048] Furthermore, based on the determined signal weights, a weighted average is calculated for the multi-channel acoustic signals. Each channel signal is multiplied by its corresponding weight and then summed to obtain the fused master signal. This master signal is an optimized signal that gathers the most valuable fault feature information from the multi-channel acoustic signals. By enhancing the contribution of channels closer to the fault source, the signal-to-noise ratio of the fault features is effectively improved, while suppressing environmental noise and irrelevant interference contained in channels far from the fault source. This provides a high-quality input signal for subsequent acoustic signature generation and fault diagnosis.

[0049] Since the aforementioned basic fusion method relies on the results of a single real-time acoustic localization, these results may be subject to interference in complex acoustic environments, resulting in a certain degree of uncertainty. To improve the robustness and accuracy of signal fusion, empirical probability weights based on historical fault statistics can be introduced on top of the basic spatial distance weights. This integrates the physical model of real-time localization with the statistical patterns of historical data, allowing the real-time fusion process to be corrected and optimized through these historical patterns.

[0050] Specifically, the method of multi-channel fusion of the multi-channel acoustic signals based on the spatial location information to obtain the main signal further includes:

[0051] Based on the probability distribution heatmap, the empirical probability weight of each channel in the multi-channel acoustic signal is determined.

[0052] Based on the spatial location information and the spatial arrangement of the multiple acoustic sensors, the spatial distance is calculated and normalized to determine the theoretical signal-to-noise ratio weight of each channel in the multi-channel acoustic signal.

[0053] The main signal is obtained by combining the empirical probability weights and the theoretical signal-to-noise ratio weights to perform a weighted average on the multi-channel acoustic signal.

[0054] First, the empirical probability weight of each channel in the multi-channel acoustic signal is determined based on the probability distribution heatmap. This empirical probability weight is a priori weight coefficient derived from historical fault statistics and is used to quantify the relative importance of each sensor's location in historical fault occurrence records. Specifically, based on the arrangement of each acoustic sensor in three-dimensional space, the probability value of the same spatial coordinate position in the probability distribution heatmap is found, and this probability value is normalized to serve as the empirical probability weight for the corresponding channel.

[0055] For example, suppose the probability values ​​of four sensors A, B, C, and D at their corresponding locations in the probability distribution heatmap are 0.15, 0.35, 0.40, and 0.10, respectively. After normalization, the empirical probability weights for each channel are: sensor A weight 0.15, sensor B weight 0.35, sensor C weight 0.40, and sensor D weight 0.10. This weight distribution reflects that the area where sensor C is located has the highest frequency of occurrence in historical fault records, therefore it should be given the largest weight in signal fusion.

[0056] Secondly, based on spatial location information and the spatial arrangement of multiple acoustic sensors, the spatial distance from the fault source to each sensor is calculated and normalized to determine the theoretical signal-to-noise ratio (SNR) weight for each channel in the multi-channel acoustic signal. This theoretical SNR weight is a weighting coefficient derived from the physical characteristics of sound wave propagation, characterizing the expected quality of fault feature components in the signals acquired by each sensor. The theoretical SNR weight is negatively correlated with spatial distance, reflecting the fundamental principle that sensors closer to the fault source have a higher SNR.

[0057] For example, suppose the distances from the four sensors to the fault source are: dA = 2.08m, dB = 1.88m, dC = 1.84m, and dD = 2.04m. The initial theoretical signal-to-noise ratio weights for each sensor are calculated using an inverse proportional function: the initial weight for sensor A is 1 / 2.08 ≈ 0.481, for sensor B it is 1 / 1.88 ≈ 0.532, for sensor C it is 1 / 1.84 ≈ 0.543, and for sensor D it is 1 / 2.04 ≈ 0.490.

[0058] The initial weight values ​​are normalized: the total is calculated to be 0.481 + 0.532 + 0.543 + 0.490 = 2.046. Each initial weight value is divided by this total to obtain the normalized theoretical signal-to-noise ratio weights: sensor A = 0.481 / 2.046 ≈ 0.235, sensor B = 0.532 / 2.046 ≈ 0.260, sensor C = 0.543 / 2.046 ≈ 0.265, sensor D = 0.490 / 2.046 ≈ 0.240.

[0059] Furthermore, the multi-channel acoustic signals are weighted and averaged by combining empirical probability weights and theoretical signal-to-noise ratio weights to obtain the main signal.

[0060] For example, a channel-by-channel multiplication fusion method is adopted, where the overall weight of each channel is equal to the product of its empirical probability weight and theoretical signal-to-noise ratio weight. Sensor A has an empirical probability weight of 0.15 and a theoretical signal-to-noise ratio weight of 0.235, resulting in an overall weight of 0.15 × 0.235 = 0.03525. Sensor B has two weights of 0.35 and 0.260, resulting in an overall weight of 0.35 × 0.260 = 0.091. Sensor C has two weights of 0.40 and 0.265, resulting in an overall weight of 0.40 × 0.265 = 0.106. Sensor D has two weights of 0.10 and 0.240, resulting in an overall weight of 0.10 × 0.240 = 0.024.

[0061] The obtained comprehensive weights are normalized to ensure that the sum of all weights is 1. Each comprehensive weight is divided by the sum 0.03525 + 0.091 + 0.106 + 0.024 = 0.25625 to obtain the final normalized weights: sensor A is 0.138, sensor B is 0.355, sensor C is 0.414, and sensor D is 0.093.

[0062] Finally, a weighted average of the multi-channel acoustic signals is calculated based on the final weights. The acoustic signal of each channel is multiplied by its corresponding final weight and then summed to generate the main signal. Specifically, this calculation process considers both historical fault statistics and the physical characteristics of sound wave propagation, ensuring that the generated main signal achieves optimal retention of key fault characteristics.

[0063] S30: Perform an adaptive window-length short-time Fourier transform on the main signal to obtain the time-frequency distribution, and combine the trainable Mel filter bank to convert the time-frequency distribution into a two-dimensional acoustic signature.

[0064] Specifically, time-frequency analysis is performed on the main signal obtained through multi-channel fusion, and the time-frequency distribution of the signal is obtained using an adaptive window-length short-time Fourier transform method. The core of this method lies in dynamically adjusting the analysis window length according to the local characteristics of the signal. A shorter window length is used to improve time resolution during periods of low signal-to-noise ratio or rapid frequency changes, while a longer window length is used to improve frequency resolution during periods of stable signal and high signal-to-noise ratio, thereby achieving an adaptive match between the accuracy of time-frequency analysis and the signal characteristics.

[0065] The obtained time-frequency distribution is then input into a trainable Mel filter bank for frequency scaling. This trainable Mel filter bank is a parameter-optimizable nonlinear frequency transformation module consisting of a set of bandpass filters covering the frequency bands sensitive to the human ear. Its initial parameters are set based on the standard Mel frequency scale and subsequently optimized using gradient descent. During training, the filter bank parameters are updated together with the subsequent diagnostic model, enabling the filter bank to adaptively learn and focus on the characteristic frequency bands most sensitive to the current transformer type and specific fault modes.

[0066] Specifically, through trainingable Mel filter banks, the linear spectrum is converted into a two-dimensional acoustic signature at the Mel frequency scale. This two-dimensional acoustic signature is a time-frequency representation optimized for perceptual characteristics and enhanced with fault features. It retains the dynamic characteristics of the sound signal over time while highlighting perceptual features valuable for fault diagnosis through Mel-scale transformation, providing optimized input feature representation for subsequent deep learning models.

[0067] The resulting two-dimensional acoustic signature map can effectively compress redundant frequency band information while amplifying characteristic frequency bands related to typical faults such as mechanical loosening and partial discharge, significantly improving the distinguishability of fault features.

[0068] S40: Perform adaptive denoising and key frequency band enhancement processing on the two-dimensional acoustic signature map to generate an enhanced acoustic signature map, and input it into a pre-trained probability distribution mapping module to obtain fault type information. Combine the enhanced acoustic signature map with the spatial location information and output it as a fault diagnosis result.

[0069] Specifically, adaptive denoising and key frequency band enhancement processing are performed on the two-dimensional acoustic signature to generate an enhanced acoustic signature, including:

[0070] Based on the pre-established background noise template library and the spatial location information, the main noise types are determined;

[0071] Dynamic noise suppression weights are configured based on the spectral overlap between the main noise type and the two-dimensional acoustic pattern, and the influence of the main noise type.

[0072] The two-dimensional speaker graph is adaptively denoised by combining spectral subtraction with the dynamic noise suppression weights.

[0073] Based on prior knowledge model learning, key frequency band information corresponding to the main noise types is obtained, and the adaptive denoising results are selectively enhanced to obtain the enhanced acoustic signature.

[0074] First, based on a pre-established background noise template library and spatial location information, the main noise types in the current scene are determined. The background noise template library is a database constructed by collecting typical noise samples from transformers under normal operating conditions over a long period and extracting their spectral characteristics. It includes common noise spectral characteristics found in transformer operating environments, such as cooling fan noise, core magnetostriction noise, and ambient electromagnetic hum. Spatial location information is used to assist in determining the physical location of the noise source. Combined with the noise template library, matching analysis is performed to accurately identify the main noise types that have the most significant impact on the current signal.

[0075] Secondly, dynamic noise suppression weights are configured based on the degree of spectral overlap between the identified main noise type and the two-dimensional acoustic signature, as well as the impact of this noise type on diagnostic accuracy in historical data. Dynamic noise suppression weight = α × spectral overlap + β × impact of the main noise type. Among them, spectral overlap represents the degree of spectral similarity between the noise type and the current voiceprint spectrum. It is obtained by calculating the correlation of the energy distribution of the two in the characteristic frequency band, and the value ranges between 0 and 1. The higher the degree of overlap, the larger the value. The influence of the main noise type represents the degree of interference of the noise type with the diagnostic accuracy. It is derived from the statistics of historical diagnostic data. Specifically, it is calculated as the average percentage decrease in diagnostic accuracy when the noise is present, normalized to the range of 0-1. The more severe the noise interference with diagnosis, the larger the value. α and β are preset weights, representing the proportion of spectral overlap and historical influence in the final weight, respectively, satisfying α+β=1. These weights are determined by the grid search method based on the noise suppression effect on the validation set, or empirically set to a fixed ratio based on the relative importance of spectral overlap and noise influence in historical data, such as α=0.6 and β=0.4.

[0076] The calculation formula assigns higher noise suppression weights to noise types with high spectral overlap, while noise types that have historically been proven to significantly interfere with fault diagnosis also receive higher noise suppression priorities. This weighting mechanism ensures the targeted nature of the noise suppression strategy.

[0077] Furthermore, by combining spectral subtraction with configured dynamic noise suppression weights, adaptive denoising processing is performed on the two-dimensional acoustic signature spectrum. Spectral subtraction is an audio noise reduction technique whose core principle is to estimate and subtract the spectral amplitude of the noise signal from the spectral amplitude of the noisy signal, thereby obtaining an estimate of the enhanced signal spectrum. During the spectral subtraction process, the suppression intensity of the noise spectrum is adjusted according to the dynamic noise suppression weights for different noise types. Specifically, for noise types with higher dynamic noise suppression weights, a larger spectral subtraction coefficient is used for deep suppression within their characteristic frequency bands; for noise types with lower dynamic noise suppression weights, a smaller spectral subtraction coefficient is used for mild suppression. Through this weight-based differentiated processing mechanism, precise suppression strategies can be implemented for different levels of noise interference, thereby achieving accurate and adaptive noise suppression and effectively avoiding excessive damage to valid fault characteristics.

[0078] Finally, based on a prior knowledge model, key frequency band information corresponding to the main noise types is learned and selectively enhanced on the adaptively denoised spectral results. The prior knowledge model is a machine learning model that associates historical fault cases with acoustic features. It is obtained by training on a large amount of labeled transformer fault acoustic signature data and is used to establish a mapping relationship between noise types and key fault feature frequency bands. This prior knowledge model can learn and acquire key frequency band information corresponding to the main noise types, thereby guiding the enhancement algorithm to accurately focus on feature regions with diagnostic value.

[0079] Furthermore, the adaptively denoised spectrum results are selectively enhanced based on key frequency band information. Specifically, this enhancement operation targets known fault characteristic frequency bands. For example, for core loosening faults, the characteristic is a harmonic group structure with a fundamental frequency of 100Hz. The enhancement operation needs to identify and lock the narrow band where each harmonic is located within the range of 50Hz to 1500Hz. For winding deformation faults, the characteristic is a specific resonant mode appearing in the range of 500Hz to 2000Hz. The precise center frequency and bandwidth need to be determined through modal analysis. Within the identified key frequency bands, a power-law based nonlinear gain function is used to locally boost the spectral amplitude. The boost intensity is dynamically adjusted according to the signal-to-noise ratio of the signal within the frequency band. At the same time, a feature sharpening algorithm based on Gabor filters is applied to enhance the edge clarity and continuity of fault features in the time-frequency domain. Then, the selectively enhanced key frequency bands are seamlessly integrated with the remaining retained frequency bands to ensure that the enhanced spectrum remains smooth and continuous in the transition region.

[0080] The enhanced voiceprint map obtained after the above processing can effectively improve the representation strength and visual saliency of fault features in the time-frequency domain, thereby providing high-quality input data with higher feature discrimination and less noise interference for subsequent deep learning models.

[0081] Furthermore, the enhanced voiceprint map is input into a pre-trained probability distribution mapping module to obtain fault type information. This probability distribution mapping module is a multi-branch decision system built on a deep neural network, used to integrate voiceprint features and spatial location information to achieve accurate fault type judgment and verification.

[0082] Specifically, such as Figure 2 As shown, the probability distribution mapping module includes at least:

[0083] The first classification branch is used to output a first probability distribution of the fault type based on the enhanced acoustic signature spectrum;

[0084] An intermediate discrimination layer is used to perform prior reliability discrimination on the first probability distribution;

[0085] The second classification branch is used to output a second probability distribution of the fault type based on the enhanced acoustic signature and the spatial location information;

[0086] The discriminant output layer is used to compare the first probability distribution with the second probability distribution and output the fault type information.

[0087] Specifically, the probability distribution mapping module includes at least the following components: The first classification branch is a feature extraction and classification network that takes the enhanced acoustic signature as input. It learns and abstracts acoustic feature patterns related to the fault from the enhanced acoustic signature and outputs a first probability distribution of the fault type, representing a preliminary diagnostic result based on acoustic signature features. The intermediate discrimination layer is a logical judgment unit based on prior knowledge rules, used to determine the reliability of the first probability distribution. Its discrimination criteria may include the consistency between the first probability distribution and the probability of the fault type derived from spatial location information. The second classification branch is a classification network supporting multi-source information fusion, used to simultaneously process the enhanced acoustic signature and spatial location information. By fusing these two types of features, it outputs a second probability distribution of the fault type, reflecting the diagnostic conclusion under the combined effect of acoustic and spatial features. The discrimination output layer is a decision optimization and fusion module, used to compare and evaluate the similarity and differences between the first and second probability distributions, and performs weighted fusion according to preset confidence rules, ultimately outputting fault type information that has undergone multi-dimensional verification and has higher reliability.

[0088] Furthermore, for example, the probability distribution mapping module employs supervised learning during training. First, a training dataset is constructed. A large amount of enhanced acoustic signature data and corresponding spatial location information are collected from prior fault records as input samples, and expert-verified fault type labels are obtained as supervision signals. The entire dataset is divided into training, validation, and test sets according to a predetermined ratio. The training process adopts a phased strategy. In the first phase, the first classification branch is trained independently, using enhanced acoustic signature data as input and fault type labels as supervision signals, optimizing network parameters by minimizing the cross-entropy loss function. In the second phase, the second classification branch is trained, simultaneously inputting enhanced acoustic signature data and spatial location information, again optimizing network parameters using fault type labels as supervision signals.

[0089] The parameters of the intermediate discriminant layer and the discriminant output layer are determined through end-to-end fine-tuning. After the initial training of the first two branches is completed, the entire probability distribution mapping module is jointly trained. At this time, the loss function comprehensively considers the consistency between the first probability distribution, the second probability distribution and the true label, as well as the similarity constraint between the two distributions.

[0090] The key hyperparameter learning rate was set to 0.001 during training, with 200 training epochs and a batch size of 32. The Adam optimizer was used for parameter updates. Model performance was monitored using a validation set. Training was terminated early when the validation set accuracy no longer improved after several consecutive training epochs. Finally, the model performance was evaluated on the test set to ensure it reached the predetermined accuracy, such as 90%, before deployment.

[0091] Furthermore, the information is input into a pre-trained probability distribution mapping module to obtain fault type information, including:

[0092] Input the enhanced voiceprint map into the first classification branch to obtain the first probability distribution;

[0093] Based on the spatial location information, the corresponding taboo fault type is obtained, and based on the taboo fault type, the cumulative taboo probability of the first probability distribution belonging to the taboo fault type is statistically analyzed.

[0094] If the cumulative taboo probability is greater than a preset taboo probability threshold, the first probability distribution is iteratively updated based on the first classification branch until it is less than the taboo probability threshold.

[0095] Specifically, the enhanced voiceprint spectrum is input into the first classification branch, where feature extraction and pattern recognition are performed, outputting a first probability distribution of the fault type. This first probability distribution is a multi-dimensional vector, with each dimension corresponding to the probability of a specific fault type, and the sum of all probabilities is 1. Next, based on the determined spatial location information, prohibited fault types incompatible with the spatial location are retrieved from a pre-built fault type and spatial location association knowledge base. This fault type and spatial location association knowledge base is constructed based on transformer structural principles and prior fault records, defining fault types that are physically impossible to occur within a specific spatial region.

[0096] Furthermore, based on the obtained list of forbidden fault types, statistical analysis is performed on the first probability distribution, and the cumulative forbidden probability is obtained by calculating the sum of the probability values ​​corresponding to all forbidden fault types. This cumulative forbidden probability reflects the degree of conflict between the preliminary diagnostic results and prior knowledge of spatial location.

[0097] Furthermore, the cumulative taboo probability is compared with a preset taboo probability threshold. This taboo probability threshold is set based on the transformer's structural characteristics and the statistical analysis results of misjudged cases in prior fault records, for example, set to 0.1. If the cumulative taboo probability is greater than this taboo probability threshold, it indicates that there are components in the preliminary diagnosis that significantly contradict physical constraints. In this case, the first probability distribution is considered unreliable, and iterative prediction is performed to update the first probability distribution until the cumulative taboo probability drops below the taboo probability threshold, ensuring that the diagnosis results conform to the physical constraints.

[0098] Furthermore, inputting the enhanced voiceprint map into a pre-trained probability distribution mapping module to obtain fault type information also includes:

[0099] If the cumulative taboo probability is less than a preset taboo probability threshold, the enhanced voiceprint map and the spatial location information are input into the second classification branch to obtain the second probability distribution;

[0100] Remove the taboo fault types from the first probability distribution and update the first probability distribution by normalization;

[0101] Calculate the distribution similarity between the updated first probability distribution and the second probability distribution. If the distribution similarity satisfies the confidence distribution similarity, then fuse the updated first probability distribution and the second probability distribution to output the fault type information. The confidence distribution similarity is a distribution similarity threshold based on confidence constraints.

[0102] Finally, based on the diagnostic accuracy of the fault diagnosis results, the trainable Mel filter bank, adaptive denoising, and probability distribution mapping module are jointly optimized using the backpropagation algorithm.

[0103] Specifically, if the cumulative taboo probability is less than a preset taboo probability threshold, the following steps are performed to obtain the final fault type information:

[0104] First, the enhanced acoustic signature and spatial location information are input into the second classification branch. This second classification branch, through the fusion processing of multi-source information, outputs a second probability distribution of the fault type, which integrates both acoustic features and spatial location evidence.

[0105] Secondly, the probability values ​​identified as taboo fault types are removed from the first probability distribution. The remaining probability values ​​are then normalized until their sum is equal to one, resulting in an updated first probability distribution. The distribution similarity between the updated first and second probability distributions is then calculated. Specifically, this similarity calculation can employ methods such as cosine similarity or Jensen-Shannon divergence to quantify the degree of consistency between the two diagnostic results.

[0106] For example, suppose the updated first probability distribution P is [0.6, 0.3, 0.1, 0.0] and the second probability distribution Q is [0.5, 0.4, 0.1, 0.0], where each dimension corresponds to one of the four fault types: winding deformation, core loosening, partial discharge, and normal state. Cosine similarity is used to calculate the degree of consistency between the two probability distributions. The calculation process includes calculating the dot product of the two vectors, calculating their respective magnitudes, and finally dividing the dot product by the product of the two magnitudes.

[0107] The dot product of the two vectors is (0.6×0.5)+(0.3×0.4)+(0.1×0.1)+(0.0×0.0)=0.43. The magnitude of the first probability distribution P is √(0.6²+0.3²+0.1²+0.0²)=√(0.36+0.09+0.01+0.00)≈0.678. The magnitude of the second probability distribution Q is √(0.5²+0.4²+0.1²+0.0²)=√(0.25+0.16+0.01+0.00)≈0.648. The similarity between the first and second probability distributions is 0.43 / (0.678×0.648)≈0.979.

[0108] Then, the calculated distribution similarity is compared with a preset confidence distribution similarity threshold. This confidence distribution similarity threshold is a threshold value derived from statistical analysis of a large amount of validation data, used to determine whether the two diagnostic results have sufficient consistency; for example, it is set to 0.85. If the distribution similarity meets the threshold requirement, the updated first probability distribution and the second probability distribution are weighted and fused. The weight coefficients are set according to the independent diagnostic accuracy of the two classification branches on the historical validation dataset. For example, the weight coefficient of the first probability distribution is set to 0.4, and the weight coefficient of the second probability distribution is set to 0.6 accordingly. Finally, the fusion result is output as reliable fault type information.

[0109] In summary, this decision-making mechanism based on dual-branch collaborative verification effectively improves the accuracy and reliability of fault diagnosis results.

[0110] Finally, based on the diagnostic accuracy of the fault diagnosis results, the trainable Mel filter bank, adaptive denoising module, and probability distribution mapping module are jointly optimized using the backpropagation algorithm to form a self-improving closed-loop system. This optimization process aims to improve the final diagnostic accuracy globally. The gradient of the diagnostic error relative to the parameters of each module is calculated using the backpropagation algorithm, and all trainable parameters are updated synchronously using gradient descent.

[0111] Specifically, the center frequency and bandwidth parameters of the trainable Mel filter bank are optimized to adaptively adjust to the frequency band most sensitive to the current transformer fault; the dynamic noise suppression weight parameters in the adaptive denoising module are optimized to more accurately balance noise suppression and feature preservation; and the network weight parameters in the probability distribution mapping module are optimized to improve its feature extraction and classification decision-making capabilities. A learning rate decay strategy is adopted during the joint optimization process, with an initial learning rate of 0.001, decaying to 0.5 of the original value every 50 training rounds. Training is terminated early when the validation set loss no longer decreases for 10 consecutive rounds. The parameters of all modules are jointly fine-tuned end-to-end under a unified training framework, achieving a globally optimal match between feature extraction, noise reduction enhancement, and fault classification. This joint optimization mechanism overcomes the limitations of isolated optimization of each module in traditional serial processing, thereby improving the performance and adaptability of the entire diagnostic system.

[0112] In summary, the embodiments of this application have at least the following technical effects:

[0113] Compared to existing technologies, this application firstly achieves accurate spatial estimation of the fault location inside the transformer through phase difference localization technology using multi-channel acoustic signals, and then performs intelligent signal fusion based on this localization information, improving the quality of the main signal used for diagnosis. Secondly, it employs adaptive time-frequency analysis and a trainable Mel filter bank to dynamically optimize the time-frequency resolution of the acoustic signature spectrum and focus on key fault frequency bands, enhancing feature representation capabilities. Thirdly, it combines spatial location information to perform guided denoising and feature enhancement on the acoustic signature spectrum, effectively suppressing background interference and highlighting fault components. Finally, it achieves deep fusion and mutual verification of fault identification and spatial localization through a dual-branch probability distribution mapping module, and establishes a closed-loop optimization mechanism to enable system parameters to continuously improve themselves with diagnostic accuracy as the goal, thereby comprehensively improving the accuracy, reliability, and adaptability of transformer fault diagnosis.

[0114] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0115] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0116] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A transformer fault diagnosis method fusing phase positioning and adaptive voiceprint mapping, characterized in that, The method comprises the following steps: Synchronously collecting multiple sound wave signals through multiple sound wave sensors arranged in space; Calculating phase deviation information based on the multiple sound wave signals, estimating the spatial position information of the internal fault source of the transformer by combining a sound source positioning algorithm, and performing multi-channel fusion on the multiple sound wave signals according to the spatial position information to obtain a main signal; Performing adaptive window length short-time Fourier transform on the main signal to obtain a time-frequency distribution, and converting the time-frequency distribution into a two-dimensional voiceprint spectrum by combining a trainable Mel filter bank; Performing adaptive denoising and key frequency band enhancement processing on the two-dimensional voiceprint spectrum to generate an enhanced voiceprint spectrum, and inputting the enhanced voiceprint spectrum into a pre-trained probability distribution mapping module to obtain fault type information, and merging the fault type information with the spatial position information to output a fault diagnosis result; The probability distribution mapping module at least comprises: A first classification branch for outputting a first probability distribution of a fault type according to the enhanced voiceprint spectrum; An intermediate discriminant layer for performing prior reliability discrimination on the first probability distribution; A second classification branch for outputting a second probability distribution of a fault type according to the enhanced voiceprint spectrum and the spatial position information; A discriminant output layer for comparing the first probability distribution and the second probability distribution to output the fault type information; The prior reliability discrimination on the first probability distribution comprises: According to the spatial position information, a taboo fault type corresponding to the spatial position information is obtained, and a cumulative taboo probability belonging to the taboo fault type in the first probability distribution is statistically analyzed according to the taboo fault type; If the cumulative taboo probability is greater than a preset taboo probability threshold, the first probability distribution is iteratively updated based on the first classification branch until it is less than the taboo probability threshold; If the cumulative taboo probability is less than the preset taboo probability threshold, the enhanced voiceprint spectrum and the spatial position information are input into the second classification branch to obtain the second probability distribution.

2. The transformer fault diagnostic method of claim 1, wherein the fusion of the phase positioning and the adaptive voiceprint map is performed by using a neural network. The calculation of the phase deviation information based on the multiple sound wave signals and the estimation of the spatial position information of the internal fault source of the transformer by combining a sound source positioning algorithm comprise: Pretreatment of the multiple sound wave signals, including filtering and synchronization; Taking any two sound wave sensor channels as a channel pair, converting the signals to the frequency domain to calculate the cross-power spectrum and perform phase normalization processing, combining the inverse transform to obtain the cross-correlation function, and taking the time delay corresponding to the peak value as the phase deviation information; Iteratively obtaining the phase deviation information of multiple channel pairs to form a phase difference matrix, and estimating the position coordinates of the sound source in the three-dimensional space by combining the spatial arrangement of multiple sound wave sensors through a geometric inversion model to output the spatial position information.

3. The transformer fault diagnostic method of claim 2, wherein the fusion of the phase positioning and the adaptive voiceprint map is performed by using a neural network. The estimation of the position coordinates of the sound source in the three-dimensional space through the geometric inversion model and the output of the spatial position information comprise: Obtaining prior fault records, and obtaining a probability distribution heat map of the fault source based on statistical analysis; Superimposing and analyzing the position coordinates output by the probability distribution heat map and the geometric inversion model to determine the spatial position information.

4. The transformer fault diagnostic method of claim 1, wherein the fusion of the phase positioning and the adaptive voiceprint map is performed by using a neural network. According to the spatial position information, the multi-channel sound wave signals are fused to obtain a main signal, including: Based on the spatial position information and the spatial arrangement of the plurality of sound wave sensors, a plurality of spatial distance information is calculated and obtained; According to the plurality of spatial distance information, a signal weight of each channel in the multi-channel sound wave signal is determined, wherein the signal weight is negatively correlated with the spatial distance information; Based on the signal weight, the multi-channel sound wave signal is weighted and averaged to obtain the main signal.

5. The transformer fault diagnostic method of claim 3, wherein the fusion of the phase positioning and the adaptive voiceprint map is performed by using a neural network. According to the spatial position information, the multi-channel sound wave signals are fused to obtain a main signal, further including: Based on the probability distribution thermodynamic map, an empirical probability weight of each channel in the multi-channel sound wave signal is determined; Based on the spatial position information and the spatial arrangement of the plurality of sound wave sensors, the spatial distance is calculated and normalized to determine the theoretical signal-to-noise ratio weight of each channel in the multi-channel sound wave signal; The empirical probability weight and the theoretical signal-to-noise ratio weight are combined to perform weighted averaging on the multi-channel sound wave signal to obtain the main signal.

6. The transformer fault diagnostic method of claim 1, wherein the fusion of the phase positioning and the adaptive voiceprint map is performed by using a neural network. The two-dimensional acoustic fingerprint spectrum is adaptively denoised and key frequency band enhanced to generate an enhanced acoustic fingerprint spectrum, including: Based on the pre-established background noise template library and the spatial position information, a main noise type is determined; According to the main noise type, the frequency spectrum overlap degree of the two-dimensional acoustic fingerprint spectrum, and the influence degree of the main noise type, a dynamic noise suppression weight is configured; The two-dimensional acoustic fingerprint spectrum is adaptively denoised by combining the spectral subtraction method and the dynamic noise suppression weight; Based on the prior knowledge model learning, the key frequency band information corresponding to the main noise type is obtained, and the adaptive denoising result is selectively enhanced to obtain the enhanced acoustic fingerprint spectrum.

7. The transformer fault diagnostic method of claim 1, wherein the fusion of the phase positioning and the adaptive voiceprint map is performed by using a neural network. Input into a pre-trained probability distribution mapping module to obtain fault type information, including: Input the enhanced acoustic fingerprint spectrum into the first classification branch to obtain the first probability distribution; The taboo fault type is excluded from the first probability distribution, and the first probability distribution is updated by normalization; The distribution similarity of the updated first probability distribution and the second probability distribution is calculated, and if the distribution similarity meets the confidence distribution similarity, the updated first probability distribution and the second probability distribution are fused, and the output is the fault type information, wherein the confidence distribution similarity is a distribution similarity threshold based on confidence constraint.

8. The transformer fault diagnostic method of claim 1, wherein the fusion of the phase positioning and the adaptive voiceprint map is performed by using a neural network. Based on the diagnostic accuracy of the fault diagnosis result, the trainable Mel filter bank, adaptive denoising, and the probability distribution mapping module are jointly optimized by a back propagation algorithm.

Images