A cross transformer acoustic vibration migration diagnostic method and system that decouples source backbone from operating condition confidence calibration

Abstract

The present application relates to the technical field of power transformer soundprint vibration signal processing and intelligent fault diagnosis, and discloses a cross-transformer sound-vibration migration diagnosis method and system with source backbone preservation and working condition decoupling confidence calibration. The method collects soundprint signals and vibration signals of source transformers and target transformers during operation, and establishes a soundprint-vibration sample group; according to time information, sampling point number or collection time length, the actual sampling rate is determined, and standardized soundprint spectrum and vibration spectrum are generated according to real time windows to form double-layer sound-vibration input; further, load level, oil temperature, environmental noise, channel signal-to-noise ratio and frequency band energy distribution and other working condition factors are extracted to construct working condition embedding vectors, and working condition disturbance and fault response in sound-vibration characteristics are decoupled; a sound-vibration diagnosis model is trained on source transformer samples, and the source model feature backbone parameter state is maintained when migrated to target transformers; the global statistics and working condition subdomain statistics of the target domain are calculated using a small amount of target samples, the target domain feature distribution is corrected through a conditional gradual calibration layer, and only the calibration layer, light adapter, working condition decoupling correction parameters and classification correction parameters are updated; combined with the migration confidence gate, unknown fault rejection and confidence sample pool closed-loop updating mechanism, the operation state or fault category is output. The present application can reduce the distribution deviation caused by different transformer installation environments, sensor positions, operation conditions and insufficient samples, and improve the stability, reliability and adaptability of cross-device sound-vibration fault migration diagnosis.

Description

Technical Field

[0001] This invention relates to the fields of power equipment state sensing, industrial acoustic signal processing, vibration signal analysis, transfer learning, and intelligent fault diagnosis. In particular, it relates to a cross-transformer acoustic vibration transfer diagnosis method, system, electronic device, and computer-readable storage medium that addresses the problems of source backbone preservation, operating condition decoupling, conditional asymptotic calibration, and confidence closed-loop update for different transformers, including differences in acoustic vibration distribution, operating conditions, and insufficient target samples. Background Technology

[0002] Power transformers play a crucial role in power transmission and distribution systems, performing voltage transformation and energy transfer. Their operational status directly impacts the reliability of the power grid. When a transformer experiences winding loosening, core clamping abnormalities, short-circuit impacts, overload operation, DC bias, or insulation abnormalities, it typically causes changes in the tank's vibration response and external acoustic radiation characteristics. Acoustic signals more readily reflect discharge, impact, and external radiation phenomena, while vibration signals more readily reflect the mechanical responses of the core, windings, and structural components. These two types of signals are complementary in fault diagnosis.

[0003] Existing acoustic-vibration fusion diagnostic solutions typically build models based on training samples from the same transformer or under the same acquisition conditions. While these solutions can achieve good recognition results on the same equipment, when the model is applied to a target transformer with different capacity, structure, installation foundation, sensor location, load level, and environmental noise, the amplitude distribution, frequency band energy ratio, and channel response relationships of the acoustic-vibration spectrum change. This leads to a decrease in diagnostic stability when the model trained on the source transformer is directly transferred.

[0004] On the other hand, obtaining fault samples from target transformers is difficult, especially since the number of early and typical fault samples is limited. Re-collecting a large number of labeled samples and retraining the complete model for each target transformer is not only costly and time-consuming, but also prone to overfitting due to the limited sample size. Therefore, a transfer diagnostic scheme is needed that can retain the general acoustic and vibration fault characteristics of the source transformer model while utilizing a small number of samples from the target transformer for distribution calibration and lightweight adaptation.

[0005] Furthermore, transformer acoustic and vibration signals are affected not only by fault type but also by load level, load current, oil temperature, cooling system operating status, installation foundation, sensor coupling status, and environmental noise. In cross-transformer migration diagnostics, even if two transformers are in the same fault category, their acoustic and vibration spectra may exhibit different frequency band energy distributions due to different operating conditions. Conversely, normal samples under different operating conditions may approach fault samples in local frequency bands, causing the model to misjudge changes in operating conditions as changes in fault conditions. Existing migration diagnostic methods typically focus on aligning the overall feature distribution between the source and target domains, with less emphasis on distinguishing the influence of non-fault factors such as load, oil temperature, environmental noise, and channel quality on acoustic and vibration characteristics. They also lack mechanisms for rejecting, verifying, and closed-loop calibrating low-confidence target samples. When the number of target transformer samples is small, the operating condition coverage is insufficient, or the acoustic and vibration channel quality is unstable, simple overall statistical calibration can easily lead to negative migration, affecting the stability and reliability of the diagnostic results. Summary of the Invention

[0006] The purpose of this invention is to provide a cross-transformer acoustic-vibration migration diagnosis method and system that maintains the source backbone and decouples the operating condition confidence calibration, so as to solve the problems of inconsistent acoustic-vibration spectrum distribution between different transformers, few labeled samples of target transformers, easy confusion between changes in operating conditions and fault characteristics, the influence of acoustic-vibration channel quality differences on fusion diagnosis, and the lack of a closed-loop calibration mechanism for low-confidence migration results.

[0007] To achieve the above objectives, this invention first pairs the acoustic signature and vibration signals of the source transformer and the target transformer, and generates acoustic signature and vibration spectra at a unified physical time scale based on the actual sampling rate and the real time window. Then, the spectra are standardized into a double-layer acoustic-vibration input of uniform size, and operating condition factors such as load level, oil temperature, ambient noise intensity, acoustic signature channel signal-to-noise ratio, vibration channel signal-to-noise ratio, and frequency band energy distribution are further extracted to construct an operating condition embedding vector. An acoustic-vibration diagnostic model, including an operating condition decoupling unit, a conditional progressive calibration layer, and a lightweight adapter, is trained on the source transformer sample. The model learns general acoustic and vibration fault characteristics while suppressing disturbances in non-faulty operating conditions. When migrating to the target transformer, the model maintains the state of the main parameters of the source model, uses a small number of samples from the target transformer to estimate the global statistics of the target domain and the statistics of the operating condition subdomain, maps the features of the target domain to the statistical range that the source model can adapt to through the conditional asymptotic calibration layer, and then only updates the target domain conditional asymptotic calibration layer, lightweight adapter, operating condition decoupling correction parameters and classification correction parameters. Finally, it uses the migration confidence gating, unknown fault rejection and confidence sample pool closed-loop update mechanism to complete the small-sample migration diagnosis of the target transformer.

[0008] Compared with the prior art, the present invention has at least the following beneficial effects:

[0009] (1) The process of pairing acoustic and vibration files, estimating sampling rate, converting time and frequency in real time window and constructing standard spectrum is retained, so that data with different sampling rates and different acquisition durations can enter the migration diagnostic model under a unified physical scale.

[0010] (2) By preserving the general acoustic and vibration fault features learned in the source transformer through the source model backbone parameter preservation method, the small number of samples of the target transformer are avoided from causing excessive updates to the complete model.

[0011] (3) The characteristic mean and standard deviation of the source domain and the target domain are statistically mapped by the target domain progressive calibration layer to reduce the distribution offset caused by differences in installation conditions, sensor response and operating load between different transformers.

[0012] (4) A small number of parameters of the target transformer are updated through a lightweight adapter, reducing the number of samples and computational resources required for transfer training.

[0013] (5) By assessing the reliability of automatic diagnostic results through transfer reliability gating, output suppression or prompting supplementary sampling is performed on low reliability samples to reduce the risk of misdiagnosis caused by negative transfer.

[0014] (6) By extracting operating condition factors and constructing operating condition embedding vectors, non-fault factors such as load level, oil temperature, environmental noise and channel quality are introduced into the target domain calibration process, so that the model can adopt different calibration intensities under different operating conditions, reducing the probability that changes in operating conditions are misjudged as changes in faults.

[0015] (7) By using the acoustic channel quality factor, vibration channel quality factor and frequency band energy distribution to decouple and enhance the dual-layer acoustic and vibration input under working conditions, the model can improve the contribution of high-quality channels and fault-sensitive frequency bands, and reduce the impact of noise channels, sensor coupling anomalies and environmental disturbances on diagnostic results.

[0016] (8) By using migration confidence gating, unknown fault rejection and confidence sample pool closed-loop update mechanism, high confidence target samples can continuously participate in target domain statistical calibration, while low confidence samples enter the supplementary sampling or manual review process, thereby reducing the risk of misdiagnosis caused by negative migration and improving the adaptability of the target transformer under long-term operating conditions. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the cross-transformer acoustic and vibration migration diagnostic method for source backbone maintenance and operating condition decoupling confidence calibration.

[0018] Figure 2 This is a schematic diagram of the source backbone maintenance and operating condition decoupled confidence calibration diagnostic network structure of the present invention.

[0019] Figure 3This is a schematic diagram of the cross-transformer acoustic and vibration migration diagnostic system module of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the described embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Where there is no conflict, features in the following embodiments can be combined with each other.

[0021] I. Accessing Voiceprint Vibration Files and Establishing Sample Groups

[0022] like Figure 1 As shown, the acoustic signature and vibration signal files of the source transformer and target transformer under different operating conditions are first obtained. The source transformer refers to a transformer with a large number of labeled acoustic and vibration samples that can be used to train the initial diagnostic model; the target transformer refers to the transformer to be migrated for application, which has fewer labeled fault samples or only a small number of calibration samples.

[0023] In one embodiment, the acoustic print files and vibration files are stored in the acoustic print directory and vibration directory, respectively. Alternatively, they can be stored in different subdirectories according to device number, operating status, acquisition date, or acquisition batch. The system reads the file path, file name, acquisition number, acquisition time, or operating status identifier to establish an acoustic print-vibration sample group. If the sample comes from the source transformer and has a category directory, the category directory is used as a fault label; if the sample comes from the target transformer but no fault label is provided, it can be used in the target domain statistical calculation and prediction stage.

[0024] II. Estimation of Actual Sampling Rate and Generation of Real Time Window Spectrum

[0025] After reading each signal file, the data in the form of a data table, array, or sequence is converted into a one-dimensional numerical sequence. If the file contains time and amplitude columns, the valid values ​​in the time column are filtered, and the actual sampling rate is determined based on the start and end times of the time column and the number of valid sampling points. If the file only contains amplitude sequences, the actual sampling rate is calculated based on the number of valid sampling points N and the preset acquisition duration T.

[0026] To ensure consistency in the physical meaning of data with different sampling rates in time-frequency analysis, this invention does not use a fixed sampling point window length, but instead uses a fixed real time window to determine the short-time Fourier transform parameters. Let the automatically identified sampling rate be fs, and the preset real time window be Tw. Then, the short-time Fourier transform window length nperseg is determined by the following formula:

[0027]

[0028] For example, when At a sampling rate of 0.25s, the window length for samples with a sampling rate of 4200Hz is approximately 1050 points, and the window length for samples with a sampling rate of 65536Hz is approximately 16384 points, thus making the frequency resolution of both types of data close. The number of overlapping points is determined by the following formula:

[0029]

[0030] in The overlap ratio is preferably 0.5. If the signal length is less than the window length, the signal is padded with zeros or the window length is made not to exceed the signal length; if the window length is too small, a minimum window length threshold is set to ensure that the short-time Fourier transform can be performed stably.

[0031] Before performing the short-time Fourier transform, the acoustic and vibration signals can be processed to remove the mean, and the amplitude scale can be adjusted according to the maximum absolute amplitude or the root mean square amplitude to reduce the amplitude deviation caused by differences in sensor sensitivity and installation position.

[0032] III. Standard Spectrum Generation and Dual-Layer Acoustic Input

[0033] To ensure consistent input dimensions for subsequent networks, this invention performs uniform processing on the initial spectral surface. First, frequency components less than or equal to 20Hz are removed to reduce the influence of DC components, low-frequency drift, and environmental low-frequency disturbances; then, the target frequency band between 20Hz and 2000Hz is retained.

[0034] The above processing is performed on the voiceprint signal and vibration signal in the same sample pair respectively to obtain a fixed time-frequency map of the voiceprint. and vibration fixed time frequency diagram Then stack them according to channel dimension:

[0035]

[0036] The first channel is the acoustic signature time-frequency graph, and the second channel is the vibration time-frequency graph.

[0037] IV. Operating condition factor extraction and operating condition decoupling enhancement

[0038] After obtaining the dual-layer acoustic and vibration input, this invention further extracts the operating condition factors corresponding to the target sample. If the acquisition system provides operating records, external operating condition factors such as load rate, load current, oil temperature, ambient temperature, acquisition period, and cooling device operating status are directly read; if the acquisition system does not provide complete operating records, internal operating condition factors are calculated from the acoustic spectrum and vibration spectrum, including the acoustic channel signal-to-noise ratio, vibration channel signal-to-noise ratio, energy proportion of different frequency bands, acoustic and vibration channel correlation coefficient, low-frequency drift intensity, and impact sparsity.

[0039] Let the set of operating condition factors corresponding to the sample be:

[0040] in, Indicates load rate, Indicates the load current. Indicates oil temperature. Indicates ambient temperature. This indicates the signal-to-noise ratio of the voiceprint channel. Indicates the signal-to-noise ratio of the vibration channel. This indicates the frequency band energy distribution of the acoustic signature channel. This indicates the frequency band energy distribution of the vibration channel. This represents the correlation coefficient between the acoustic channel and the vibration channel.

[0041] After normalizing the above operating condition factors, an operating condition embedding vector is constructed:

[0042] in, This indicates normalization processing. and For learnable parameters, This represents the operating condition embedding vector. The operating condition embedding vector is not directly used as the basis for fault category discrimination, but rather to control the fusion weights of the acoustic and vibration channels, the frequency band response correction weights, and the target domain conditional asymptotic calibration coefficients, enabling the model to employ different feature calibration intensities under different load, oil temperature, and noise conditions.

[0043] Furthermore, based on the working condition embedding vector Voiceprint channel quality factor and vibration channel quality factor Generate the fusion weights for the voiceprint channel and the vibration channel:

[0044]

[0045] in, Indicates the voiceprint channel fusion weights. Indicates the vibration channel fusion weight. and These are learnable parameters.

[0046] Generate frequency band response correction weights based on the operating condition embedding vector and the prior knowledge of the fault-sensitive frequency band:

[0047]

[0048] Where A represents the frequency band response correction weight. This indicates a priori information about the fault-sensitive frequency band. W represents the Sigmoid function. a and b a These are learnable parameters.

[0049] The input for acoustic and vibration enhancement after decoupling under operating conditions is represented as follows:

[0050] Where X′ represents the acoustic-vibration enhancement input after decoupling of the operating conditions. In this way, when the acoustic channel is greatly affected by environmental noise, the contribution of the acoustic channel in the fusion diagnosis is reduced; when the vibration sensor is loosely installed or the vibration channel has strong low-frequency drift, the contribution of the vibration channel in the fusion diagnosis is reduced, thereby avoiding the deviation of the overall diagnostic results caused by the abnormality of a single channel.

[0051] V. Source Model Training and Backbone Parameter Preservation

[0052] An acoustic-vibration fault diagnosis model is trained using labeled acoustic-vibration enhancement inputs from a source transformer. The model includes a feature backbone, a condition decoupling unit, a target domain conditional asymptotic calibration layer, a lightweight adapter, and a classification output layer. During the source training phase, the feature backbone, condition decoupling unit, and classification output layer can be trained simultaneously, enabling them to learn the acoustic signature spectrum, vibration spectrum, condition embedding vector, and the correspondence between them.

[0053] The feature backbone includes a spectral texture extraction unit, an acoustic-vibration channel interaction unit, and a frequency-time region focusing unit; the operating condition decoupling unit adjusts the fusion contribution of the acoustic and vibration channels based on the operating condition embedding vector, the acoustic channel quality factor, and the vibration channel quality factor, and corrects the response weights of different frequency bands.

[0054] After the source model is trained, record the source domain global statistical benchmark and source domain operating condition subdomain statistical benchmark of the source model at multiple feature layers. For the l-th feature layer, the source domain global feature mean and source domain global feature standard deviation are calculated using the following formula:

[0055]

[0056] Where represents the output of the i-th source domain sample in the l-th feature layer, and GAP represents global average pooling. Indicates the number of source transformer samples. and Let $\mathbf{l}$ and $\mathbf{l}$ represent the mean and standard deviation of the global features in the source domain of the l-th feature layer, respectively.

[0057] For the k-th source domain operating condition subdomain, its statistical baseline is calculated using the following formula:

[0058]

[0059] Where k represents the operating condition subdomain number, This represents the number of samples in the k-th source domain operating condition subdomain. and These represent the feature mean and feature standard deviation of the k-th source domain operating condition subdomain on the l-th feature layer, respectively.

[0060] When migrating to the target transformer, the source model feature backbone maintains the parameter state of the source training phase and does not perform overall reverse updates during the target domain few-sample update process.

[0061] VI. Target Domain Statistical Calculation and Conditional Asymptotic Calibration Layer

[0062] A small number of acoustic and vibration samples of the target transformer are input into the feature backbone of the source model and the operating condition decoupling unit to obtain the target domain feature output. The global statistics of the target domain and the operating condition subdomain statistics of the target domain are calculated based on the operating condition embedding vector of the target samples.

[0063] For the l-th feature layer, the mean and standard deviation of the global features in the target domain are calculated using the following formula:

[0064] in, Indicates the number of target transformer samples. This represents the output of the j-th target domain sample in the l-th feature layer.

[0065] For the k-th target domain operating condition subdomain, its characteristic mean and characteristic standard deviation are calculated using the following formula:

[0066]

[0067] in, This represents the number of samples in the k-th target domain subdomain.

[0068] The target domain conditional asymptotic calibration layer maps target domain features based on the source domain global statistical benchmark, the source domain operating condition subdomain statistical benchmark, the target domain global statistics, and the target domain operating condition subdomain statistics. For the original features of the target domain at layer l... First, perform global statistical calibration:

[0069]

[0070] in and These are learnable calibration parameters.

[0071] For the target domain features belonging to the kth working condition subdomain, further statistical calibration of working conditions is performed:

[0072]

[0073] in, and The learnable calibration parameters corresponding to the kth operating condition subdomain.

[0074] The final calibration features are:

[0075]

[0076] in, This represents the conditional asymptotic calibration coefficient.

[0077] The conditional asymptotic calibration coefficients are determined by the number of samples in the target domain, the statistical distance between the source and target domains, the similarity of operating conditions, and the channel quality factor.

[0078]

[0079] in, Let represent the statistical distance between the source domain and the target domain under the k-th working condition subdomain, and let represent the similarity between the working conditions of the source domain and the working conditions of the target domain. This represents the overall acoustic-vibration channel quality factor. These are the weighting coefficients.

[0080] When the number of samples in the target domain is small, the quality of the acoustic and vibration channels is low, or the statistical distance between the source domain and the target domain is large, the calibration intensity is reduced to avoid excessive traction of the model by a few abnormal samples. When the number of samples in the target domain increases and the operating conditions are sufficiently covered, the calibration intensity is increased to gradually bring the characteristics of the target domain closer to the statistical range that the source model is adapted to. Through the above conditional progressive calibration method, this invention does not simply align all samples of the target transformer to the source domain distribution, but performs distribution calibration under the premise of considering the differences in operating conditions, thereby reducing the probability that load changes, oil temperature changes, and channel quality changes are misjudged as fault changes.

[0081] VII. Lightweight Adapter and Target Domain Update

[0082] To further improve the adaptability under conditions of a small number of target transformer samples, this invention sets a lightweight adapter after the feature backbone or before the classification output layer. The lightweight adapter adopts a residual structure and contains only a small number of dimensionality reduction and dimensionality increase parameters, as shown in the following equation:

[0083]

[0084] in, This represents the dimensionality reduction mapping of the lightweight adapter. This represents the up-dimensional mapping of the lightweight adapter. Represents a non-linear activation function. This represents the features after the target domain has been calibrated. This indicates the output characteristics of the lightweight adapter.

[0085] During the target transformer migration phase, only the target domain conditional progressive calibration layer, lightweight adapter, operating condition decoupling correction parameters, and classification correction parameters are updated. The overall parameters of the source model feature backbone are not updated, thereby reducing the number of training parameters in the target domain and lowering the risk of overfitting with few samples.

[0086] The target domain update loss is shown in the following formula:

[0087]

[0088] in, This represents the classification loss for a small number of labeled samples of the target transformer. This represents the statistical alignment loss between the source and target domains; This indicates that the source model output maintains constraints; This indicates a constraint on the consistency of output between the acoustic channel and the vibration channel. This represents a decoupling constraint for operating conditions, used to reduce the direct interference of operating condition factors on the fault category discrimination results; λ1, λ2, λ3, λ4, and λ5 represent the regularization terms for the lightweight adapter parameters; λ1, λ2, λ3, λ4, and λ5 are weighting coefficients.

[0089] If the target domain sample has no label, a pseudo-label can be generated based on the high-confidence prediction result to participate in the classification loss calculation. When the outputs of the voiceprint channel and the vibration channel are significantly conflicting, pseudo-labels are not forced to be generated; instead, the weight of that sample in the target domain update is reduced.

[0090] VIII. Migration Reliability Gating and Diagnostic Output

[0091] To mitigate the risk of negative migration when the target transformer differs significantly from the source transformer, this invention employs a migration confidence gating system. The migration confidence gating system comprehensively considers the maximum probability of the category, the statistical distance between the source and target domains, the acoustic signature channel quality factor, the vibration channel quality factor, and the consistency of the acoustic and vibration channel outputs.

[0092] Let the class probability output by the model be... The statistical distance between the source domain and the target domain is The sound and vibration integrated channel quality factor is The acoustic vibration channel output consistency factor is The migration confidence C is then calculated using the following formula:

[0093]

[0094] in, Indicates the maximum probability of the category. This represents the statistical distance between the source and target domains. This represents the combined quality factor of the acoustic channel and the vibration channel. This represents the consistency factor between the acoustic channel and the vibration channel outputs.

[0095] The acoustic-vibration channel output consistency factor is determined by the following formula:

[0096]

[0097] in Indicates the model's predicted category. Indicates the voiceprint channel for predicted category The probability, Indicates the vibration channel's relationship to the predicted category. The probability of.

[0098] If the migration confidence C is greater than or equal to the first preset threshold If the migration confidence level is lower than the second preset threshold, then the fault category, category number, and probability of each category of the target transformer will be output. If the output results of the acoustic channel and the vibration channel are significantly inconsistent, the system will not output a definite fault category, but will instead output an unknown fault, insufficient sample, sensor abnormality, request for supplementary target transformer sample, or manual verification prompt.

[0099] In one embodiment, the present invention establishes a confidence sample pool. A target sample and its predicted label are added to the confidence sample pool when the target sample meets the following conditions:

[0100]

[0101] in, Indicates the migration confidence threshold. Indicates the category probability threshold. This indicates the quality threshold of the voiceprint channel. This indicates the quality threshold of the vibration channel.

[0102] When a target sample meets the following conditions, it will proceed to either rejection or manual review:

[0103]

[0104] in, This indicates a low confidence threshold. This represents the acoustic vibration channel consistency threshold.

[0105] The samples in the confidence sample pool are not used to update the backbone features of the source model, but rather to periodically re-estimate the target domain statistics and update the target domain conditional asymptotic calibration layer, lightweight adapter, operating condition decoupling correction parameters, and classification correction parameters. In this way, the present invention can gradually accumulate reliable samples during the long-term operation of the target transformer, while avoiding negative migration and the propagation of erroneous false labels.

[0106] IX. System Structure

[0107] like Figure 3 As shown, this invention also provides a cross-transformer acoustic-vibration migration diagnostic system, including a file access and pairing module, a sampling rate estimation module, a true window parameter generation module, a spectrum generation module, a spectrum normalization module, a working condition factor extraction module, a working condition decoupling enhancement module, a source model training module, a backbone parameter preservation module, a target statistical calculation module, a conditional asymptotic calibration module, a lightweight adaptation update module, a migration confidence gating module, a confidence sample pool update module, an unknown fault rejection module, and a diagnostic output module. The file access and pairing module is used to acquire acoustic signature signal files and vibration signal files of the source transformer and the target transformer, and establish an acoustic signature-vibration sample group; the sampling rate estimation module is used to determine the actual sampling rate of each signal file; the true window parameter generation module is used to determine the time-frequency conversion parameters based on the actual sampling rate and the true time window; the spectrum generation module and the spectrum normalization module are used to perform target frequency band preservation, grid resampling, logarithmic compression, and amplitude normalization to generate a dual-layer acoustic-vibration input. The operating condition factor extraction module is used to extract operating condition factors such as load level, oil temperature, ambient noise intensity, channel signal-to-noise ratio, and frequency band energy distribution; the operating condition decoupling enhancement module is used to perform channel quality weighting and frequency band response correction on the dual-layer acoustic and vibration input based on the operating condition embedding vector, acoustic channel quality factor, and vibration channel quality factor; the source model training module is used to train the source transformer acoustic and vibration diagnostic model; and the backbone parameter preservation module is used to preserve the state of the source model's characteristic backbone parameters during the target domain migration stage. The target statistics calculation module and the conditional asymptotic calibration module are used to complete the target domain feature mapping based on the global statistics of the target domain and the statistics of the operating condition subdomain; the lightweight adaptation update module is used to perform minor parameter updates on the target domain conditional asymptotic calibration layer, lightweight adapter, operating condition decoupling correction parameters, and classification correction parameters; the migration confidence gating module is used to calculate the migration confidence of the target samples; the confidence sample pool update module is used to store high-confidence target samples and periodically update the target domain calibration parameters; the unknown fault rejection module is used to output a verification prompt when the migration confidence is insufficient or the acoustic vibration channel diagnostic results are inconsistent; the diagnostic output module is used to output the operating status, fault category, and migration confidence of the target transformer.

[0108] 10. Examples of Electronic Devices and Storage Media

[0109] The present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can perform the above-mentioned steps such as acoustic vibration file pairing, sampling rate estimation, spectrum generation, source model training, target domain progressive calibration, lightweight adaptation update, and diagnostic output.

[0110] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it can implement the trans-transformer acoustic-vibration migration diagnostic method for source backbone maintenance and operating condition decoupling confidence calibration described in this embodiment.

Claims

1. A method and system for cross-transformer acoustic and vibration migration diagnosis based on source backbone maintenance and decoupled confidence calibration of operating conditions, characterized by: S1. Obtain the acoustic and vibration signal files corresponding to the operation of the source transformer and the target transformer, and establish an acoustic-vibration sample group based on the file path, device number, acquisition number, acquisition time, or operating status identifier; S2. Analyze the time information and amplitude information in the voiceprint signal file and vibration signal file respectively. When there is no time information in the file, estimate the actual sampling rate based on the number of valid sampling points and the preset sampling duration. S3. Calculate the short-time Fourier transform window length based on the actual sampling rate and the preset real time window, and calculate the number of overlapping points between adjacent analysis windows based on the preset overlap ratio. S4. Perform mean removal, amplitude scaling adjustment, and short-time Fourier transform on the acoustic signature signal and vibration signal respectively to obtain the initial frequency spectrum of the acoustic signature and the initial frequency spectrum of the vibration. S5. Perform target frequency band preservation, spectrum grid resampling, logarithmic compression and amplitude normalization on the initial spectral surface of the acoustic text and the initial spectral surface of the vibration to obtain the standard spectral surface of the acoustic text and the standard spectral surface of the vibration, and combine them into a dual-layer acoustic-vibration input. S6. Extract the operating condition factor based on the dual-layer acoustic and vibration input and the corresponding operation record, construct the operating condition embedding vector, and perform operating condition decoupling enhancement on the acoustic standard spectrum and the vibration standard spectrum based on the operating condition embedding vector, the acoustic channel quality factor and the vibration channel quality factor to obtain the acoustic and vibration enhancement input. S7. Train the acoustic and vibration fault diagnosis model using the labeled acoustic and vibration enhancement input of the source transformer to obtain the source diagnosis model including the feature backbone, the operating condition decoupling unit, the target domain condition progressive calibration layer, the lightweight adapter and the classification output layer, and save the source domain global statistical benchmark and the source domain operating condition subdomain statistical benchmark. S8. When the source diagnostic model is transferred to the target transformer, the feature backbone parameters are kept in the parameter state of the source training stage. The global statistics of the target domain and the statistics of the target domain operating condition subdomain are calculated using a small number of samples of the target transformer. The target domain features are statistically mapped by the target domain conditional asymptotic calibration layer. S9. Update the target domain only for the target domain conditional progressive calibration layer, lightweight adapter, operating condition decoupling correction parameters and classification correction parameters, and output at least one diagnostic result from the target transformer's operating status, fault category, category number, category probability, migration confidence and review prompt based on migration confidence gating, unknown fault rejection rules and confidence sample pool closed-loop update mechanism.

2. The method as described in claim 1, characterized in that, In steps S1 and S2, the establishment of the acoustic print-vibration sample group includes: reading files with the same device number, acquisition batch, operating status identifier, or time identifier in the acoustic print file directory and vibration file directory, and forming a one-to-one correspondence according to the file sequence number, acquisition timestamp, or natural sorting position; the source transformer sample is used for model training and source domain statistical benchmark establishment, and the target transformer sample is used for target domain migration calibration and diagnostic output; when the signal file contains a time series, the actual sampling rate is determined according to the start and end points of the time series and the number of effective sampling points; when the signal file only contains an amplitude sequence, the actual sampling rate is determined according to the ratio of the number of effective sampling points to the preset acquisition duration, and the estimated result is compared with the candidate sampling rate set to correct the file recording error.

3. The method as described in claim 1, characterized in that, In steps S3 to S5, the short-time Fourier transform window length is determined by the product of the actual sampling rate and the real time window, and the number of overlapping points is determined by the product of the window length and the overlap ratio. When the signal length is less than the window length, zero padding or adaptive window length shortening is performed on the signal; frequency components less than or equal to 20Hz are removed, the target frequency band between 20Hz and 2000Hz is retained, and the initial frequency spectrum of the acoustic pattern and the initial frequency spectrum of the vibration are resampled to standard frequency spectrums with the same number of frequency coordinate points and time coordinate points, respectively; the dual-layer acoustic and vibration input includes an acoustic pattern channel and a vibration channel.

4. The method as described in claim 1, characterized in that, The operating condition factors in step S6 include external operating condition factors obtained directly from the operation records and internal operating condition factors calculated from the acoustic and vibration signals; the external operating condition factors include one or more of the following: load rate, load current, oil temperature, ambient temperature, acquisition period, and cooling device operating status. The internal operating condition factors include one or more of the following: acoustic channel signal-to-noise ratio, vibration channel signal-to-noise ratio, acoustic channel frequency band energy distribution, vibration channel frequency band energy distribution, acoustic-vibration channel correlation coefficient, low-frequency drift intensity, and impact sparsity; the operating condition embedding vector is obtained by splicing or mapping the normalized external and internal operating condition factors, and is used to control the operating condition decoupling intensity and the target domain calibration intensity.

5. The method as described in claim 1, characterized in that, The operating condition decoupling enhancement in step S6 includes channel quality weighting and frequency band response correction. The channel quality weighting determines the contribution weights of the acoustic and vibration channels in the fusion diagnosis based on the acoustic channel quality factor and the vibration channel quality factor. The frequency band response correction generates correction weights for each frequency band based on the operating condition embedding vector and the fault-sensitive frequency band prior, and applies the correction weights to the acoustic standard spectrum and the vibration standard spectrum. The fault-sensitive frequency band prior includes at least two frequency bands from the low-frequency structural vibration frequency band, the mid-frequency mechanical response frequency band, and the high-frequency impact or discharge response frequency band, which are used to enhance the frequency-time regions related to winding loosening, core abnormalities, short-circuit impacts, DC bias, or insulation abnormalities, and suppress the frequency-time regions corresponding to environmental noise or operating condition disturbances.

6. The method as described in claim 1, characterized in that, In step S7, the characteristic backbone of the acoustic and vibration fault diagnosis model includes a spectral texture extraction unit, an acoustic and vibration channel interaction unit, a frequency-time region focusing unit, and a working condition decoupling unit. The classification output layer is used to output the operating status or fault category of the source transformer; After training, the mean of global features in the source domain, the standard deviation of global features in the source domain, the mean of features in the subdomain of operating conditions in the source domain, and the standard deviation of features in the subdomain of operating conditions in the source domain are saved as statistical benchmarks when the target transformer is migrated. When migrating to the target transformer, the core feature parameters maintain the parameter state of the source training phase and are not updated in reverse as a whole during the target domain few-sample update process.

7. The method as described in claim 1, characterized in that, The target domain conditional asymptotic calibration layer in step S8 includes a global statistical calibration branch and a sub-domain statistical calibration branch. The global statistical calibration branch calculates the global mean and global standard deviation of the target domain based on all samples in the target domain. The sub-domain statistical calibration branch divides the target domain samples into at least two sub-domains based on load level, oil temperature, ambient noise intensity, channel quality factor, or frequency band energy distribution, and calculates the mean and standard deviation of each sub-domain. The target domain conditional asymptotic calibration layer performs feature distribution calibration based on the source domain statistical benchmark, the global statistics of the target domain, and the sub-domain statistics of the target domain, and controls the degree to which the target domain features approach the source domain statistical benchmark through asymptotic coefficients. The asymptotic coefficients are jointly determined by the number of target domain samples, the statistical distance between the source and target domains, the operating condition similarity, and the channel quality factor.

8. The method as described in claim 1, characterized in that, In step S9, the lightweight adapter is set after the feature backbone or before the classification output layer using a residual structure. The number of parameters of the lightweight adapter is smaller than that of the feature backbone. It is updated together with the target domain conditional asymptotic calibration layer, operating condition decoupling correction parameters and classification correction parameters during the target transformer migration stage. The target domain update loss includes the classification loss of a small number of labeled samples of the target transformer, the statistical alignment loss between the source domain and the target domain, the source model output preservation constraint, the acoustic and vibration channel consistency constraint and the adapter parameter regularization term. If the target domain samples are unlabeled, pseudo-labels are generated based on the high-confidence prediction results and used to calculate the classification loss.

9. The method as described in claim 1, characterized in that, The migration confidence gating, unknown fault rejection rule, and confidence sample pool closed-loop update mechanism in step S9 include: calculating migration confidence based on the maximum probability of the category, the statistical distance between the source and target domains, the acoustic channel quality factor, the vibration channel quality factor, and the consistency of the acoustic and vibration channel outputs; when the migration confidence is greater than or equal to the first preset threshold, outputting the fault category, category number, and category probability of the target transformer; when the maximum probability of the target sample's category, the acoustic channel quality factor, the vibration channel quality factor, and the migration confidence all meet preset conditions, adding the target sample and its predicted label to the confidence sample pool; when the number of samples in the confidence sample pool reaches a preset number or a preset time interval, re-estimating the target domain statistics based on the confidence sample pool, and updating the target domain conditional progressive calibration layer, lightweight adapter, operating condition decoupling correction parameters, and classification correction parameters; when the migration confidence is lower than the second preset threshold, or when the output results of the acoustic channel and the vibration channel are inconsistent, not adding the target sample to the confidence sample pool, and outputting prompts such as unknown fault, insufficient samples, sensor anomaly, supplementing target transformer samples, or manual review.

10. A trans-transformer acoustic-vibration migration diagnostic system, electronic device, and computer-readable storage medium for source backbone maintenance and operating condition decoupling confidence calibration, characterized in that, The system includes a file access and pairing module, a sampling rate estimation module, a true window parameter generation module, a spectrum generation module, a spectrum standardization module, a working condition factor extraction module, a working condition decoupling enhancement module, a source model training module, a backbone parameter preservation module, a target statistical calculation module, a conditional asymptotic calibration module, a lightweight adaptation update module, a migration confidence gating module, a confidence sample pool update module, an unknown fault rejection module, and a diagnostic output module. The file access and pairing module is used to acquire the acoustic signature and vibration signal files of the source and target transformers and establish an acoustic signature-vibration sample group. The sampling rate estimation module is used to determine the actual sampling rate of each signal file. The true window parameter generation module is used to calculate the actual sampling rate of each signal file based on the actual sampling rate of the source and target transformers. The short-time Fourier transform parameters are determined by the actual sampling rate and the real time window; the spectrum generation module and the spectrum normalization module are used to generate the acoustic signature standard spectrum and the vibration standard spectrum, forming a dual-layer acoustic-vibration input; the operating condition factor extraction module is used to extract one or more operating condition factors from the load level, oil temperature, ambient noise intensity, channel signal-to-noise ratio, and frequency band energy distribution; the operating condition decoupling enhancement module is used to perform channel quality weighting and frequency band response correction on the dual-layer acoustic-vibration input according to the operating condition embedding vector, acoustic signature channel quality factor, and vibration channel quality factor; the source model training module is used to train the source transformer acoustic-vibration fault diagnosis model; the backbone parameter preservation module is used to preserve the state of the source model's characteristic backbone parameters during the target domain migration stage; The target statistical calculation module and the conditional asymptotic calibration module are used to complete the target domain feature mapping based on the global statistics of the target domain and the statistics of the operating condition subdomain. The lightweight adaptation update module is used to update the target domain conditional progressive calibration layer, lightweight adapter, operating condition decoupling correction parameters, and classification correction parameters; the migration confidence gating module is used to calculate the migration confidence of the target sample. The confidence sample pool update module is used to store high-confidence target samples and periodically update the target domain calibration parameters; the unknown fault rejection module is used to output a verification prompt when the migration confidence is insufficient or the acoustic vibration channel diagnostic results are inconsistent. The diagnostic output module is used to output at least one diagnostic result from the following: the target transformer's operating status, fault category, category number, category probability, migration confidence level, and review prompt; the electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the method according to any one of claims 1 to 9; the computer-readable storage medium stores a computer program, which, when executed by the processor, implements the method according to any one of claims 1 to 9.

Images