A Transformer Anomaly Diagnosis Method and System Integrating Voiceprint and Electrical Parameters

Abstract

This invention relates to a method and system for transformer anomaly diagnosis that integrates acoustic signature and electrical parameters. The method includes: acquiring acoustic signature time-series data and electrical parameter time-series data of the target transformer; extracting features from the preprocessed acoustic signature time-series data and electrical parameter time-series data to obtain acoustic signature features and electrical parameter features, and resampling based on a time offset parameter; fusing the resampled acoustic signature features and electrical parameter features based on a cross-attention mechanism to generate fused features; inputting the fused features into anomaly category classification branch and operating condition regression branch to generate logarithmic values ​​and operating condition regression vectors for each anomaly category; converting the logarithmic values ​​of each anomaly category into a classification probability sequence; inputting the operating condition regression vector into a causal consistency calibrator to obtain a causal consistency score; recalibrating the causal consistency score and classification probability sequence using a monotonic recalibration function to obtain the recalibrated classification probability, and obtaining the final diagnostic result.

Description

Technical Field

[0001] This invention relates to the field of transformer fault detection technology, and mainly to a method and system for diagnosing transformer anomalies by integrating acoustic signatures and electrical parameters. Background Technology

[0002] As a core device for energy conversion and transmission in a power system, transformers are susceptible to various operational anomalies during long-term service, including load fluctuations, grid disturbances, environmental changes, and equipment aging. These anomalies can lead to overexcitation, mechanical abnormalities, insulation degradation, and abnormal electromagnetic noise. Failure to detect and address these anomalies in a timely manner can result in escalating faults, equipment damage, and even large-scale power outages. Therefore, achieving accurate and real-time online detection of various transformer anomalies is of great significance for improving grid operation and maintenance efficiency, reducing fault risks, and ensuring the reliable operation of the power system.

[0003] Chinese invention patent application CN118197359A discloses a transformer fault identification method and system based on the fusion of acoustic signature and electromagnetic features. The technical solution includes collecting acoustic signature time-series data and electromagnetic information during transformer operation; converting the acoustic signature time-series data into an acoustic signature spectrum using wavelet transform; preprocessing the electromagnetic information and acoustic signature spectrum, then extracting electromagnetic features and acoustic signature features; merging the acoustic signature feature vector composed of acoustic signature features and the electromagnetic feature vector composed of electromagnetic features to form a feature set; inputting the feature set into a feature fusion module based on adaptive convolutional weight learning for fusion to obtain fused features; inputting the fused features into a deep neural network, and the classification layer of the deep neural network outputs diagnostic results, thereby identifying the fault type and location; however, the above technical solution does not consider the temporal relationship between acoustic signature data and electromagnetic information. The above-mentioned technical solutions suffer from several drawbacks. First, they fail to address differences in sampling frequency and device clock drift by designing an adaptive alignment scheme. Second, conventional preprocessing alone is insufficient to compensate for time delays, impacting diagnostic accuracy. Third, they do not consider the different mechanisms of transformer anomalies, directly merging acoustic signature vectors and electromagnetic signature vectors and indiscriminately fusing them through adaptive convolution weights. This results in the fused features being unable to distinguish between anomaly types with similar features, leading to a high risk of misjudgment. Fourth, the above-mentioned technical solutions lack learnable time delay modeling and causal constraints, which may introduce future information or time sequence mismatches during the fusion stage, affecting the reliability of online diagnosis. Finally, they only perform a single classification task, outputting only the fault type and location without any regression output of operating condition parameters. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a transformer anomaly diagnosis method and system that integrates acoustic signature and electrical parameters.

[0005] The technical solution of the present invention is as follows: On the one hand, this invention proposes a transformer anomaly diagnosis method that integrates acoustic signature and electrical parameters, the method comprising: Acquire acoustic signature timing data and electrical parameter timing data of the target transformer, and perform preprocessing. Features are extracted from the preprocessed audioprint time-series data and electrical parameter time-series data to obtain audioprint features and electrical parameter features. Feature alignment is achieved by resampling based on the time offset parameter. The resampled audioprint features and electrical parameter features are fused based on the cross-attention mechanism to generate fused features. The fused features are input into the anomaly category classification branch and the operating condition regression branch to generate the logarithmic values ​​of each anomaly category and the operating condition regression vector; the logarithmic values ​​of each anomaly category are converted into a classification probability sequence; the operating condition regression vector is input into the causal consistency calibrator to obtain the causal consistency score. The causal consistency score and classification probability sequence are recalibrated using a monotonic recalibration function to obtain the recalibrated classification probability; the final diagnostic result is output based on the recalibrated classification probability.

[0006] Preferably, resampling is performed based on the time offset parameter, and the specific steps are as follows: Map voiceprint features and electrical parameter features to a unified time axis; The mapped voiceprint features and electrical parameter features are resampled using a time offset parameter to obtain preliminary resampled voiceprint features and electrical parameter features. Linear interpolation is used to calculate the corresponding feature values ​​at non-integer sampling points, and mirror filling is used to fill the corresponding feature values ​​at the boundaries of voiceprint features and electrical parameter features. A causal mask matrix is ​​applied to the initial resampled voiceprint and electrical parameter features to obtain the resampled voiceprint and electrical parameter features.

[0007] Preferably, the maximum time delay limit is calculated based on the distance between the acoustic timing data acquisition device and the transformer, the ambient sound speed, and the preset safety margin of the device clock drift; the time offset parameter is obtained based on the maximum time delay limit.

[0008] Preferably, the resampled voiceprint features and electrical parameter features are fused based on a cross-attention mechanism, and the specific steps are as follows: Construct a set of mechanism labeling vectors, including heavy overload mechanism labeling vectors, overexcitation mechanism labeling vectors, and mechanical anomaly mechanism labeling vectors; The resampled voiceprint features and the resampled electrical parameter features are concatenated to obtain joint features; the joint features are projected based on the preset key matrix, value matrix, and weight matrix to obtain key vectors and value vectors. The cross-attention fusion weights are calculated using the mechanism label vector set as the query vector for cross-attention. The value vector is weighted based on the cross-attention fusion weights to obtain the corresponding fusion features; All fusion features are concatenated to generate a fusion feature.

[0009] Preferably, the fused features are subjected to linear and nonlinear transformations using the anomaly category classification branch to obtain the logarithmic values ​​of each anomaly category, wherein the anomaly categories include heavy overload, overexcitation, and mechanical anomalies. Linear regression of the fused features is performed on the working condition regression branch to generate the working condition regression vector.

[0010] Preferably, the logarithmic values ​​of each anomaly category are converted into a classification probability sequence, wherein a minimum residence time constraint is applied to the heavily overloaded state in the anomaly category. The specific steps are as follows: If the target transformer is not in a heavy overload state, the target transformer will be switched to a heavy overload state only when the duration of the detected initial probability of heavy overload exceeding the preset trigger threshold reaches the minimum dwell time. If the target transformer is in a state of heavy overload, the current state will be maintained until the duration reaches the minimum dwell time; when the duration reaches the minimum dwell time and the initial probability of heavy overload is less than the preset release threshold, the heavy overload state will be exited. After fulfilling the above duration constraints, the logarithmic values ​​of each anomaly category are converted into an initial classification probability sequence using the Softmax function. All initial classification probability sequences are then normalized to obtain the normalized initial classification probability sequence, i.e., the classification probability sequence, which includes the classification probability sequences of heavy overload, overexcitation, and mechanical anomaly categories.

[0011] Preferably, the shortest residence time is adaptively adjusted based on ambient temperature, historical load rate moving average, and transformer capacity.

[0012] Preferably, the classification probability sequence and the working condition regression vector are input into the causal consistency calibrator. The specific steps are as follows: The operating condition regression vector includes load factor, magnetic flux density, and winding temperature rise slope. The current load rate and winding temperature rise slope are compared with the corresponding preset thresholds to obtain the causal consistency score of the heavy overload state. The current magnetic flux density is compared with the corresponding preset threshold to obtain the causal consistency score of the overexcitation state. The statistical independence of the high-frequency band energy ratio, narrow-band spectral peak stability, and load rate obtained from the voiceprint features is compared with the corresponding preset thresholds to obtain the causal consistency score of the mechanical state.

[0013] Preferably, the monotonically recalibrated function satisfies the monotonically non-decreasing constraint.

[0014] On the other hand, the present invention also provides a transformer anomaly diagnosis system that integrates voiceprint and electrical parameters, the system comprising: The data acquisition module is used to collect the acoustic signature time series data and electrical parameter time series data of the target transformer and perform preprocessing. The feature processing module is used to extract features from the preprocessed audioprint time-series data and electrical parameter time-series data to obtain audioprint features and electrical parameter features, and to achieve feature alignment by resampling based on the time offset parameter; the resampled audioprint features and electrical parameter features are fused based on the cross-attention mechanism to generate fused features; The scoring module is used to input the fused features into the anomaly category classification branch and the working condition regression branch to generate the logarithmic value of each anomaly category and the working condition regression vector; to convert the logarithmic value of each anomaly category into a classification probability sequence; and to input the working condition regression vector into the causal consistency calibrator to obtain the causal consistency score. The diagnostic module is used to recalibrate the causal consistency score and classification probability sequence using a monotonic recalibration function to obtain the recalibrated classification probability; and outputs the final diagnostic result based on the recalibrated classification probability.

[0015] The present invention has the following beneficial effects: 1. This invention integrates voiceprint time-series data and electrical parameter time-series data to overcome the limitations of a single data dimension; it constructs three types of mechanism labeling vectors: heavy overload, overexcitation, and mechanical anomaly, and achieves fusion through a cross-attention mechanism, so that feature extraction fits the physical essence of the fault occurrence; it improves the accuracy of anomaly category identification, especially the distinguishing ability between heavy overload and overexcitation, because the electrical parameter features of the two are similar but the voiceprint features are significantly different, and mechanism-guided fusion can accurately capture the differences; 2. This invention reduces the false positive rate through dynamic constraints and causal consistency calibration. The dynamic constraints introduce a minimum residence time to constrain the heavy overload state, and the minimum residence time can be dynamically and adaptively adjusted based on ambient temperature, historical load rate moving average, and transformer capacity. The causal consistency calibration verifies whether the classification result conforms to physical logic by calculating a score by weighting the operating condition regression vector and classification probability. This improves the stability of the diagnostic results and enhances robustness in complex environments. 3. This invention, through a multi-task branch design, takes into account both classification and working condition perception, and realizes anomaly category classification and working condition regression, forming a dual-output mode that combines anomaly identification and working condition verification. This solves the problem that a single classification task cannot provide feedback on the working condition background, provides quantitative working condition support for the confirmation of abnormal states, and improves the accuracy of abnormal state switching. Attached Figure Description

[0016] Figure 1 This is a detailed flowchart of an embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0019] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0020] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0021] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0022] Example 1: See Figure 1 This invention provides a method for diagnosing transformer anomalies by integrating acoustic signatures and electrical parameters, the method comprising: S1. Collect the acoustic signature timing data and electrical parameter timing data of the target transformer, and perform preprocessing. S11. Set up a microphone or microphone array near the target transformer, collect air sound pressure signals according to the set sampling rate and bit depth, and generate time-stamped voiceprint timing data. The electrical parameter time series data corresponding to the time period of the voiceprint time series data are obtained from the monitoring system. The electrical parameter time series data includes at least a continuous time series of current, load rate, magnetic flux density and winding temperature. S12. The preprocessing specifically includes: Noise suppression, bandpass filtering, silence removal, amplitude normalization and timestamp calibration are performed on the voiceprint time series data to align the voiceprint time series data with a unified time axis; Outlier removal, missing value completion, scale normalization, and timestamp calibration are performed on the electrical parameter time series data to align the electrical parameter time series data with a unified time axis. S2. Use a feature extraction encoder to extract features from the preprocessed acoustic print time series data and electrical parameter time series data to obtain acoustic print features and electrical parameter features. Based on the preprocessed voiceprint time-series data, voiceprint features are extracted in the voiceprint encoder. The voiceprint features are vector sequences arranged along a unified time axis. Based on the preprocessed electrical parameter timing data, electrical parameter features are extracted in the electrical parameter encoder. The electrical parameter features are vector sequences arranged along a uniform time axis. S3. Feature alignment is achieved by resampling based on time offset parameters, that is, mapping voiceprint features and electrical parameter features to a unified time axis. S31. Resample the mapped voiceprint features and electrical parameter features using the time offset parameter to obtain the preliminary resampled voiceprint features and electrical parameter features. Linear interpolation is used to calculate the corresponding feature values ​​at non-integer sampling points, and mirror filling is used to fill the corresponding feature values ​​at the boundaries of voiceprint features and electrical parameter features. A causal mask matrix is ​​applied to the initial resampled voiceprint features and electrical parameter features to obtain the resampled voiceprint features and electrical parameter features. The application of the causal mask matrix only allows historical time windows to participate in subsequent attention calculation and fusion, and prohibits future time windows from participating in the calculation, thereby ensuring real-time performance and causal consistency. S32. The maximum time delay limit is calculated based on the distance between the acoustic signature timing data acquisition device and the transformer, the ambient sound velocity, and the preset safety margin for device clock drift; the time offset parameter is obtained based on the maximum time delay limit, and the specific calculation method is as follows: ; ; In the formula, Indicates the maximum time delay limit; Indicates the time offset parameter; This indicates the distance between the voiceprint timing data acquisition device and the transformer, wherein the voiceprint timing data acquisition device is a microphone or a microphone array; Indicates the speed of ambient sound; This indicates the preset safety margin for device clock drift; S4. Based on the cross-attention mechanism, the resampled voiceprint features and electrical parameter features are fused to generate fused features; A set of mechanism labeling vectors is constructed, including heavy overload mechanism labeling vectors, overexcitation mechanism labeling vectors, and mechanical anomaly mechanism labeling vectors; the mechanism labeling vectors are learnable query vectors, whose feature dimensions are consistent with the feature dimensions of the resampled acoustic features and electrical parameter features, and are aligned with a unified time axis; The resampled voiceprint features and resampled electrical parameter features are concatenated to obtain joint features. The joint features are then projected onto a pre-defined key matrix, value matrix, and weight matrix to obtain key vectors and value vectors. The calculation method is as follows: ; ; ; In the formula, Indicates joint features; This represents a predefined key matrix; This represents a preset value matrix; Represents the key vector; Represents a value vector; This indicates the characteristics of the electrical parameters after resampling; This represents the voiceprint features after resampling; Indicates a connector; Using the set of mechanism-labeled vectors as the query vector for cross-attention, the cross-attention fusion weight is calculated as follows: ; In the formula, Indicates the first The cross-attention fusion weights of the mechanism label vectors, where When, it indicates a heavy overload. At that time, it indicates over-excitation. When this occurs, it indicates a mechanical malfunction; This represents the softmax function; Indicates the first A mechanism marker vector; Represents the preset projection vector; Indicates the transpose operation; The value vector is weighted based on cross-attention fusion weights to obtain the corresponding fusion features. The calculation method is as follows: ; In the formula, Indicates the first The fusion features of individual mechanism label vectors; All fused features are concatenated to generate a fused feature, calculated as follows: ; In the formula, Indicates fusion characteristics; Indicates the join function; S5. Input the fused features into the anomaly category classification branch and the working condition regression branch to generate the logarithmic value and working condition regression vector for each anomaly category. The multi-task branch includes an anomaly category classification branch and a working condition regression branch, wherein: The fused features are subjected to linear and nonlinear transformations using anomaly category classification branches to obtain the logarithmic values ​​of each anomaly category, wherein the anomaly categories include heavy overload, overexcitation, and mechanical anomalies. Linear regression is performed on the fused features using the working condition regression branch to generate the working condition regression vector; S6. Convert the logarithmic values ​​of each anomaly category into a classification probability sequence, wherein the anomaly categories include heavy overload, overexcitation, and mechanical anomaly; The engineering definition of heavy overload is usually not "instantaneously exceeding the rated value", but is strongly related to the duration: the increase in load - the increase in copper loss - the increase in winding temperature requires time to accumulate, and has obvious thermal inertia and hysteresis; therefore, the heavy overload state in real-world scenarios often manifests as a continuous residence for a period of time, rather than a pulse of one or two sampling points. The overexcitation may occur due to short-term fluctuations in system voltage / frequency, and may be an event that can occur in a short time. The aforementioned mechanical anomalies (loosening, resonance, friction, etc.) may acoustically manifest as intermittent, impact-related, or condition-related short-term abrupt changes. S61. Apply a minimum residence time constraint to the heavy overload state in the anomaly category; Setting a minimum dwell time constraint for overexcitation or mechanical anomalies may lead to missed alarms / alarm delays (especially detrimental to anomalies requiring rapid response), and the output classification probability will fluctuate under noise, transient disturbances, and alignment errors. If the heavy overload state is not constrained, the probability of heavy overload may suddenly spike at a certain moment, triggering an alarm, and then return to normal the next moment, forming a false alarm of "probability flashing". The effect of the minimum dwell time constraint is that only when the heavy overload is highly probable and lasts for a certain duration is it allowed to enter / maintain the heavy overload state; if it is not reached, the handover is suppressed, thereby significantly reducing false alarms. The specific steps are as follows: If the target transformer is not in a heavy overload state, the target transformer will be switched to a heavy overload state only when the duration of the detected initial probability of heavy overload exceeding the preset trigger threshold reaches the minimum dwell time. If the target transformer is in a state of heavy overload, the current state will be maintained until the duration reaches the minimum dwell time; when the duration reaches the minimum dwell time and the initial probability of heavy overload is less than the preset release threshold, the heavy overload state will be exited. After completing the above duration constraints, the logarithmic values ​​of each anomaly category are converted into an initial classification probability sequence through the Softmax function. All initial classification probability sequences are normalized to obtain the normalized initial classification probability sequence, i.e., the classification probability sequence, which includes the classification probability sequences of heavy overload, overexcitation and mechanical anomaly categories. The above steps are used to further convert the logarithmic values ​​output by the multi-task classification branch at each time step into a heavy overload state that can be used in engineering, and to impose a minimum dwell time constraint on the entry and exit of heavy overload, thereby suppressing probabilistic jitter, alarm flashing and false triggering caused by short-term impact load, noise disturbance or alignment error. The reason for "switching to heavy overload state" is that the heavy overload state is not a numerical judgment at a single point in time, but an operating state that needs to be maintained for a certain period of time: only when the initial probability of heavy overload continuously exceeds the trigger threshold and accumulates to the minimum dwell time is the transformer considered to have truly entered the heavy overload state and the corresponding status / alarm is output; similarly, exiting also usually requires meeting conditions such as dwell time and being below the release threshold to avoid fluctuating back and forth around the threshold (hysteresis); this "state switching" transforms the instantaneous probability on the continuous time series into a "entry-hold-exit" state decision, which facilitates the subsequent output of a stable classification probability sequence, graded alarms and rejection results, and conforms to the physical mechanism of the time-scale physical aging and temperature rise accumulation of insulation caused by heavy overload; S62. The minimum residence time is adaptively adjusted based on ambient temperature, historical load rate moving average, and transformer capacity, and is calculated as follows: ; In the formula, Indicates the shortest length of stay; This represents an adaptive adjustment function; Indicates ambient temperature; This represents the historical moving average of the load factor. Indicates transformer capacity; When the ambient temperature rises or the historical load rate increases, the minimum residence time is increased to enhance the continuous constraint of heavy overload. S7. Input the operating condition regression vector into the causal consistency calibrator to obtain the causal consistency score; the operating condition regression vector includes load factor, magnetic flux density and winding temperature rise slope; S71. Compare the current load factor and winding temperature rise slope with the corresponding preset thresholds to obtain the causal consistency score of the heavy overload state, specifically: like and If the condition is overloaded, the score is set to 1; otherwise, it is 0. Indicates the current load factor. This indicates the preset load factor threshold. Indicates the slope of the winding temperature rise. This indicates the preset winding temperature rise slope threshold; S72. Compare the current magnetic flux density with the corresponding preset threshold to obtain the causal consistency score of the overexcitation state, specifically: like If the overexcitation state score is 1, then the score is 0; otherwise, it is 0. Indicates the current magnetic flux density. This indicates the preset magnetic flux density threshold; S73. The statistical independence of the high-frequency band energy ratio, narrow-band spectral peak stability, and load rate obtained from the voiceprint features is compared with the corresponding preset thresholds to obtain the causal consistency score of the mechanical state, specifically: like , and If the mechanical condition score is 1, then the mechanical condition score is set to 1; otherwise, it is 0. Indicates the high-frequency band energy ratio. This indicates the preset high-frequency band energy ratio threshold. Indicates the stability of narrowband spectral peaks. This indicates the preset narrowband spectral peak stability threshold. Indicates the statistical independence of load factors. Indicates the statistical independence threshold for the preset load factor; S8. Use the monotonic recalibration function to recalibrate the causal consistency score and classification probability sequence to obtain the recalibrated classification probability. The monotonic recalibration function satisfies the monotonic non-decreasing constraint, which ensures that when the causal consistency score increases, the recalibrated classification probability of the corresponding anomaly category does not decrease, and when the causal consistency score decreases, the recalibrated classification probability of the corresponding anomaly category does not increase. S9. Output the final diagnostic result based on the recalibrated classification probability; the diagnostic result includes timestamp, recalibrated classification probability, alarm signal, alarm level, rejection result, etc. The alarm signal specifically sets an alarm threshold for each anomaly category; if the recalibrated classification probability of the current time step exceeds its corresponding threshold, an alarm signal for the corresponding anomaly category is generated. The alarm levels are specifically divided into multiple levels (such as early warning, general alarm, and severe alarm) based on the duration and probability intensity of the alarm signal. The rejection result is specifically set as a global rejection threshold; if, within a given time window, the recalibrated classification probability of all abnormal categories is lower than their respective alarm thresholds, and the maximum probability value is also lower than the global rejection threshold, then the rejection result is output, indicating that the current state is no different from a diagnosable abnormality. In a specific embodiment, the classification probability sequence [0.83, 0.11, 0.06] is obtained through steps S1-S6; Perform state score calculation, where: Heavy overload condition rating: At this time and If so, the heavy overload condition score is 1; Overexcitation state score: At this time If so, the overexcitation state score is 0; Mechanical abnormal condition scoring: If The mechanical condition score is then 0. Obtain the causal consistency score of the heavy overload state. ; Causal consistency score of over-excitation state obtained ; Obtain the causal consistency score of the mechanical abnormal state. ; Using a monotonically recalibrated function The causal consistency score was recalibrated, where: Heavy overload ; Overexcitation ; Mechanical abnormality ; The output diagnostic results are as follows: timestamp, classification probability: heavy overload 0.90, overexcitation 0.15, mechanical abnormality 0.13, alarm signal: heavy overload alarm (probability 0.90 > alarm threshold 0.7), alarm level: severe alarm (duration 120s ≥ 60s, probability ≥ 0.8).

[0023] Example 2: This embodiment provides a transformer anomaly diagnosis system that integrates acoustic signature and electrical parameters. The system includes: The data acquisition module is used to collect the acoustic signature time series data and electrical parameter time series data of the target transformer and perform preprocessing. The feature processing module is used to extract features from the preprocessed audioprint time-series data and electrical parameter time-series data to obtain audioprint features and electrical parameter features, and to achieve feature alignment by resampling based on the time offset parameter; the resampled audioprint features and electrical parameter features are fused based on the cross-attention mechanism to generate fused features; The scoring module is used to input the fused features into the anomaly category classification branch and the working condition regression branch to generate the logarithmic value of each anomaly category and the working condition regression vector; to convert the logarithmic value of each anomaly category into a classification probability sequence; and to input the working condition regression vector into the causal consistency calibrator to obtain the causal consistency score. The diagnostic module is used to recalibrate the causal consistency score and classification probability sequence using a monotonic recalibration function to obtain the recalibrated classification probability; and outputs the final diagnostic result based on the recalibrated classification probability.

[0024] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0025] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0026] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0027] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0028] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for diagnosing transformer anomalies by integrating voiceprint and electrical parameters, characterized in that, The method includes: Acquire acoustic signature timing data and electrical parameter timing data of the target transformer, and perform preprocessing. Features are extracted from the preprocessed audioprint time-series data and electrical parameter time-series data to obtain audioprint features and electrical parameter features. Feature alignment is achieved by resampling based on the time offset parameter. The resampled audioprint features and electrical parameter features are fused based on the cross-attention mechanism to generate fused features. The fused features are input into the anomaly category classification branch and the operating condition regression branch to generate the logarithmic values ​​of each anomaly category and the operating condition regression vector; the logarithmic values ​​of each anomaly category are converted into a classification probability sequence; the operating condition regression vector is input into the causal consistency calibrator to obtain the causal consistency score. The causal consistency score and classification probability sequence are recalibrated using a monotonic recalibration function to obtain the recalibrated classification probability; the final diagnostic result is output based on the recalibrated classification probability.

2. The transformer anomaly diagnosis method integrating acoustic signature and electrical parameters according to claim 1, characterized in that, Resampling based on the time offset parameter involves the following steps: Map voiceprint features and electrical parameter features to a unified time axis; The mapped voiceprint features and electrical parameter features are resampled using a time offset parameter to obtain preliminary resampled voiceprint features and electrical parameter features. Linear interpolation is used to calculate the corresponding feature values ​​at non-integer sampling points, and mirror filling is used to fill the corresponding feature values ​​at the boundaries of voiceprint features and electrical parameter features. A causal mask matrix is ​​applied to the initial resampled voiceprint and electrical parameter features to obtain the resampled voiceprint and electrical parameter features.

3. The transformer anomaly diagnosis method integrating acoustic signature and electrical parameters according to claim 2, characterized in that, The maximum time delay limit is calculated based on the distance between the acoustic timing data acquisition device and the transformer, the ambient sound speed, and the preset safety margin for device clock drift; the time offset parameter is obtained based on the maximum time delay limit.

4. The transformer anomaly diagnosis method integrating acoustic signature and electrical parameters according to claim 1, characterized in that, The resampled speaker features and electrical parameter features are fused based on a cross-attention mechanism. The specific steps are as follows: Construct a set of mechanism labeling vectors, including heavy overload mechanism labeling vectors, overexcitation mechanism labeling vectors, and mechanical anomaly mechanism labeling vectors; The resampled voiceprint features and the resampled electrical parameter features are concatenated to obtain joint features; The joint features are projected based on the preset key matrix, value matrix, and weight matrix to obtain the key vector and value vector; The cross-attention fusion weights are calculated using the mechanism label vector set as the query vector for cross-attention. The value vector is weighted based on the cross-attention fusion weights to obtain the corresponding fusion features; All fusion features are concatenated to generate a fusion feature.

5. The transformer anomaly diagnosis method integrating acoustic signature and electrical parameters according to claim 1, characterized in that, The fused features are subjected to linear and nonlinear transformations using anomaly category classification branches to obtain the logarithmic values ​​of each anomaly category, wherein the anomaly categories include heavy overload, overexcitation, and mechanical anomalies. Linear regression of the fused features is performed on the working condition regression branch to generate the working condition regression vector.

6. The transformer anomaly diagnosis method integrating acoustic signature and electrical parameters according to claim 5, characterized in that, The logarithmic values ​​of each anomaly category are transformed into a classification probability sequence, wherein a minimum residence time constraint is applied to the heavily overloaded state in the anomaly category. The specific steps are as follows: If the target transformer is not in a heavy overload state, the target transformer will be switched to a heavy overload state only when the duration of the detected initial probability of heavy overload exceeding the preset trigger threshold reaches the minimum dwell time. If the target transformer is in a state of heavy overload, the current state will be maintained until the duration reaches the minimum dwell time; when the duration reaches the minimum dwell time and the initial probability of heavy overload is less than the preset release threshold, the heavy overload state will be exited. After fulfilling the above duration constraints, the logarithmic values ​​of each anomaly category are converted into an initial classification probability sequence using the Softmax function. All initial classification probability sequences are then normalized to obtain the normalized initial classification probability sequence, i.e., the classification probability sequence, which includes the classification probability sequences of heavy overload, overexcitation, and mechanical anomaly categories.

7. The transformer anomaly diagnosis method integrating acoustic signature and electrical parameters according to claim 6, characterized in that, The minimum residence time is adaptively adjusted based on ambient temperature, historical load rate moving average, and transformer capacity.

8. The transformer anomaly diagnosis method integrating acoustic signature and electrical parameters according to claim 1, characterized in that, The classification probability sequence and the working condition regression vector are input into the causal consistency calibrator. The specific steps are as follows: The operating condition regression vector includes load factor, magnetic flux density, and winding temperature rise slope. The current load rate and winding temperature rise slope are compared with the corresponding preset thresholds to obtain the causal consistency score of the heavy overload state. The current magnetic flux density is compared with the corresponding preset threshold to obtain the causal consistency score of the overexcitation state. The statistical independence of the high-frequency band energy ratio, narrow-band spectral peak stability, and load rate obtained from the voiceprint features is compared with the corresponding preset thresholds to obtain the causal consistency score of the mechanical state.

9. The transformer anomaly diagnosis method integrating acoustic signature and electrical parameters according to claim 1, characterized in that, The monotonically recalibrated function satisfies the monotonically non-decreasing constraint.

10. A transformer anomaly diagnosis system integrating voiceprint and electrical parameters, characterized in that, The system includes: The data acquisition module is used to collect the acoustic signature time series data and electrical parameter time series data of the target transformer and perform preprocessing. The feature processing module is used to extract features from the preprocessed audioprint time-series data and electrical parameter time-series data to obtain audioprint features and electrical parameter features, and to achieve feature alignment by resampling based on the time offset parameter; the resampled audioprint features and electrical parameter features are fused based on the cross-attention mechanism to generate fused features; The scoring module is used to input the fused features into the anomaly category classification branch and the working condition regression branch to generate the logarithmic value of each anomaly category and the working condition regression vector; to convert the logarithmic value of each anomaly category into a classification probability sequence; and to input the working condition regression vector into the causal consistency calibrator to obtain the causal consistency score. The diagnostic module is used to recalibrate the causal consistency score and classification probability sequence using a monotonic recalibration function to obtain the recalibrated classification probability; and outputs the final diagnostic result based on the recalibrated classification probability.

Images