A method and system for diagnosing faults of a high-frequency transformer

Abstract

The application relates to the technical field of fault diagnosis, and provides a high-frequency transformer fault diagnosis method and system, which comprises the following steps: collecting a target signal with a time stamp and extracting corresponding features, simultaneously relying on a transformer structure, material parameters and physical rules to build a digital twin model, simulating insulation and structure degradation equivalent working conditions, solving multi-physical field data and generating multi-physical field mechanism samples; then, the mechanism samples and field measured data are fused through a generative adversarial network to expand the fault sample data set and solve the sample scarcity problem; in the running stage, the digital twin model is updated in real time through parameter online inversion, a feature dynamic graph representing multi-physical coupling is constructed by combining the parameter deviation of internal mechanism degradation and the multi-source features of external working conditions, finally, the feature aggregation and time sequence reasoning are completed through the graph neural network and the time sequence neural network trained offline, the fault probability is output, and the optimal diagnosis result is determined; thereby, the fault recognition accuracy and the robustness in the running stage are improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and more specifically, to a method and system for diagnosing high-frequency transformer faults. Background Technology

[0002] The content in this section only provides background information related to this invention and may not constitute prior art.

[0003] High-frequency transformers are widely used in rail transit, power electronic conversion, and high-power-density energy conversion. They operate at high frequencies and experience rapid voltage and current changes, and typically use solid insulation systems as the primary insulating medium. Under long-term combined electro-thermal-mechanical stress, solid insulation materials are prone to aging and performance degradation, accompanied by phenomena such as enhanced partial discharge, localized temperature rise, structural vibration, and abnormal sound radiation. In severe cases, this can lead to insulation breakdown or inter-turn short circuits.

[0004] In existing technologies, high-frequency transformer condition monitoring and fault diagnosis methods mainly include feature analysis methods based on single or a few physical quantities and data-driven intelligent diagnostic methods. The former often analyzes single or a few physical quantities such as partial discharge, temperature, vibration, or acoustics, making it difficult to fully reflect the coupled evolution characteristics of multiple physical quantities during solid insulation degradation. Furthermore, its diagnostic stability and accuracy are limited in complex electromagnetic interference and noise environments. The latter typically relies on a large number of labeled samples and lacks physical mechanism constraints, making it prone to decreased generalization performance and insufficient physical consistency of inference results when samples are insufficient or operating conditions change.

[0005] Digital twin technology can realize the virtual-real mapping and inference of equipment status through mechanism models, but existing applications are mostly at the offline or static model stage, making it difficult to combine multi-source monitoring data to realize online inversion and real-time updating of parameters, and also lacking the ability to express and intelligently reason about multi-physical coupling relationships.

[0006] Therefore, there is an urgent need for a technical solution that can address the multi-source heterogeneous monitoring data of high-frequency transformers, combine multi-physics mechanism constraints and support online updates, and explicitly characterize the multi-physics coupling relationships between components to achieve fault diagnosis and location, so as to improve diagnostic accuracy, robustness and interpretability. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for diagnosing high-frequency transformer faults, thereby improving the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a method for diagnosing high-frequency transformer faults, including: The system collects time-stamped target signals from the transformer during operation. These target signals include partial discharge signals, temperature signals, voltage and current signals, vibration signals, insulation material decomposition gas signals, and acoustic array signals. Feature extraction is performed on the target signal to obtain target features; target features include partial discharge features, temperature features, vibration features, gas features, and acoustic features; A digital twin model is established based on the structural parameters, material parameters, and corresponding physical rules of the transformer. Equivalent operating conditions representing solid insulation degradation and structural degradation are set in the digital twin model, and numerical solutions are used to obtain the local electric field distribution, temperature field distribution, and vibration response data of key components inside the transformer. Based on the local electric field distribution and a preset insulation breakdown criterion, the potential region of partial discharge within the transformer is determined. An equivalent sound source is constructed based on the potential region and the energy evolution of partial discharge. The equivalent sound source is used as the volume sound source term in the acoustic wave equation, and the vibration response data is used as the acoustic boundary excitation to calculate the equivalent acoustic characteristics corresponding to the equivalent operating conditions. An equivalent dynamic model of the insulation degradation gas is established, and the temperature field distribution is correlated with the partial discharge characteristics to obtain the theoretical concentration characteristics of at least one degradation gas. The local electric field distribution, temperature field distribution, vibration response, equivalent acoustic characteristics, and theoretical concentration characteristics are integrated to form a multi-physics mechanism sample for characterizing solid insulation degradation. A generative adversarial network is constructed, and the mechanism samples and field measured data are used as inputs respectively. The generative adversarial network is trained adversarially to generate extended samples that match the distribution of field data, forming an extended dataset. During the operation phase, the measured voltage, current and ambient temperature are used as inputs to perform online inversion of the indirect measurable parameters inside the digital twin model, obtain the optimal parameters of the indirect measurable parameters under the current time window, and use the optimal parameters to update the digital twin model in real time. The target deviation between the optimal parameters and the corresponding actual parameters is calculated. The key components of the transformer are defined as graph nodes, and the target features and target deviations are used as node features. The conduction paths of the preset physical field are used as connecting edges. Based on the node features and connecting edges, a feature dynamic graph reflecting the multi-physical coupling relationship is constructed. Based on an extended dataset, a multi-dimensional edge feature sample dynamic graph covering normal and various fault conditions is constructed and labeled. The labeled data is then used for offline training of a graph neural network. Using the feature dynamic graph as input, the trained graph neural network is invoked to update the node features in the feature dynamic graph, thus completing the representation of abnormal patterns. The updated node features are aggregated through a graph-level readout mechanism to obtain a time series of system-level global feature vectors. The time series is then input into a pre-trained temporal neural network model for temporal inference, outputting probability vectors for each fault type. The fault type with the highest probability among the probability vectors of each fault type is selected as the diagnostic result.

[0008] Furthermore, the steps for acquiring time-stamped target signals during transformer operation specifically include: A unified time reference signal is used to synchronize the monitoring data of each channel and to mark the monitoring data of each channel with a timestamp. Using sudden changes in partial discharge amplitude and / or sudden changes in vibration energy as triggering conditions, monitoring data of preset durations are extracted before and after the triggering time to form an event time window; The monitoring data of signals with frequencies lower than the preset frequency in the target signal are mapped to a unified time axis of the event time window through interpolation and / or resampling, while signals with frequencies higher than the preset frequency are aligned in time.

[0009] Furthermore, the steps for feature extraction of the target signal specifically include: Multi-layer wavelet decomposition is used to extract the maximum discharge amount, average number of discharges, and discharge phase distribution within a preset time window for the partial discharge signal. Infrared thermal imaging is obtained through temperature signals. An adaptive threshold segmentation method is used to extract high-temperature regions that exceed the temperature threshold, and the highest temperature and temperature rise gradient of the high-temperature regions are extracted. Perform a fast Fourier transform on the vibration signal to extract the fundamental frequency and harmonic energy ratio; The concentration of pyrolysis gas in insulating materials was normalized and the gas concentration and rate of change were extracted. A beamforming algorithm is used to generate a sound pressure distribution map for the acoustic array signal, and the coordinates of the sound source center and the sound pressure energy characteristics are extracted.

[0010] Furthermore, an equivalent sound source is constructed based on the potential region and the energy evolution of partial discharge, specifically including: Within the potential region, a time-varying function of the partial discharge energy is constructed by combining an equivalent gap capacitance model with a preset partial discharge current or charge waveform. Based on the time-varying function, the partial discharge process within the potential region is equivalent to the volume sound source term in the acoustic wave equation, thus completing the construction of the equivalent sound source.

[0011] Furthermore, the steps for adversarial training of the generative adversarial network specifically include: The Generative Adversarial Network (GAN) adopts a Wasserstein GAN with gradient penalty. The GAN consists of a generator subnetwork and a discriminator subnetwork. A gradient penalty term is set at the output of the discriminator subnetwork. During training, the mechanism sample is input into the generator subnetwork to generate samples. The generated samples and the field test data are then input into the discriminator subnetwork for discrimination. The gradient penalty term constrains the gradient norm of the discriminator function in the discriminator subnetwork.

[0012] Furthermore, the step of performing online inversion on the indirectly measurable parameters within the digital twin model to obtain the optimal parameters for the indirectly measurable parameters under the current time window specifically includes: Indirectly measurable parameters that are highly correlated with the insulation state of the transformer are selected as the parameter vector to be inverted. The parameters to be inverted include thermal resistance, equivalent capacitance, mechanical stiffness, equivalent dielectric constant, dielectric loss, and equivalent breakdown field strength. The particle swarm optimization algorithm is used to iteratively update the parameter vector. The goal of the iterative optimization is to minimize the preset error index. By iteratively solving the particle velocity and position, the optimal parameters corresponding to the indirect measurable parameters under the current time window are obtained.

[0013] Furthermore, the steps for offline training of the graph neural network specifically include: During the training of the graph neural network, a physical consistency constraint loss term is constructed. The constraint loss term is determined based on the change in the parameter vector obtained by inversion within adjacent time windows. It is used to penalize the graph neural network when the change exceeds a preset threshold, so as to limit the rate of change of the parameter vector and make the fault diagnosis results consistent with the actual thermo-mechanical evolution process of the transformer.

[0014] Furthermore, the step of aggregating the updated node features through a graph-level readout mechanism specifically includes: Graph-level readout is implemented using differentiable pooling. Based on the node parameter deviation rate, corresponding aggregation weights are adaptively generated. The updated node features are then weighted and aggregated according to the aggregation weights to obtain the system-level global feature vector.

[0015] Furthermore, after selecting the fault type with the highest probability from the probability vectors of each fault type as the diagnostic result, the following steps are also included: Sensitivity analysis is performed on the nodes output by the graph neural network to characterize the contribution of each node to fault determination. Establish a mapping relationship between nodes with a contribution greater than a preset value and the digital twin model, map the diagnostic results of the corresponding nodes to the corresponding components of the digital twin model, and display them in a preset display window.

[0016] Secondly, this application also provides a diagnostic system for high-frequency transformer faults, comprising: The data acquisition module is used to acquire time-stamped target signals of the transformer during operation. These target signals include partial discharge signals, temperature signals, voltage and current signals, vibration signals, insulation material pyrolysis gas signals, and acoustic array signals. The feature extraction module is used to extract features from the target signal to obtain target features; target features include partial discharge features, temperature features, vibration features, gas features, and acoustic features. The twin modeling module is used to establish a digital twin model based on the structural parameters, material parameters, and corresponding physical rules of the transformer. In the digital twin model, equivalent operating conditions representing solid insulation degradation and structural degradation are set, and numerical solutions are performed to obtain the local electric field distribution, temperature field distribution, and vibration response data of key components inside the transformer. Based on the local electric field distribution and a preset insulation breakdown criterion, the potential region of partial discharge within the transformer is determined. An equivalent sound source is constructed based on the potential region and the energy evolution of partial discharge. The equivalent sound source is used as the volume sound source term in the acoustic wave equation, and the vibration response data is used as the acoustic boundary excitation to calculate the equivalent acoustic characteristics corresponding to the equivalent operating conditions. An equivalent dynamic model of insulation degradation gas is established, and the temperature field distribution is correlated with the partial discharge characteristics to obtain the theoretical concentration characteristics of at least one degradation gas. The local electric field distribution, temperature field distribution, vibration response, equivalent acoustic characteristics, and theoretical concentration characteristics are integrated to form a multi-physics mechanism sample for characterizing solid insulation degradation. The sample expansion module is used to construct a generative adversarial network. It takes the mechanism samples and the field measured data as inputs, respectively, and performs adversarial training on the generative adversarial network to generate expanded samples that match the distribution of field data, thus forming an expanded dataset. The online parameter inversion module is used to perform online inversion of the indirectly measurable parameters inside the digital twin model during the operation phase, using measured voltage, current and ambient temperature as inputs, to obtain the optimal parameters of the indirectly measurable parameters under the current time window, and to update the digital twin model in real time using the optimal parameters. The graph construction module is used to calculate the target deviation between the optimal parameters and the corresponding actual parameters. It defines the key components of the transformer as graph nodes, and uses the target features and target deviations as node features. It uses the conduction paths of the preset physical fields as connecting edges. Based on the node features and connecting edges, it constructs a feature dynamic graph that reflects the multi-physical coupling relationship. The diagnostic module, based on an extended dataset, constructs and annotates a multi-dimensional dynamic graph of side feature samples covering normal and various fault conditions. The annotated data is then used for offline training of a graph neural network. Using the dynamic feature graph as input, the trained graph neural network updates the node features in the dynamic feature graph, completing the representation of abnormal patterns. A graph-level readout mechanism aggregates the updated node features to obtain a time series of system-level global feature vectors. This time series is then input into a pre-trained temporal neural network model for temporal inference, outputting probability vectors for each fault type. The results module is used to select the fault type with the highest probability from the probability vectors of each fault type as the diagnostic result.

[0017] The beneficial effects of this invention are as follows: This invention first collects time-stamped signals from multiple target types, including partial discharge, temperature, voltage and current, vibration, insulation degradation gas, and acoustic arrays, and extracts corresponding features. Simultaneously, it builds a digital twin model based on transformer structure, material parameters, and physical rules to simulate equivalent working conditions of insulation and structural degradation, solving for multi-physics field data and generating multi-physics field mechanism samples. Then, it expands the fault sample dataset by fusing mechanism samples with field-measured data using a generative adversarial network to address the sample scarcity problem. During operation, the digital twin model is updated in real-time through online parameter inversion. A dynamic feature graph representing multi-physics coupling is constructed by combining parameter deviations and multi-source features. Finally, feature aggregation and temporal inference are completed using offline-trained graph neural networks and temporal neural networks, outputting the fault probability and determining the optimal diagnostic result. This solves the problems of difficulty in synchronizing heterogeneous multi-source data, weak early insulation degradation symptoms that are easily affected by noise, insufficient model generalization ability due to scarce fault samples, lack of physical consistency constraints, and insufficient interpretability of location results, thereby improving fault identification accuracy and robustness during operation. Attached Figure Description

[0018] Figure 1 A flowchart of a high-frequency transformer fault diagnosis method provided by the present invention; Figure 2 This is a diagram of the architecture of the multi-source heterogeneous sensing and digital twin interaction system in this invention; Figure 3 This is a schematic diagram of the physical-data hybrid driven dynamic graph feature mapping principle in this invention; Figure 4 This is a diagram of the hierarchical graph neural network reasoning and three-dimensional diagnostic architecture in this invention; Figure 5 A schematic diagram of a high-frequency transformer fault diagnosis system provided for the invention.

[0019] In the diagram: 201, Data Acquisition Module; 202, Feature Extraction Module; 203, Twin Modeling Module; 204, Sample Expansion Module; 205, Online Parameter Inversion Module; 206, Graph Construction Module; 207, Diagnosis Module; 208, Results Module. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] like Figure 1 and 2As shown in the embodiment of the present invention, a method for diagnosing high-frequency transformer faults includes: S101 collects target signals with timestamps during transformer operation. The target signals include partial discharge signals, temperature signals, voltage and current signals, vibration signals, insulation material pyrolysis gas signals, and acoustic array signals.

[0022] Specifically, the solid insulation degradation and fault evolution of high-frequency transformers are the result of the coupling effects of multiple physical fields, including electricity, heat, force, sound, and air. A single physical quantity signal cannot fully capture the full-dimensional signs of fault initiation and development. However, by simultaneously acquiring the aforementioned multiple types of target signals, the operating status of the transformer insulation and structure can be comprehensively characterized from multiple dimensions, such as discharge characteristics, thermal state, electrical operation, mechanical vibration, insulation degradation products, and acoustic radiation characteristics. In actual implementation, target signals can be continuously acquired and cover typical operating conditions such as no-load, rated load, and overload. For example, in the embodiment, the aforementioned target signals were continuously acquired for 3 hours, covering operating conditions such as no-load, rated load, and 110% overload, effectively ensuring the comprehensive coverage of the acquired data on the characteristics of different stages of insulation fault evolution.

[0023] In the specific implementation of acquiring time-stamped target signals from transformers during operation, a unified time reference signal is first used to synchronize the monitoring data of each channel, and a timestamp is marked on the monitoring data of each channel. The principle is that the acquisition equipment and signal transmission links of each monitoring channel have inherent differences, which can easily cause acquisition delays. Furthermore, the fault-related signals of high-frequency transformers, such as partial discharge and vibration, have significant transient characteristics. If the multi-source signals are inaccurate in the time dimension, the spatiotemporal correlation characteristics of multi-physics coupling will be directly lost, failing to accurately reflect the temporal correspondence between faults and various physical quantities. Therefore, a unified time reference is used to achieve time synchronization of each channel and mark the timestamps, ensuring the spatiotemporal consistency of multi-source heterogeneous monitoring data. In specific implementation, a dual synchronization mechanism of BeiDou time synchronization and local crystal oscillator is used to mark the data of each channel with a timestamp accuracy of ±0.3μs. Through the calibration effect of this high-precision timestamp, the acquisition delay of multiple devices is effectively eliminated, providing an accurate time reference for subsequent fusion analysis of multi-source data.

[0024] Secondly, using sudden changes in partial discharge amplitude and / or vibration energy as triggering conditions, monitoring data of preset durations are extracted before and after the triggering moment to form an event time window. The principle is that early insulation degradation and fault occurrence in high-frequency transformers are usually accompanied by a sudden increase in partial discharge amplitude and abnormal changes in vibration energy. These sudden events are the core time nodes for fault initiation and evolution. Compared to continuously collecting and processing all monitoring data, using such fault-related sudden events as triggering conditions to extract data for a specific duration can accurately focus on the effective data segments related to the fault and filter out redundant data in fault-free states. In specific implementation, a trigger threshold is set based on the partial discharge amplitude or vibration energy, and a preset duration is extracted before the triggering moment. The preset duration is 1.0s, and the time is truncated afterward. With a duration of 1.5 seconds, an event time window of 2.5 seconds was formed, successfully achieving accurate extraction of insulation anomaly-related event data and locking in the effective data range for subsequent feature extraction and fusion inference.

[0025] Finally, the monitoring data of signals with frequencies lower than the preset frequency in the target signal are mapped to a unified time axis of the event time window through interpolation and / or resampling. Signals with frequencies higher than the preset frequency are aligned in time. The principle is that the target signal contains low-frequency signals such as temperature and gas from the cracking of insulating materials. The sampling rate of these signals is much lower than that of high-frequency signals such as partial discharge, vibration, and acoustic arrays. If multi-source signal fusion is performed directly, the significant difference in time resolution will lead to spatiotemporal misalignment of multimodal features, which cannot accurately reflect the spatiotemporal correspondence of multi-physics coupling. Therefore, interpolation and / or resampling are used for time axis mapping for low-frequency signals, and high-frequency signals are directly aligned in time, thereby unifying the time axis reference of multi-source signals and eliminating the feature misalignment problem caused by the difference in sampling rate. In practice, for target signals such as temperature and gas concentration that are below the preset frequency, interpolation, preservation, or segmented statistics are used to establish a correspondence with high-frequency signals such as partial discharge and vibration. For high-frequency signals, event-triggered window sampling and feature extraction are used to achieve temporal alignment. Ultimately, feature-level alignment and fusion of multi-source signals are achieved, forming structured fusionable data units. At the same time, the collected data is cleaned during the process, filtering out invalid data such as damaged data files, deduplicating based on file hash values, and storing them in a temporary buffer. Information such as channel number, sampling rate, trigger timestamp, sensor location, and operating condition identifier are recorded synchronously to ensure the integrity and traceability of the time-aligned data.

[0026] S102, extract features from the target signal to obtain target features; target features include partial discharge features, temperature features, vibration features, gas features and acoustic features.

[0027] Specifically, firstly, multi-level wavelet decomposition is applied to the partial discharge signal to extract the maximum discharge quantity, average number of discharges, and discharge phase distribution within a preset time window. The principle is that partial discharge signals are easily affected by noise such as electromagnetic interference from the power system during acquisition. Wavelet decomposition, with its multi-scale time-frequency analysis operator, can achieve multi-level decomposition and soft-threshold denoising of the partial discharge signal. After reconstructing the signal, the core discharge features are preserved, and features are extracted within a preset event time window focusing on fault-related data. This allows for precise screening of discharge information directly related to insulation degradation, avoiding interference from redundant data in fault-free states. The extracted maximum discharge quantity, average number of discharges, and discharge phase distribution can directly quantify the intensity, frequency, and phase patterns of partial discharge, which are core electrical characteristic signs of aging and breakdown faults in solid insulation, and can intuitively reflect the internal discharge state of the insulation. In specific implementation, the db6 wavelet basis function is used to perform multi-level wavelet decomposition on the partial discharge signal within a 2.5s preset time window. An improved soft-threshold denoising method is used for the low-frequency approximation coefficients. The threshold is calculated using an adaptive adjustment formula, as follows: (1) In the formula, An adaptive threshold; The standard deviation of noise; This is the signal length.

[0028] After denoising, the reconstructed signal successfully extracted the maximum discharge amount. 650pC, average number of discharges The frequency was 11.2 times / ms. At the same time, the phase distribution characteristics of the discharge phase were mainly concentrated on the rising and falling edges of the voltage. The results of this characteristic are highly consistent with the partial discharge law during the aging of inter-turn insulation of high-frequency transformers, providing key discharge characteristic basis for subsequent identification of insulation fault types.

[0029] Secondly, infrared thermal imaging is obtained through temperature signals. An adaptive threshold segmentation method is used on the infrared thermal imaging to extract high-temperature areas exceeding the temperature threshold, and to extract the highest temperature and temperature rise gradient of the high-temperature areas. Since solid insulation degrades with the increase of local dielectric loss, resulting in local temperature rise anomalies, the temperature signal is a key physical quantity reflecting the thermal aging of transformer insulation. Infrared thermal imaging generated from the temperature signal can intuitively present the temperature distribution of various components of the transformer. The adaptive threshold segmentation method, instead of a fixed threshold, can automatically determine the temperature threshold based on the grayscale distribution characteristics of the infrared thermal imaging, adapting to the temperature distribution differences under different operating conditions such as no-load, rated load, and overload of the transformer, effectively avoiding missed or false detections of high-temperature areas caused by fixed thresholds. Furthermore, extracting the highest temperature and temperature rise gradient of the high-temperature areas can quantify the severity of local thermal anomalies and the uneven distribution of the temperature field, respectively. The combination of these two can accurately characterize the thermal aging characteristics of the contact area between the insulation layer and the winding. In practice, the Otsu method (adaptive threshold segmentation algorithm) is used to set a temperature threshold of 83℃. High-temperature areas are extracted from infrared thermal imaging, and combined with temperature signals from distributed fiber optic temperature measurement, the highest temperature of the insulation layer is calculated. The temperature gradient is 91℃. The temperature was 13.5℃ / cm, thus capturing the local thermal anomaly characteristics caused by insulation thermal aging.

[0030] Third, a Fast Fourier Transform (FFT) is performed on the vibration signal to extract the fundamental frequency and harmonic energy ratios. Since the vibration signal of a high-frequency transformer is generated by the mechanical vibration of key components such as the windings and core, solid insulation degradation alters the coupling characteristics of the winding-insulation structure, leading to significant changes in the frequency domain characteristics of the vibration signal. Therefore, the FFT can convert the time-domain vibration signal into a frequency-domain signal, enabling precise extraction of frequency domain characteristics. This overcomes the limitation that time-domain vibration signals are difficult to characterize changes in structural characteristics. The extracted fundamental frequency and harmonic energy ratios are core indicators reflecting changes in vibration characteristics and can quantify the changes in winding-insulation structure coupling characteristics caused by insulation aging. In specific implementation, the FFT is performed on the vibration signal within the event time window, yielding an energy ratio of 0.58 for the 30kHz fundamental frequency component, 0.32 for the 60kHz second harmonic, and 0.10 for the 90kHz third harmonic.

[0031] Fourth, the concentration of the cracked gas in the insulation material is normalized, and the gas concentration and rate of change are extracted. Since high-frequency transformers use a solid insulation system, the insulation material will crack under the combined stress of electricity, heat, and partial discharge, producing characteristic cracked gases such as H2, CO, and CH4. The gas concentration and rate of change directly reflect the degree and rate of insulation cracking and are core indicators characterizing the chemical degradation of the insulation. Therefore, normalizing the gas concentration can eliminate the interference of external factors such as ambient temperature and humidity, and the accuracy of the acquisition equipment on the collected data, making the gas characteristics comparable and measurable, facilitating subsequent error construction with the simulated gas characteristics of the digital twin model. In practice, the Z-score normalization method was used to process the concentration data of pyrolysis gases such as H2, CO, and CH4 in the insulating material. The extracted values ​​were: H2 concentration 7.8 ppm with a change rate of 0.45 ppm / h; CO concentration 11.2 ppm with a change rate of 0.28 ppm / h; and CH4 concentration 3.3 ppm with a change rate of 0.12 ppm / h. These gas characteristics are highly consistent with the thermal pyrolysis pattern of epoxy insulation, providing important chemical characteristic basis for insulation aging diagnosis.

[0032] Fifth, a beamforming algorithm is used to generate a sound pressure distribution map from the acoustic array signal, extracting the coordinates of the sound source center and the sound pressure energy characteristics. Since partial discharge of high-frequency transformers generates characteristic acoustic signals, and structural vibrations also cause abnormal sound radiation, using acoustic array signals can comprehensively capture the sound radiation characteristics around the transformer. Furthermore, the beamforming algorithm can perform spatial filtering and focusing on the multi-channel signals of the acoustic array, generating an accurate sound pressure distribution map. This enables spatial localization of the sound source and quantitative analysis of sound pressure energy. The extracted coordinates of the sound source center and sound pressure energy characteristics can correlate acoustic anomalies with the physical spatial location of the transformer, quantify the intensity of the acoustic anomalies, and establish a physical correlation between acoustic features and partial discharge of insulation, providing crucial acoustic spatial characteristic basis for subsequent spatial localization of fault locations. In specific implementation, a coherent beamforming algorithm is used to process the acoustic array signal and generate a sound pressure distribution map. The extracted coordinates of the sound source center are X=0.18m, Y=0.08m, Z=0.22m relative to the transformer center point, and the peak value of the sound pressure energy spectrum is 1.15×10⁻⁶. -3 Pa 2 Furthermore, the signal-to-noise ratio of the acoustic signal is ≥28dB, effectively establishing a physical correlation between acoustic characteristics and partial discharge in insulation, providing key acoustic characteristic support for fault reasoning and localization.

[0033] S103. A digital twin model is established based on the structural parameters, material parameters, and corresponding physical rules of the transformer. Equivalent operating conditions representing solid insulation degradation and structural degradation are set in the digital twin model, and numerical solutions are used to obtain the local electric field distribution, temperature field distribution, and vibration response data of key components inside the transformer. Based on the local electric field distribution and a preset insulation breakdown criterion, the potential region of partial discharge within the transformer is determined. An equivalent sound source is constructed based on the potential region and the energy evolution of partial discharge. The equivalent sound source is used as the volume sound source term in the acoustic wave equation, and the vibration response data is used as the acoustic boundary excitation to calculate the equivalent acoustic characteristics corresponding to the equivalent operating conditions. An equivalent dynamic model of the insulation decomposition gas is established, and the temperature field distribution is correlated with the partial discharge characteristics to obtain the theoretical concentration characteristics of at least one decomposition gas. The local electric field distribution, temperature field distribution, vibration response, equivalent acoustic characteristics, and theoretical concentration characteristics are integrated to form a multi-physics mechanism sample for characterizing solid insulation degradation.

[0034] Specifically, the analysis is based first on the actual structural parameters of the transformer (e.g., the winding arrangement, core and clamp structure dimensions of a high-frequency transformer with a rated capacity of 1000kVA, primary voltage of 6kV, secondary voltage of 480V, and operating frequency of 30kHz), and solid insulation material parameters (e.g., equivalent dielectric constant 4.6, dielectric loss tangent 0.0015, equivalent breakdown field strength 23kV / mm, equivalent thermal resistance 19K / W, and clamp equivalent mechanical stiffness 5.2×10⁻⁶). 5 By combining the fundamental physical rules of electromagnetic induction, heat conduction, and elastic vibration (N / m, etc.), an electromagnetic-thermal-mechanical coupled digital twin model is constructed. This modeling method achieves a precise virtual-real mapping between the physical entity of the transformer and the virtual model, enabling the virtual model to reproduce the multi-physics interactions in actual operation. Subsequently, equivalent operating conditions characterizing solid insulation degradation (such as reduced dielectric constant, thinner insulation thickness, and decreased equivalent breakdown field strength) and structural degradation (such as reduced mechanical stiffness of clamping components) are set in the digital twin model (for example, setting the dielectric constant of the third section of the high-voltage winding insulation to decrease by about 32%, the insulation thickness of the fifth section to decrease by about 18%, and the equivalent mechanical stiffness of the clamping components to decrease by about 16%). The multi-physics coupled model is numerically solved to obtain the local electric field distribution, temperature field distribution, and vibration response data of key components such as the core, windings, and clamping components inside the transformer (e.g., For example, the maximum electric field strength obtained by solving is about 21.8 kV / mm, located in the fifth insulation layer of the high-voltage winding, with a maximum simulated temperature of about 87.5℃ and a fundamental frequency of about 30kHz and a vibration response amplitude of about 3.1 μm. This process transforms the abstract insulation and structural degradation into quantifiable physical field distribution data, which not only restores the multi-physics field change law in the fault evolution process, but also provides calculable physical data support for the subsequent determination of potential partial discharge regions and the derivation of acoustic and gas characteristics, thus avoiding the problem of the lack of actual physical law support for mechanism samples.

[0035] Next, the local electric field distribution obtained from the numerical solution is compared with the preset insulation breakdown criterion (the criterion is the local electric field strength). ≥ Equivalent breakdown field strength By comparing the data, the potential regions of partial discharge within the transformer can be determined. The definition of the potential regions is as follows: (2) In the formula, This represents the local electric field strength. This is the equivalent breakdown field strength threshold.

[0036] This determination method is based on the physical mechanism of insulation breakdown and presupposes... The value is 23kV / mm. Based on this, the fifth section of the high-voltage winding insulation layer is determined to be a potential partial discharge region. This ensures the accuracy and rationality of the determination of the potential partial discharge region, allowing subsequent equivalent sound source construction and acoustic characteristic calculations to be carried out around the core fault region, thus improving the matching degree between the mechanism sample and the actual fault characteristics.

[0037] Based on the determined potential region and the evolution of partial discharge energy, an equivalent sound source is constructed. Since the energy released during the partial discharge process is an important source of transformer acoustic radiation anomalies, the potential partial discharge region is used as the spatial basis. Combined with the evolution law of partial discharge energy over time, an equivalent sound source is constructed to achieve the coupling mapping of electrical characteristics (partial discharge) and acoustic characteristics, thus overcoming the defect of the separation between electrical and acoustic physical fields in traditional models. Specifically, within the potential region, the time-varying function of partial discharge energy is constructed by combining the equivalent gap capacitance model with the preset partial discharge current or charge waveform. The time-varying function is as follows: (3) In the formula, For a moment Below, the instantaneous partial discharge energy released during the partial discharge process within the potential partial discharge region of a high-frequency transformer; The initial partial discharge energy released within the potential partial discharge region at the initial moment of partial discharge occurrence; is the time constant.

[0038] Based on this time-varying function, the partial discharge process in the potential region is equivalent to the volume sound source term in the acoustic wave equation, thus completing the construction of the equivalent sound source. This involves using the partial discharge charge waveform in the form of exponential decay and setting the time constant. An energy-time variation function was constructed for 50μs and equivalently represented as a volumetric sound source term. This process follows the physical mechanism of partial discharge sound generation, ensuring the consistency of the equivalent sound source with the acoustic signal generated by actual partial discharge in terms of physical laws.

[0039] The constructed equivalent sound source was then used as the volume sound source term in the acoustic wave equation. Simultaneously, the vibration response data of key components obtained from numerical solutions were used as acoustic boundary excitations to solve the acoustic wave equation in the external air domain of the transformer. The equivalent acoustic characteristics corresponding to the equivalent operating condition were calculated. This calculation method simultaneously considers the combined contributions of partial discharge and structural vibration to the acoustic characteristics, fully reconstructing the generation mechanism of the transformer's acoustic characteristics. The sound pressure energy characteristic was then calculated to be 1.08 × 10⁻⁶. -3 Pa 2 This solves the problem of acoustic characteristic distortion caused by considering only partial discharge or vibration.

[0040] Then, an equivalent kinetic model of the insulation pyrolysis gas is established, and the temperature field distribution is quantitatively correlated with the partial discharge characteristics. Because the solid insulation material of the high-frequency transformer will pyrolyze under the combined stress of electro-thermal-partial discharge, producing characteristic pyrolysis gases such as H2, CO, and CH4, the generation rate of these gases is related to the temperature field distribution. Partial discharge characteristics (such as maximum discharge amount) Highly relevant, for example, setting the gas generation rate. Proportional to The theoretical concentration characteristics of at least one cracked gas such as H2, CO, and CH4 can be obtained through this model (H2=7.5ppm, CO=10.8ppm, CH4=3.1ppm). This modeling method realizes the coupled mapping of thermal characteristics, electrical characteristics and chemical characteristics (gases), filling the gap in the traditional mechanism model that lacks insulation chemical degradation characteristics.

[0041] Finally, the local electric field distribution, temperature field distribution, vibration response obtained by numerical solution, as well as the derived equivalent acoustic characteristics and theoretical concentration characteristics are integrated to form a multi-physics mechanism sample for characterizing the degradation of solid insulation.

[0042] S104. Construct a generative adversarial network (GAN). Use the mechanism samples and the field measured data as inputs to train the GAN against adversarial forces, generating extended samples that match the distribution of the field data, thus forming an extended dataset.

[0043] Specifically, while the mechanism samples possess physical consistency, they exhibit differences in noise texture and sampling path distribution compared to the field-measured data, and there is a scarcity of labeled samples from high-frequency transformer fault sites. Therefore, adversarial training is used to achieve style transfer of mechanism samples, preserving physical regularities while ensuring that the generated samples conform to the statistical characteristics of the field data, fundamentally compensating for the deficiency of small fault samples. In specific implementation, 1200 sets of multi-physics mechanism samples and 200 sets of field insulation anomaly measured data are used as network inputs for training. The steps for adversarial training of the generative adversarial network specifically include: The generative adversarial network (GAN) employs a Wasserstein GAN with gradient penalty. This network comprises a generator subnetwork and a discriminator subnetwork. This network was chosen because traditional GANs are prone to training instability and mode collapse, while the Wasserstein GAN with gradient penalty can constrain the gradient norm of the discriminator function through a gradient penalty term, effectively improving the stability of adversarial training and reducing the risk of mode collapse. In specific implementation, the generator subnetwork uses an 8-layer convolutional structure, and the discriminator subnetwork uses a 6-layer convolutional structure, with a gradient penalty coefficient λ≈10 to align the distribution of mechanistic samples with field samples. During training, mechanistic samples are input into the generator subnetwork to generate samples, and then these generated samples, along with field-measured data, are input into the discriminator subnetwork for discrimination. A gradient penalty term is set at the output of the discriminator subnetwork, which constrains the gradient norm of the discriminator function, forcing it to satisfy the Lipschitz continuity condition, making the update direction of the generator subnetwork more stable. Simultaneously, to prevent style transfer from damaging the core physical features of the mechanistic samples, a physical feature operator is introduced. The physical consistency loss is constructed, and the specific formula is as follows: (4) In the formula, This represents a loss of physical consistency. This is a physical feature operator used to extract key features characterizing the physical mechanism of a device from input samples; the key features preferably include one or more of the following: partial discharge phase distribution features, temperature gradient features, vibration response features, acoustic propagation features, and gas characterization features; The mapping function for generating subnetworks is used to map the input mechanism samples to generated samples that are closer to the distribution of field samples; Mechanism samples generated from digital twin models or multiphysics mechanism models; This is a physical feature index used to characterize the m-th type of key physical features; The physical feature set represents the set of all key physical feature categories that participate in the consistency constraints.

[0044] This constrains the deviation between the generated samples and the original mechanism samples in key physical characteristics such as partial discharge phase distribution and temperature gradient, ensuring that the consistency error is ≤3.5% and guaranteeing the physical rationality of the generated samples. During training, batch size=64, learning rate=0.0002, 10000 iterations, and the Adam optimizer were used for training. After training, the maximum mean difference (MMD) was used to measure the data distribution difference. This metric accurately characterizes the degree of matching between the generated samples and the field-measured samples in the feature space. The data distribution MMD value decreased from the initial 0.28 to 0.04, proving the significant distribution alignment effect. Finally, the trained generative model was used to generate 800 sets of extended samples that conform to the field statistical characteristics. These were then merged with the field-measured samples to form an extended dataset. This dataset combines physical consistency and field statistical characteristics, becoming the feature mapping benchmark for subsequent dynamic graph construction and the core data foundation for offline training of hierarchical graph neural networks.

[0045] S105, during the operation phase, uses measured voltage, current and ambient temperature as inputs to perform online inversion of the indirectly measurable parameters inside the digital twin model, obtains the optimal parameters of the indirectly measurable parameters under the current time window, and uses the optimal parameters to update the digital twin model in real time.

[0046] Specifically, measured voltage, current, and ambient temperature are used as inputs. These parameters are directly measurable external boundary condition parameters of transformer operation. Using them as driving inputs for the digital twin model ensures that the model input boundary is completely consistent with the actual operating conditions on site, avoiding simulation result distortion caused by operating condition deviations and providing accurate boundary constraints for subsequent parameter inversion. Online inversion is performed on indirect measurable parameters within the digital twin model (such as thermal resistance, equivalent capacitance, mechanical stiffness, equivalent dielectric constant, dielectric loss, and equivalent breakdown field strength) to obtain the optimal parameters for indirect measurable parameters within the current time window. This is because core parameters such as insulation thermal resistance and equivalent dielectric constant of high-frequency transformers directly reflect the health status of insulation and structure, but cannot be directly measured by sensors. They are core indicators characterizing internal degradation of equipment, and traditional methods cannot obtain their real-time values. This step achieves real-time quantification of indirect measurable parameters through online inversion, establishing a mapping link between measurable external signals and internal degradation states.

[0047] In specific implementation, such as Figure 3As shown, firstly, indirect measurable parameters highly correlated with the transformer insulation state are selected as the parameter vector to be inverted. These parameters include thermal resistance, equivalent capacitance, mechanical stiffness, equivalent dielectric constant, dielectric loss, and equivalent breakdown field strength. These parameters correspond to the thermal, electrical, and mechanical properties of the insulation, respectively, and are directly and strongly correlated with the solid insulation degradation process. A decrease in the equivalent dielectric constant and an increase in dielectric loss directly characterize the aging of the insulation material; an increase in thermal resistance reflects the deterioration of the insulation's thermal conductivity; and a decrease in mechanical stiffness corresponds to the loosening or degradation of the winding-insulation structure. Integrating these parameters into a parameter vector allows for a comprehensive and multi-dimensional characterization of the real-time state of the transformer's internal insulation and structure, avoiding the one-sided characterization caused by single-parameter inversion. For example, selecting a parameter vector... =[Insulation thermal resistance] Equivalent capacitance Mechanical stiffness Insulation dielectric constant Dielectric loss Equivalent breakdown field strength Set the initial value θ0 = [19 K / W, 22 nF, 5.2 × 10⁻⁶]. 5 [N / m, 4.6, 0.0015, 23kV / mm], achieving comprehensive characterization of the thermal, electrical, and mechanical properties of insulation.

[0048] Subsequently, a particle swarm optimization (PSO) algorithm is used to iteratively update the parameter vector, with the goal of minimizing a preset error index. By iteratively solving for particle velocity and position, the optimal parameters corresponding to indirect measurable parameters within the current time window are obtained. The PSO algorithm possesses advantages such as strong global optimization capability, fast convergence speed, and no need for gradient information. It is suitable for inversion scenarios involving multiple parameters and multiple physics fields. By minimizing the preset error between multimodal measured features and twin model simulation features, it ensures that the inverted parameter vector matches the digital twin model's output to the greatest extent possible with the actual measured state. The preset error can be constructed in various ways, as detailed below: Method 1: (5) In the formula, For indexing, For modal sets; These are the weighting coefficients; The parameter vector to be inverted; For the first Measured feature vectors or scalar features of a mode; For in the parameter vector The corresponding simulation features are calculated from the digital twin model.

[0049] To ensure that parameter evolution conforms to the thermal inertia and structural inertia of the equipment, a physical consistency constraint is introduced in this implementation: (6) In the formula, , Adjacent time windows and -1 The parameter vector obtained by inversion; This is the upper limit threshold for parameter variation, used to describe the maximum allowable change in parameters of the device within an adjacent time window under constraints of thermal and structural inertia, and is expressed as... Driver parameters updated.

[0050] Method 2: Calculate simulation features using a digital twin model. Compared with measured features The weighted average yields the preset error index: (7) In the formula, For indexing, It is a combination of partial discharge, temperature, vibration, acoustics, and gas. These are the weighting coefficients. =[0.3, 0.25, 0.15, 0.15, 0.15], highlighting the error weights of the core insulation features.

[0051] Taking Method 2 as an example, a weighted comprehensive error index was constructed that includes partial discharge, temperature, vibration, acoustics, and gas modes, highlighting the error weights of the core insulation characteristics. Simultaneously, an improved particle swarm optimization was employed, and a physical consistency constraint term was introduced to limit parameter abrupt changes between adjacent time windows. The physical consistency constraint term is as follows: (8) In the formula, , Adjacent time windows and -1 inversion obtained parameter vector; and with Driver parameters updated.

[0052] The optimal parameters are finally obtained through inversion. =[18.3K / W, 23.7nF, 5.0×10 5 [N / m, 3.9, 0.0021, 21.6kV / mm] reduced the overall model error from 16.2% to 2.8%.

[0053] S106, calculate the target deviation between the optimal parameters and the corresponding actual parameters, define the key components of the transformer as graph nodes, and use the target features and target deviations as node features; use the transmission path of the preset physical field as the connecting edge; based on the node features and connecting edge, construct a feature dynamic graph that reflects the multi-physical coupling relationship.

[0054] Specifically, first calculate the optimal parameters. The target deviation is calculated by comparing the actual parameters with the target deviation of the corresponding actual parameters. The actual parameters are the initial nominal parameters corresponding to the healthy state of indirect measurable parameters (such as thermal resistance, equivalent capacitance, mechanical stiffness, equivalent dielectric constant, dielectric loss, equivalent breakdown field strength, etc.). The deviation rate of this deviation can quantify the degree of invisible insulation and structural degradation inside the equipment at the physical parameter level, providing a core physical degradation characterization for the graph nodes. For example, it can be used to invert the dielectric constant of the insulation. The optimal value is 3.9, corresponding to an initial nominal value of 4.6. The calculation yields... Deviation rate -15.2%, dielectric loss The optimal value is 0.0021, corresponding to an initial nominal value of 0.0015. The calculation yields... The deviation rate was 40.0%, which quantified the degree of insulation aging.

[0055] Secondly, key components of the transformer are defined as graph nodes. This means that core physical components prone to insulation faults, such as the high-voltage winding, low-voltage winding, core, and clamping parts, are defined as independent graph nodes. This ensures a one-to-one correspondence between graph nodes and the physical entities of the transformer, guaranteeing the physical interpretability of subsequent fault location. For example, the node set V = {high-voltage winding ( ), low voltage winding ( ), iron core ( ), clamps ( This process fully covers the core components associated with the fault. Then, the target features and target deviations are used as node features. Specifically, the partial discharge, temperature, vibration, gas, and acoustic features of the corresponding node components extracted in step S102 are combined with the aforementioned target deviation rate to form node features. This allows nodes to simultaneously integrate external measurable signs and internal physical degradation information, comprehensively depicting the insulation operation state of the components and avoiding the one-sidedness of single-feature representation. Furthermore, the conduction paths of the preset physical fields are used as connecting edges. Based on the heat conduction, electric field coupling, mechanical force transmission, sound radiation, and pyrolysis gas migration paths given by the digital twin model, edge connections are established between nodes with physical coupling relationships. This ensures that the graph edge connections fully comply with the objective laws of multi-physics coupling, guaranteeing the physical consistency of the reasoning process from a structural perspective. For example, establish an edge set E={(V1-V2) (electrical coupling), (V1-V3) (heat conduction), (V1-V4) (mechanical force transmission), (V2-V3) (heat conduction), (V1-air) (sound radiation), (V2-air) (sound radiation)} to completely cover the core coupling path of the insulation state evolution.

[0056] Finally, based on node features and connecting edges, a dynamic feature graph reflecting multi-physical coupling relationships is constructed, that is, a graph with optimal parameters is constructed for each edge. The multidimensional edge feature vector is calculated using a preset physical mapping function. =[ =39.5W / (m·K) =17.2℃ =0.85W =2.3%], where the edge features reflect the coupling strength and consistency between virtual and real in real time, and change with the parameter vector. The system updates synchronously and iterates. This allows the node and edge features to be updated synchronously and iterated in real time with the optimal parameters, forming a time-varying graph structure that couples multiple physics such as electricity, heat, force, sound, and air, thus enabling an explicit representation of the transformer fault evolution process.

[0057] S107: Based on an extended dataset, a multi-dimensional side feature sample dynamic graph covering normal and various fault conditions is constructed and labeled. The labeled data is then used for offline training of a graph neural network. Using the feature dynamic graph as input, the trained graph neural network is called to update the node features in the feature dynamic graph, thus completing the representation of abnormal patterns. The updated node features are aggregated through a graph-level readout mechanism to obtain the time series of the system-level global feature vector. The time series is then input into a pre-trained temporal neural network model for temporal inference, outputting the probability vector of each fault type.

[0058] Specifically, this step is divided into two related stages: offline model training and online temporal inference, fully realizing end-to-end inference from multi-physical coupled dynamic graphs to fault type probability output. First, based on the extended dataset generated by S104, a multi-dimensional side feature sample dynamic graph covering normal and various fault conditions is constructed and the fault type is labeled. Then, offline training of the hierarchical graph neural network is completed based on the labeled data, such as... Figure 4 As shown, the extended dataset combines consistency in physical mechanisms with statistical characteristics of field data. Using this dataset to construct dynamic sample graphs, the multi-physical coupling relationships between transformer components can be transformed into structured data that graph neural networks can process. This addresses the limitation of traditional neural networks in explicitly depicting the physical relationships between components. Full-condition sample coverage ensures the model's generalization ability across various fault modes. For example, based on the extended dataset containing 1000 samples, dynamic sample graphs corresponding to normal, inter-turn insulation aging, localized insulation wear, and insulation dampness are constructed in batches, and fault type labels are added, providing a benchmark for supervised model training.

[0059] During offline training, this invention simultaneously constructs a physical consistency constraint loss term. This loss term is determined based on the change in the inversion parameter vector within adjacent time windows. It is used to penalize the model when the change exceeds a preset threshold, limiting the rate of change of the parameter vector and ensuring that the diagnostic results are consistent with the actual thermo-mechanical evolution process of the transformer. The principle is that the insulation degradation and structural deterioration of high-frequency transformers are gradual processes that conform to thermal inertia and structural inertia. The internal physical parameters do not undergo abrupt changes within adjacent windows. The constraint term can penalize abnormal outputs caused by parameter abrupt changes, forcing the model to learn features that conform to physical laws, avoiding overfitting that violates the mechanism, and improving the physical interpretability and robustness of the diagnostic results. After offline training is completed, the feature dynamic map constructed by S106 is used as input, and the trained graph neural network is called to update the node features, completing the component-level abnormal pattern representation.

[0060] In detail, during the message passing phase of the graph neural network, edge-gating weights generated from multi-dimensional edge features are introduced to adjust the transmission strength of neighbor node information to the target node, thereby achieving joint perception of anomalous patterns of multi-physics coupling involving "electricity, heat, force, sound, and air". For any node... Its set of neighboring nodes is denoted as Then by node with neighboring nodes The gating weights generated from the multidimensional edge features between them can be expressed as: (9) In the formula, For trainable parameters, For the edge In the time window The multidimensional edge eigenvectors. It is a non-linear activation function used to normalize the values ​​of the gating weights to the range of 0 to 1.

[0061] This gating weight enables the information transmission strength of corresponding neighbor nodes to be increased in fault-related scenarios such as enhanced thermal coupling, enhanced electrical coupling, or enhanced mechanical coupling. When the coupling strength of the coupling path is weak or the consistency between virtual and real is poor, the information transmission of the corresponding path is automatically weakened, effectively filtering out interference from irrelevant paths.

[0062] After obtaining the edge gating weights, the node representation is iteratively updated through a gating-weighted message passing mechanism, achieving multi-dimensional fusion and enhancement of local anomaly information. Target node In the graph neural network The representation update of a layer can be abstractly represented as: (10) In the formula, , These are the mapping matrices that the model can train; It is a non-linear activation function; For the target node In the The output representation of the layer; For the target node In the Output representation of layer +1; For the target node In the The output representation of the layer.

[0063] Through this representation update method, the target node The new representation integrates its own historical features with the features of all neighboring nodes, and the contribution of neighboring node information is adaptively adjusted by the edge gating weights that reflect the strength of multi-physical coupling, thus realizing the full participation of multi-dimensional edge features in the inference process of the graph neural network.

[0064] Specifically, the inference model is constructed using a four-layer hierarchical graph convolutional network. The node feature dimension evolves along the path of 128→256→512→256. LeakyReLU with a negative slope of 0.01 is used as a non-linear activation function between layers to complete the iterative update of node representations and the effective extraction of abnormal patterns.

[0065] After completing the node-level feature update, the updated node features are aggregated through a graph-level readout mechanism to obtain a system-level global feature vector reflecting the overall operating status of the transformer. A corresponding time series is then constructed based on the global feature vector within a continuous time window. Graph-level readout is implemented using differentiable pooling, with aggregation weights adaptively generated based on the node parameter deviation rate. The node features are then weighted and aggregated according to these weights to obtain the global feature vector. Differentiable pooling enables end-to-end training optimization of the model, and adaptive aggregation weights ensure that the global features highlight information from key components that contribute more to anomalies, avoiding the dilution of core fault features by normal features, while fully preserving the node anomaly contribution information, providing support for subsequent fault localization. Finally, this step inputs the global feature vector time series from the continuous time window into the pre-trained temporal neural network model for temporal inference, outputting probability vectors for each fault type. It can be represented as: (11) In the formula, These are trainable parameters; This is the temporal feature representation vector output by the temporal neural network in the t-th time window; This is the trainable bias vector for the output layer.

[0066] By modeling with temporal neural networks, the evolution patterns of faults can be extracted, and transient disturbances can be distinguished from real faults, thereby improving diagnostic accuracy and anti-interference capabilities. The output probability vector can intuitively quantify the confidence of various faults. By selecting global feature sequences of 20 consecutive time windows and inputting them into a 3-layer gated recurrent unit network to complete temporal modeling, the fault type probability vector is output, enabling accurate identification of fault types.

[0067] S108, select the fault type with the highest probability from the probability vectors of each fault type as the diagnostic result.

[0068] Specifically, selecting the fault type with the highest probability as the diagnostic result allows for the ability of time-series modeling to capture the gradual evolution of faults, distinguishing between transient electromagnetic interference and actual insulation degradation faults, reducing the risk of misjudgment, and improving diagnostic stability. In a specific embodiment, the fault type probability vector output by the time-series neural network is [0.02 (normal), 0.95 (inter-turn insulation aging), 0.02 (partial insulation wear), 0.01 (insulation dampness)]. Based on this, the "inter-turn insulation aging" with the highest probability is selected as the final diagnostic result with a confidence level of 95%. This highly matches the physical degradation characteristics of decreased insulation dielectric constant and increased dielectric loss obtained from online inversion, ensuring the physical interpretability of the diagnostic result.

[0069] After obtaining the diagnosis, the following is also included: Sensitivity analysis is performed on the nodes output by the graph neural network to characterize the contribution of each node to fault determination. A node importance index is constructed based on gradient sensitivity. The calculation formula is as follows: (12) In the formula, For the target fault category The corresponding model output score; For the last layer of the graph neural network The output representation of each node; This formula quantifies the impact of node state changes on fault determination by calculating the absolute value of the partial derivative of the target fault output score with respect to the node characterization, directly representing the contribution of each component node to fault determination. In a specific embodiment, based on the target fault category of inter-turn insulation aging, the importance index of each node is calculated as [0.94 (high voltage winding V1), 0.03 (low voltage winding V2), 0.02 (iron core V3), 0.01 (clamping component V4)], clearly identifying the high voltage winding as the core fault contributing node, thus achieving precise identification of the root cause component of the fault.

[0070] Secondly, a mapping relationship is established between nodes with a contribution greater than a preset value and the digital twin model. The diagnostic results of the corresponding nodes are mapped to the corresponding components in the digital twin model and displayed in a preset display window. This step pre-establishes a one-to-one mapping relationship Ψ between graph structure nodes and the geometric objects of physical components such as high-voltage windings and low-voltage windings in the digital twin 3D model. Core fault nodes are screened through a preset contribution threshold, and low-contribution irrelevant components are filtered out to focus on the core fault area. At the same time, combined with multimodal spatial anchor point information such as temperature sensor location and acoustic array sound source coordinates, a 3D heat distribution is constructed to complete the visualization mapping. For any spatial point r in the 3D model, its heat value is calculated using the following formula: (13) In the formula, For a set of nodes, To represent nodes The One spatial anchor point; This is a diffusion scale parameter used to control the spatial extent of the influence of heat.

[0071] This formula uses a spatial anchor point as the center and Gaussian diffusion to map the importance of nodes to three-dimensional space, achieving unified alignment of multi-source spatial information. It refines positioning accuracy from the component level to the region level, and then normalizes the heat values ​​before mapping them to visual attributes, forming a fault heat map overlaid on the corresponding component's display window in the twin model, providing intuitive guidance for on-site maintenance. In specific implementation, a contribution threshold of 0.1 is preset to select high-voltage winding V1 as the core fault node. The insulation layers of each of the three phases of the high-voltage winding are divided into 6 sub-regions. Combining the spatial anchor point and setting diffusion scale parameters, the normalized heat value of the 5th insulation layer of phase A of the high-voltage winding reaches 0.96. The fault heat map is overlaid in the display window using RGB color mapping (heat value ≥ 0.9 is dark red), simultaneously outputting the fault type, confidence level, and key insulation parameters, providing a direct basis for prioritizing maintenance.

[0072] like Figure 5 As shown, based on the same inventive concept, this embodiment provides a diagnostic system for high-frequency transformer faults, including: The data acquisition module 201 is used to acquire target signals with timestamps during the operation of the transformer. The target signals include partial discharge signals, temperature signals, voltage and current signals, vibration signals, insulation material pyrolysis gas signals, and acoustic array signals. The feature extraction module 202 is used to extract features from the target signal to obtain target features; the target features include partial discharge features, temperature features, vibration features, gas features, and acoustic features; The twin modeling module 203 is used to establish a digital twin model based on the structural parameters, material parameters, and corresponding physical rules of the transformer. In the digital twin model, equivalent operating conditions representing solid insulation degradation and structural degradation are set, and numerical solutions are performed to obtain the local electric field distribution, temperature field distribution, and vibration response data of key components inside the transformer. Based on the local electric field distribution and a preset insulation breakdown criterion, the potential region of partial discharge within the transformer is determined. An equivalent sound source is constructed based on the potential region and the energy evolution of partial discharge. The equivalent sound source is used as the volume sound source term in the acoustic wave equation, and the vibration response data is used as the acoustic boundary excitation to calculate the equivalent acoustic characteristics corresponding to the equivalent operating conditions. An equivalent dynamic model of the insulation decomposition gas is established, and the temperature field distribution is correlated with the partial discharge characteristics to obtain the theoretical concentration characteristics of at least one decomposition gas. The local electric field distribution, temperature field distribution, vibration response, equivalent acoustic characteristics, and theoretical concentration characteristics are integrated to form a multi-physics mechanism sample for characterizing solid insulation degradation. The sample expansion module 204 is used to construct a generative adversarial network. It takes the mechanism sample and the field measured data as inputs, respectively, to perform adversarial training on the generative adversarial network, generate expanded samples that match the distribution of field data, and form an expanded dataset. The online parameter inversion module 205 is used to perform online inversion of the indirectly measurable parameters inside the digital twin model during the operation phase, using measured voltage, current and ambient temperature as inputs, to obtain the optimal parameters of the indirectly measurable parameters under the current time window, and to update the digital twin model in real time using the optimal parameters. The graph construction module 206 is used to calculate the target deviation between the optimal parameters and the corresponding actual parameters. It defines the key components of the transformer as graph nodes, and uses the target features and target deviations as node features. It uses the transmission paths of the preset physical field as connecting edges. Based on the node features and connecting edges, it constructs a feature dynamic graph that reflects the multi-physical coupling relationship. The diagnostic module 207, based on an extended dataset, constructs and annotates a multi-dimensional dynamic graph of side feature samples covering normal and various fault conditions. It then trains an offline graph neural network on the annotated data. Using the dynamic feature graph as input, it calls the trained graph neural network to update the node features in the dynamic feature graph, thus representing abnormal patterns. Through a graph-level readout mechanism, it aggregates the updated node features to obtain a time series of system-level global feature vectors. Finally, it inputs the time series into a pre-trained temporal neural network model for temporal inference, outputting probability vectors for each fault type. The result module 208 is used to select the fault type with the highest probability from the probability vectors of each fault type as the diagnostic result.

[0073] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0074] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for diagnosing faults in a high-frequency transformer, characterized in that, include: The system collects time-stamped target signals during transformer operation, including partial discharge signals, temperature signals, voltage and current signals, vibration signals, insulation material pyrolysis gas signals, and acoustic array signals. Feature extraction is performed on the target signal to obtain target features; the target features include partial discharge features, temperature features, vibration features, gas features, and acoustic features; A digital twin model is established based on the structural parameters, material parameters, and corresponding physical rules of the transformer. Equivalent operating conditions representing solid insulation degradation and structural degradation are set in the digital twin model, and numerical solutions are performed to obtain the local electric field distribution, temperature field distribution, and vibration response data of key components inside the transformer. Based on the local electric field distribution and a preset insulation breakdown criterion, the potential region of partial discharge within the transformer is determined. An equivalent sound source is constructed based on the potential region and the energy evolution of partial discharge. The equivalent sound source is used as the volume sound source term in the acoustic wave equation, and the vibration response data is used as the acoustic boundary excitation to calculate the equivalent acoustic characteristics corresponding to the equivalent operating condition. An equivalent dynamic model of the insulation pyrolysis gas is established, and the temperature field distribution is correlated with the partial discharge characteristics to obtain the theoretical concentration characteristics of at least one pyrolysis gas. The local electric field distribution, temperature field distribution, vibration response, equivalent acoustic characteristics and theoretical concentration characteristics are integrated to form a multi-physics mechanism sample for characterizing solid insulation degradation. A generative adversarial network is constructed, and the mechanism samples and field measured data are used as inputs respectively. The generative adversarial network is trained adversarially to generate extended samples that match the distribution of field data, forming an extended dataset. During the operation phase, the measured voltage, current and ambient temperature are used as inputs to perform online inversion on the indirect measurable parameters inside the digital twin model, to obtain the optimal parameters of the indirect measurable parameters under the current time window, and the digital twin model is updated in real time using the optimal parameters. Calculate the target deviation between the optimal parameters and the corresponding actual parameters, define the key components of the transformer as graph nodes, and use the target features and the target deviation as node features; Use the transmission path of the preset physical field as the connecting edge; Based on the node features and the connecting edges, a feature dynamic graph reflecting multi-physical coupling relationships is constructed; Based on the extended dataset, a multi-dimensional side feature sample dynamic graph covering normal and various fault conditions is constructed and labeled. The labeled data is then used for offline training of a graph neural network. Using the feature dynamic graph as input, the trained graph neural network is invoked to update the node features in the feature dynamic graph, thus completing the abnormal mode representation. The updated node features are aggregated through a graph-level readout mechanism to obtain a time series of system-level global feature vectors. The time series is then input into a pre-trained temporal neural network model for temporal inference, outputting probability vectors for each fault type. The fault type with the highest probability among the probability vectors of each fault type is selected as the diagnostic result.

2. The method for diagnosing high-frequency transformer faults according to claim 1, characterized in that, The step of acquiring the time-stamped target signal during the operation of the transformer specifically includes: A unified time reference signal is used to synchronize the monitoring data of each channel and to mark the monitoring data of each channel with a timestamp. Using sudden changes in partial discharge amplitude and / or sudden changes in vibration energy as triggering conditions, monitoring data of preset durations are extracted before and after the triggering time to form an event time window; The monitoring data of signals with frequencies lower than the preset frequency in the target signal are mapped to the unified time axis of the event time window through interpolation and / or resampling, and signals with frequencies higher than the preset frequency are aligned in time.

3. The method for diagnosing high-frequency transformer faults according to claim 1, characterized in that, The step of extracting features from the target signal specifically includes: The partial discharge signal is subjected to multi-layer wavelet decomposition to extract the maximum discharge amount, average number of discharges and discharge phase distribution within a preset time window; Infrared thermal imaging is obtained through the temperature signal. An adaptive threshold segmentation method is used on the infrared thermal imaging to extract high-temperature regions that exceed the temperature threshold, and the highest temperature and temperature rise gradient of the high-temperature regions are extracted. Perform a fast Fourier transform on the vibration signal to extract the fundamental frequency and harmonic energy ratio; The concentration of the pyrolysis gas in the insulating material was normalized and the gas concentration and rate of change were extracted. A beamforming algorithm is used to generate a sound pressure distribution map for the acoustic array signal, and the coordinates of the sound source center and the sound pressure energy characteristics are extracted.

4. The method for diagnosing high-frequency transformer faults according to claim 1, characterized in that, Constructing an equivalent sound source based on the potential region and the energy evolution of partial discharge specifically includes: Within the potential region, a time-varying function of partial discharge energy is constructed by combining an equivalent gap capacitance model with a preset partial discharge current or charge waveform. Based on the time-varying function, the partial discharge process within the potential region is equivalent to a volume sound source term in the acoustic wave equation, thus completing the construction of the equivalent sound source.

5. The method for diagnosing high-frequency transformer faults according to claim 1, characterized in that, The steps for adversarial training of the generative adversarial network specifically include: the generative adversarial network adopts a Wasserstein generative adversarial network with gradient penalty, the generative adversarial network includes a generator subnetwork and a discriminator subnetwork, a gradient penalty term is set at the output of the discriminator subnetwork, during training, the mechanism sample is input into the generator subnetwork to generate samples, and then the generated samples and the field measured data are input into the discriminator subnetwork for discrimination, the gradient penalty term constrains the gradient norm of the discriminator function in the discriminator subnetwork.

6. The method for diagnosing high-frequency transformer faults according to claim 1, characterized in that, The step of performing online inversion on the indirectly measurable parameters within the digital twin model to obtain the optimal parameters for the indirectly measurable parameters under the current time window specifically includes: Indirectly measurable parameters that are highly correlated with the insulation state of the transformer are selected as the parameter vector to be inverted. The parameters to be inverted include thermal resistance, equivalent capacitance, mechanical stiffness, equivalent dielectric constant, dielectric loss, and equivalent breakdown field strength. The parameter vector is iteratively updated using a particle swarm optimization algorithm. The goal of the iterative optimization is to minimize the preset error index. By iteratively solving for particle velocity and position, the optimal parameters corresponding to the indirect measurable parameters under the current time window are obtained.

7. The method for diagnosing high-frequency transformer faults according to claim 6, characterized in that, The steps for offline training of the graph neural network specifically include: During the training of the graph neural network, a physical consistency constraint loss term is constructed. The constraint loss term is determined based on the change in the parameter vector obtained by inversion within adjacent time windows. It is used to penalize the graph neural network when the change exceeds a preset threshold, so as to limit the rate of change of the parameter vector and make the fault diagnosis result consistent with the actual thermo-mechanical evolution process of the transformer.

8. The method for diagnosing high-frequency transformer faults according to claim 1, characterized in that, The steps for aggregating the updated node features using a graph-level readout mechanism specifically include: Graph-level readout is implemented using differentiable pooling. Based on the node parameter deviation rate, corresponding aggregation weights are adaptively generated. The updated node features are then weighted and aggregated according to the aggregation weights to obtain a system-level global feature vector.

9. The method for diagnosing high-frequency transformer faults according to claim 1, characterized in that, After selecting the fault type with the highest probability from the probability vectors of each fault type as the diagnostic result, the process further includes: Sensitivity analysis is performed on the nodes output by the graph neural network to characterize the contribution of each node to fault determination. Establish a mapping relationship between nodes with a contribution greater than a preset value and the digital twin model, map the diagnostic results of the corresponding nodes to the corresponding components of the digital twin model, and display them in a preset display window.

10. A diagnostic system for high-frequency transformer faults, based on the diagnostic method for high-frequency transformer faults according to claim 1, characterized in that, include: The data acquisition module is used to acquire target signals with timestamps during the operation of the transformer. The target signals include partial discharge signals, temperature signals, voltage and current signals, vibration signals, insulation material pyrolysis gas signals, and acoustic array signals. The feature extraction module is used to extract features from the target signal to obtain target features; the target features include partial discharge features, temperature features, vibration features, gas features, and acoustic features; The twin modeling module is used to establish a digital twin model based on the structural parameters, material parameters, and corresponding physical rules of the transformer. In the digital twin model, equivalent operating conditions representing solid insulation degradation and structural degradation are set, and numerical solutions are performed to obtain the local electric field distribution, temperature field distribution, and vibration response data of key components inside the transformer. Based on the local electric field distribution and a preset insulation breakdown criterion, the potential region of partial discharge within the transformer is determined. An equivalent sound source is constructed based on the potential region and the energy evolution of partial discharge. The equivalent sound source is used as the volume sound source term in the acoustic wave equation, and the vibration response data is used as the acoustic boundary excitation to calculate the equivalent acoustic characteristics corresponding to the equivalent operating conditions. An equivalent dynamic model of the insulation pyrolysis gas is established, and the temperature field distribution is correlated with the partial discharge characteristics to obtain the theoretical concentration characteristics of at least one pyrolysis gas. The local electric field distribution, temperature field distribution, vibration response, equivalent acoustic characteristics and theoretical concentration characteristics are integrated to form a multi-physics mechanism sample for characterizing solid insulation degradation. The sample expansion module is used to construct a generative adversarial network. It takes the mechanism sample and the field measured data as inputs, respectively, and performs adversarial training on the generative adversarial network to generate expanded samples that match the distribution of field data, thus forming an expanded dataset. The online parameter inversion module is used to perform online inversion of the indirectly measurable parameters inside the digital twin model during the operation phase, using measured voltage, current and ambient temperature as inputs, to obtain the optimal parameters of the indirectly measurable parameters under the current time window, and to update the digital twin model in real time using the optimal parameters. The graph construction module is used to calculate the target deviation between the optimal parameters and the corresponding actual parameters, define key components of the transformer as graph nodes, and use the target features and the target deviation as node features. Use the transmission path of the preset physical field as the connecting edge; Based on the node features and the connecting edges, a feature dynamic graph reflecting multi-physical coupling relationships is constructed; The diagnostic module, based on the extended dataset, constructs and labels a multi-dimensional dynamic graph of side feature samples covering normal and various fault conditions. It then trains an offline graph neural network on the labeled data. Using the dynamic feature graph as input, it calls the trained graph neural network to update the node features in the dynamic feature graph, thus completing the abnormal pattern representation. Through a graph-level readout mechanism, it aggregates the updated node features to obtain a time series of system-level global feature vectors. Finally, it inputs the time series into a pre-trained temporal neural network model for temporal inference, outputting probability vectors for each fault type. The results module is used to select the fault type with the highest probability from the probability vectors of each fault type as the diagnostic result.

Images