A fault recognition method based on acoustic amplitude characteristic change

Abstract

This application provides a fault identification method based on acoustic amplitude feature changes, belonging to the field of acoustic fault identification technology. The method includes: during transformer operation, acquiring high-frequency vibration signals and partial discharge current signals using ultrasonic sensors and high-frequency current sensors, and obtaining multimodal raw samples after preprocessing; calling a digital twin sample generation module to generate multimodal virtual samples, and fusing them to obtain an enhanced training dataset; constructing a fault identification model based on transfer learning, and obtaining a transformer fault identifyr through fine-tuning training; inputting the real-time acquired acoustic amplitude sequence and electrical pulse sequence into the transformer fault identifyr, and outputting preliminary identification results of fault type and aging level; quantizing and calibrating the preliminary identification results, and outputting the final fault identification result. This method solves the technical problems of low accuracy and weak generalization ability in identifying early, weak faults in existing transformer fault identification methods based on acoustic amplitude feature changes.

Description

Technical Field

[0001] This invention relates to the field of acoustic fault identification, and more particularly to a fault identification method based on changes in acoustic amplitude characteristics. Background Technology

[0002] As a core component of the power system, the operational reliability of transformers directly determines the stability of the power grid. Early partial discharge faults caused by insulation aging are a primary cause of unplanned transformer outages. Acoustic amplitude-based fault detection technology, due to its ease of operation and non-invasiveness, has become a commonly used method for identifying early partial discharge faults in transformers.

[0003] However, existing fault identification methods that rely on acoustic amplitude features have significant shortcomings: the acoustic signal amplitude generated by early partial discharge is extremely low and is easily masked by environmental noise such as transformer cooling fans, oil pump vibration, and power grid electromagnetic radiation. This makes it difficult to collect effective samples on-site and the cost of sample labeling is high. As a result, the fault identification model may overfit due to insufficient effective training data, leading to low accuracy and weak generalization ability in identifying early and weak faults, which makes it difficult to meet the power grid's demand for early warning of transformer faults.

[0004] Therefore, there is an urgent need for a fault identification method based on acoustic amplitude feature changes that can overcome the bottleneck of small sample size and improve the reliability of early fault identification. Summary of the Invention

[0005] This invention addresses the technical problems of low accuracy and weak generalization ability in identifying early and subtle faults in existing transformer fault identification methods based on changes in acoustic amplitude characteristics. It provides a fault identification method based on changes in acoustic amplitude characteristics.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: This invention provides a fault identification method based on acoustic amplitude characteristic changes, comprising: During transformer operation, high-frequency vibration signals of the tank wall are collected by an ultrasonic sensor, and partial discharge current signals are collected synchronously by a high-frequency current sensor. After preprocessing the two types of signals, multimodal raw samples containing acoustic amplitude sequences and electrical pulse sequences are obtained. The digital twin sample generation module is invoked, transformer insulation-related parameters are input, and multimodal virtual samples with the same feature distribution as the original multimodal samples are generated. The original multimodal samples and the multimodal virtual samples are fused together as an enhanced training dataset. The transformer insulation-related parameters include the degree of polymerization of insulating paper, air gap size, and operating temperature. A fault identification model is constructed based on transfer learning, and the fault identification model is fine-tuned and trained using the enhanced training dataset to obtain a transformer fault identifyer. The acoustic amplitude sequence and electrical pulse sequence, which are collected and preprocessed in real time, are input into the transformer fault identifier, and the preliminary identification results of fault type and aging level are output. The preliminary identification results are quantified and calibrated to output the final fault identification results.

[0007] The beneficial effects of this invention are: Compared to existing technologies, this application firstly collects high-frequency vibration signals from the tank wall using an ultrasonic sensor and simultaneously collects partial discharge current signals using a high-frequency current sensor during transformer operation. After preprocessing these two types of signals, multimodal raw samples containing acoustic amplitude sequences and electrical pulse sequences are obtained, providing comprehensive basic sample support for subsequent fault type identification and insulation aging level determination. Secondly, a digital twin sample generation module is invoked, inputting transformer insulation-related parameters to generate multimodal virtual samples with feature distributions consistent with the multimodal raw samples. The multimodal raw samples and multimodal virtual samples are fused as an enhanced training dataset. The digital twin simulates the coupling process of multiple physical fields (electric, acoustic, and chemical), providing sufficient and high-quality training data for subsequent transfer learning models. Thirdly, a fault identification model is constructed based on transfer learning. The enhanced training dataset is used to fine-tune the fault identification model, resulting in a transformer fault identifyer. This solves the problem of insufficient feature learning under small sample sizes, allowing the transformer fault identifyer to adapt to specific transformer fault patterns and achieve high accuracy and high generalization in fault identification. Furthermore, the real-time acquired and preprocessed acoustic amplitude sequence and electrical pulse sequence are input into the transformer fault identifier, which outputs preliminary identification results of fault type and aging level, providing a data foundation for subsequent accurate diagnosis by combining data such as dissolved gases in the oil. Finally, the preliminary identification results are quantitatively calibrated to output the final fault identification result, improving the reliability and accuracy of single fault identification. Long-term adaptability is enhanced through model iteration, ultimately achieving high-precision and robust fault identification for the transformer fault identifier.

[0008] Through the above technical solutions, this application uses ultrasonic sensors and high-frequency current sensors to collect multimodal data, overcoming the deficiency of single acoustic signals being easily masked by environmental noise. A large number of virtual samples with matching features are generated using a digital twin sample generation module, overcoming the dilemma of scarce effective samples and high annotation costs associated with small samples in early faults, providing sufficient data support for model training. Based on transfer learning, the application reuses general high-frequency vibration feature learning results to reduce dependence on transformer-specific samples. Furthermore, fine-tuning the training to adapt to specific transformer fault patterns improves the generalization ability of the transformer fault identifyr. Finally, quantization calibration further corrects errors. In this way, accurate identification of early partial discharge fault types and aging levels in transformers is achieved, improving the scenario adaptability of early fault identification and meeting the power grid's needs for early warning of transformer insulation status and ensuring the reliable operation of core hub equipment. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a fault identification method based on acoustic amplitude characteristic changes provided by the present invention; Figure 2 This is a flowchart illustrating the fine-tuning training process of a transformer fault identifier in a fault identification method based on acoustic amplitude characteristic changes provided by the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0012] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0013] Examples, such as Figure 1 As shown, this embodiment of the invention provides a fault identification method based on changes in acoustic amplitude characteristics, including: S10: During the operation of the transformer, high-frequency vibration signals of the tank wall are collected by an ultrasonic sensor, and partial discharge current signals are collected synchronously by a high-frequency current sensor. After preprocessing the two types of signals, multimodal raw samples containing acoustic amplitude sequences and electrical pulse sequences are obtained.

[0014] Partial discharge is one of the core manifestations of transformer insulation aging. The tiny shock waves generated during discharge are transmitted through the transformer oil to the tank wall, triggering high-frequency vibrations in a specific frequency band. The characteristics of these high-frequency vibrations can reflect the intensity of the discharge energy. However, acoustic signals are affected by the propagation path, such as uneven tank wall thickness or air bubbles in the oil, which may cause characteristic attenuation or distortion, making it difficult to fully and accurately reflect the nature of the discharge. The essence of partial discharge is the rapid transfer of charge within the insulation gap, which forms a weak pulse current in the transformer grounding circuit (such as the neutral point grounding line). The acquisition of this type of electrical pulse signal is not affected by mechanical noise and has stronger anti-interference capabilities, effectively compensating for the shortcomings of acoustic characteristics.

[0015] To address the aforementioned issues, this application acquires high-frequency vibration signals from the tank wall using an ultrasonic sensor and simultaneously acquires partial discharge current signals using a high-frequency current sensor during transformer operation. After preprocessing the two types of signals, a multimodal raw sample containing acoustic amplitude sequences and electrical pulse sequences is obtained.

[0016] For example, during data acquisition, a high-frequency response ultrasonic sensor with a center frequency matching the frequency band should be selected. The ultrasonic sensor should be installed magnetically in the area corresponding to the winding on the tank wall, ensuring the original vibration characteristics are preserved to the greatest extent possible in the winding area. Simultaneously, a wideband high-frequency current sensor with a bandwidth ≥1MHz and a response time <10ns can be selected and mounted on the neutral-point grounding line to capture pulse signals.

[0017] Furthermore, step S10 of the method also includes: The high-frequency vibration signal acquired by the ultrasonic sensor is filtered to extract the peak amplitude, kurtosis, and dominant frequency characteristics of the signal, and then arranged in chronological order to form an acoustic amplitude sequence. Baseline correction is performed on the partial discharge current signal acquired by the high-frequency current sensor, and the pulse peak value, rise time, and pulse interval characteristics of the signal are extracted and arranged in chronological order to form an electrical pulse sequence. The acoustic amplitude sequence and electrical pulse sequence are aligned by timestamp, and combined with the dissolved gas detection data in transformer oil, the corresponding fault type label and aging level label are marked to form a multimodal original sample.

[0018] In this embodiment, the high-frequency vibration signal of the oil tank wall collected by the ultrasonic sensor is mixed with interference signals such as low-frequency noise from the cooling fan. These interference signals will mask the weak vibration characteristics generated by partial discharge in the early stage of transformer insulation aging. Therefore, it is necessary to filter the high-frequency vibration signal collected by the ultrasonic sensor, extract the peak amplitude, kurtosis and main frequency characteristics of the signal, and organize them into an acoustic amplitude sequence in chronological order.

[0019] Among these, the peak amplitude, kurtosis, and dominant frequency characteristics of the signal are extracted because they can directly or indirectly characterize the partial discharge fault attributes caused by transformer insulation aging, and are strongly correlated with the severity of the fault, the authenticity of the signal, and the type of fault. Peak amplitude is a physical quantity that reflects the maximum intensity of the vibration signal. The essence of transformer insulation aging is the deterioration of the insulating medium (such as insulating paper, oil gaps, etc.) leading to partial discharge. The more severe the aging, the higher the partial discharge energy, and the greater the vibration intensity generated by the impact on the surrounding medium during the discharge process. The corresponding peak amplitude also increases. Therefore, the severity of aging can be indirectly judged by the peak amplitude. Kurtosis is a statistical measure of the sharpness of the signal pulse. During normal operation, the environmental noise of the transformer (such as fan vibration and core noise) As a smooth random signal, the kurtosis is approximately 3. However, partial discharge is a short-duration, high-energy pulsed discharge, which causes the vibration signal to exhibit obvious sharp pulse characteristics, with the corresponding kurtosis increasing to 5-15 and positively correlated with the discharge energy. Therefore, by using the difference in kurtosis values, fault pulses can be effectively distinguished from random environmental noise, avoiding misjudging noise as a fault signal. The dominant frequency is the frequency component where the vibration signal energy is most concentrated. Different types of partial discharges produce significantly different vibration frequency distributions due to different discharge mechanisms. For example, the vibration energy of air gap discharge is mostly concentrated in 100-120kHz, while the vibration energy of tip discharge is concentrated in 120-150kHz. Therefore, by identifying the range of the dominant frequency, the specific fault type can be determined.

[0020] For example, for high-frequency vibration signals, a 20-200kHz bandpass filter can be used to filter out low-frequency mechanical noise (<20kHz) such as cooling fans and oil pumps, as well as external electromagnetic interference (>200kHz), retaining only high-frequency signals related to partial discharge. Simultaneously, a 50Hz notch filter is superimposed to eliminate power grid frequency and harmonic interference, achieving signal purification. Then, based on the purified signal, the maximum value within a fixed time window is selected in the time domain as the peak amplitude. In the statistical domain, the ratio of the fourth-order central moment to the square of the variance is used as the kurtosis. In the frequency domain, a fast Fourier transform is used to identify the dominant frequency. Finally, the peak amplitude, kurtosis, and dominant frequency characteristics at each time point are organized into a structured acoustic amplitude sequence according to the timestamp order, for example, {(t1: 50μPa, 8, 110kHz), (t2: 55μPa, 8.5, 112kHz), (t3: 62μPa, 9, 115kHz), ...}. In this way, the model can capture the dynamic evolution of vibration intensity gradually increasing, pulse characteristics becoming more prominent, and main frequency steadily shifting during the fault development process, rather than relying solely on isolated static feature values, thereby improving the sensitivity of early fault identification.

[0021] Secondly, the partial discharge current signal collected by the high-frequency current sensor may be offset due to sensor zero drift and electromagnetic interference of the line, which may cause the real pulse signal to be masked or misjudged. Therefore, it is necessary to perform baseline correction on the partial discharge current signal collected by the high-frequency current sensor, extract the pulse peak value, rise time and pulse interval characteristics of the signal, and organize them into an electrical pulse sequence in chronological order.

[0022] Among these, the pulse peak value, rise time, and pulse interval characteristics of the signal are extracted because they can describe the characteristics of partial discharge from three dimensions: energy intensity, signal authenticity, and development law. The pulse peak value can reflect the energy intensity of partial discharge. Partial discharge is essentially a charge transfer process in the insulation gap. The higher the discharge energy, the larger the current pulse peak value generated. Moreover, the pulse peak value and peak amplitude are significantly positively correlated. The two can jointly reflect the degree of insulation aging. The rise time can distinguish between fault signals and interference. Fault discharge (such as air gap breakdown and tip discharge) is an instantaneous breakdown process of the insulating medium. The rise time of the current pulse from the baseline to the peak value is extremely short, usually <5μs. In contrast, environmental interference is mostly a slowly changing signal, and the rise time is generally >5μs. The pulse interval can reflect the frequency law of partial discharge. In the early stage of insulation aging, partial discharge is limited by insulation strength and occurs at a low frequency, with a longer pulse interval. As aging intensifies, discharge conditions are more easily met, and the interval gradually shortens. The dynamic change of the pulse interval can intuitively reflect the development trend of the fault and help the model judge the evolution stage of the aging level.

[0023] For example, for partial discharge current signals, the baseline value can be subtracted from the original signal using a sliding window averaging method or polynomial fitting, so that the pulse signal presents a true positive and negative symmetrical shape centered on the zero axis. Subsequently, in the time domain, the peak value of each complete pulse is identified as the pulse peak value, the time difference between the pulse rising from the 10% peak value to the 90% peak value is calculated as the rise time, and the time interval between two adjacent effective pulses (amplitude exceeding 3 times the noise threshold) is counted as the pulse interval. Finally, the pulse peak value, rise time, and pulse interval characteristics at each time point are organized into a structured electrical pulse sequence according to the timestamp order, for example, {(t1 time: 30pC, 0.4μs, 25ms), (t2 time: 35pC, 0.35μs, 22ms), (t3 time: 42pC, 0.3μs, 18ms), ...}. In this way, the electrical pulse sequence can be precisely aligned with the acoustic amplitude sequence in the same time dimension, forming an acoustic-electric feature pair at the same fault moment. This provides reliable data support for subsequent model learning of the correlation rules of cross-modal features and effectively improves the robustness of fault identification.

[0024] Finally, since the acoustic amplitude sequence and the electrical pulse sequence both originate from the same fault process, but there is a slight time difference in sensor acquisition, the acoustic amplitude sequence and the electrical pulse sequence at the same timestamp can be aligned using a GPS synchronization clock (error ≤ 1ms) to ensure that they correspond to the same fault state.

[0025] Meanwhile, during partial discharge, high-energy electrons bombard transformer oil molecules, causing their chemical bonds to break and decompose to produce characteristic gases such as H2, CH4, and C2H2. These gases gradually dissolve in the oil, and their concentration is positively correlated with the discharge energy and duration, exhibiting a cumulative effect: in the early insulation aging stage, the partial discharge energy is low and the duration is short, so the concentration of dissolved gases in the transformer oil is usually low. As aging intensifies, the concentration of dissolved gases in the transformer oil continues to increase. This cumulative characteristic complements the acoustic amplitude sequence and electrical pulse sequence. Therefore, the data on dissolved gases in transformer oil can be used as a third-party verification basis, thereby improving the accuracy of fault identification.

[0026] For example, by combining acoustic amplitude sequences, electrical pulse sequences, and dissolved gas detection data in transformer oil, corresponding fault type labels and aging level labels are manually marked. For instance, "main frequency 100-120kHz, rise time 0.4μs, H2 concentration 120μL / L, peak amplitude 200μPa, pulse peak 80pC" is labeled as "air gap discharge, level two aging", and "main frequency 130kHz, rise time 0.2μs, H2 concentration 40μL / L, peak amplitude 80μPa, pulse peak 40pC" is labeled as "point discharge, level one aging". Finally, a multimodal original sample containing acoustic, electrical, and gas temporal features and corresponding fault type and aging level labels is generated. This not only ensures the physical correlation between features and labels, but also provides a real feature distribution template for the digital twin module to generate high-fidelity virtual samples, ensuring that the virtual samples are consistent with the real fault patterns.

[0027] In summary, compared to existing technologies, this application, during transformer operation, acquires high-frequency vibration signals from the tank wall using an ultrasonic sensor and simultaneously acquires partial discharge current signals using a high-frequency current sensor. After preprocessing these two types of signals, multimodal raw samples containing acoustic amplitude sequences and electrical pulse sequences are obtained. In this way, acoustic and electrical dual-modal raw data reflecting the partial discharge state of the transformer are simultaneously acquired, providing comprehensive basic sample support for subsequent fault type identification and insulation aging level determination.

[0028] S20: Call the digital twin sample generation module, input transformer insulation-related parameters, generate multimodal virtual samples with the same feature distribution as the original multimodal samples, and fuse the original multimodal samples with the multimodal virtual samples as an enhanced training dataset. The transformer insulation-related parameters include the degree of polymerization of insulating paper, air gap size, and operating temperature.

[0029] Early fault signals caused by insulation aging in transformers have extremely low amplitudes and are easily masked by environmental noise. It is difficult to collect effective samples on-site and the labeling cost is high. Traditional fault identification models often rely on supervised learning modes with a large number of labeled samples. When the number of training samples is insufficient, the model cannot fully learn the fault characteristics of different insulation aging stages, and is prone to overfitting and insufficient generalization ability.

[0030] To address the aforementioned issues, this application calls a digital twin sample generation module, inputs transformer insulation-related parameters, generates multimodal virtual samples with the same feature distribution as the original multimodal samples, and fuses the original multimodal samples with the multimodal virtual samples as an enhanced training dataset.

[0031] Specifically, step S20 in the method includes: Based on the three-dimensional structural model of the transformer, a digital twin sample generation module is constructed, which includes electrical simulation unit, acoustic simulation unit, and chemical simulation unit. Input the degree of polymerization of the transformer's insulating paper, the air gap size, and the operating temperature into the digital twin sample generation module; The electrical simulation unit generates a virtual current sequence. The discharge energy is calculated using the acoustic simulation unit to generate a virtual acoustic amplitude sequence. The chemical simulation unit synchronously generates virtual dissolved gas concentration data in oil. By integrating the virtual current sequence, virtual acoustic amplitude sequence, and virtual dissolved gas concentration data in oil, a multimodal virtual sample is formed that is consistent with the feature distribution of the original multimodal sample. The original multimodal samples and the virtual multimodal samples are fused together at a preset ratio to form an enhanced training dataset.

[0032] In this embodiment, firstly, based on actual CAD drawings or laser scanning data of the transformer, the three-dimensional structural details such as the transformer winding arrangement, core size, tank volume, and oil passage width are reconstructed. Then, based on the multi-physics coupling characteristics of partial discharge (electrical, acoustic, and chemical), a digital twin sample generation module is constructed, comprising an electrical simulation unit, an acoustic simulation unit, and a chemical simulation unit. The electrical simulation unit corresponds to the current signal of the real sample, the acoustic simulation unit corresponds to the vibration signal, and the chemical simulation unit corresponds to the dissolved gas data in the oil. The discharge energy output by the electrical simulation unit serves as the input to the acoustic and chemical simulation units, ensuring that the correlation of multi-modal features in the virtual sample is consistent with the real fault patterns, avoiding feature disconnect caused by single-physics simulation. It should be noted that the digital twin technology itself and the basic construction framework of the digital twin sample generation module belong to the prior art, and the relevant technical details have been fully disclosed in existing published documents or patents. Therefore, this application will not elaborate on the basic construction principle and general technical details of this module.

[0033] Secondly, the transformer's insulation paper polymerization degree, air gap size, and operating temperature are input into the digital twin sample generation module. These three parameters are key physical boundary conditions for partial discharge characteristics. The insulation paper polymerization degree (DP value) reflects the degree of insulation aging; a decrease in DP value leads to a decrease in the insulation breakdown field strength, directly affecting the discharge initiation voltage. The air gap size determines the discharge energy intensity; the thinner the air gap, the lower the discharge energy. Operating temperature accelerates insulation degradation and gas diffusion; as the temperature rises, the transformer oil breakdown field strength decreases, and the gas diffusion coefficient increases. For example, the transformer's insulation paper polymerization degree, air gap size, and operating temperature can be obtained through laboratory testing and online monitoring. For instance, the insulation paper polymerization degree can be detected by gel permeation chromatography, the air gap size can be obtained through ultrasonic flaw detection of the tank wall or partial discharge localization technology, and the operating temperature can be collected in real time by an online platinum resistance sensor on the top of the tank. These three parameters cover the entire physical process of partial discharge, from its occurrence conditions, energy intensity, and evolution law, ensuring that the virtual sample in the digital twin simulation closely matches the physical mechanism of real faults, rather than being abstract data detached from reality.

[0034] Next, a virtual current sequence is generated through the electrical simulation unit. For example, the electrical simulation unit conducts simulations based on Paschen's law and the RLC equivalent circuit model: first, the discharge initiation voltage is calculated based on the input insulation paper polymerization degree and air gap size to determine whether discharge has occurred; then, the charge transfer process in the air gap during discharge is simulated, and the time-domain waveform of the ionization current is calculated; finally, the pulse peak value, rise time, and pulse interval features consistent with the real sample are extracted from the waveform and organized into a virtual current sequence according to the timestamp. By controlling the statistical characteristic deviation between the virtual current and the real current to ≤10%, it can be ensured that it can replace scarce real electrical samples for model training.

[0035] Furthermore, the discharge energy is calculated using an acoustic simulation unit to generate a virtual acoustic amplitude sequence. For example, the acoustic simulation unit starts with the discharge energy output from the electrical simulation unit, converts the energy into shock wave pressure using a fluid dynamics algorithm, and then simulates the propagation process of the pressure in the transformer oil. Considering the attenuation effect of the oil's viscosity and density on the wave, when the shock wave reaches the tank wall, finite element analysis is used to calculate the vibration response of the tank wall, obtaining the vibration frequency and amplitude. Finally, the peak amplitude, kurtosis, and dominant frequency characteristics are extracted, and a virtual acoustic amplitude sequence is formed according to timestamps, ensuring that its vibration pattern completely matches the vibration caused by real partial discharge.

[0036] Furthermore, the chemical simulation unit synchronously generates virtual dissolved gas concentration data in the oil. For example, the chemical simulation unit conducts simulations based on chemical reaction kinetics and diffusion equations. First, it calculates the decomposition rate of transformer oil molecules based on the discharge energy, simulating the process of alkanes and alkenes decomposing into characteristic gases such as H2, CH4, and C2H2. Then, combined with the input operating temperature, it calculates the dissolution and diffusion patterns of gases in the oil using diffusion equations, ultimately outputting gas concentration data that changes over time, thus obtaining virtual dissolved gas concentration data in the oil. This ensures that the changes in gas composition and concentration conform to the chemical characteristics of real faults.

[0037] Furthermore, the virtual current sequence, virtual acoustic amplitude sequence, and virtual dissolved gas concentration data in the oil at the same time are bound together to form a multimodal virtual sample that is consistent with the feature distribution of the original multimodal sample.

[0038] Finally, the original multimodal samples and the virtual multimodal samples are fused at a preset ratio to form an enhanced training dataset. This enhanced training dataset retains the physical authenticity of the original multimodal samples while covering fault scenarios under different combinations of insulation parameters through the virtual multimodal samples. This addresses the overfitting problem caused by sample scarcity in traditional models and provides data support for high-accuracy training of subsequent fault identification models. The preset ratio can be dynamically set according to the sample size of the original multimodal samples. This application recommends fusing the original multimodal samples and the virtual multimodal samples at a preset ratio of 1:3 to 1:5. If the proportion of the original multimodal samples is too low (e.g., 1:10), the model will be affected by simulation bias; if the proportion of the original multimodal samples is too high (e.g., 1:1), it will not effectively supplement the sample size.

[0039] In summary, compared to existing technologies, this application utilizes a digital twin sample generation module, inputs transformer insulation-related parameters, and generates multimodal virtual samples with a feature distribution consistent with the original multimodal samples. The original multimodal samples and the virtual multimodal samples are then fused together to form an enhanced training dataset. In this way, by simulating the coupling process of multiple physical fields (electricity, sound, and chemistry) through digital twins, sufficient and high-quality training data is provided for subsequent transfer learning models.

[0040] S30: Construct a fault identification model based on transfer learning, and fine-tune the fault identification model using the enhanced training dataset to obtain a transformer fault identifyr.

[0041] Traditional fault identification models face significant small sample problems in transformer fault identification scenarios, while transfer learning can solve the small sample problem and improve the model's generalization ability through the logic of general knowledge reuse and scenario fine-tuning.

[0042] To address the aforementioned issues, this application constructs a fault identification model based on transfer learning, and fine-tunes the fault identification model using the enhanced training dataset to obtain a transformer fault identifyer.

[0043] Specifically, step S30 in the method includes: Acquire general acoustic data of high-frequency mechanical vibration as a sample general acoustic dataset; A one-dimensional convolutional feature extraction network containing convolutional layers and pooling layers is constructed. The network is trained using the sample general acoustic dataset as input until it converges. Freeze the parameters of the convolutional and pooling layers of the one-dimensional convolutional feature extraction network to obtain a transferable shared feature extractor, wherein the shared feature extractor is used to extract general time-frequency features of acoustic amplitude sequences; The shared feature extractor is connected to construct an acoustic branch, which takes the acoustic amplitude sequence as input and outputs an acoustic feature vector. An LSTM network is used to construct the electrical branch, which takes an electrical pulse sequence as input and outputs an electrical feature vector. A fusion classification layer is constructed to weight and fuse the acoustic feature vector and the electrical feature vector, and output a joint feature vector. A classification output layer is connected after the fusion classification layer. Taking the joint features as input, the fault type and aging level are output to obtain the fault identification model.

[0044] In this embodiment, high-frequency mechanical vibration general acoustic data is first acquired as a sample general acoustic dataset. For example, high-frequency mechanical vibration general acoustic data, such as vibration signals from motors, gearboxes, and other equipment in the 50-200kHz high-frequency band, can be selected to form a sample general acoustic dataset with a scale of tens of thousands of sets. The sample general acoustic data shares common time-frequency characteristics with the high-frequency vibration signals of transformer partial discharge. For instance, the vibration signals of early wear of motor bearings (dominant frequency 80-120kHz) and the vibration signals of transformer air gap discharge (dominant frequency 100-120kHz) are similar in pulse amplitude variation trends and kurtosis distribution, providing sufficient general feature learning samples for subsequent pre-training.

[0045] Secondly, a one-dimensional convolutional feature extraction network containing convolutional and pooling layers is constructed. For example, 3-5 convolutional layers with kernel sizes of 3-7 and pooling layers with a stride of 2 are constructed. Convolutional layers are good at capturing local features in time-series data, while pooling layers reduce parameters by dimensionality reduction to avoid overfitting. Then, using a sample general acoustic dataset as input, the parameters of the one-dimensional convolutional feature extraction network, such as kernel weights and biases, are optimized through backpropagation until the feature reconstruction error on the sample general acoustic dataset decreases by <0.001 for 10 consecutive iterations, which is considered convergence. The one-dimensional convolutional feature extraction network can extract general time-frequency features from the sample general acoustic dataset, such as pulse sharpness and frequency distribution patterns. These features can be transferred to the processing of transformer acoustic signals.

[0046] Furthermore, the one-dimensional convolutional feature extraction network trained on a general acoustic dataset can identify basic features such as pulse morphology and frequency changes of high-frequency vibrations. These features are universal for transformer acoustic amplitude sequences. Therefore, by freezing the parameters of the convolutional and pooling layers of the one-dimensional convolutional feature extraction network, a transferable shared feature extractor is obtained. The shared feature extractor is used to extract general time-frequency features of the acoustic amplitude sequence, avoiding learning basic features from scratch in the small sample scenario of transformers and reducing dependence on the number of fault samples.

[0047] Furthermore, the shared feature extractor is used as the basic module of the acoustic branch. The input end is connected to the shared feature extractor, and 1-2 fully connected layers can be added to the output end to obtain the acoustic branch. The acoustic branch can use the general knowledge of the shared feature extractor to transform the input acoustic amplitude sequence into an acoustic feature vector output with better fault identification.

[0048] Furthermore, since electrical pulse sequences are time-dependent data, and LSTM networks are adept at capturing long- and short-term time-dependent relationships through gating mechanisms, an electrical branch containing 2-3 layers of LSTM network can be constructed. Taking the electrical pulse sequence as input, after mapping through a fully connected layer, a fixed-dimensional electrical feature vector is output. The role of the electrical branch is to extract time-series features such as pulse energy change trends and discharge frequency patterns in the electrical pulse sequence.

[0049] Furthermore, a fusion classification layer is constructed, which uses weighted summation or attention mechanisms to weight and fuse acoustic and electrical feature vectors, outputting a joint feature vector. For example, if the acoustic signal has a high signal-to-noise ratio (e.g., clear vibration amplitude), the acoustic feature weight is increased; if the electrical signal is more stable (e.g., clear pulse rising edge), the electrical feature weight is increased. The weights are dynamically allocated by calculating the fault contribution of the acoustic and electrical feature vectors, ultimately outputting a joint feature vector that fuses dual-modal information.

[0050] Finally, after the fusion classification layer, a classification output layer containing a fully connected layer (for feature compression) and a softmax activation function (for multi-classification) is connected. Taking the joint features as input, the output fault type and aging level, such as air gap discharge and first-level aging, are obtained to obtain the fault identification model. The fault identification model reuses general knowledge through transfer learning, reduces the dependence on real transformer fault samples, and improves feature recognition through dual-branch fusion, laying the foundation for subsequent fine-tuning.

[0051] It should be noted that the core technical framework of transfer learning, as well as the basic network structures such as one-dimensional convolutional networks, LSTM networks, and fully connected layers involved in this application, are all existing technologies known in the field. The relevant technical principles and implementation details have been fully disclosed in academic literature such as "Introduction to Transfer Learning" and published patents. Therefore, this application will not elaborate on the general details such as the construction logic of the above-mentioned basic model structure and the parameter initialization method. Those skilled in the art can complete the construction and debugging of the basic model without additional disclosure based on the existing knowledge of transfer learning technology and the scenario-based design ideas disclosed in this application.

[0052] Furthermore, such as Figure 2 As shown, the step of "fine-tuning the fault identification model using the enhanced training dataset to obtain a transformer fault identifyer" includes: The augmented training dataset is divided into a training subset and a validation subset; The training subset is input into the fault identification model in batches, the parameters of the shared feature extractor in the fault identification model are fixed, and only the electrical branch, the fusion classification layer and the classification output layer are trained. Perform iterative training, and calculate the fault identification accuracy of the model on the validation subset in each preset round of iteration; Once convergence is verified, the model parameters are saved to obtain the transformer fault identifier.

[0053] In this embodiment, the augmented training dataset is first divided into a training subset and a validation subset according to a preset ratio (e.g., 7:3). The training subset is used for updating model parameters, and the validation subset is used to evaluate the model's generalization ability.

[0054] Secondly, the training subset is input into the fault recognition model in batches. The parameters of the shared feature extractor in the fault recognition model are fixed, and only the electrical branch, fusion classification layer, and classification output layer are trained. For example, the training subset is input into the fault recognition model in batches (e.g., 32 groups per batch). During training, the parameters of the convolutional and pooling layers of the shared feature extractor in the fault recognition model are fixed, and only the parameters of the electrical branch, fusion classification layer, and classification output layer are adjusted. The electrical branch, fusion classification layer, and classification output layer learn the acoustic-electric correlation laws specific to transformers. For example, there is a strong correlation between the acoustic dominant frequency of 100-120kHz and the electrical rise time <0.5μs in the air gap discharge of transformers. These laws are scene-specific and need to be trained with an enhanced training dataset.

[0055] Next, iterative training is performed, with the model's fault identification accuracy calculated on the validation subset after a preset number of iterations. For example, a gradient descent algorithm (such as Adam) is used for iterative training, with a preset number of iterations, such as 10, and the model's fault identification accuracy is calculated using the validation subset.

[0056] Finally, after validation convergence, the model parameters are saved to obtain the transformer fault identifyr. For example, a convergence criterion is set; for instance, when the accuracy of the validation subset remains stable for multiple consecutive rounds (e.g., 20 rounds) and the fault identification accuracy is >95%, the model is considered converged. At this point, the model has learned the fault patterns in the augmented training dataset and possesses good generalization ability. The model parameters at this point are saved to obtain the final transformer fault identifyr. The transformer fault identifyr reuses general knowledge through transfer learning to solve the small sample problem, while training on the augmented training dataset ensures the accuracy of identifying specific fault types and aging levels of transformers.

[0057] In summary, compared to existing technologies, this application constructs a fault identification model based on transfer learning, and fine-tunes the model using the enhanced training dataset to obtain a transformer fault identifyer. This solves the problem of insufficient feature learning with small sample sizes, allowing the transformer fault identifyer to adapt to specific transformer fault patterns and achieve high accuracy and high generalization in fault identification.

[0058] S40: Input the real-time acquired and preprocessed acoustic amplitude sequence and electrical pulse sequence into the transformer fault identifier, and output the preliminary identification results of fault type and aging level.

[0059] This application inputs the real-time acquired and preprocessed acoustic amplitude sequence and electrical pulse sequence into the transformer fault identifier, and outputs preliminary identification results of fault type and aging level.

[0060] For example, following the method in step S10, during transformer operation, high-frequency vibration signals of the tank wall are collected by an ultrasonic sensor, and partial discharge current signals are collected synchronously by a high-frequency current sensor. After preprocessing, acoustic amplitude sequences and electrical pulse sequences are obtained. After inputting the acoustic amplitude sequences and electrical pulse sequences into the transformer fault identifier, the acoustic amplitude sequences flow to the acoustic branch, reuse general high-frequency vibration features and combine them with the specificity of the transformer scenario to transform them into acoustic feature vectors. The electrical pulse sequences capture the pulse timing dependency through the electrical branch to generate electrical feature vectors. Subsequently, the fusion classification layer dynamically allocates weights according to the signal-to-noise ratio of the dual-mode signals. If the acoustic signal is free of interference, the weight is increased to 65%; if the electrical signal is stable, the weight is increased to 70%. The weighted fusion yields a joint feature vector. Finally, the classification output layer calculates the probability through the softmax function and outputs a preliminary identification result with confidence, such as fault type: air gap discharge (confidence 92%), aging level: level one aging (confidence 88%). In this way, the impact of single signal interference is reduced by dual-modal fusion, and early fault warnings are provided to maintenance personnel with high real-time performance, providing a data foundation for subsequent accurate diagnosis by combining data such as dissolved gas in oil.

[0061] S50: Quantitatively calibrate the preliminary identification results and output the final fault identification results.

[0062] In the process of outputting fault type and aging level by transformer fault identifier, the acoustic, electrical, and gaseous multimodal characteristics of early faults are weak and easily masked by environmental noise. In addition, the initial state of insulation materials of different individual devices and the dynamic changes in operating conditions lead to unstable feature distribution, resulting in inherent output errors of transformer fault identifier. Especially for the identification of aging level, because the aging process is continuous, the feature differences between different levels are subtle and there is a lot of overlap, which leads to a significantly higher misjudgment rate of aging level than fault type identification, which may lead to deviations in operation and maintenance decisions.

[0063] To address the aforementioned issues, this application performs quantitative calibration on the preliminary identification results and outputs the final fault identification results.

[0064] Specifically, step S50 in the method includes: Extract the measured acoustic amplitude feature vector, measured electrical pulse feature vector, and measured dissolved gas detection data in transformer oil corresponding to the preliminary identification results; Based on historical data database and preliminary identification results, several historical samples of the same model, fault type and aging level as the target transformer are selected. Correspondingly, several historical acoustic amplitude features, historical electrical pulse features and historical dissolved gas detection data in transformer oil are extracted. The mean values ​​are calculated to obtain historical reference acoustic amplitude feature vector, historical reference electrical pulse feature vector and historical reference dissolved gas detection data in transformer oil. The similarity between the measured acoustic amplitude feature vector and the historical reference acoustic amplitude feature vector is calculated and used as the first consistency coefficient. The similarity between the measured electrical pulse feature vector and the historical reference electrical pulse feature vector is calculated and used as the second consistency coefficient. The similarity between the measured dissolved gas detection data in transformer oil and the historical benchmark dissolved gas detection data in transformer oil is calculated and used as the third consistency coefficient. If the first consistency coefficient is greater than or equal to the preset first consistency coefficient threshold, the second consistency coefficient is greater than or equal to the preset second consistency coefficient threshold, and the third consistency coefficient is greater than or equal to the preset third consistency coefficient threshold, the preliminary identification result is directly output as the final fault identification result; If the first consistency coefficient is less than the preset first consistency coefficient threshold, or the second consistency coefficient is less than the preset second consistency coefficient threshold, or the third consistency coefficient is less than the preset third consistency coefficient threshold, calculate the aging level calibration coefficient, adjust the aging level in the preliminary identification result based on the aging level calibration coefficient, and output the final fault identification result.

[0065] In this embodiment, the measured acoustic amplitude feature vector, measured electrical pulse feature vector, and measured dissolved gas detection data in transformer oil corresponding to the preliminary identification results are first extracted based on the data collected in step S10. In this way, measured data in three dimensions—acoustic, electrical, and chemical—are obtained, which can more comprehensively reflect the true fault state of the equipment.

[0066] Secondly, based on the historical database and preliminary identification results, several historical samples of the same model, fault type, and aging level as the target transformer were selected. Correspondingly, several historical acoustic amplitude features, historical electrical pulse features, and historical dissolved gas detection data in transformer oil were extracted, and their mean values ​​were calculated to obtain historical baseline acoustic amplitude feature vectors, historical baseline electrical pulse feature vectors, and historical baseline dissolved gas detection data in transformer oil. The same model, fault type, and aging level were ensured to guarantee the relevance and comparability of the selected historical data. For example, several historical samples of the same model, fault type, and aging level are selected from the historical database. Then, the mean values ​​are extracted and calculated to obtain historical reference acoustic amplitude feature vectors, historical reference electrical pulse feature vectors, and historical reference dissolved gas detection data in transformer oil. For example, the historical reference acoustic amplitude feature vector is [52μPa, 8.3, 111kHz], the historical reference electrical pulse feature vector is [33pC, 0.36μs, 23ms], and the historical reference dissolved gas detection data in transformer oil is [H2: 42μL / L, CH4: 20μL / L, C2H2: 0.3μL / L]. These reference data represent the typical characteristic level of the target transformer under specific faults and aging levels, and can provide quantifiable comparison standards for measured data.

[0067] Next, the similarity between the measured acoustic amplitude feature vector and the historical reference acoustic amplitude feature vector is calculated as the first consistency coefficient. For example, the similarity can be calculated using cosine similarity. The first consistency coefficient quantifies the degree of matching between the measured acoustic amplitude feature vector and the historical reference acoustic amplitude feature vector, which can reflect the typical feature level.

[0068] Furthermore, following the same method as the first consistency coefficient, the similarity between the measured electrical pulse feature vector and the historical benchmark electrical pulse feature vector is calculated as the second consistency coefficient. The second consistency coefficient quantifies the degree of matching between the measured electrical pulse feature vector and the historical benchmark electrical pulse feature vector, which can reflect the typical feature level.

[0069] Furthermore, the similarity between the measured dissolved gas detection data in transformer oil and the historical benchmark dissolved gas detection data in transformer oil is calculated as a third consistency coefficient. For example, since the gas components have different sensitivities to faults, the relative deviation amplitude can be calculated, and then a weighted fusion method can be used to obtain the third consistency coefficient. For instance, weights of 0.4, 0.3, and 0.3 are assigned to H2, CH4, and C2H2, respectively. If the measured dissolved gas detection data in transformer oil shows an H2 concentration of 45 μL / L, a CH4 concentration of 22 μL / L, and a C2H2 concentration of 0.5 μL / L, and the historical benchmark dissolved gas detection data shows [H2: 42 μL / L]... [CH4: 20 μL / L, C2H2: 0.3 μL / L], then the relative deviation of H2 is 7%, the relative deviation of CH4 is 9%, and the relative deviation of C2H2 is 40%. Therefore, the third consistency coefficient = 1 - (0.4 × 7% + 0.3 × 9% + 0.3 × 40%) = 0.825. The third consistency coefficient quantifies the degree of matching between the measured dissolved gas detection data in transformer oil and the historical benchmark dissolved gas detection data in transformer oil that can reflect typical characteristic levels.

[0070] Furthermore, if the first consistency coefficient is greater than or equal to the preset first consistency coefficient threshold, the second consistency coefficient is greater than or equal to the preset second consistency coefficient threshold, and the third consistency coefficient is greater than or equal to the preset third consistency coefficient threshold, the preliminary identification result is directly output as the final fault identification result. The preset first, second, and third consistency coefficient thresholds can be dynamically adjusted according to actual needs. Preferably, since acoustics is susceptible to mechanical interference, the preset first consistency coefficient threshold should be set relatively high, such as 0.85-0.9; since electrical systems are less susceptible to electromagnetic interference, the preset second consistency coefficient threshold can be set relatively low, such as 0.8-0.85; and since gases have high stability, the preset third consistency coefficient threshold should be set moderately, such as 0.8. For example, only when the first, second, and third consistency coefficients are simultaneously ≥ the corresponding preset thresholds does it indicate that the measured acoustic amplitude feature vector, the measured electrical pulse feature vector, and the measured dissolved gas detection data in transformer oil all highly match the typical characteristics, and the preliminary identification result is reliable. In this case, the preliminary identification result is directly output as the final result.

[0071] Finally, if any consistency coefficient does not meet the corresponding preset threshold, for example, if the first consistency coefficient is less than the preset first consistency coefficient threshold, or the second consistency coefficient is less than the preset second consistency coefficient threshold, or the third consistency coefficient is less than the preset third consistency coefficient threshold, it indicates that there is a deviation between the measured features and the typical features, such as individual differences in equipment, temporary interference, inherent errors in the model, etc. Therefore, it is necessary to calculate the aging level calibration coefficient, adjust the aging level in the preliminary identification result based on the aging level calibration coefficient, and output the final fault identification result.

[0072] Specifically, the step of "calculating the aging level calibration coefficient, adjusting the aging level in the preliminary identification result based on the aging level calibration coefficient, and outputting the final fault identification result" includes: Based on the first consistency coefficient, the second consistency coefficient, and the third consistency coefficient, the aging level calibration coefficient is calculated by weighting. The aging level in the preliminary identification results is quantified and assigned a value to obtain the aging level quantification value; By combining the aging level calibration coefficient and the aging level quantification value, the corrected aging level quantification value is calculated. Boundary constraints are applied to the corrected aging level quantization values, and then mapped to the corresponding final aging level.

[0073] In this embodiment, the aging level calibration coefficient is first calculated by weighting the first consistency coefficient, the second consistency coefficient, and the third consistency coefficient. The aging level calibration coefficient = α × first consistency coefficient + β × second consistency coefficient + γ × third consistency coefficient, where α + β + γ = 1, and γ > α and γ > β. This is because the dissolved gas data in transformer oil reflects the long-term aging trend, and its consistency has a greater impact on the aging level judgment, thus requiring higher weighting. Those skilled in the art can dynamically assign values ​​to α, β, and γ according to the actual situation. For example, if α is 0.25, β is 0.25, γ is 0.5, the first consistency coefficient is 0.75, the second consistency coefficient is 0.88, and the third consistency coefficient is 0.75, then the aging level calibration coefficient = 0.25 × 0.75 + 0.25 × 0.78 + 0.5 × 0.825 = 0.795. The aging level calibration coefficient can reflect the comprehensive consistency of the three modal data. The closer the aging level calibration coefficient is to 1, the higher the comprehensive consistency of the three modal data and the smaller the aging level deviation.

[0074] Secondly, the aging levels in the preliminary identification results are quantified and assigned values ​​to obtain quantified aging level values. For example, Level 1 aging is assigned a value of 1, Level 2 aging is assigned a value of 2, and Level 3 aging is assigned a value of 3. In this way, the level label is converted into a value that can be corrected by the aging level calibration coefficient, so that the deviation can be directly reflected in the level adjustment.

[0075] Next, combining the aging level calibration coefficient and the aging level quantification value, the corrected aging level quantification value is calculated. The corrected aging level quantification value = aging level calibration coefficient × aging level quantification value. For example, if the aging level quantification value in the initial identification result is 2 and the aging level calibration coefficient is 0.795, then the corrected aging level quantification value = 0.795 × 2 = 1.59. During the correction process, the greater the deviation of the aging level calibration coefficient from 1, the greater the combined deviation between real-time features and typical features, and the lower the reliability of the aging level. The aging level quantification value needs to be adaptively reduced (or increased) to avoid misjudgment.

[0076] Finally, boundary constraints are applied to the corrected aging level quantization value, and then mapped to the corresponding final aging level. For example, the boundary constraints for the corrected aging level quantization value are based on the upper and lower limits of the aging level quantization value. For instance, if the upper and lower limits of the aging level quantization value are 3 and 1 respectively, then the boundary constraints for the corrected aging level quantization value are 3 and 1. If the corrected aging level quantization value is < 1, it is constrained to be 1; if the corrected aging level quantization value is > 3, it is constrained to be 3. For example, the corrected aging level quantization value is then mapped to the corresponding final aging level using a preset mapping rule. For example, the preset mapping rule is: 1.0-1.5 → Level 1 aging, 1.6-2.5 → Level 2 aging, 2.6-3.0 → Level 3 aging. If the corrected aging level quantization value is 1.59, according to this mapping rule, the output final aging level is Level 1 aging. Thus, the susceptible aging level is corrected.

[0077] Furthermore, step S50 of the method, when the aging level calibration coefficient is less than a preset calibration coefficient threshold, further includes: The acoustic amplitude sequence, electrical pulse sequence, and dissolved gas detection data in transformer oil that were acquired and preprocessed in real time were marked as low-consistency samples. The digital twin sample generation module is invoked to generate a preset number of similar virtual samples based on the insulation-related parameters corresponding to the low-consistency samples. Based on the similar virtual samples, update the enhanced training dataset and perform incremental fine-tuning training on the transformer fault identifyr until convergence. Save the converged model parameters to complete the iterative optimization of the transformer fault identifier.

[0078] In this embodiment, a preset calibration coefficient threshold is set for the aging level calibration coefficient. When the aging level calibration coefficient is less than the preset calibration coefficient threshold, it indicates that the measured acoustic amplitude feature vector, measured electrical pulse feature vector, and measured dissolved gas detection data in transformer oil exceed the learning range of the existing model. Therefore, the acoustic amplitude sequence, electrical pulse sequence, and dissolved gas detection data in transformer oil that were collected and preprocessed in real time are marked as low-consistency samples and used as model optimization resources. The preset calibration coefficient threshold can be dynamically set according to the actual situation, such as 0.7-0.75.

[0079] Secondly, the digital twin sample generation module is invoked to generate a preset number (e.g., 50-100 sets) of similar virtual samples based on the insulation-related parameters corresponding to the low-consistency samples. The similar virtual samples include virtual current sequences, virtual acoustic amplitude sequences, and virtual dissolved gas concentration data in oil. Multiple sets of similar virtual samples are generated through digital twin to form a new scene sample set.

[0080] Next, based on similar virtual samples, the enhanced training dataset is updated, and the transformer fault identifyer is incrementally fine-tuned until convergence. For example, low-consistency samples and similar virtual samples are added to the enhanced training dataset to form an updated dataset. Following the same training method as step S30, the transformer fault identifyer is incrementally fine-tuned. For instance, the training focuses on the fusion classification layer and the classification output layer, optimizing the weighted logic. During fine-tuning, the recognition accuracy is calculated using a validation subset every 10-20 iterations until the fluctuation is less than 1% for 5 consecutive iterations, at which point convergence is considered achieved. This ensures that the updated transformer fault identifyer can learn the feature patterns of new scenarios.

[0081] Finally, after convergence, the converged model parameters are saved to complete the iterative optimization of the transformer fault identifier. The optimized transformer fault identifier can more accurately identify the aging level when encountering similar new scenarios in the future.

[0082] In summary, compared to existing technologies, this application quantifies and calibrates the preliminary identification results to output the final fault identification results. Thus, by employing a three-modal consistency verification method, the limitations of single-modal calibration are addressed. Through dynamic correction of aging levels, precise quantitative adjustment of deviations is achieved, transforming low-consistency samples into optimization resources. This drives the model to continuously adapt to new scenarios, improving the reliability and accuracy of single fault identification. Furthermore, model iteration enhances long-term adaptability, ultimately achieving high-precision and robust fault identification for transformer fault detectors.

[0083] In summary, the embodiments of this application have at least the following technical effects: Compared to existing technologies, this application first acquires high-frequency vibration signals from the tank wall using an ultrasonic sensor and simultaneously acquires partial discharge current signals using a high-frequency current sensor during transformer operation. After preprocessing these two types of signals, multimodal raw samples containing acoustic amplitude sequences and electrical pulse sequences are obtained. In this way, acoustic and electrical dual-modal raw data reflecting the partial discharge state of the transformer are acquired simultaneously, providing comprehensive basic sample support for subsequent fault type identification and insulation aging level determination.

[0084] Secondly, this application calls the digital twin sample generation module, inputs transformer insulation-related parameters, and generates multimodal virtual samples with the same feature distribution as the original multimodal samples. The original multimodal samples and the multimodal virtual samples are then fused together as an enhanced training dataset. In this way, by simulating the coupling process of multiple physical fields such as electricity, sound, and chemistry through digital twins, sufficient and high-quality training data is provided for subsequent transfer learning models.

[0085] Furthermore, this application constructs a fault identification model based on transfer learning, and fine-tunes the model using the enhanced training dataset to obtain a transformer fault identifyer. This solves the problem of insufficient feature learning with small sample sizes, allowing the transformer fault identifyer to adapt to specific transformer fault patterns and achieve high accuracy and high generalization in fault identification.

[0086] Furthermore, this application inputs the real-time acquired and preprocessed acoustic amplitude sequence and electrical pulse sequence into the transformer fault identifier, outputting preliminary identification results of fault type and aging level. Thus, by reducing the impact of single-signal interference through dual-modal fusion, and providing early fault warning references for maintenance personnel with high real-time performance, it provides a data foundation for subsequent accurate diagnosis by combining data such as dissolved gases in the oil.

[0087] Finally, this application performs quantitative calibration on the preliminary identification results and outputs the final fault identification results. Thus, by using a three-modal consistency verification method, the limitations of single-modal calibration are resolved. Through dynamic correction of aging levels, precise quantitative adjustment of deviations is achieved, transforming low-consistency samples into optimization resources, driving the model to continuously adapt to new scenarios, improving the reliability and accuracy of single fault identification, and enhancing long-term adaptability through model iteration. Ultimately, this achieves high-precision and robust fault identification for transformer fault detectors.

[0088] Through the above technical solutions, this application uses ultrasonic sensors and high-frequency current sensors to collect multimodal data, overcoming the deficiency of single acoustic signals being easily masked by environmental noise. A large number of virtual samples with matching features are generated using a digital twin sample generation module, overcoming the dilemma of scarce effective samples and high annotation costs associated with small samples in early faults, providing sufficient data support for model training. Based on transfer learning, the application reuses general high-frequency vibration feature learning results to reduce dependence on transformer-specific samples. Furthermore, fine-tuning the training to adapt to specific transformer fault patterns improves the generalization ability of the transformer fault identifyr. Finally, quantization calibration further corrects errors. In this way, accurate identification of early partial discharge fault types and aging levels in transformers is achieved, improving the scenario adaptability of early fault identification and meeting the power grid's needs for early warning of transformer insulation status and ensuring the reliable operation of core hub equipment.

[0089] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0090] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0092] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0093] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0094] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0095] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A fault identification method based on acoustic amplitude characteristic changes, characterized in that, include: During transformer operation, high-frequency vibration signals of the tank wall are collected by an ultrasonic sensor, and partial discharge current signals are collected synchronously by a high-frequency current sensor. After preprocessing the two types of signals, multimodal raw samples containing acoustic amplitude sequences and electrical pulse sequences are obtained. The digital twin sample generation module is invoked, transformer insulation-related parameters are input, and multimodal virtual samples with the same feature distribution as the original multimodal samples are generated. The original multimodal samples and the multimodal virtual samples are fused together as an enhanced training dataset. The transformer insulation-related parameters include the degree of polymerization of insulating paper, air gap size, and operating temperature. A fault identification model is constructed based on transfer learning, and the fault identification model is fine-tuned and trained using the enhanced training dataset to obtain a transformer fault identifyer. The acoustic amplitude sequence and electrical pulse sequence, which are collected and preprocessed in real time, are input into the transformer fault identifier, and the preliminary identification results of fault type and aging level are output. The preliminary identification results are quantified and calibrated to output the final fault identification results.

2. The method according to claim 1, characterized in that, After preprocessing the two types of signals, multimodal raw samples containing acoustic amplitude sequences and electrical pulse sequences are obtained, including: The high-frequency vibration signal acquired by the ultrasonic sensor is filtered to extract the peak amplitude, kurtosis, and dominant frequency characteristics of the signal, and then arranged in chronological order to form an acoustic amplitude sequence. Baseline correction is performed on the partial discharge current signal acquired by the high-frequency current sensor, and the pulse peak value, rise time, and pulse interval characteristics of the signal are extracted and arranged in chronological order to form an electrical pulse sequence. The acoustic amplitude sequence and electrical pulse sequence are aligned by timestamp, and combined with the dissolved gas detection data in transformer oil, the corresponding fault type label and aging level label are marked to form a multimodal original sample.

3. The method according to claim 1, characterized in that, The digital twin sample generation module is invoked, transformer insulation-related parameters are input, and multimodal virtual samples with the same feature distribution as the original multimodal samples are generated. The original multimodal samples and the multimodal virtual samples are then fused to form an enhanced training dataset, including: Based on the three-dimensional structural model of the transformer, a digital twin sample generation module is constructed, which includes electrical simulation unit, acoustic simulation unit, and chemical simulation unit. Input the degree of polymerization of the transformer's insulating paper, the air gap size, and the operating temperature into the digital twin sample generation module; The electrical simulation unit generates a virtual current sequence. The discharge energy is calculated using the acoustic simulation unit to generate a virtual acoustic amplitude sequence. The chemical simulation unit synchronously generates virtual dissolved gas concentration data in oil. By integrating the virtual current sequence, virtual acoustic amplitude sequence, and virtual dissolved gas concentration data in oil, a multimodal virtual sample is formed that is consistent with the feature distribution of the original multimodal sample. The original multimodal samples and the virtual multimodal samples are fused together at a preset ratio to form an enhanced training dataset.

4. The method according to claim 1, characterized in that, Fault identification models are built based on transfer learning, including: Acquire general acoustic data of high-frequency mechanical vibration as a sample general acoustic dataset; A one-dimensional convolutional feature extraction network containing convolutional layers and pooling layers is constructed. The network is trained using the sample general acoustic dataset as input until it converges. Freeze the parameters of the convolutional and pooling layers of the one-dimensional convolutional feature extraction network to obtain a transferable shared feature extractor, wherein the shared feature extractor is used to extract general time-frequency features of acoustic amplitude sequences; The shared feature extractor is connected to construct an acoustic branch, which takes the acoustic amplitude sequence as input and outputs an acoustic feature vector. An LSTM network is used to construct the electrical branch, which takes an electrical pulse sequence as input and outputs an electrical feature vector. A fusion classification layer is constructed to weight and fuse the acoustic feature vector and the electrical feature vector, and output a joint feature vector. A classification output layer is connected after the fusion classification layer. Taking the joint features as input, the fault type and aging level are output to obtain the fault identification model.

5. The method according to claim 1, characterized in that, The fault identification model is fine-tuned and trained using the enhanced training dataset to obtain a transformer fault identifyer, including: The augmented training dataset is divided into a training subset and a validation subset; The training subset is input into the fault identification model in batches, the parameters of the shared feature extractor in the fault identification model are fixed, and only the electrical branch, the fusion classification layer and the classification output layer are trained. Perform iterative training, and calculate the fault identification accuracy of the model on the validation subset in each preset round of iteration; Once convergence is verified, the model parameters are saved to obtain the transformer fault identifier.

6. The method according to claim 1, characterized in that, The preliminary identification results are quantified and calibrated to output the final fault identification results, including: Extract the measured acoustic amplitude feature vector, measured electrical pulse feature vector, and measured dissolved gas detection data in transformer oil corresponding to the preliminary identification results; Based on the historical database and preliminary identification results, several historical samples of the same model, fault type and aging level as the target transformer were selected. Correspondingly, several historical acoustic amplitude features, historical electrical pulse features and historical dissolved gas detection data in transformer oil were extracted. The mean values ​​were calculated to obtain the historical reference acoustic amplitude feature vector, historical reference electrical pulse feature vector and historical reference dissolved gas detection data in transformer oil. The similarity between the measured acoustic amplitude feature vector and the historical reference acoustic amplitude feature vector is calculated and used as the first consistency coefficient. The similarity between the measured electrical pulse feature vector and the historical reference electrical pulse feature vector is calculated and used as the second consistency coefficient. The similarity between the measured dissolved gas detection data in transformer oil and the historical benchmark dissolved gas detection data in transformer oil is calculated and used as the third consistency coefficient. If the first consistency coefficient is greater than or equal to the preset first consistency coefficient threshold, the second consistency coefficient is greater than or equal to the preset second consistency coefficient threshold, and the third consistency coefficient is greater than or equal to the preset third consistency coefficient threshold, the preliminary identification result is directly output as the final fault identification result; If the first consistency coefficient is less than the preset first consistency coefficient threshold, or the second consistency coefficient is less than the preset second consistency coefficient threshold, or the third consistency coefficient is less than the preset third consistency coefficient threshold, calculate the aging level calibration coefficient, adjust the aging level in the preliminary identification result based on the aging level calibration coefficient, and output the final fault identification result.

7. The method according to claim 6, characterized in that, Calculate the aging level calibration coefficient, adjust the aging level in the preliminary identification result based on the aging level calibration coefficient, and output the final fault identification result, including: Based on the first consistency coefficient, the second consistency coefficient, and the third consistency coefficient, the aging level calibration coefficient is calculated by weighting. The aging level in the preliminary identification results is quantified and assigned a value to obtain the aging level quantification value; By combining the aging level calibration coefficient and the aging level quantification value, the corrected aging level quantification value is calculated. Boundary constraints are applied to the corrected aging level quantization values, and then mapped to the corresponding final aging level.

8. The method according to claim 7, characterized in that, When the aging level calibration coefficient is less than the preset calibration coefficient threshold, it also includes: The acoustic amplitude sequence, electrical pulse sequence, and dissolved gas detection data in transformer oil that were acquired and preprocessed in real time were marked as low-consistency samples. The digital twin sample generation module is invoked to generate a preset number of similar virtual samples based on the insulation-related parameters corresponding to the low-consistency samples. Based on the similar virtual samples, update the enhanced training dataset and perform incremental fine-tuning training on the transformer fault identifyr until convergence. Save the converged model parameters to complete the iterative optimization of the transformer fault identifier.

Images