Power transformer fault automatic identification and grade determination method

Abstract

The invention discloses a power transformer fault automatic identification and grade determination method, and relates to the technical field of power system equipment state monitoring and fault diagnosis, and the method comprises the following steps: 1, carrying out the cross-domain fusion of multi-modal signal collection, including sensor type selection and layout, time-space synchronous triggering and signal transmission; 2, constructing a transfer learning framework for cross-field coupling feature mining, wherein the transfer learning framework comprises digital twin model construction, transfer learning algorithm optimization and coupling feature modulation and demodulation; according to the invention, through cross-domain fusion of multi-modal signal acquisition, cross-field coupling information is comprehensively reserved; coupling features are accurately mined based on a digital twin model and transfer learning optimization; the hidden fault recognition accuracy and the grade judgment adaptability are improved through an attention enhancement framework; the limitation of traditional single physical field analysis is broken through on the whole, the cost is reduced, meanwhile, early warning and full-life-cycle accurate management and control of various hidden faults are achieved, and support is provided for power grid safety.

Description

Technical Field

[0001] This invention relates to the field of power system equipment condition monitoring and fault diagnosis technology, specifically a method for automatic identification and level determination of power transformer faults. Background Technology

[0002] As a core piece of equipment in the power system, the safe and stable operation of power transformers is directly related to the reliability of the power grid. Transformer operation involves the dynamic coupling of multiple physical fields such as electromagnetic, mechanical, and thermal fields, and the types of faults are complex and diverse. Among them, latent faults (such as loose core, winding deformation, and local overheating) have weak early characteristics and concealed development, making them potential sources of risk for major accidents.

[0003] The essential characteristics of latent faults are significantly unique: their evolution mechanism relies on cross-physical field coupling effects such as "vibration-electromagnetism" and "thermal-structural"—core loosening is manifested through the coupling of mechanical vibration and electromagnetic noise; winding deformation causes distortion of the correlation between vibration and magnetic field through changes in electromagnetic force distribution; and local overheating exacerbates winding stress and electromagnetic parameter fluctuations through thermal expansion. These fault characteristics are not manifested through direct electrical variables (such as current and temperature) of a single physical field, but rather indirectly reflected through cross-field coupling signals, which is the core attribute of latent faults.

[0004] Existing technologies have always been limited to the mindset of "direct electrical variable analysis of a single physical field": technical solutions focus on the extraction of features from a single electrical signal or a single field, and achieve fault identification through electrical variable measurement such as partial discharge detection and oil chromatography analysis, or single field parameter monitoring such as infrared thermometry and magnetic field strength monitoring. Typical paths include direct measurement of electrical signals, analysis of a single physical field, and traditional feature mapping.

[0005] This technical approach fundamentally contradicts the core attributes of latent faults, specifically manifesting as two related aspects of the same problem: At the data acquisition level: Traditional equipment is only equipped with electrical parameters or a single field sensor, without simultaneously acquiring multi-modal signals such as vibration, electromagnetic noise, and temperature, resulting in the direct loss of cross-field coupling information; At the feature analysis level: the mapping relationship between cross-field coupling features and fault types has not been established, and coupling variables such as "electromagnetic noise modulated by vibration" and "electromagnetic parameter fluctuations caused by thermal expansion" have not been used as core identification criteria, resulting in the omission of key features of latent faults.

[0006] These two aspects are mutually restrictive: the lack of cross-field signal acquisition prevents the establishment of cross-field mapping relationships; and the lack of understanding of the importance of cross-field characteristics solidifies the technical approach of single-field measurement. Ultimately, existing technologies fail to capture the essential cross-field coupling characteristics of latent faults, resulting in these faults remaining unidentified for a long time, posing a serious threat to the safe operation of transformers.

[0007] Therefore, there is an urgent need to provide an automatic fault identification and fault level determination method for power transformers, so as to break through the limitations of "direct electrical variable analysis of a single physical field" and achieve effective capture of cross-field coupling characteristics and fault identification. Summary of the Invention

[0008] The purpose of this invention is to provide a method for automatic identification and severity determination of power transformer faults, so as to solve the problems mentioned in the background art.

[0009] To solve the above-mentioned technical problems, the present invention provides a method for automatic identification and level determination of power transformer faults, comprising the following steps: Step 1: Cross-domain fusion of multimodal signal acquisition, including sensor selection and deployment, spatiotemporal synchronization triggering and signal transmission; Step 2: Construction of a transfer learning framework for cross-field coupled feature mining, including digital twin model construction, transfer learning algorithm optimization, and coupled feature modulation and demodulation; Step 3: Fault identification and level determination based on attention enhancement architecture, including deep belief network optimization, cross-field attention mechanism design and fuzzy cognitive graph reasoning.

[0010] Furthermore, the sensor selection and deployment in step one includes: A piezoelectric thin-film vibration sensor is used, with a serpentine electrode pattern. The electromagnetic noise sensor uses a square array of four MEMS microphones spaced 5cm apart. The temperature sensor uses an array of 16 surface-mount NTC thermistors mounted in a 4×4 matrix. The sensors adopt a layered staggered layout, with a sensor layer set every 30cm along the height direction on the inner wall of the transformer tank. Vibration sensors, electromagnetic noise sensors, and temperature sensors are distributed alternately in each layer, with a spacing of 15cm between adjacent sensors.

[0011] Furthermore, the spatiotemporal synchronization triggering mechanism in step one includes: Using crystal oscillator frequency doubling synchronization technology, based on the transformer's 50Hz operating frequency, the crystal oscillator frequency is multiplied to 100MHz through a phase-locked loop circuit to generate a nanosecond-level synchronization pulse signal. Each sensor is connected to a 1MB FIFO buffer, where signals are temporarily stored and then uploaded once the synchronization pulse arrives.

[0012] Furthermore, in step one, the signal transmission adopts twisted-pair differential transmission, and the transmission cable is a twisted-pair cable with an aluminum foil shielding layer.

[0013] Furthermore, in step two, the construction of the digital twin model adopts a hybrid finite element-boundary element modeling approach, with the transformer core modeled using solid elements, the windings using line elements, and the tank modeled using shell elements. The model introduces fault excitation functions: a gap vibration function for core loosening, a radial displacement function for winding deformation, and a heat power density function for local overheating.

[0014] Step two, the optimization of the transfer learning algorithm, includes: A feature alignment attention mechanism is adopted, which assigns weights by calculating the cosine similarity of features between virtual samples and real samples; A sample augmentation strategy is adopted to stretch the actual samples in the time domain and add noise in the frequency domain.

[0015] Furthermore, the coupling feature modulation and demodulation module in step two includes: To address the electromagnetic noise characteristics of vibration modulation, variable step-size wavelet packet decomposition is employed, with a step size of 1Hz for the low-frequency band and 10Hz for the high-frequency band. To address the electromagnetic parameter fluctuation characteristics induced by thermal expansion, Hilbert marginal spectrum analysis is used to extract the fluctuation period and amplitude characteristics.

[0016] Furthermore, the deep belief network optimization described in step three includes: The input layer employs adaptive normalization for cross-field coupling features; The hidden layer introduces a skip connection mechanism, connecting the output of layer 1 to the input of layer 3, and the output of layer 2 to the input of layer 4.

[0017] Furthermore, in step three, the cross-field attention mechanism is designed as a fault type-oriented attention weight calculation, with a preset initial weight matrix that is dynamically adjusted according to the recognition accuracy. The nodes in the fuzzy cognitive graph construction include cross-field coupling feature nodes and fault level nodes. The node connection weights are optimized using the least squares method, and the inference process sets dynamic thresholds based on the transformer operating time.

[0018] Compared with the prior art, the beneficial effects of the present invention are: I. Data Acquisition Level: Fully preserve cross-field coupling information to improve signal quality and synchronization accuracy: Multimodal signal acquisition across all dimensions: Through a layered and staggered deployment of piezoelectric thin-film vibration sensors (serpentine electrodes), a 4×4 matrix NTC temperature sensor array, and a square-distributed MEMS microphone array (one layer every 30cm along the height, with adjacent layers spaced 15cm apart), spatial full-coverage acquisition of multimodal signals including vibration, electromagnetic noise, and temperature is achieved. Compared to traditional single-field sensors, the sampling density is significantly improved, enabling the capture of subtle cross-field characteristics such as micron-level vibrations from loose iron cores and temperature gradients from localized overheating, thus solving the problem of cross-field coupling information loss caused by traditional single-field measurements.

[0019] Nanosecond-level time and space synchronization: It adopts crystal oscillator frequency doubling synchronization technology (multiplying the frequency from 50Hz to 100MHz) to generate nanosecond-level synchronization pulses, and combines a 1MB FIFO buffer to realize signal temporary storage and unified uploading. The synchronization accuracy is improved compared with traditional GPS time synchronization, avoiding time misalignment caused by weak GPS signal when installed indoors, and ensuring strict alignment of multi-modal signals in the time dimension.

[0020] High-stability signal transmission: Differential transmission with twisted pair cable with aluminum foil shielding layer is adopted. The grounding resistance of the shielding layer is ≤1Ω, which effectively suppresses strong electromagnetic interference from transformers, reduces signal transmission distortion, and solves the problem of signal instability in traditional single-ended transmission under strong electromagnetic environment.

[0021] II. Feature Mining Level: Accurately extracting cross-field coupling features to overcome the bottleneck of hidden fault identification: Deep mining of cross-field coupling features: A digital twin model is constructed by finite element-boundary element hybrid modeling, and fault excitation functions such as core gap vibration, winding radial displacement, and local thermal power density are introduced to simulate the four-field coupling process of "vibration-electromagnetism-thermal-strain" and generate diverse virtual fault samples. By combining the feature alignment attention mechanism (based on cosine similarity weighting) and the sample enhancement strategy of time-domain stretching and frequency-domain noise addition, the problem of scarcity of actual fault samples is solved, and the feature alignment accuracy is significantly improved.

[0022] Efficient extraction of coupling features: Variable step size wavelet packet decomposition (1Hz for low frequency and 10Hz for high frequency) is used for electromagnetic noise features of vibration modulation, which improves the efficiency of high frequency band while ensuring the extraction accuracy of low frequency band; Hilbert marginal spectrum analysis is used for electromagnetic parameter fluctuations caused by thermal expansion to accurately extract the fluctuation period and amplitude features, which breaks through the limitation of traditional single physical field analysis that cannot capture cross-field coupling features.

[0023] III. Fault Identification Level: Improving the accuracy of hidden fault identification and the adaptability of fault rating determination: Accurate identification of weak features: The input layer of the deep belief network adopts adaptive normalization to eliminate training bias caused by differences in feature magnitude, thereby improving the convergence speed; the hidden layer introduces a skip connection mechanism (layer 1-3, layer 2-4 connection) to solve the gradient vanishing problem in deep networks and enhance the ability to identify weak coupling features of latent faults such as core loosening and local overheating.

[0024] Dynamic focus on key features: The cross-field attention mechanism presets an initial weight matrix for different fault types (such as focusing on vibration-electromagnetic coupling features for loose iron cores) and dynamically adjusts it according to the recognition accuracy, so that the model accurately focuses on the core coupling features related to specific faults, which significantly improves the feature allocation accuracy compared with the traditional random weight mechanism.

[0025] Adaptive fault rating determination: The fuzzy cognitive graph constructs cross-field coupling features and fault rating nodes (minor / moderate / severe), and optimizes node weights using the least squares method; the reasoning process sets dynamic thresholds based on the transformer's operating time (0-5 years / 5-10 years / over 10 years) to adapt to changes in fault characteristics during equipment aging, achieving accurate fault rating determination throughout the entire life cycle, and can simultaneously identify multiple latent faults such as core loosening and winding deformation.

[0026] IV. Overall Benefits: Breaking through the limitations of traditional technologies, achieving effective control of hidden faults: Breaking away from the traditional framework of "direct electrical variable analysis of a single physical field," this approach employs a technical path of "multimodal cross-domain acquisition - cross-field coupled feature mining - attention-enhanced model recognition," resolving the core contradictions of existing technologies such as cross-field information loss and omission of key features. In practical applications, signal-to-noise ratio, signal synchronization accuracy, and fault identification accuracy are all significantly improved. Furthermore, by using low-cost sensors and conventional transmission technologies, it reduces system costs while achieving early warning and accurate severity determination of latent transformer faults, providing strong support for the safe and stable operation of the power grid. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of step one in the automatic fault identification and level determination method for power transformers of the present invention. Figure 2 This is a schematic diagram of step two in the automatic fault identification and level determination method for power transformers of the present invention. Figure 3 This is a schematic diagram of step three in the automatic fault identification and level determination method for power transformers of the present invention. Detailed Implementation

[0028] 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.

[0029] Please see Figures 1-3 The present invention provides a technical solution: See Figures 1-3 The following is an embodiment of a method for automatic fault identification and severity determination in power transformers: Detailed Embodiments of an Automatic Fault Identification and Level Determination Method for Power Transformers I. Cross-domain fusion of multimodal signal acquisition: 1. Sensor selection and deployment: The selected piezoelectric thin film vibration sensor, unlike traditional piezoelectric ceramic sensors, can be attached to the arc-shaped area of ​​the transformer core surface. When converting weak mechanical vibrations into electrical signals using the piezoelectric effect, a serpentine electrode pattern is used to improve the sensor's sensitivity, enabling it to capture micron-level vibration signals caused by core loosening.

[0030] The electromagnetic noise sensor employs a MEMS microphone array, consisting of four square-distributed MEMS microphones spaced 5cm apart. Beamforming algorithms are used for spatial filtering of electromagnetic noise, suppressing external environmental noise interference from the transformer and improving the signal-to-noise ratio. Traditional electromagnetic noise monitoring often uses a single microphone, which is easily affected by environmental noise.

[0031] The temperature sensor uses a surface-mount NTC thermistor array, with 16 NTC thermistors mounted in a 4×4 matrix on the surface of the winding insulation layer. Each resistor is only 2×2mm in size. Multi-point temperature synchronous acquisition is achieved through a series voltage divider circuit, which improves the sampling density compared to traditional single-point temperature measurement and can capture temperature gradient changes caused by local overheating.

[0032] The sensors are deployed in a layered, staggered layout. A sensor layer is set every 30cm along the height direction on the inner wall of the transformer tank. Vibration sensors, electromagnetic noise sensors, and temperature sensors are distributed alternately in each layer, with a spacing of 15cm between adjacent sensors. This ensures that the collected signals can fully reflect the cross-field coupling situation at different locations inside the transformer.

[0033] 2. Spatiotemporal synchronization triggering mechanism: Employing crystal oscillator frequency doubling synchronization technology, using the transformer's own 50Hz operating frequency as a reference, the crystal oscillator frequency is multiplied to 100MHz through a phase-locked loop circuit to generate nanosecond-level synchronization pulse signals, controlling the sampling trigger time of each sensor. Compared to traditional GPS time synchronization, it is not affected by outdoor signal obstruction, improves synchronization accuracy, and solves the problem of inaccurate synchronization caused by weak GPS signals when the transformer is installed indoors.

[0034] The design incorporates a signal buffer storage module, with each sensor connected to a 1MB FIFO buffer. When the synchronization pulse has not arrived, the signals collected by the sensors are temporarily stored in the buffer and then uploaded uniformly after the synchronization pulse arrives. This avoids time misalignment caused by signal transmission delay and ensures strict alignment of multimodal signals in the time dimension.

[0035] 3. Signal transmission: Differential transmission using twisted-pair cable converts the electrical signal output by the sensor into a differential signal for transmission. The transmission cable is a twisted-pair cable with an aluminum foil shield, and the grounding resistance of the shield is controlled below 1Ω, effectively suppressing electromagnetic interference and reducing signal distortion during transmission. Unlike traditional single-ended signal transmission, this method significantly improves signal transmission stability in strong electromagnetic environments with transformers.

[0036] II. Transfer Learning Framework for Cross-Field Coupled Feature Mining: 1. Digital Twin Model Construction: Based on a hybrid finite element-boundary element modeling approach, the transformer core is modeled using solid elements, the windings using wire elements, and the tank using shell elements. By setting the material properties of each component (such as the permeability of the core, the resistivity of the windings, and the thermal conductivity of the tank) and boundary conditions (such as the convective heat transfer coefficient of the tank's outer surface), the model accurately simulates the coupled processes of vibration, electromagnetic, thermal, and strain within the transformer. Traditional transformer models often use a single finite element model, while this hybrid modeling method can more accurately simulate energy exchange at the field boundaries.

[0037] The model introduces fault excitation functions. For core loosening faults, the excitation function is a time-varying gap vibration function (simulating the relative motion between core laminations); for winding deformation faults, the excitation function is a radial displacement function of the winding (simulating the degree of winding deformation); and for local overheating faults, the excitation function is a localized heat power density function (simulating heat release from the overheated area). By adjusting the excitation function parameters, the cross-field coupling dynamic process under different fault levels can be simulated, generating diverse virtual fault samples.

[0038] 2. Optimization of transfer learning algorithm: A feature alignment attention mechanism is employed. During the feature alignment process between virtual and real samples, the cosine similarity of the features of the two classes of samples in a high-dimensional space is calculated, and features with high similarity are automatically assigned higher weights. This makes the aligned feature space more closely resemble the actual fault feature distribution. Traditional transfer learning algorithms often use simple feature mapping, but the introduction of this attention mechanism improves the accuracy of feature alignment.

[0039] The designed sample augmentation strategy performs time-domain stretching (stretching ratio 0.8-1.2) and frequency-domain noise addition (signal-to-noise ratio 20-30dB) on the actual collected fault samples to increase sample diversity and solve the problem of scarcity of actual fault samples. This strategy, combined with virtual sample generation, increases the amount of training samples for the model.

[0040] In the transfer learning framework for cross-field coupled feature mining, the introduction of a feature alignment attention mechanism improves feature alignment accuracy. This effect is not a subjective judgment but can be precisely proven through a quantitative formula. This formula clearly reflects the degree of alignment between virtual sample features and actual sample features, as well as the role of attention weights in improving alignment accuracy. Formula definition: ; In the formula: For feature alignment precision; The number of feature samples; For the first Attention weights for each feature; ; for The cosine similarity between the features of the first virtual sample and the features of the actual sample in high-dimensional space; For the first The angle between the feature vector of a virtual sample and the feature vector of the actual sample.

[0041] Parameter explanation: : Feature alignment precision, with a value range of [0,1]. The closer to 1, the better the alignment effect.

[0042] The number of feature samples involved in alignment is determined based on the actual sample size and the virtual sample size. generally ≥100.

[0043] : No. The attention weight of each feature is determined by the cosine similarity of that feature; the higher the similarity, the greater the weight.

[0044] Cosine similarity measures the degree of similarity between two feature vectors; the closer to 1, the more similar they are.

[0045] The angle between feature vectors; the smaller the angle, the greater the cosine similarity.

[0046] Proof process: The formula for feature alignment accuracy in traditional feature mapping methods is: ; Due to this method Zhongyu Positive correlation, meaning features with high similarity receive higher weights, therefore: ; When there are differences in feature similarity, therefore This demonstrates that the feature alignment attention mechanism effectively improves feature alignment accuracy.

[0047] 3. Coupling Feature Modulation / Demodulation Module: To address the characteristic of "vibration-modulated electromagnetic noise," a variable-step wavelet packet decomposition algorithm is employed. This algorithm automatically adjusts the step size of the wavelet packet decomposition based on the frequency variations of the electromagnetic noise signal. A small step size (1Hz) is used in the low-frequency range (30-100Hz), while a large step size (10Hz) is used in the high-frequency range (100-1000Hz). This improves feature extraction efficiency while maintaining the accuracy of low-frequency vibration coupling feature extraction. Traditional wavelet packet decomposition often uses a fixed step size.

[0048] For the characteristic of "electromagnetic parameter fluctuations caused by thermal expansion", Hilbert marginal spectrum analysis is used to calculate the changes in the instantaneous frequency and instantaneous amplitude of the electromagnetic parameter signal over time, and the periodic and amplitude characteristics of electromagnetic parameter fluctuations caused by thermal expansion are extracted. This characteristic cannot be obtained in traditional single physical field analysis.

[0049] In the coupled feature modulation and demodulation module, the variable step size wavelet packet decomposition algorithm improves feature extraction efficiency while ensuring extraction accuracy. This advantage can be demonstrated by designing a special decomposition performance evaluation formula, which comprehensively considers both decomposition time and feature extraction accuracy.

[0050] Formula definition: Formula for comprehensive evaluation of decomposition performance: ; Decomposition time calculation formula: ; in: As a comprehensive evaluation index for decomposition performance; , These are the weighting coefficients. , here , ; Decomposition time; To improve feature extraction accuracy; ; ; ; To decompose the step size function; when hour, ;when hour, ; Parameter explanation: The overall performance of the decomposition is evaluated; the higher the value, the better the performance.

[0051] , Weighting coefficients reflect the relative importance of efficiency and accuracy, with accuracy receiving a higher weight.

[0052] : Decomposition time; the smaller the value, the higher the efficiency.

[0053] Feature extraction accuracy: the closer to 1, the higher the accuracy.

[0054] , : Analyze the upper and lower limits of the frequency range, covering the main frequency range of fault characteristics.

[0055] Decompose the step size function and dynamically adjust it to balance efficiency and accuracy.

[0056] Proof process: Traditional fixed step size Decomposition time: , Variable step size decomposition time: , Obviously This improves efficiency. Meanwhile, the small step size in the low-frequency band ensures... Approaching 1, therefore (Evaluation index of traditional methods) proves that variable step size decomposition achieves a balance between efficiency and accuracy.

[0057] III. Attention-enhanced architecture of the fault identification model: 1. Deep Belief Network Optimization: The network input layer employs feature normalization, using an adaptive normalization algorithm for different types of cross-field coupling features (such as the amplitude of vibration signals, the frequency of electromagnetic noise, and the value of temperature). The normalization parameters are automatically adjusted based on the probability distribution of the features, ensuring that all types of features participate in network training at the same magnitude, thus avoiding training bias caused by differences in feature magnitude. Traditional normalization methods often use fixed parameters; this adaptive method improves the network convergence speed.

[0058] The hidden layer introduces a skip connection mechanism, connecting the output of the first layer to the input of the third layer, and the output of the second layer to the input of the fourth layer. This allows the network to retain more shallow feature information during training, solves the gradient vanishing problem during deep network training, and improves the network's ability to recognize weakly coupled features.

[0059] 2. Cross-field attention mechanism: The design employs a fault type-oriented attention weight calculation method. An initial attention weight matrix is ​​preset for different fault types (e.g., vibration-electromagnetic noise coupling features have higher weights in the initial weight matrix for core loosening faults). During model training, the weight matrix is ​​dynamically adjusted based on the fault identification accuracy, enabling the attention mechanism to more accurately focus on coupling features related to specific faults. Traditional attention mechanisms often use randomly set initial weight values; this fault type-oriented design improves the accuracy of feature weight allocation.

[0060] 3. Construction of fuzzy cognitive maps: The nodes of the fuzzy cognitive graph are defined as cross-field coupling feature nodes and fault level nodes. Cross-field coupling feature nodes include "vibration-electromagnetic noise coupling strength" and "temperature-strain coupling coefficient," while fault level nodes are divided into minor faults, moderate faults, and severe faults. The connection weights between nodes are determined by learning the correlation between features and fault levels in historical fault cases. The least squares method is used to optimize the weight values, thereby reducing the inference error of the fuzzy cognitive graph.

[0061] The reasoning process employs dynamic threshold adjustment, setting different fault level judgment thresholds based on the transformer's operating time (e.g., 0-5 years, 5-10 years, and over 10 years). The longer the operating time, the lower the threshold is appropriately, in order to adapt to the changes in fault characteristics during transformer aging and achieve adaptive level judgment for different operating stages.

[0062] It should be noted here that: By coupling and analyzing the micron-level vibration signal of the iron core captured by the piezoelectric thin film vibration sensor with the electromagnetic noise signal collected by the MEMS microphone array, it is possible to identify the precursors of the fault before it occurs, and the early warning time is earlier than that of the existing technology.

[0063] By employing a layered, staggered sensor deployment and multi-feature fusion analysis, this method can simultaneously identify two or more latent faults within a transformer, such as simultaneously identifying core loosening and localized overheating faults, and determine their respective fault levels. Existing technologies often only identify single faults; this ability to simultaneously identify multiple faults provides transformer operation and maintenance with more comprehensive fault information.

[0064] The piezoelectric thin film sensors, MEMS microphone arrays, and NTC thermistor arrays used are all low-cost devices, and the signal transmission adopts conventional technologies such as twisted-pair differential transmission. The overall detection system cost is lower than that of traditional systems that use high-end laser sensors and other equipment, while the fault identification accuracy is improved, achieving a balance between low cost and high precision.

[0065] The attention-enhanced architecture of the fault identification model significantly improves the fault identification accuracy. To accurately quantify this effect, a fault identification accuracy formula that considers the attention weight of fault type is designed. This formula can more reasonably evaluate the identification effect of different fault types.

[0066] Formula definition: ; in: For fault identification accuracy; Number of fault types; For the first Attention weights for different types of faults , For the first The number of faults correctly identified; For the first The total number of samples for each type of fault.

[0067] Parameter explanation: Fault identification accuracy: the closer to 1, the better the identification effect.

[0068] Number of fault types, here (Loose iron core, winding deformation, local overheating).

[0069] Fault attention weight, set according to the importance of the fault, such as , , .

[0070] , : respectively the first The number of faults correctly identified and the total number of samples.

[0071] Proof process: Traditional model accuracy formula: ; because Assigning higher weight to critical faults improves the accuracy of critical fault identification. The increase is greater than The improvement in fault identification accuracy can be verified through this formula, and This demonstrates that the attention-enhanced architecture improves recognition accuracy.

[0072] Further explanation is needed: At the data acquisition level: through sensor selection, layered and staggered deployment, crystal oscillator frequency doubling synchronous triggering, and twisted-pair differential transmission, the cross-field coupling information of multi-modal signals such as vibration, electromagnetic noise, and temperature is fully acquired and retained, completely solving the problem of cross-field coupling information loss caused by traditional single-field acquisition.

[0073] At the feature analysis level: Virtual samples are generated using a digital twin model with finite element-boundary element hybrid modeling. Feature alignment is performed using a transfer learning algorithm with feature alignment attention mechanism. Then, cross-field coupling features are mined using algorithms such as variable step size wavelet packet decomposition. A precise mapping relationship between cross-field coupling features and fault types is established, which solves the problem of missing key features at the feature analysis level.

[0074] This method breaks away from the traditional framework of "direct electrical variable analysis of a single physical field" and forms a new technical path of "multimodal cross-domain acquisition - cross-field coupling feature mining - attention-enhanced model recognition". It realizes full-cycle, high-precision identification and level determination of latent faults in transformers and solves the problem that existing technologies cannot effectively identify latent faults.

[0075] Summarize: From a data acquisition perspective, traditional equipment is only equipped with electrical parameter or single-field sensors, failing to simultaneously acquire multimodal signals such as vibration, electromagnetic noise, and temperature, resulting in the direct loss of cross-field coupling information. This method, however, employs multiple types of sensors, including piezoelectric thin-film vibration sensors, MEMS microphone arrays, and patch-type NTC thermistor arrays, using a layered, staggered layout to achieve multimodal signal acquisition. Simultaneously, it utilizes crystal oscillator frequency doubling synchronization technology and a signal buffer storage module to ensure signal synchronization, and employs twisted-pair differential transmission to ensure signal stability, comprehensively resolving the problem of cross-field coupling information loss during data acquisition.

[0076] In feature analysis, traditional techniques fail to establish a mapping relationship between cross-field coupling features and fault types, and do not use coupling variables such as "electromagnetic noise modulated by vibration" and "electromagnetic parameter fluctuations caused by thermal expansion" as core identification criteria, leading to the omission of key features of latent faults. This method constructs a transfer learning framework for cross-field coupling feature mining. Virtual samples are generated through a digital twin model using a finite element-boundary element hybrid modeling approach. The feature-aligned attention mechanism and sample augmentation strategy address the sample scarcity problem. Variable step-size wavelet packet decomposition and Hilbert marginal spectrum analysis are then used to mine coupling features, establishing a precise mapping relationship between cross-field coupling features and fault types, effectively solving the problem of missing key features in the feature analysis stage.

[0077] Furthermore, traditional techniques have always been confined to the framework of "direct electrical variable analysis of a single physical field," with solutions focusing on the extraction of features from a single electrical signal or a single field. This method, however, breaks this traditional framework, forming a novel technical path of "multimodal cross-domain acquisition - cross-field coupling feature mining - attention-enhanced model recognition." This approach effectively captures the essential cross-field coupling characteristics of latent faults, enabling effective identification and severity determination of latent faults, thus solving the problem of traditional techniques' inability to effectively identify latent faults.

Claims

1. A method for automatic fault identification and severity determination in power transformers, characterized in that, Includes the following steps: Step 1: Cross-domain fusion of multimodal signal acquisition, including sensor selection and deployment, spatiotemporal synchronization triggering and signal transmission; Step 2: Construction of a transfer learning framework for cross-field coupled feature mining, including digital twin model construction, transfer learning algorithm optimization, and coupled feature modulation and demodulation; Step 3: Fault identification and level determination based on attention enhancement architecture, including deep belief network optimization, cross-field attention mechanism design and fuzzy cognitive graph reasoning.

2. The method for automatic fault identification and level determination of power transformers as described in claim 1, characterized in that: Step one, sensor selection and deployment, includes: A piezoelectric thin-film vibration sensor is used, with a serpentine electrode pattern. The electromagnetic noise sensor uses a square array of four MEMS microphones spaced 5cm apart. The temperature sensor uses an array of 16 surface-mount NTC thermistors mounted in a 4×4 matrix. The sensors adopt a layered staggered layout, with a sensor layer set every 30cm along the height direction on the inner wall of the transformer tank. Vibration sensors, electromagnetic noise sensors, and temperature sensors are distributed alternately in each layer, with a spacing of 15cm between adjacent sensors.

3. The method for automatic fault identification and level determination of power transformers as described in claim 1, characterized in that: The spatiotemporal synchronization triggering mechanism in step one includes: Using crystal oscillator frequency doubling synchronization technology, based on the transformer's 50Hz operating frequency, the crystal oscillator frequency is multiplied to 100MHz through a phase-locked loop circuit to generate a nanosecond-level synchronization pulse signal. Each sensor is connected to a 1MB FIFO buffer, where signals are temporarily stored and then uploaded once the synchronization pulse arrives.

4. The method for automatic fault identification and level determination of power transformers as described in claim 1, characterized in that: In step one, signal transmission uses twisted-pair differential transmission, and the transmission cable is a twisted-pair cable with an aluminum foil shielding layer.

5. The method for automatic fault identification and level determination of power transformers as described in claim 1, characterized in that: In step two, the digital twin model is constructed using a hybrid finite element-boundary element modeling approach. The transformer core is modeled using solid elements, the windings using line elements, and the tank using shell elements. The model introduces fault excitation functions: a gap vibration function for core loosening, a radial displacement function for winding deformation, and a heat power density function for local overheating.

6. The method for automatic fault identification and level determination of power transformers as described in claim 1, characterized in that: Step two, the optimization of the transfer learning algorithm, includes: A feature alignment attention mechanism is adopted, which assigns weights by calculating the cosine similarity of features between virtual samples and real samples; A sample augmentation strategy is adopted to stretch the actual samples in the time domain and add noise in the frequency domain.

7. The method for automatic fault identification and level determination of power transformers as described in claim 1, characterized in that: The coupling feature modulation and demodulation module in step two includes: To address the electromagnetic noise characteristics of vibration modulation, variable step-size wavelet packet decomposition is employed, with a step size of 1Hz for the low-frequency band and 10Hz for the high-frequency band. To address the electromagnetic parameter fluctuation characteristics induced by thermal expansion, Hilbert marginal spectrum analysis is used to extract the fluctuation period and amplitude characteristics.

8. The method for automatic fault identification and level determination of power transformers as described in claim 1, characterized in that: Step 3, the deep belief network optimization, includes: The input layer employs adaptive normalization for cross-field coupling features; The hidden layer introduces a skip connection mechanism, connecting the output of layer 1 to the input of layer 3, and the output of layer 2 to the input of layer 4.

9. The method for automatic fault identification and level determination of power transformers as described in claim 1, characterized in that: In step three, the cross-field attention mechanism is designed to calculate attention weights based on fault type, with a preset initial weight matrix that is dynamically adjusted according to the recognition accuracy. The nodes in the fuzzy cognitive graph construction include cross-field coupling feature nodes and fault level nodes. The node connection weights are optimized using the least squares method, and the inference process sets dynamic thresholds based on the transformer operating time.

Images