A transformer winding health state evaluation method and system based on current fingerprint features

Abstract

The present application relates to the technical field of power system, disclose a kind of transformer winding health state evaluation method and system based on current fingerprint feature, the present application is excited by ross coil to collect transformer inrush current or short-circuit impulse current waveform, using anti-aliasing filter, wavelet threshold denoising and zero correction are carried out signal pretreatment, extract time-domain feature, frequency domain feature and time-frequency domain feature to constitute multidimensional feature parameter.Characteristic parameter is input into one-dimensional convolutional neural network, bidirectional long short-term memory network and the deep learning model of attention mechanism fusion, output winding loose grade, insulation aging index and three evaluation indexes of residual life prediction.The system uses cloud edge collaborative architecture, edge terminal completes local preprocessing and inference, cloud platform executes global model training and parameter update, and is equipped with benchmark fingerprint library and alarm management module.The present application can realize the beneficial effects of early warning advance, high evaluation accuracy, low implementation cost, operation and maintenance mode upgrade and strong generalization ability.

Description

Technical Field

[0001] This invention relates to the field of power system technology, specifically to a method and system for assessing the health status of transformer windings based on current fingerprint characteristics. Background Technology

[0002] As a core hub in the power grid system, the power transformer undertakes the crucial functions of electromagnetic energy conversion and voltage level transformation. Its operating status directly affects the safety, stability, and reliability of the power grid. The transformer windings, as the core component for electromagnetic induction, inevitably endure multiple stresses during long-term operation, including thermal aging, mechanical vibration, and short-circuit electrodynamics. This leads to progressive damage such as loosening of the winding structure, decreased elasticity of the spacers, deterioration of insulation materials, and even inter-turn short circuits, seriously threatening the reliable operation of the transformer. How to monitor and assess the health status of transformer windings in real time, accurately, and effectively has become a critical technical challenge that urgently needs to be addressed in the field of power system operation and maintenance.

[0003] Currently, transformer winding condition monitoring technology has developed into various methods, mainly including dissolved gas analysis in oil, frequency response analysis, infrared thermography detection, and vibration-based online monitoring.

[0004] Dissolved gas analysis in insulating oil identifies internal faults by analyzing changes in the content of characteristic gases. However, this method lacks sensitivity to slow deterioration processes such as winding mechanical loosening and pad compression, and suffers from a lag in gas accumulation, making it unable to provide early warning.

[0005] Frequency response analysis methods measure the frequency response characteristics of the winding transfer function by applying low-voltage pulse excitation, but this requires the transformer to be taken out of operation, has insufficient ability to detect dynamic faults such as inter-turn short circuits, and is complicated to operate on-site and difficult to achieve online monitoring.

[0006] Infrared thermal imaging detection methods indirectly identify internal hot spots by detecting the temperature distribution on the surface of the transformer casing. However, they can only identify faults that have formed stable hot spots and are powerless against early deterioration where the contact resistance gradually increases but has not yet formed hot spots.

[0007] Vibration monitoring methods determine the tightness of the pads by collecting vibration signals from the windings. However, the vibration signals are significantly affected by ambient noise, and the sensor installation location has a significant impact on the monitoring results.

[0008] The existing monitoring methods mentioned above all rely on external excitation signals or indirect physical quantities. They suffer from common defects such as indirect signal acquisition, insufficient early warning lead time, and lack of multi-dimensional feature fusion. They cannot assess the mechanical fastening status of windings and the degree of insulation aging in real time, accurately, and economically. As a result, transformer faults are mostly dealt with by post-failure emergency repairs, which seriously affects the safe and reliable operation of the power grid. There is an urgent need for a new monitoring and evaluation method that can directly sense changes in the electrical characteristics of windings, has early warning capabilities, and can achieve multi-dimensional feature fusion. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for assessing the health status of transformer windings based on current fingerprint features. This invention aims to solve the technical problems of indirect signal acquisition, insufficient early warning lead time, and lack of multi-dimensional feature fusion in existing transformer winding health status monitoring methods, and to achieve online, real-time, and accurate assessment of the health status of transformer windings.

[0010] To achieve the above objectives, according to one aspect of the present invention, a method for assessing the health status of transformer windings based on current fingerprint characteristics is provided, comprising the following steps: Step 1: Acquire transient current waveform data generated during transformer operation: A wideband Rogowski coil current sensor is connected in series in the secondary circuit of the current transformer on the high-voltage or low-voltage side of the transformer to acquire the transformer inrush current waveform or short-circuit impact current waveform. When the acquisition module detects an inrush current event or a short-circuit impact current event, it triggers a high-speed acquisition window and continuously records transient current waveform data for a preset duration. Unlike frequency response analysis methods, which require the active injection of a frequency sweep signal, this step achieves non-intrusive monitoring without the need for an additional excitation source by passively acquiring the inrush current or short-circuit impact current generated during the operation of the equipment itself.

[0011] Step 2: Perform signal preprocessing on the transient current waveform data: Use an anti-aliasing filter to filter the acquired signal, remove high-frequency noise interference, remove the DC component in the signal, perform wavelet threshold denoising to eliminate random noise, perform zero-point correction to eliminate zero-point offset of the acquisition channel, and obtain the preprocessed current waveform data.

[0012] Step 3: Extract multi-dimensional feature parameters from the preprocessed current waveform data: Perform time-domain feature extraction, including peak current, rise time, oscillation attenuation coefficient, and waveform duration; perform frequency-domain feature extraction, extracting the fundamental to preset harmonic amplitudes using Fast Fourier Transform, and calculating the total harmonic distortion and main harmonic frequency; perform time-frequency domain feature extraction, extracting the energy distribution of each frequency band through wavelet packet decomposition, calculating the high-frequency energy proportion, and extracting the phase angle between the inrush current and the system voltage as the phase feature. Unlike existing methods that rely on single-dimensional feature extraction, this step constructs a multi-dimensional feature parameter system combining time, frequency, and time-frequency domains.

[0013] Step 4: Input the multi-dimensional feature parameters into a pre-trained deep neural network model for health status assessment: A one-dimensional convolutional neural network is used to extract local time-domain features of the current waveform, a bidirectional long short-term memory network is used to model the temporal dependencies of the waveform sequence, and an attention mechanism is introduced to weight and fuse key feature dimensions. The model outputs three assessment indicators: winding looseness level, insulation aging index, and remaining life prediction. Unlike the threshold judgment and traditional machine learning assessment methods in existing technologies, this step achieves end-to-end automatic assessment through a deep learning model.

[0014] Preferably, in step 1, the triggering condition for the inrush current event is a sudden increase in the effective value of the current to more than 150% of the rated current; the triggering condition for the short-circuit impact current event is the detection of a current peak exceeding 8 times the rated current and a duration of less than 200 milliseconds. The trigger detection employs a dual-threshold comparison logic. First, the real-time effective value of the current is calculated and compared with a first threshold (150% of the rated current). When the effective value exceeds the first threshold, the trigger determination state is entered. Then, it is detected whether the maximum envelope of the current peak exceeds a second threshold (8 times the rated current) and whether the duration is less than 200 milliseconds. When the above conditions are met, the trigger event is confirmed and a high-speed acquisition window is started. The acquisition window duration is set to no less than 500ms, preferably 800ms, where the pre-trigger duration is set to 100ms to record waveform data before the trigger event occurs. The sampling rate is set to 100kS / s, the total number of sampling points within the acquisition window is 80,000, and the data format is 32-bit floating-point numbers.

[0015] Preferably, the Rogowski coil current sensor has a bandwidth of 10Hz to 10kHz and a sampling rate of not less than 50kS / s, and more preferably a sampling rate of 100kS / s. The acquisition module incorporates an anti-aliasing filter and a 16-bit analog-to-digital converter, ensuring quantization accuracy that meets the high-fidelity acquisition requirements of transient current waveforms. The cutoff frequency of the anti-aliasing filter is 15kHz.

[0016] Preferably, the signal conditioning circuit includes a differential amplifier, a low-pass filter, and a programmable gain amplifier. The differential amplifier converts the single-ended signal output from the Rogowski coil into a differential signal. The cutoff frequency of the low-pass filter is set to 15kHz. The programmable gain amplifier automatically adjusts the amplification factor according to the amplitude range of the input signal.

[0017] Preferably, in step 2, the signal preprocessing includes: calculating the arithmetic mean of the sampled sequence as the DC offset and subtracting it from each sampling point to eliminate the influence of zero-point drift; the wavelet threshold denoising process uses a soft threshold function, the number of wavelet decomposition layers is set to 5, the wavelet basis function is selected as the db8 wavelet, the threshold selection adopts the unbiased likelihood estimation criterion, and the optimal threshold is calculated for the high-frequency coefficients of each wavelet decomposition layer; the zero-point offset is calculated for baseline correction in the stable interval of the first 50ms of the waveform start segment; the denoised signal is resampled to the standard sampling rate of 50kS / s to unify the input format of subsequent feature extraction, and the resampling adopts an algorithm combining an anti-aliasing low-pass filter and decimation / interpolation.

[0018] Preferably, in step 3, the rise time is defined as the time taken for the peak value to rise from 10% to 90%; the oscillation decay coefficient is obtained by fitting the exponential decay characteristic of the waveform envelope, and the exponential decay model of the waveform envelope is expressed as follows: The oscillation damping coefficient is taken as: ,in The time interval is dimensionless and ranges from 0 to 1. An increase in rise time reflects a change in winding inductance parameters (leakage flux changes due to winding loosening), while a decrease in the oscillation attenuation coefficient reflects a change in the damping characteristics of the winding's mechanical structure (loose pads, reduced clamping force). The total harmonic distortion is defined as the ratio of the square root of the sum of the squares of the amplitudes of each harmonic to the fundamental amplitude; the main harmonic frequency is defined as the frequency position corresponding to the harmonic component with the largest amplitude in the harmonic amplitude spectrum.

[0019] Preferably, in step 3, the time-frequency domain feature extraction employs a wavelet packet decomposition algorithm with a db10 wavelet basis and a decomposition level of 5, decomposing the signal into 32 frequency bands, each with a bandwidth of 781.25Hz. The energy proportion of each frequency band is calculated, and the high-frequency energy proportion is defined as the ratio of the energy of the high-frequency band (frequency ≥ 2kHz, corresponding to frequency band numbers 16 to 31) to the total signal energy. An increase in the high-frequency energy proportion reflects the enhanced modulation effect of mechanical vibration caused by winding loosening on the high-frequency components of the current waveform.

[0020] Preferably, in step 3, the phase feature extraction targets the inrush current event, extracting the phase angle between the inrush current and the system voltage. Simultaneously with the current fingerprint acquisition, the acquisition module synchronously acquires the bus voltage as a reference phase signal. The phase angle is calculated by first detecting the moment the voltage crosses zero, then detecting the moment corresponding to the peak value of the inrush current, and calculating the time difference between the two moments. The phase angle is , where T is the power grid period of 20ms.

[0021] Preferably, in step 4, the one-dimensional convolutional neural network comprises three convolutional layers. The first convolutional layer contains 64 convolutional kernels with a kernel length of 16 sampling points, a stride of 1, and a "same" padding method. The activation function is a modified linear unit (MRU). The second convolutional layer contains 128 convolutional kernels with a kernel length of 8 sampling points, a stride of 2, and the activation function is a modified linear unit (MRU). The third convolutional layer contains 256 convolutional kernels with a kernel length of 4 sampling points, a stride of 2, and the activation function is a modified linear unit (MRU). Each convolutional layer is followed by a batch normalization layer and a max pooling layer. The pooling window length of the max pooling layer is 2.

[0022] Preferably, in step 4, the bidirectional long short-term memory network includes two long short-term memory layers, each containing 256 hidden units. The forward and reverse long short-term memory units process the sequence data independently, and the output layer splices and fuses the forward and reverse hidden states, with the fused feature vector having a dimension of 512.

[0023] Preferably, in step 4, the attention mechanism adopts a scaled dot product attention mechanism. First, the fused feature vector is mapped to a query vector and a key vector through a fully connected layer. The dimensions of the query vector and the key vector are both 64. Then, the dot product of the query vector and the key vector is calculated and divided by √64 for scaling. The scaled dot product result is normalized by the softmax function to obtain the attention weight. Finally, the attention weight and the value vector are weighted and summed to obtain the attention-weighted feature vector.

[0024] Preferably, in step 4, the winding looseness level is a graded output from 1 to 5, where level 1 represents normal condition, level 2 represents slight looseness, level 3 represents moderate looseness, level 4 represents severe looseness, and level 5 represents extreme looseness; the insulation aging index is a regression output from 0 to 1, where 0 represents brand new insulation and 1 represents fully aged insulation; the remaining life estimate is a regression output in years. The model training employs a multi-task learning framework, and the loss function is a weighted sum of the cross-entropy loss of the classification task and the mean squared error loss of the regression task.

[0025] On the other hand, the present invention provides a transformer winding health status assessment system based on current fingerprint characteristics, comprising: The current fingerprint acquisition subsystem is used to acquire transient current waveform data generated during transformer operation. The current fingerprint acquisition subsystem includes a wideband Rogowski coil current sensor, a signal conditioning circuit, and an analog-to-digital conversion module, which are connected in series in the secondary circuit of the transformer bushing current transformer. The signal preprocessing module is used to perform anti-aliasing filtering, DC component removal, wavelet threshold denoising, and zero-point correction on the transient current waveform data, and output the preprocessed current waveform data. The feature extraction module is used to extract multi-dimensional feature parameters from the preprocessed current waveform data. The multi-dimensional feature parameters include time-domain features, frequency-domain features, and time-frequency-domain features. The deep neural network evaluation module includes a one-dimensional convolutional neural network submodule, a bidirectional long short-term memory network submodule, and an attention mechanism submodule, which are used to evaluate the health status of the transformer windings based on the multi-dimensional feature parameters. The output module is used to output the health status assessment results generated by the deep neural network assessment module.

[0026] Preferably, the system further includes a cloud-edge collaborative architecture: the edge computing terminal is deployed on the intelligent fusion terminal in the transformer area to complete local data preprocessing and lightweight model inference, and completes the entire processing flow from raw waveform acquisition to evaluation result output within a 2-second response time threshold; the cloud platform is used to aggregate data from multiple transformers, perform global model training and parameter updates; a data synchronization channel is set between the edge computing terminal and the cloud platform to realize the periodic synchronization of model parameters, and the periodic synchronization period is 24 hours.

[0027] Preferably, the edge inference engine is optimized and pruned based on a deep neural network model trained in the cloud. It uses knowledge distillation technology to compress the large-scale model in the cloud into a lightweight version. The pruning strategy includes network weight quantization, inter-layer connection pruning, and feature dimension compression. The quantization precision is compressed from 32-bit floating-point numbers to 8-bit fixed-point numbers.

[0028] Preferably, the system further includes a reference fingerprint library module for storing reference current fingerprint features of the same model transformers from factory tests. The reference current fingerprint features include reference time-domain feature parameters, reference frequency-domain feature parameters, and reference time-frequency-domain feature parameters. The reference fingerprint library module supports comparison and analysis of evaluation results with reference values, and identifies the gradual change trend of winding status through deviation rate calculation. The deviation rate is calculated as the percentage of the difference between the measured value and the reference value divided by the reference value.

[0029] Preferably, the system further includes an alarm management module for automatically generating alarm work orders based on the health status assessment results. Alarm triggering conditions include: winding looseness level reaching 3 or above, or insulation aging index exceeding 0.6, or estimated remaining lifespan less than 10 years. Alarm levels are divided into three levels: a yellow alarm corresponds to winding looseness level 3 or insulation aging index of 0.6 to 0.7, requiring on-site inspection to be completed within 30 days; an orange alarm corresponds to winding looseness level 4 or insulation aging index of 0.7 to 0.85, requiring maintenance to be completed within 15 days; a red alarm corresponds to winding looseness level 5 or insulation aging index exceeding 0.85 or estimated remaining lifespan less than 5 years, requiring immediate initiation of transformer replacement or overhaul procedures.

[0030] Preferably, the cloud platform includes a data aggregation module, a global model training module, and a parameter synchronization module; the data aggregation module is responsible for receiving feature data and evaluation results reported by all edge terminals within its jurisdiction and establishing a regional transformer health management database; the global model training module uses a distributed GPU cluster to train and update the deep learning model, and the training adopts a data parallel strategy; the parameter synchronization module distributes the updated model parameters to each edge terminal.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes an event-triggered acquisition mechanism to automatically capture transient current waveforms each time a transformer experiences inrush current or short-circuit impact. By extracting characteristic parameters highly sensitive to changes in winding condition, such as the proportion of high-frequency energy, it can detect subtle changes in current fingerprints in the early stages of slow deterioration in mechanical conditions, such as winding loosening and pad compression. Furthermore, the deep learning model in this invention, through the fusion analysis of multi-dimensional features, can issue early warnings in the initial stages of fault development. The synergistic effect of these technical features enables this invention to provide early warnings before faults occur, offering a sufficient time window for preventative maintenance and effectively preventing transformer faults from developing into sudden accidents. This invention constructs a multi-dimensional feature parameter system combining time domain, frequency domain, and time-frequency domain. The rise time is directly related to the winding inductance parameters, the oscillation attenuation coefficient is directly related to the winding mechanical damping characteristics, and the high-frequency energy proportion is directly related to the winding mechanical vibration modulation effect. Each feature parameter has a clear physical meaning. Furthermore, it extracts local time-domain features through a one-dimensional convolutional neural network, models temporal dependencies using a bidirectional long short-term memory network, and weights and fuses key feature dimensions using an attention mechanism, achieving in-depth mining and accurate identification of current fingerprint features. Compared to existing methods that rely on single-dimensional features and traditional threshold judgments, the evaluation results of this invention are more accurate and reliable. This invention employs a wideband Rogowski coil current sensor connected in series in the secondary circuit of the transformer bushing current transformer, eliminating the need to modify the transformer's primary equipment and simplifying the installation process. Compared to frequency response analysis methods that require the transformer to be taken out of service, and vibration methods that require sensors to be installed on the tank surface, this invention is suitable for upgrading existing transformers, effectively reducing the deployment cost of the monitoring system. This invention provides maintenance personnel with a comprehensive equipment health profile based on three evaluation indicators: winding looseness level, insulation aging index, and remaining life estimate. Through a tiered alarm mechanism and differentiated maintenance strategies in the alarm management module, it achieves a shift from scheduled maintenance to condition-based maintenance. Maintenance plans are formulated based on real-time evaluation results, avoiding over-maintenance due to "mandatory maintenance upon expiration" and safety hazards caused by "failure to maintain upon expiration," reducing unnecessary investment of manpower and resources, and lowering the total lifecycle maintenance cost. This invention utilizes a cloud-edge collaborative architecture, where the cloud platform aggregates multi-source data to train a global model. It also employs transfer learning technology to adapt to different transformer models and operating conditions. The collaborative update mechanism between edge terminals and the cloud ensures that the model always maintains optimal performance and supports cross-model and cross-regional model migration applications. Attached Figure Description

[0032] Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the deep neural network model in the invention; Figure 3 This is a schematic diagram of the multi-level interaction relationship and data flow of the cloud-edge collaborative evaluation architecture in this invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer and more complete, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention in any way. Example

[0034] A transformer winding health status assessment system based on current fingerprint features includes a current fingerprint acquisition subsystem, a signal preprocessing module, a feature extraction module, a deep neural network evaluation module, an output module, a cloud-edge collaborative architecture, a benchmark fingerprint library module, and an alarm management module.

[0035] The current fingerprint acquisition subsystem consists of a wideband Rogowski coil current sensor, a signal conditioning circuit, and a high-speed analog-to-digital converter module. The Rogowski coil current sensor has a bandwidth of 10Hz to 10kHz and a sampling rate of 100kS / s. The acquisition module incorporates an anti-aliasing filter and a 16-bit analog-to-digital converter. The signal conditioning circuit includes a differential amplifier (converting the single-ended signal output from the Rogowski coil into a differential signal), a low-pass filter (cutoff frequency 15kHz), and a programmable gain amplifier (automatically adjusting the amplification factor according to the input signal amplitude).

[0036] The edge computing terminal is deployed in the intelligent converged terminal of the distribution area (ARM Cortex-A processor, clock speed no less than 1.2GHz, 1GB RAM, 8GB storage), completing local data preprocessing and lightweight model inference, completing the entire process from data acquisition to output within 2 seconds. The edge inference engine uses knowledge distillation technology to compress the cloud model into a lightweight version, with pruning strategies including network weight quantization (compressing 32-bit floating-point numbers to 8-bit fixed-point numbers), inter-layer connection pruning, and feature dimension compression. The edge terminal has a self-assessment capability for network outages, automatically uploading cached data after network recovery.

[0037] The cloud platform comprises a data aggregation module, a global model training module, and a parameter synchronization module. The data aggregation module receives feature data and evaluation results reported by all edge terminals within its jurisdiction, establishing a regional transformer health management database. This database includes three types of data tables: a transformer equipment ledger, a feature parameter storage table, and an evaluation history record table. The global model training module uses a distributed GPU cluster for training and updating the deep learning model. Training employs a data parallel strategy, distributing training samples across multiple GPUs for parallel computation to accelerate convergence. After training, the updated model parameters are distributed to each edge terminal via the parameter synchronization module. Parameter synchronization between the cloud platform and edge terminals uses a combination of periodic and event-triggered synchronization. The periodic synchronization cycle is set to 24 hours. At the end of each synchronization cycle, the edge terminal uploads its locally accumulated feature data and evaluation results to the cloud platform. The cloud platform then aggregates the data, executes global model training, and sends the updated parameters back to the edge terminals.

[0038] The benchmark fingerprint database stores the benchmark current fingerprint characteristics of transformers of the same model from factory tests. The database is categorized and managed by transformer model, with each model maintaining an independent set of benchmark parameters. The benchmark parameters included in the database are: benchmark time-domain characteristic parameters, covering the benchmark value of peak current and its allowable deviation range, the benchmark value of rise time and its allowable deviation range, the benchmark value of oscillation attenuation coefficient and its allowable deviation range, and the benchmark value of waveform duration and its allowable deviation range; benchmark frequency-domain characteristic parameters, covering the benchmark amplitude and allowable deviation range of the fundamental to 50th harmonic, the benchmark value of total harmonic distortion and its allowable deviation range, and the benchmark position of the main harmonic frequency; and benchmark time-frequency-domain characteristic parameters, covering the benchmark energy distribution and high-frequency energy proportion benchmark values ​​of the 32 frequency bands after wavelet packet decomposition. The benchmark fingerprint database supports comparative analysis between evaluation results and benchmark values. It identifies the gradual change trend of winding status through deviation rate calculation. The deviation rate is calculated as the percentage of the difference between the measured value and the benchmark value divided by the benchmark value. When the deviation rate exceeds a preset threshold, it is marked as an abnormal feature.

[0039] The alarm management module automatically generates alarm work orders based on the winding looseness level, insulation aging index, and estimated remaining life. Alarm triggering conditions include the following three scenarios: an alarm is triggered when the winding looseness level reaches level 3 or above, or when the insulation aging index exceeds 0.6, or when the estimated remaining life is less than 10 years. Alarm levels are divided into three categories: yellow alarms correspond to winding looseness level 3 or an insulation aging index between 0.6 and 0.7, requiring on-site inspection within 30 days; orange alarms correspond to winding looseness level 4 or an insulation aging index between 0.7 and 0.85, requiring repair within 15 days; red alarms correspond to winding looseness level 5 or an insulation aging index exceeding 0.85 or an estimated remaining life of less than 5 years, requiring immediate initiation of transformer replacement or overhaul procedures. After alarm work orders are generated, they are automatically dispatched to the responsible work team, and the system tracks the work order status in real time from dispatched, confirmed, processing, to closed loop.

[0040] A method for assessing the health status of transformer windings based on current fingerprint characteristics includes the following steps: Step 1: Event-triggered transient current waveform acquisition.

[0041] To extract features of the transformer winding state, it is first necessary to capture the complete current waveform when a transient event occurs. Therefore, this step involves connecting a wideband Rogowski coil current sensor in series in the secondary circuit of the bushing current transformer on the high-voltage or low-voltage side of the transformer, and setting up dual-threshold triggering logic.

[0042] The acquisition module continuously monitors the effective current value. When the effective current value suddenly increases to more than 150% of the rated current, it is determined to be an inrush current event. When the peak current value exceeds 8 times the rated current and the duration is less than 200 milliseconds, it is determined to be a short circuit event. The trigger detection adopts a dual threshold comparison logic: first, the real-time effective current value is calculated and compared with the first threshold (150% of the rated current). If it exceeds the threshold, the trigger judgment state is entered. Then, it checks whether the current peak envelope exceeds the second threshold (8 times the rated current) and the duration is less than 200 milliseconds. If the conditions are met, the trigger is confirmed and the high-speed acquisition window is started.

[0043] The acquisition window duration is set to 800ms, including a 100ms pre-trigger duration, a sampling rate of 100kS / s, and 80,000 sampling points. The output signal from the Rogowski coil sensor is amplified, filtered, and impedance matched by a signal conditioning circuit before being sent to a 16-bit analog-to-digital converter for digital processing. This triggering acquisition mechanism ensures that the complete transient current waveform is automatically captured each time the transformer experiences inrush current or short-circuit impact, providing a data foundation for subsequent analysis.

[0044] Step 2: Preprocessing of multi-stage cascaded signals.

[0045] The raw waveform acquired in Step 1 contains interference components such as DC offset, random noise, and channel zero-point drift, which directly affect the accuracy of subsequent feature extraction. Therefore, this step performs cascaded preprocessing on the raw signal before feature extraction, specifically including four sequentially executed steps.

[0046] Remove the DC component, calculate the arithmetic mean of the sampled sequence as the DC offset, and then subtract this offset value from each sample point to make the signal mean zero.

[0047] Wavelet thresholding denoising uses a 5-level decomposition with the db8 wavelet to effectively suppress random noise. Wavelet decomposition divides the signal into low-frequency approximation coefficients and high-frequency detail coefficients. The high-frequency coefficients contain noise components. The decomposed high-frequency coefficients are then processed using a soft thresholding function, the expression of which is: ; in, These are wavelet coefficients. For threshold parameters, The sign function is used. The threshold selection employs an unbiased likelihood estimation criterion, which is based on an adaptive threshold selection method that minimizes the risk function. The optimal threshold is calculated for the high-frequency coefficients of each wavelet decomposition layer. The threshold calculation process includes taking the absolute value of the wavelet coefficients and arranging them in ascending order, calculating the risk estimation vector, selecting the index position corresponding to the minimum risk, and multiplying the coefficient value at that index position by the maximum coefficient as the threshold for that layer. After wavelet thresholding and denoising, the processed low-frequency approximation coefficients and high-frequency detail coefficients are reconstructed using wavelet reconstruction to recover the denoised time-domain signal.

[0048] Baseline correction identifies the stable interval of the first 50ms of the waveform, calculates the average sampling value of this interval as the zero-point offset, and subtracts this zero-point offset from the full waveform data to eliminate the zero-point offset of the acquisition channel.

[0049] Resampling processing involves resampling the denoised signal to a standard sampling rate of 50 kS / s to unify the input format for subsequent feature extraction. The resampling employs an algorithm combining an anti-aliasing low-pass filter and decimation / interpolation to ensure signal integrity. The preprocessed current waveform data is stored in a standardized format, including fields such as timestamp, sampling rate, number of sampling points, and waveform data array.

[0050] The above four steps are executed sequentially, with the output of the previous step serving as the input of the next step, gradually purifying various interferences in the signal and providing high-quality waveform data for feature extraction.

[0051] Step 3: Extract multi-dimensional feature parameters from the preprocessed current waveform data.

[0052] The preprocessed current waveform contains multi-dimensional information reflecting the health status of the winding. This step extracts feature parameters from three dimensions: time domain, frequency domain, and time-frequency domain, respectively, to form the input feature vector for subsequent model evaluation. Time domain feature extraction includes the following four parameters: The peak current parameter is obtained by traversing the sampling sequence to identify the maximum absolute value; The rise time parameter is the time taken for the waveform to rise from the 10% peak value to the 90% peak value. First, determine the peak value of the waveform and its time. Then, search forward to find the time corresponding to the 10% peak value and the time corresponding to the 90% peak value. The difference between the two times is the rise time. The oscillation attenuation coefficient is obtained by fitting the exponential decay characteristic of the waveform envelope. The attenuation coefficient is calculated by taking the logarithm of the positive half of the waveform envelope and then performing a linear fit; the slope of the fitted line is the attenuation coefficient. The exponential decay model of the waveform envelope is expressed as: Normalization is performed with a time interval of 1 second, and the oscillation decay coefficient is defined as having a value of [value missing]. This value is dimensionless and ranges from (0,1). The waveform duration parameter is the time it takes for the waveform to decay from the starting point (10% peak position) to 5% of the peak value.

[0053] Frequency domain feature extraction includes the following three parameters: The frequency resolution of the Fast Fourier Transform is Where T is the sampling duration of 500ms, the transformation result extracts the amplitude values ​​of the frequency points corresponding to the fundamental wave to the 50th harmonic, the fundamental wave frequency is 50Hz (the rated frequency of the power grid), and the frequencies of each harmonic are n×50Hz (n=1,2,...,50). The fundamental wave amplitude is denoted as... ; The total harmonic distortion parameter is the ratio of the square root of the sum of the squares of the amplitudes of each harmonic to the amplitude of the fundamental frequency. The calculation formula is as follows:

[0054] The main harmonic frequency parameter is the frequency position corresponding to the harmonic component with the largest amplitude in the harmonic amplitude spectrum. The frequency value f = n×50Hz is obtained by searching for the maximum value of the 50th harmonic amplitude and recording its corresponding harmonic number n.

[0055] Time-frequency domain feature extraction employs wavelet packet decomposition to decompose the signal into multiple frequency bands to characterize the modulation effect of mechanical vibration caused by winding loosening on the high-frequency components of the current waveform. Wavelet packet decomposition uses the db10 wavelet basis, with a decomposition level of 5, decomposing the signal into 32 frequency bands, each with a bandwidth of [missing value]. ,in The standard sampling rate is used. After reconstructing the signal for each frequency band, the energy value of that band is calculated. The energy value is the sum of the squares of the reconstructed signal sampling points within the band. The energy proportion of the 32 frequency bands is calculated, which is the ratio of the energy of each frequency band to the total energy. The high-frequency energy proportion parameter is defined as the ratio of the sum of the energy of high-frequency bands (frequency ≥ 2kHz, corresponding to frequency band numbers 16 to 31) to the total signal energy. The calculation formula is:

[0056] The increased proportion of high-frequency energy reflects the enhanced modulation effect of mechanical vibration caused by winding loosening on the high-frequency components of the current waveform. Furthermore, as a supplement to the time-frequency domain features, this embodiment also extracts phase features for inrush current events, namely the phase angle between the inrush current and the system voltage. Simultaneously with current fingerprint acquisition, the acquisition module synchronously acquires the bus voltage as a reference phase signal. The voltage signal is coupled through a voltage transformer and then sent to an analog-to-digital converter for digital processing. The phase angle is calculated by first detecting the moment when the voltage crosses zero, then detecting the moment corresponding to the peak value of the inrush current, and calculating the time difference between the two moments. The phase angle is , where T is the power grid period of 20ms.

[0057] Step 4: Input the multi-dimensional feature parameters into the pre-trained deep neural network model to assess the health status.

[0058] The deep neural network model adopts a hybrid architecture design, consisting of three core modules cascaded together: a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism layer. The three modules process features in the following order: feature extraction, temporal modeling, and weighted key features. The one-dimensional convolutional neural network is responsible for extracting local temporal features, the bidirectional long short-term memory network is responsible for modeling temporal dependencies, and the attention mechanism is responsible for weighted fusion of key feature dimensions.

[0059] The one-dimensional convolutional neural network module comprises three convolutional layers to extract multi-scale local temporal features from preprocessed waveform data. The first convolutional layer contains 64 kernels with a kernel length of 16 sampling points and a stride of 1. It uses the same padding method to ensure the output length matches the input, and employs a rectified linear unit (CLU) activation function. This layer extracts basic local features from the original waveform. The second convolutional layer contains 128 kernels with a kernel length of 8 sampling points and a stride of 2. It also uses a CLU activation function and achieves layer-by-layer feature abstraction through downsampling. The third convolutional layer contains 256 kernels with a kernel length of 4 sampling points and a stride of 2. It also uses a CLU activation function and outputs a multi-scale local feature vector. After each convolutional layer, batch normalization and max pooling operations are performed sequentially. The batch normalization layer stabilizes the gradient distribution during training, while the max pooling layer has a pooling window length of 2 to reduce feature dimensionality and enhance feature robustness.

[0060] The bidirectional Long Short-Term Memory (LSTM) network module comprises two LSM layers, each containing 256 hidden units, used to model the temporal dependencies of waveform sequences. The LSM units employ a gating mechanism to selectively remember and forget long-term information. The gating structure includes three gates: a forget gate, an input gate, and an output gate. The forget gate determines how much information from the previous time step is discarded, the input gate determines how much of the current input information is added to the memory unit, and the output gate determines how much information from the memory unit state is output. The bidirectional structure means that sequence data is processed simultaneously by the forward and backward LSM networks. The forward LSM network processes data from the start to the end of the sequence, while the backward LSM network processes data from the end to the start. The outputs from both directions are concatenated and fused along the channel dimension. The output layer concatenates and fuses the forward and backward hidden states to obtain a fused feature vector with a dimension of 512. This bidirectional processing enables the model to simultaneously capture both forward and backward dependencies in the waveform sequence.

[0061] The attention mechanism layer calculates attention weights on the fused feature vectors to highlight key feature dimensions that significantly impact the evaluation results. The attention weight calculation employs a scaled dot product attention mechanism, which includes three vectors: a query vector Q, a key vector K, and a value vector V. First, a fully connected layer maps the fused feature vectors to query and key vectors, both with a dimension of 64. Then, the dot product of the query and key vectors is calculated and divided by [the matrix]. Scaling is performed, where =64 represents the dimension of the key vector. The scaling factor is used to prevent the dot product result from becoming too large, causing the softmax function to enter the saturation region. The scaled dot product result is normalized by the softmax function to obtain the attention weights. Each element in the weight vector represents the importance of the corresponding feature dimension. Finally, the attention weights are weighted and summed with the value vector to obtain the attention-weighted feature vector. This vector highlights the key feature dimensions while suppressing the interference of redundant features. The attention-weighted feature vector is flattened and then connected to the fully connected output layer. The output layer contains three branches, each corresponding to the predicted values ​​of the three evaluation metrics.

[0062] The model outputs three evaluation metrics: winding looseness level, insulation aging index, and remaining life prediction. The winding looseness level is a classification result, ranging from 1 to 5, where level 1 represents normal condition, level 2 represents slight looseness, level 3 represents moderate looseness, level 4 represents severe looseness, and level 5 represents extreme looseness requiring immediate repair. The insulation aging index is a regression result, ranging from 0 to 1, where 0 represents brand new insulation and 1 represents completely aged insulation requiring replacement. The remaining life prediction is a regression result in years, representing the estimated remaining service life of the transformer winding under current operating conditions. Model training employs a multi-task learning framework to simultaneously optimize both classification and regression tasks. The loss function is a weighted sum of the cross-entropy loss from the classification task and the mean squared error loss from the regression task, with weights set at 0.4 for classification loss and 0.6 for regression loss to balance the learning difficulty of the two tasks.

[0063] Example 2 This embodiment provides a specific implementation scheme for a multi-transformer health status management platform.

[0064] The cloud platform aggregates current fingerprint data from all connected transformers within its jurisdiction, and constructs a regional transformer health profile based on the aggregated data. The platform adopts a microservice architecture, with each functional module deployed independently and communicating decoupled through message queues. The message queues employ a high-throughput distributed messaging system to support the needs of large-scale concurrent data processing.

[0065] The platform's functional modules include the following five core modules.

[0066] The equipment ledger management module is responsible for maintaining a detailed equipment information ledger for all connected transformers. Ledger information includes basic parameters such as transformer model, rated capacity, rated voltage, and commissioning date; installation location including latitude and longitude coordinates and substation number; historical maintenance records including maintenance date, maintenance type, maintenance content, and maintenance personnel; and sensor configuration information including sensor model, installation location, and calibration parameters. Ledger data is stored in a relational database, and data in each table can be queried using a unique equipment code. It supports functions such as adding new equipment registrations, changing equipment information, retrieving historical records, and exporting ledgers.

[0067] The real-time monitoring dashboard module provides a visual interface displaying the health status of each transformer within the managed area. The dashboard automatically calculates a comprehensive health score for each transformer, ranging from 0 to 100 points. A higher score indicates a better condition. The comprehensive score is calculated by weighting three evaluation indicators: winding looseness level, insulation aging index, and remaining life estimate. The weighting ratios are: winding looseness level 0.4, insulation aging index 0.35, and remaining life estimate 0.25. The dashboard displays the current health status score and level classification for each transformer, categorized into five levels: Excellent (90-100 points), Good (80-89 points), Average (70-79 points), Poor (60-69 points), and Dangerous (below 60 points). The dashboard also displays historical trend curves of each transformer's score to reflect deterioration trends. Maintenance personnel can filter and view monitoring data for specific substations or specific transformer models, and the dashboard supports exporting evaluation reports.

[0068] The alarm management module receives alarm events reported by each edge terminal, highlights the number of currently unprocessed alarms on the dashboard, classifies and statistically analyzes alarms by alarm level and type, and provides access to view alarm details and work order processing. The alarm processing flow covers the entire lifecycle management from work order generation, work order dispatch, work order confirmation, work order processing to work order closure.

[0069] The model iteration and update module periodically collects operational feedback, adds confirmed degradation cases to the training dataset, and retrains the deep learning model to improve evaluation accuracy. The model training adopts an incremental learning strategy, fine-tuning the existing model parameters each time to adapt to the new data distribution. After training is completed, the updated model parameters are distributed to each edge terminal through a parameter synchronization mechanism.

[0070] The predictive maintenance recommendation module automatically generates annual maintenance plans based on the estimated remaining lifespan of each transformer, marking transformers requiring priority attention and their recommended maintenance times. Simultaneously, the module generates a regional transformer asset health analysis report, statistically analyzing the overall health status distribution of transformers within its jurisdiction, health degradation trends, and a list of transformers requiring priority maintenance, providing data support and decision-making assistance for power grid asset management.

[0071] To verify the effectiveness of the technical solution of the present invention, actual testing and verification were carried out on a distribution transformer based on Embodiment 1.

[0072] A distribution transformer, model S13-M-400 / 10, with a rated capacity of 400kVA, was used as the test object. The high-voltage side bushing current transformer had a transformation ratio of 600:5 and an accuracy class of 5P20. A Rogowski coil current sensor was installed on the secondary circuit of the high-voltage side bushing current transformer, ensuring the coil center was coaxial with the conductor to obtain optimal measurement accuracy. The sensor signal line was connected to the analog input interface of the intelligent fusion terminal in the distribution area, and the acquisition parameters were configured as follows: the trigger threshold was set to 150% of the rated current (corresponding to a secondary current of 5A), the acquisition window duration was set to 800ms, the pre-trigger duration was set to 100ms, and the sampling rate was set to 100kS / s. After configuration, the communication link was debugged to confirm that the data could be uploaded to the cloud platform.

[0073] During normal operation of the distribution transformer, the acquisition module continuously monitors the effective value of the current. When an inrush current event or a short-circuit impact current event is detected, a high-speed acquisition window is triggered. The acquisition module synchronously records the current waveform data for 100ms before and after the trigger, with a total of 80,000 sampling points. After acquisition, local signal preprocessing is performed: the mean of the sampling sequence is calculated as the DC offset for removal; denoising is performed using a 5-level decomposition of the db8 wavelet combined with a soft thresholding function; and zero-point offset is calculated and corrected for the stable interval of the first 50ms of the waveform's initial segment.

[0074] The preprocessed waveform data is fed into the feature extraction module to extract feature parameters in various dimensions. The time-domain feature extraction results under normal conditions are: peak current 7.2kA, rise time 6.5ms, oscillation attenuation coefficient 0.78, and waveform duration approximately 380ms. The frequency-domain feature extraction results are: total harmonic distortion 15.2%, with the main harmonic frequency located at 2kHz. The time-frequency domain feature extraction results are: high-frequency band energy proportion 8.3%. The above normal-state data is stored in the benchmark fingerprint database as the benchmark value for this type of transformer.

[0075] After the transformer experienced multiple short-circuit impacts, the system again collected the current fingerprint waveform. The time-domain feature extraction results were: peak current 6.8kA (reference value 7.2kA, deviation rate -5.6%), rise time 8.2ms (reference value 6.5ms, deviation rate +26.2%), oscillation attenuation coefficient 0.62 (reference value 0.78, deviation rate -20.5%), and waveform duration approximately 350ms. The frequency-domain feature extraction results were: total harmonic distortion 28.3% (reference value 15.2%, deviation rate +86.2%), with the main harmonic frequency located at 2kHz. The time-frequency domain feature extraction results were: high-frequency band energy proportion 18.7% (reference value 8.3%, deviation rate +125.3%). The phase feature extraction results were: the phase angle between the inrush current and the system voltage was 72 degrees.

[0076] The extracted multi-dimensional feature parameters are input into a pre-trained deep neural network model. The edge inference engine completes local inference within 2 seconds and outputs the following evaluation results: winding looseness level is 3 (indicating moderate looseness), insulation aging index is 0.58, and estimated remaining lifespan is 8.2 years. The system automatically compares the winding looseness level 3 with the alarm threshold level 2, determines that the transformer status is an alarm level requiring attention, automatically generates an alarm work order and pushes it to maintenance personnel. The alarm content includes equipment information, alarm time, evaluation results, and handling suggestions. The handling suggestion is to perform winding tightening maintenance on the transformer within 30 days.

[0077] After maintenance personnel completed the overhaul, the system again collected the current fingerprint waveform of the transformer and extracted its features. The feature parameters after the overhaul were compared and analyzed with those before the overhaul. The feature parameters after the overhaul are as follows: peak current deviation rate improved from -5.6% to -1.2%, rise time deviation rate improved from +26.2% to +5.8%, oscillation attenuation coefficient deviation rate improved from -20.5% to -3.1%, total harmonic distortion deviation rate improved from +86.2% to +12.5%, and high-frequency energy ratio deviation rate improved from +125.3% to +15.2%. After the overhaul, the winding looseness level decreased from level 3 to level 1 (indicating a normal state), the insulation aging index improved from 0.58 to 0.52, and the remaining life estimate increased from 8.2 years to 10.5 years.

[0078] The above verification results show that the deviation rates of all characteristic parameters significantly improved after maintenance. The high-frequency energy ratio was the most sensitive to winding loosening (deviation rate decreased from +125.3% to +15.2%), followed by the oscillation attenuation coefficient (deviation rate improved from -20.5% to -3.1%), verifying the effectiveness of the multi-dimensional characteristic parameter system. The winding loosening level assessment results were consistent with the actual state changes before and after maintenance, verifying the accuracy of the deep learning model's evaluation. The system updated the data after maintenance to the benchmark fingerprint database, and the first data collected after maintenance was set as the new benchmark value for subsequent evaluation reference.

[0079] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for assessing the health status of transformer windings based on current fingerprint characteristics, characterized in that, Includes the following steps: Step 1: Connect a wideband Rogowski coil current sensor in series in the secondary circuit of the transformer bushing current transformer to collect the transformer excitation inrush current waveform or short-circuit impact current waveform. When the acquisition module detects an excitation inrush current event or a short-circuit impact current event, it triggers a high-speed acquisition window and continuously records transient current waveform data for a preset duration. Step 2: The acquired signal is filtered using an anti-aliasing filter to remove high-frequency noise interference, remove the DC component from the signal, perform wavelet threshold denoising to eliminate random noise, perform zero-point correction to eliminate zero-point offset of the acquisition channel, and obtain the preprocessed current waveform data. Step 3: Extract multi-dimensional feature parameters from the preprocessed current waveform data. These multi-dimensional feature parameters include three dimensions: time-domain features, frequency-domain features, and time-frequency-domain features. The extraction is performed as follows: Time-domain feature extraction is performed to obtain feature parameters that describe the time-domain morphology of the current waveform. The time-domain features include peak current, rise time, oscillation decay coefficient, and waveform duration. The peak current reflects the magnitude of the current surge; the rise time reflects the steepness of the current waveform; the oscillation decay coefficient reflects the decay rate of the free oscillation component of the current; and the waveform duration reflects the duration of the transient process. Frequency domain feature extraction is performed. The time domain waveform is converted into a spectrum by fast Fourier transform. The amplitude of harmonics from the fundamental wave to the preset order is extracted. The total harmonic distortion and the main harmonic frequency are calculated. The total harmonic distortion reflects the degree of waveform distortion, and the main harmonic frequency reflects the concentrated frequency band of harmonic energy. Time-frequency domain feature extraction is performed. The signal is decomposed into multiple frequency bands by wavelet packet decomposition. The energy distribution of each frequency band is extracted, and the high-frequency energy ratio is calculated. The high-frequency energy ratio reflects the proportion of high-frequency components in the total signal energy and is used to characterize the modulation effect of mechanical vibration caused by winding loosening on the current waveform. Step 4: Input the multi-dimensional feature parameters into a pre-trained deep neural network model for health status assessment. The deep neural network model includes a one-dimensional convolutional neural network module, a bidirectional long short-term memory network module, and an attention mechanism module. The one-dimensional convolutional neural network module is used to extract local time-domain features of the current waveform. The bidirectional long short-term memory network module is used to model the temporal dependencies of the waveform sequence. The attention mechanism module is used to perform weighted fusion of key feature dimensions. The model outputs the health status assessment result of the transformer winding.

2. The method for assessing the health status of transformer windings based on current fingerprint characteristics according to claim 1, characterized in that, In step 1, the triggering condition for the excitation inrush current event is that the effective value of the current suddenly increases to more than a preset multiple of the rated current; the triggering condition for the short-circuit impact current event is that the peak current is detected to exceed a preset multiple of the rated current and the duration is less than a preset time threshold.

3. The method for assessing the health status of transformer windings based on current fingerprint characteristics according to claim 1, characterized in that, In step 2, the wavelet threshold denoising process uses a soft threshold function, the number of wavelet decomposition layers and the wavelet basis function are preset values, the threshold selection adopts the unbiased likelihood estimation criterion, and the threshold is calculated for the high-frequency coefficients of each wavelet decomposition layer; the zero-point correction calculates the zero-point offset and performs baseline correction by identifying the stationary interval of the waveform's initial segment.

4. The method for assessing the health status of transformer windings based on current fingerprint characteristics according to claim 1, characterized in that, In step 3, the time-domain features include at least one of peak current, rise time, oscillation attenuation coefficient, and waveform duration; the frequency-domain features include at least one of the fundamental to preset harmonic amplitude, total harmonic distortion, and main harmonic frequency extracted by fast Fourier transform; and the time-frequency domain features include at least one of the energy distribution of each frequency band, high-frequency energy ratio, and phase angle between excitation inrush current and system voltage extracted by wavelet packet decomposition.

5. The method for assessing the health status of transformer windings based on current fingerprint characteristics according to claim 1, characterized in that, In step 4, the one-dimensional convolutional neural network module contains three convolutional layers, each using a modified linear unit activation function, followed by a batch normalization layer and a max pooling layer; the bidirectional long short-term memory network module contains two long short-term memory layers, each containing 256 hidden units, with forward and backward long short-term memory units processing sequential data independently, and the output layer concatenating and fusing the forward and backward hidden states; the attention mechanism module uses a scaled dot product attention mechanism, highlighting key feature dimensions that significantly affect the evaluation results through weighted summation.

6. The method for assessing the health status of transformer windings based on current fingerprint characteristics according to claim 1, characterized in that, In step 4, the health status assessment results include three assessment indicators: winding looseness level, insulation aging index, and remaining life prediction. The winding looseness level is a graded output from 1 to 5, and the insulation aging index and the remaining life prediction are regression outputs.

7. A transformer winding health status assessment system based on current fingerprint characteristics, characterized in that, include: The current fingerprint acquisition subsystem is used to acquire transient current waveform data generated during transformer operation. The current fingerprint acquisition subsystem includes a wideband Rogowski coil current sensor, a signal conditioning circuit, and an analog-to-digital conversion module, which are connected in series in the secondary circuit of the transformer bushing current transformer. The signal preprocessing module is used to perform anti-aliasing filtering, DC component removal, wavelet threshold denoising, and zero-point correction on the transient current waveform data, and output the preprocessed current waveform data. The feature extraction module is used to extract multi-dimensional feature parameters from the preprocessed current waveform data. The multi-dimensional feature parameters include time-domain features, frequency-domain features, and time-frequency-domain features. The deep neural network evaluation module includes a one-dimensional convolutional neural network submodule, a bidirectional long short-term memory network submodule, and an attention mechanism submodule, which are used to evaluate the health status of the transformer windings based on the multi-dimensional feature parameters. The output module is used to output the health status assessment results generated by the deep neural network assessment module.

8. The transformer winding health status assessment system based on current fingerprint characteristics according to claim 7, characterized in that, It also includes a cloud-edge collaborative architecture: edge computing terminals are deployed on the intelligent fusion terminals in the transformer substation to complete local data preprocessing and model inference, and complete the processing flow from raw waveform acquisition to evaluation result output within a 2-second response time threshold; the cloud platform is used to aggregate data from multiple transformers and perform global model training and parameter updates; a data synchronization channel is set between the edge computing terminals and the cloud platform to realize the periodic synchronization of model parameters, and the periodic synchronization period is 24 hours.

9. The transformer winding health status assessment system based on current fingerprint characteristics according to claim 7, characterized in that, It also includes a reference fingerprint library module, which stores the reference current fingerprint characteristics of the same type of transformer from the factory test. The reference current fingerprint characteristics include reference time domain feature parameters, reference frequency domain feature parameters, and reference time and frequency domain feature parameters. The reference fingerprint library module supports the comparison and analysis of the evaluation results with the reference values, and identifies the gradual change trend of the winding state by calculating the deviation rate.

10. The transformer winding health status assessment system based on current fingerprint characteristics according to claim 7, characterized in that, It also includes an alarm management module, which is used to automatically generate alarm work orders based on the health status assessment results. The alarm management module is set with multiple alarm levels, and different alarm levels correspond to different maintenance strategies.

Images