Transformer fault prediction method based on frequency domain and time domain combined feature extraction
By combining frequency and time domain feature extraction methods with swept frequency response and transient current and voltage data, and utilizing principal component analysis and random forest models, the problem of identifying and locating early-stage weak faults in transformers was solved, achieving efficient fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-03-24
AI Technical Summary
Existing transformer fault diagnosis methods are not sensitive to early and minor faults, and it is difficult to identify multiple fault types and locate faults simultaneously. In particular, they are not effective in diagnosing inter-turn short circuits and winding deformations in 110kV oil-immersed transformers.
By employing a joint feature extraction method in the frequency and time domains, combined with swept frequency response curves and transient voltage and current waveform data, and through principal component analysis and random forest machine learning models, we can achieve joint identification, severity assessment, and localization of inter-turn short circuits and winding deformations.
It improves the sensitivity to early and subtle faults, realizes the integrated identification and localization of fault types, and enhances diagnostic accuracy and generalization ability.
Smart Images
Figure CN121723367A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of power equipment fault diagnosis, and in particular to a 110kV oil-immersed transformer inter-turn short circuit and winding deformation fault prediction method based on joint feature extraction in frequency domain and time domain. BACKGROUND
[0002] As the core equipment of the power system, the operation state of the power transformer is directly related to the safety and stability of the power grid. The oil-immersed transformer is widely used in the 110kV and above voltage level power grid due to its good insulation performance and large capacity. However, due to long-term operation in a complex electromagnetic, mechanical and thermal stress environment, the transformer winding is prone to inter-turn short circuit and deformation faults. If these faults cannot be detected and handled in time, they may develop into serious insulation breakdown accidents, leading to large-scale power outages.
[0003] At present, the transformer fault diagnosis methods mainly include the following categories:
[0004] (1) Method based on electrical parameters: such as winding current, impedance measurement, etc., but these methods are not sensitive to early weak inter-turn short circuits, especially when the number of short-circuit turns is less than 5%, the detection effect is poor.
[0005] (2) Method based on dissolved gas analysis (DGA) in oil: diagnose the fault type by analyzing the characteristic gases dissolved in the transformer oil, but this method cannot locate the fault and the response to early mechanical faults is lagging.
[0006] (3) Method based on frequency response analysis (FRA): diagnose winding deformation by detecting the frequency response changes of the transformer winding, but this method mainly focuses on structural faults and has limited ability to diagnose inter-turn short circuits, and is easily affected by external interference.
[0007] (4) Method based on temperature monitoring: assess the insulation state by monitoring the temperature changes of the transformer hot spots, but the response to sudden faults is not timely.
[0008] (5) Method based on single time domain or frequency domain signal: the existing technology only uses time domain signals, and the early fault feature extraction is not sufficient.
[0009] The main defects of the existing technology include: ① Most methods rely on a single type of signal and cannot capture the changes in transformer structure and electromagnetic characteristics simultaneously; ② Not sensitive to early latent faults, especially weak inter-turn short circuits and initial winding deformation; ③ Difficult to identify multiple fault types and locate the fault position simultaneously; ④ The fault diagnosis accuracy needs to be improved, especially the generalization ability under the condition of sample imbalance.
[0010] In recent years, some researchers have tried to fuse multi-source information for transformer fault diagnosis. These methods still do not fully combine the frequency domain and time domain features to jointly analyze the transformer structure and electromagnetic characteristics, and the diagnosis ability for composite faults is limited. SUMMARY
[0011] In order to solve the existing problems, the present application provides a transformer fault prediction method based on frequency domain and time domain joint feature extraction, specifically provides a 110kV oil-immersed transformer inter-turn short circuit and winding deformation fault prediction method based on frequency domain and time domain joint feature extraction, and the specific scheme is as follows:
[0012] The transformer fault prediction method based on frequency domain and time domain joint feature extraction comprises the following steps:
[0013] S1: synchronously collecting the frequency response curve and transient voltage and current wave recording data of the transformer as the basic data source;
[0014] S2: performing multi-dimensional feature extraction in parallel, including extracting structure sensitivity features from the frequency domain curve, and extracting electromagnetic sensitivity features from the time domain waveform;
[0015] S3: fusing the above heterogeneous features to form a high-dimensional joint feature vector, and performing dimension reduction and feature enhancement by using principal component analysis method;
[0016] S4: inputting the fusion feature vector into a random forest machine learning model trained by a large number of simulations and historical cases, and completing the joint identification, severity evaluation and positioning in the winding space position of the inter-turn short circuit and winding deformation fault by the model.
[0017] Preferably, the step S1 of synchronously collecting the frequency response curve and transient voltage and current wave recording data specifically comprises:
[0018] S11: using a frequency sweep impedance analyzer to apply a constant amplitude sinusoidal voltage excitation signal in a logarithmic step manner in the frequency range of 1kHz-1MHz, measuring the voltage and current response of the transformer winding end, and calculating the frequency response curve;
[0019] S12: using a high-frequency recording device to synchronously collect the three-phase voltage and current transient waveform of the transformer during normal operation or sudden short circuit event, and the sampling rate is not less than 100kHz.
[0020] Preferably, the step S2 of extracting structure sensitivity features from the frequency domain curve comprises:
[0021] S21: resonance point frequency shift and amplitude change feature extraction: identifying the main resonance points and amplitude in the frequency response curve, comparing with the transformer health benchmark curve, and calculating the resonance frequency shift and amplitude change rate;
[0022] S22: Sub-band energy distribution feature extraction: Divide the 1kHz-1MHz frequency band into multiple sub-bands, calculate the energy proportion of the frequency response curve in each sub-band, and form a sub-band energy distribution vector.
[0023] Preferably, the extraction of electromagnetic susceptibility features from the time-domain waveform in step S2 includes:
[0024] S23: Electrical symmetry feature extraction: Calculate the amplitude deviation and phase asymmetry of the three-phase current and voltage of the transformer;
[0025] S24: Differential Current Characteristic Extraction: Calculate the ratio of differential current value to braking current in the transformer differential protection circuit, and extract the harmonic content characteristics in the differential current waveform;
[0026] S25: High-frequency transient energy extraction: Perform wavelet packet transform on transient current recording data to extract the energy distribution of high-frequency transient components in each frequency band.
[0027] Preferably, the feature fusion and dimensionality reduction processing in step S3 includes:
[0028] S31: Combine the frequency domain structural sensitivity features and the time domain electromagnetic sensitivity features into a high-dimensional joint feature vector according to the weights;
[0029] S32: Principal component analysis is used to reduce the dimensionality of the high-dimensional joint feature vector, retaining the top k principal components with the highest contribution rates to form the dimensionality-reduced fused feature vector.
[0030] Preferably, the random forest machine learning model training in step S4 includes:
[0031] S41: Based on digital twin technology, establish a digital twin of the physical entity of the transformer, and use multiphysics simulation to deduce the variation law of the electrothermal characteristic parameters of the digital twin under different operating sections and different inter-turn fault conditions;
[0032] S42: Collect historical transformer fault case data and construct a sample library containing inter-turn short circuit location, short circuit turns ratio, winding deformation degree and corresponding feature vectors;
[0033] S43: The random forest algorithm is used to train a multi-task classifier with fused feature vectors as input and fault type, severity and location as output.
[0034] S44: Optimize the hyperparameters of random forests through grid search algorithm to improve the generalization ability of fault diagnosis models.
[0035] Preferably, the joint fault identification and localization in step S4 includes:
[0036] S45: Fault type identification: After processing the real-time acquired transformer data through steps S1-S3, input it into the trained random forest model and output the identification results of inter-turn short circuit, winding deformation or compound fault.
[0037] S46: Fault Severity Assessment: Based on the probability distribution output by the model, combined with regression analysis, assess the short-circuit turns ratio or the degree of winding deformation of the inter-turn short circuit.
[0038] S47: Fault location: By analyzing the resonant point distribution pattern in the frequency domain feature vector and the electromagnetic transient propagation characteristics in the time domain feature vector, the spatial location of the fault in the winding is determined.
[0039] Preferably, it also includes visualization and early warning functions for fault diagnosis results:
[0040] S5: Build a transformer condition monitoring dashboard to display fault identification results, severity assessment and location information in real time;
[0041] S6: Generates different levels of early warning signals based on the severity of the fault and pushes them to maintenance personnel.
[0042] The present invention also discloses an electronic device and a computer-readable storage medium, wherein the medium includes a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that the processor, when executing the program, implements any of the methods described above. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods described above.
[0043] The beneficial effects of this invention are as follows:
[0044] 1. Multi-source feature fusion enhances the comprehensiveness of fault identification: By simultaneously acquiring frequency domain and time domain data and extracting structural sensitivity and electromagnetic sensitivity features, the limitations of single signal type analysis are overcome, and the transformer status can be reflected more comprehensively.
[0045] 2. High sensitivity to early and weak faults: The frequency shift characteristics of the frequency domain resonant point are sensitive to winding deformation, and the time domain differential current characteristics and high-frequency transient energy are sensitive to inter-turn short circuits. The combination of the two can detect early and weak faults that are difficult to identify by traditional methods.
[0046] 3. Achieve integrated fault type identification and location: By training a random forest multi-task classifier, it can simultaneously complete fault type identification, severity assessment and location, providing more comprehensive fault information.
[0047] 4. High diagnostic accuracy: Based on digital twin technology and historical cases, a large-scale sample library is built. Combined with principal component dimensionality reduction and random forest algorithm, the overall fault diagnosis accuracy is significantly higher than that of traditional electrothermal characteristic fusion methods and single signal analysis methods.
[0048] 5. Strong generalization ability: Principal component analysis for dimensionality reduction and grid search optimization are used to ensure that the model maintains good performance under conditions of small sample size and imbalanced datasets, thus solving the problem of difficult sample collection for transformer faults. Attached Figure Description
[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0050] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation
[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.
[0052] like Figure 1 This invention discloses a transformer fault prediction method based on joint feature extraction in the frequency and time domains. Specifically, it discloses a method for predicting inter-turn short circuit and winding deformation faults in a 110kV oil-immersed transformer based on joint feature extraction in the frequency and time domains, comprising the following steps:
[0053] S1: Synchronously acquire the frequency sweep response curve and transient voltage and current waveform data of the transformer as the basic data source.
[0054] In this step, two types of data need to be collected simultaneously: frequency domain sweep response curves and time domain transient voltage and current waveform data.
[0055] In specific implementation, step S11 includes: using a swept-frequency impedance analyzer (such as Keysight E4990A) to apply a sinusoidal voltage excitation signal of constant amplitude (such as 10V) in a logarithmic step manner within a frequency range of 1kHz-1MHz, measuring the voltage and current response at the ends of the transformer winding, and calculating the frequency response curve. This frequency response curve reflects the physical structural characteristics of the transformer winding and is very sensitive to winding deformation and displacement.
[0056] Step S12 includes: synchronously acquiring the three-phase voltage and current transient waveforms of the transformer during normal operation or a sudden short-circuit event using a high-frequency recording device (such as a Power Quality Recorder), with a sampling rate of not less than 100kHz to ensure that high-frequency transient components can be captured. These time-domain waveforms reflect changes in the electromagnetic characteristics of the transformer and are extremely sensitive to faults such as inter-turn short circuits.
[0057] S2: Parallel multidimensional feature extraction
[0058] Extracting structural sensitivity features from frequency domain curves, including:
[0059] S21: Resonance Point Frequency Shift and Amplitude Change Feature Extraction. Identify the main resonant points and amplitudes in the frequency response curve, compare them with the transformer health baseline curve, and calculate the resonant frequency shift and amplitude change rate. When the winding undergoes deformation or an inter-turn short circuit, the distributed parameters (inductance, capacitance) of the winding will change, leading to changes in the resonant frequency and amplitude.
[0060] S22: Sub-band energy distribution feature extraction. The 1kHz-1MHz frequency band is divided into multiple sub-bands (e.g., low-frequency band 1k-10kHz, mid-frequency band 10k-100kHz, high-frequency band 100k-1MHz), and the energy proportion of the frequency response curve within each sub-band is calculated to form a sub-band energy distribution vector. Faults at different locations will cause changes in the energy distribution of different frequency bands.
[0061] Extracting electromagnetic susceptibility features from time-domain waveforms, including:
[0062] S23: Electrical Symmetry Feature Extraction. Calculate the amplitude deviation and phase asymmetry of the three-phase current and voltage of the transformer. Under normal circumstances, the three-phase electrical quantities of the transformer should be basically symmetrical; however, this symmetry is disrupted when an inter-turn short circuit occurs.
[0063] S24: Differential Current Characteristic Extraction. Calculate the ratio of differential current to braking current in the transformer differential protection circuit, and extract the harmonic content characteristics of the differential current waveform. Inter-turn short circuits will cause an increase in differential current and changes in harmonic content.
[0064] S25: High-frequency transient energy extraction. Wavelet packet transform is performed on transient current recording data to extract the energy distribution of high-frequency transient components (e.g., above 10kHz) in each frequency band. Inter-turn short circuits generate specific high-frequency electromagnetic transient phenomena.
[0065] S3: Feature Fusion and Dimensionality Reduction
[0066] This step includes:
[0067] S31: Combine the frequency domain structural sensitivity features (resonance point frequency shift and amplitude variation, sub-band energy distribution) and the time domain electromagnetic sensitivity features (electrical symmetry, differential current characteristics, high-frequency transient energy) into a high-dimensional joint feature vector according to weights. The initial feature dimension may reach tens or even hundreds of dimensions.
[0068] S32: Principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional joint feature vector, retaining the top k principal components with the highest contribution rates to form a dimensionality-reduced fused feature vector. This reduces data redundancy while retaining most of the effective information, improving the efficiency and quality of subsequent model training.
[0069] S4: Random Forest Model Training and Fault Diagnosis
[0070] This step includes two parts: model training and fault diagnosis.
[0071] Model training phase:
[0072] S41: A digital twin of the transformer's physical entity is established based on digital twin technology. Multiphysics simulation (such as COMSOL Multiphysics) is used to deduce the variation law of electrothermal characteristic parameters of the digital twin under different operating sections and different inter-turn fault conditions. A large number of fault samples are generated through simulation to solve the problem of scarce real fault samples.
[0073] S42: Collect historical transformer fault case data and construct a sample library containing inter-turn short circuit location, short circuit turns ratio, winding deformation degree and corresponding feature vectors.
[0074] S43: The Random Forest algorithm is used to train a multi-task classifier, taking the fused feature vector as input and the fault type, severity, and location as output. Random Forest is an ensemble learning algorithm that constructs multiple decision trees and combines their prediction results, offering advantages such as fast training speed and strong resistance to overfitting.
[0075] S44: Optimize the hyperparameters of the random forest (such as the number of decision trees, maximum depth, minimum number of samples per node) through a grid search algorithm, evaluate the performance of different parameter combinations using cross-validation, select the optimal parameters, and improve the generalization ability of the fault diagnosis model.
[0076] Fault diagnosis phase:
[0077] S45: Fault type identification: After processing the real-time acquired transformer data through steps S1-S3, input it into the trained random forest model and output the identification results of inter-turn short circuit, winding deformation or compound fault.
[0078] S46: Fault Severity Assessment: Based on the probability distribution output by the model, combined with regression analysis, assess the degree of short-circuit turns ratio or winding deformation of inter-turn short circuits. For example, the severity of short circuits is divided into three levels: mild (turns ratio <2%), moderate (2%-5%), and severe (>5%).
[0079] S47: Fault location: By analyzing the resonant point distribution pattern in the frequency domain characteristic vector and the electromagnetic transient propagation characteristics in the time domain characteristic vector, the spatial location of the fault in the winding (such as the upper, middle or lower part of the high voltage winding) is determined.
[0080] S5 and S6: Visualization and Early Warning of Fault Diagnosis Results
[0081] A transformer condition monitoring dashboard is built to display fault identification results, severity assessment, and location information in real time. Different levels of warning signals (e.g., Normal - Green, Caution - Yellow, Abnormal - Orange, Danger - Red) are generated based on the severity of the fault and pushed to maintenance personnel.
[0082] The present invention also discloses a computer-readable storage medium and a computer system. The medium stores a computer program, which, upon execution, performs the method described above. A computer system includes a processor and a storage medium. The storage medium stores a computer program, and the processor reads and runs the computer program from the storage medium to perform the method described above.
[0083] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.
[0084] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A transformer fault prediction method based on joint frequency-domain and time-domain feature extraction, characterized in that, Includes the following steps: S1: Synchronously acquire the frequency sweep response curve and transient voltage and current recording data of the transformer as the basic data source; S2: Perform multi-dimensional feature extraction in parallel, including extracting structural sensitivity features from frequency domain curves and electromagnetic sensitivity features from time domain waveforms; S3: The above heterogeneous features are fused to construct a high-dimensional joint feature vector, and principal component analysis is used for dimensionality reduction and feature enhancement. S4: Input the fused feature vector into a random forest machine learning model that has been pre-trained through a large number of simulations and historical cases. The model then performs joint identification of inter-turn short circuits and winding deformation faults, severity assessment, and location of them in the winding space.
2. The method according to claim 1, characterized in that, The specific steps in step S1 for synchronously acquiring the swept frequency response curve and transient voltage and current waveform data include: S11: A sweep frequency impedance analyzer is used to apply a sinusoidal voltage excitation signal with a constant amplitude in a logarithmic step manner within the frequency range of 1kHz-1MHz. The voltage and current responses at the ends of the transformer windings are measured, and the frequency response curve is calculated. S12: Use a high-frequency waveform recording device to synchronously acquire the transient waveforms of three-phase voltage and current of the transformer during normal operation or a sudden short-circuit event, with a sampling rate of not less than 100kHz.
3. The method according to claim 1, characterized in that, The extraction of structural sensitivity features from the frequency domain curve in step S2 includes: S21: Resonance point frequency shift and amplitude change feature extraction: Identify the main resonant points and amplitudes in the frequency response curve, compare them with the transformer health reference curve, and calculate the resonant frequency shift and amplitude change rate. S22: Sub-band energy distribution feature extraction: Divide the 1kHz-1MHz frequency band into multiple sub-bands, calculate the energy proportion of the frequency response curve in each sub-band, and form a sub-band energy distribution vector.
4. The method according to claim 1, characterized in that, The step S2, which involves extracting electromagnetic susceptibility features from the time-domain waveform, includes: S23: Electrical symmetry feature extraction: Calculate the amplitude deviation and phase asymmetry of the three-phase current and voltage of the transformer; S24: Differential Current Characteristic Extraction: Calculate the ratio of differential current value to braking current in the transformer differential protection circuit, and extract the harmonic content characteristics in the differential current waveform; S25: High-frequency transient energy extraction: Perform wavelet packet transform on transient current recording data to extract the energy distribution of high-frequency transient components in each frequency band.
5. The method according to claim 1, characterized in that, The feature fusion and dimensionality reduction processing in step S3 includes: S31: Combine the frequency domain structural sensitivity features and the time domain electromagnetic sensitivity features into a high-dimensional joint feature vector according to the weights; S32: Principal component analysis is used to reduce the dimensionality of the high-dimensional joint feature vector, retaining the top k principal components with the highest contribution rates to form the dimensionality-reduced fused feature vector.
6. The method according to claim 1, characterized in that, The random forest machine learning model training in step S4 includes: S41: Based on digital twin technology, establish a digital twin of the physical entity of the transformer, and use multiphysics simulation to deduce the variation law of the electrothermal characteristic parameters of the digital twin under different operating sections and different inter-turn fault conditions; S42: Collect historical transformer fault case data and construct a sample library containing inter-turn short circuit location, short circuit turns ratio, winding deformation degree and corresponding feature vectors; S43: The random forest algorithm is used to train a multi-task classifier with fused feature vectors as input and fault type, severity and location as output. S44: Optimize the hyperparameters of random forests through grid search algorithm to improve the generalization ability of fault diagnosis models.
7. The method according to claim 6, characterized in that, The joint fault identification and localization in step S4 includes: S45: Fault type identification: After processing the real-time acquired transformer data through steps S1-S3, input it into the trained random forest model and output the identification results of inter-turn short circuit, winding deformation or compound fault. S46: Fault Severity Assessment: Based on the probability distribution output by the model, combined with regression analysis, assess the short-circuit turns ratio or the degree of winding deformation of the inter-turn short circuit. S47: Fault location: By analyzing the resonant point distribution pattern in the frequency domain feature vector and the electromagnetic transient propagation characteristics in the time domain feature vector, the spatial location of the fault in the winding is determined.
8. The method according to claim 1, characterized in that, It also includes visualization and early warning functions for fault diagnosis results: S5: Build a transformer condition monitoring dashboard to display fault identification results, severity assessment and location information in real time; S6: Generates different levels of early warning signals based on the severity of the fault and pushes them to maintenance personnel.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.