A transformer health state evaluation method and system based on double-source time-frequency characteristics

Abstract

A method and system for assessing the health status of transformers based on dual-source time-frequency features are disclosed. The method acquires vibration and acoustic emission signals from power transformers and preprocesses the acquired signals. Then, time-domain and frequency-domain features are extracted to construct a multi-dimensional time-frequency feature set. A feature decoupling model based on a generative adversarial network (GAN) is established, and corresponding constraints are set in conjunction with the known characteristics of power grid harmonics. Effective separation of power grid harmonic interference is achieved through adversarial iterative training. The decoupled time-frequency features are weighted and fused, and then input into a convolutional neural network (CNN) assessment model to complete health status classification and output the power transformer health status assessment level. This invention fully leverages the advantages of GANs in signal decoupling and the capabilities of CNNs in feature learning and classification, enabling accurate assessment of the health status of power transformers under power grid harmonic interference conditions, providing a reliable basis for power transformer operation, maintenance, and condition-based overhaul.

Description

Technical Field

[0001] This invention belongs to the field of power transformer health status assessment technology, and particularly relates to a transformer health status assessment method and system based on dual-source time-frequency characteristics. Background Technology

[0002] As a core hub in the power generation, transmission, and distribution chain, the operating status of power transformers directly determines the reliability, quality, and maintenance costs of the power grid. With the grid integration of new energy sources and the large-scale application of power electronic devices and nonlinear loads, harmonic pollution from the power grid is becoming increasingly severe. Harmonic currents in the power grid excite high-frequency alternating electromagnetic forces within the transformer, directly causing intensified magnetostriction of the core, winding deformation vibration, and tank resonance noise. These harmonic-derived signals, combined with vibration and acoustic emission signals caused by transformer faults, result in a significant decrease in the signal-to-noise ratio of the status signals, severely masking fault characteristics and interfering with health status assessments.

[0003] Current power transformer health status assessment technologies are mainly divided into two categories: traditional detection methods and data-driven methods. Neither can adequately address the precise assessment requirements under harmonic interference scenarios. Traditional methods require offline electrical testing, which not only interrupts grid power supply continuity and causes significant power generation losses, but also fails to capture potential faults under harmonic coupling during dynamic operation. Infrared thermal imaging and dissolved gas analysis in oil can only reflect local fault characteristics, showing low sensitivity to structural faults in the core and windings, and are unable to distinguish between fault signals and harmonic interference signals. Data-driven methods are gradually being adopted due to the advantages of online monitoring, but their reliance on a single signal source has inherent limitations: vibration monitoring is susceptible to interference from the superposition of harmonic electromagnetic forces and environmental vibrations, making it difficult to distinguish between false harmonic features and genuine fault characteristics; while acoustic emission monitoring can capture information on the evolution of internal defects, the strong interference from the superposition of environmental noise and harmonic noise leads to significant fluctuations in assessment accuracy under complex grid conditions, resulting in a high misjudgment rate.

[0004] In recent years, deep learning technology has provided a new direction for the condition assessment of power transformers. Although LSTM-based single-signal assessment models have improved feature extraction capabilities, their adaptability to harmonic interference scenarios remains insufficient. The core limitations are threefold: First, most models rely on a single data source, failing to utilize the complementary characteristics of vibration signals' sensitivity to structural damage and acoustic emission signals' timely response to internal defects. This results in incomplete fault characterization dimensions and difficulty in handling complex signal patterns under harmonic interference. Second, there is a lack of targeted interference decoupling mechanisms. For example, the patent document with publication number CN117688444A only suppresses interference through traditional methods such as low-pass filtering and wavelet denoising, failing to fundamentally separate harmonic interference from the equipment's own condition characteristics. When harmonic frequencies overlap with fault characteristic frequencies, misjudgments and omissions are highly likely. Third, existing dual-source signal fusion methods are mostly simple feature splicing. For example, the patent document with publication number CN121071444A does not perform spatiotemporal alignment and deep collaborative fusion, failing to uncover the intrinsic correlation between signals. Furthermore, the superposition of interference features may further reduce the reliability of the assessment.

[0005] The complexity of power grid harmonic interference and the heterogeneity of dual-source signals place higher demands on the health status assessment of power transformers. While fusing vibration and acoustic emission signals from both sources can achieve multi-dimensional complementarity of fault characteristics, it requires addressing two core issues: harmonic interference decoupling and signal synergistic fusion. GAN models, with their adversarial iterative mechanism, possess powerful feature decoupling capabilities, while CNN models exhibit excellent efficiency in feature classification. Combining these two approaches can specifically address the assessment challenges under harmonic interference. Therefore, there is an urgent need for an assessment method that integrates dual-source time-frequency features and accurately decouples harmonic interference to improve the accuracy and stability of power transformer health status assessment in complex scenarios, providing technical support for intelligent operation and maintenance of power systems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for assessing the health status of transformers based on dual-source time-frequency features. The health status assessment method of this invention includes the following steps: extracting the time-domain and frequency-domain features of power transformer vibration and acoustic emission signals to construct a multi-dimensional time-frequency feature set; using a GAN model to decouple the signals and separate the power grid harmonic interference; and utilizing the efficient classification capability of a CNN model to output the health status level of the power transformer.

[0007] The present invention adopts the following technical solution, and its specific steps are as follows: S1, through the sensor to collect vibration signals and acoustic emission signals of the power transformer, and preprocess the vibration signals and acoustic emission signals; S2, extract the time-domain and frequency-domain features of vibration and acoustic emission signals respectively, and perform spatiotemporal alignment processing to construct a multi-dimensional time-frequency feature set; S3. Input the multi-dimensional time-frequency feature set into the GAN-based decoupling model, set the constraints, perform iterative training, separate the power grid harmonic interference, and obtain the decoupled time-frequency features. S4 inputs the decoupled time-frequency features into a pre-defined CNN evaluation model after weighted fusion to complete the health status classification and outputs a quantified power transformer health status evaluation level.

[0008] More preferably, in S2, the time-domain and frequency-domain features of the vibration and acoustic emission signals are extracted respectively, specifically as follows: Temporal characteristics include root mean square, variance, standard deviation, skewness, and kurtosis; Frequency domain characteristics include dominant frequency amplitude, bandwidth, spectral entropy, spectral flatness, and spectral roll-off point.

[0009] More preferably, in S2, spatiotemporal alignment processing is performed to construct a multidimensional time-frequency feature set. Specific methods include: Time alignment is based on the timestamp of the vibration signal to calibrate the time axis of the acoustic emission signal; By establishing an internal coordinate system based on the location of the power transformer structure, spatial alignment is achieved by spatially associating the acquisition locations of the core and winding vibration signals with the sensor locations of the acoustic emission signals.

[0010] More preferably, in S3, the multi-dimensional time-frequency feature set is input into the GAN-based decoupling model, constraints are set, iterative training is performed, and the power grid harmonic interference is separated to obtain the decoupled time-frequency features, specifically including: The GAN-based decoupling model includes a generator and a discriminator; The generator adopts a convolutional neural network structure. The input is a multi-dimensional time-frequency feature set, and the output is the separated power grid harmonic interference features and the power transformer's own state features. The discriminator adopts a fully connected neural network structure. The separated power grid harmonic interference features and the power transformer's own state features are input into the discriminator, and the feature category judgment result is output.

[0011] More preferably, in S3, the constraints specifically include: Frequency constraints are set based on the fixed frequency characteristics of power grid harmonics; Amplitude constraints are set based on the distribution law that harmonic amplitude decreases with increasing frequency; The correlation coefficient between the decoupled power grid harmonic interference characteristics and the power transformer's own state characteristics is lower than the set threshold.

[0012] More preferably, in S3, iterative training is performed to separate the power grid harmonic interference to obtain the decoupled time-frequency characteristics. Specific steps include: Initialize the model parameters, and the generator outputs the characteristics of power grid harmonic interference and the state characteristics of the power transformer itself. The discriminator identifies and judges the grid harmonic interference characteristics output by the generator and the power transformer's own state characteristics, respectively. Update the parameters of the generator and discriminator using the backpropagation algorithm; Iterate through the training until the loss function converges, then stop training and save the model parameters.

[0013] More preferably, in S4, the step of weighted fusion includes: The objective weights of each time-frequency feature after decoupling are calculated using the entropy weight method. The subjective weights of each time-frequency feature in the decoupled system are determined using the analytic hierarchy process (AHP). The final comprehensive weight of the corresponding time-frequency feature is obtained by linearly summing the objective weight and the subjective weight; The decoupled time-frequency features are summed based on the final comprehensive weights to obtain the input of the CNN evaluation model.

[0014] More preferably, in S4, the CNN evaluation model specifically includes: The CNN evaluation model consists of convolutional layers, pooling layers, fully connected layers, and an output layer. Convolutional layers extract comprehensive fault features, pooling layers reduce feature dimensionality while retaining fault information, ReLU activation function is used for nonlinear fitting, fully connected layers integrate and summarize features and transform them into feature vectors suitable for classification decisions, and the output layer completes health status classification through the Softmax function.

[0015] This invention also proposes a transformer health status assessment method based on dual-source time-frequency characteristics, including a data acquisition and processing module, a dataset construction module, a feature decoupling module, and a health assessment module: The data acquisition and processing module collects vibration and acoustic emission signals from the power transformer through sensors and performs preprocessing on the vibration and acoustic emission signals. The dataset construction module extracts the time-domain and frequency-domain features of vibration and acoustic emission signals respectively, and performs spatiotemporal alignment processing to construct a multi-dimensional time-frequency feature set; The feature decoupling module inputs a multi-dimensional time-frequency feature set into a GAN-based decoupling model, sets constraints, performs iterative training, separates the power grid harmonic interference, and obtains the decoupled time-frequency features. The health assessment module takes the decoupled time-frequency features, performs weighted fusion, and inputs them into a preset CNN assessment model to complete the health status classification and output a quantitative assessment level of the power transformer's health status.

[0016] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0017] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a complementary and synergistic multidimensional time-frequency feature set by fusing vibration and acoustic emission signals from two sources and performing spatiotemporal alignment processing. This effectively overcomes the deficiency of incomplete information representation of a single signal source under complex working conditions and provides a more reliable data foundation for subsequent accurate evaluation.

[0019] 2. This invention introduces a decoupling model based on generative adversarial networks and sets constraints based on the harmonic characteristics of the power grid, thereby effectively separating harmonic interference components from aliased signals. This solves the core problem of inaccurate evaluation caused by the aliasing of harmonics and fault characteristics, and significantly improves the characteristic signal-to-noise ratio.

[0020] 3. This invention achieves more accurate and stable automated assessment and classification of the health status of power transformers under strong harmonic interference by intelligently weighting and fusing the decoupled pure features and using convolutional neural networks for deep pattern recognition. Attached Figure Description

[0021] Figure 1 This is a flowchart of a transformer health status assessment method based on dual-source time-frequency characteristics according to the present invention; Figure 2 This is a flowchart illustrating the operational process of a transformer health status assessment method based on dual-source time-frequency characteristics according to the present invention. Figure 3 This is a flowchart illustrating the signal preprocessing and multi-dimensional time-frequency feature set construction process of the present invention. Figure 4 This is a flowchart illustrating the business process of health status assessment using a GAN-based decoupling model and a CNN-based evaluation model, as described in this invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0023] like Figure 1 As shown, this invention proposes a transformer health status assessment method based on dual-source time-frequency characteristics, comprising: S1, through the sensor to collect vibration signals and acoustic emission signals of the power transformer, and preprocess the vibration signals and acoustic emission signals; S2, extract the time-domain and frequency-domain features of vibration and acoustic emission signals respectively, and perform spatiotemporal alignment processing to construct a multi-dimensional time-frequency feature set; In S2, the time-domain and frequency-domain features of the vibration and acoustic emission signals are extracted respectively, specifically as follows: Temporal characteristics include root mean square, variance, standard deviation, skewness, and kurtosis; Frequency domain characteristics include dominant frequency amplitude, bandwidth, spectral entropy, spectral flatness, and spectral roll-off point.

[0024] In S2, spatiotemporal alignment is performed to construct a multidimensional time-frequency feature set. Specific methods include: Time alignment is based on the timestamp of the vibration signal to calibrate the time axis of the acoustic emission signal; By establishing an internal coordinate system based on the location of the power transformer structure, spatial alignment is achieved by spatially associating the acquisition locations of the core and winding vibration signals with the sensor locations of the acoustic emission signals.

[0025] S3. Input the multi-dimensional time-frequency feature set into the GAN-based decoupling model, set the constraints, perform iterative training, separate the power grid harmonic interference, and obtain the decoupled time-frequency features. In S3, the multi-dimensional time-frequency feature set is input into the GAN-based decoupling model, constraints are set, and iterative training is performed to separate the power grid harmonic interference, obtaining the decoupled time-frequency features, specifically including: The GAN-based decoupling model includes a generator and a discriminator; The generator adopts a convolutional neural network structure. The input is a multi-dimensional time-frequency feature set, and the output is the separated power grid harmonic interference features and the power transformer's own state features. The discriminator adopts a fully connected neural network structure. The separated power grid harmonic interference features and the power transformer's own state features are input into the discriminator, and the feature category judgment result is output.

[0026] In S3, the constraints specifically include: Frequency constraints are set based on the fixed frequency characteristics of power grid harmonics; Amplitude constraints are set based on the distribution law that harmonic amplitude decreases with increasing frequency; The correlation coefficient between the decoupled power grid harmonic interference characteristics and the power transformer's own state characteristics is lower than the set threshold.

[0027] In S3, iterative training is performed to separate the harmonic interference from the power grid, obtaining the decoupled time-frequency characteristics. Specific steps include: Initialize the model parameters, and the generator outputs the characteristics of power grid harmonic interference and the state characteristics of the power transformer itself. The discriminator identifies and judges the grid harmonic interference characteristics output by the generator and the power transformer's own state characteristics, respectively. Update the parameters of the generator and discriminator using the backpropagation algorithm; Iterate through the training until the loss function converges, then stop training and save the model parameters.

[0028] S4 inputs the decoupled time-frequency features into a pre-defined CNN evaluation model after weighted fusion to complete the health status classification and outputs a quantified power transformer health status evaluation level.

[0029] In S4, the step of weighted fusion includes: The objective weights of each time-frequency feature after decoupling are calculated using the entropy weight method. The subjective weights of each time-frequency feature in the decoupled system are determined using the analytic hierarchy process (AHP). The final comprehensive weight of the corresponding time-frequency feature is obtained by linearly summing the objective weight and the subjective weight; The decoupled time-frequency features are summed based on the final comprehensive weights to obtain the input of the CNN evaluation model.

[0030] In S4, the CNN evaluation model specifically includes: The CNN evaluation model consists of convolutional layers, pooling layers, fully connected layers, and an output layer. Convolutional layers extract comprehensive fault features, pooling layers reduce feature dimensionality while retaining fault information, ReLU activation function is used for nonlinear fitting, fully connected layers integrate and summarize features and transform them into feature vectors suitable for classification decisions, and the output layer completes health status classification through the Softmax function.

[0031] This invention also proposes a transformer health status assessment method based on dual-source time-frequency characteristics, including a data acquisition and processing module, a dataset construction module, a feature decoupling module, and a health assessment module: The data acquisition and processing module collects vibration and acoustic emission signals from the power transformer through sensors and performs preprocessing on the vibration and acoustic emission signals. The dataset construction module extracts the time-domain and frequency-domain features of vibration and acoustic emission signals respectively, and performs spatiotemporal alignment processing to construct a multi-dimensional time-frequency feature set; The feature decoupling module inputs a multi-dimensional time-frequency feature set into a GAN-based decoupling model, sets constraints, performs iterative training, separates the power grid harmonic interference, and obtains the decoupled time-frequency features. The health assessment module takes the decoupled time-frequency features, performs weighted fusion, and inputs them into a preset CNN assessment model to complete the health status classification and output a quantitative assessment level of the power transformer's health status.

[0032] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0033] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0034] Example 1 This embodiment provides a transformer health status assessment method based on dual-source time-frequency characteristics, such as... Figure 2 As shown, the method includes the following steps: signal acquisition and preprocessing, extraction of time-frequency features of vibration and acoustic emission signals, construction of a multi-dimensional time-frequency feature set, spatiotemporal alignment processing, separation of power grid harmonic interference using a GAN-based decoupling model, and assessment of the health status of power transformers using a pre-set CNN model. The specific implementation plan is as follows: S1 uses sensors to collect vibration and acoustic emission signals from the power transformer and performs preprocessing on the vibration and acoustic emission signals.

[0035] like Figure 3 As shown, vibration signals of the power transformer are acquired by accelerometers located at the upper and lower yokes of the transformer core, the end plates of the windings, and the corresponding projection areas of the core and windings on the side wall of the oil tank. Acoustic emission signals of the power transformer are acquired by sensors located in the upper middle part of the side of the oil tank, the area near the center of the core on the top cover of the oil tank, and near the winding bushings. The signals are preprocessed, specifically including DC component removal, power frequency notch filtering, wavelet threshold denoising, and normalization of the vibration signals; and pre-amplification, bandpass filtering, time-domain segmentation, and normalization of the acoustic emission signals.

[0036] For DC component removal, since the frequency of the DC component is 0Hz, a high-pass filter is used. Its characteristic is that it allows high-frequency signals to pass while suppressing low-frequency and DC signals. A commonly used method is a first-order RC high-pass filter. The transfer function of a first-order RC high-pass filter in the analog domain is:

[0037] in, This is the resistance value. This is the capacitance value. Given a complex frequency, its discretization yields the digital domain transfer function:

[0038] in, These are the filter coefficients. This is the filter cutoff frequency. The sampling period.

[0039] The transfer function is transformed into a difference equation, and the processed signal is... The recursive formula is:

[0040] in, This is the filtered output from the previous time step. This is the input signal at the current moment.

[0041] For normalization, the Z-score normalization method is used:

[0042] in, These are the original data points. It is the mean of the sample data. It is the standard deviation of the sample data.

[0043] S2 extracts the time-domain and frequency-domain features of vibration and acoustic emission signals respectively, and performs spatiotemporal alignment processing to construct a multi-dimensional time-frequency feature set.

[0044] Time-domain characteristics: root mean square, variance, standard deviation, skewness, kurtosis; Frequency domain characteristics: dominant frequency amplitude, bandwidth, spectral entropy, spectral flatness, and spectral roll-off point.

[0045] First, based on the preprocessed vibration and acoustic emission dual-source signals, corresponding time-domain and frequency-domain feature parameters are extracted to form two types of single-source feature subsets. Time alignment is performed using the timestamp of the vibration signal as a reference to calibrate the time axis of the acoustic emission signal; spatial alignment is achieved by establishing a power transformer structural coordinate mapping model to spatially correlate the acquisition locations of the core and winding vibration signals with the sensor locations of the acoustic emission signals.

[0046] Specifically, the structural coordinate mapping model of a power transformer is established based on the physical design drawings or 3D model of the power transformer. This model uses a unified internal coordinate system defined by key components such as the transformer core and windings. Each vibration sensor and acoustic emission sensor has a defined coordinate within this system. During spatial alignment, based on the principle of proximity or projection relationship, signal features collected from different sensors but with similar coordinates (e.g., a vibration sensor at a point outside the tank and an acoustic emission sensor at a corresponding point inside the tank) are paired, treating them as observations of the same local area. This establishes a correlation at the feature level, providing a spatially consistent foundation for subsequent dual-source feature fusion.

[0047] As an improved implementation method, spatial alignment can also be achieved in the following ways: Methods include mapping based on topological connectivity, mapping based on signal propagation characteristics, and mapping based on data-driven learning.

[0048] S3 inputs a multi-dimensional time-frequency feature set into a GAN-based decoupling model, sets constraints based on the known characteristics of power grid harmonics, and uses the adversarial iterative training of the GAN model to separate the power grid harmonic interference.

[0049] like Figure 4 As shown, a GAN-based decoupling model is used to process the multi-dimensional time-frequency feature set, achieving the separation of power grid harmonic interference characteristics from the power transformer's own state characteristics. This model employs a generator (…). ) and discriminator ( Collaborative adversarial iterative architecture: The generator undertakes the core task of eigenvalue decomposition, receiving a multi-dimensional time-frequency feature set. Output candidate state features With candidate harmonic interference characteristics The following features are superimposed:

[0050] Based on the distribution characteristics of power grid harmonics in the dual-source time-frequency characteristics, a harmonic distribution constraint term is introduced. Avoid misinterpreting fault characteristics.

[0051] Specifically, frequency constraints are set based on the fixed frequency characteristics of power grid harmonics, limiting the frequency range of interference characteristics to include only harmonic frequencies and harmonics; amplitude constraints are set based on the distribution law that the amplitude of harmonics decreases with increasing frequency, so that the amplitude of the separated interference characteristics satisfies the exponential decay relationship; at the same time, it is ensured that the correlation coefficient between the decoupled harmonic interference characteristics and the state characteristics of the power transformer itself is less than 0.1, so as to achieve complete separation between the two.

[0052] The generator optimization objective, based on the harmonic distribution constraint, is:

[0053] in, For the generator's basic loss, To constrain the weights, This is used to measure the degree of fit between the candidate harmonic interference features and the actual harmonic distribution. The adversarial loss function commonly used in generative adversarial networks can be used as the generator's basic loss; for example, the cross-entropy loss form of the original GAN ​​can be adopted.

[0054] The discriminator adopts a dual-parallel structure, consisting of a harmonic discrimination branch. State feature discrimination branch Both branches output the probability of judging the true value of the sample. ,in Characterizes the probability that the candidate harmonic interference features are the actual power grid harmonic features. It represents the probability that the candidate state features are the actual device state features.

[0055] With multidimensional time-frequency feature set To prepare training data, a real power grid harmonic feature sample set was prepared simultaneously. Sample set of real device status features As a discrimination criterion, after initializing the generator and discriminator network structures and basic parameters, the iterative process is initiated, and the following operations are performed in each iteration: Step 1: The generator processes the input features Decompose the data to generate candidate state features. With candidate harmonic interference characteristics ; Step 2: The discriminator simultaneously compares candidate features with real samples using both branches, and calculates the discriminator loss. The discriminator loss uses binary cross-entropy loss, which combines the discrimination errors of the two branches:

[0056] Step 3: Calculate the generator loss The objective is to minimize the discriminator's recognition ability, and the generator loss formula is:

[0057] Step 4: Backpropagate the loss value to the generator and discriminator networks. The generator is based on... and constraints Adjust feature mapping strategy; discriminator based on Optimize the discrimination criteria to improve the sample differentiation ability; Step 5: Introduce preset power grid harmonic constraints Calibrate the training direction, among which To set an allowable deviation threshold, the model is prevented from deviating from the decoupling objective; Finally, repeat steps 1-5 until... , The state characteristics output by the generator tend to stabilize, and the discriminator accuracy approaches random levels. Harmonic interference from the power grid has been isolated to the greatest extent possible, and characteristic decoupling has been completed.

[0058] S4 inputs the decoupled time-frequency features into a preset CNN evaluation model after weighted fusion to complete the health status classification and output the quantitative power transformer health status evaluation level.

[0059] The weighted fusion method uses a combination of entropy weighting and analytic hierarchy process to determine feature weights. First, the objective weights of each time-frequency feature are calculated using the entropy weight method. Then, the subjective weights are determined using the analytic hierarchy process. Finally, the fused features are calculated using a linear weighted summation formula.

[0060] Specifically, in order to perform entropy weight calculation, an evaluation matrix is ​​first constructed based on historical data or samples for each decoupled time-frequency feature value under each health status level.

[0061] Furthermore, the specific steps of the entropy weight method include: Data standardization: The eigenvalues ​​in the evaluation matrix are normalized to eliminate the influence of dimensions.

[0062] Calculate the weight of each feature: Calculate the weight of each sample value under each feature.

[0063] Calculate information entropy: Combine the weighting and calculate the information entropy of each feature according to the information entropy formula.

[0064] Calculate the difference coefficient: Calculate the difference coefficient for each feature based on information entropy. The smaller the information entropy, the larger the difference coefficient of the feature, indicating that the feature plays a greater role in the evaluation.

[0065] Determine objective weights: Calculate objective weights based on the difference coefficients of each feature.

[0066] Furthermore, the specific steps of the analytic hierarchy process include: Establish a hierarchical model: the target layer is the optimal feature fusion, the criterion layer may include "feature sensitivity", "feature stability" and "correlation with fault", and the scheme layer is the decoupled time-frequency features.

[0067] Constructing the judgment matrix: Based on the experience of domain experts, the importance of each feature within the same level to a certain criterion of the previous level is compared pairwise, and the 1-9 scale method is used to assign values ​​to form the judgment matrix.

[0068] Calculate subjective weights: Calculate the largest eigenvalue of the judgment matrix and its corresponding eigenvector. After normalizing the eigenvector, the subjective weight vector of each feature is obtained.

[0069] Consistency check: Calculate the consistency ratio. If it is less than the set threshold, the consistency of the judgment matrix is ​​acceptable; otherwise, the judgment matrix needs to be adjusted.

[0070] Finally, a linear combination method is used to synthesize the subjective and objective weights: The final comprehensive weight of each time-frequency feature after decoupling is the weighted sum of objective weight and subjective weight. The specific values ​​of the weights can be adjusted according to the actual evaluation effect, or according to the emphasis on data objectivity and expert experience.

[0071] The decoupled time-frequency features are fused together using the final comprehensive weights to obtain the fused features, which are then used as input to the CNN evaluation model.

[0072] A pre-defined CNN model is used to assess the health status of the decoupled power transformer state features (SG). This model consists of convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional layers rely on sliding operations of the convolutional kernels to capture multi-dimensional correlation information of the features. Let the input feature map have a dimension of... The kernel size is The formula for convolution is:

[0073] in, To output the feature map of the th The eigenvalue at the channel and coordinate (i, j) These are the convolution kernel weight parameters. For bias terms, The input feature map contains the feature values ​​at the corresponding locations.

[0074] The feature map after convolution operation Activation function processing breaks the limitations of linear fitting. The function formula is:

[0075] The activated feature map is fed into a pooling layer for downsampling and dimensionality reduction, using max pooling with a pooling window size of [size missing]. The calculation formula is:

[0076] in, This is the feature value at the corresponding position in the pooled feature map, achieving the goal of removing redundant information and retaining core features.

[0077] Flatten the feature map output from the pooling layer into a one-dimensional feature vector. The features are fed into a fully connected layer for feature integration. The calculation formula for the fully connected layer is as follows:

[0078] in, This is the weight matrix of the fully connected layer. For bias vectors, This is the integrated feature vector.

[0079] The output layer incorporates a Softmax function, transforming the feature vector Y into probability distributions corresponding to various health states. The Softmax function formula is as follows:

[0080] in, For the number of health status categories, For feature vectors The One element, For the target object to belong to the first The probability of a health state satisfies .

[0081] The probability distributions are sorted, and the health status corresponding to the maximum probability is taken as the final health status assessment result of the power transformer.

[0082] The health status of the power transformer described in this invention can be divided into multiple discrete levels according to industry standards or actual operation and maintenance needs. For example, it can typically be defined into the following four categories: Normal: All characteristic parameters are normal, and the equipment is operating stably.

[0083] Note: Some characteristic parameters show a slight deterioration trend, requiring enhanced monitoring.

[0084] Abnormal: Multiple characteristic parameters deviate significantly from the normal range, indicating a potential fault.

[0085] Critical: The characteristic parameters are severely abnormal, indicating a clear fault that requires immediate repair.

[0086] Example 2 This invention also proposes a transformer health status assessment method based on dual-source time-frequency characteristics, including a data acquisition and processing module, a dataset construction module, a feature decoupling module, and a health assessment module: The data acquisition and processing module collects vibration and acoustic emission signals from the power transformer through sensors and performs preprocessing on the vibration and acoustic emission signals. The dataset construction module extracts the time-domain and frequency-domain features of vibration and acoustic emission signals respectively, and performs spatiotemporal alignment processing to construct a multi-dimensional time-frequency feature set; The feature decoupling module inputs a multi-dimensional time-frequency feature set into a GAN-based decoupling model, sets constraints, performs iterative training, separates the power grid harmonic interference, and obtains the decoupled time-frequency features. The health assessment module takes the decoupled time-frequency features, performs weighted fusion, and inputs them into a preset CNN assessment model to complete the health status classification and output a quantitative assessment level of the power transformer's health status.

[0087] The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0088] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0089] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0090] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0091] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0092] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for assessing the health status of a transformer based on dual-source time-frequency characteristics, characterized in that, include: S1, through the sensor to collect vibration signals and acoustic emission signals of the power transformer, and preprocess the vibration signals and acoustic emission signals; S2, extract the time-domain and frequency-domain features of vibration and acoustic emission signals respectively, and perform spatiotemporal alignment processing to construct a multi-dimensional time-frequency feature set; S3. Input the multi-dimensional time-frequency feature set into the GAN-based decoupling model, set the constraints, perform iterative training, separate the power grid harmonic interference, and obtain the decoupled time-frequency features. S4 inputs the decoupled time-frequency features into a pre-defined CNN evaluation model after weighted fusion to complete the health status classification and outputs a quantified power transformer health status evaluation level.

2. The transformer health status assessment method based on dual-source time-frequency characteristics according to claim 1, characterized in that: In S2, the time-domain and frequency-domain features of the vibration and acoustic emission signals are extracted respectively, specifically as follows: Temporal characteristics include root mean square, variance, standard deviation, skewness, and kurtosis; Frequency domain characteristics include dominant frequency amplitude, bandwidth, spectral entropy, spectral flatness, and spectral roll-off point.

3. The transformer health status assessment method based on dual-source time-frequency characteristics according to claim 1, characterized in that: In S2, spatiotemporal alignment is performed to construct a multidimensional time-frequency feature set. Specific methods include: Time alignment is based on the timestamp of the vibration signal to calibrate the time axis of the acoustic emission signal; By establishing an internal coordinate system based on the location of the power transformer structure, spatial alignment is achieved by spatially associating the acquisition locations of the core and winding vibration signals with the sensor locations of the acoustic emission signals.

4. The transformer health status assessment method based on dual-source time-frequency characteristics according to claim 1, characterized in that: In S3, the multi-dimensional time-frequency feature set is input into the GAN-based decoupling model, constraints are set, and iterative training is performed to separate the power grid harmonic interference, obtaining the decoupled time-frequency features, specifically including: The GAN-based decoupling model includes a generator and a discriminator; The generator adopts a convolutional neural network structure. The input is a multi-dimensional time-frequency feature set, and the output is the separated power grid harmonic interference features and the power transformer's own state features. The discriminator adopts a fully connected neural network structure. The separated power grid harmonic interference features and the power transformer's own state features are input into the discriminator, and the feature category judgment result is output.

5. The transformer health status assessment method based on dual-source time-frequency characteristics according to claim 4, characterized in that: In S3, the constraints specifically include: Frequency constraints are set based on the fixed frequency characteristics of power grid harmonics; Amplitude constraints are set based on the distribution law that harmonic amplitude decreases with increasing frequency; The correlation coefficient between the decoupled power grid harmonic interference characteristics and the power transformer's own state characteristics is lower than the set threshold.

6. The transformer health status assessment method based on dual-source time-frequency characteristics according to claim 4, characterized in that: In S3, iterative training is performed to separate the harmonic interference from the power grid, obtaining the decoupled time-frequency characteristics. Specific steps include: Initialize the model parameters, and the generator outputs the characteristics of power grid harmonic interference and the state characteristics of the power transformer itself. The discriminator identifies and judges the grid harmonic interference characteristics output by the generator and the power transformer's own state characteristics, respectively. Update the parameters of the generator and discriminator using the backpropagation algorithm; Iterate through the training until the loss function converges, then stop training and save the model parameters.

7. The transformer health status assessment method based on dual-source time-frequency characteristics according to claim 1, characterized in that: In S4, the step of weighted fusion includes: The objective weights of each time-frequency feature after decoupling are calculated using the entropy weight method. The subjective weights of each time-frequency feature in the decoupled system are determined using the analytic hierarchy process (AHP). The final comprehensive weight of the corresponding time-frequency feature is obtained by linearly summing the objective weight and the subjective weight; The decoupled time-frequency features are summed based on the final comprehensive weights to obtain the input of the CNN evaluation model.

8. The transformer health status assessment method based on dual-source time-frequency characteristics according to claim 1, characterized in that: In S4, the CNN evaluation model specifically includes: The CNN evaluation model consists of convolutional layers, pooling layers, fully connected layers, and an output layer. Convolutional layers extract comprehensive fault features, pooling layers reduce feature dimensionality while retaining fault information, ReLU activation function is used for nonlinear fitting, fully connected layers integrate and summarize features and transform them into feature vectors suitable for classification decisions, and the output layer completes health status classification through the Softmax function.

9. A transformer health status assessment based on dual-source time-frequency characteristics using the method described in any one of claims 1-8, comprising a data acquisition and processing module, a dataset construction module, a feature decoupling module, and a health assessment module, characterized in that: The data acquisition and processing module collects vibration and acoustic emission signals from the power transformer through sensors and performs preprocessing on the vibration and acoustic emission signals. The dataset construction module extracts the time-domain and frequency-domain features of vibration and acoustic emission signals respectively, and performs spatiotemporal alignment processing to construct a multi-dimensional time-frequency feature set; The feature decoupling module inputs a multi-dimensional time-frequency feature set into a GAN-based decoupling model, sets constraints, performs iterative training, separates the power grid harmonic interference, and obtains the decoupled time-frequency features. The health assessment module takes the decoupled time-frequency features, performs weighted fusion, and inputs them into a preset CNN assessment model to complete the health status classification and output a quantitative assessment level of the power transformer's health status.

10. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-8.

Images