Transformer winding fault monitoring method and system based on multivariate information fusion and deep learning

Abstract

The application discloses a kind of based on multivariate information fusion and transformer winding fault monitoring method and system of deep learning, method includes: obtaining original vibration signal in the process of transformer operation, input to the transformer fault diagnosis model constructed in advance, output transformer winding fault type;The process for being constructed in advance by transformer fault diagnosis model is: obtaining original vibration signal in the process of transformer operation, constructs initial signal data set and is divided into multiple data samples;Composite feature vector of fusion time domain, frequency domain and time-frequency domain features is extracted from data sample;Convolutional neural network is constructed, and channel attention mechanism and spatial attention mechanism are introduced in convolutional neural network, and the deep learning model of fusion attention mechanism is constructed;Composite feature vector is used to train and evaluate deep learning model, and the final transformer winding fault diagnosis model is obtained.The application has the advantages of high fault diagnosis precision.

Description

Technical Field

[0001] This invention mainly relates to the field of transformer fault monitoring technology, specifically to a transformer winding fault monitoring method and system based on multi-source information fusion and deep learning. Background Technology

[0002] As a key energy hub in ultra-high voltage (UHV) and extra-high voltage (UHV) power transmission networks, the reliability of the mechanical state of oil-immersed power transformer windings directly determines the continuity of power supply and the operational safety margin. With the evolution of "dual-high" power grids (high proportion of renewable energy and high proportion of power electronic equipment), traditional fault criteria based on threshold limits or simplified physical models are no longer sufficient to meet the demand for refined perception of latent defects. Currently, research on winding mechanical fault diagnosis mainly faces the following three technical bottlenecks: The non-stationary coupling characteristics are prominent: the vibration signal of the transformer is excited by both the magnetostriction of the iron core and the electrodynamic force of the winding, exhibiting wide bandwidth, nonlinearity and non-stationary characteristics. Conventional time-frequency analysis (such as short-time Fourier transform and wavelet transform) has inherent contradictions in terms of time-frequency resolution and energy concentration, making it difficult to deeply explore weak fault components.

[0003] Strong background noise interference: The measured signal is mixed with multiple sources of disturbance such as cooling system, tap changer operation and external short circuit impact, which leads to a significant reduction in the signal-to-noise ratio of fault characteristics. Traditional filtering or blind source separation methods are prone to characteristic distortion when lacking prior knowledge.

[0004] High dependence on feature engineering: Manually designed statistics (peak-to-peak value, kurtosis, energy entropy, etc.) are extremely sensitive to expert experience, and their generalization performance drops sharply when faced with multi-condition and multi-fault coupled scenarios, making it difficult to achieve adaptive feature selection.

[0005] In recent years, intelligent diagnostic technologies, represented by deep representation learning, have offered new solutions to the aforementioned challenges. Convolutional Neural Networks (CNNs), with their local receptive fields and weight-sharing mechanisms, can progressively abstract the spatial hierarchical features of signals; the attention mechanism, through adaptive weight allocation, significantly enhances the discriminative representation of key fault modes. However, existing research still has the following shortcomings: Simple structure: Most algorithms directly use one-dimensional CNN or simple stacked convolutional blocks, without fully considering the coupling and correlation of vibration signals in the time-frequency-space three-dimensional domain, resulting in insufficient feature extraction.

[0006] Multi-source heterogeneous features are fragmented: time-domain statistical features, frequency-domain resonance features and acoustic-electric-vibration multimodal information lack a unified embedding space, resulting in severe information silos and limiting diagnostic robustness.

[0007] Imbalanced sample distribution: In actual operation and maintenance, fault samples are scarce and the categories are extremely unbalanced, which makes the network prone to falling into the "overfitting-minority class" trap, and the generalization performance under complex working conditions is difficult to guarantee.

[0008] To address the aforementioned technical challenges, there is an urgent need to construct an end-to-end diagnostic framework for deep fusion of multi-source information, achieving the following core objectives: 1. Automatically capture the deep spatiotemporal-spectral coupling characteristics of winding faults in non-stationary, high-noise environments; 2. Establish an adaptive feature selection mechanism that coordinates channel and space to suppress redundant information and enhance key discriminative features; 3. It still exhibits high robustness and generalization ability under imbalanced sample and working condition transfer scenarios. Summary of the Invention

[0009] To address the technical problems existing in the prior art, this invention provides a transformer winding fault monitoring method and system with high diagnostic accuracy based on multi-source information fusion and deep learning.

[0010] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A method for monitoring transformer winding faults based on multi-source information fusion and deep learning includes the following steps: The original vibration signal during transformer operation is acquired, and the original vibration signal is preprocessed. The preprocessed vibration signal is then input into a pre-built transformer fault diagnosis model, and the transformer winding fault type is output. The process of pre-constructing the transformer fault diagnosis model is as follows: The original vibration signals during transformer operation are acquired, and the original vibration signals are preprocessed to construct an initial signal dataset and divide it into multiple data samples. Extract composite feature vectors that fuse time-domain, frequency-domain, and time-frequency-domain features from the data samples, use them as input sample data, and divide them into pre-training set, validation set, and test set; Construct a convolutional neural network containing convolutional layers, pooling layers, and fully connected layers, and introduce channel attention mechanism and spatial attention mechanism into the convolutional neural network to build a deep learning model that integrates attention mechanisms; The deep learning model is trained using the pre-training set and validation set to obtain a pre-trained transformer fault diagnosis model. The pre-trained transformer fault diagnosis model is then tested and its performance evaluated using the test set. Based on the evaluation results, the transformer fault diagnosis model is optimized and adjusted to obtain the final transformer winding fault diagnosis model.

[0011] Preferably, the specific process of extracting a composite feature vector from the data sample, which integrates time-domain, frequency-domain, and time-frequency-domain features, is as follows: For each vibration signal sample, its effective value, kurtosis, peak factor, impulse factor, waveform factor, and margin factor are calculated to construct a primary time-domain feature subset characterizing the fundamental features of the signal; The vibration signal is subjected to a fast Fourier transform to obtain the amplitude spectrum. Based on the amplitude spectrum, the spectral centroid, variance, entropy, amplitude and frequency of the first N main harmonic components, and the energy proportion of the preset fault frequency band are obtained, and an intermediate feature subset reflecting the frequency domain characteristics of the signal is constructed. The vibration signal is subjected to continuous wavelet transform to generate a high-resolution wavelet scale map, and then deep compression coding is performed to obtain a low-dimensional latent vector, which serves as a high-level time-frequency domain feature subset. The primary time-domain feature subset, the intermediate frequency-domain feature subset, and the high-level time-frequency domain feature subset are adaptively fused to obtain a composite feature vector.

[0012] Preferably, the specific process for obtaining the composite feature vector is as follows: Dynamic fusion weights are generated for the primary time-domain feature subset, the intermediate frequency-domain feature subset, and the advanced time-frequency-domain feature subset using independent multilayer perceptrons; The primary time-domain feature subset, intermediate frequency-domain feature subset, and advanced time-frequency-domain feature subset are projected into the shared semantic space, and the cross-attention of the frequency-domain feature subset and the time-frequency-domain feature subset relative to the time-domain feature subset is calculated to generate the corresponding compensation feature vector. Based on the dynamic fusion weights, each feature subset is weighted and summed, and the summation result is fused with the compensation feature vector through a gating vector to obtain the fused features. The fused features are subjected to nonlinear transformation and dimensionality reduction to output the final composite feature vector.

[0013] Preferably, the convolutional neural network includes multiple layers of convolutional blocks connected in series, with each convolutional block internally stacked in the order of convolutional layer, batch normalization layer, activation function layer and pooling layer; wherein the activation function is the ReLU function.

[0014] Preferably, the process of introducing channel attention and spatial attention mechanisms into a convolutional neural network is as follows: After each convolutional block, SE channel attention modules and CBAM spatial attention modules are inserted in parallel. The channel attention module adaptively assigns weights to each feature channel through global average pooling, fully connected layer mapping, and Sigmoid activation. The spatial attention module performs max pooling and average pooling on the feature map along the channel axis respectively, concatenates the results, and then generates a two-dimensional spatial weight map through convolution and sigmoid activation. The outputs of the two attention modules are multiplied element-wise to complete the feature recalibration.

[0015] Preferably, the preprocessing includes normalization, which employs the Z-Score normalization method, with the specific formula as follows:

[0016] in, For the first i The result of normalizing the signals collected by each sensor For the first i Each sensor collects signals. and The first i The mean and standard deviation of the signal sequences collected by each sensor. i =1, 2... N N represents the total number of sensors.

[0017] The present invention also discloses a computer program product, comprising a computer program that, when executed by a processor, performs the steps of the method described above.

[0018] The present invention further discloses a computer-readable storage medium having a computer program stored thereon, the computer program executing the steps of the method described above when run by a processor.

[0019] The present invention also discloses a transformer winding fault monitoring system based on multi-source information fusion and deep learning, including a memory and a processor connected to each other. The memory stores a computer program, which executes the steps of the method described above when run by the processor.

[0020] Compared with the prior art, the advantages of the present invention are as follows: The transformer winding fault monitoring method of this invention collects multi-dimensional vibration signals during transformer operation through a distributed sensor network, constructs a composite feature set integrating time-domain and frequency-domain features, and preprocesses the input features using a standardization method. In terms of model architecture, this invention innovatively introduces a deep convolutional neural network and a channel-space dual attention mechanism to construct an adaptive feature weighting module, enabling the model to autonomously focus on key fault information and enhance the targeting of feature selection. Simultaneously, a multi-stage dataset partitioning strategy is adopted to balance model training efficiency and generalization ability, effectively improving diagnostic stability. The embedding of the attention mechanism not only strengthens the identification ability of key features but also significantly improves the model's convergence performance.

[0021] The innovative advantages of the method of this invention are mainly reflected in the following aspects: First, through the synergistic mechanism of multi-source information fusion and attention weighting, the distinguishability and identification accuracy of fault features are significantly improved; Second, the optimized lightweight network structure greatly improves training efficiency while ensuring feature extraction depth; Third, the introduction of a multi-stage cross-validation mechanism enhances the reliability and robustness of diagnostic results. Attached Figure Description

[0022] Figure 1 This is a flowchart of transformer winding fault diagnosis based on correlation coefficient matrix and ACNN in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the ACNN structure in an embodiment of the present invention.

[0024] Figure 3 This is a curve showing the change in the recognition accuracy of transformer winding fault diagnosis results based on the correlation coefficient matrix and ACNN in an embodiment of the present invention.

[0025] Figure 4 This is a graph showing the change in the loss function of the transformer winding fault diagnosis results based on the correlation coefficient matrix and ACNN in an embodiment of the present invention.

[0026] Figure 5 This is the classification confusion matrix diagram of the ACNN model of this invention for the test set.

[0027] Figure 6 This is a graph showing the change in recognition accuracy based on different sensing channels in this invention.

[0028] Figure 7 This is a graph showing the variation of loss function values ​​based on different sensing channels in this invention.

[0029] Figure 8 This is a flowchart of an embodiment of the transformer winding fault monitoring method of the present invention. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0031] like Figure 1 and Figure 8 As shown in the figure, the transformer winding fault monitoring method based on multi-source information fusion and deep learning provided by the embodiments of the present invention includes the following steps: S1. Construct an experimental test platform for winding faults in an oil-immersed transformer and build a multi-channel signal acquisition system; install vibration sensors at key locations on the surface of the transformer test model's oil tank to collect the raw vibration signals generated during transformer operation in real time; S2. The collected raw vibration signal data is preprocessed to construct an initial signal dataset, and the dataset is normalized. Then, the normalized dataset is divided into multiple data samples for subsequent modeling and training. Specifically, the original vibration signal is represented in the form of a one-dimensional time series, denoted as . x ( t ),in t The time variable is used; based on the vibration response characteristics corresponding to different fault types, a fault feature set is constructed to characterize the transformer operating state.

[0032] The normalization process uses the Z-Score standardization method, and the specific formula is as follows:

[0033] in, For the first i The result of normalizing the signals collected by each sensor and These represent the mean and standard deviation of the signal sequence acquired by the sensor, respectively. i =1, 2... N N represents the total number of sensors.

[0034] S3. Extract effective feature vectors from the data samples and use them as input sample data; further divide the input sample data into pre-training set, validation set and test set for model training and performance evaluation; Specifically, the hierarchical multidimensional feature pyramid construction process is executed to extract composite feature vectors that fuse the time domain, frequency domain, and time-frequency domain from the data samples, including: Primary Time-Domain Feature Extraction Layer: This layer directly extracts statistically significant features with clear physical meaning from the signal waveform to characterize its basic energy and impact morphology. For each normalized vibration signal sample, its effective value, kurtosis, peak factor, impulse factor, waveform factor, and margin factor are calculated. These six features describe the signal from different perspectives, including average energy, distribution steepness, peak prominence, impact characteristics, waveform shape, and relative peak level, respectively, and together constitute the primary time-domain feature subset characterizing the signal's fundamental properties. ; Intermediate Frequency Domain Feature Analysis Layer: This layer reveals the harmonic structure and energy distribution patterns of a signal through frequency domain transformation. This information is directly related to the physical state of mechanical components. First, a Fast Fourier Transform (FFT) is performed on the signal to obtain the amplitude spectrum. Then, the spectral centroid, variance, and entropy are calculated to describe the overall characteristics of the energy distribution. The amplitudes and frequencies of the top N major harmonic components are extracted to capture the dominant oscillation modes. Finally, the energy proportion of preset key fault frequency bands is calculated to specifically monitor characteristic frequency bands that may experience resonance or anomalies. These features collectively constitute an intermediate frequency domain feature subset reflecting the signal's frequency domain characteristics. ; Advanced Time-Frequency Domain Feature Condensation Layer: To capture the non-stationary and transient components in the signal, deep feature learning is performed in the joint time-frequency domain. First, a continuous wavelet transform is applied to the signal to generate a high-resolution wavelet scale map, which fully preserves the energy evolution information of the signal in the time-frequency plane. Then, a pre-trained lightweight 2D convolutional autoencoder is used for deep compression encoding of the image: the encoder progressively extracts abstract features through convolutional layers and compresses them into a low-dimensional latent vector; the decoder attempts to reconstruct the image. During training, the model is optimized using both reconstruction error and perceptual loss to ensure that the latent vector effectively preserves the key patterns of the time-frequency structure. This latent vector serves as the condensed advanced time-frequency domain feature subset. ; Adaptive Feature Fusion Layer: This layer organically integrates and optimizes feature subsets from the three different levels of abstraction and physical dimensions mentioned above. It uses the primary temporal feature subset... intermediate frequency domain feature subset and high-level time-frequency domain feature subsets As input, dynamic fusion weights are first computed for each feature subset via a modality-specific weight generation path. Specifically, each subset is processed by an independent multilayer perceptron (MLP), which outputs a scalar weight value. (After normalization using the Sigmoid function). This process can be formally represented as follows: ,in i ∈{ time , freq , ftime}, σ This is a sigmoid function. These weights reflect the relative importance of each modality feature in the current sample, achieving automation and adaptability in "feature selection".

[0035] Simultaneously, a parallel cross-attention and compensation path is activated to model and mitigate potential conflicts between features from different modalities. This path first projects three feature subsets into a shared semantic space, then calculates the cross-attention of frequency-domain features and time-frequency-domain features relative to time-domain features (as reference anchors). Taking the time-domain to frequency-domain transition as an example, the calculation process is as follows: The obtained attention-weighted features and These are considered "compensation features," encoding information from other modal features that is related to or complementary to the core temporal features. Finally, in the gated collaborative fusion and dimensionality reduction stage, the initial weighted features... This is then fused with the aggregated compensation features. A learnable gating mechanism is introduced here: The compensation information is controlled by this gating vector, and the final output is... This design achieves an adaptive balance between original and compensated information. Ultimately, By performing nonlinear transformation and dimensionality reduction through a bottleneck layer (such as a fully connected layer with shrinking dimensions), a unified, compact, and highly discriminative final composite feature vector is output. This information is then used to power subsequent convolutional neural networks for classification decisions.

[0036] This fusion layer addresses the question of "what to fuse" through dynamic weight allocation, solves the problem of "how to coordinate" through a cross-compensation mechanism, and achieves efficient information integration through "gated fusion and bottleneck compression." It not only enhances the expression of key features but also structurally improves the model's adaptive fusion capability and diagnostic robustness against multi-source heterogeneous features under complex and variable operating conditions.

[0037] S4. Construct a Convolutional Neural Network (ACNN) model as a deep learning classifier to classify the operating state of transformers; Specifically, a convolutional neural network model includes convolutional layers, pooling layers, and fully connected layers: The convolutional layer extracts local features from the input data by performing convolution operations with the convolutional kernel and then processing them through an activation function. Its output can be represented as:

[0038] in, The representative is the first j Each convolution outputs a feature map. As input features, For convolution kernel weights, The term is the bias term, * represents the convolution operation, and f(.) is the activation function; Pooling layers are used to downsample the output of convolutional layers, employing max pooling, and their expression is:

[0039] in, R For pooling regions, Let be the activation value of the neuron at position (m,n).p j For pooled output; Multiple convolutional modules and pooling layers are stacked alternately to extract high-level features layer by layer, and finally a classification output is achieved through a fully connected layer, with classification probabilities. Calculated using the Softmax function:

[0040] in, z j For the output score of the j-th class, K This represents the total number of categories.

[0041] The convolutional layers employ non-linear activation functions to enhance the network's ability to model non-linear features; preferably, the ReLU (Rectified Linear Unit) activation function is used, and its expression is: .

[0042] In the fully connected layer, the input is mapped to the range (0, 1), which is converted into a probability distribution that sums to 1. The label corresponding to the highest probability is the classification result.

[0043] S5. Introduce channel attention mechanism and spatial attention mechanism into the convolutional neural network structure to construct a transformer deep learning model that integrates attention mechanism to enhance the model's ability to perceive key feature regions; Specifically, the channel-space attention mechanism comprises two independent sub-modules: a channel attention module and a spatial attention module, which analyze channel and spatial features respectively. The specific process is as follows: First, channel attention mechanism is used to process the input features. ; Features Input respectively based on and The global max pooling layer and the global average pooling layer are used to obtain two layers of size. The features are then input into a fully connected layer for computation and summation to obtain a one-dimensional channel attention. ,Will Input features Multiplication The process formula is as follows:

[0044] In the above formula, for The output weights after channel attention; This represents the weighted multiplication operator for the feature map.

[0045] Then, the features obtained in the previous step As input features for the spatial attention module. Channel-pair-based input features. Performing global max pooling and global average pooling yields two results. The features are concatenated in the channel dimension and then convolutional to obtain the two-dimensional spatial attention. ,Will and The output is obtained by element-wise multiplication. Its formula can be expressed as:

[0046] In the above formula for The output weights of the channel attention.

[0047] S6. Use the pre-training set and validation set obtained in step S3 to train the deep learning model with the fusion attention mechanism to obtain a transformer fault diagnosis model that has been preliminarily trained. S7. Use the test set obtained in step S3 to test and evaluate the performance of the initially trained transformer fault diagnosis model. Based on the evaluation results, optimize and adjust the model structure and parameters to finally obtain a transformer winding fault diagnosis model with high accuracy and robustness, so as to realize the diagnosis of transformer winding faults. The classification effect of the model is judged by whether the network has reached the maximum number of iterations and the classification accuracy.

[0048] The present invention provides a method for monitoring winding faults in oil-immersed transformers that integrates multi-source information embedding and attention-enhanced convolutional neural networks. By constructing an integrated architecture of "data acquisition - deep representation - attention weighting - decision output", it achieves high sensitivity and high reliability in identifying latent mechanical defects, providing theoretical support and technical path for intelligent operation and maintenance of highly resilient power grids.

[0049] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0050] S1. Vibration Signal Acquisition and Experimental Platform Construction: A test platform capable of simulating and approximating real-world operating conditions was constructed in a laboratory environment. Its core is an oil-immersed transformer test piece designed to artificially induce various typical defects such as winding loosening, insufficient core clamping, and partial discharge. A multi-channel synchronous signal acquisition system was deployed around this test piece to achieve wideband, high-sensitivity measurement of surface vibrations in the oil tank. Sensor placement followed the principle of "equal angle + equal density," with ICP-type accelerometers attached to key vibration transmission paths such as the top, front, and sides of the oil tank to ensure complete capture of complex vibration modes generated by the coupling of windings, core, and internal structure. Acquisition parameters were uniformly set to a sampling frequency of no less than 20kHz, adjustable gain, and a preamplifier bandpass filter of 10Hz–10kHz to accommodate both low-frequency mechanical fault characteristics and high-frequency discharge pulses.

[0051] S2. The original vibration signal obtained in step S1 is segmented in the manner of "time window + overlapping sliding window", with each window segment being 2s long and overlapping by 50%, forming a large-scale signal segment library; then normalization (Z-Score) is performed to eliminate the difference in dimensions and improve the convergence stability of the model; finally, the normalized segments are randomly shuffled and divided into training set, validation set and test set in a ratio of 7:2:1, while ensuring the balance of fault categories in each set.

[0052] S3. Feature Extraction and Sample Labeling: For each signal segment, the effective value, kurtosis, peak factor, and impulse factor are first calculated in the time domain to characterize the impact and energy distribution. Then, the FFT spectrum centroid, spectral entropy, and key frequency band energy ratio are extracted in the frequency domain to capture the resonance and energy migration patterns. Finally, a high-resolution energy map is obtained in the time and frequency domain through continuous wavelet transform. Its L2 norm is taken and fed into a two-dimensional convolutional autoencoder to condense it into a 128-dimensional deep representation. The features of these three domains are concatenated to form a 256-dimensional fusion vector, and corresponding fault category labels are assigned: normal, winding loose, core loose, partial discharge, etc., thus forming samples that can be directly input into the network.

[0053] S4. Based on the architectural design concept of the Convolutional Neural Network (CNN) baseline model, the constructed Attention Convolutional Neural Network (ACNN) retains the hierarchical feature learning framework and achieves targeted enhancement and focus on key features by embedding an adaptive attention mechanism. For example... Figure 1As shown, the ACNN model also uses three concatenated convolutional blocks as its backbone. Each block performs the following operations: convolution (Conv) - batch normalization (BN) - activation (ReLU) - channel attention weighting (ChannelAttention) - max pooling (MaxPool). The convolutional kernel size is set layer by layer along the depth direction to 5×1, 3×1, and 3×1 to extract frequency domain features from coarse to fine. The number of channels increases from 32 to 64 and 128 to gradually improve the feature representation capability. The convolution stride is kept at 1, and the pooling kernel is uniformly 2×1 to ensure that spatial information is preserved in reasonable downsampling. The core difference from the baseline model lies in the fact that each convolutional block embeds a lightweight channel attention module before the pooling operation. This module first aggregates the spatial information of each channel through global average pooling, then generates the weight coefficients of each channel through a fully connected layer containing dimensionality reduction and a sigmoid activation function. Finally, it performs channel-level recalibration on the feature map, enabling the network to autonomously strengthen feature responses highly correlated with faults and suppress irrelevant or noise-dominated channels. This "convolutional extraction" Through alternating stacking of "attention filtering," the model expands its receptive field and refines abstract features layer by layer, while simultaneously achieving adaptive optimization selection of fault-sensitive features. Finally, the deep features obtained through global average pooling are integrated by a fully connected layer, and the probability distribution of each fault category is output through the Softmax function, thereby constructing an intelligent model for fault diagnosis of oil-immersed transformer windings that combines feature discriminativeness and model interpretability.

[0054] S5. Channel-Spatial Attention Embedding: such as Figure 2 As shown, SE channel attention and CBAM spatial attention are inserted in parallel after each convolutional block of the baseline ACNN to form a dual-branch collaborative structure. The channel attention first compresses the spatial dimension with global average pooling, and then maps and activates it with a fully connected layer. It adaptively assigns weights to each channel, so that the features most sensitive to faults are highlighted. The spatial attention performs max pooling and average pooling on the same feature map along the channel axis respectively. After concatenation, it generates a two-dimensional spatial weight map through 7×7 convolution and sigmoid to suppress irrelevant regions and strengthen key positions. The outputs of the two branches are multiplied element-wise to complete the "channel-space" joint recalibration. The whole constitutes the transformer deep learning model TS-CNN-Attn with significantly improved performance.

[0055] S6. Based on the training and validation sets constructed in step S3, the transformer deep learning model is iteratively trained to obtain the trained model. During the training process, the trend of fault diagnosis accuracy with training rounds is as follows: Figure 3 As shown, the trend of the loss function is as follows: Figure 4 As shown.

[0056] S7. Model Evaluation and Iterative Optimization: Figure 5 The classification confusion matrix of the ACNN model on the test set is presented. This matrix intuitively presents the model's classification performance, where each column represents the class predicted by the model, and each row corresponds to the actual true class of the sample. The results show that the ACNN model exhibits excellent classification accuracy and stability, with only two misclassified cases across the entire test set, indicating strong feature learning capabilities and good generalization performance. This result provides a clear direction for further model optimization. Future work will focus on continuously improving the model's diagnostic accuracy and robustness through misclassified case analysis, hyperparameter tuning, and feature engineering improvements.

[0057] S8. To demonstrate the advantages of the correlation coefficient matrix over the correlation coefficient vector, a comparative experiment was set up. Using a specific sensor as a reference point, the correlation coefficients between the remaining measurement points and the reference point were calculated to construct a feature dataset. Since the correlation coefficient calculation results are independent of the order of the two measurement points, i.e. Therefore, the first half of the measuring points were selected as reference points in sequence. Figure 6 Based on the accuracy variation curves of different sensor channels and Figure 7 Based on the curves showing the changes in loss function values ​​for different sensing channels, it can be seen that compared to calculating the correlation coefficient vector using a certain measurement point as a reference point, the ACNN model trains faster and requires fewer iterations to converge when the correlation coefficient matrix is ​​used as the feature dataset. Furthermore, the convergence speed of the model varies depending on the reference point selected.

[0058] The transformer winding fault monitoring method of this invention collects multi-dimensional vibration signals during transformer operation through a distributed sensor network, constructs a composite feature set integrating time-domain and frequency-domain features, and preprocesses the input features using a standardization method. In terms of model architecture, this invention innovatively introduces a deep convolutional neural network and a channel-space dual attention mechanism to construct an adaptive feature weighting module, enabling the model to autonomously focus on key fault information and enhance the targeting of feature selection. Simultaneously, a multi-stage dataset partitioning strategy is adopted to balance model training efficiency and generalization ability, effectively improving diagnostic stability. The embedding of the attention mechanism not only strengthens the identification ability of key features but also significantly improves the model's convergence performance.

[0059] The innovative advantages of the method of this invention are mainly reflected in the following aspects: First, through the synergistic mechanism of multi-source information fusion and attention weighting, the distinguishability and identification accuracy of fault features are significantly improved; Second, the optimized lightweight network structure greatly improves training efficiency while ensuring feature extraction depth; Third, the introduction of a multi-stage cross-validation mechanism enhances the reliability and robustness of diagnostic results.

[0060] The method of this invention is particularly applicable to real-time status monitoring and early fault warning of key equipment such as oil-immersed transformers in power systems, providing effective technical support for realizing intelligent operation and maintenance of transformers.

[0061] The method of this invention automatically captures the deep spatiotemporal-spectral coupling characteristics of winding faults in non-stationary high-noise environments; establishes an adaptive feature selection mechanism for channel-space coordination to suppress redundant information and enhance key discriminative features; and still has high robustness and generalization ability in scenarios of sample imbalance and operating condition migration.

[0062] The transformer winding fault monitoring method of the present invention can improve the accuracy of fault diagnosis and assessment, and effectively prevent equipment damage and unplanned downtime losses caused by missed or misdiagnosed faults in oil-immersed transformers.

[0063] The present invention also discloses a computer program product, comprising a computer program that, when executed by a processor, performs the steps of the method described above.

[0064] The present invention further discloses a computer-readable storage medium having a computer program stored thereon, the computer program executing the steps of the method described above when run by a processor.

[0065] The present invention also discloses a transformer winding fault monitoring system based on multi-source information fusion and deep learning, including a memory and a processor connected to each other. The memory stores a computer program, which executes the steps of the method described above when run by the processor.

[0066] The products, media, and systems of the present invention, corresponding to the methods described above, also possess the advantages described above.

[0067] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0068] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring transformer winding faults based on multi-source information fusion and deep learning, characterized in that, Including the following steps: The original vibration signal during transformer operation is acquired, and the original vibration signal is preprocessed. The preprocessed vibration signal is then input into a pre-built transformer fault diagnosis model, and the transformer winding fault type is output. The process of pre-constructing the transformer fault diagnosis model is as follows: The original vibration signals during transformer operation are acquired, and the original vibration signals are preprocessed to construct an initial signal dataset and divide it into multiple data samples. Extract composite feature vectors that fuse time-domain, frequency-domain, and time-frequency-domain features from the data samples, use them as input sample data, and divide them into pre-training set, validation set, and test set; Construct a convolutional neural network containing convolutional layers, pooling layers, and fully connected layers, and introduce channel attention mechanism and spatial attention mechanism into the convolutional neural network to build a deep learning model that integrates attention mechanisms; The deep learning model is trained using the pre-training set and validation set to obtain a pre-trained transformer fault diagnosis model. The pre-trained transformer fault diagnosis model is then tested and its performance evaluated using the test set. Based on the evaluation results, the transformer fault diagnosis model is optimized and adjusted to obtain the final transformer winding fault diagnosis model.

2. The transformer winding fault monitoring method based on multi-source information fusion and deep learning according to claim 1, characterized in that, The specific process of extracting a composite feature vector that fuses time-domain, frequency-domain, and time-frequency-domain features from the data sample is as follows: For each vibration signal sample, its effective value, kurtosis, peak factor, impulse factor, waveform factor, and margin factor are calculated to construct a primary time-domain feature subset characterizing the fundamental features of the signal; The vibration signal is subjected to a fast Fourier transform to obtain the amplitude spectrum. Based on the amplitude spectrum, the spectral centroid, variance, entropy, amplitude and frequency of the first N main harmonic components, and the energy proportion of the preset fault frequency band are obtained, and an intermediate feature subset reflecting the frequency domain characteristics of the signal is constructed. The vibration signal is subjected to continuous wavelet transform to generate a high-resolution wavelet scale map, and then deep compression coding is performed to obtain a low-dimensional latent vector, which serves as a high-level time-frequency domain feature subset. The primary time-domain feature subset, the intermediate frequency-domain feature subset, and the high-level time-frequency domain feature subset are adaptively fused to obtain a composite feature vector.

3. The transformer winding fault monitoring method based on multi-source information fusion and deep learning according to claim 2, characterized in that, The specific process of obtaining the composite feature vector is as follows: Dynamic fusion weights are generated for the primary time-domain feature subset, the intermediate frequency-domain feature subset, and the advanced time-frequency-domain feature subset using independent multilayer perceptrons; The primary time-domain feature subset, intermediate frequency-domain feature subset, and advanced time-frequency-domain feature subset are projected into the shared semantic space, and the cross-attention of the frequency-domain feature subset and the time-frequency-domain feature subset relative to the time-domain feature subset is calculated to generate the corresponding compensation feature vector. Based on the dynamic fusion weights, each feature subset is weighted and summed, and the summation result is fused with the compensation feature vector through a gating vector to obtain the fused features. The fused features are subjected to nonlinear transformation and dimensionality reduction to output the final composite feature vector.

4. The transformer winding fault monitoring method based on multi-source information fusion and deep learning according to claim 1, 2, or 3, characterized in that, The convolutional neural network comprises multiple layers of convolutional blocks connected in series. Each convolutional block is stacked in the order of convolutional layer, batch normalization layer, activation function layer and pooling layer; wherein the activation function is the ReLU function.

5. The transformer winding fault monitoring method based on multi-source information fusion and deep learning according to claim 4, characterized in that, The activation function is the ReLU function.

6. The transformer winding fault monitoring method based on multi-source information fusion and deep learning according to claim 4, characterized in that, The process of introducing channel attention and spatial attention mechanisms into convolutional neural networks is as follows: After each convolutional block, SE channel attention modules and CBAM spatial attention modules are inserted in parallel. The channel attention module adaptively assigns weights to each feature channel through global average pooling, fully connected layer mapping, and Sigmoid activation. The spatial attention module performs max pooling and average pooling on the feature map along the channel axis respectively, concatenates the results, and then generates a two-dimensional spatial weight map through convolution and sigmoid activation. The outputs of the two attention modules are multiplied element-wise to complete the feature recalibration.

7. The transformer winding fault monitoring method based on multi-source information fusion and deep learning according to claim 1, 2, or 3, characterized in that, Preprocessing includes normalization, which uses the Z-Score normalization method, with the specific formula as follows: in, For the first i The result of normalizing the signals collected by each sensor For the first i Each sensor collects signals. and The first i The mean and standard deviation of the signal sequences collected by each sensor. i =1, 2... N ; N This represents the total number of sensors.

8. A computer program product, comprising a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-7.

10. A transformer winding fault monitoring system based on multi-source information fusion and deep learning, comprising an interconnected memory and a processor, wherein the memory stores a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-7.

Images