Intelligent Diagnosis Method for Partial Discharge in Transformers Based on Improved Transformation Network

Abstract

The present invention discloses an intelligent diagnosis method for partial discharge of a transformer based on an improved transformation network, which relates to the technical field of transformer fault diagnosis. The method includes collecting partial discharge signals of the transformer, preprocessing the partial discharge signals to obtain a time series; using the Markov transition field (MTF) algorithm, the Gram angular field (GASF) algorithm, and the Gram angular difference field (GADF) algorithm to convert the time series into feature images respectively; processing the feature images through an improved transformation network model to obtain a fused feature vector; and inputting the fused feature vector into a fault diagnosis model to output a fault diagnosis result of the partial discharge of the transformer. By combining an interpretable graph reasoning mechanism of environmental parameters and signal features, the present invention improves the recognition rate of oil gap discharge in tests, enabling the fault diagnosis to leap from the traditional threshold-based judgment mode to an intelligent mode based on causal logic reasoning.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer fault diagnosis, and particularly to an intelligent diagnosis method for transformer partial discharge based on an improved transformation network. Background Art

[0002] With the continuous expansion of the power grid scale, as a core device of the power system, the health status of transformers directly affects the safe and stable operation of the power grid. Partial discharge is one of the important manifestations of internal insulation defects in transformers. Long-term partial discharge may lead to the deterioration of insulation materials and ultimately cause equipment failures. Accurately identifying partial discharge patterns and conducting intelligent diagnosis is crucial for ensuring the reliability of the power grid.

[0003] Currently, the detection and diagnosis methods for partial discharge mainly include:

[0004] 1) Methods based on mathematical models: Analyze partial discharge characteristics by establishing a physical model of the transformer. However, due to the complex structure of the transformer, it is difficult to accurately model, resulting in limited diagnostic accuracy. In addition, mathematical models rely on a large number of physical parameters and assumptions and are easily affected by environmental changes.

[0005] 2) Methods based on signal processing: Use time-frequency analysis methods (such as wavelet transform, short-time Fourier transform, etc.) to extract discharge characteristics, but are sensitive to noise and lack the ability to effectively model complex patterns. Existing signal processing methods often struggle to adapt to the complex partial discharge signals under various operating conditions of transformers.

[0006] 3) Methods based on machine learning: Use neural networks or deep learning for fault classification. However, traditional deep learning methods have high computational complexity, strong dependence on large-scale labeled data, and insufficient generalization ability. In addition, many deep learning methods lack interpretability, making it difficult for diagnostic results to directly guide maintenance and fault prevention.

[0007] To address the above problems, the present invention proposes an intelligent diagnosis method for transformer partial discharge based on an improved transformation network, which can improve the feature extraction ability of discharge signals and enhance the generalization performance of the model. Summary of the Invention

[0008] In view of the problems that the existing transformer partial discharge diagnosis methods generally have insufficient feature extraction ability, single feature fusion method, and low diagnostic accuracy, the present invention is proposed.

[0009] Therefore, the problem to be solved by the present invention is how to improve the feature extraction and feature fusion capabilities of partial discharge signals by improving the transformation network model, so as to achieve efficient and accurate partial discharge fault diagnosis under complex operating conditions.

[0010] To solve the above technical problems, the present invention provides the following technical solutions:

[0011] In a first aspect, the present invention provides an intelligent diagnosis method for partial discharge of a transformer based on an improved transformation network, which includes:

[0012] Collect the partial discharge signals of the transformer, and preprocess the partial discharge signals to obtain a time series;

[0013] Use the Markov transition field MTF algorithm, the Gram angular field GASF algorithm, and the Gram angular difference field GADF algorithm to convert the time series into feature images respectively;

[0014] Process the feature images through an improved transformation network model to obtain a fused feature vector;

[0015] Input the fused feature vector into a fault diagnosis model, and output the fault diagnosis result of the partial discharge of the transformer.

[0016] As a preferred solution of the intelligent diagnosis method for partial discharge of a transformer based on an improved transformation network of the present invention, wherein: introduce a self-supervised-graph reasoning dual-drive framework to enhance the generalization ability of the improved transformation network model, and form a fault diagnosis model, including:

[0017] Randomly mask a preset proportion of time-frequency blocks of the feature image to obtain a masked feature image;

[0018] Reconstruct the masked feature image through a Transformer decoder to obtain a reconstructed feature image;

[0019] Apply random noise and scale transformation to the same partial discharge signal to generate positive sample pairs, and use different types of partial discharge signals as negative samples to construct a contrast learning sample set;

[0020] Based on the contrast learning sample set, train the improved transformation network model to obtain a fault diagnosis model;

[0021] Input the fused feature vector into the fault diagnosis model, and construct a dynamic causal graph network in the graph reasoning stage;

[0022] Set the discharge pulse characteristics, time-frequency energy distribution, and environmental parameters as graph nodes;

[0023] Calculate the causal relationship strength between the graph nodes through an attention mechanism to obtain node relationship weights;

[0024] Based on the node relationship weights, iteratively update the graph nodes, establish a fault propagation path, and output the fault diagnosis result.

[0025] As a preferred solution of the intelligent diagnosis method for transformer partial discharge based on the improved transformation network of the present invention, the method includes: processing the feature image through the improved transformation network model to obtain a fused feature vector, including:

[0026] Input the feature image into the improved transformation network model, where the improved transformation network model includes a time-frequency analysis branch and a time-domain analysis branch; the time-frequency analysis branch uses a spatial attention mechanism to extract three-dimensional time-frequency features; the time-domain analysis branch captures the time-domain features of partial discharge pulses through a depthwise separable dilated convolutional network, and constructs the improved transformation network model using a BiLSTM temporal attention.

[0027] Based on the three-dimensional time-frequency features and the time-domain features, adopt a heterogeneous feature alignment and fusion mechanism, and map the spatial dimension of the three-dimensional time-frequency features to the temporal dimension of the time-domain features through deformable convolution.

[0028] Adopt a multi-head cross-attention mechanism to establish cross-modal feature associations, and achieve dynamic weighted fusion of the three-dimensional time-frequency features and the time-domain features through attention weight calculation.

[0029] Perform residual connection on the feature after obtaining cross-attention and the time-frequency feature after dimension alignment, and obtain a fused feature vector through layer normalization processing.

[0030] As a preferred solution of the intelligent diagnosis method for transformer partial discharge based on the improved transformation network of the present invention, the feature image includes an MTF feature image, a GASF feature image, and a GADF feature image; the conversion method of the MTF feature image is as follows:

[0031] Discretize the time series into Q states, calculate the state transition probability, and generate a state transition matrix P, where the element of the state transition matrix represents the probability of transitioning from state i to state j.

[0032] Arrange all state transition probabilities in the order of state numbers to construct a Markov transition matrix.

[0033] As a preferred solution of the intelligent diagnosis method for transformer partial discharge based on the improved transformation network of the present invention, the conversion method of the GASF feature image is as follows:

[0034] Perform angle mapping on the signal of the time series to obtain a first angle-mapped signal ;

[0035] Based on the first angle-mapped signal , construct a Gram sum matrix, and the calculation formula of the elements of the Gram sum matrix is as follows:

[0036] ;

[0037] where, is the element in the a-th row and b-th column of the Gram angle sum field matrix, is the time point corresponding angle, is the time point corresponding angle, is the timestamp of the i-th data point in the original time series, is the timestamp of the j-th data point in the original time series.

[0038] As a preferred scheme of the intelligent diagnosis method for transformer partial discharge based on the improved transformation network described in the present invention, wherein: the conversion method of the GADF feature image is

[0039] Convert the signal of the time series into a second angle mapping signal through the inverse cosine function;

[0040] Based on the second angle mapping signal, construct a Gram difference matrix, and the calculation formula of the elements of the Gram difference matrix is as follows:

[0041] ;

[0042] where, is the element in the a-th row and b-th column of the Gram angle difference field matrix, is the time point corresponding angle, is the time point corresponding angle, is the timestamp of the i-th data point in the original time series, is the timestamp of the j-th data point in the original time series.

[0043] As a preferred scheme of the intelligent diagnosis method for transformer partial discharge based on the improved transformation network described in the present invention, wherein: the acquisition method of the time series is

[0044] Synchronously collect the transformer partial discharge signals by using ultrasonic detection, ultra-high frequency UHF detection and electromagnetic wave detection methods, and sequentially perform Gaussian filtering and adaptive denoising processing on the partial discharge signals to obtain the denoised time-domain signal;

[0045] Perform dynamic threshold segmentation on the time-domain signal, extract the discharge pulse segments, and generate a time series through wavelet transform.

[0046] In a second aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program instructions are executed by the processor, the steps of the intelligent diagnosis method for transformer partial discharge based on an improved transformation network as described in the first aspect of the present invention are implemented.

[0047] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and: when the computer program instructions are executed by the processor, the steps of the intelligent diagnosis method for transformer partial discharge based on an improved transformation network as described in the first aspect of the present invention are implemented.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] By adopting a multi-modal time-frequency feature joint representation method, fusing three representation mechanisms of Markov transition field, Gram angular field and Gram angular difference field, and deeply analyzing the signal from the dimensions of state transition probability, phase change law and energy difference respectively, a robust representation of partial discharge signals under strong noise interference is realized, effectively improving the discriminant ability of the signals and providing more discriminative input features for subsequent diagnosis;

[0050] An improved heterogeneous transformation network structure is designed. The time-frequency analysis branch based on the spatial attention mechanism and the time-domain analysis branch combining depthwise separable dilated convolution and temporal attention mechanism are processed collaboratively. The deformable convolution is used to align the branch features, and then the multi-head cross-attention mechanism is used to achieve dynamic weighted fusion. This structure can fully capture the complex coupling characteristics of partial discharge signals in the spatial and time domains, significantly improving the model's recognition ability for fine-grained features and the overall diagnosis accuracy;

[0051] Self-supervised learning and graph reasoning mechanisms are introduced. Masked reconstruction and contrast learning are used to enhance the discriminability of features, and a dynamic causal graph network is constructed to iteratively update the fault propagation path. This dual-drive mechanism not only improves the model's generalization ability for unknown devices or scenarios, but also significantly enhances the accurate positioning effect of the partial discharge source location, effectively supporting the cross-domain diagnosis requirements in a multi-device environment;

[0052] At the engineering application level, multiple signal preprocessing modules such as Gaussian filtering, adaptive denoising and wavelet transform are integrated, enhancing the adaptability to complex background noise; in the face of actual ultra-high frequency signals and strong signal-to-noise ratio interference, the system can still maintain a high recognition rate and a low false alarm rate, ensuring the stable operation of the system in a real power grid environment. Description of the Drawings

[0053] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings. Among them:

[0054] Figure 1 It is a flowchart of the intelligent diagnosis method for partial discharge of a transformer based on an improved transformation network in Embodiment 1. Specific embodiments

[0055] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings of the specification.

[0056] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0057] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or selectively exclusive embodiment from other embodiments.

[0058] Embodiment 1

[0059] Refer to Figure 1 , which is the first embodiment of the present invention. This embodiment provides an intelligent diagnosis method for partial discharge of a transformer based on an improved transformation network, including:

[0060] S1: Collect the partial discharge signals of the transformer and preprocess the partial discharge signals to obtain a time series.

[0061] Specifically, ultrasonic detection, ultra-high frequency (UHF) detection, and electromagnetic wave detection methods are used to synchronously collect the partial discharge signals of the transformer, and the partial discharge signals are sequentially subjected to Gaussian filtering and adaptive denoising processing to obtain the denoised time-domain signal.

[0062] Furthermore, dynamic threshold segmentation is performed on the time-domain signal to extract the discharge pulse segments, and a time series is generated through wavelet transform.

[0063] It should be noted that the time series includes the energy distribution information of the signal in the frequency band of 0.1 - 100 MHz.

[0064] S2: Use the Markov Transition Field (MTF) algorithm, Gram Angular Summation Field (GASF) algorithm, and Gram Angular Difference Field (GADF) algorithm to convert the time series into feature images respectively.

[0065] Specifically, the feature images include MTF feature images, GASF feature images, and GADF feature images.

[0066] Furthermore, the conversion method of the MTF feature image is to discretize the time series into Q states, calculate the state transition probability, and generate the state transition matrix P, where the elements of the state transition matrix represent the probability of transitioning from state i to state j.

[0067] It should be noted that the value range of the time series is evenly divided into Q intervals, and each interval represents a state. Assuming Q = 4, the time series will be divided into four intervals, and each interval corresponds to a state {1, 2, 3, 4}.

[0068] Preferably, by traversing all the time series, count the number of times transitioning from state i to state j, and calculate the transition probability of each state. The specific formula for the state transition probability is as follows:

[0069] ;

[0070] where, is the probability of transitioning from state i to state j, N is the total length of the time series, that is, the number of observed data points, and are two discrete states in the state space respectively, is the state of the system at time point t, is the indicator function, which takes the value of 1 when the condition is true, otherwise 0.

[0071] Exemplarily, for the numerator: calculate the number of times transitioning from state to state in the partial discharge signal, that is, the number of all transitions in the time series that satisfy "being in state i at the current moment and transitioning to state j at the next moment"; for the denominator: calculate the number of times state appears in the partial discharge signal, that is, count the total number of times state i appears in the time series (excluding the last moment t = N because there is no subsequent transition).

[0072] Even further, arrange all the state transition probabilities according to the state numbers to construct the Markov transition matrix. The elements of the Markov transition matrix are calculated by the following formula:

[0073] ;

[0074] where, is the state transition probability from coding time point i to coding time point j, forming a two-dimensional feature image. is the basic state transition probability, obtained through historical data statistics. is the discrete state of the i-th transition state in the partial discharge signal. is the discrete state of the j-th transition state in the partial discharge signal.

[0075] Specifically, the conversion method of the GASF feature image is to map the signal of the time series to obtain the first angle mapping signal , and the specific formula is as follows:

[0076] .

[0077] Furthermore, based on the first angle mapping signal , construct the Gram sum matrix, and the calculation formula of the elements of the Gram sum matrix is as follows:

[0078] ;

[0079] where, is the element in the a-th row and b-th column of the Gram angle sum field matrix, is the angle corresponding to the time point , is the angle corresponding to the time point , is the time stamp of the i-th data point in the original time series, is the time stamp of the j-th data point in the original time series.

[0080] Specifically, the conversion method of the GADF feature image is to convert the signal of the time series to the second angle mapping signal through the inverse cosine function, and the specific formula is as follows:

[0081] .

[0082] Furthermore, based on the second angle mapping signal, construct the Gram difference matrix, and the calculation formula of the elements of the Gram difference matrix is as follows:

[0083] ;

[0084] where, is the element in the a-th row and b-th column of the Gram angle difference field matrix, is the angle corresponding to the time point , is the angle corresponding to the time point , is the timestamp of the i-th data point in the original time series, is the timestamp of the j-th data point in the original time series.

[0085] Furthermore, through the spatial attention mechanism, the MTF feature image, GASF feature image, and GADF feature image are weighted and fused to output a three-dimensional time-frequency - time-domain joint feature image.

[0086] S3: Process the feature image through the improved transformation network model to obtain a fused feature vector.

[0087] Specifically, input the feature image into the improved transformation network model, where the improved transformation network model has a time-frequency analysis branch and a time-domain analysis branch.

[0088] Preferably, the time-frequency analysis branch uses the spatial attention mechanism to extract three-dimensional time-frequency features; the time-domain analysis branch captures the time-domain features of partial discharge pulses through a depthwise separable dilated convolutional network, and uses BiLSTM temporal attention to construct the improved transformation network model.

[0089] Further, based on the three-dimensional time-frequency features and time-domain features, an heterogeneous feature alignment and fusion mechanism is adopted, and the spatial dimension of the three-dimensional time-frequency features is mapped to the temporal dimension of the time-domain features through deformable convolution of the time-domain features.

[0090] It should be noted that the query matrix , the key matrix and the value matrix are calculated, where , , are weight matrices, which are optimized through end-to-end training.

[0091] Furthermore, a multi-head cross-attention mechanism is adopted to establish cross-modal feature associations, and dynamic weighted fusion of three-dimensional time-frequency features and time-domain features is achieved through attention weight calculation.

[0092] Excellent, the relevant formula for cross-modal feature association is as follows:

[0093] ;

[0094] where, is the output result of the attention mechanism, is the query matrix, which is used to retrieve information related to the current task in the input sequence, is the key matrix, which is matched with the query matrix to determine the weight allocation at different positions, is the transposed form of the key matrix, which is used for matrix multiplication operation with the query matrix, is a value matrix that stores the actual information content of the input sequence. is a key vector The dimension size of is the normalized exponential function that maps a real vector to a probability distribution.

[0095] Specifically, perform a residual connection on the cross-attention obtained features and the time-frequency features after dimension alignment, and obtain the fused feature vector through layer normalization processing to achieve a three-dimensional joint representation of time-frequency-space.

[0096] Preferably, the specific formula for the fused feature vector is as follows:

[0097] ;

[0098] Among them, is the fused feature vector, is the layer normalization operation, is the three-dimensional time-frequency feature, is the cross-modal or cross-level cross feature.

[0099] S4: Input the fused feature vector into the fault diagnosis model and output the fault diagnosis result of the transformer partial discharge.

[0100] Specifically, the construction method of the fault diagnosis model is to introduce a self-supervised-graph reasoning dual-driven framework to enhance the generalization ability of the improved transformation network model to form the fault diagnosis model.

[0101] Furthermore, randomly mask a preset proportion of time-frequency patches of the feature image to obtain the masked feature image; reconstruct the masked feature image through the Transformer decoder to obtain the reconstructed feature image.

[0102] Preferably, the relevant formula for reconstructing the original signal through the Transformer decoder to force the network to learn the inherent pattern of the discharge signal is as follows:

[0103] ;

[0104] Among them, is the reconstruction ability loss of the model for the input data, is the total number of samples in the training dataset, is the original time-frequency feature vector of the i-th sample, is the reconstructed time-frequency feature vector of the i-th sample.

[0105] Even further, apply random noise and scale transformation to the same partial discharge signal to generate positive sample pairs, use different categories of partial discharge signals as negative samples, construct a contrastive learning sample set, and optimize the feature discriminability. The relevant formula is as follows:

[0106] ;

[0107] Among them, is the contrast loss function value, B is the total number of samples in the training batch, is the projected feature vector of the i-th positive sample pair generated after different data augmentations of the same original sample, is the projected feature vector of the j-th positive sample pair generated after different data augmentations of the same original sample, is the projected feature vector of other samples in the current training batch, is the temperature hyperparameter, used to adjust the distribution of similarity scores.

[0108] Specifically, based on the contrast learning sample set, the improved transformation network model is trained to obtain the fault diagnosis model.

[0109] Furthermore, the fused feature vector is input into the fault diagnosis model, and a dynamic causal graph network is constructed in the graph inference stage; the discharge pulse feature, time-frequency energy distribution, and environmental parameters are set as graph nodes.

[0110] Even further, the causal relationship strength between graph nodes is calculated through the attention mechanism to obtain the node relationship weight. The relevant formula is as follows:

[0111] ;

[0112] Among them, is the attention score between the i-th element and the j-th element in the input sequence, is the permutation of the feature vector of the i-th element in the input sequence, is the learnable weight matrix, is the feature vector of the j-th element in the input sequence, is the vector and is the dimension of.

[0113] Specifically, based on the node relationship weight, the graph nodes are iteratively updated to establish the fault propagation path and output the fault diagnosis result.

[0114] Preferably, the fault diagnosis result includes the fault type probability distribution and the fault location result; the node representation is iteratively updated to capture the fault propagation path. The specific formula is as follows:

[0115] ;

[0116] Among them, is the feature representation of the node at the l+1 layer after the l-th layer of graph convolution operation, is the non-linear activation function, is the degree matrix, which is a diagonal matrix, is the adjacency matrix, is a learnable parameter matrix used for linearly transforming node features.

[0117] In an alternative embodiment, accurate determination of the fault diagnosis result is achieved through a dynamic causal graph network, including: the input is a 512-dimensional fused feature vector, and the output is the probability distribution of 5 types of faults (such as insulation deterioration, winding deformation, etc.); during the processing, the fused feature is decomposed into three types of graph nodes: discharge pulses (256-dimensional time-domain features), time-frequency energy distribution (128-dimensional frequency-band energy features), and environmental parameters (128-dimensional sensor data); by constructing an initial node matrix, these nodes are used as inputs for subsequent processing; the causal relationship weights between nodes are calculated using the multi-head attention mechanism, and the edges with weights greater than 0.5 are selected as the key causal paths; through the iterative update of three graph convolutional layers, the node features are gradually optimized; the updated node features are concatenated with the global features and input into the classification layer to generate the fault probability output.

[0118] Exemplarily, for a certain 220 kV transformer under high-temperature (45 °C) conditions, the fused feature analysis shows that the high-frequency energy accounts for 85% and the pulse amplitude is 120 mV; through the analysis of the dynamic causal graph network, the fault propagation path is obtained as: "temperature → oil gap insulation degradation → high-frequency energy surge → pulse mutation"; after the analysis of this path, the output probability of oil gap discharge is 85%, and the false alarm rate is lower than 2%, and the positioning error is only ±5 cm, which is three times higher than the positioning accuracy of the traditional method (±15 cm).

[0119] In summary, through the interpretable graph reasoning mechanism that combines environmental parameters and signal features, the present invention achieves an oil gap discharge recognition rate of 98% in the test. This technology enables the fault diagnosis to leap from the traditional threshold-based judgment mode to the intelligent mode based on causal logic reasoning, effectively improving the fault recognition ability and diagnosis efficiency of the system.

[0120] This embodiment also provides a computer device applicable to the case of the intelligent diagnosis method for partial discharge of a transformer based on an improved transformation network, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the intelligent diagnosis method for partial discharge of a transformer based on the improved transformation network as proposed in the above embodiment.

[0121] The computer device may be a terminal, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner, and the wireless manner can be achieved through WIFI, carrier network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0122] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, the following steps are implemented: collecting partial discharge signals of a transformer, and preprocessing the partial discharge signals to obtain a time series; respectively converting the time series into feature images by using the Markov transition field MTF algorithm, the Gram angle field GASF algorithm, and the Gram angle difference field GADF algorithm; processing the feature images through an improved transformation network model to obtain a fused feature vector; inputting the fused feature vector into a fault diagnosis model to output a fault diagnosis result of the partial discharge of the transformer.

[0123] Embodiment 2

[0124] Referring to Tables 1 to 4, this is the second embodiment of the present invention. This embodiment provides an intelligent diagnosis method for partial discharge of a transformer based on an improved transformation network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0125] Specifically, taking the partial discharge diagnosis of a 500kV substation transformer as an example for verification; three detection methods (ultrasonic, ultra-high frequency UHF, electromagnetic wave) are used in the test to synchronously collect partial discharge signals to form a multi-modal data set, including 1,200 groups of samples from a certain 500kV substation of the State Grid (covering 5 types of faults such as oil bubbles, floating discharges, conical corona, creeping discharges, and surface discharges) and 3,500 groups of samples from the IEEE PES public data set (8 types of fault types); in the signal preprocessing link, Gaussian filtering (cutoff frequency 15kHz) combined with empirical mode decomposition (EEMD) is used to achieve noise suppression, and time-frequency features are extracted through wavelet transform. When the signal is converted into a feature image, the MTF algorithm uses 32 discrete state partitions, and both GASF and GADF use a resolution of 256×256.

[0126] Furthermore, the time-frequency analysis branch of the improved transformation network model uses 3 layers of depthwise separable convolutions with attention enhancement (kernel size 3×3, dilation rates 1, 2, 4), and the time-domain analysis branch uses 2 layers of BiLSTM (number of hidden units 128); the training uses a self-supervised-graph reasoning dual-drive framework, the masking ratio is set to 15%, and the construction of positive sample pairs uses Gaussian noise (σ = 0.1) and random scaling (0.8 - 1.2 times) perturbations. 21 nodes are set in the dynamic causal graph network, including 12 discharge pulse feature nodes, 6 time-frequency energy distribution nodes, and 3 environmental parameter nodes (temperature, humidity, oil pressure), and multi-head attention (number of heads = 8) is used to calculate the relationship weights between nodes.

[0127] Furthermore, as shown in Table 1, the present invention is significantly superior to the prior art methods in terms of multiple key indicators. In terms of diagnostic accuracy, the present invention reaches 98.7% in a 20dB noise environment, a 20.4 percentage point increase compared to the traditional network method, and a 5.2 percentage point increase compared to the state-of-the-art ViT-Base model; in the more challenging 10dB noise environment, the accuracy of the present invention is 95.3%, showing extremely strong anti-noise performance, 11.3 percentage points higher than ViT-Base. Especially in the cross-device generalization ability test, the accuracy of the present invention is 92.3%, far higher than 72.3% of the second-place ViT-Base, indicating that the self-supervised-graph reasoning dual-drive framework of the present invention has significant advantages in terms of generalization ability.

[0128] Table 1. Data comparison table between the present invention and the prior art

[0129]

[0130] Specifically, in terms of fault location accuracy, the average location error of the present invention is only ±4.2 cm, a 77.3% reduction compared to ±18.5 cm of the traditional network method. Specifically for different fault types, as shown in Table 2, the location errors of the present invention for oil-immersed discharge, conical corona, and creeping discharge are ±2.0 cm, ±3.5 cm, and ±4.8 cm respectively, all significantly lower than the traditional methods. This benefits from the collaborative reasoning mechanism between the environmental parameter nodes and the time-frequency energy nodes in the dynamic causal graph network.

[0131] Table 2. Comparison table of fault location accuracy

[0132]

[0133] Furthermore, in terms of real-time performance and resource consumption, as shown in Table 3, the single-sample inference time of the present invention is only 20 ms, which is 86.7% lower than 150 ms of the traditional network method and 83.3% lower than 120 ms of CNN-LSTM; the GPU video memory occupancy is 1.2 GB, which is only 31.6% of CNN-LSTM; the CPU utilization rate is 35%, far lower than 95% of the traditional network method and 80% of CNN-LSTM. This indicates that the present invention has successfully achieved model lightweighting through depthwise separable dilated convolution and BiLSTM temporal attention mechanism, and the number of parameters is only 18.7 MB, which is 5.7% of ResNet-50 and 4.5% of ViT-Base.

[0134] Table 3. Comparison Table of Real-time Performance and Resource Consumption

[0135]

[0136] Furthermore, in the verification of the public dataset, as shown in Table 4, the present invention has improved by 5.6%, 5.7% and 5.8% respectively in the three indicators of accuracy, recall rate and F1-Score compared with the current SOTA method, fully verifying the advancement and effectiveness of the present method.

[0137] Table 4. Verification Table of Public Dataset

[0138]

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An intelligent diagnosis method for partial discharge of transformers based on an improved transformation network, characterized in that: including, collecting partial discharge signals of a transformer, preprocessing the partial discharge signals to obtain a time series; using the Markov transition field (MTF) algorithm, the Gram angle field (GASF) algorithm, and the Gram angle difference field (GADF) algorithm to convert the time series into feature images respectively; processing the feature images through an improved transformation network model to obtain fused feature vectors; inputting the fused feature vectors into a fault diagnosis model to output a fault diagnosis result of the partial discharge of the transformer; introducing a self-supervised graph reasoning dual-drive framework to enhance the generalization ability of the improved transformation network model to form a fault diagnosis model, including: randomly masking a preset proportion of time-frequency patches of the feature images to obtain masked feature images; reconstructing the masked feature images through a Transformer decoder to obtain reconstructed feature images; generating positive sample pairs by applying random noise and scale transformation to the same partial discharge signal, and using different categories of partial discharge signals as negative samples to construct a contrast learning sample set; training the improved transformation network model based on the contrast learning sample set to obtain a fault diagnosis model; inputting the fused feature vectors into the fault diagnosis model and constructing a dynamic causal graph network in the graph reasoning stage; setting discharge pulse features, time-frequency energy distribution, and environmental parameters as graph nodes; calculating the causal relationship strength between the graph nodes through an attention mechanism to obtain node relationship weights; iteratively updating the graph nodes based on the node relationship weights, establishing a fault propagation path, and outputting a fault diagnosis result; processing the feature images through an improved transformation network model to obtain fused feature vectors, including: inputting the feature images into the improved transformation network model, where the improved transformation network model has a time-frequency analysis branch and a time-domain analysis branch; the time-frequency analysis branch uses a spatial attention mechanism to extract three-dimensional time-frequency features; the time-domain analysis branch captures the time-domain features of partial discharge pulses through a depthwise separable dilated convolutional network and constructs the improved transformation network model using a bidirectional long short-term memory (BiLSTM) time series attention; Based on the three-dimensional time-frequency features and the time-domain features, a heterogeneous feature alignment and fusion mechanism is adopted, and the spatial dimension of the three-dimensional time-frequency features is mapped to the time sequence dimension of the time-domain features through deformable convolution ; ​ using a multi-head cross-attention mechanism to establish cross-modal feature associations, and realizing dynamic weighted fusion of the three-dimensional time-frequency features and time-domain features through attention weight calculation; performing residual connection on the features after cross-attention and the time-frequency features after dimension alignment, and obtaining fused feature vectors through layer normalization processing; the feature images include MTF feature images, GASF feature images, and GADF feature images; the conversion method of the MTF feature images is Discretize the time series into Q states, calculate the state transition probability, and generate a state transition matrix P, where the elements of the state transition matrix represent the probability of transitioning from state i to state j; arranging all state transition probabilities according to state numbers to construct a Markov transition matrix.

2. The intelligent diagnosis method for partial discharge of transformer based on improved transformation network according to claim 1, characterized in that: The conversion method of the GASF feature images is For the signal of the time series perform angle mapping to obtain the first angle-mapped signal ; Based on the first angle mapping signal , a Gram summation matrix is constructed, and the calculation formula for the elements of the Gram summation matrix is as follows: ； · Among them, is the element at the a-th row and b-th column of the Gram angle and field matrix, is the time point corresponding angle, is the time point corresponding angle, is the timestamp of the i-th data point in the original time series, is the timestamp of the j-th data point in the original time series.

3. The intelligent diagnosis method for partial discharge of transformer based on improved transformation network according to claim 1, characterized in that: The conversion method of the GADF feature images is Convert the signal of the time series into a second angle mapping signal through the arccosine function; Based on the second angle mapping signal, constructing a Gram difference matrix, where the calculation formula of the elements of the Gram difference matrix is as follows: ； wherein, is the element at the a-th row and b-th column of the Gram angular difference field matrix, is the time point corresponding angle, is the time point corresponding angle, is the timestamp of the i-th data point in the original time series, is the timestamp of the j-th data point in the original time series.

4. The intelligent diagnosis method for partial discharge of transformer based on improved transformation network according to claim 3, characterized in that: The acquisition method of the time series is The partial discharge signals of the transformer are synchronously collected by using ultrasonic detection, ultra-high frequency (UHF) detection and electromagnetic wave detection methods, and the partial discharge signals are successively subjected to Gaussian filtering and adaptive denoising processing to obtain the denoised time-domain signal; The time-domain signal is subjected to dynamic threshold segmentation to extract the discharge pulse segments, and a time series is generated through wavelet transform.

Images