Transformer early fault diagnosis method based on cross fusion of three-mode monitoring data

Abstract

The invention relates to a transformer early-stage fault diagnosis method based on cross fusion of three-mode monitoring data, and the method comprises the steps: collecting the three-mode monitoring data, and extracting a magnetic flux leakage feature, a vibration feature and an ultrasonic feature; respectively inputting the vibration features and the ultrasonic features into dense connection networks with the same structure, respectively generating intermediate features of the corresponding levels, and fusing the intermediate features of the corresponding levels by using an attention mechanism between the dense connection networks to guide the generation of the intermediate features of the next level, so as to achieve the fusion of the intermediate features of the next level. Respectively generating end features based on the output of the last level; inputting the magnetic flux leakage characteristics into a residual neural network to generate end characteristics; and performing end fusion on all the end features to generate a fusion vector, and generating an early fault diagnosis result of the transformer based on the fusion vector. Compared with the prior art, the method has the advantages that deep feature fusion is carried out on the three-modal data, so that more early signals of the transformer can be reserved by the fused features, and more accurate and reliable fault diagnosis is realized.

Description

Technical Field

[0001] This invention relates to the field of transformer fault diagnosis technology, and in particular to a method for early fault diagnosis of transformers based on cross-fusion of three-modal monitoring data. Background Technology

[0002] For early fault diagnosis of transformers, the following are some common monitoring techniques and analysis methods: (A) Oil chromatography analysis: Insulating oil gas analysis mainly assesses the transformer's operating status by detecting the types and concentration changes of characteristic gases dissolved in the insulating oil. During normal operation, the oil and insulating materials of the transformer will slowly age and release trace amounts of gases, such as hydrogen, methane, acetylene, ethylene, ethane, and carbon monoxide. At this time, the oxygen and nitrogen content in the gas composition is relatively high. When an internal abnormality occurs, the content of these gases will rise rapidly. By separating the gas components through oil chromatography technology, combining sensor data, analyzing their concentration and comparing it with the benchmark value, an alarm is triggered and maintenance is prompted once an abnormality is detected; (B) Frequency response analysis: The frequency response method obtains the response curve by injecting signals of different frequencies into the transformer windings. When the winding is deformed, loosened, or the support structure is abnormal, its equivalent electrical parameters (such as inductance, capacitance, and coupling coefficient) will change, which will affect the amplitude and phase distribution of the frequency response curve. By comparing the difference between the response curves under normal and fault conditions, the winding fault can be detected and located. (C) Online temperature monitoring method: Temperature is a key indicator of the thermal state of the transformer. By arranging temperature sensors in the transformer windings, oil tank, and other parts and collecting data, it is possible to quickly identify and warn of abnormal temperature rise. This method can judge potential fault risks through temperature change trends and automatically alarm based on set thresholds; (D) Partial discharge online monitoring technology: Partial discharge monitoring is an important means of evaluating the insulation performance of transformers. By installing specific sensors, partial discharge signals are collected in real time and trend analysis is performed, which can effectively detect insulation aging and abnormal phenomena. Commonly used technical means include pulse current method, ultra-high frequency method, radio frequency method, ultrasonic method, optical method, gas chromatography method, infrared thermal imaging method, etc. Among them, the pulse current method is widely used because of its rich information and strong quantification ability; (E) Vibration signal online monitoring method: Vibration signal monitoring is used for online condition assessment of power transformers. It is a non-invasive, non-destructive testing method that does not require shutdown. The main vibration sources inside the transformer are the core and windings: the silicon steel sheets of the core generate magnetostrictive vibration under the action of alternating magnetic flux, and the windings generate Ampere force due to the interaction between current and leakage magnetic field, resulting in periodic mechanical vibration. The main frequencies of the vibrations of both are concentrated in 100Hz. Vibration characteristics are closely related to the structural state of the winding. When faults such as winding turn displacement or weakening of preload occur, they usually cause an increase in vibration amplitude and enhancement of high-frequency components. (F) Distributed optical fiber online monitoring technology: Distributed optical fiber monitoring technology is an emerging means of power equipment condition sensing. It relies on optical fiber as the sensing medium and uses nonlinear optical effects such as Brillouin Raman scattering or Rayleigh scattering to achieve long-distance, high-resolution temperature and strain measurement. It has strong anti-electromagnetic interference capability, small size and flexible installation, and is suitable for transformer condition monitoring in complex environments. In particular, time domain reflectometry (BOTDR) based on Brillouin scattering is widely used for temperature / strain monitoring.

[0003] As can be seen, the above analyses are mostly based on single-type data and lack the joint utilization of multi-source information. To address this issue, Chinese patent application CN111289829A provides a transformer fault identification method that integrates transformer vibration signals, ultrasonic signals, and winding temperatures. Although it can solve the problem of low reliability caused by relying on a single mode for fault identification, it calculates the probability of each mode independently and then fuses the probabilities when processing multi-modal data, without considering the complementarity or correlation between modes. Furthermore, this process relies on manually set thresholds and probability functions, which means that the generated fused features cannot retain early fault signals of the transformer, resulting in low reliability and accuracy of the final fault diagnosis results.

[0004] Therefore, there is an urgent need for a method for early fault diagnosis of transformers that integrates multiple types of monitoring data and has deep feature interaction capabilities, so as to improve the detection performance of early faults in transformers. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for early fault diagnosis of transformers based on cross-fusion of three-modal monitoring data. By fusing intermediate-level features and terminal features of the three-modal data respectively, the early fault features of the transformer can be deeply correlated and complemented, thereby improving the accuracy and reliability of early fault diagnosis.

[0006] The objective of this invention can be achieved through the following technical solutions: This invention provides a method for early fault diagnosis of transformers based on cross-fusion of three-modal monitoring data, the method comprising: Three-mode monitoring data of the transformer during operation are collected, including leakage flux signal, vibration signal and ultrasonic signal, and feature extraction is performed to obtain leakage flux feature, vibration feature and ultrasonic feature; Vibration features and ultrasonic features are respectively input into a densely connected network with the same structure to generate intermediate features at the corresponding level. The attention mechanism between the densely connected networks is used to fuse the intermediate features at the corresponding level to guide the generation of intermediate features at the next level. Based on the output of the last level, vibration end features and ultrasonic end features are generated. The magnetic leakage features are input into a residual neural network to generate magnetic leakage end features. The leakage magnetic field end features, vibration end features and ultrasonic end features are fused to generate a fusion vector, and the early fault diagnosis results of the transformer are generated based on the fusion vector.

[0007] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention provides a method for early transformer fault detection. First, a dense connection network and a residual network are used to extract deep features from vibration, ultrasonic, and leakage magnetic signals, respectively. Then, intermediate layer feature fusion and terminal feature fusion are performed. Specifically, for vibration and ultrasonic signals that are highly correlated at the physical level, intermediate layer feature interaction is prioritized, so that the terminal features of the two modes extracted at the end can retain richer early micro-fault features. At the end of the three network models, multi-head attention execution is introduced to allow the features of the three modes to take turns as queries, keys, and values ​​for multiple cross-calculations, realizing cross-physical field feature enhancement, fully exploring the complementarity and correlation between modes, thereby generating a highly integrated fusion representation of early transformer fault features. This avoids the dependence on manual threshold setting, solves the problem of insufficient information utilization in complex fault mode recognition by traditional networks, and significantly improves the fault diagnosis accuracy.

[0008] 2) This invention addresses the problem of scarce early-stage transformer fault samples by proposing the use of Conditional Gradient Penalized Adversarial Network (WCGAN-GP) to generate high-quality fault samples. It also uses 1-Nearest Neighbor (1-NN) classification accuracy and Dynamic Time Warping (DTW) average distance as evaluation metrics. The aim is to provide a large number of reliable training datasets for the network model, enabling the network model to be trained more fully and effectively. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the network framework of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a schematic diagram of the transformer measuring points of the present invention; Figure 4 This is a schematic diagram of the transformer leakage flux signal processing of the present invention; Figure 5 This is a schematic diagram of the vibration signal of the present invention; (5a) shows the vibration signal waveform; (5b) shows the RPM encoding junction; Figure 6 This is a schematic diagram of the ultrasonic signal of the present invention; (6a) represents the ultrasonic signal waveform diagram; (6b) represents the CWT time-frequency diagram; Figure 7 This is a schematic diagram of the ECA channel attention of the present invention; Figure 8 This is a schematic diagram of the dense network embedding ECA module of the present invention; Figure 9 (9a) is a schematic diagram of the residual neural network embedded ECA module of the present invention; (9b) is a schematic diagram of the identity mapping residual block embedded ECA module; Figure 10 This is a diagram of the WCGAN-GP network structure of the present invention; Figure 11 This is the wiring diagram for the dynamic model experiment of the present invention; Figure 12 This is a comparison chart of measured and simulated magnetic flux leakage of the present invention; Figure 13 The curve showing the change of the loss function of the WCGAN-GP network in this invention; Figure 14 This is an example of a two-dimensional feature map of vibration RPM and ultrasonic CWT according to the present invention; Figure 15 (15a) represents the accuracy and loss variation curves of the test set in this invention; (15b) represents the accuracy variation curves of the training set and the test set; Figure 16 The t-SNE clustering visualization results of the test set diagnostic results of the present invention are shown in Figure 16a; (16b) shows the visualization results of the original input features; and (16a) shows the visualization results of the final output layer features. Figure 17 This is a schematic diagram of the confusion matrix of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0011] Deep learning neural networks for transformer fault diagnosis suffer from several drawbacks, including a lack of real-world fault samples, limited training sample scalability, and insufficient exploitation of the correlation and complementary features of multimodal training data. Therefore, this invention proposes a training sample expansion method based on the Wasserstein distance-based conditional gradient penalty adversarial network WCGAN-GP and the DTW-1NN verification method. Simultaneously, a transformer early fault diagnosis model is constructed using a cascaded network of DenseNet-121 and ResNet-18 embedded with a multi-layer cross-attention mechanism. After training with the expanded training set, this model is used for early transformer fault diagnosis. The overall process framework is as follows: Figure 1 As shown, the model includes a model training section and a fault diagnosis section. The detailed process is described in detail using the following embodiments.

[0012] Example 1 This embodiment provides a method for early fault diagnosis of transformers based on cross-fusion of three-modal monitoring data. The process is as follows: Figure 2As shown, the method includes: S1. Collect three-mode monitoring data of the transformer during operation, including leakage magnetic field signal, vibration signal and ultrasonic signal, and perform feature extraction to obtain leakage magnetic field feature, vibration feature and ultrasonic feature.

[0013] S11, Extraction of magnetic flux leakage features.

[0014] When a transformer malfunctions, the distribution of the spatial leakage magnetic field is affected by the winding current density, which contains rich amplitude and phase information. To accurately determine the transformer's operating status, three measuring points are used to monitor its condition. The distribution of these measuring points is as follows: Figure 3 As shown; leakage magnetic field signals were collected from the upper, middle, and lower parts of the transformer winding. A sliding window method was used to collect one cycle of data from each of the upper, middle, and lower parts, and the data were then stitched together. Mean-minimum normalization was then applied to eliminate the influence of amplitude differences on the model, resulting in: , in, Indicates the first The amplitude of each leakage magnetic field sampling point; This represents the average value of the leakage magnetic field amplitude; This indicates the maximum value of the leakage flux amplitude; This represents the minimum value of the leakage flux amplitude.

[0015] A schematic diagram of the leakage magnetic field feature extraction method is shown below. Figure 4 As shown, waveform data of the leakage magnetic field at three measuring points (upper, middle, and lower) of the transformer were simultaneously acquired. The aforementioned sliding window splicing and normalization method was applied to both normal and fault states to construct a leakage magnetic field sample dataset. In the figure, the dashed box represents the range of the sliding window, and the solid line represents the time node of the fault occurrence; the period before the solid line represents the normal operating state of the transformer, and the period after the solid line represents the fault state. The leakage magnetic field data after feature extraction retains the temporal information of the leakage magnetic field at each measuring point, which helps improve the accuracy of fault identification.

[0016] S12, Vibration feature extraction.

[0017] The vibration signals of a transformer mainly come from the transformer windings and core, and the main frequencies of the vibration signals of both are concentrated at 100Hz, which are obviously periodic signals. In order to better extract the time structure and amplitude information in the transformer vibration signals, this embodiment uses the relative position matrix method to convert the one-dimensional vibration signal of a single cycle into a two-dimensional image form, which is then used as the input for the subsequent multimodal neural network branch.

[0018] Assume the original acquired transformer vibration signal of one cycle is as follows: To ensure uniformity of scale, each vibration signal was z-score normalized to obtain a standardized signal. To reduce computational load and control image size, the standardized vibration signal is dimensionality-reduced using a piecewise aggregation approximation. The entire time series is divided into several equal-length segments, and then the average is taken, which can be expressed as: , in, This represents the compression factor, which is set to 2 in this embodiment, meaning that every two sampling points are combined into one mean. This represents the amplitude of the vibration signal after dimensionality reduction.

[0019] Based on the above processing, the dimensionality-reduced sequence is obtained. Construct the RPM matrix using a dimensionality-reduced sequence: , Each element in the matrix represents the relative position between two timestamps.

[0020] Finally, the RPM is linearly normalized and mapped to the image pixel space. The resulting matrix F is the RPM image. The normalization formula is shown below: , This represents the element in the i-th row and j-th column of the RPM matrix.

[0021] The final result is as follows Figure 5 As shown, it represents a comparison between converting the data into a color two-dimensional image using RPM encoding and generating vibration data. Figure (5a) represents the original vibration signal, with the horizontal axis representing the number of sampling points and the vertical axis representing the amplitude of the vibration signal. Figure (5b) represents the RPM encoding result. The depth of the color in the encoded two-dimensional image represents the amplitude comparison relationship between different sampling points. The red area indicates that the current point has a larger amplitude relative to other sampling points, the blue area indicates a smaller amplitude, and the yellow and green areas correspond to the intermediate amplitude. Through the regularity and symmetry of the color distribution, the local changes, periodic structure, and overall trend characteristics in the generated vibration signal can be reflected.

[0022] S13, Ultrasonic Feature Extraction.

[0023] Ultrasonic signals are most prominent when local faults such as partial discharge and inter-turn short circuits occur during transformer operation. These signals typically exhibit short-term non-stationarity and transient impact characteristics, with a wide frequency range and rapid time-domain fluctuations. To more effectively extract the time-frequency structural features from the transformer's ultrasonic signals, a continuous wavelet transform (CWT) is used to convert the one-dimensional ultrasonic signal of the transformer into a two-dimensional time-frequency diagram.

[0024] Let the original acquired ultrasonic signal of the transformer be... First, perform Z-score normalization to obtain Then, using the Morlet wavelet, which is suitable for processing oscillating signals, as the mother wavelet function, a continuous wavelet transform is performed on the standardized signal to obtain the complex form of the time-frequency coefficient matrix. : ,in, This represents the standardized ultrasound signal; Represents the concomitant complex conjugate of the mother wavelet function; and These represent the scale parameter and the translation parameter, respectively.

[0025] Similarly, to facilitate input into the neural network model, the modulus values ​​of the normalized wavelet transform coefficients are color-mapped to obtain a two-dimensional CWT time-frequency image. A comparison of the transformer ultrasonic waveform and the two-dimensional time-frequency image generated using CWT is provided. Figure 6 As shown, Figure (6a) represents the original ultrasonic signal, and Figure (6b) is a CWT image. The image is plotted with time on the horizontal axis and frequency on the vertical axis. The brightness of the colors reflects the signal energy at different frequencies. It can intuitively reflect the non-stationarity, frequency distribution and local abnormalities of the ultrasonic signal, and has good time-frequency resolution.

[0026] S2, Trimodal Feature Processing.

[0027] After extracting features from leakage magnetic field signals, vibration signals, and ultrasonic signals, a ResNet-DenseNet-MLCA transformer fault diagnosis model was established.

[0028] Among them, the leakage magnetic mode is extracted by a one-dimensional residual neural network ResNet-18, while the vibration mode and ultrasonic mode are processed by two densely connected networks with the same structure, DenseNet-121. The fault diagnosis model introduces ECA attention mechanism in the residual neural network and densely connected network to enhance the expression of key features, and adopts multi-layer cross attention fusion (MLCA) strategy to achieve deep information interaction between multimodal features, and finally completes the accurate classification of fault state.

[0029] In detail, the ECA attention mechanism is a lightweight channel attention mechanism that adaptively models inter-channel dependencies through local one-dimensional convolution, improving feature representation capabilities while maintaining low computational cost. The principle diagram of the ECA module is shown below. Figure 7As shown, the main steps include the following: (1) Global average pooling: Perform global average operation on each channel of the input feature map to extract statistical information at the channel level and form a channel description vector; (2) One-dimensional convolution and Sigmoid: Use one-dimensional convolution to capture the local channel dependencies and generate normalized channel weight coefficients through the Sigmoid function; (3) Channel weighting: Multiply the generated channel weights with the original feature map channel by channel to achieve adaptive enhancement of the feature map in the channel dimension.

[0030] The detailed feature processing procedures for each modality are as follows: S21. Input the vibration features and ultrasonic features into a densely connected network with the same structure to generate intermediate features at the corresponding level. Use the attention mechanism between the densely connected networks to fuse the intermediate features at the corresponding level to guide the generation of intermediate features at the next level. Generate vibration end features and ultrasonic end features based on the output of the last level.

[0031] In this embodiment, the densely connected network includes multiple dense blocks and transition layers. Unlike the cross-layer connections of ResNet, the densely connected network achieves efficient feature reuse and propagation through dense connections within each dense block. Specifically, each dense block includes dense layers, with an ECA module embedded after each dense layer. Each dense layer not only receives the output of the previous layer but also receives the feature maps of all preceding layers as input, thereby enhancing information flow and gradient propagation based on feature cascading, improving network performance and training efficiency. The internal structure of each dense block is as follows: Figure 8 As shown, this structure enables the network to effectively enhance the propagation and reuse of features, alleviate the gradient vanishing problem, and improve the model's expressive power and training efficiency.

[0032] To extract the time-frequency features contained in the ultrasonic and vibration signals of the transformer, the RPM image of the vibration signal and the CWT image of the ultrasonic signal are used as inputs. Two DenseNet-121 networks with identical structures are used for feature extraction. In this embodiment, the entire DenseNet-121 network consists of an initial convolutional layer, four dense blocks, three transition layers, global average pooling, and a fully connected classification layer. The four dense blocks are composed of 6, 12, 24, and 16 dense units, respectively. The transition layers achieve feature compression and downsampling through 1×1 convolutional kernels and 2×2 average pooling, thereby effectively controlling the model complexity and reducing the risk of overfitting.

[0033] The detailed intermediate fusion method is as follows: S211, For two densely connected networks, the first... The intermediate features output from each dense block are average-pooled and projected into a low-dimensional representation of the same dimension; the intermediate features include vibration intermediate features and ultrasound intermediate features, and the low-dimensional representation includes vibration low-dimensional representation and ultrasound low-dimensional representation, the expression of which is: , , in, This indicates a global average pooling operation; The linear mapping matrix represents the channel dimension. and Become ; This indicates that the sequence dimension is increased to meet the input requirements of the Multilevel Cross-Attention (MHCA) mechanism.

[0034] S212. In each densely connected network, the first... The low-dimensional representations of vibration and ultrasound are used alternately as keys and values, respectively. Correspondingly, either the low-dimensional ultrasound representation or the low-dimensional vibration representation is used as the query. Multi-head cross-attention fusion is then performed, and the fusion result is compared with the first... The intermediate features output from the first dense block are used to generate enhanced features through residual connections, which serve as the first... The input of a dense block, to enhance its cross-modal sensing capability, is: , , in, Indicates vibration enhancement characteristics; Indicates the first The vibrational intermediate characteristics of the output of a dense block; Indicates the first A low-dimensional representation of the vibration of a dense block; Indicates the first A low-dimensional representation of a dense block of ultrasound; Indicates ultrasound enhancement characteristics; Indicates the first The intermediate ultrasonic features output by a dense block; This represents a multi-head attention fusion of the query and the low-dimensional representation of vibration as the key, and the low-dimensional representation of ultrasound as the value. This represents a multi-head attention fusion where the ultrasound low-dimensional representation is used as the key, and the vibration low-dimensional representation is used as the value and query.

[0035] Multi-head attention fusion maps input features to query vectors (Query, Q), key vectors (Key, K), and value vectors (Value, V), respectively. It then weights V by calculating the similarity between Q and K, thereby achieving information fusion between different inputs. The specific calculation formula is as follows: , in, Representing the key-value dimension, multi-head attention enhances expressive power through attention operations across multiple subspaces, with the overall form as follows: , , , , The learnable parameter matrix; Indicates a splicing operation; Represents the weights of a linear mapping; The first indicator of bullish attention Size.

[0036] S22. Input the leakage magnetic field feature into the residual neural network to generate the leakage magnetic field terminal feature.

[0037] To effectively extract the deep temporal features of transformer leakage magnetic field, a lightweight one-dimensional residual neural network (1D-ResNet-18) is used as the feature extraction sub-network for leakage magnetic field mode. It retains the core ideas of residual connections and batch normalization from the two-dimensional ResNet. The 1D-ResNet-18 network uses one-dimensional residual blocks as basic building blocks. Each residual block consists of two one-dimensional convolutional layers, a BN layer, and a ReLU activation function. Cross-layer connections are achieved through identity mapping or projection mapping. When the number of input and output channels and the sequence length are consistent, identity mapping is used. The output of the residual block is: , where x represents the input of the residual block. This represents the nonlinear transformation within the residual block. In a residual neural network, if the dimensions or lengths of the input and output are inconsistent, a projection mapping is introduced to ensure that the dimensions are the same. In this case, the output of the residual block can be represented as: , It typically consists of a 1×1 one-dimensional convolutional layer with stride and a BN layer, used to adjust the number of channels and length of the input x to keep it consistent in dimension with the output of the residual main branch.

[0038] In this embodiment, the residual neural network includes multiple residual blocks, each containing multiple convolutional layers and activation functions. The residual blocks are divided into identity mapping residual blocks and projection mapping residual blocks. An ECA module is embedded between the last convolutional layer and the activation function in all residual blocks, such as... Figure 9 As shown, (9a) represents the identity mapping residual block structure, and (9b) represents the projection mapping residual block structure.

[0039] S3. The leakage magnetic field end features, vibration end features and ultrasonic end features are fused to generate a fusion vector, and the early fault diagnosis results of the transformer are generated based on the fusion vector.

[0040] After extracting the complete local semantic features, three types of terminal feature vectors are obtained: vibration terminal features, ultrasonic terminal features, and magnetic flux leakage terminal features. To achieve deeper fusion between modes, MHCA is introduced at the end of each sub-network. Multiple rounds of repeated interactive fusion are performed on the features between each pair of modes, for a total of 6 MHCA operations. The detailed steps include: S31. Using the magnetic flux leakage end features, vibration end features, and ultrasonic end features alternately as keys and values, and correspondingly using the magnetic flux leakage end features, vibration end features, or ultrasonic end features as queries, perform multiple multi-head cross-attention fusions. The expression is as follows: , in, Indicates the first vibrational fusion characteristic; Indicates the characteristics of the vibration end; Indicates the characteristics of the ultrasound terminal; This indicates the second vibrational fusion characteristic; Indicates the characteristics of the leakage flux end; Indicates the first ultrasound fusion feature; Indicates the second ultrasound fusion feature; This indicates the first characteristic of magnetic flux leakage fusion; This indicates the second leakage magnetic flux fusion characteristic; This indicates multi-head attention fusion, in order to For example, it represents the characteristics of the vibration end. As a key, the ultrasound end features Multi-head attention fusion for values ​​and queries.

[0041] S32. The second vibration fusion feature, the second vibration fusion feature, and the second leakage magnetic flux fusion feature are spliced ​​together to obtain the fusion feature. , This indicates the fusion feature.

[0042] S33. Input the fused features into the classifier of the fault diagnosis model to generate fault diagnosis results: .

[0043] Example 2 This invention also provides a training method for densely connected networks and residual neural networks. However, when training for early fault diagnosis of transformers, there are problems such as a lack of actual fault samples and a lack of scalability in the training samples. Therefore, this embodiment proposes a model training method based on an extended dataset. First, for early faults such as arc discharge, winding structure deformation, and minor inter-turn short circuits generated in the transformer dynamic model experiment, the collected vibration and ultrasonic signals are expanded through the WCGAN-GP network. Then, the leakage magnetic field data from the dynamic model experiment is fused and expanded with the leakage magnetic field data from the transformer simulation model. Finally, all the expanded samples are processed to construct a multimodal fault sample set for deep neural network training, and the densely connected network and residual neural network are trained. The detailed steps include: A1. Collect fault vibration and fault ultrasonic signals during early transformer faults, and use the WCGAN-GP network to expand the samples to obtain expanded vibration signal data and expanded ultrasonic signal data.

[0044] In this embodiment, to further enhance the convergence and stability of Wasserstein GAN (WGAN) training, the Lipschitz condition implementation method of the discriminator is improved. A gradient penalty (GP) term is introduced into the loss function. By applying gradient norm constraints to the sample points interpolated between real and generated data, the weight clipping operation in the original WGAN is replaced, thereby achieving a more stable and continuous 1-Lipschitz function constraint. This effectively avoids gradient explosion and mode collapse problems during training. The improved objective function is: , in, Indicates a generator; Indicates the discriminator; This represents a real sample collected from the data distribution of fault vibration signals or fault ultrasonic signals. This represents the discriminator results generated based on real samples; This represents a sample collected from the data distribution of the initial noisy sequence; This represents the discriminator results generated based on fake samples; Indicates the generation of samples; Indicates the gradient penalty strength coefficient; This represents a sample collected from the data distribution of the interpolated samples; This represents the discriminator results generated based on the difference samples; This represents the gradient of the discriminator result generated based on the difference samples relative to the interpolated samples; This indicates the calculation of the L2 norm.

[0045] In this embodiment, a Conditional Generative Adversarial Network (CGAN) is introduced to generate data related to early transformer fault diagnosis. That is, conditional information is introduced into the aforementioned generative adversarial network. By introducing conditional constraints, CGAN significantly enhances the controllability and relevance of the generated results, improving the practical value of the model in real-world applications, resulting in the final WGAN-GP network. In this embodiment, the WGAN-GP network includes a generator and a discriminator, with the structure as follows: Figure 10 As shown, considering that both transformer vibration signals and ultrasonic signals are one-dimensional time-series signals, a one-dimensional convolutional network is used as the network structure for the generator. The discriminator adopts a simpler and easier-to-train multilayer perceptron (MLP). The discriminator receives the flattened input sequence and embeds the corresponding conditional labels. After concatenating the two, it is fed into a linear layer with Spectral Normalization (SN) applied to ensure that the 1-Lipschitz continuity condition required by WGAN is met. After each linear transformation, the LeakyReLU activation function is used to enhance the nonlinear expressive power of the model. The discriminator finally outputs a scalar to judge the authenticity of the input signal. The above structure makes the network more likely to converge stably in adversarial training. The parameters of the generator and discriminator are shown in Table 1.

[0046] Table 1 WCGAN-GP Network Parameters Iteratively execute the following steps to expand the sample: A11. Based on the fault vibration signal or fault ultrasonic signal, generate a 10,000-dimensional one-dimensional high-dimensional random noise sequence and embed a conditional label vector to obtain the initial noise sequence.

[0047] A12. Based on the initial noise sequence, a generator is used to obtain generated samples.

[0048] from Figure 10 As can be seen, the initial noise sequence is first mapped to a high-dimensional space through a linear layer; then, through a series of feature recovery modules consisting of upsampling, one-dimensional convolution, batch normalization (BatchNorm, BN) layers and LeakyReLU activation function, the structural information of the time series signal is gradually restored; finally, a 1D convolution layer combined with the Tanh activation function normalizes the output to [-1,1], and an adaptive average pooling layer ensures that the length of the output signal is consistent with the target, so that the generator can learn and restore the complex structure and nonlinear features in the one-dimensional signal.

[0049] A13. After embedding the fault vibration signal or fault ultrasonic signal into the condition variable, it is used as the input of the discriminator to obtain the first discriminant scalar.

[0050] A14. After embedding the generated samples into the condition variables, use them as input to the discriminator to obtain the second discriminant scalar.

[0051] A15. The generated sample is mixed with the fault vibration signal or fault ultrasonic signal according to a preset ratio to generate an interpolated sample. The interpolated sample is then embedded into the condition variable and used as the input of the discriminator to obtain the third discriminant scalar.

[0052] A16. Calculate the objective function value based on the first, second, and third discriminant scalars: , in, Indicates a generator; Indicates the discriminator; This refers to samples collected from the data distribution of fault vibration signals or fault ultrasonic signals. Indicates the first discriminant scalar; This represents a sample collected from the data distribution of the initial noisy sequence; Indicates the second discriminant scalar; Indicates the generation of samples; Indicates the gradient penalty strength coefficient; This represents a sample collected from the data distribution of the interpolated samples; Indicates the third discriminant scalar; This represents the gradient of the third discriminant scalar relative to the interpolated sample; This indicates the calculation of the L2 norm.

[0053] Gradient penalty term in the objective function The negative sign is because the discriminator aims for maximum, so this term serves as a deduction to satisfy the 1-Lipschitz constraint, thus limiting the discriminator from increasing the Wasserstein distance indefinitely.

[0054] A17. Update the generator parameters based on the objective function value and determine whether the termination condition is met. If it is met, stop the iteration; otherwise, regenerate the initial noise sequence.

[0055] Furthermore, this embodiment introduces a DTW-1NN-based method to evaluate the accuracy of the expanded samples. This method can objectively measure the similarity and separability of the generated samples and the real samples in terms of temporal structure without training an additional discriminant model. DTW is a similarity measurement method commonly used for time series alignment. Its goal is to find the optimal alignment path between two time series, thereby minimizing their nonlinear differences on the time axis. The detailed steps are as follows: A171. Construct a real sample set based on fault vibration signals or fault ultrasonic signals, construct a fake sample set based on generated samples, and merge the real sample set and the fake sample set to obtain a fused sample set.

[0056] A172. Construct an initial similarity matrix between each sample in the fused sample set and all other samples. Each element in the initial similarity matrix is ​​the Euclidean distance between data points in the sample. , Indicates sample The Middle Data points, Indicates sample The Middle Data points.

[0057] A173, Regarding the sample Calculate samples based on the initial similarity matrix The Middle Data points and samples The Middle The cumulative distances of the data points are used to construct a cumulative distance matrix, and the sample distances are calculated based on the cumulative distance matrix. and samples The DTW value.

[0058] The method for calculating the cumulative distance is as follows: , in, Indicate data pairs arrive The minimum cumulative Euclidean distance, in order to For example, it represents data pairs arrive The minimum cumulative Euclidean distance; Represents the data points in the initial similarity matrix To data point The Euclidean distance, where x represents the true time series. This indicates a spurious time series.

[0059] The DTW value is calculated as follows: ,in, Indicate data pairs arrive The minimum cumulative Euclidean distance, and This represents the total number of data points in the sample. This value indicates the minimum matching cost of two samples after optimal time warp alignment, reflecting the similarity of their temporal patterns.

[0060] A174. Sample selection based on DTW value The 1-nearest neighbors are used to generate nearest neighbor labels.

[0061] 1-NN accuracy is an evaluation method based on nearest neighbor classifiers, used to measure whether generated samples and real samples are separable, assuming the set of real samples is... Each real sample is labeled as 1; the generated sample set is... Each generated sample is labeled with a 0; the two are then merged into a unified fusion sample set. For the training set Samples in Its nearest neighbor sample is defined as: , in, express and The sample with the smallest DTW distance.

[0062] A175. Obtain the label for each data point, calculate the accuracy based on the label and the nearest neighbor labels, and evaluate the expanded sample based on the accuracy.

[0063] The method for calculating accuracy is as follows: , , , in, Indicates the size of the fused sample set; The label representing data point X; Represents the nearest neighbor labels of data point X; Indicates an indicator function; Indicates the nearest neighbor of data point X1; Represents the real sample set; This indicates a fake sample set.

[0064] When the 1-NN classification accuracy approaches 50%, it indicates that the generated samples are difficult to distinguish from the real samples, demonstrating a good fit. This method does not rely on discriminant network training. Compared to traditional Euclidean distance, DTW possesses non-linear time alignment capabilities, effectively measuring the true similarity of two time series in waveform structure. Therefore, DTW-1NN evaluation is used as an external validation method to verify the effectiveness of the generated samples at the classification level.

[0065] A2. Collect fault leakage flux data from the transformer dynamic model experiment and the transformer simulation model, and fuse and expand the two types of fault leakage flux data to obtain extended leakage flux signal data.

[0066] A3. Data processing is performed on the extended magnetic flux leakage signal data, extended vibration signal data, and extended ultrasonic signal data to obtain a multimodal fault sample dataset.

[0067] A4. Train densely connected networks and residual neural networks using a multimodal fault sample dataset.

[0068] During training, the cross-entropy loss function was selected as the loss function, and the AdamW adaptive moment estimator was used as the optimizer with a weight decay coefficient of 0.00001, an initial learning rate of 0.0001, and a batch size of 32.

[0069] Example 3 To verify the feasibility of the method provided by this invention, a transformer dynamic model experiment was conducted in this embodiment. The wiring of the entire dynamic model system is as follows: Figure 11 As shown, the experimental transformer is a three-phase, two-winding, dry-type transformer with a Yn / d11 connection. The transformer parameters are: rated capacity 50kVA, rated voltage 1 / 0.4kV, and winding parameters referred to the high-voltage side are as follows. , , , The parameters of the infinite power supply are: The transmission line parameters are as follows: , Adjustable load parameters are set as follows: .

[0070] A vibration sensor is mounted on the upper part of the transformer's clamps, with a frequency measurement range of 0~10kHz. An ultrasonic sensor is positioned 50 cm horizontally from the center of the winding, with a frequency measurement range of 15~70kHz. The vibration and ultrasonic sensors are connected to a constant current source, and their outputs are connected to two independent channels of an oscilloscope with a sampling frequency of 1GHz. A leakage magnetic field sensor, based on the Faraday magneto-optical effect, is connected via optical fiber to a self-developed magnetic quantity protection unit, which has an online leakage magnetic field monitoring function and a sampling frequency of 1600Hz. Finally, all measured waveform data is transmitted to a host computer.

[0071] Considering the limitations of the types of faults that can be simulated by the dynamic model transformer, this patent sets five operating states for dry-type transformers under rated voltage conditions: normal operation, inter-turn short circuit in the middle of the high-voltage winding, inter-turn short circuit in the lower part of the high-voltage winding, 10% axial deformation in the upper part of the high-voltage winding, and arc discharge in the high-voltage winding. Each state is labeled 0, 1, 2, 3, and 4 respectively. The inter-turn short circuit fault is set as a 2-turn short circuit, with a total of 201 turns in the transformer's high-voltage winding. The discharge fault is simulated using an arc generator, specifically by connecting two wires in parallel to different turns of the transformer winding to simulate an arc discharge fault inside the winding.

[0072] Based on the structural parameters of the transformer used in the dynamic model experiment, a 1:1 transformer simulation model was constructed using finite element simulation software. The specific structural parameters are shown in Table 2.

[0073] Table 2 Simulated Transformer Structural Parameters To verify the consistency between the simulation model and the results of the actual transformer, under the same normal operating conditions and secondary load, the radial leakage magnetic field amplitudes at three measuring points (upper, middle, and lower) of the transformer were plotted for both the actual and simulated values. The comparison results are shown below. Figure 12 As shown, through Figure 12 It can be seen that the radial leakage flux at each measurement point in the simulation is basically consistent with that in the actual measurement, which shows the correctness of the simulation model. The leakage magnetic field data obtained in the actual measurement can be expanded through the finite element simulation model.

[0074] Furthermore, the WCGAN-GP model was used to augment the transformer vibration and ultrasonic signals for subsequent training of the multimodal fault diagnosis model. The training run was set to 10,000 rounds. The changes in the loss function values ​​of the generator and discriminator during model training are as follows: Figure 13 As shown, both the generator and discriminator have high losses in the early stages of training. As the number of training iterations increases, the generator loss gradually decreases while the discriminator loss gradually increases. Eventually, both tend to stabilize. This process reflects the dynamic adversarial game relationship between the generator and discriminator when expanding transformer vibration and ultrasonic samples.

[0075] After training, to objectively evaluate the effectiveness of the proposed generation method, the DTW-1NN method provided in Example 2 was used to evaluate the quality of the generated samples and compared with several typical generation methods. The evaluation results of each method are shown in Table 3. The vibration and ultrasonic signals generated by WCGAN-GP outperformed other methods in terms of evaluation metrics, with its DTW-1NN index value closer to 50%, indicating that the difference between the generated samples and real samples is small, demonstrating high realism and indistinguishability. Furthermore, the generated samples from 6000 training rounds showed better evaluation results than those from 2000 rounds, and the difference was not significant compared to those from 10000 rounds. Considering the training time cost, this patent ultimately selected 6000 rounds as the training round number for WCGAN-GP.

[0076] Table 3. Evaluation results of DTW-1NN on vibration and ultrasonic signals using different generation methods. After data augmentation, RPM and CWT methods were used to perform feature transformation on vibration and ultrasonic signals, mapping one-dimensional time-domain signals to two-dimensional feature maps, and constructing corresponding category datasets. Each augmented fault category contained 200 samples, and after conversion to two-dimensional images, 1000 image samples were generated for each vibration and ultrasonic mode. Figure 14 Examples of two-dimensional feature maps after modality transformation are shown.

[0077] The generated data is used to train the residual neural network and the dense connection network, specifically: For three modes—magnetic leakage, vibration, and ultrasound—datasets containing 2000 samples each were constructed. First, the original samples were proportionally divided into training and test sets, with the test set containing a fixed 80 samples per class, and the remaining samples serving as the training set to ensure independence and class balance. Then, the training sets for vibration and ultrasound modes were augmented using WCGAN-GP and converted into two-dimensional image data; the training set for the magnetic leakage mode was fused 1:1 with measured and simulated data from the dynamic model, and preprocessed according to the procedure provided in Example 1 to ensure the sufficiency and diversity of the model training samples. Finally, the number of samples in the training and test sets for each mode under each operating state were 320 and 80, respectively. Figure 15 The model's accuracy gradually increases during training, as shown in Figure (15a), while its loss continuously decreases, as shown in Figure (15b). It exhibits good convergence and generalization capabilities, with an accuracy approaching 100% within 10 rounds.

[0078] After training, the model with the highest accuracy is saved and used for diagnostics on the test set. The t-SNE algorithm and confusion matrix are used to visualize and analyze the model's classification performance (see...). Figure 16 and Figure 17), Figure 16 The changes in feature distribution before and after model training are shown in Figure (16a). The original fused data shows significant class overlap, while Figure (16b) shows that the clustering effect of the output features of the fault diagnosis model after training is good, with clear classification boundaries; and the confusion matrix is... Figure 16 The model's classification accuracy was verified. A small number of faults were misjudged due to the similarity between vibration and ultrasonic signals in the upper winding deformation fault, but other faults were accurately identified, proving that the proposed method has good fault discrimination ability.

[0079] To further evaluate the effectiveness and generalization of the proposed method, its performance was compared with that of each modality using its own network and a trimodal attention-free fusion network. All methods were trained using 5-fold cross-validation, and accuracy, precision, recall, and F1 score were used as evaluation metrics. The calculation of the evaluation metrics is as follows: . The evaluation results of different methods are shown in Table 4. It can be seen that among the single-mode networks, 1D-ResNet based on the magnetic flux leakage mode performs best, with accuracy and other indicators higher than those of the vibration and ultrasonic modes, indicating that the magnetic flux leakage signal contains richer fault features. In contrast, the three-mode fusion network (without MLCA) shows significant improvement in all indicators. The method proposed in this patent performs best in all indicators, achieving an accuracy of 99.5%, demonstrating that by introducing the MLCA mechanism, the model can better integrate the features of different modes, thereby significantly improving the accuracy and stability of fault diagnosis, and possessing strong generalization ability and application value.

[0080] Table 4 Evaluation results under different methods In summary, this invention employs a multi-layer cross-attention fusion method that interacts between an improved DenseNet and an improved ResNet, enabling the full fusion of vibration, ultrasonic, and magnetic flux leakage data related to fault features. Compared to a single-layer convolutional neural network cross-attention fusion method, the accuracy of the neural network's diagnosis is significantly improved. Furthermore, since both vibration and ultrasonic signals belong to mechanical vibration-related modes, initial fusion can form a complete mechanical fault feature vector, improving feature complementarity. Subsequently, the final fusion with the magnetic flux leakage signal can introduce electromagnetic characteristic information, achieving cross-physical field feature enhancement. Experimental results show that, compared to the traditional single-layer convolutional network cross-attention fusion method, the method provided by this invention can fully utilize inter-modal complementary information, solving the problem of insufficient information utilization in complex fault mode recognition by traditional networks, and significantly improving fault diagnosis accuracy.

[0081] Furthermore, this invention utilizes Conditional Gradient Penalized Adversarial Network (WCGAN-GP) for sample augmentation and employs 1-Nearest Neighbor (1-NN) classification accuracy and Dynamic Time Warping (DTW) average distance as evaluation metrics. Its feasibility has been verified through limited simulation and dynamic model experiments. By combining actual early fault samples of transformers to generate sufficient data samples for training a deep neural network model for transformer fault diagnosis, the deep neural network model is fully trained, and the model achieves a correct diagnosis rate of over 95% for the trained faults.

[0082] Example 4 Furthermore, the present invention provides an electronic device including a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). Various programs and data required for device operation can also be stored in the RAM. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0083] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0084] The processing unit executes the various methods and processes described above, such as methods S1-S3 and methods A1-A4. For example, in some embodiments, methods S1-S3 and methods A1-A4 may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of methods S1-S3 and methods A1-A4 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute methods S1-S3 and methods A1-A4 by any other suitable means (e.g., by means of firmware).

[0085] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0086] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0087] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0088] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A transformer early fault diagnosis method based on three-mode monitoring data cross fusion, characterized in that, The method comprises: Collecting three-modal monitoring data of the transformer during operation, including leakage magnetic signals, vibration signals and ultrasonic signals, and performing feature extraction to obtain leakage magnetic features, vibration features and ultrasonic features; The vibration features and the ultrasonic features are respectively input into the dense connection networks with the same structure to respectively generate corresponding levels of intermediate features, the attention mechanism between the dense connection networks is used to fuse the corresponding levels of intermediate features to guide the generation of the next level of intermediate features, and the vibration end features and the ultrasonic end features are generated based on the output of the last level; The leakage magnetic end features, the vibration end features and the ultrasonic end features are fused to generate a fusion vector, and a transformer early fault diagnosis result is generated based on the fusion vector.

2. The transformer early fault diagnosis method based on three-mode monitoring data cross fusion according to claim 1, characterized in that, The residual neural network comprises a plurality of residual blocks, each of the residual blocks comprises a plurality of convolution layers and an activation function, and an ECA module is embedded between the last convolution layer and the activation function; the dense connection network comprises a plurality of dense blocks, each of the dense blocks comprises a plurality of dense layers, and an ECA module is embedded after each dense layer.

3. The transformer early fault diagnosis method based on three-mode monitoring data cross fusion according to claim 2, characterized in that, The intermediate fusion method is: For the intermediate features outputted by the two dense connection networks average pooling and projecting into a low-dimensional representation of the same dimension; the intermediate features include vibration intermediate features and ultrasound intermediate features, and the low-dimensional representation includes vibration low-dimensional representation and ultrasound low-dimensional representation; In each of the dense connection networks, the first layer outputs the vibration low-dimensional representation and the ultrasound low-dimensional representation alternately as keys and values, and the corresponding ultrasound low-dimensional representation or vibration low-dimensional representation is taken as a query to perform multi-head cross attention fusion, and the fusion result is connected with the intermediate features output by the first dense block to generate enhanced features as the input of the second dense block, which is: ， ， wherein, represents a vibration enhanced feature; represents a vibration intermediate feature of a th dense block output; represents a vibration low-dimensional representation of a th dense block; represents a ultrasound low-dimensional representation of a th dense block; represents an ultrasound enhanced feature; represents an ultrasound intermediate feature of a th dense block output; represents a multi-head attention fusion querying with the vibration low-dimensional representation as key and the ultrasound low-dimensional representation as value; represents a multi-head attention fusion querying with the ultrasound low-dimensional representation as key and the vibration low-dimensional representation as value.

4. The transformer early fault diagnosis method based on three-mode monitoring data cross fusion according to claim 2, characterized in that, The end fusion method is: The leakage magnetic end features, the vibration end features and the ultrasonic end features are alternately used as keys and values, and the leakage magnetic end features, the vibration end features or the ultrasonic end features are used as queries for multiple multi-head cross-attention fusion, and the expression is: ， wherein, represents a first vibration fusion feature; represents a vibration end feature; represents an ultrasound end feature; represents a second vibration fusion feature; represents a magnetic flux leakage end feature; represents a first ultrasound fusion feature; represents a second ultrasound fusion feature; represents a first magnetic flux leakage fusion feature; represents a second magnetic flux leakage fusion feature; represents multi-head attention fusion to as an example, represents a vibration end feature as a key, an ultrasound end feature as a value and multi-head attention fusion of a query; The second vibration fusion features, the second vibration fusion features and the second leakage magnetic fusion features are fused to obtain fusion features.

5. The transformer early fault diagnosis method based on three-modal monitoring data cross fusion according to claim 1, characterized in that, The training method of the dense connection network and the residual neural network comprises: Collecting fault vibration signals and fault ultrasonic signals during early transformer faults, and using a WCGAN-GP network to expand the samples to obtain expanded vibration signal data and expanded ultrasonic signal data; Collecting fault leakage magnetic data in a transformer dynamic model experiment and fault leakage magnetic data of a transformer simulation model, fusing and expanding the two kinds of fault leakage magnetic data to obtain expanded leakage magnetic signal data; Performing data processing on the expanded leakage magnetic signal data, the expanded vibration signal data and the expanded ultrasonic signal data to obtain a multi-modal fault sample data set; Using the multi-modal fault sample data set to train the dense connection network and the residual neural network.

6. The transformer early fault diagnosis method based on three-modal monitoring data cross fusion according to claim 5, characterized in that, The WCGAN-GP network comprises a generator and a discriminator, and the following steps are iteratively performed to realize sample expansion: Based on the fault vibration signals or the fault ultrasonic signals, a random noise sequence is generated, and a condition label vector is embedded to obtain an initial noise sequence; Based on the initial noise sequence, a generator is used to obtain a generated sample; The fault vibration signals or the fault ultrasonic signals are embedded into a condition variable to be input into the discriminator to obtain a first discrimination scalar; The generated sample is embedded into a condition variable to be input into the discriminator to obtain a second discrimination scalar; The interpolation sample is generated by mixing the generated sample and the fault vibration signal or the fault ultrasonic signal according to a preset proportion, the interpolation sample is embedded in a conditional variable to serve as an input of the discriminator, and a third discriminant scalar is obtained; A target function value is calculated based on the first discriminant scalar, the second discriminant scalar and the third discriminant scalar; Generator parameters are updated based on the target function value, and it is judged whether an end condition is met; if yes, iteration is stopped; otherwise, an initial noise sequence is regenerated.

7. The transformer early fault diagnosis method based on three-modal monitoring data cross fusion according to claim 6, characterized in that, The method for calculating the target function value is: ， wherein, denotes a generator; denotes a discriminator; denotes a sample drawn from a data distribution of a fault vibration signal or a fault ultrasonic signal; denotes a first discriminant scalar; denotes a sample drawn from a data distribution of an initial noise sequence; denotes a second discriminant scalar; denotes a generated sample; denotes a gradient penalty strength coefficient; denotes a sample drawn from a data distribution of an interpolated sample; denotes a third discriminant scalar; denotes a gradient of the third discriminant scalar with respect to the interpolated sample; denotes a two-norm computation.

8. The transformer early fault diagnosis method based on three-modal monitoring data cross fusion according to claim 6, characterized in that, The sample expansion further includes DTW-1NN precision evaluation: A real sample set is constructed based on the fault vibration signal or the fault ultrasonic signal, a false sample set is constructed based on the generated sample, and the real sample set and the false sample set are merged to obtain a fusion sample set; constructing an initial similarity matrix for each sample in the fusion sample set with respect to all other samples in the fusion sample set, each element in the initial similarity matrix being an Euclidean distance between data points in the sample, , denotes the i-th data point in the sample denotes the i-th data point in the sample denotes the i-th data point in the sample denotes the i-th data point in the sample denotes the i-th data point in the sample denotes the i-th data point in the sample For a sample , compute the cumulative distance between the th data point in sample and the th data point in sample , construct a cumulative distance matrix, compute the DTW value between sample and sample based on the cumulative distance matrix. selecting samples based on the DTW values a 1-neighbor based on the 1-neighbor, generating a near-neighbor label based on the 1-neighbor; A label of each data point is acquired, an accuracy rate is calculated based on the label and the near neighbor label, and the expanded sample is evaluated based on the accuracy rate.

9. The transformer early fault diagnosis method based on three-modal monitoring data cross fusion according to claim 8, characterized in that, The method for calculating the cumulative distance is: ， wherein, denotes the minimum cumulative Euclidean distance of data pair to , for example, denotes the minimum cumulative Euclidean distance of data pair to ; denotes the Euclidean distance of data point to data point in the initial similarity matrix, x denotes the real time series, denotes the false time series.​ 10. The transformer early fault diagnosis method based on three-modal monitoring data cross fusion according to claim 8, characterized in that, The method for calculating the accuracy rate is: ， ， ， wherein, denotes the size of the fusion sample set; denotes the label of data point X; denotes the near neighbor label of data point X; denotes the indicator function; denotes the data point X1-near neighbor; denotes the true sample set; denotes the false sample set.

Images