A transformer body vibration signal analysis and fault diagnosis method, device and medium

Abstract

The present application relates to the technical field of transformer fault diagnosis, and in particular to a transformer body vibration signal analysis and fault diagnosis method, device and medium. The present application combines the kurtosis characteristics of the vibration signal, adopts a semi-soft threshold function wavelet denoising method based on the threshold selection method of the 3sigm rule, and achieves better denoising effect. Through the fault diagnosis model of the organic fusion of the two improved HHT transforms-mobilenetV2 model, combined with different feature extraction methods, it is more conducive to retaining the effective features of the vibration signal; the improved mobilenetV2 model designs a multi-scale deep convolution model, introduces a channel attention mechanism before the channel-by-channel convolution, and after multi-scale deep convolution feature extraction, a multi-source data attention mechanism is introduced; without affecting the safe and reliable operation of the transformer, the intelligent diagnosis of the vibration state fault of the transformer is realized.

Description

Technical Field

[0001] This invention relates to the field of transformer fault diagnosis, specifically a method, device, and medium for analyzing transformer body vibration signals and diagnosing faults. Summary of the Invention

[0002] The technical problem to be solved by the transformer vibration signal analysis and fault diagnosis method based on the improved HHT transform-Mobilenet V2 model is to enhance the feature extraction capability of vibration signals and improve the fault diagnosis accuracy of transformer vibration signals by improving the feature extraction method and fault diagnosis model. This achieves intelligent fault diagnosis of transformer vibration state without affecting the safe and reliable operation of the transformer.

[0003] The overall approach of this invention is to improve the vibration signal feature extraction method by combining statistics with traditional algorithms. By incorporating the kurtosis characteristics of the vibration signal and employing a threshold selection method based on the 3sigm rule, a semi-soft threshold function wavelet denoising method is used, achieving better denoising results. A fault diagnosis model organically integrating two improved HHT transform-MobileNetV2 models, combined with different feature extraction methods, is more conducive to preserving effective vibration signal features. The improved MobileNetV2 model incorporates a multi-scale deep convolution model, introducing a channel attention mechanism before channel-by-channel convolution and a multi-source data attention mechanism after multi-scale deep convolution feature extraction. The data feature fusion operation of the improved WT-HHT transform-MobileNetV2 model and the VMD-HHT transform-MobileNetV2 model employs feature dimension expansion and splicing operations, ensuring the accuracy of the classification model while effectively improving the overall performance of the diagnostic model. This achieves intelligent diagnosis of transformer vibration state faults without affecting the safe and reliable operation of the transformer.

[0004] Specifically, the technical solution for implementing this invention is: a method for analyzing vibration signals and diagnosing faults in a transformer body, comprising the following steps: S01) respectively acquire the vibration signals of transformer winding and lead faults, insulating oil faults, core and metal faults and normal conditions; S02) Feature extraction is performed on the vibration signal collected in step S01). Two different improved HHT transform feature extraction methods are used, including the improved WT-HHT transform and the improved VMD-HHT transform. The improved WT-HHT transform first uses a semi-soft threshold signal denoising method based on the 3sigm rule, and then combines the kurtosis characteristics of the vibration signal for denoising. Then, the improved HHT transform is used to extract the feature values ​​of the transformer body vibration signal. The improved VMD-HHT transform first uses the variational mode decomposition algorithm VMD to preprocess the transformer body vibration signal data, and then uses the HHT transform to extract the feature values ​​of the transformer body vibration signal. S03) The vibration signal feature values ​​extracted in step S02) are made into a sample set, labeled, and the sample set is divided into a training set and a test set. S04) Establish a transformer body vibration fault diagnosis model based on the improved HHT converter-mobilenetV2; S05) Use the preprocessed training set to train the initial model of the transformer body vibration fault diagnosis model to obtain the transformer fault diagnosis model; use the validation set to verify the diagnostic performance of the transformer fault diagnosis model. S06) Real-time acquisition of vibration signals from transformer equipment, interception and processing of vibration signals to obtain feature value samples, inputting the samples into the transformer fault diagnosis model to obtain the fault diagnosis results of the transformer equipment.

[0005] Furthermore, the process of feature extraction using the improved WT-HHT transform is as follows: A1) Select a Haar wavelet basis that is similar to the waveform of the transformer body vibration signal and perform n-level wavelet decomposition on it; A2) Combining the kurtosis value of the vibration signal, and based on the threshold selection method of the 3sigm rule, a semi-soft threshold function is used to process the wavelet coefficients of each layer obtained from the decomposition; let ,in This represents the threshold set for the wavelet coefficients of the j-th layer. This represents the standard deviation of the wavelet coefficients at the j-th level. If the absolute value of the wavelet coefficients at this level is greater than... If a signal is considered valid and needs to be retained, a semi-soft thresholding function is used to process the wavelet coefficients; otherwise, it is considered noise and is set to 0. Vibration signals with kurtosis values ​​greater than 3 are removed, and those with kurtosis values ​​less than 3 are subjected to wavelet thresholding for noise reduction. A3) The retained signal is reconstructed using inverse wavelet transform to obtain the denoised signal; A4) Perform empirical mode decomposition on the denoised vibration signal to obtain a series of IMF components; A5) Use the KS hypothesis testing method to remove spurious components from a series of IMF components, and finally obtain the effective IMF components; A 6) Perform Hilbert transform on the final IMF components to obtain the time spectrum of the measured signal.

[0006] Furthermore, in step A2), the formula for calculating the semi-soft threshold function is: , in Let J be the wavelet coefficients of the j-th layer of the vibration signal decomposition. These are the wavelet coefficients of the j-th layer after correction using a semi-soft thresholding function. It is a threshold set for the wavelet coefficients of the j-th layer.

[0007] Furthermore, the feature extraction process of the improved VMD-HHT transform is as follows: B1) Obtain the raw data A(t) from the vibration sensor, preprocess the raw data A(t) to obtain several decomposed intrinsic mode functions IMFAs. n (t); The preprocessing involves using the Variational Mode Decomposition (VMD) algorithm to decompose the raw data A(t) from the vibration sensor, obtaining the corresponding mode functions for each mode component. ,in The number of decomposition layers in the Variational Mode Decomposition (VMD) algorithm. The decomposed components possess eigenmode functions; B2) Perform Hilbert transform on the corresponding mode functions of each decomposed mode component to obtain the time spectrum of the measured signal.

[0008] Furthermore, the transformer body vibration fault diagnosis model based on the improved HHT converter-mobilenetV2 includes the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model. Both the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model adopt the improved mobileenetV2 model. The MobileNetV2 model includes a first convolutional layer, which is connected to a RuLU6 activation function to increase the dimensionality of the input, thus obtaining a dimension-upgrading layer. The channel attention layer, connected to the first convolutional layer, uses a channel attention mechanism to adaptively weight the features of different channels. First, the input is subjected to max pooling and average pooling respectively, resulting in two one-dimensional vectors representing the max pooling features and average pooling features. Then, the two vectors are input into a shared multilayer perceptron for learning. The two vectors output by the multilayer perceptron are added together and activated using the sigmoid activation function to obtain two channel attention feature maps. The channel attention feature maps are multiplied with the original input feature map to obtain the feature map after channel attention. A multi-scale deep convolutional module is connected to the channel attention layer, and multiple convolutional modules are connected in parallel. LeakyRelu is used instead of Relu6 in MobileNetV2 as the activation function to obtain the feature extraction layer. The multi-source data attention layer is connected to the multi-scale deep convolution module. It uses multiple KANs to compress the output features of the multi-scale deep convolution module. The number of KANs is the same as the number of convolution modules connected in parallel with the multi-scale deep convolution module. Each KAN compresses the output features of the multi-scale deep convolution module into a 1×1 output. Then, the compressed features are concatenated, and the normalized exponential function is used to calculate the concatenated features to convert the output value of each data source feature into a probability value between [0, 1]. The sum of the three weights is made equal to 1. The weights are multiplied by the corresponding features to obtain the weighted features of the convolution signals of different sizes. Then, the three weighted features are fused. The second convolutional layer is connected to the multi-source data attention layer. The second convolutional layer is connected to the Linear activation function, which performs pointwise convolution to fuse cross-channel information.

[0009] Furthermore, based on the improved HHT-mobilenetV2 transformer vibration fault diagnosis model, feature fusion is performed on the data features output by the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model. Let F1 and F2 be the one-dimensional expanded sequences of the feature sets output by the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model, respectively, and g be the fused features used for classification. The feature fusion process is expressed as follows: , , in For feature expansion and splicing operations, For mapping operations in fully connected layers, , These are the j-th local feature value in the output feature set of the WT-HHT-mobilenetV2 model and the j-th local feature value in the output feature set of the VMD-HHT-mobilenetV2 model, respectively.

[0010] Furthermore, the transformer body vibration fault diagnosis model adopts the cross-entropy loss function as the optimization objective function. The expression of the cross-entropy loss function is as follows: , in For the target value, This is the predicted value output by softmax.

[0011] Furthermore, the multi-scale depth convolution module connects three convolution modules in parallel. The dimensions of the three convolution modules are 1×1, 3×3, and two 3×3 convolutions stacked together. The multi-scale depth convolution module combines different convolution kernels together to expand the model width and increase the receptive field.

[0012] The present invention also discloses a transformer body vibration signal analysis and fault diagnosis device, including a processor and a memory storing program instructions. The processor is configured to execute the transformer body vibration signal analysis and fault diagnosis method as described above when running the program instructions.

[0013] The present invention also discloses a storage medium storing program instructions, which, when executed, perform the transformer body vibration signal analysis and fault diagnosis method as described above.

[0014] Compared with the prior art, the present invention has the following advantages: (1) Feature extraction: Combining different feature extraction methods, each with its own advantages and disadvantages, is more conducive to preserving the effective features of vibration signals.

[0015] (2) Signal denoising: Combining statistics with traditional algorithms, the method for extracting vibration signal features was improved. Based on the kurtosis characteristics of the vibration signal and the threshold selection method of the 3sigm rule, a semi-soft threshold function wavelet denoising method was adopted to achieve better denoising effect.

[0016] (3) Fault diagnosis: The improved MobilenetV2 model is designed with a multi-scale deep convolution model. Before the channel-by-channel convolution, a channel attention mechanism is introduced. After the multi-scale deep convolution feature extraction, a multi-source data attention mechanism is introduced. The data feature fusion operation of the output of the improved WT-HHT transform-MobilenetV2 model and the VMD-HHT transform-MobilenetV2 model adopts the feature expansion and splicing operation, which ensures the accuracy of the classification model and effectively improves the overall performance of the diagnostic model. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method; Figure 2 This is a schematic diagram of the improved MobilenetV2 model. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1 This embodiment discloses a method for analyzing and diagnosing transformer vibration signals. This method is based on the combination of an improved WT-HHT transform and an improved MobilenetV2 model, and the combination of an improved VMD-HHT transform and an improved MobilenetV2 model. By organically integrating the two improved HHT transforms and MobilenetV2 models, a transformer vibration fault diagnosis model based on deep learning theory is further constructed, which improves the stability and generalization ability of the fault diagnosis model, thereby improving the accuracy of online monitoring and identification of transformer vibration status.

[0020] like Figure 1 As shown, this method includes the following steps: Step S1: Obtain vibration signals for winding and lead faults, insulating oil faults, core and metal faults, and normal conditions of the transformer equipment, respectively. Step S2: Feature extraction is performed on the above vibration signal using two different improved HHT transform feature extraction methods, including the improved WT-HHT transform and the improved VMD-HHT transform. Step S3: Perform preprocessing operations on the above vibration signals, create a sample set from the vibration signals, add labels, and divide the sample set into training samples and test samples. Step S4: Establish a transformer body vibration fault diagnosis model based on the improved HHT converter-mobilenetV2; Step S5: Train the initial model using the preprocessed training set to obtain the transformer fault diagnosis model; use the validation set to verify the diagnostic performance of the transformer fault diagnosis model. Step S6: Real-time acquisition of vibration signals from the transformer equipment, extraction of samples from the vibration signals, input of the samples into the transformer fault diagnosis model to obtain fault diagnosis results for the transformer equipment, thereby realizing real-time monitoring of transformer vibration status, intelligent fault diagnosis, and intelligent fault early warning.

[0021] In step S2, two different improved HHT transform feature extraction methods are used: The first method is an improved WT-HHT transform. It first employs a semi-soft thresholding method based on the 3sigm rule for signal denoising, while also incorporating the kurtosis characteristics of the vibration signal for further denoising. Then, an improved HHT transform is used to extract the transformer's vibration signal. This includes the following steps: A1) Select a Haar wavelet basis that is similar to the waveform of the transformer body vibration signal and perform n-level wavelet decomposition on it; in this embodiment, when the waveform of the transformer body vibration signal is more than 70% similar to the Haar wavelet basis, it is considered to be relatively similar.

[0022] A2) Combining the kurtosis value of the vibration signal, and based on the threshold selection method of the 3sigm rule, a semi-soft threshold function is used to process the wavelet coefficients of each layer obtained from the decomposition; let ,in This represents the threshold set for the wavelet coefficients of the j-th layer. This represents the standard deviation of the wavelet coefficients at the j-th level. If the absolute value of the wavelet coefficients at this level is greater than... If a signal is considered a valid signal that needs to be retained, a semi-soft thresholding function is used to process the wavelet coefficients; otherwise, it is considered a noise signal and is set to 0. Vibration signals with a kurtosis value greater than 3 are removed, and those with a kurtosis value less than 3 are subjected to wavelet thresholding for noise reduction.

[0023] In this step, kurtosis is a dimensionless parameter describing the peak sharpness of a vibration signal. The kurtosis index is highly sensitive to the impact characteristics of the signal. When a fault-free transformer is operating normally, the kurtosis value of the vibration signal is generally less than 3. If the kurtosis value is greater than 3, it indicates a fault, and the more severe the fault, the greater the kurtosis value. Therefore, kurtosis can be used as a method for optimizing IMF components. Specifically, vibration signals with kurtosis values ​​greater than 3 are discarded, while those with kurtosis values ​​less than 3 undergo wavelet thresholding for noise reduction.

[0024] In this step, the formula for calculating the semi-soft threshold function is: , in Let J be the wavelet coefficients of the j-th layer of the vibration signal decomposition. These are the wavelet coefficients of the j-th layer after correction using a semi-soft thresholding function. It is a threshold set for the wavelet coefficients of the j-th layer.

[0025] Compared to the soft thresholding function, the semi-soft thresholding function preserves the continuity of the soft thresholding function. At that time, its input and output exhibit a non-linear relationship, and as... As the threshold increases, the deviation between input and output gradually decreases, eliminating the constant bias problem inherent in soft thresholding functions and preserving effective features to the greatest extent. Simultaneously, the 3Sigma threshold selection method combines statistical concepts with traditional algorithms to achieve better denoising results.

[0026] A3) The retained signal is reconstructed using inverse wavelet transform to obtain the denoised signal.

[0027] A4) Perform empirical mode decomposition on the denoised vibration signal to obtain a series of IMF components.

[0028] A5) Use the KS hypothesis testing method to remove spurious components from a series of IMF components, and finally obtain the valid IMF components.

[0029] A6) Perform Hilbert transform on the final IMF components to obtain the time spectrum of the measured signal.

[0030] In this embodiment, the improved VMD-HHT transform first uses the Variational Mode Decomposition (VMD) algorithm to preprocess the transformer body vibration signal data, and then uses the HHT transform to extract the transformer body vibration signal. The feature extraction process is as follows: B1) Obtain the raw data A(t) from the vibration sensor, preprocess the raw data A(t) to obtain several decomposed intrinsic mode functions IMFAs. n (t); The preprocessing involves using the Variational Mode Decomposition (VMD) algorithm to decompose the raw data A(t) from the vibration sensor, obtaining the corresponding mode functions for each mode component. ,in The number of decomposition layers in the Variational Mode Decomposition (VMD) algorithm. The decomposed components possess several intrinsic mode functions.

[0031] B2) Perform Hilbert transform on the corresponding mode functions of each decomposed mode component to obtain the time spectrum of the measured signal.

[0032] In this embodiment, the transformer body vibration fault diagnosis model based on the improved HHT converter-mobilenetV2 includes the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model. Both the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model adopt the improved mobileenetV2 model. Figure 2 As shown, the improved MobileNetV2 model structure includes: The first convolutional layer has a dimension of 1*1. The first convolutional layer is connected to the RuLU6 activation function to increase the dimension of the input, thus obtaining the dimension-upgrading layer.

[0033] The channel attention layer, connected to the first convolutional layer, employs a channel attention mechanism to adaptively weight the features of different channels before channel-wise convolution. First, the input is subjected to max pooling and average pooling respectively, resulting in two one-dimensional vectors representing max pooling and average pooling features. Then, these two vectors are input into a shared multilayer perceptron for learning. The two vectors output by the multilayer perceptron are added together, and the sigmoid activation function is used for activation, resulting in two channel attention feature maps. The channel attention feature maps are multiplied by the original input feature map to obtain the feature map after channel attention.

[0034] A multi-scale deep convolutional module is connected to the channel attention layer and consists of multiple convolutional modules in parallel. LeakyReLU is used instead of ReLU6 in MobileNetV2 as the activation function to obtain the feature extraction layer. In this embodiment, the multi-scale deep convolutional module connects three convolutional modules in parallel, with dimensions of 1×1, 3×3, and two stacked 3×3 convolutions, respectively. The multi-scale deep convolutional module combines different convolutional kernels to expand the model width and increase the receptive field, thereby improving the model's feature extraction capability and enhancing the robustness of the neural network.

[0035] The multi-source data attention layer, connected to the multi-scale deep convolutional module, aims to avoid the low recognition accuracy of neural networks caused by the equal-weighted fusion of multi-scale deep convolutional data features. First, multiple KANs are used to compress the output features of the multi-scale deep convolutional module. The number of KANs is the same as the number of convolutional modules connected in parallel with the multi-scale deep convolutional module. In this embodiment, there are 3 KANs, each compressing 3 different scale deep convolutional input features into a 1×1 output. Then, the compressed features are concatenated, and the normalized exponential function (i.e., the Softmax function) is used to calculate the concatenated features, converting the output value of each data source feature into a probability value between [0, 1], and ensuring the sum of the three weights equals 1. The weights are multiplied by the corresponding features to obtain the weighted features of the convolutional signals at different sizes. Finally, feature fusion is performed on the three weighted features.

[0036] The second convolutional layer, with a dimension of 1*1, is connected to the multi-source data attention layer. The second convolutional layer is connected to the Linear activation function, which performs pointwise convolution to fuse cross-channel information.

[0037] In this embodiment, based on the improved HHT-mobilenetV2 transformer vibration fault diagnosis model, the data features output by the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model are fused. Let F1 and F2 be the one-dimensional expanded sequences of the feature sets output by the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model, respectively, and g be the fused feature used for classification. The feature fusion process is expressed as follows: , , in For feature expansion and splicing operations, For mapping operations in fully connected layers, , These are the j-th local feature value in the output feature set of the WT-HHT-mobilenetV2 model and the j-th local feature value in the output feature set of the VMD-HHT-mobilenetV2 model, respectively.

[0038] After constructing the model, the cross-entropy loss function is used as the objective function to optimize the model, and the classification result is output. The objective function of the model is shown below: , in For the target value, This is the predicted value output by softmax.

[0039] In this example, the improved MobileNetV2 model introduces channel attention and multi-source data attention mechanisms. The differences are as follows: ① The channel attention mechanism performs max pooling and average pooling on the input, resulting in two one-dimensional vectors. These vectors are then fed into a shared multilayer perceptron (MLP) for learning, and the two output vectors of the MLP are summed. The multi-source data attention mechanism compresses the input features from three different scale convolutions (including 1×1, 3×3, and a stack of two 3×3 convolutions) using three KANs, compressing them into a 1×1 output. The compressed features are then concatenated. ② The channel attention mechanism uses the Sigmoid activation function to activate the input, resulting in two channel attention feature maps, which are multiplied by the original input feature map. The multi-source data attention mechanism uses a normalization function (i.e., the Softmax function) to calculate the concatenated features, transforming the output value of each data source feature into a probability value between [0, 1], which sums to 1. Multiplying the weights by the corresponding features yields the weighted features of convolutional signals at different scales and depths.

[0040] Example 2 This embodiment discloses a transformer body vibration signal analysis and fault diagnosis device, including a processor and a memory storing program instructions. The processor is configured to execute the transformer body vibration signal analysis and fault diagnosis method as described in Embodiment 1 when running the program instructions.

[0041] Example 3 This embodiment discloses a storage medium storing program instructions, which, when executed, perform the transformer body vibration signal analysis and fault diagnosis method as described in Embodiment 1.

[0042] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0043] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0044] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0045] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

Claims

1. A method for analyzing vibration signals and diagnosing faults in a transformer body, characterized in that: Includes the following steps: S01) respectively acquire vibration signals of transformer winding and lead faults, insulating oil faults, core and metal faults and normal conditions; S02) Feature extraction is performed on the vibration signal collected in step S01). Two different improved HHT transform feature extraction methods are used, including the improved WT-HHT transform and the improved VMD-HHT transform. The improved WT-HHT transform first uses a semi-soft threshold signal denoising method based on the 3sigm rule, and then combines the kurtosis characteristics of the vibration signal for denoising. Then, the improved HHT transform is used to extract the feature values ​​of the transformer body vibration signal. The improved VMD-HHT transform first uses the variational mode decomposition algorithm VMD to preprocess the transformer body vibration signal data, and then uses the HHT transform to extract the feature values ​​of the transformer body vibration signal. S03) The vibration signal feature values ​​extracted in step S02) are made into a sample set, labeled, and the sample set is divided into a training set and a test set. S04) Establish a transformer body vibration fault diagnosis model based on the improved HHT converter-mobilenetV2; The transformer body vibration fault diagnosis model based on the improved HHT converter-mobilenetV2 includes the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model. Both the WT-HHT-mobilenetV2 model and the VMD-HHT-mobilenetV2 model adopt the improved mobileenetV2 model. The MobileNetV2 model includes a first convolutional layer, which is connected to a RuLU6 activation function to increase the dimensionality of the input, thus obtaining a dimension-upgrading layer. The channel attention layer, connected to the first convolutional layer, uses a channel attention mechanism to adaptively weight the features of different channels. First, the input is subjected to max pooling and average pooling respectively, resulting in two one-dimensional vectors representing the max pooling features and average pooling features. Then, the two vectors are input into a shared multilayer perceptron for learning. The two vectors output by the multilayer perceptron are added together and activated using the sigmoid activation function to obtain two channel attention feature maps. The channel attention feature maps are multiplied with the original input feature map to obtain the feature map after channel attention. A multi-scale deep convolutional module is connected to the channel attention layer, and multiple convolutional modules are connected in parallel. LeakyRelu is used instead of Relu6 in MobileNetV2 as the activation function to obtain the feature extraction layer. The multi-source data attention layer is connected to the multi-scale deep convolution module. It uses multiple KANs to compress the output features of the multi-scale deep convolution module. The number of KANs is the same as the number of convolution modules connected in parallel with the multi-scale deep convolution module. Each KAN compresses the output features of the multi-scale deep convolution module into a 1×1 output. Then, the compressed features are concatenated, and the normalized exponential function is used to calculate the concatenated features to convert the output value of each data source feature into a probability value between [0, 1]. The sum of the three weights is made equal to 1. The weights are multiplied by the corresponding features to obtain the weighted features of the convolution signals of different sizes. Then, the three weighted features are fused. The second convolutional layer is connected to the multi-source data attention layer. The second convolutional layer is connected to the linear activation function, and pointwise convolution is used to fuse cross-channel information. S05) Use the preprocessed training set to train the initial model of the transformer body vibration fault diagnosis model to obtain the transformer fault diagnosis model; use the validation set to verify the diagnostic performance of the transformer fault diagnosis model. S06) Real-time acquisition of vibration signals from transformer equipment, interception and processing of vibration signals to obtain feature value samples, inputting the samples into the transformer fault diagnosis model to obtain the fault diagnosis results of the transformer equipment.

2. The method for analyzing transformer body vibration signals and diagnosing faults according to claim 1, characterized in that: The process of feature extraction using the improved WT-HHT transform is as follows: A1) Select a Haar wavelet basis that is similar to the waveform of the transformer body vibration signal and perform n-level wavelet decomposition on it; A2) Combining the kurtosis value of the vibration signal, based on the threshold selection method of the 3sigm rule, a semi-soft threshold function is used to process the wavelet coefficients of each layer obtained by decomposition. make ,in This represents the threshold set for the wavelet coefficients of the j-th layer. This represents the standard deviation of the wavelet coefficients at the j-th level. If the absolute value of the wavelet coefficients at this level is greater than... If a signal is considered valid and needs to be retained, a semi-soft thresholding function is used to process the wavelet coefficients; otherwise, it is considered noise and is set to 0. Vibration signals with kurtosis values ​​greater than 3 are removed, and those with kurtosis values ​​less than 3 are subjected to wavelet thresholding for noise reduction. A3) The retained signal is reconstructed using inverse wavelet transform to obtain the denoised signal; A4) Perform empirical mode decomposition on the denoised vibration signal to obtain a series of IMF components; A5) Use the KS hypothesis testing method to remove spurious components from a series of IMF components, and finally obtain the effective IMF components; A 6) Perform Hilbert transform on the final IMF components to obtain the time spectrum of the measured signal.

3. The method for analyzing transformer body vibration signals and diagnosing faults according to claim 2, characterized in that: In step A2), the formula for calculating the semi-soft threshold function is: ， in Let J be the wavelet coefficients of the j-th layer of the vibration signal decomposition. These are the wavelet coefficients of the j-th layer after correction using a semi-soft thresholding function. It is a threshold set for the wavelet coefficients of the j-th layer.

4. The method for analyzing and diagnosing transformer body vibration signals according to claim 1, characterized in that: The process of feature extraction using the improved VMD-HHT transform is as follows: B1) Obtain the raw data A(t) from the vibration sensor, preprocess the raw data A(t), and obtain several decomposed data with intrinsic mode functions IMFAs. n (t); The preprocessing involves using the Variational Mode Decomposition (VMD) algorithm to decompose the raw data A(t) from the vibration sensor, obtaining the corresponding mode functions for each mode component. ,in The number of decomposition layers in the Variational Mode Decomposition (VMD) algorithm. Several intrinsic mode functions after decomposition B2) Perform Hilbert transform on the corresponding mode functions of each decomposed mode component to obtain the time spectrum of the measured signal.

5. The method for analyzing transformer body vibration signals and diagnosing faults according to claim 1, characterized in that: A transformer vibration fault diagnosis model based on the improved HHT-mobilenetV2 model is used. The data features output by the WT-HHT-mobilenetV2 and VMD-HHT-mobilenetV2 models are fused. Let F1 and F2 be the one-dimensional expanded sequences of the feature sets output by the WT-HHT-mobilenetV2 and VMD-HHT-mobilenetV2 models, respectively, and g be the fused features used for classification. The expression for the feature fusion process is as follows: ， ， in For feature expansion and splicing operations, For mapping operations in fully connected layers, , These are the j-th local feature value in the output feature set of the WT-HHT-mobilenetV2 model and the j-th local feature value in the output feature set of the VMD-HHT-mobilenetV2 model, respectively.

6. The method for analyzing transformer body vibration signals and diagnosing faults according to claim 1, characterized in that: The transformer body vibration fault diagnosis model uses the cross-entropy loss function as the optimization objective function. The expression for the cross-entropy loss function is as follows: ， in For the target value, This is the predicted value output by softmax.

7. The method for analyzing transformer body vibration signals and diagnosing faults according to claim 1, characterized in that: The multi-scale depth convolution module connects three convolution modules in parallel. The dimensions of the three convolution modules are 1×1, 3×3, and two 3×3 convolutions stacked together. The multi-scale depth convolution module combines different convolution kernels to expand the model width and increase the receptive field.

8. A transformer body vibration signal analysis and fault diagnosis device, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute the transformer body vibration signal analysis and fault diagnosis method as described in any one of claims 1 to 7 when running the program instructions.

9. A storage medium storing program instructions, characterized in that, When the program instructions are executed, they perform the transformer body vibration signal analysis and fault diagnosis method as described in any one of claims 1 to 7.

Images