Improved MobileNetV2 voltage transformer fault diagnosis method based on multi-channel feature image

Abstract

According to the multi-channel feature image-based improved MobileNetV2 voltage transformer fault diagnosis method, a partial overlapping sliding window technology is used to carry out data enhancement on a secondary side voltage signal so as to construct images capable of representing different features. Through fusion of multi-channel feature images representing structure information, transient change and frequency domain characteristics, combined extraction of voltage signal multi-dimensional features is realized. On the basis, an improved MobileNetV2 diagnosis model is provided, and the nonlinear modeling capability of the network to a complex feature mode is enhanced by introducing a Mish activation function; and a normalization-based attention module (NAM) is introduced to realize dynamic focusing of fault features. Meanwhile, a receptive field of a convolutional layer is expanded by using multi-scale expansion convolution, so that correlation modeling between features is enhanced. According to the framework, the diagnosis precision of the voltage transformer is remarkably improved while the light weight of the model is kept.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis, and specifically discloses a voltage transformer fault diagnosis method based on a multi-channel feature image-improved MobileNetV2. Background Art

[0002] With the continued and in-depth development of my country's power system, its architecture and efficiency are undergoing profound changes. The grid's coverage continues to expand, and the capacity of power facilities is steadily increasing. In this process, voltage transformers, as core electrical components of substations, play a crucial role that cannot be ignored. In the operating environment of high-voltage transmission lines and substations, voltage transformers are responsible for converting high voltage to low voltage in a precise ratio, providing accurate measurement and monitoring data for relay protection systems and other systems. Therefore, voltage transformer fault diagnosis is crucial to the safe and stable operation of the power system and the accuracy of energy metering.

[0003] Currently, with the deep penetration of artificial intelligence (AI) technology in the industrial sector, fault diagnosis technology is evolving towards intelligentization. This phase of the fault diagnosis process integrates mathematical modeling, signal processing, and AI algorithms. By combining signal data collected by sensors with techniques such as spectrum analysis, feature extraction, and pattern recognition, this enables faster fault detection and diagnosis. Current intelligent diagnostic technology not only identifies the fault type but also analyzes the cause and, in some cases, locates the fault. With the rapid advancement of machine learning and deep learning, AI-driven diagnostic methods have become mainstream, effectively addressing complex fault diagnosis challenges.

[0004] In the field of industrial equipment fault diagnosis, traditional data-driven approaches often rely on large-scale training data with a balanced sample distribution. However, this premise is difficult to meet in real-world engineering scenarios. Due to the complex and variable operating conditions of equipment, the number of samples of different fault types is often significantly unevenly distributed, causing models trained on big data to encounter bottlenecks in generalization performance in engineering applications. These issues also lead to limitations in traditional feature extraction methods and diagnostic models, necessitating the search for new diagnostic methods that can be applied to real-world scenarios with small sample sizes and imbalanced classifications between samples.

[0005] Compared with the prior art, the differences are as follows:

[0006] Technical comparison with patent CN115438576A "Electronic voltage transformer error prediction method based on Prophet, self-attention mechanism and time series convolutional network"

[0007] Patent CN115438576A proposes a method for predicting electronic voltage transformer errors based on Prophet, self-attention mechanism, and time series convolutional network. It mainly performs time series error prediction on electronic voltage transformers, aiming to predict their future error trends and provide a basis for power system equipment calibration. Its application scenarios focus on the operation monitoring and prediction of electronic voltage transformers in power systems. By analyzing historical ratio difference data to predict future error trends, it assists in formulating equipment calibration plans and improves power system stability. This patent proposes a voltage transformer fault diagnosis method based on a multi-channel feature image-improved MobileNetV2. By analyzing the abnormal characteristics of the secondary side voltage signal, it identifies the fault type in real time, solving the problems of a small number of fault data samples and unbalanced classification. Its application scenarios focus on fault detection and diagnosis of voltage transformers, ensuring the safe and stable operation of the power system and the accuracy of electricity metering. The application scenarios of the two are essentially different.

[0008] Patent CN115438576A collects and normalizes the ratio error data of an electronic voltage transformer. The normalized data is then fed into a Prophet model for decomposition, and trend components are fitted to output predictions. The training data is then fed into a time series convolutional network model for feature extraction and prediction. The output of the time series convolutional network model is then summed with the periodic terms decomposed from the Prophet model and fed into a self-attention mechanism layer for feature extraction. Finally, a fully connected layer performs dimensionality reduction to output predicted data for the ratio error of the electronic voltage transformer. This patent is based on a voltage transformer fault diagnosis model based on a multi-channel feature image-improved MobileNetV2. The model performs data augmentation by performing sliding window slicing on the secondary voltage time series signal of the voltage transformer and adding Gaussian white noise. The enhanced subsequences of the time series data are then used to generate Gram angle fields, recurrence graphs, and continuous wavelet transform graphs. Bilinear interpolation is used to align the image scales, and multi-channel feature images are fused through three channels. Training is performed using an improved MobileNetV2 diagnostic model to enhance the ability to focus on and model complex feature patterns. There are essential differences in the technical solutions of the two.

[0009] Patent CN115438576A directly processes raw time series data, collecting ratio-difference data from electronic voltage transformers. This data is then z-score normalized to remove outliers. The Prophet model is then used to train the processed training data and fit the trend components. The training data is then fed into a time series convolutional network (TCN), which uses residual blocks and dilated causal convolution for high-dimensional feature extraction. This combined model achieves error prediction. This patent converts the time series signal into an image, slices the secondary voltage of the voltage transformer using partially overlapping sliding windows, and adds Gaussian white noise for data enhancement and overfitting prevention. Polar coordinate mapping is used to generate a Gram matrix from the subsequences to characterize the signal structure. Time-delay embedding is used to generate a recurrence graph to characterize the signal's transient characteristics. Morlet wavelets are used to generate a time-frequency graph to extract multi-scale information in the frequency domain. These three feature maps are then fed into the R, G, and B channels to represent multidimensional features using images. The two approaches differ fundamentally in their technical approach.

[0010] Patent CN115438576A introduces a time series convolutional network through time series model fusion and dynamic weight optimization, strengthening the feature extraction capability of time series data. It also introduces a self-attention mechanism to better explore and utilize the potential relationships in the data sequence, thereby improving the prediction effect and being able to better predict the future error change trend of electronic voltage transformers. This patent improves the lightweight network, uses the Mish activation function in the shallow layer to alleviate the gradient truncation problem, adopts multi-scale dilated convolution to capture local details and long-range periodic features, and combines dilated convolution with a normalized attention mechanism to improve the inverted residual block to balance computational efficiency and diagnostic accuracy. While maintaining the lightweight model, it improves the diagnostic accuracy of the voltage transformer, improves the accuracy of fault diagnosis, and improves the model's adaptability to special data scenarios, especially in the case of small samples and imbalanced fault classification. There is an essential difference in the technical effects of the two. Summary of the Invention

[0011] To solve the above technical problems, the present invention proposes a voltage transformer fault diagnosis method based on a multi-channel feature image-improved MobileNetV2, which can improve the accuracy of fault diagnosis and solve the problem of poor diagnostic network feature recognition in the case of small samples and unbalanced fault classification.

[0012] To achieve the above object, the technical solution adopted by the present invention is:

[0013] The voltage transformer fault diagnosis method based on the improved MobileNetV2 with multi-channel feature images includes the following steps:

[0014] 1) Perform partially overlapping sliding window slicing on the voltage transformer secondary side voltage time series signal:

[0015] 2) Generate feature images for the subsequences after time series data enhancement;

[0016] 3) Generation of multi-channel feature maps;

[0017] 4) The generated multi-channel feature map is input into the improved MobileNetV2 network for training.

[0018] As a further improvement of the present invention, the step 1) of performing sliding window slicing on the time series signal is as follows:

[0019] Step 1: Use the collected voltage signal time series, recorded as X = {x1, x2, ..., x N}; Use partially overlapping window slicing with a window size of w, a step size of s, and an overlap ratio of p to generate multiple subsequences:

[0020] X i ={x i ,x i+1 ,...,x i+w-1},i=1,2,...N-w+1

[0021] Step 2: To prevent the deep learning model from overfitting, add Gaussian white noise with a signal-to-noise ratio of N to the subsequence to obtain the subsequence after time series data enhancement. i is a Gaussian white noise sample with a standard deviation of ε, then the subsequence after time series data enhancement is:

[0022] X i ={x i +ξ i ,x i+1 +ξ i+1 ,...,x i+w-1 +ξ i+w-1},i=1,2,...N-w+1.

[0023] As a further improvement of the present invention, the step 2) of generating a feature image for the subsequence after the time series data enhancement is as follows:

[0024] Step 1: Construct the Gram Angle Field (GAF) of the subsequence i , scale the subsequence to the interval [0,1] by extreme value normalization, convert the normalized data into polar coordinate representation, and further generate the Gram matrix through polar coordinate angle operation to obtain the Gram angle field of the subsequence. Let I be the unit row vector and X be the subsequence. The Gram matrix can be expressed as:

[0025]

[0026] Step 2: Construct the recursive graph RP of the subsequencei The time series is converted into phase space trajectory by time delay embedding method. The embedding dimension is set to m, which determines the complexity of phase space reconstruction; the delay time is τ, which affects the expansion degree of phase space trajectory; the i-th phase space vector is obtained as Let Θ be the step function, ||·|| be the norm, and ε be the similarity threshold, which controls the similarity determination of the state. Then the elements of the recursive matrix are defined as:

[0027]

[0028] Step 3: Construct the continuous wavelet transform time-frequency graph CWT of the subsequence i Let a be the scale parameter that controls the expansion and contraction of the wavelet, and b be the translation parameter that determines the position of the wavelet on the time axis. ψ(t) is the wavelet basis function, and the Morlet wavelet is selected. Then the elements of the continuous wavelet transform time-frequency diagram are defined as:

[0029]

[0030] As a further improvement of the present invention, the step of generating the multi-channel feature map in step 3) is as follows:

[0031] Step 1: Use bilinear interpolation to align the scale of the Gram angle field, recurrence graph, and continuous wavelet transform time-frequency graph of each subsequence. Assume that the pixel values of the image are known at the integer coordinates (x, y), Δx and Δy are fractions between 0 and 1, and P(·) represents the pixel value of the coordinate point. Then, the aligned image has each pixel value as follows:

[0032] P(x+Δx,y+Δy)=(1-Δx)(1-Δy)·P(x,y)+Δx(1-Δy)·P(x+1,y)

[0033] +(1-Δx)Δy·P(x,y+1)+ΔxΔy·P(x+1,y+1)

[0034] Step 2: For each sample, the GAF image, RP image, and CWT time-frequency image are respectively calculated as X, Y, and Z channel signals. Then, the X channel signal is input into the R channel of the RGB image, the Y channel signal is input into the G channel of the RGB image, and the Z channel signal is input into the B channel of the RGB image. Feature diversity is enhanced through three-channel fusion.

[0035] As a further improvement of the present invention, the step of inputting the generated multi-channel feature map into the improved MobileNetV2 network for training in step 4) is as follows:

[0036] Step 1: Divide the multi-channel feature images into training set, validation set, and test set in a ratio of 8:1:1;

[0037] Step 2: Define the improved MobileNetV2 submodule: multi-scale dilated convolution block, which uses convolution kernels with different dilation rates to extract multi-scale features. Let the input be X, the output be Y, the convolution kernel size be k×k, the dilation rate be d, the step size be s, and c be the number of output channels. For d = [1, 2, 4], the convolution operation is expressed as:

[0038] Y=MDConv(X,W,d)

[0039] Where W is the convolution kernel weight, d is the expansion rate, and the activation function is Mish, which is calculated as follows:

[0040] Mish(x)=x*tanh(ln(1+e x ))

[0041] Step 3: Define the improved MobileNetV2 submodule: the improved inverted residual block, and set the input feature to

[0042] Step 4: For the output of the improved inverted residual block, use 1×1 convolution to increase the dimension, and use ReLU6 as the activation function. Let W 1×1 is the 1×1 convolution kernel weight. Its specific calculation formula is:

[0043] Y = Conv2D(X,W)

[0044] Step 5: The final output of the module is input to the global average pooling layer, which uses global average pooling to compress the spatial dimension to 1×1 and focus on effective information. Assuming the shape of the input feature map X is H×W×C, the specific calculation formula is:

[0045]

[0046] Step 6: Use the fully connected classification layer to map the pooling result to the category space. The specific calculation formula is:

[0047] Y=FC(X)=W·X+b

[0048] Among them, W is the weight matrix, b is the bias term, and the activation function is Softmax;

[0049] Step 7: Obtain the final output sequence through the above network model.

[0050] As a further improvement of the present invention, the inverted residual block improved in step 3 in the step 4) training step has the following specific structure:

[0051] (1) Dilated convolution;

[0052] Use 1×1 convolution to expand the number of input channels to a higher dimension and enhance the feature expression capability. The calculation formula is:

[0053] X expand =Mish(W expand *X)

[0054] in, t is the expansion factor, * represents a 1×1 convolution operation, and Mish activation enhances nonlinearity and avoids gradient truncation;

[0055] (2) Multi-scale dilated convolution;

[0056] Use multiple dilated convolutions with different dilation rates to extract features of different scales and fuse them through 1×1 convolution:

[0057] Y1=MDConv(X expand ,dilation=1)

[0058] Y2=MDConv(X expand ,dilation=2)

[0059] Y3=MDConv(X expand ,dilation=4)

[0060]

[0061] in:

[0062] α i =Softmax(W fusion ([Y1,Y2,Y3]))

[0063] The local details and long-range periodic features are captured by deep convolution blocks with different dilation rates, and the dynamic weight α i Generated by lightweight 1×1 convolution, the shallow layer activation function uses Mish. When Mish is only used in the shallow layer, it alleviates problems in the deep layer, such as feature redundancy and overfitting, while retaining Mish's ability to extract useful features in the shallow layer. The deep network switches to the ReLU6 activation function to avoid gradient vanishing and balance training stability.

[0064] (3) NAM module;

[0065] Inserting the NAM module after the multi-scale dilated convolution dynamically adjusts the channel weights and enhances the feature expression capability:

[0066] Y attn =Y fused ⊙(γ·σ(W fc (GAP(Y fused )))

[0067] in, is the BN layer scaling factor, σ is the Sigmoid function, which retains the channel attention effect while introducing only a few parameters;

[0068] (4) Linear bottleneck;

[0069] Use 1×1 convolution to compress the number of channels back to the original dimension and reduce computational complexity:

[0070]

[0071] 1×1 convolution compression channel without activation function, retaining complete feature information;

[0072] (5) Residual connection;

[0073] If the input and output have the same number of channels and the stride is 1, the input is directly added to the output via a residual connection.

[0074] X final =X in +X out .

[0075] Beneficial effects:

[0076] (1) Based on the time series signal slicing and multimodal image fusion strategy, the one-dimensional signal is mapped into a multi-channel composite feature map;

[0077] (2) An improved MobileNetV2 lightweight network is proposed, which enhances the nonlinear expression ability through the Mish function, captures fault-sensitive areas with NAM attention, and extracts cross-band hierarchical features with multi-scale dilated convolution, which significantly improves the model's adaptability to special data scenarios while ensuring computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 Algorithm flow chart;

[0079] Figure 2 Multi-channel feature map construction process;

[0080] Figure 3 Improved inverted residual network model. DETAILED DESCRIPTION

[0081] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0082] The present invention provides a voltage transformer fault diagnosis algorithm based on a multi-channel feature image-improved MobileNetV2, which can improve the accuracy of fault diagnosis and solve the problem of poor diagnostic network feature recognition in the case of small samples and unbalanced fault classification.

[0083] As an embodiment of the present invention, the present invention provides a voltage transformer fault diagnosis algorithm based on a multi-channel feature image-improved MobileNetV2, wherein the flowchart of the voltage transformer fault diagnosis algorithm based on a multi-channel feature image-improved MobileNetV2 proposed by the present invention is as follows: Figure 1 As shown in , the construction of multi-channel feature maps is as follows Figure 2 As shown in , the improved inverted residual network model is as follows Figure 3 As shown, and achieved by the following steps:

[0084] 1) Perform partially overlapping sliding window slicing on the voltage transformer secondary side voltage time series signal:

[0085] The steps of performing sliding window slicing on the time series signal are as follows:

[0086] Step 1: Use the collected voltage signal time series, recorded as X = {x1, x2, ..., x N}; Using partially overlapping window slicing with a window size of w, a step size of s, and an overlap ratio of p, multiple subsequences can be generated:

[0087] X i ={x i ,x i+1 ,...,x i+w-1},i=1,2,...N-w+1

[0088] Step 2: To prevent the deep learning model from overfitting, add Gaussian white noise with a signal-to-noise ratio of N to the subsequence to obtain the subsequence after time series data enhancement. i is a Gaussian white noise sample with a standard deviation of ε, then the subsequence after time series data enhancement is:

[0089] X i ={x i +ξ i ,x i+1 +ξ i+1 ,...,x i+w-1 +ξ i+w-1},i=1,2,...N-w+1

[0090] 2) Generate feature images for the subsequences after time series data enhancement;

[0091] The steps of generating feature images for the subsequence after time series data enhancement are as follows:

[0092] Step 1: Construct the Gram Angle Field (GAF) of the subsequence i, scale the subsequence to the range [0,1] through extreme value normalization, convert the normalized data into polar coordinate representation, and further generate the Gram matrix through polar coordinate angle operation to obtain the Gram angle field of the subsequence. Let I be the unit row vector and X be the subsequence, then the Gram matrix can be expressed as:

[0093]

[0094] Step 2: Construct the recursive graph RP of the subsequence i , the time series is converted into phase space trajectory by time delay embedding method. Let embedding dimension be m, which determines the complexity of phase space reconstruction; delay time is τ, which affects the expansion degree of phase space trajectory; the i-th phase space vector is obtained as Let Θ be the step function, ||·|| be the norm, and ε be the similarity threshold, which controls the similarity determination of the state. Then the elements of the recursive matrix are defined as:

[0095]

[0096] Step 3: Construct the continuous wavelet transform time-frequency graph CWT of the subsequence i , let a be the scale parameter, which controls the expansion and contraction of the wavelet, and b be the translation parameter, which determines the position of the wavelet on the time axis. ψ(t) is the wavelet basis function, and Morlet wavelet is selected. The elements of the continuous wavelet transform time-frequency diagram are defined as:

[0097]

[0098] 3) Generation of multi-channel feature maps;

[0099] The steps for generating the multi-channel feature map are:

[0100] Step 1: Use bilinear interpolation to align the scale of the Gram angle field, recurrence graph, and continuous wavelet transform time-frequency graph of each subsequence. Assume that the pixel values of the image are known at the integer coordinates (x, y), Δx and Δy are fractions between 0 and 1, and P(·) represents the pixel value at the coordinate point. Then, after alignment, each pixel value is:

[0101] P(x+Δx,y+Δy)=(1-Δx)(1-Δy)·P(x,y)+Δx(1-Δy)·P(x+1,y)

[0102] +(1-Δx)Δy·P(x,y+1)+ΔxΔy·P(x+1,y+1)

[0103] Step 2: For each sample, the GAF image, RP image, and CWT time-frequency image are calculated as X, Y, and Z channel signals respectively. Then, the X channel signal is input into the R channel of the RGB image, the Y channel signal is input into the G channel of the RGB image, and the Z channel signal is input into the B channel of the RGB image. Feature diversity is enhanced through three-channel fusion;

[0104] 4) The generated multi-channel feature map is input into the improved MobileNetV2 network for training.

[0105] Step 1: Divide the multi-channel feature images into training set, validation set, and test set in a ratio of 8:1:1;

[0106] Step 2: Define the improved MobileNetV2 submodule: the multi-scale dilated convolution block, which uses convolution kernels with different dilation rates to extract multi-scale features. Let the input be X, the output be Y, the convolution kernel size be k×k, the dilation rate be d, the stride be s, and c be the number of output channels. For d = [1, 2, 4], the convolution operation is expressed as:

[0107] Y=MDConv(X,W,d)

[0108] Where W is the convolution kernel weight, d is the expansion rate, and the activation function is Mish, which is calculated as follows:

[0109] Mish(x)=x*tanh(ln(1+e x ))

[0110] Step 3: Define the improved MobileNetV2 submodule: the improved inverted residual block, and set the input feature to The specific structure of the improved inverted residual module is as follows:

[0111] (1) Dilated convolution

[0112] Use 1×1 convolution to expand the number of input channels to a higher dimension and enhance the feature expression capability. The calculation formula is:

[0113] X expand =Mish(W expand *X)

[0114] in, t is the expansion factor, * represents a 1×1 convolution operation, and Mish activation enhances nonlinearity and avoids gradient truncation.

[0115] (2) Multi-scale dilated convolution

[0116] Use multiple dilated convolutions with different dilation rates to extract features of different scales and fuse them through 1×1 convolution:

[0117] Y1=MDConv(X expand ,dilation=1)

[0118] Y2=MDConv(X expand ,dilation=2)

[0119] Y3=MDConv(X expand ,dilation=4)

[0120]

[0121] in:

[0122] α i =Softmax(W fusion ([Y1,Y2,Y3]))

[0123] The local details and long-range periodic features are captured by deep convolution blocks with different dilation rates, and the dynamic weight α i Generated by lightweight 1×1 convolutions. Mish is used as the activation function for shallow layers. When Mish is used only in shallow layers, it can alleviate issues in deeper layers, such as feature redundancy and overfitting, while retaining Mish's ability to extract useful features in shallow layers. Deep networks use the ReLU6 activation function to prevent vanishing gradients and maintain training stability.

[0124] (3) NAM module;

[0125] Inserting the NAM module after the multi-scale dilated convolution dynamically adjusts the channel weights and enhances the feature expression capability:

[0126] Y attn =Y fused ⊙(γ·σ(W fc (GAP(Y fused )))

[0127] in, is the BN layer scaling factor, and σ is the Sigmoid function, which retains the channel attention effect while introducing only a few parameters.

[0128] (4) Linear bottleneck;

[0129] Use 1×1 convolution to compress the number of channels back to the original dimension and reduce computational complexity:

[0130]

[0131] The 1×1 convolution without activation function compresses the channel and retains the complete feature information.

[0132] (5) Residual connection;

[0133] If the input and output have the same number of channels and the stride is 1, the input is directly added to the output via a residual connection.

[0134] X final =X in +X out

[0135] Step 4: For the output of the improved inverted residual block, use 1×1 convolution to increase the dimension, and use ReLU6 as the activation function. Let W 1×1 is the 1×1 convolution kernel weight. Its specific calculation formula is:

[0136] Y = Conv2D(X,W)

[0137] Step 5: The final output of the module is input to the global average pooling layer, which uses global average pooling to compress the spatial dimension to 1×1 and focus on effective information. Assuming the shape of the input feature map X is H×W×C, the specific calculation formula is:

[0138]

[0139] Step 6: Use the fully connected classification layer to map the pooling result to the category space. The specific calculation formula is:

[0140] Y=FC(X)=W·X+b

[0141] Among them, W is the weight matrix, b is the bias term, and the activation function is Softmax.

[0142] Step 7: Obtain the final output sequence through the above network model.

[0143] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. Based on multi-channel feature image-improved MobileNetV2 voltage transformer fault diagnosis method, the characteristics are: The steps include: 1) Perform partially overlapping sliding window slicing on the voltage transformer secondary side voltage time series signal: 2) Generate feature images for the subsequences after time series data enhancement; 3) Generation of multi-channel feature maps; 4) The generated multi-channel feature map is input into the improved MobileNetV2 network for training.

2. The voltage transformer fault diagnosis method based on multi-channel feature image-improved MobileNetV2 according to claim 1 is characterized in that: The step 1) of performing sliding window slicing on the time series signal is as follows: Step 1: Use the collected voltage signal time series, recorded as X = {x1, x2, ..., x N }; Use partially overlapping window slicing with a window size of w, a step size of s, and an overlap ratio of p to generate multiple subsequences: X i ={x i ,x i+1 ,...,x i+w-1 },i=1,2,...N-w+1 Step 2: To prevent the deep learning model from overfitting, add Gaussian white noise with a signal-to-noise ratio of N to the subsequence to obtain the subsequence after time series data enhancement. i is a Gaussian white noise sample with a standard deviation of σ, then the subsequence after time series data enhancement is: X i ={x i +ξ i ,x i+1 +ξ i+1 ,...,x i+w-1 +ξ i+w-1 },i=1,2,...N-w+1.

3. The voltage transformer fault diagnosis method based on multi-channel feature image-improved MobileNetV2 according to claim 1 is characterized in that: The step 2) generates a feature image for the subsequence after the time series data enhancement is as follows: Step 1: Construct the Gram Angle Field (GAF) of the subsequence i , scale the subsequence to the interval [0,1] by extreme value normalization, convert the normalized data into polar coordinate representation, and further generate the Gram matrix through polar coordinate angle operation to obtain the Gram angle field of the subsequence. Let I be the unit row vector and X be the subsequence. The Gram matrix can be expressed as: Step 2: Construct the recursive graph RP of the subsequence i The time series is converted into phase space trajectory by time delay embedding method. The embedding dimension is set to m, which determines the complexity of phase space reconstruction; the delay time is τ, which affects the expansion degree of phase space trajectory; the i-th phase space vector is obtained as Let Θ be the step function, ||·|| be the norm, and ε be the similarity threshold, which controls the similarity determination of the state. Then the elements of the recursive matrix are defined as: Step 3: Construct the continuous wavelet transform time-frequency graph CWT of the subsequence i Let a be the scale parameter that controls the expansion and contraction of the wavelet, and b be the translation parameter that determines the position of the wavelet on the time axis. ψ(t) is the wavelet basis function, and the Morlet wavelet is selected. Then the elements of the continuous wavelet transform time-frequency diagram are defined as:

4. The voltage transformer fault diagnosis method based on multi-channel feature image-improved MobileNetV2 according to claim 1 is characterized in that: The steps for generating the multi-channel feature map in step 3) are: Step 1: Use bilinear interpolation to align the scale of the Gram angle field, recurrence graph, and continuous wavelet transform time-frequency graph of each subsequence. Assume that the pixel values of the image are known at the integer coordinates (x, y), Δx and Δy are fractions between 0 and 1, and P(·) represents the pixel value of the coordinate point. Then, the aligned image has each pixel value as follows: P(x+Δx,y+Δy)=(1-Δx)(1-Δy)·P(x,y)+Δx(1-Δy)·P(x+1,y) +(1-Δx)Δy·P(x,y+1)+ΔxΔy·P(x+1,y+1) Step 2: For each sample, the GAF image, RP image, and CWT time-frequency image are respectively calculated as X, Y, and Z channel signals. Then, the X channel signal is input into the R channel of the RGB image, the Y channel signal is input into the G channel of the RGB image, and the Z channel signal is input into the B channel of the RGB image. Feature diversity is enhanced through three-channel fusion.

5. The voltage transformer fault diagnosis method based on multi-channel feature image-improved MobileNetV2 according to claim 1 is characterized in that: In step 4), the generated multi-channel feature map is input into the improved MobileNetV2 network for training. The training steps are: Step 1: Divide the multi-channel feature images into training set, validation set, and test set in a ratio of 8:1:1; Step 2: Define the improved MobileNetV2 submodule: multi-scale dilated convolution block, which uses convolution kernels with different dilation rates to extract multi-scale features. Let the input be X, the output be Y, the convolution kernel size be k×k, the dilation rate be d, the step size be s, and c be the number of output channels. For d = [1, 2, 4], the convolution operation is expressed as: Y=MDConv(X,W,d) Where W is the convolution kernel weight, d is the expansion rate, and the activation function is Mish, which is calculated as follows: Mish(x)=x*tanh(ln(1+e x )) Step 3: Define the improved MobileNetV2 submodule: the improved inverted residual block, and set the input feature to Step 4: For the output of the improved inverted residual block, use 1×1 convolution to increase the dimension, and use ReLU6 as the activation function. Let W 1×1 is the 1×1 convolution kernel weight. Its specific calculation formula is: Y = Conv2D(X,W) Step 5: The final output of the module is input to the global average pooling layer, which uses global average pooling to compress the spatial dimension to 1×1 and focus on effective information. Assuming the shape of the input feature map X is H×W×C, the specific calculation formula is: Step 6: Use the fully connected classification layer to map the pooling result to the category space. The specific calculation formula is: Y=FC(X)=W·X+b Among them, W is the weight matrix, b is the bias term, and the activation function is Softmax; Step 7: Obtain the final output sequence through the above network model.

6. The voltage transformer fault diagnosis method based on multi-channel feature image-improved MobileNetV2 according to claim 5 is characterized in that: The inverted residual block improved in step 3 in the step 4) training step has the following specific structure: (1) Dilated convolution; Use 1×1 convolution to expand the number of input channels to a higher dimension and enhance the feature expression capability. The calculation formula is: X expand =Mish(W expand *X) in, t is the expansion factor, * represents a 1×1 convolution operation, and Mish activation enhances nonlinearity and avoids gradient truncation; (2) Multi-scale dilated convolution; Use multiple dilated convolutions with different dilation rates to extract features of different scales and fuse them through 1×1 convolution: Y1=MDConv(X expand ,dilation=1) Y2=MDConv(X expand ,dilation=2) Y3=MDConv(X expand ,dilation=4) in: α i =Softmax(W fusion ([Y1,Y2,Y3])) The local details and long-range periodic features are captured by deep convolution blocks with different dilation rates, and the dynamic weight α i Generated by lightweight 1×1 convolution, the shallow layer activation function uses Mish. When Mish is only used in the shallow layer, it alleviates problems in the deep layer, such as feature redundancy and overfitting, while retaining Mish's ability to extract useful features in the shallow layer. The deep network switches to the ReLU6 activation function to avoid gradient vanishing and balance training stability. (3) NAM module; Inserting the NAM module after the multi-scale dilated convolution dynamically adjusts the channel weights and enhances the feature expression capability: Y attn =Y fused ⊙(γ·σ(W fc (GAP(Y fused ))) in, is the BN layer scaling factor, σ is the Sigmoid function, which retains the channel attention effect while introducing only a few parameters; (4) Linear bottleneck; Use 1×1 convolution to compress the number of channels back to the original dimension and reduce computational complexity: 1×1 convolution compression channel without activation function, retaining complete feature information; (5) Residual connection; If the input and output have the same number of channels and the stride is 1, the input is directly added to the output via a residual connection. X final =X in +X out 。

Images