A mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning

Abstract

The present invention discloses a mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning. First, the actual fault data of the mechanical equipment to be diagnosed is obtained, and adaptive decomposition is performed based on the improved VMD algorithm to obtain multiple intrinsic mode components, thereby reducing the degree of manual intervention. Secondly, based on the preset signal evaluation index, the multiple intrinsic mode components are screened and reconstructed to obtain a reconstructed signal, which can selectively retain those components rich in fault information and effectively reduce noise. Finally, the reconstructed signal is fused to obtain fused fault data, which is input into a trained fault diagnosis model, and the corresponding fault classification result is output. The fault diagnosis model adopts a dual encoder structure, which has strong noise resistance, excellent feature extraction performance, good generalization ability and high classification accuracy. Finally, the present invention realizes high-precision fault identification and improves the efficiency of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to a mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning, belonging to the technical field of mechanical equipment fault diagnosis. Background Art

[0002] As a core component of the national defense industry, mechanical equipment is widely used in aviation, navigation, aerospace and other fields. In addition, they also play a vital role in high-tech fields such as electronic warfare, missile defense systems, unmanned vehicles, satellite communications, and deep-sea exploration. During the operation of mechanical equipment, the probability of mechanical equipment failure increases significantly due to factors such as wear, fatigue and overload. Therefore, fault diagnosis is of great significance to reducing the maintenance cost of mechanical equipment and improving production efficiency. With the continuous advancement of intelligent manufacturing and industrial Internet of Things technology, many accelerometers and acoustic sensors are deployed on mechanical equipment to collect operation data. However, the information provided by a single sensor is often insufficient to fully describe the operating status of mechanical equipment, and its data output is easily interfered by surrounding environmental factors. In order to more comprehensively obtain the physical and spatial characteristics of mechanical equipment, it is necessary to fuse data from multiple sensors. Therefore, collecting multi-dimensional vibration data obtained by accelerometers plays a vital role in identifying the fault type of mechanical equipment. This approach can significantly improve the understanding of the operating status of mechanical equipment and improve the accuracy and robustness of fault diagnosis.

[0003] In the field of fault diagnosis of mechanical equipment, vibration analysis is a commonly used technology. Since the collected vibration signals usually contain multiple components and exhibit non-stationary and nonlinear characteristics, these signals contain rich fault information, which is crucial for fault identification. However, the mixed noise in the signal makes it difficult to extract useful fault features. Therefore, developing effective signal preprocessing methods and extracting fault features from non-stationary vibration signals interfered by noise are key challenges in realizing fault diagnosis and early warning of mechanical equipment. Solving this problem is of great significance to improving the reliability and safety of mechanical equipment.

[0004] In the study of reducing noise interference in non-stationary vibration signals, researchers have proposed a variety of techniques including time domain analysis, short-time Fourier transform (STFT), discrete wavelet transform (DWT) and variational mode decomposition (VMD). With the increase in data size, data-driven deep learning methods, especially convolutional neural networks (CNNs), have been widely used in the field of fault diagnosis due to their excellent feature extraction and characterization capabilities. However, in practical applications, the noise in the signal often interferes with the accurate extraction of fault features by neural networks, and the accuracy of classification is also affected when the training samples are limited. Although multi-sensor data can theoretically improve the accuracy of fault diagnosis, most studies are still limited to single sensor data and fail to fully utilize the feature information of multiple positions of mechanical equipment. In addition, simply combining or splicing data may lead to poor classification performance. In order to solve these problems, it is necessary to further study and develop new methods to enhance the application effect of multi-source data in fault diagnosis. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning. First, based on the improved VMD algorithm, the adaptive decomposition of the position channel signal is completed to reduce the degree of manual intervention. Secondly, based on the preset signal evaluation index, those components rich in fault information can be selectively retained and the noise can be effectively reduced. Finally, the constructed fault diagnosis model adopts a dual encoder structure, and processes multi-source data based on a sparse self-attention mechanism and a gating unit. It has strong noise resistance, excellent feature extraction performance, good generalization ability and high classification accuracy.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] The present invention discloses a mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning, comprising the following steps:

[0008] Acquire actual fault data of the mechanical equipment to be diagnosed; wherein the actual fault data includes a plurality of position channel signals;

[0009] For any position channel signal, the optimal decomposition number and the optimal penalty factor are determined based on the improved VMD algorithm for adaptive decomposition to obtain multiple intrinsic mode components;

[0010] Based on the preset signal evaluation index, multiple intrinsic mode components are screened and reconstructed to obtain a reconstructed signal;

[0011] According to the reconstructed signals corresponding to the signals of multiple position channels, multi-source data fusion is performed to obtain fused fault data;

[0012] Inputting the fused fault data into a trained fault diagnosis model and outputting a corresponding fault classification result;

[0013] Wherein, the fault diagnosis model includes:

[0014] A step encoder is used to perform temporal feature encoding based on a sparse self-attention mechanism and a feedforward mechanism according to the fused fault data to obtain temporal feature data;

[0015] A channel encoder is used to perform channel feature encoding based on a sparse self-attention mechanism and a feedforward mechanism according to the fused fault data to obtain channel feature data;

[0016] The gating unit is used to perform linear mapping, splicing and normalization processing according to the time characteristic data and the channel characteristic data to obtain a corresponding fault classification result.

[0017] Furthermore, the step of determining the optimal decomposition quantity and the optimal penalty factor based on the improved VMD algorithm for adaptive decomposition includes the following steps:

[0018] For any position channel signal, initialize parameters, the parameters including the frequency domain representation of the intrinsic mode component, the frequency domain representation of the center frequency of the intrinsic mode component, the frequency domain value of the penalty term, the number of intrinsic mode component decompositions and the second-order penalty factor;

[0019] Based on the improved VMD algorithm, an iterative cycle is performed until a preset termination condition is met, and an optimal decomposition number and an optimal penalty factor are output; the corresponding position channel signal is decomposed according to the optimal decomposition number and the optimal penalty factor to obtain multiple intrinsic mode components;

[0020] Each cycle includes the following steps:

[0021] Calculate the signal decomposition quality of the eigenmode components in the current iteration;

[0022] In response to the signal decomposition quality of the intrinsic mode component in the current iteration being less than a preset signal decomposition quality threshold, and the current frequency variable being greater than a preset frequency variable threshold, the parameters are updated and the calculation step of the signal decomposition quality is repeated to perform the next iteration loop; otherwise, the loop is terminated and the eigenmode component decomposition quantity and the second-order penalty factor of the current iteration are taken as the optimal decomposition quantity and the optimal penalty factor.

[0023] Furthermore, the expression of the signal decomposition quality is as follows:

[0024] ;

[0025] In the formula, represents the signal decomposition quality of the kth eigenmode component; represents the calculation index weight of the kth eigenmode component; Represents the calculation index weight of the kth eigenmode component. When k=1, The value is 0;

[0026] represents the first coefficient; represents the kurtosis value of the kth eigenmode component; represents the second coefficient; represents the correlation coefficient of the kth eigenmode component; represents the third coefficient; represents the information entropy of the kth eigenmode component.

[0027] Further, the updating parameters comprises the following steps:

[0028] Update the decomposition quantity of the eigenmode component, the expression is as follows:

[0029] ;

[0030] In the formula, Indicates the number of eigenmode component decompositions at the n+1th iteration; Indicates the number of eigenmode component decompositions at the nth iteration;

[0031] Update the second-order penalty factor, the expression is as follows:

[0032] ;

[0033] In the formula, Represents the second-order penalty factor at the n+1th iteration; Represents the second-order penalty factor at the nth iteration;

[0034] Update the frequency domain representation of the eigenmode component, the expression is as follows:

[0035] ;

[0036] In the formula, represents the frequency domain representation of the kth eigenmode component obtained at the n+1th iteration, Indicates the current frequency variable; represents the frequency domain representation of the position channel signal f(t); represents the sum of the frequency domain representations of all eigenmode components except the kth eigenmode component; represents the frequency domain value of the penalty term; represents the second-order penalty factor; represents the center frequency of the kth eigenmode component;

[0037] Update the center frequency of the eigenmode component, the expression is as follows:

[0038] ;

[0039] In the formula, represents the center frequency of the kth eigenmode component at the n+1th iteration; Indicates the current frequency variable; Indicates the current frequency variable The differential of

[0040] Update the penalty term frequency domain value , the expression is as follows:

[0041] ;

[0042] In the formula, Represents the frequency domain value of the penalty term at the n+1th iteration; Represents the frequency domain value of the penalty term at the nth iteration; represents the noise tolerance parameter; represents the frequency domain representation of the position channel signal f(t); represents the sum of the frequency domain representations of the eigenmode components obtained at the n+1th iteration.

[0043] Furthermore, the signal screening and reconstruction of multiple intrinsic mode components based on the preset signal evaluation index includes the following steps:

[0044] Calculate the kurtosis sparsity value of each eigenmode component;

[0045] In response to the kurtosis sparsity value of any eigenmode component being greater than a preset kurtosis sparsity threshold, the eigenmode component is used as a modal component to be reconstructed; otherwise, the eigenmode component is not used as a modal component to be reconstructed;

[0046] According to all the modal components to be reconstructed, splicing processing is performed to obtain the reconstructed signal.

[0047] Furthermore, the expression of the kurtosis sparse value is as follows:

[0048]

[0049] In the formula, represents the kth eigenmode component; represents the number of eigenmode components; represents the sparsity of the kth eigenmode component; represents the kurtosis value of the kth eigenmode component; Represents the dot product value of the sparsity and kurtosis value of the kth eigenmode component; Represents the kurtosis sparse value of the kth eigenmode component.

[0050] Furthermore, the training method of the fault diagnosis model is as follows:

[0051] Acquire original fault data of mechanical equipment; wherein the original fault data includes multiple position channel signals, and the original fault data carries a fault label;

[0052] For any position channel signal, the number of decompositions and penalty factors are determined based on the iterative optimization method to perform adaptive decomposition and obtain multiple intrinsic mode components;

[0053] Based on the preset signal evaluation index, multiple intrinsic mode components are screened and reconstructed to obtain a reconstructed signal;

[0054] According to the reconstructed signals corresponding to the signals of multiple position channels, multi-source data fusion and sliding window sampling are performed to obtain a training set;

[0055] The pre-built fault diagnosis model is trained according to the training set to obtain a trained fault diagnosis model.

[0056] Furthermore, the step size encoder comprises:

[0057] A first embedding layer, used for performing vectorization processing according to the fused fault data to obtain a first embedding vector;

[0058] A first sparse self-attention layer, configured to perform sparse self-attention calculation according to the first embedding vector to obtain a weighted first embedding vector;

[0059] A first residual regularization layer, used for performing residual connection and regularization processing according to the weighted first embedding vector to obtain a regularized first embedding vector;

[0060] A first feedforward network, used to perform a nonlinear transformation of a feedforward neural network according to the regularized first embedding vector to obtain a first embedding vector processed by the feedforward network;

[0061] The second residual regularization layer is used to perform residual connection and regularization processing according to the first embedding vector processed by the feedforward network to obtain final time feature data.

[0062] Furthermore, the channel encoder comprises:

[0063] A second embedding layer, used for performing vectorization processing according to the fused fault data to obtain a second embedding vector;

[0064] A second sparse self-attention layer is used to perform sparse self-attention calculation according to the second embedding vector of each channel to obtain a weighted second embedding vector;

[0065] A third residual regularization layer, used for performing residual connection and regularization processing according to the second embedding vector calculated by the self-attention, so as to obtain a second embedding vector after residual regularization;

[0066] A second feedforward network is used to perform a nonlinear transformation of a feedforward neural network according to the second embedding vector data after residual regularization to obtain a second embedding vector processed by the feedforward network;

[0067] The fourth residual regularization layer is used to perform residual connection and regularization processing according to the second embedding vector processed by the feedforward network to obtain final channel feature data.

[0068] Furthermore, the gate control unit includes:

[0069] A Gated layer is used to perform a nonlinear activation operation according to the time feature data and the channel feature data to obtain a corresponding time output vector and a channel output vector;

[0070] A Linear layer is used to perform a weighted concatenation operation and then a linear mapping according to the time output vector and the channel output vector to obtain a weighted vector;

[0071] The Softmax layer is used to perform a normalization operation according to the weighted vector to obtain a step encoder weight and a channel encoder weight; and to perform weighted feature fusion of two channels according to the time output vector, the channel output vector, the step encoder weight and the channel encoder weight to obtain a corresponding fault classification result.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] The mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning of the present invention, first, completes the adaptive decomposition of the position channel signal based on the improved VMD algorithm, performs the optimal parameter search operation, and reduces the degree of manual intervention. Secondly, based on the preset signal evaluation index, multiple intrinsic mode components are screened and reconstructed, and those components rich in fault information are selectively retained. This process not only realizes the reconstruction of the signal, but also simultaneously achieves the purpose of noise reduction, thereby improving the signal quality and accuracy of fault diagnosis. Finally, the present invention designs an efficient fault diagnosis model, adopts the structure of a dual encoder of a step encoder and a channel encoder, respectively extracts the features of the time step and channel information, and adopts a sparse self-attention mechanism and a gating unit to process multi-source data to achieve accurate fault identification, while improving the efficiency of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a flow chart of a mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning provided in Example 1;

[0075] Figure 2 is a schematic diagram of the training process of the fault diagnosis model provided in Example 1;

[0076] Figure 3 is a structural diagram of the fault diagnosis model provided in Example 1;

[0077] Figure 4 is a time domain diagram and a frequency domain diagram of the modal components after the signal decomposition provided in Example 2;

[0078] Figure 5 Schematic diagram of the confusion matrix provided in Example 2. DETAILED DESCRIPTION

[0079] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0080] Example 1

[0081] This embodiment 1 provides a mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning, comprising the following steps:

[0082] Acquire actual fault data of the mechanical equipment to be diagnosed; wherein the actual fault data includes a plurality of position channel signals;

[0083] For any position channel signal, the number of decompositions and penalty factors are determined based on the improved VMD algorithm for adaptive decomposition to obtain multiple intrinsic mode components.

[0084] Based on the preset signal evaluation index, multiple intrinsic mode components are screened and reconstructed to obtain a reconstructed signal;

[0085] According to the reconstructed signals corresponding to the signals of multiple position channels, multi-source data fusion is performed to obtain fused fault data;

[0086] Input the fused fault data into the trained fault diagnosis model and output the corresponding fault classification results;

[0087] Among them, the fault diagnosis model includes:

[0088] A step encoder is used to encode the temporal features based on the sparse self-attention mechanism and the feedforward mechanism according to the fused fault data to obtain the temporal feature data;

[0089] A channel encoder is used to encode channel features based on the fused fault data and a sparse self-attention mechanism and a feedforward mechanism to obtain channel feature data;

[0090] The gating unit is used to perform linear mapping, splicing and normalization processing according to the time feature data and the channel feature data to obtain the corresponding fault classification result.

[0091] The design concept of the present invention is as follows: first, based on the improved VMD algorithm, the adaptive decomposition of the position channel signal is completed, the optimal parameter search operation is performed, and the degree of manual intervention is reduced. Secondly, based on the preset signal evaluation index, the signal screening and reconstruction of multiple intrinsic mode components are performed, and those components rich in fault information are selectively retained. This process not only realizes the reconstruction of the signal, but also simultaneously achieves the purpose of noise reduction, thereby improving the signal quality and accuracy of fault diagnosis. Finally, the present invention designs an efficient fault diagnosis model, adopts the structure of a dual encoder of a step encoder and a channel encoder, respectively extracts the features of the time step and channel information, and adopts a sparse self-attention mechanism and a gating unit to process multi-source data to achieve accurate fault identification, while improving the efficiency of fault diagnosis.

[0092] like Figure 1 As shown, the specific steps are as follows:

[0093] Step 1: Acquire actual fault data of the mechanical equipment to be diagnosed; wherein the actual fault data includes multiple position channel signals.

[0094] Step 2: For any position channel signal, the optimal decomposition number and the optimal penalty factor are determined based on the improved VMD algorithm to perform adaptive decomposition to obtain multiple intrinsic mode components.

[0095] The invention performs optimal parameter search operation based on an improved VMD algorithm, determines the optimal decomposition quantity and the optimal penalty factor, adaptively decomposes the original position channel signal, and decomposes it into a series of intrinsic mode components with limited bandwidth.

[0096] The specific principles are as follows:

[0097] (1) Define position channel signal is composed of K finite bandwidth eigenmode components The center frequency of each eigenmode component is , the constraint condition is that the intrinsic mode component and the position channel signal are equal to the input, and the corresponding variational constraint model is constructed. The specific operations are as follows:

[0098] First, the eigenmode components are transformed using the Hilbert transform Analyze and obtain its unilateral spectrum for . Secondly, through exponential mixing modulation, that is, with the operator Multiply them together to modulate the spectrum of each eigenmode component to the response baseband. Finally, the modulated eigenmode component is smoothed using a Gaussian function to ensure the smoothness of the function in the H1 space, thereby estimating the eigenmode component. The bandwidth is the square of the 2-norm of the gradient. represents the decomposed eigenmode component, Represents the center frequency of each eigenmode component.

[0099] The expression of the variational constraint model is as follows:

[0100]

[0101]

[0102] In the formula, represents the kth eigenmode component at time t; represents the instantaneous amplitude of the kth eigenmode component at time t; represents the phase angle of the kth eigenmode component at time t; represents the partial derivative operator; represents the Dirac function, which is used to represent the unit pulse; represents the imaginary unit, ; represents a complex exponential function, which is used to represent the frequency component of the kth eigenmode component. represents the center frequency of the kth eigenmode component at time t; Indicates subject to, that is, the conditions are met.

[0103] (2) In order to find the optimal solution to the constrained variational problem, we first introduce the Lagrange multiplier and the second-order penalty factor , transforming the variational constraint problem into an unconstrained variational problem. The second-order penalty factor It can ensure the accuracy of signal reconstruction in Gaussian noise environment. Lagrange multiplier The strictness of the constraints can be guaranteed. Solving this problem can convert the constrained optimization problem into an unconstrained optimization problem. The extended Lagrangian expression is as follows:

[0104]

[0105] same, represents the decomposed eigenmode component, represents the center frequency of each eigenmode component, This ensures the rigor of the variational model. The accuracy of signal decomposition and reconstruction is guaranteed.

[0106] (3) Continuously update each component and its center frequency through the alternating direction multiplier method (ADMM) , and finally the saddle point of the unconstrained model is obtained, which is the optimal solution to the original problem. All components can be obtained according to the frequency domain space by the following formula;

[0107]

[0108] In the formula, represents the frequency domain representation of the kth eigenmode component obtained at the n+1th iteration, Indicates the current frequency variable; represents the frequency domain representation of the position channel signal f(t); represents the sum of the frequency domain representations of all eigenmode components except the kth eigenmode component; represents the Lagrange multiplier The frequency domain representation of is used to adjust the weighting function of the importance of different frequency components, that is, the frequency domain value of the penalty term; represents the second-order penalty factor; represents the center frequency of the kth eigenmode component.

[0109]

[0110] In the formula, represents the center frequency of the kth eigenmode component at the n+1th iteration; Indicates the current frequency variable; Indicates the current frequency variable The differential of .

[0111] The present invention re-estimates the centroid frequency according to the centroid of the power spectrum of each eigenmode component, and simultaneously calculates the kurtosis value, correlation coefficient, information entropy, and value of the index weight QualityIndex, and performs threshold judgment to adaptively determine the optimal decomposition quantity and the optimal penalty factor. The specific operations are as follows:

[0112] For any position channel signal, initialize the parameters, including the frequency domain representation of the intrinsic mode component 、 Frequency domain representation of the center frequency of the eigenmode component 、 Frequency domain value of penalty term , the number of eigenmode component decompositions and the second-order penalty factor .

[0113] Based on the improved VMD algorithm, an iterative cycle is performed until the preset termination condition is met, and the optimal decomposition number and the optimal penalty factor are output; according to the optimal decomposition number and the optimal penalty factor, the corresponding position channel signal is decomposed to obtain multiple intrinsic mode components.

[0114] Each cycle includes the following steps:

[0115] Calculate the signal decomposition quality of the intrinsic mode component in the current iteration, as follows:

[0116]

[0117] In the formula, represents the signal decomposition quality of the kth eigenmode component; represents the calculation index weight of the kth eigenmode component; Represents the calculation index weight of the kth eigenmode component. When k=1, The value is 0;

[0118] represents the first coefficient; represents the kurtosis value of the kth eigenmode component; represents the second coefficient; represents the correlation coefficient of the kth eigenmode component; represents the third coefficient; represents the information entropy of the kth eigenmode component.

[0119] In response to the signal decomposition quality of the intrinsic mode component in the current iteration being less than a preset signal decomposition quality threshold of 5%, and the current frequency variable When it is greater than the preset frequency variable threshold 0, the parameters are updated and the calculation of the signal decomposition quality is repeated to perform the next iteration cycle; otherwise, the loop is terminated and the number of intrinsic mode component decompositions K and the second-order penalty factor at this time are output. .

[0120] The specific update parameters are as follows:

[0121] Update the decomposition quantity of the eigenmode component, the expression is as follows:

[0122]

[0123] In the formula, Indicates the number of eigenmode component decompositions at the n+1th iteration; Represents the number of eigenmode component decompositions at the nth iteration.

[0124] Update the second-order penalty factor, the expression is as follows:

[0125]

[0126] In the formula, Represents the second-order penalty factor at the n+1th iteration; Represents the second-order penalty factor at the nth iteration.

[0127] Update the frequency domain representation of the eigenmode component, the expression is as follows:

[0128]

[0129] In the formula, represents the frequency domain representation of the kth eigenmode component obtained at the n+1th iteration, Indicates the current frequency variable; represents the frequency domain representation of the position channel signal f(t); represents the sum of the frequency domain representations of all eigenmode components except the kth eigenmode component; represents the frequency domain value of the penalty term; represents the second-order penalty factor; represents the center frequency of the kth eigenmode component.

[0130] Update the center frequency of the eigenmode component, the expression is as follows:

[0131]

[0132] In the formula, represents the center frequency of the kth eigenmode component at the n+1th iteration; Indicates the current frequency variable; Indicates the current frequency variable The differential of .

[0133] Update the penalty term frequency domain value , the expression is as follows:

[0134]

[0135] In the formula, Represents the frequency domain value of the penalty term at the n+1th iteration; Represents the frequency domain value of the penalty term at the nth iteration; represents the noise tolerance parameter; represents the frequency domain representation of the position channel signal f(t); represents the sum of the frequency domain representations of the eigenmode components obtained at the n+1th iteration.

[0136] Step 3: Based on the preset signal evaluation index, multiple intrinsic mode components are screened and reconstructed to obtain a reconstructed signal.

[0137] The present invention performs signal screening and analysis on multiple intrinsic mode components based on preset signal evaluation indicators, and while completing noise reduction, more fault information may be retained, and finally the signal is reconstructed, including the following steps:

[0138] Calculate the kurtosis sparsity value of each eigenmode component;

[0139] In response to the kurtosis sparsity value of any eigenmode component being greater than a preset kurtosis sparsity threshold, the eigenmode component is used as a modal component to be reconstructed; otherwise, it is not used as a modal component to be reconstructed;

[0140] According to all the modal components to be reconstructed, splicing processing is performed to obtain a reconstructed signal;

[0141] The expression of the kurtosis sparse value is as follows:

[0142]

[0143] In the formula, represents the kth eigenmode component; represents the number of eigenmode components; represents the sparsity of the kth eigenmode component; represents the kurtosis value of the kth eigenmode component; Represents the dot product value of the sparsity and kurtosis value of the kth eigenmode component; Represents the kurtosis sparse value of the kth eigenmode component.

[0144] In this embodiment, the kurtosis sparsity threshold is set to 0, that is, when the kurtosis sparsity value of any eigenmode component is greater than 0, the eigenmode component is used as the mode component to be reconstructed; otherwise, it is not used as the mode component to be reconstructed.

[0145] Step 4: Perform multi-source data fusion based on the reconstructed signals corresponding to the multiple position channel signals to obtain fused fault data.

[0146] Through SAVMD processing and the screening and reconstruction of the corresponding components, we obtain a single column of reconstructed signal values ​​that best match the original signal sequence; the same processing of different sensor signals at the same position will result in a column of reconstructed signals; we perform a splicing and fusion operation on these single column reconstructed signals to obtain fused fault data, in which there is a strong correlation between the columns.

[0147] Step 5: Input the fused fault data into the trained fault diagnosis model and output the corresponding fault classification results.

[0148] 5.1. Construction of fault diagnosis model.

[0149] like Figure 3 As shown, the fault diagnosis model designed by the present invention receives multi-channel data as input and outputs fault classification results. The dual encoder structure of the step encoder and the channel encoder is used to process the time step and channel information respectively. The dual encoder structure uses the attention mechanism and mask strategy to simultaneously focus on the information of the time and space dimensions, and reduces the computational cost through the sparse self-attention mechanism. At the end of the model, the gating mechanism and the Softmax fully connected layer are introduced to finally output the probability distribution of the fault classification, that is, the fault type of the mechanical equipment, which specifically includes the following;

[0150] 5.1.1. Step encoder.

[0151] To effectively encode the temporal features of time series data, we design a specialized step encoder. The encoder uses scaled dot products inside a multi-head sparse self-attention layer to identify key time-step attention weights. Similar to the standard Transformer model, position information is incorporated into feed-forward networks that are stacked on top of the self-attention layer to enhance feature representation. Residual connections and layer normalization are applied around each sub-layer, ensuring the model's direct information and gradient flow are passed and stabilized during processing.

[0152] The sparse self-attention mechanism focuses on elements with more information by introducing masks or constraints. In this mechanism, different queries pay different attention to the key, and queries can be divided into active values ​​and lazy values. By filtering out the lazy values, the query vector matrix can be sparse. Specifically, the most active elements are filtered out from the original matrix. , forming a new matrix , to replace the original matrix This method not only improves the effect of the attention mechanism, but also reduces the time. Finally, the obtained sparse self-attention formula can process data more efficiently and improve the performance of the model, as shown below.

[0153]

[0154] in, For the filtered activity query matrix and key-value matrix, is a value matrix, and softmax is a fully connected layer; Represents the output data of the attention mechanism; Represents the dimension of the input vector, which is a positive number;

[0155] Specifically, the step encoder includes:

[0156] A first embedding layer, used for performing vectorization processing according to the fused fault data to obtain a first embedding vector;

[0157] A first sparse self-attention layer, used to perform sparse self-attention calculation according to the first embedding vector to obtain a weighted first embedding vector;

[0158] A first residual regularization layer, used for performing residual connection and regularization processing according to the weighted first embedding vector to obtain a regularized first embedding vector;

[0159] A first feedforward network is used to perform a nonlinear transformation of the feedforward neural network according to the regularized first embedding vector to obtain the first embedding vector processed by the feedforward network;

[0160] The second residual regularization layer is used to perform residual connection and regularization processing according to the first embedding vector processed by the feedforward network to obtain the final time feature data.

[0161] 5.1.2. Channel encoder.

[0162] The design concept of the channel encoder is similar to that of the time step encoder. It analyzes the information of each channel in all time steps through a sparse self-attention mechanism and determines and calculates the attention weight matrix accordingly. Since the position information of the channel in the multidimensional signal data is not crucial, just as changing the order of the channels does not affect the essence of the time series, this mechanism enables the model to focus on the relationship between channels and ignore the actual position of the channels. Such a design helps the model capture the deep connection between channels more accurately.

[0163] Specifically, the channel encoder includes:

[0164] The second embedding layer is used to perform vectorization processing according to the fused fault data to obtain a second embedding vector;

[0165] A second sparse self-attention layer is used to perform sparse self-attention calculation according to the second embedding vector of each channel to obtain a weighted second embedding vector;

[0166] The third residual regularization layer is used to perform residual connection and regularization processing according to the second embedding vector calculated by self-attention to obtain the second embedding vector after residual regularization;

[0167] A second feedforward network is used to perform a nonlinear transformation of the feedforward neural network according to the second embedding vector data after residual regularization to obtain a second embedding vector processed by the feedforward network;

[0168] The fourth residual regularization layer is used to perform residual connection and regularization processing according to the second embedding vector processed by the feedforward network to obtain the final channel feature data;

[0169] 5.1.3. Gate control unit.

[0170] The present invention designs a gating mechanism to learn the weight of each encoder. After obtaining the output of each encoder, a Gated layer with nonlinear activation is included to output the time output vector and channel output vector Then, we concatenate them through a linear mapping to get the weighted vector After normalization, the step encoder weights are calculated and the channel encoder weights , and then the weights are added to the corresponding encoder output to calculate the final fault classification result.

[0171] The specific calculation expression is as follows:

[0172]

[0173] In the formula, represents the weight vector; represents the splicing coefficient; Represents a splicing operation; represents the time output vector; represents the channel output vector; Indicates the offset; represents the step encoder weight; represents the channel encoder weight; Represents a normalization operation; Indicates the fault classification result.

[0174] Specifically, the gate control unit includes:

[0175] The Gated layer is used to perform nonlinear activation operations based on the time feature data and channel feature data to obtain the corresponding time output vector and channel output vector ;

[0176] Linear layer, used to output vectors according to time and channel output vector , perform weighted concatenation and linear mapping to obtain the weighted vector ;

[0177] Softmax layer, used to calculate the weighted vector , perform normalization operation to obtain the step encoder weight and the channel encoder weights ; Output vector according to time , channel output vector , step encoder weight and the channel encoder weights , perform dual-channel weighted feature fusion to obtain the corresponding fault classification results.

[0178] 5.2. Training of fault diagnosis model.

[0179] like Figure 2 As shown, the training method of the fault diagnosis model is as follows:

[0180] 5.2.1. Obtain original fault data of mechanical equipment; wherein the original fault data includes multiple position channel signals, and the original fault data carries a fault label.

[0181] This example uses the bearing dataset in the university mechanical equipment test bench dataset for in-depth analysis. The dataset contains five different working conditions: rotor fault, inner ring fault, outer ring fault, compound fault, and normal state. In order to maintain the consistency of the research, we focus on the bearing dataset, which consists of eight channels of signals, namely motor vibration signals, vibration signals in three directions, motor torque signals, and reducer vibration signals in three directions. The fault types are shown in the following table;

[0182]

[0183] The fault characteristics of mechanical equipment can be represented by m position channels. The data (size is N) is modeled as m single-position channel data (size N*1), and then the m single-position channel data are merged to obtain multi-position channel fusion data (size N*m). The mechanical equipment test bench data of a certain university provided here contains 8 position channel signals. The signals are spliced ​​to obtain 8-position channel fusion data to constitute the original fault data.

[0184] 5.2.2. For any position channel signal, the number of decompositions and the penalty factor are determined based on the iterative optimization method to perform adaptive decomposition and obtain multiple intrinsic mode components.

[0185] The principle is the same as step 2 and will not be repeated here.

[0186] 5.2.3. Based on the preset signal evaluation index, multiple intrinsic mode components are screened and reconstructed to obtain the reconstructed signal.

[0187] The principle is the same as step 3 and will not be repeated here.

[0188] 5.2.4. Based on the reconstructed signals corresponding to the signals of multiple position channels, multi-source data fusion and sliding window sampling are performed to obtain a training set.

[0189] The principle of multi-source data fusion is the same as step 4 and will not be repeated here.

[0190] The principle of sliding window sampling is as follows:

[0191] The speed change of mechanical equipment can be quantified by the number of revolutions per minute (RPM), and the number of revolutions per second (RPS) can be derived from this to determine the rotation period of the mechanical equipment. The relationship between the sampling frequency and the rotation period can be used to determine the number of sampling points captured within a rotation period. Considering that the generation of vibration signals of mechanical equipment is directly related to rotation, the vibration signals of abnormal bearings often show periodic fluctuations that match the motor rotation period. Based on this assumption, the length of each sample is set to the number of sampling points covering one or more complete rotation periods, and the length L of each sample is defined to capture the periodic characteristics of the vibration signal.

[0192]

[0193] Where RPM is the number of revolutions per minute of the mechanical equipment, F is the sampling frequency of the mechanical equipment of the test bench, T is the rotation period, is a random positive real number.

[0194] Given that the total number of sampling points in the test bench dataset is limited and fixed, in order to alleviate the problem of insufficient training samples, data enhancement techniques are usually used in fault classification tasks, among which sliding window technology is a commonly used strategy. Therefore, from the original vibration signal, samples can be sampled according to the set sample length L and sliding step length S. In this study, the sample length is set to the number of sampling points collected by the bearing in one or more rotation cycles, and finally a sample set is formed. Specifically, the sample length L is set to 500 and the sliding step length S is 250. In this way, the number of training samples can be effectively increased, and the generalization ability and classification performance of the model can be improved.

[0195] sample _ set = { x e , y g }, i ∈ { 1 , N } x e = { x 1 e , x 2 e , … } x g e = [ F e start , … , F e start + L ] x g + 1 e = [ F e start + S , … , F e start + L + S ]

[0196] Among them, start represents the starting point position, L represents the sampling length, and S represents the sliding step length;

[0197] represents the training sample set; represents the eigenvalue of the e-th sample; Represents the label value of the e-th sample; Indicates the sliding window range where the e-th sample is located; Indicates the data point where the e-th sample starts to be collected.

[0198] According to the optimal number of decompositions and the optimal penalty factor determined in step 5.2.2, the selected IMF components are used for signal reconstruction, as shown in the following table.

[0199]

[0200] Next, the 8-dimensional signals of each fault type are fused, and the periodic sampling method is used to determine the number of each sample point to be 500 and the sliding step size to be 250. The reconstructed signal is divided into training set, validation set and test set in the ratio of 0.75:0.15:0.15, and their dimensions are (5000, 500, 8), (1000, 500, 8), and (1000, 500, 8), respectively.

[0201] 5.2.5. According to the training set, the pre-built fault diagnosis model is trained to obtain a trained fault diagnosis model.

[0202] Set the fault classification model training parameters as follows:

[0203] (1) d-model, the dimension of the feature vector in the model, is usually used to represent the size of the hidden layer in the model. In this embodiment, d-model is set to 512.

[0204] (2) d_hidden: refers to the size of the hidden layer, which is usually used for fully connected layers or certain types of neural network layers. In this embodiment, d_hidden is set to 1024.

[0205] (3) d_input: The feature dimension of the input data, that is, the number of neurons in the input layer. In this embodiment, d_input is set to 500.

[0206] (4) d_channel: The number of filters in a convolutional layer. In this embodiment, d_channel is set to 8, which means that each convolutional layer will have 8 filters.

[0207] (5) d_output: The dimension of the output layer, which usually indicates the number of categories in the classification task. In this example, d_output is set to 5.

[0208] (6) q and v: In the attention mechanism, q usually represents the dimension of the query, and v represents the dimension of the value. In this embodiment, q and v are set to 8.

[0209] (7) Dropout: A regularization technique used to prevent overfitting, which means that in each training batch, the output of each neuron has a 20% probability of being set to zero. In this example, dropout is set to 0.2.

[0210] (8) EPOCH: A cycle in the training process, which indicates the number of times the entire data set is completely traversed and trained by the model. In this embodiment, EPOCH is set to 100, that is, when the number of training times reaches 100, the training is terminated.

[0211] (9) BATCH_SIZE: The number of samples used in one gradient descent update. In this example, BATCH_SIZE is set to 3.

[0212] (10) LR: learning rate, used to control the step size of parameter update in the gradient descent algorithm. In this embodiment, LR is set to 1e-4.

[0213] 5.2.6. Model testing.

[0214] The trained fault diagnosis model obtained in step 5.2.5 was tested on the test set obtained in step 5.2.4, and finally achieved an average accuracy of 99.1% on the test set.

[0215] It can be seen that the mechanical equipment fault diagnosis method and device based on signal self-decomposition and deep learning proposed in the present invention achieves signal noise reduction by adaptively decomposing the mechanical equipment sensor data and using customized evaluation indicators to screen modal components rich in more characteristic information. At the same time, a fault diagnosis model is constructed to process time and channel information, which well achieves the purpose of distinguishing the fault type and greatly improves the accuracy and efficiency of fault diagnosis.

[0216] 5.3. Use of fault diagnosis model.

[0217] Input the fused fault data in step 4 into the fault diagnosis model trained in step 5.2 and output the corresponding fault classification results.

[0218] Example 2

[0219] This embodiment 2 adopts the mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning of embodiment 1, and uses the bearing data set in the university mechanical equipment test bench data set for specific testing.

[0220] Step 1: Acquire actual fault data of the mechanical equipment to be diagnosed; wherein the actual fault data includes multiple position channel signals.

[0221] Step 2: For any position channel signal, the number of decompositions and the penalty factor are determined based on the improved VMD algorithm to perform adaptive decomposition to obtain multiple intrinsic mode components, such as Figure 4 As shown, including IMF1, IMF2, IMF3, IMF4, IMF5, IMF6, IMF7, and IMF8.

[0222] Step 3: Based on the preset signal evaluation index, multiple intrinsic mode components are screened and reconstructed to obtain a reconstructed signal.

[0223] The specific calculation data are shown in the following table:

[0224]

[0225] Therefore, IMF1, IMF2, IMF3, IMF4, and IMF5 are taken as modal components to be reconstructed and are spliced ​​to obtain the reconstructed signal.

[0226] Step 4: Perform multi-source data fusion according to the reconstructed signals corresponding to the signals of multiple position channels to obtain fused fault data;

[0227] Step 5: Input the fused fault data into the trained fault diagnosis model and output the corresponding fault classification results.

[0228] like Figure 5 As shown in the figure, for the 1,000 test fault data we input, there are 5 fault types, and each fault type contains 200 test data. Finally, after the test output of the fault diagnosis model, the number of correct fault classifications of each type is basically maintained at 198, with a few deviations, and the average classification accuracy reaches 99%.

[0229] In summary, the present invention proposes a mechanical equipment fault diagnosis method and device based on signal self-decomposition and deep learning, which achieves signal noise reduction by adaptively decomposing the sensor data of the mechanical equipment and using customized evaluation indicators to screen modal components rich in more characteristic information. At the same time, a fault diagnosis model is constructed to process time steps and channel information, which effectively achieves the purpose of distinguishing the fault type and greatly improves the accuracy and efficiency of fault diagnosis.

[0230] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0231] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0232] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0233] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0234] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning, characterized in that: The steps include: Acquire actual fault data of the mechanical equipment to be diagnosed; wherein the actual fault data includes a plurality of position channel signals; For any position channel signal, the optimal decomposition number and the optimal penalty factor are determined based on the improved VMD algorithm for adaptive decomposition to obtain multiple intrinsic mode components; Based on the preset signal evaluation index, multiple intrinsic mode components are screened and reconstructed to obtain a reconstructed signal; According to the reconstructed signals corresponding to the signals of multiple position channels, multi-source data fusion is performed to obtain fused fault data; Inputting the fused fault data into a trained fault diagnosis model and outputting a corresponding fault classification result; Wherein, the fault diagnosis model includes: A step encoder is used to perform temporal feature encoding based on a sparse self-attention mechanism and a feedforward mechanism according to the fused fault data to obtain temporal feature data; A channel encoder is used to perform channel feature encoding based on a sparse self-attention mechanism and a feedforward mechanism according to the fused fault data to obtain channel feature data; A gating unit, used for performing linear mapping, splicing and normalization processing according to the time characteristic data and the channel characteristic data to obtain a corresponding fault classification result; The signal screening and reconstruction of multiple intrinsic mode components based on the preset signal evaluation index includes the following steps: Calculate the kurtosis sparsity value of each eigenmode component; In response to the kurtosis sparsity value of any eigenmode component being greater than a preset kurtosis sparsity threshold, the eigenmode component is used as a modal component to be reconstructed; otherwise, the eigenmode component is not used as a modal component to be reconstructed; According to all the modal components to be reconstructed, splicing processing is performed to obtain the reconstructed signal.

2. The mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning according to claim 1 is characterized in that: Determining the optimal decomposition quantity and the optimal penalty factor for adaptive decomposition based on the improved VMD algorithm comprises the following steps: For any position channel signal, initialize parameters, the parameters including the frequency domain representation of the intrinsic mode component, the frequency domain representation of the center frequency of the intrinsic mode component, the frequency domain value of the penalty term, the number of intrinsic mode component decompositions and the second-order penalty factor; Iterate and loop based on the improved VMD algorithm until a preset termination condition is met, and output an optimal decomposition number and an optimal penalty factor; decompose the corresponding position channel signal according to the optimal decomposition number and the optimal penalty factor to obtain multiple intrinsic mode components; Each cycle includes the following steps: Calculate the signal decomposition quality of the eigenmode components in the current iteration; In response to the signal decomposition quality of the intrinsic mode component in the current iteration being less than a preset signal decomposition quality threshold, and the current frequency variable being greater than a preset frequency variable threshold, the parameters are updated and the calculation step of the signal decomposition quality is repeated to perform the next iteration loop; otherwise, the loop is terminated and the eigenmode component decomposition quantity and the second-order penalty factor of the current iteration are taken as the optimal decomposition quantity and the optimal penalty factor.

3. The mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning according to claim 2 is characterized in that: The expression of the signal decomposition quality is as follows: Where Q(k) represents the signal decomposition quality of the kth intrinsic mode component; QualityIndex(k) represents the calculation index weight of the kth intrinsic mode component; QualityIndex(k-1) represents the calculation index weight of the kth intrinsic mode component. When k=1, QualityIndex(0) takes the value of 0. q1 represents the first coefficient; kurtosis(k) represents the kurtosis value of the kth eigenmode component; q2 represents the second coefficient; correlation(k) represents the correlation coefficient of the kth eigenmode component; q3 represents the third coefficient; entropy(k) represents the information entropy of the kth eigenmode component.

4. The mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning according to claim 3 is characterized in that: The updating parameters comprises the following steps: Update the decomposition quantity of the eigenmode component, the expression is as follows: K n+1 ←K n +1; In the formula, K n+1 K represents the number of eigenmode component decompositions at the n+1th iteration; n Indicates the number of eigenmode component decompositions at the nth iteration; Update the second-order penalty factor, the expression is as follows: α n+1 ←α n +50； In the formula, α n+1 represents the second-order penalty factor at the n+1th iteration; α n Represents the second-order penalty factor at the nth iteration; Update the frequency domain representation of the eigenmode component, the expression is as follows: In the formula, represents the frequency domain representation of the kth intrinsic mode component obtained at the n+1th iteration, w represents the current frequency variable; f(w) represents the frequency domain representation of the position channel signal f(t); represents the sum of the frequency domain representations of all eigenmode components except the kth eigenmode component; represents the frequency domain value of the penalty term; α represents the second-order penalty factor; w k represents the center frequency of the kth eigenmode component; Update the center frequency of the eigenmode component, the expression is as follows: In the formula, represents the center frequency of the kth intrinsic mode component at the n+1th iteration; w represents the current frequency variable; dw represents the differential of the current frequency variable w; Update the penalty term frequency domain value The expression is as follows: In the formula, Represents the frequency domain value of the penalty term at the n+1th iteration; represents the frequency domain value of the penalty term at the nth iteration; τ represents the noise tolerance parameter; f(w) represents the frequency domain representation of the position channel signal f(t); represents the sum of the frequency domain representations of the eigenmode components obtained at the n+1th iteration.

5. The mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning according to claim 1 is characterized in that: The expression of the kurtosis sparse value is as follows: In the formula, u k (t) represents the kth eigenmode component; K represents the number of eigenmode components; S(k) represents the sparsity of the kth eigenmode component; C(k) represents the kurtosis value of the kth eigenmode component; W(k) represents the dot product of the sparsity of the kth eigenmode component and the kurtosis value; I(k) represents the kurtosis sparsity value of the kth eigenmode component.

6. The mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning according to claim 1 is characterized in that: The training method of the fault diagnosis model is as follows: Acquire original fault data of mechanical equipment; wherein the original fault data includes multiple position channel signals, and the original fault data carries a fault label; For any position channel signal, the number of decompositions and penalty factors are determined based on the iterative optimization method to perform adaptive decomposition and obtain multiple intrinsic mode components; Based on the preset signal evaluation index, multiple intrinsic mode components are screened and reconstructed to obtain a reconstructed signal; According to the reconstructed signals corresponding to the signals of multiple position channels, multi-source data fusion and sliding window sampling are performed to obtain a training set; The pre-built fault diagnosis model is trained according to the training set to obtain a trained fault diagnosis model.

7. The mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning according to claim 1 is characterized in that: The step size encoder comprises: A first embedding layer, used for performing vectorization processing according to the fused fault data to obtain a first embedding vector; A first sparse self-attention layer, configured to perform sparse self-attention calculation according to the first embedding vector to obtain a weighted first embedding vector; A first residual regularization layer, used for performing residual connection and regularization processing according to the weighted first embedding vector to obtain a regularized first embedding vector; A first feedforward network, used to perform a nonlinear transformation of a feedforward neural network according to the regularized first embedding vector to obtain a first embedding vector processed by the feedforward network; The second residual regularization layer is used to perform residual connection and regularization processing according to the first embedding vector processed by the feedforward network to obtain final time feature data.

8. The mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning according to claim 1 is characterized in that: The channel encoder comprises: A second embedding layer, used for performing vectorization processing according to the fused fault data to obtain a second embedding vector; A second sparse self-attention layer is used to perform sparse self-attention calculation according to the second embedding vector of each channel to obtain a weighted second embedding vector; A third residual regularization layer, used for performing residual connection and regularization processing according to the second embedding vector calculated by the self-attention, so as to obtain a second embedding vector after residual regularization; A second feedforward network is used to perform a nonlinear transformation of a feedforward neural network according to the second embedding vector data after residual regularization to obtain a second embedding vector processed by the feedforward network; The fourth residual regularization layer is used to perform residual connection and regularization processing according to the second embedding vector processed by the feedforward network to obtain final channel feature data.

9. The mechanical equipment fault diagnosis method based on signal self-decomposition and deep learning according to claim 1 is characterized in that: The gate control unit comprises: A Gated layer is used to perform a nonlinear activation operation according to the time feature data and the channel feature data to obtain a corresponding time output vector and a channel output vector; A Linear layer is used to perform a weighted concatenation operation and then a linear mapping according to the time output vector and the channel output vector to obtain a weighted vector; The Softmax layer is used to perform a normalization operation according to the weighted vector to obtain a step encoder weight and a channel encoder weight; and to perform weighted feature fusion of two channels according to the time output vector, the channel output vector, the step encoder weight and the channel encoder weight to obtain a corresponding fault classification result.

Images