A fault diagnosis method based on intelligent oversampling of linear modulation of timing characteristics

Abstract

The present invention discloses a fault diagnosis method based on intelligent oversampling of linear modulation of time series characteristics, which improves the signal sampling rate, supplements the high-frequency band details, and enhances the fault diagnosis performance. The steps include: obtaining the original vibration signal. Preprocessing the original vibration signal, and obtaining a low sampling rate signal by downsampling. Pairing the low sampling rate signal and the original vibration signal for intelligent oversampling network training. Constructing an oversampling rate model to achieve the predicted output of the oversampling rate signal. Constructing a fault diagnosis model to learn the characteristics of signal fault classification. Inputting the training set into the oversampling rate fault diagnosis model to train the model. Inputting the vibration signal into the trained model to obtain an oversampling rate signal, classifying the oversampling rate signal, and obtaining its health status type. The fault diagnosis model of the present invention better solves the problem of poor diagnostic performance when facing low sampling rate signals, and at the same time more realistically restores the high-frequency information of the signal.

Description

Technical Field

[0001] The invention belongs to the technical field of bearing fault diagnosis, and in particular relates to a fault diagnosis method based on time series characteristic linear modulation intelligent oversampling. Background Art

[0002] In industrial production, mechanical fault diagnosis plays a vital role in ensuring industrial safety and promoting industrial development. With the continuous advancement of artificial intelligence technology, intelligent fault diagnosis methods based on deep learning and other technologies have become a hot topic in the industry. Compared with traditional fault diagnosis methods, intelligent fault diagnosis has a higher level of automation and diagnostic accuracy, which can effectively improve the safety and efficiency of industrial production.

[0003] In industrial production, real vibration data is usually acquired through monitoring by instruments such as vibration sensors, which means that the parameters of the sensor have an important impact on the quality of the collected vibration data. Generally speaking, the parameters of the sensor determine the sampling frequency of the vibration signal. However, due to the limitations of sampling frequency and quantization accuracy, low sampling rate signals cannot fully capture the details and high-frequency components of the signal. This will cause some valuable information to be lost, affecting subsequent analysis and application. Due to information loss, the accuracy and reliability of the results of analysis and processing based on low sampling rate signals, such as spectrum analysis and pattern recognition, will be affected. This may lead to incorrect diagnosis and prediction.

[0004] The characteristics of vibration signals mainly include information such as signal amplitude, frequency and phase. Compared with low sampling rate signals, high sampling rate signals have more complete information, higher analysis accuracy, stronger anti-interference ability and higher data utilization. For example, the part with larger amplitude has higher corresponding signal strength, indicating that the characteristics are more obvious. The part with richer frequency components indicates that the signal characteristics are more complex. These characteristics will be very obvious in high sampling rate vibration signals.

[0005] On the other hand, with the in-depth research of data-driven methods in the field of fault diagnosis, the demand for real vibration data has greatly increased, which means that when transferring vibration data, the load on the equipment will also increase year-on-year. At this time, if the vibration signal is downsampled and the low sampling rate signal is saved, it can be restored to a high sampling rate signal through oversampling technology after the signal data transfer is completed. This can reduce the load on the equipment without losing signal information.

[0006] Traditional oversampling methods mainly include interpolation-based methods and subspace-based methods. Traditional oversampling methods can improve the quality of low-sampling rate signals to a certain extent, but they also have some limitations. For example, for complex vibration signals, traditional methods often find it difficult to effectively recover the details of high-sampling rate signals, so the effect may not be satisfactory in practical applications. With the development of technologies such as machine learning and deep learning, traditional oversampling methods have gradually been challenged. Using deep learning technology, we can better capture the subtle changes and contextual structure information in vibration signals, so as to more accurately recover the details and features of high-sampling rate signals. At the same time, deep learning methods can effectively integrate multi-modal and multi-scene data features, and have more comprehensive generalization, whether it is a real industrial production scene or a laboratory scene. However, there are still some problems with existing methods. Some oversampling models may introduce artifacts or distortion when generating high-resolution signals, especially in complex textures or detail areas; at the same time, the noise in the signal is also a factor that affects the quality of the final generated signal; secondly, the overfitting problem has always been a challenge faced by deep learning. Summary of the invention

[0007] Purpose of the invention: In view of the above problems, the present invention proposes a fault diagnosis method based on intelligent oversampling of linear modulation of timing characteristics, which improves the diagnosis efficiency by restoring high-sampling rate signals to supplement high-frequency band information, solves the problem of low-sampling rate signals losing high-frequency band information, and improves the details and quality of vibration signals.

[0008] Technical solution: To achieve the purpose of the present invention, the technical solution adopted by the present invention is: a fault diagnosis method based on time series characteristic linear modulation intelligent oversampling, comprising the following steps:

[0009] Step 1: Get the original vibration signal, divide the original signal into samples to get a high sampling rate signal, then downsample the original signal, perform bicubic interpolation and padding on the downsampled signal to make its sample length equal to that of the high sampling rate signal, and get a low sampling rate signal;

[0010] Step 2: Pair the high sampling rate signal and the low sampling rate signal so that they correspond one to one in the time series, and then combine the high sampling rate signal and the low sampling rate signal to construct a training set, where the low sampling rate signal is the input data and the high sampling rate signal is the label data;

[0011] Step 3: Build an oversampling rate network, including downsampling layer, bottleneck layer, upsampling layer and output layer; combine the oversampling rate network and long short-term memory network to build a fault diagnosis model;

[0012] Step 4: Training the fault diagnosis model. The process includes:

[0013] First, the training set is input into the oversampling rate network for training, and the network is adjusted according to the loss and evaluation indicators;

[0014] Input the low sampling rate signal that needs to be oversampled into the oversampling rate network to obtain the predicted oversampling rate signal; then input the oversampling rate signal into the fault diagnosis model to train the model, adjust the model parameters according to the loss value, and obtain the trained fault diagnosis model;

[0015] Step 5: Obtain a low sampling rate vibration signal through equipment collection, then input it into the oversampling rate network to obtain a high sampling rate signal, and then input it into the fault diagnosis network to output its health status type.

[0016] Furthermore, the process of building the downsampling layer and bottleneck layer of the supersampling rate network is as follows:

[0017] The downsampling layer contains Downsampling Block , each downsampling block contains The length is And a convolution filter with a stride of 2;

[0018] In each downsampling block, the interpolated low-resolution signal is max-pooled and activated after each convolution filter to obtain a one-dimensional convolution activation tensor, and then the signal is linearly modulated by the temporal features;

[0019] The timing characteristic linear modulation process is as follows:

[0020] Input 1D convolution activation tensor , Represents the time dimension, Indicates the number of channels;

[0021] (1) The tensor Reshape into block tensor , defined as:

[0022] ,

[0023] in Indicates the block length, Indicates block tensors, Indicates The time dimension, Indicates channels;

[0024] (2) By pooling the channels within each block, we obtain the pooled representation of the block tensor , defined as:

[0025] ,

[0026] in Refers to operating on block tensors at all time steps;

[0027] (3) Calculate the normalizer sequence , and apply it to the pooled block tensor using an RNN:

[0028] ,

[0029] in, is the scaling parameter, is the translation parameter, The time steps The hidden state of Refers to the pooling block tensor on all channels. The entire cycle of the RNN network starts with ;

[0030] (4) Calculate the normalized block tensor , the formula is:

[0031] ,

[0032] in, and Respectively Scaling and translation parameters on each channel;

[0033] (5) Normalize the block tensor Reshape to get the adaptive normalized activation tensor , the formula is:

[0034] ,

[0035] in, That is the adaptive normalized activation tensor , yes The time step index of Indicates channels; Calculate the current time step The batch index, Calculate the current time step Relative position in the batch;

[0036] (6) Output adaptive normalized activation tensor ;

[0037] After filtering through the downsampling layer, the signal enters the bottleneck layer. After convolution and pooling operations, the Dropout operation is added. Finally, after activation, the signal is linearly modulated using temporal features again.

[0038] Furthermore, the process of constructing the upsampling layer on the supersampling rate network is as follows:

[0039] The upsampling layer reconstructs the signal. The upsampling layer contains upsampling blocks, each upsampling block Include The length is Filters;

[0040] In the upsampling layer, the signal time dimension is doubled and the filter size is reduced by the operation; after the convolution tensor in the upsampling block is activated, the signal is sub-pixel shuffled; the sub-pixel shuffling method is as follows:

[0041] For the input tensor of the upsampling block, the shape is , is the batch size, For the The sample length of the upsampling block input, For the The number of channels of the upsampling block input;

[0042] After the upsampling block convolution activation, the output is a shape of The sub-pixel layer reorders this tensor, and the shape after sorting is ; The final output tensor shape is ;

[0043] The next step is to perform linear modulation of the time series characteristics on the sub-pixel shuffled signal. The process is as follows:

[0044] Input sub-pixel shuffled 1D convolution activation tensor , Represents the time dimension, Indicates the number of channels;

[0045] (1) The tensor Reshape into block tensor , defined as:

[0046] ,

[0047] in Indicates the block length, Indicates block tensors, Indicates The time dimension, Indicates channels;

[0048] (2) By pooling the channels within each block, we obtain the pooled representation of the block tensor , defined as:

[0049] ,

[0050] in Refers to operating on block tensors at all time steps;

[0051] (3) Calculate the normalizer sequence , and apply it to the pooled block tensor using an RNN:

[0052] ,

[0053] in, is the scaling parameter, is the translation parameter, The time steps The hidden state of Refers to the pooling block tensor on all channels. The entire cycle of the RNN network starts with ;

[0054] (4) Calculate the normalized block tensor , the formula is:

[0055] ,

[0056] in, and Respectively Scaling and translation parameters on each channel;

[0057] (5) Normalize the block tensor Reshape to get the adaptive normalized activation tensor , the formula is:

[0058] ,

[0059] in, That is the adaptive normalized activation tensor , yes The time step index of Indicates channels; Calculate the current time step The batch index, Calculate the current time step Relative position in the batch;

[0060] (6) Output adaptive normalized activation tensor ;

[0061] After the linear modulation operation of the timing feature, a skip connection is added to The feature tensor of the downsampling block is The feature tensors of the upsampling blocks are stacked together to obtain the final upsampling layer output result.

[0062] Furthermore, the last output layer of the supersampling network performs a convolution on the signal input to this layer, performs sub-pixel shuffling, and then adds an additional residual connection from the input to the final output.

[0063] Furthermore, the training process of the supersampling rate network uses the mean square error loss to converge, and the formula is:

[0064] ,

[0065] Where MSE is the mean square error, which represents the average square difference between the predicted value and the true value; represents the number of samples, represents the true value, represents the predicted value;

[0066] The logarithmic spectral distance indicator is used to evaluate the reconstruction quality of the oversampled signal; the logarithmic spectral distance is measured for a single frequency, and the formula is as follows:

[0067] ,

[0068] in is the log spectral distance, represents the index of the current timeframe, Indicates the number of time frames; Indicates the index of the current frequency point, Indicates the number of frequency points; represents the original signal, represents the reconstruction signal, and They are and The logarithmic spectral power amplitude of is defined as:

[0069] ,

[0070] In the formula Represents the short-time Fourier transform of a signal.

[0071] Furthermore, the fault diagnosis model first uses one-hot encoding to encode the oversampling rate signal output by the oversampling rate network. The encoded signal passes through the long short-term memory network layer, convolution layer, pooling layer and fully connected layer to output the diagnosis result; the model training uses the cross entropy loss function, the formula is as follows:

[0072] ,

[0073] in is the cross entropy loss, represents the number of samples, represents the sample index, Represents the category index, represents the number of categories, is the true label, indicating that samples, Class label, if samples belong to class ,but ,otherwise , Represents the model for samples belong to class The predicted probability of .

[0074] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0075] The method of the present invention uses oversampling technology, and the vibration signal has more structure and detail information. When combined with fault diagnosis, the diagnostic model can obtain more comprehensive and accurate fault feature information. The features learned by the model are closer to the original signal value and have more semantic information. The downsampling layer of the oversampling model halves the spatial dimension and doubles the feature size, while the upsampling layer does the opposite. Considering that the vibration signal is time series data, linear modulation of the time series features is added to the model. It can model the time series features, thereby capturing the time dependency in the input sequence, enhancing the expressiveness of the features, and enabling the subsequent upsampling layer to better reconstruct the high sampling rate output.

[0076] The present invention adds a bottleneck layer between the downsampling layer and the upsampling layer. This bottleneck structure is similar to a convolutional autoencoder, which encourages the model to learn the hierarchical structure of features. In the final output, the present invention adds a jump connection to stack the downsampled feature tensor with the upsampled feature tensor, and also adds an additive residual connection from the input to the final output, which will increase the training speed. During the entire mapping process, the model can better preserve the temporal structure and correlation of the signal, and after the mapping is completed, it can also make corresponding adjustments when restoring the details. This helps to generate more realistic and natural high-sampling rate vibration signals, thereby more accurately analyzing and interpreting the features in the vibration signals.

[0077] The application of this invention is expected to provide more reliable data support and less-loaded data transfer for equipment fault diagnosis. It has great potential and application prospects in the field of intelligent fault detection. It is expected to provide more intelligent, efficient and reliable fault detection solutions for industrial production, bringing significant economic and benefit improvements to industrial production, and taking an important step in promoting the development of industrial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a flow chart of the fault diagnosis model based on intelligent super sampling of linear modulation of timing characteristics of the present invention.

[0079] Figure 2 It is the oversampling rate model diagram. DETAILED DESCRIPTION

[0080] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0081] The present invention discloses a fault diagnosis method based on time series characteristic linear modulation intelligent oversampling, the process is as follows: Figure 1 As shown, the specific implementation includes the following steps:

[0082] Step 1: Get the original vibration signal. Time domain vibration signal As input, It is a signal of a working condition of a type of fault, containing data points; is the total number of categories of different working conditions under different faults; Fault type label for each class of faults.

[0083] Step 2: Training set construction. First, divide the original signal into samples to obtain a high sampling rate signal, then downsample the original signal, perform bicubic interpolation and padding on the downsampled signal to make its sample length equal to that of the high sampling rate signal, thus obtaining a low sampling rate signal. The bicubic interpolation formula is as follows.

[0084] ,

[0085] in, is the function value at the position to be interpolated, is the function value at a known point, is the interpolation basis function, is the position coordinate that needs to be interpolated, are the coordinates of the known points.

[0086] The high sampling rate signal and the low sampling rate signal are paired so that they correspond one to one in the time series, and then the high sampling rate signal and the low sampling rate signal are combined to construct a training set, in which the low sampling rate signal is the input data and the high sampling rate signal is the label data.

[0087] Step 3: Build the downsampling layer of the supersampling rate network.

[0088] The supersampling model consists of a downsampling layer, a bottleneck layer, an upsampling layer, and an output layer, such as Figure 2 The downsampling layer contains Downsampling Block , each downsampling block contains The length is And the convolution filter with a step size of 2. After passing through each convolution filter, the signal will be max-pooled and activated to obtain a one-dimensional convolution activation tensor, and then the signal will be linearly modulated by the temporal features.

[0089] The timing characteristic linear modulation process is as follows:

[0090] Input 1D convolution activation tensor , Represents the time dimension, Indicates the number of channels.

[0091] (1) The tensor Reshape into block tensor , defined as:

[0092] ,

[0093] in Indicates the block length, Indicates block tensors, Indicates The time dimension, Indicates channels.

[0094] (2) By pooling the channels within each block, we obtain the pooled representation of the block tensor , defined as:

[0095] ,

[0096] in Refers to operating on block tensors over all time steps.

[0097] (3) Calculate the normalizer sequence , and apply it to the pooled block tensor using an RNN:

[0098] ,

[0099] in, is the scaling parameter, is the translation parameter, The time steps The hidden state of Refers to the pooling block tensor on all channels. The entire cycle of the RNN network starts with .

[0100] (4) Calculate the normalized block tensor , the formula is:

[0101] ,

[0102] in, and Respectively Scaling and translation parameters on each channel.

[0103] (5) Normalize the block tensor Reshape to get the adaptive normalized activation tensor , the formula is:

[0104] ,

[0105] in, That is the adaptive normalized activation tensor , yes The time step index of Indicates channels; Calculate the current time step The batch index, Calculate the current time step The relative position in the batch.

[0106] (6) Output adaptive normalized activation tensor .

[0107] Step 4: Build the oversampling rate network bottleneck layer.

[0108] After filtering in the downsampling layer, the signal enters the bottleneck layer. After convolution and pooling operations, the Dropout operation is added to effectively extract the key features of the input data and provide a more effective feature representation for the subsequent upsampling layer. At the same time, the addition of Dropout can help the model learn more robust and effective feature representations to avoid overfitting. Finally, after activation, the signal is linearly modulated using the time series features again.

[0109] Step 5: Build the sampling layer on the supersampling rate network.

[0110] The upsampling layer reconstructs the signal. The upsampling layer is similar to the downsampling layer. The upsampling layer contains upsampling blocks, each upsampling block Include The length is filter.

[0111] In the upsampling layer, the signal temporal dimension is doubled and the filter size is reduced by the operation; after the activation of the convolution tensor in the upsampling block, the signal is sub-pixel shuffled.

[0112] In order to increase the signal time dimension during upsampling reconstruction, a one-dimensional version of sub-pixel shuffling is introduced in the upsampling layer, which has been shown to be effective in overcoming artifacts. For the input tensor of the upsampling block, the shape is , is the batch size, For the The sample length of the upsampling block input, For the The number of channels of the upsampling block input; after the upsampling block convolution activation, the output shape is The sub-pixel layer reorders this tensor, and the shape after sorting is ; The final output tensor shape is .

[0113] The next step is to perform linear modulation of the temporal features on the sub-pixel shuffled signal, and the process is the same as the upsampling layer. Considering that the input data has a low sampling rate signal and target data high sampling rate signal The sampling rates are different, but the overall trends are similar. It can be said that the features extracted from the downsampling layer are also applicable to the upsampling layer. Therefore, a skip connection is added after the modulation operation. The main process is to convert the first The tensor of the downsampled features and the The tensors of upsampled features are stacked together.

[0114] The final output layer performs a simple convolution on the signal followed by sub-pixel shuffling, and then adds an additional residual connection from the input to the final output, so the model only needs to learn , which accelerates the convergence of the model and improves the training effect.

[0115] Step 6: Build the fault diagnosis network model.

[0116] In order to further improve the performance of the oversampling rate model in fault diagnosis, considering the characteristics of the vibration signal time series, a long short-term memory network layer is chosen to be added when building the fault diagnosis network.

[0117] The fault diagnosis network model first uses one-hot encoding to encode the oversampling rate signal. The diagnosis network consists of a long short-term memory network layer, a convolutional layer, a pooling layer, and a fully connected layer. The loss function uses the cross entropy loss function, and the formula is as follows:

[0118] ,

[0119] in is the cross entropy loss, represents the number of samples, represents the sample index, Represents the category index, represents the number of categories, is the true label, indicating that samples, Class label, if samples belong to class ,but ,otherwise , Represents the model for samples belong to class The predicted probability of .

[0120] Step 7: Train the oversampling rate fault diagnosis network model.

[0121] First, the training set is input into the oversampling rate model to train the model, and the model is adjusted according to the loss and evaluation indicators. The training process of the oversampling rate model uses the mean square error loss for convergence, and the formula is:

[0122] ,

[0123] Where MSE is the mean square error, which represents the average square difference between the predicted value and the true value; represents the number of samples, represents the true value, Represents the predicted value.

[0124] In order to evaluate the performance of the oversampling model, the log spectral distance is used to evaluate the quality of the oversampling signal reconstruction. The log spectral distance is measured for a single frequency and the formula is as follows:

[0125] ,

[0126] in is the log spectral distance, represents the index of the current timeframe, Indicates the number of time frames; Indicates the index of the current frequency point, Indicates the number of frequency points; represents the original signal, represents the reconstruction signal, and They are and The logarithmic spectral power amplitude of is defined as:

[0127] ,

[0128] In the formula Represents the short-time Fourier transform of the signal, and its formula is:

[0129] ,

[0130] The low sampling rate signal that needs to be oversampled is input into the oversampling rate model to obtain the predicted oversampling rate signal. The oversampling rate signal is then input into the fault diagnosis network model to train the model, and the model is adjusted according to the loss value to guide the model convergence.

[0131] Step 8: Troubleshooting.

[0132] The low sampling rate vibration signal collected by the device is then input into the oversampling rate network to obtain the high sampling rate signal, which is then input into the fault diagnosis network to output its health status type.

[0133] The oversampling rate fault diagnosis model of the present invention improves the sampling rate of the signal, and improves the fault diagnosis performance through the generated high sampling rate signal. An intelligent method for improving the sampling rate of the vibration signal is designed to effectively reduce the demand for high sampling rate in the actual vibration signal acquisition, which can better reduce the cost of vibration signal acquisition and storage, and improve the low-cost application prospect of the fault diagnosis system. The oversampling rate fault diagnosis model learns the characteristics of the high sampling rate signal by using a bottleneck network structure that downsamples first and then upsamples. This structure prompts the model to learn the hierarchical structure of the features. The oversampling rate fault diagnosis model modulates the activation of the convolution model by using linear modulation of the time series features, effectively captures the long-range input dependency in the input sequence, and learns the overall trend of the signal. The oversampling rate fault diagnosis model uses jump connections between the upsampling block and the downsampling block, and uses residual connections between the input and output to learn the correlation between the upsampling and downsampling layer features, which accelerates the convergence speed of the model. The oversampling rate fault diagnosis model uses one-dimensional sub-pixel shuffling to achieve high-frequency information recovery while avoiding artifacts.

Claims

1. A fault diagnosis method based on time series characteristic linear modulation intelligent oversampling, comprising the following steps: Step 1: Get the original vibration signal, divide the original signal into samples to get a high sampling rate signal, then downsample the original signal, perform bicubic interpolation and padding on the downsampled signal to make its sample length equal to that of the high sampling rate signal, and get a low sampling rate signal; Step 2: Pair the high sampling rate signal and the low sampling rate signal so that they correspond one to one in the time series, and then combine the high sampling rate signal and the low sampling rate signal to construct a training set, where the low sampling rate signal is the input data and the high sampling rate signal is the label data; Step 3: Build an oversampling rate network, including downsampling layer, bottleneck layer, upsampling layer and output layer; combine the oversampling rate network and long short-term memory network to build a fault diagnosis model; Step 4: Training the fault diagnosis model includes: First, the training set is input into the oversampling rate network for training, and the network is adjusted according to the loss and evaluation indicators; The training process of the super-sampling rate network uses the mean square error loss to converge, the formula is: Where MSE is the mean square error, which represents the average square difference between the predicted value and the true value; n represents the number of samples, y i represents the true value, represents the predicted value; The logarithmic spectral distance indicator is used to evaluate the reconstruction quality of the oversampled signal; the logarithmic spectral distance is measured for a single frequency, and the formula is as follows: Where LSD(x, y) is the logarithmic spectrum distance, t represents the index of the current time frame, T represents the number of time frames; k represents the index of the current frequency point, K represents the number of frequency points; x represents the original signal, y represents the reconstructed signal, X and Y are the logarithmic spectrum power amplitudes of x and y respectively, and X is defined as: X=log|S| 2 , Where S represents the short-time Fourier transform of the signal; Input the low sampling rate signal that needs to be oversampled into the oversampling rate network to obtain the predicted oversampling rate signal; then input the oversampling rate signal into the fault diagnosis model to train the model, adjust the model parameters according to the loss value, and obtain the trained fault diagnosis model; Step 5: Obtain a low sampling rate vibration signal through equipment collection, then input it into the oversampling rate network to obtain a high sampling rate signal, and then input it into the fault diagnosis network to output its health status type.

2. The fault diagnosis method based on time series characteristic linear modulation intelligent oversampling according to claim 1 is characterized in that: The process of building the downsampling layer and bottleneck layer of the supersampling rate network is as follows: The downsampling layer contains K downsampling blocks k=1, 2, ..., K, each downsampling block contains min(2 6+k , 512) with a length of max(2 7-k +1, 9) and a convolution filter with a step size of 2; in each downsampling block, the interpolated low-resolution signal is subjected to maximum pooling and activation after passing through each convolution filter to obtain a one-dimensional convolution activation tensor, and then the signal is subjected to temporal feature linear modulation; The timing characteristic linear modulation process is as follows: Input 1D convolution activation tensor T represents the time dimension, and C represents the number of channels; (1) Reshape the tensor F into a block tensor Defined as: Where B represents the block length, b represents the b-th block tensor, t represents the t-th time dimension, and c represents the c-th channel; (2) By pooling the channels within each block, we obtain the pooled representation of the block tensor Defined as: in Refers to operating on block tensors at all time steps; (3) Calculate the normalizer sequence Use an RNN to apply this to the pooled block tensor: Among them, γ b is the scaling parameter, β b is the translation parameter, h b ,h b-1 are the hidden states at time steps b and b-1 respectively, Refers to the pooling block tensor on all channels. The entire cycle of the RNN network starts with (4) Calculate the normalized block tensor The formula is: Among them, γ b，c and β b，c are the scaling and translation parameters on the cth channel respectively; (5) Normalize the block tensor F norm Reshape to get the adaptive normalized activation tensor The formula is: Among them, F′ l，c That is, it is the adaptive normalized activation tensor F′, l is the time step index of F′, and c represents the cth channel; Calculate the batch index of the current time step t, and t mod B calculates the relative position of the current time step t in the batch; (6) Output adaptive normalized activation tensor After filtering through the downsampling layer, the signal enters the bottleneck layer. After convolution and pooling operations, the Dropout operation is added. Finally, after activation, the signal is linearly modulated using temporal features again.

3. The fault diagnosis method based on time series characteristic linear modulation intelligent oversampling according to claim 1 is characterized in that: The process of building the upsampling layer on the supersampling rate network is as follows: The upsampling layer reconstructs the signal. The upsampling layer contains K upsampling blocks, each upsampling block k=1,2,…,K contains min(2 7+(K-k+1) , 512) with a length of max(2 7-(K-k+1) +1,9) filter; In the upsampling layer, the signal time dimension is doubled and the filter size is reduced by the operation; after the convolution tensor in the upsampling block is activated, the signal is sub-pixel shuffled; the sub-pixel shuffling method is as follows: For the input tensor of the upsampling block, the shape is N×T k ×C k , N is the batch size, T k is the sample length of the kth upsampling block input, C k The number of channels input to the kth upsampling block; After the upsampling block convolution activation, the output is a shape of N×T k ×C k / 2 tensor; the sub-pixel layer reorders this tensor, and the shape after sorting is N×2T k ×C k / 4; the final output tensor shape is N×2T k ×C k / 4 The next step is to perform linear modulation of the time series characteristics on the sub-pixel shuffled signal. The process is as follows: Input sub-pixel shuffled 1D convolution activation tensor T represents the time dimension, and C represents the number of channels; (1) Reshape the tensor F into a block tensor Defined as: Where B represents the block length, b represents the b-th block tensor, t represents the t-th time dimension, and c represents the c-th channel; (2) By pooling the channels within each block, we obtain the pooled representation of the block tensor Defined as: in Refers to operating on block tensors at all time steps; (3) Calculate the normalizer sequence Use an RNN to apply this to the pooled block tensor: Among them, γ b is the scaling parameter, β b is the translation parameter, h b ,h b-1 are the hidden states at time steps b and b-1 respectively, Refers to the pooling block tensor on all channels. The entire cycle of the RNN network starts with (4) Calculate the normalized block tensor The formula is: Among them, γ b，c and β b，c are the scaling and translation parameters on the cth channel respectively; (5) Normalize the block tensor F norm Reshape to get the adaptive normalized activation tensor The formula is: Among them, F′ l，c That is, it is the adaptive normalized activation tensor F′, l is the time step index of F′, and c represents the cth channel; Calculate the batch index of the current time step t, and t mod B calculates the relative position of the current time step t in the batch; (6) Output adaptive normalized activation tensor A skip connection is added after the linear modulation operation of the temporal features, and the feature tensor of the kth downsampling block is stacked with the feature tensor of the (K-k+1)th upsampling block to obtain the final upsampling layer output result.

4. The fault diagnosis method based on time series characteristic linear modulation intelligent oversampling according to claim 1, 2 or 3, characterized in that: The last output layer of the supersampling network performs a convolution on the signal input to the output layer, performs sub-pixel shuffling, and then adds an additional residual connection from the input to the final output.

5. The fault diagnosis method based on time series characteristic linear modulation intelligent oversampling according to claim 1 is characterized in that: The fault diagnosis model first uses one-hot encoding to encode the oversampling rate signal output by the oversampling rate network. The encoded signal passes through the long short-term memory network layer, convolution layer, pooling layer and fully connected layer to output the diagnosis result. The model training uses the cross entropy loss function, and the formula is as follows: Where L is the cross entropy loss, N is the number of samples, i is the sample index, c is the category index, M is the number of categories, and y is the ic is the true label, which means that for the i-th sample, the label of class c, if the i-th sample belongs to class c, then y ic =1, otherwise y ic =0, p ic Represents the model's predicted probability that the i-th sample belongs to class c.

Images