Bearing residual service life prediction method based on improved Attention-GRU model

Abstract

The invention relates to the technical field of bearing service life prediction, in particular to a bearing residual service life prediction method based on an improved Attention-GRU model, and the method comprises the steps: monitoring full-life vibration data of a bearing when the bearing runs until the bearing is damaged, carrying out the preprocessing of the data, and dividing the data into a training set and a test set; time-frequency domain features are extracted from the vibration data and feature fusion is carried out to obtain a fused feature set; and constructing a bearing residual life prediction model. And inputting the fused feature set into a life prediction model for training and prediction to obtain the remaining service life of the bearing. According to the residual life prediction model in the scheme, time-frequency domain information extraction is carried out on bearing data through discrete wavelet down-sampling and Fourier frequency domain feature extraction, and then the receptive field of the bearing data in the channel and space is improved through channel and space attention feature enhancement; and finally, predicting and outputting the health index of the predicted time step through Attention-GRU regression, and combining the health index with the full-life data to obtain the predicted residual life.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing service life prediction, and in particular to a method for predicting the remaining service life of a bearing based on an improved Attention-GRU model. Background Art

[0002] As a key component of rotating machinery, the health status of rolling bearings is directly related to the working status of the entire rotating machinery. Once a bearing fails, it will seriously affect the safety of the mechanical system and cause serious economic losses. However, rolling bearings often work in harsh environments such as aircraft engines, and are prone to failure, which can cause engine failure. Therefore, the application of remaining useful life (RUL) prediction technology in rolling bearings has always been a key research issue.

[0003] The existing life prediction methods mainly include methods based on statistical models, data-driven methods based on condition monitoring data, and hybrid prediction methods. Among them, although the hybrid prediction method combines the first two methods, it fails to extract features from vibration data information at multiple scales. For example, a pure data-driven method based on an encoder-decoder framework is proposed in the paper (A novel deep learning method based on attention mechanism for bearing remaining useful life prediction). The method includes three steps: feature extraction, calculation of HI through RNN, and prediction of azimuth RUL through linear regression. In the first step, the energy values ​​of the five sub-band spectra are extracted from the original signal as features. Then, with the features as input, a sequence of HI values ​​between 0 and 1 is obtained as the output of the prediction neural network. Finally, the last HI value is scaled up by linear regression to calculate the RUL value at the end of the collected signal. The shortcoming of this method is that it does not conduct a deeper exploration of the features contained in the initial vibration data information, and only uses one data set for feature extraction, but does not explore the versatility of the model. Therefore, there is an urgent need for a bearing remaining useful life prediction method that can extract features from vibration data information at different scales. Summary of the invention

[0004] The technical problem solved by the present invention is to provide a bearing remaining service life prediction method based on an improved Attention-GRU model to realize feature extraction of vibration data information at different scales to improve the safety of rolling bearing failures in harsh environments and reduce the probability of engine failure.

[0005] The basic solution provided by the present invention is a method for predicting the remaining service life of a bearing based on an improved Attention-GRU model, comprising the following steps: S1: monitors the full life vibration data of the bearing until it is damaged; S2: Extract time-frequency domain features from vibration data and perform feature fusion to obtain input sequence; S3: Extract features of the input sequence through discrete wavelet downsampling to obtain data representation after feature learning; S4: extracting features from the discrete wavelet downsampled output data through Fourier frequency domain feature extraction, and further obtaining frequency domain features in the vibration data; S5: Improving the channel and spatial receptive field of vibration data through channel and spatial attention feature enhancement; S6: Using Attention-GRU regression prediction, the outputs of discrete wavelet downsampling, Fourier frequency domain feature extraction, and channel and spatial attention feature enhancement are fused and connected to obtain fused features , and calculate the predicted remaining lifespan by combining the health indicators of the predicted time step with the full lifespan data.

[0006] Further, the S2 includes: The time-frequency domain features are obtained by calculating one or more of the following features of the bearing vibration data: mean, root mean square, kurtosis, skewness, peak, variance, crest factor, impulse factor, margin factor, shape factor, root mean square amplitude and form factor; The obtained time-frequency domain features are fused to obtain the input sequence.

[0007] Further, the discrete wavelet downsampling comprises the steps of: S311: Input sequence Discrete Wavelet Transform Decompose it into low frequency parts and high frequency part :

[0008] S312: The high frequency part is further decomposed into horizontal , vertical direction and diagonal direction Three high frequency parts:

[0009]

[0010]

[0011] In the formula, represents the spatial dimensions (rows and columns), Indicates the number of channels, Represents the selection of high-frequency components in the horizontal direction. Represents the selection of high-frequency components in the vertical direction. Represents the selection of high-frequency components in the diagonal direction; S303: The low frequency part and high frequency part , as well as Splicing along the channel dimension, channel fusion after the reshaping Convolution, batch normalization and Feature learning including activation operation:

[0012] In the formula, is the output of discrete wavelet downsampling, is the batch normalization operation, is the convolution operation.

[0013] Further, the Fourier frequency domain feature extraction comprises the steps of: S401: Perform a 1D convolution operation on the input and compress the features:

[0014] In the formula, is the output of the discrete wavelet downsampling, for Output after 1D convolution operation; S402: Reshape the tensor and perform Fourier transform. After extracting the frequency domain information, take the inverse Fourier transform and reshape it to the original shape and then perform residual connection with the initial input:

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022] In the formula, The shape is , Yes The shape is shaped as , is the sequence length, is the number of channels, is the feature dimension, for The projection of for The projection of is the bias term, where Range is , Range is , Range is , Range is , For The result of Fourier transform is is the Fourier transform function, is the weight function, is a function that converts the specified variable into a complex number, is the initial weight, is a random function, To specify the shape of the output array, is the intermediate result calculated in the frequency domain, For The result of inverse Fourier transform is is the intermediate result calculated in the frequency domain, is the inverse Fourier transform function, is the initial weight, is to reshape the tensor into , is the sequence length, for The projection of Preprocessing output for vibration data.

[0023] Further, the channel attention feature enhancement comprises the steps of: S501: The input feature maps are concatenated together along the channel dimension to form a new feature map :

[0024] In the formula, is the splicing operation function, , , Represent the output of one time step respectively, Indicates the omitted input. Represents the set of outputs of N time steps; S502: Integrate multiple input features into a tensor and concatenate them into a new feature map Split into Block , Shape , for features and (Shape is ) performs global average pooling to obtain the channel descriptor :

[0025] In the formula, is the sequence length, is the number of features, To all The fused features; S503: Channel descriptor By having The fully connected layer calculates the attention weight of each block , the calculated attention weight Perform weighted summation on each block feature map to obtain the final output feature map :

[0026]

[0027] In the formula, and represents the weight matrix used for linear transformation, represent Activation function, is the activation function.

[0028] Further, the spatial attention feature enhancement comprises the steps of: S511: Stack and sum the feature maps after the input is processed by different convolution kernels:

[0029]

[0030] In the formula, is the convolution result, represents a one-dimensional convolution operation on the input x. Represents different convolution kernel sizes, is the activation function, represents the output of feature map fusion, Represents the output after the input undergoes a one-dimensional convolution operation; S512: Perform global average pooling on the fused features and reduce the number of channels through the fully connected layer:

[0031]

[0032] In the formula, is the output of the global average pooling layer, is the sequence length, is the output of feature map fusion, is the output of the fully connected layer, and are the weights and biases of the fully connected layer; S513: Calculate the attention weight of each channel and perform weighted summation on the multi-scale feature map:

[0033]

[0034] In the formula, and are the weights and biases of the kth fully connected layer, For the The attention weight of each channel, is the output of the module.

[0035] Furthermore, the Attention-GRU regression prediction includes the steps of: Encoder operation phase: S601: The outputs of discrete wavelet downsampling, Fourier frequency domain feature extraction, and channel and spatial attention feature enhancement are fused and connected to obtain fused features ; S602: Initial hidden state Connect to BIGRU together to get the encoder output and the hidden state at the current time step :

[0036] In the formula, represents the encoder output and the hidden state at the current time step, is a function of a bidirectional gating unit; S603: encoder output and the hidden state at the current time step Input into the attention mechanism of the decoder and calculate the attention weight and context vector :

[0037]

[0038] In the formula, Indicates transposing the matrix. represents the hyperbolic tangent function, Denotes the decoder at time step The hidden state of Denotes the decoder at time step The hidden state of is the i-th attention weight, represents the hidden state of the i-th decoder; Decoder operation phase: S604: Output the encoder of the previous time step and context vector After splicing, input GRU to get the output :

[0039] In the formula, is the function of the gating unit, is the decoder output of the previous time step; S605: and Concatenate and connect to the linear layer to get the decoder output of the current time step:

[0040] In the formula, is a linear transformation function.

[0041] The principles and advantages of the present invention are as follows: the bearing life prediction model in this scheme is composed of a discrete wavelet downsampling feature extraction module, a Fourier frequency domain feature extraction module, a channel and spatial attention feature enhancement module, and an Attention-GRU regression prediction module. By extracting features through discrete wavelet downsampling, the number of parameters in the input model is reduced, and the gradient explosion phenomenon is effectively avoided. The Fourier frequency domain extraction feature can further extract the frequency domain features in the vibration signal, while the channel attention feature extraction and the spatial attention feature extraction can obtain a more comprehensive receptive field, and the feature mining of the vibration signal is more in-depth and enhanced. While realizing multi-scale feature extraction, the channel attention feature extraction is also used to weigh the contribution of the input features to the RUL prediction, reducing the impact of irrelevant data on the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1The present invention is a flowchart of the steps of an embodiment of a method for predicting the remaining useful life of a bearing based on an improved Attention-GRU model.

[0043] Figure 2 This is a flow chart of a discrete wavelet downsampling feature extraction module of an embodiment of a bearing remaining service life prediction method based on an improved Attention-GRU model of the present invention.

[0044] Figure 3 This is a Fourier frequency domain feature extraction flow chart of an embodiment of a bearing remaining service life prediction method based on an improved Attention-GRU model of the present invention.

[0045] Figure 4 This is a flow chart of a channel attention feature extraction module of an embodiment of a bearing remaining service life prediction method based on an improved Attention-GRU model of the present invention.

[0046] Figure 5 This is a flow chart of the spatial attention feature extraction module of an embodiment of a bearing remaining service life prediction method based on an improved Attention-GRU model of the present invention. DETAILED DESCRIPTION

[0047] The following is further described in detail through specific implementation methods: Embodiment 1: A bearing remaining service life prediction method based on the improved Attention-GRU model, such as Figure 1 As shown, the following steps are included: S1: Monitor the vibration data of the bearing throughout its life until it is damaged, and pre-process the vibration data.

[0048] In this solution, the full life vibration data of the bearing running for a long time until it is damaged is monitored. The data includes the vibration data of the X and Y axes of the bearing. After the vibration data is obtained, it is preprocessed to clean the original vibration signal, remove noise and standardize it. The vibration data in this solution is divided into training set and test set and packaged into a python storage file for easy calling.

[0049] S2: Extract time-frequency domain features from vibration data and perform feature fusion to obtain input sequence. Time-frequency domain features include mean, root mean square, kurtosis, skewness, peak, variance, peak factor, pulse factor, margin factor, shape factor, root square amplitude and waveform factor, etc. The operation status of the bearing is reflected through time-frequency domain features.

[0050] S3: Extract features of the input sequence through discrete wavelet downsampling to obtain the data representation after feature learning.

[0051] S4: Feature extraction is performed on the output data of discrete wavelet downsampling through Fourier frequency domain feature extraction to further obtain frequency domain features in the vibration data.

[0052] S5: Improving the channel and spatial receptive field of vibration data through channel and spatial attention feature enhancement.

[0053] S6: Using Attention-GRU regression prediction, the outputs of discrete wavelet downsampling, Fourier frequency domain feature extraction, and channel and spatial attention feature enhancement are fused and connected to obtain fused features , and calculate the predicted remaining lifespan by combining the health indicators of the predicted time step with the full lifespan data.

[0054] Specifically, in this embodiment, a bearing remaining life prediction model is constructed according to the bearing remaining service life prediction method based on the improved Attention-GRU model, and its network architecture consists of an output layer, an encoder layer, a feature extraction layer, a feature normalization layer, a decoder layer and an output layer. In the feature extraction layer of the network structure, a discrete wavelet downsampling feature extraction module, a Fourier frequency domain feature extraction module, and a channel and space attention feature enhancement module are added to comprehensively enhance the detailed feature extraction of vibration. In terms of training methods, the model automatically extracts multidimensional features in the time domain and frequency domain, such as mean, kurtosis, skewness, etc., to avoid the limitations of traditional manual feature selection, and uses a dynamic attention mechanism to capture dependencies in time series data, solving the gradient disappearance problem in long sequence data processing. In addition, the normalization process ensures that the scale of each eigenvalue is consistent, and improves the convergence speed and stability of model training. In terms of training sets, the model designs a multi-dimensional feature extraction method, combines the sliding window method to generate samples, and enhances the adaptability and generalization ability of the model to different working conditions. By preprocessing the data set, the signal characteristics are enhanced, ensuring that the model can achieve high-precision prediction effects in complex bearing life prediction tasks. The discrete wavelet downsampling and Fourier frequency domain feature extraction module extracts time-frequency domain information from the bearing data to reduce the impact of environmental noise. The channel and spatial attention feature enhancement module obtains a more comprehensive receptive field for the bearing data in channel and space. The encoder-decoder regression prediction module outputs the HI of the predicted time step and combines it with the full life data to obtain the predicted remaining life.

[0055] As attached Figure 2 As shown, the operation steps of the discrete wavelet downsampling feature extraction module include: S301: Input Discrete Wavelet Transform Decompose it into low frequency parts and high frequency part ,in It is the fusion feature sequence after the time-frequency domain features of the bearing vibration data are calculated:

[0056] S302: Further decompose the high frequency part into horizontal , vertical direction and diagonal direction Three high frequency parts:

[0057]

[0058]

[0059] In the formula, represents the spatial dimensions (rows and columns), Indicates the number of channels Represents the selection of high-frequency components in the horizontal direction. Represents the selection of high-frequency components in the vertical direction. Represents the selection of high-frequency components in the diagonal direction.

[0060] S303: The low frequency part and high frequency part , as well as Splicing along the channel dimension, channel fusion after the reshaping Convolution , Batch Normalization and Feature learning including activation operation:

[0061] In the formula, is the data representation after feature learning, that is, the output of the discrete wavelet downsampling, is the batch normalization operation, is a convolution operation. This scheme can capture features of different scales by decomposing the signal into low-frequency and high-frequency parts. In bearing data analysis, this helps to separate fault-related features while reducing the impact of noise. The low-frequency part usually contains trends and periodic components, while the high-frequency part may reveal signs of instantaneous impact or damage. By further decomposing the high-frequency part into horizontal, vertical, and diagonal sub-bands, the directional characteristics of the vibration signal can be analyzed in more detail. In addition, by extracting features through discrete wavelet downsampling, the number of parameters of the input model is also reduced, which can effectively avoid the gradient explosion phenomenon.

[0062] Afterwards, as attached Figure 3As shown, the Fourier frequency domain feature extraction module performs feature extraction based on the output of the discrete wavelet downsampling module, specifically including the following steps: S401: Perform a 1D convolution operation on the input and compress the features:

[0063] In the formula, is the output of the discrete wavelet downsampling, for Output after a 1D convolution operation.

[0064] S402: Reshape the tensor and perform Fourier transform. After extracting the frequency domain information, take the inverse Fourier transform and reshape it to the original shape and then perform residual connection with the initial input:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] In the formula, The shape is , Yes The shape is shaped as , is the sequence length, is the number of channels, is the feature dimension, for The projection of for The projection of is the bias term, where Range is , Range is , Range is , Range is , For The result of Fourier transform is is the Fourier transform function, is the weight function, is a function that converts the specified variable into a complex number, is the initial weight, is a random function, To specify the shape of the output array, is the intermediate result calculated in the frequency domain, For The result of inverse Fourier transform is is the intermediate result calculated in the frequency domain, is the inverse Fourier transform function, is the initial weight, is to reshape the tensor into , is the sequence length, for The projection of Preprocessing output for vibration data.

[0073] Specifically, the output of the discrete wavelet downsampling module is compressed by 1D convolution to generate the feature sequence The shape is , represents the batch size, that is, the number of samples processed simultaneously. represents the length of the sequence, Indicates the number of channels, Yes The shape is shaped as After the features are reshaped, Fourier transform is performed, where Range is , Range is , Range is , The range is ,in Indicates the number of features selected during processing, Indicates the number of groups and slices of the sequence in the new dimension, represents the index of the feature channel, express The transpose of express The transpose of is the Fourier transform function, It is the feature sequence after the frequency domain features are extracted through Fourier transform. is the weight matrix used in the frequency domain, is the initial weight, Indicates that the variable Treated as plural operations, Generate a size of A random matrix, and Respectively represent the height and width of the generated matrix, Represents the dimension of the matrix, 2 represents the real and imaginary parts of the complex number, Indicates that the randomly generated weight matrix is ​​scaled to a range of 0.02 to prevent gradient explosion. is the intermediate result calculated in the frequency domain, is the inverse Fourier transform function, is the result of inverse Fourier transform Reshape into , The vibration data preprocessing output after discrete wavelet downsampling and convolution Fourier operation enhances the time and frequency domains and reduces the output of parameters.

[0074] In this scheme, the Fourier transform can reveal the dominant frequency components in the vibration signal, which are often closely related to the failure mode of the bearing. For example, a specific type of damage may cause harmonics of a specific frequency to appear. By adding a frequency domain feature extraction step to the model, these key frequency information can be captured, thereby providing more accurate fault diagnosis and life prediction. In addition, the processed signal is converted back to the time domain through the inverse Fourier transform and connected with the original input residual, which ensures the integrity of the information and avoids information loss during frequency domain processing.

[0075] The channel and spatial attention feature enhancement module includes channel attention feature enhancement and spatial attention feature enhancement. Figure 4 As shown, the steps include: S501: The input feature maps are concatenated together along the channel dimension to form a new feature map :

[0076] In the formula, is the splicing operation function, , , Represent the output of one time step respectively, Indicates the omitted input. Represents the set of outputs of N time steps, feature maps Shape .

[0077] S502: Integrate multiple input features into a tensor and concatenate them into a new feature map Split into Block , Shape , for features and (Shape is ) performs global average pooling to obtain the channel descriptor :

[0078] In the formula, is the sequence length, is the number of features, To all The fused features.

[0079] S503: Channel descriptor By having The fully connected layer calculates the attention weight of each block , the calculated attention weight Perform weighted summation on each block feature map to obtain the final output feature map :

[0080]

[0081] In the formula, and represents the weight matrix used for linear transformation, represent Activation function, is the activation function.

[0082] The channel attention mechanism allows the model to automatically learn the importance of features on different channels, that is, to determine which features are more important for the current prediction task. In the case of bearing data, different sensor channels may capture different types of information, and the information of some channels may be more critical for predicting the remaining service life. By introducing the channel attention mechanism, the model can dynamically adjust the weights of different channel features to optimize the combination of features and improve prediction performance. This adaptive feature weighting strategy helps improve the model's ability to identify bearing performance degradation under complex working conditions.

[0083] like Figure 5 As shown, the spatial attention feature enhancement module includes the following steps: S511: Stack and sum the feature maps after the input is processed by different convolution kernels:

[0084]

[0085] In the formula, is the convolution result, represents a one-dimensional convolution operation on the input x. Represents different convolution kernel sizes, is the activation function; Represents the output of feature map fusion, the shape is ; Represents the output after the input undergoes a one-dimensional convolution operation.

[0086] S512: Perform global average pooling on the fused features and reduce the number of channels through the fully connected layer:

[0087]

[0088] In the formula, is the output of the global average pooling layer, is the sequence length, and are the weights and biases of the fully connected layer, is the output of the fully connected layer, and are the weights and biases of the fully connected layers.

[0089] S513: Calculate the attention weight of each channel and perform weighted summation on the multi-scale feature map:

[0090]

[0091] In the formula, and are the weights and biases of the kth fully connected layer, For the The attention weight of each channel, is the output of the module. The spatial attention mechanism adopted in this scheme weights the spatial distribution of features, emphasizing the contribution of features at different positions to the current task. When processing high-dimensional data such as images, spatial attention can help the model focus on local areas that contain rich information. Although bearing vibration data does not show the spatial structure as intuitively as images, the feature map processed by the convolution layer still has spatial dimensions. The spatial attention mechanism enables the model to pay more attention to those areas that are more critical to fault diagnosis and life prediction, thereby improving the accuracy of prediction.

[0092] Attention-GRU regression prediction includes the following steps: Encoder operation phase: S601: The outputs of discrete wavelet downsampling, Fourier frequency domain feature extraction, and channel and spatial attention feature enhancement are fused and connected to obtain fused features .

[0093] S602: Initial hidden state Connect to BIGRU together to get the encoder output and the hidden state at the current time step :

[0094] In the formula, represents the encoder output and the hidden state at the current time step, is the function of the bidirectional gating unit.

[0095] S603: Encoder output and the hidden state at the current time step Input into the attention mechanism of the decoder and calculate the attention weight and context vector :

[0096]

[0097] In the formula, Indicates transposing the matrix. represents the hyperbolic tangent function, Denotes the decoder at time step The hidden state of Denotes the decoder at time step The hidden state of is the i-th attention weight, represents the hidden state of the i-th decoder.

[0098] Decoder operation phase: S604: Output the encoder of the previous time step and context vector After splicing, input GRU to get the output :

[0099] In the formula, is the function of the gating unit, is the decoder output of the previous time step.

[0100] S605: and Concatenate and connect to the linear layer to get the decoder output of the current time step:

[0101] In the formula, is a linear transformation function. The GRU (Gated Recurrent Unit) regression model combined with the attention mechanism is the core part of this solution. As a variant of recurrent neural network, GRU is particularly suitable for processing sequence data and can capture long-term dependencies in time series. In bearing life prediction, long-term trends in historical monitoring data are crucial for predicting future performance degradation. At the same time, the introduction of the attention mechanism enables the model to dynamically focus on the most relevant time steps under the encoder-decoder framework, thereby more accurately predicting the remaining service life. This mechanism is particularly suitable for variable-length sequence data and can adapt to historical monitoring records of different lengths.

[0102] In summary, the bearing life prediction model in this scheme extracts features through multi-scale and multi-directional wavelet transform, uses Fourier transform to capture frequency domain information, combines channel and spatial attention mechanisms to strengthen key features, and finally uses the Attention-GRU regression model for effective sequence modeling and life prediction. This method comprehensively utilizes the advantages of signal processing and deep learning, can improve the accuracy and robustness of bearing remaining service life prediction, and is of great significance for the maintenance and management of industrial equipment.

[0103] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the relevant field know all the common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for predicting the remaining useful life of a bearing based on an improved Attention-GRU model, characterized in that: The following steps are involved: S1: monitors the full life vibration data of the bearing until it is damaged; S2: Extract time-frequency domain features from vibration data and perform feature fusion to obtain input sequence; S3: Extract features of the input sequence through discrete wavelet downsampling to obtain data representation after feature learning; S4: extracting features from the discrete wavelet downsampled output data through Fourier frequency domain feature extraction, and further obtaining frequency domain features in the vibration data; S5: Improving the channel and spatial receptive field of vibration data through channel and spatial attention feature enhancement; S6: Using Attention-GRU regression prediction, the outputs of discrete wavelet downsampling, Fourier frequency domain feature extraction, and channel and spatial attention feature enhancement are fused and connected to obtain fused features , and calculate the predicted remaining lifespan by combining the health indicators of the predicted time step with the full lifespan data.

2. A method for predicting the remaining useful life of a bearing based on an improved Attention-GRU model according to claim 1, characterized in that: The S2 includes: The time-frequency domain features are obtained by calculating one or more of the following features of the bearing vibration data: mean, root mean square, kurtosis, skewness, peak, variance, crest factor, impulse factor, margin factor, shape factor, root mean square amplitude and form factor; The obtained time-frequency domain features are fused to obtain the input sequence.

3. The method for predicting the remaining useful life of a bearing based on an improved Attention-GRU model according to claim 1, characterized in that: The discrete wavelet downsampling comprises the steps of: S311: Input sequence Discrete Wavelet Transform Decompose it into low frequency parts and high frequency part : S312: The high frequency part is further decomposed into horizontal , vertical direction and diagonal direction Three high frequency parts: In the formula, represents the spatial dimensions (rows and columns), Indicates the number of channels, Represents the selection of high-frequency components in the horizontal direction. Represents the selection of high-frequency components in the vertical direction. Represents the selection of high-frequency components in the diagonal direction; S303: The low frequency part and high frequency part , as well as Splicing along the channel dimension, channel fusion after the reshaping Convolution, batch normalization and Feature learning including activation operation: In the formula, is the output of discrete wavelet downsampling, is the batch normalization operation, is the convolution operation.

4. The method for predicting the remaining useful life of a bearing based on an improved Attention-GRU model according to claim 3 is characterized in that: The Fourier frequency domain feature extraction comprises the steps of: S401: Perform a 1D convolution operation on the input and compress the features: In the formula, is the output of the discrete wavelet downsampling, for Output after 1D convolution operation; S402: Reshape the tensor and perform Fourier transform. After extracting the frequency domain information, take the inverse Fourier transform and reshape it to the original shape and then perform residual connection with the initial input: In the formula, The shape is , Yes The shape is shaped as , is the sequence length, is the number of channels, is the feature dimension, for The projection of for The projection of is the bias term, where Range is , Range is , Range is , Range is , For The result of Fourier transform is is the Fourier transform function, is the weight function, is a function that converts the specified variable into a complex number, is the initial weight, is a random function, To specify the shape of the output array, is the intermediate result calculated in the frequency domain, For The result of inverse Fourier transform is is the intermediate result calculated in the frequency domain, is the inverse Fourier transform function, is the initial weight, is to reshape the tensor into , is the sequence length, for The projection of Preprocessing output for vibration data.

5. The method for predicting the remaining useful life of a bearing based on the improved Attention-GRU model according to claim 4 is characterized in that: The channel attention feature enhancement includes step: S501: The input feature maps are concatenated together along the channel dimension to form a new feature map : In the formula, is the splicing operation function, , , Represent the output of one time step respectively, Indicates the omitted input. Represents the set of outputs of N time steps; S502: Integrate multiple input features into a tensor and concatenate them into a new feature map Split into Block , Shape , for features and (Shape is ) performs global average pooling to obtain the channel descriptor : In the formula, is the sequence length, is the number of features, To all The fused features; S503: Channel descriptor By having The fully connected layer calculates the attention weight of each block , the calculated attention weight Perform weighted summation on each block feature map to obtain the final output feature map : In the formula, and represents the weight matrix used for linear transformation, represent Activation function, is the activation function.

6. The method for predicting the remaining useful life of a bearing based on an improved Attention-GRU model according to claim 5, characterized in that: The spatial attention feature enhancement includes step: S511: Stack and sum the feature maps after the input is processed by different convolution kernels: In the formula, is the convolution result, represents a one-dimensional convolution operation on the input x. Represents different convolution kernel sizes, is the activation function, represents the output of feature map fusion, Represents the output after the input undergoes a one-dimensional convolution operation; S512: Perform global average pooling on the fused features and reduce the number of channels through the fully connected layer: In the formula, is the output of the global average pooling layer, is the sequence length, is the output of feature map fusion, is the output of the fully connected layer, and are the weights and biases of the fully connected layer; S513: Calculate the attention weight of each channel and perform weighted summation on the multi-scale feature map: In the formula, and are the weights and biases of the kth fully connected layer, For the The attention weight of each channel, is the output of the module.

7. The method for predicting the remaining useful life of a bearing based on an improved Attention-GRU model according to claim 6, characterized in that: The Attention-GRU regression prediction includes the following steps: Encoder operation phase: S601: The outputs of discrete wavelet downsampling, Fourier frequency domain feature extraction, and channel and spatial attention feature enhancement are fused and connected to obtain fused features ; S602: Initial hidden state Connect to BIGRU together to get the encoder output and the hidden state at the current time step : In the formula, represents the encoder output and the hidden state at the current time step, is the function of the bidirectional gating unit; S603: Encoder output and the hidden state at the current time step Input into the attention mechanism of the decoder and calculate the attention weight and context vector : In the formula, Indicates transposing the matrix. represents the hyperbolic tangent function, Denotes the decoder at time step The hidden state of Denotes the decoder at time step The hidden state of is the i-th attention weight, represents the hidden state of the i-th decoder; Decoder operation phase: S604: Output the encoder of the previous time step and context vector After splicing, input GRU to get the output : In the formula, is the function of the gating unit, is the decoder output of the previous time step; S605: and Concatenate and connect to the linear layer to get the decoder output of the current time step: In the formula, is a linear transformation function.

Images