Bearing residual service life estimation model design method based on Conv-LSTM and CBAM

Abstract

The invention discloses a bearing residual service life estimation model design method based on Conv-LSTM and CBAM. The method can accurately capture the time dependence and spatial characteristics in the performance degradation process of the new energy automobile speed reducer, and achieves the efficient estimation of the RUL of the new energy automobile speed reducer. A Conv-LSTM-based network structure is designed and is used for processing time series data such as vibration signals of the new energy automobile speed reducer and extracting performance degradation characteristics of the new energy automobile speed reducer. A CBAM module is introduced, the attention mechanism of the network to the features is enhanced, and the CBAM module is automatically focused on a signal part which is most critical to RUL estimation. The network structure is adaptively adjusted, network parameters are dynamically optimized according to real-time monitoring data, and the generalization ability and prediction accuracy of the model are improved. According to the method, Conv-LSTM and CBAM are combined, key features are effectively extracted and emphasized, and the accuracy of new energy automobile reducer RUL estimation is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of industrial fault detection and relates to a bearing remaining service life estimation model design method based on Conv-LSTM and CBAM. Background Art

[0002] As one of the key components of the vehicle, the reducer bearing of new energy vehicles undertakes the important task of wheel rotation and vehicle driving. However, during long-term use or under the influence of the external environment, the bearing may suffer from wear, fatigue, cracks and other faults, which will cause abnormal sounds and vibrations during vehicle operation, and may even cause safety hazards such as vehicle loss of control. Therefore, early detection and accurate diagnosis of the fault of the reducer bearing of new energy vehicles is crucial to ensure vehicle safety and extend vehicle service life.

[0003] Traditional fault diagnosis methods such as threshold judgment, expert systems and machine learning methods generally have problems such as poor generalization ability, difficulty in capturing complex fault features and limited deployment scenarios. In recent years, deep learning algorithms have been widely used in the field of fault diagnosis due to their powerful feature extraction capabilities and good generalization performance. However, most existing fault diagnosis methods based on deep learning usually adopt a two-stage model, that is, first extract features through signal processing methods, and then input the features into the neural network for training. The disadvantage of this method is that the dynamic adjustment ability of the neural network cannot be well fed back to the front-end signal processing stage. Summary of the invention

[0004] In view of the deficiencies in the existing technology, the present invention provides a method for designing a remaining useful life estimation model for intelligent bearings based on Conv-LSTM (convolutional long short-term memory network) and CBAM (convolutional block attention module).

[0005] The present invention comprises the following steps:

[0006] Step 101: collect and prepare a new energy vehicle reducer vibration data set, including a training set and a test set; each sample includes the new energy vehicle reducer vibration data and a corresponding label value, that is, the remaining life of the current state.

[0007] This data will be used to train machine learning models to predict the health status and potential failure time of the reducer, thereby achieving preventive maintenance. In order to ensure the diversity and representativeness of the dataset, vibration data is collected from new energy vehicles and their operating environment and conditions are recorded.

[0008] Step 102: Feature Layer Design:

[0009] As a part of the neural network, the feature layer is responsible for converting time series waveform data into feature coefficients.

[0010] The feature layer uses dilated convolution with learnable parameters to extract features in the time and frequency domains.

[0011] Step 103: Intermediate feature extraction:

[0012] The output of the feature layer is represented as frequency domain features and combined with the original time domain features to form the input of the network.

[0013] This step aims to preserve as much information as possible while taking advantage of frequency domain features.

[0014] Dilated convolution can expand the receptive field without adding additional parameters, thereby capturing a wider range of contextual information;

[0015] Step 104: Use the CBAM structure to screen the feature importance of the feature output of the dilated convolutional layer:

[0016] Using the CBAM (Convolutional Block Attention Module) structure to screen the feature importance of the feature output of the dilated convolutional layer can significantly improve the performance and robustness of the model.

[0017] CBAM is a simple yet effective attention mechanism module that automatically enhances the attention to useful information in the input feature map while suppressing irrelevant information by calculating the importance of the feature map in both spatial and channel dimensions.

[0018] Specifically, applying CBAM after the dilated convolutional layer can more accurately capture the key areas and channels in the feature map, allowing subsequent network layers to focus more on features that are critical to task objectives (such as classification or detection), thereby improving the recognition accuracy and generalization ability of the entire model.

[0019] Step 105: CBAM Attention Module:

[0020] The CBAM attention module is integrated into the network to enhance the network's attention to important features.

[0021] The CBAM module can automatically identify and emphasize important timing and frequency domain features and reduce the interference of irrelevant information.

[0022] This helps the model focus on key areas in the input data and ignore irrelevant or redundant information.

[0023] Step 106: CBAM can learn the importance of each feature channel through the channel attention mechanism and adjust the weight of each channel accordingly, so as to highlight those feature channels that are more important for a specific task.

[0024] Using the CBAM (Convolutional Block Attention Module) structure to screen the feature importance of the feature output of the dilated convolutional layer can significantly improve the performance and robustness of the model. CBAM is a simple and effective attention mechanism module that automatically enhances the attention to useful information in the input feature map while suppressing irrelevant information by calculating the importance of the feature map in both spatial and channel dimensions. Specifically, applying CBAM after the dilated convolutional layer can more accurately capture the key areas and channels in the feature map.

[0025] Step 107: Adaptive Activation Layer

[0026] Adaptive activation layers, including learnable parameters such as scaling factors and offset parameters.

[0027] A parameterized GELU activation function with adjustable slope and offset to increase the flexibility and expressiveness of the model.

[0028] Further, the feature layer design is designed as follows:

[0029] Step 101: Construct a dilated convolution layer: Dilated convolution is a special convolution operation that allows the convolution kernel to move in a jumpy manner on the input data, thereby expanding the receptive field without increasing the computational cost. The input of this layer is the fault timing waveform data.

[0030] Step 102: Multi-scale feature extraction: Multi-scale feature extraction uses dilated convolution kernels with different dilation rates to capture features of different scales, effectively covering a wide frequency range. By setting multiple dilation rates, multiple convolution kernels can simultaneously extract local details and global context information, enhance the model's understanding and adaptability to input data, and improve its performance in complex tasks.

[0031] Step 103: Adaptive feature transformation: The input signal is subjected to feature transformation through the learned feature function, aiming to capture the frequency domain features of the signal so as to better characterize the information.

[0032] Step 104: Extract frequency domain features: Use spectrum analysis method to convert time domain signal into frequency domain signal, and then apply dilated convolution to extract frequency domain features.

[0033] Step 105: Feature fusion: Merge the time domain features and frequency domain features to form a comprehensive feature representation. Feature fusion can be achieved through simple concatenation or more advanced fusion techniques (such as attention mechanism).

[0034] Step 106: Feature normalization: The extracted features are normalized to ensure the scale consistency and numerical stability of the features. Batch normalization or other normalization techniques are used to adjust the distribution of the features.

[0035] Step 107: Feature enhancement: Transform the features through data enhancement techniques to improve the generalization ability of the model. Use random cropping, rotation, flipping and other techniques to transform the features.

[0036] Step 108: Adaptive parameter adjustment: The learnable parameters in the dilated convolution layer are adjusted to adapt to different failure modes and working conditions. The parameters in the dilated convolution layer are adjusted by back propagation during the training process.

[0037] Step 109 may also be included: feature encoding optimization: This step optimizes the feature encoding process by adding more convolutional layers or adjusting the convolution kernel size to improve the accuracy of feature extraction. Specific methods include enhancing the depth and expression ability of the model, using convolution kernels of different sizes to capture multi-scale features, and exploring combinations of different convolution kernel sizes and numbers of layers to find the best feature encoding strategy, thereby significantly improving the accuracy of feature encoding and the generalization ability of the model.

[0038] Furthermore, the intermediate feature extraction is designed as follows:

[0039] Step 101: different dilation rates: by setting different dilation rates, a plurality of dilated convolution kernels are used to perform convolution operations on the frequency domain signal to extract frequency domain features.

[0040] Step 102: Time-frequency features: directly use multiple dilated convolution kernels to perform convolution operations on the time domain signal to extract time domain features.

[0041] Step 103: Normalization: Use global normalization techniques to adjust the distribution of features.

[0042] Furthermore, the Conv-LSTM layer module is designed as follows:

[0043] Step 101: Conv-LSTM unit design: The Conv-LSTM unit is the basic building block of the Conv-LSTM layer, which combines the convolution operation and the memory unit of LSTM. At each time step, the Conv-LSTM unit receives input and updates its internal state while producing output.

[0044] Step 102: Input data preparation: Input the time domain and frequency domain features processed by the feature layer to the Conv-LSTM layer. Make sure that the format of the input data is suitable for processing by the Conv-LSTM layer, usually a three-dimensional tensor (batch_size, time_steps, features).

[0045] Step 103: Initialize states: Initialize the Conv-LSTM layer hidden states and cell states with zero matrices.

[0046] Step 104: Time series processing: The input data of each time step is processed through the Conv-LSTM layer. After each time step, the Conv-LSTM unit updates its hidden state and unit state and generates output.

[0047] Step 105: Output consolidation: Consolidate the outputs of all time steps into a continuous sequence to form the final output of the Conv-LSTM layer. Use the last item of the sequence as the final output, or perform further processing on the entire sequence.

[0048] Furthermore, the CBAM attention module aims to improve the ability to learn long-distance dependencies in signal processing. The design is as follows:

[0049] Step 101: Channel attention mechanism: Perform global average pooling and global maximum pooling on the output of the Conv-LSTM layer to obtain the importance score of each channel. Generate the weight of each channel through a fully connected layer and an activation function (such as Sigmoid), and then multiply these weights with the original feature map.

[0050] Step 102: Spatial Attention Mechanism: Perform max pooling and average pooling operations on the output of the Conv-LSTM layer to capture the importance of different positions. The weights of the spatial positions are generated through the convolutional layer and activation function (such as Sigmoid), and then these weights are multiplied with the original feature map.

[0051] Step 103: Feature fusion: Fuse the results of channel attention and spatial attention.

[0052] Step 104: Feature normalization: Normalize the features processed by the CBAM module to ensure the scale consistency and numerical stability of the features. Use batch normalization or other normalization techniques to adjust the distribution of the features.

[0053] Step 105: Feature propagation: Pass the features processed by the CBAM module to the next layer of the network. Directly use the features as the input of the next layer, or further enhance the feature propagation through residual connections.

[0054] Step 106: Parameter Optimization: Adjust the parameters in the CBAM module through back-propagation to improve the accuracy of feature selection. Update the weights in the CBAM module using an optimization algorithm such as gradient descent.

[0055] Furthermore, the adaptive activation layer is designed as follows:

[0056] Step 101: Selection of parameterized activation function: Determine which parameterized activation function to use, such as parameterized ReLU (PReLU) or parameterized GELU (PGELU). Parameterized activation functions can automatically adjust the slope and offset according to the input data, thereby improving the nonlinear expression ability of the model.

[0057] Step 102: Initialize parameters: Initialize the learnable parameters in the activation function, such as slope α and offset β. Use small positive numbers or zero to initialize these parameters.

[0058] Step 103: Calculate activation output: Initialize the learnable parameters in the activation function, such as slope α and offset β. Use small positive numbers or zero to initialize these parameters.

[0059] Step 104: Parameter update: Update the parameters in the adaptive activation layer by back propagation during the training process. Use an optimization algorithm such as gradient descent to update the parameters according to the gradient of the loss function.

[0060] Step 105: Feature enhancement: By adjusting the parameters of the activation function, the network's response to key features is enhanced. The parameterized activation function can automatically adjust the slope and offset according to the input data to highlight important features.

[0061] Based on the above designed model, the present invention also provides the application of the above model in estimating the remaining service life of a bearing.

[0062] Beneficial effects of the present invention:

[0063] The present invention uses deep learning technology to accurately predict the remaining useful life (RUL) of bearings by real-time monitoring and analyzing the data of new energy vehicle reducer bearings. This not only helps to improve the reliability and stability of new energy vehicle reducer bearing systems, but also reduces maintenance costs and improves overall safety and economy. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0065] Figure 1 This is a flow chart of a method for detecting bearing faults of a reducer of a new energy vehicle according to the present invention;

[0066] Figure 2 Flowchart of the spatiotemporal convolutional layer and attention mechanism of the present invention; DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] Remaining useful life (RUL) prediction plays a vital role in the stability and reliability of mechanical equipment: RUL prediction can help mechanical systems understand the service life of components in advance, including new energy vehicle reducer bearings, transmission systems, etc. By timely predicting the RUL of equipment, maintenance can be arranged in advance to prevent failures and improve system stability. Equipment operation and maintenance personnel can better plan maintenance time based on RUL prediction results, thereby reducing downtime and reducing the impact on production efficiency.

[0069] RUL prediction helps improve the reliability of mechanical equipment. Through predictive and real-time RUL prediction, the system can better cope with sudden failures and ensure continuous and reliable operation of the equipment. Failures in mechanical equipment often trigger chain reactions, leading to more extensive system failures. RUL prediction can provide timely warnings to prevent the expansion of chain reactions and help maintain the overall stability of the system. Mechanical equipment may face a variety of complex failure situations, including multi-point failures, equipment combination failures, etc. The application of RUL prediction technology can better cope with these complex situations and improve the stability of the system.

[0070] As a key component in mechanical equipment, the reducer bearing of new energy vehicles is responsible for supporting the rotating shaft and reducing friction. The RUL prediction of the reducer bearing of new energy vehicles can identify potential failure points in advance, prevent the chain reaction of failures in the mechanical system, and affect the reliability and stability of the equipment. By timely predicting the RUL of the reducer bearing of new energy vehicles, the stability of mechanical equipment can be improved and the risk of chain reaction of failures can be reduced. The reducer bearing of new energy vehicles is a key node in mechanical equipment, and its safe operation is crucial to the overall operation of the equipment. Through RUL prediction, potential problems can be discovered in advance, accidents can be prevented, and the safe operation of the equipment can be ensured.

[0071] Compared with traditional threshold judgment methods and machine learning methods, deep learning has some significant advantages in the field of RUL prediction of new energy vehicle reducer bearings, which are mainly reflected in the following aspects: Traditional methods usually rely on domain experts to manually extract features, which may face difficulties for complex nonlinear problems and multimodal data processing. Deep learning can automatically learn features in the data without manually designing feature extractors, and can better adapt to different prediction modes.

[0072] The prediction of RUL of reducer bearings for new energy vehicles usually requires processing a large amount of multimodal data such as vibration data, temperature data, and sound signals. Deep learning models are good at processing large-scale data and can better mine potential patterns in the data and improve the accuracy of predictions. Deep learning models can learn complex nonlinear relationships, which is crucial for the various complex patterns that exist in the prediction of RUL of reducer bearings for new energy vehicles. Traditional methods may not be able to capture these complex relationships, while deep learning can better model nonlinear relationships.

[0073] Transfer learning technology enables deep learning models trained in one field to be transferred to other related fields, thereby improving the generalization ability of the model on new tasks. This is of great significance to the data scarcity problem in the field of RUL prediction of new energy vehicle reducer bearings. RUL prediction of new energy vehicle reducer bearings may involve multiple data sources, such as vibration, sound, temperature, etc. Deep learning can easily integrate multimodal information and improve the effect of comprehensive prediction, while traditional methods may need to process each type of data independently.

[0074] Deep learning supports end-to-end learning and can learn task-related mappings directly from raw data. This eliminates the tedious steps of manually designing features and processing flows, and simplifies the model building and training process. Deep learning models are generally robust to noise and complex backgrounds. This is of great significance for the various interferences and complex conditions that exist in the actual operating environment of new energy vehicle reducer bearings, and can improve the reliability of the prediction model.

[0075] The present invention mainly involves the following details: 1) an adaptive deep neural network based on Conv-LSTM and CBAM, comprising: an input layer for receiving an input signal; a Conv-LSTM layer for extracting spatiotemporal features of the input signal; a CBAM layer (Convolutional Block Attention Module) for enhancing feature representation; an output layer for outputting processed results; 2) an adaptive feature transformation layer; 3) an Embedding layer with nonlinear mapping; 4) an encoder module combining Conv-LSTM and CBAM; 5) a long-distance information capture mechanism based on spatiotemporal-aware convolution and attention mechanism (Attention); 6) data enhancement of intermediate features of the feature layer; 7) an adaptive output layer activation function.

[0076] The present invention provides an adaptive deep neural network for estimating the remaining service life of intelligent bearings based on Conv-LSTM and CBAM. The network can directly learn task-related mappings from raw data, simplifying the model construction and training process. In addition, through the application of technologies such as Conv-LSTM and CBAM, the network can effectively process the vibration signals of the reducer bearings of new energy vehicles and provide accurate remaining service life estimation, which helps to improve the reliability, safety and economy of the system.

[0077] Example:

[0078] like Figure 1 and Figure 2 As shown, this application uses an end-to-end deep learning network structure to predict the remaining useful life (RUL) of the reducer bearing of a new energy vehicle. The network structure is specially designed for the RUL prediction of the reducer bearing of a new energy vehicle:

[0079] Feature transformation layer:

[0080] The feature transformation layer is introduced to perform multi-scale feature decomposition on the original vibration signal of the reducer bearing of new energy vehicles and extract the fault characteristics of the reducer bearing of new energy vehicles at different frequencies.

[0081] Implementation: Use dilated convolutions with learnable parameters to extract time and frequency domain features.

[0082] Feature encoding layer:

[0083] Feature encoding: The multi-scale features are passed to the feature encoding layer, and the features are encoded using a convolutional neural network (CNN) or a fully connected layer to capture high-level local features in the vibration signal of the reducer bearing of new energy vehicles.

[0084] Conv-LSTM layer:

[0085] Conv-LSTM unit: The Conv-LSTM layer is introduced to capture the spatiotemporal characteristics in the vibration signal of the reducer bearing of new energy vehicles.

[0086] Implementation: The signal sequence at each feature scale is fed into a separate Conv-LSTM unit, enabling the model to effectively learn long-term dependencies in time series data.

[0087] CBAM attention module:

[0088] Multi-head attention: A CBAM attention module is introduced in the Conv-LSTM layer to capture the associations of different positions in the signal sequence.

[0089] Position encoding: Use position encoding to inject temporal information into the sequence and enhance the model’s ability to understand the signal change process.

[0090] Residual connections and layer normalization:

[0091] Residual Connection: Residual connections and layer normalization are introduced in Conv-LSTM layers to promote gradient flow and stability of network training.

[0092] Pooling layer:

[0093] Pooling: A pooling layer is introduced after the Conv-LSTM layer to reduce the dimension of features and retain the most important information, thereby improving the computational efficiency of the network.

[0094] Fully connected layer and output layer:

[0095] Fully connected layer: The fully connected layer is connected after the pooling layer and finally connected to the output layer to predict the remaining useful life (RUL) of the reducer bearing of new energy vehicles.

[0096] Furthermore, this application includes the following improvements:

[0097] Improvements to the feature transformation layer:

[0098] The feature transformation layer is introduced to perform multi-scale feature decomposition on the original vibration signal of the reducer bearing of new energy vehicles and extract the fault characteristics of the reducer bearing of new energy vehicles at different frequencies.

[0099] Implementation: Use dilated convolution with learnable parameters to extract time domain and frequency domain features to more effectively capture feature information at different frequencies.

[0100] Optimization of feature encoding layer:

[0101] Feature encoding: Pass multi-scale features to the feature encoding layer, and use a convolutional neural network (CNN) or a fully connected layer to encode the features to capture high-level local features in the vibration signal of the reducer bearing of new energy vehicles. Use multiple convolutional layers to extract features from the input signal, and each convolutional layer is followed by batch normalization and an activation function (such as ReLU).

[0102] Improvement: Optimize the feature encoding process by adding more convolutional layers or adjusting the convolution kernel size to improve the accuracy of feature extraction.

[0103] Optimization of Conv-LSTM layer:

[0104] Conv-LSTM unit: The Conv-LSTM layer is introduced to capture the spatiotemporal characteristics of the vibration signal of the reducer bearing of new energy vehicles. The signal sequence of each feature scale is fed into an independent Conv-LSTM unit, enabling the model to effectively learn long-term dependencies in time series data.

[0105] Improvement: Optimize the structure of the model by increasing the number of Conv-LSTM layers or adjusting the number of hidden units to better capture the long-term dependencies in time series data.

[0106] Enhancements to the CBAM attention module:

[0107] Multi-head attention: A CBAM attention module is introduced in the Conv-LSTM layer to capture the associations of different positions in the signal sequence.

[0108] Improvement: Enhance the model's focus on important features by increasing the number of attention heads or optimizing the position encoding strategy.

[0109] Optimization of residual connections and layer normalization:

[0110] Residual Connection: Residual connections and layer normalization are introduced in Conv-LSTM layers to promote gradient flow and stability of network training.

[0111] Improvement: Further improve the stability of network training by adjusting the residual connection method or optimizing the parameters of layer normalization.

[0112] Optimization of pooling layer:

[0113] Pooling: A pooling layer is introduced after the Conv-LSTM layer to reduce the dimension of features and retain the most important information, thereby improving the computational efficiency of the network.

[0114] Improvement: Optimize the feature dimensionality reduction process by trying different pooling strategies (such as global average pooling, global maximum pooling) or adaptive pooling.

[0115] Optimization of the fully connected layer and output layer:

[0116] Fully connected layer: The fully connected layer is connected after the pooling layer and finally connected to the output layer to predict the remaining useful life (RUL) of the reducer bearing of new energy vehicles.

[0117] Improvement: The output layer of the model is optimized by adjusting the number of layers or units in the fully connected layer to improve the accuracy of RUL prediction.

[0118] Furthermore, this embodiment also provides dilated convolution feature coding in RUL, and its specific implementation and improvement points are as follows:

[0119] Design of dilated convolutional feature encoding layer:

[0120] Purpose: To introduce a dilated convolutional feature encoding layer to more effectively extract time and frequency domain features from the vibration signal of the reducer bearing of new energy vehicles.

[0121] Implementation: Dilated convolutions with learnable parameters are used to extract time and frequency domain features to capture information at different scales.

[0122] Multi-scale feature extraction:

[0123] Multi-scale: Use dilated convolution kernels with different dilation rates to extract features of different scales to cover a wider frequency range.

[0124] Implementation: By setting different dilation rates, multiple dilated convolution kernels are used to perform convolution operations on the input signal to extract multi-scale features.

[0125] Feature Fusion:

[0126] Fusion: Combine time domain features and frequency domain features to form a comprehensive feature representation.

[0127] Method: Feature fusion can be achieved through simple concatenation or more advanced fusion techniques such as attention mechanism.

[0128] Feature Normalization:

[0129] Standardization: The extracted features are standardized to ensure the scale consistency and numerical stability of the features.

[0130] Method: Use Batch Normalization or other normalization techniques to adjust the distribution of features.

[0131] Feature Enhancements:

[0132] Enhancement: Transform features through data augmentation techniques to improve the generalization ability of the model.

[0133] Method: Use random cropping, rotation, flipping and other techniques to transform features.

[0134] Adaptive parameter adjustment:

[0135] Tuning: Adapt to different failure modes and operating conditions by adjusting the learnable parameters in the dilated convolutional layers.

[0136] Method: Parameters in dilated convolutional layers are adjusted via back-propagation during training.

[0137] Feature encoding optimization:

[0138] Optimization: Optimize the feature encoding process by adding more convolutional layers or adjusting the convolution kernel size to improve the accuracy of feature extraction.

[0139] Method: Explore different combinations of convolution kernel sizes and number of layers to find the best feature encoding strategy.

[0140] Furthermore, this embodiment also provides a Conv-LSTM structure in RUL, and its specific implementation and improvements are as follows:

[0141] Conv-LSTM cell design:

[0142] Purpose: To introduce the Conv-LSTM layer to capture the spatiotemporal characteristics of the vibration signal of the reducer bearing of new energy vehicles.

[0143] Implementation: At each time step, a Conv-LSTM cell receives input and updates its internal state while producing output.

[0144] Input data preparation:

[0145] Input: The time domain and frequency domain features processed by the feature layer are input into the Conv-LSTM layer.

[0146] Format: Make sure the input data is in a format suitable for processing by the Conv-LSTM layer, usually a 3D tensor (batch_size, time_steps, features).

[0147] Initialization state:

[0148] State: Conv-LSTM layer requires initialization of hidden state and cell state.

[0149] Method: Usually both states are initialized with zero matrices.

[0150] Timing Processing:

[0151] Processing: Process the input data at each time step through the Conv-LSTM layer.

[0152] Update: After each time step, the Conv-LSTM unit updates its hidden state and cell state and produces output.

[0153] Output Integration:

[0154] Integration: The outputs of all time steps are integrated into a continuous sequence to form the final output of the Conv-LSTM layer.

[0155] Method: You can use the last item of the sequence as the final output, or perform further processing on the entire sequence.

[0156] Residual connections and layer normalization:

[0157] Residual Connection: Residual connections are introduced in Conv-LSTM layers to promote gradient flow and stability of network training.

[0158] Layer Normalization: Layer normalization is introduced to stabilize the gradient during training.

[0159] Feature encoding optimization:

[0160] Optimization: Optimize the structure of the model by adding more Conv-LSTM layers or adjusting the number of hidden units to better capture the long-term dependencies in time series data.

[0161] Method: Explore different combinations of number of layers and hidden units to find the best feature encoding strategy.

[0162] Adaptive parameter adjustment:

[0163] Tuning: By adjusting the learnable parameters in the Conv-LSTM layer to adapt to different failure modes and operating conditions.

[0164] Method: Parameters in the Conv-LSTM layer are adjusted via back-propagation during training.

[0165] Feature Fusion:

[0166] Fusion: The output of the Conv-LSTM layer is fused with the output of other feature layers to form a comprehensive feature representation.

[0167] Method: Feature fusion can be achieved through simple concatenation or more advanced fusion techniques such as attention mechanism.

[0168] Furthermore, this embodiment also provides a CBAM structure in RUL, and its specific implementation and improvement points are as follows:

[0169] Design of CBAM attention module:

[0170] Purpose: To integrate the CBAM attention module into the neural network to enhance the network's attention to important features.

[0171] Implementation: Use the CBAM module to automatically identify and emphasize important timing and frequency domain features and reduce the interference of irrelevant information.

[0172] Channel Attention Mechanism:

[0173] Channel attention focuses on which channels are more important.

[0174] Implementation: Perform global average pooling and global max pooling on the output of the Conv-LSTM layer to get the importance score of each channel.

[0175] Method: Generate weights for each channel through a fully connected layer and an activation function (such as Sigmoid), and then multiply these weights with the original feature map.

[0176] Spatial Attention Mechanism:

[0177] Spatial attention focuses on which spatial locations are more important.

[0178] Implementation: Max pooling and average pooling operations are performed on the output of the Conv-LSTM layer to capture the importance of different positions.

[0179] Method: Generate weights of spatial positions through convolutional layers and activation functions (such as Sigmoid), and then multiply these weights with the original feature map.

[0180] Feature Fusion:

[0181] Fusion: Fusion the results of channel attention and spatial attention.

[0182] Method: Feature fusion can be achieved through simple multiplication operations or more advanced fusion techniques such as weighted summation.

[0183] Feature Normalization:

[0184] Standardization: The features processed by the CBAM module are standardized to ensure the scale consistency and numerical stability of the features.

[0185] Method: Use Batch Normalization or other normalization techniques to adjust the distribution of features.

[0186] Feature Propagation:

[0187] Propagation: Pass the features processed by the CBAM module to the next layer of the network.

[0188] Method: Directly use the features as the input of the next layer, or further enhance the propagation of features through residual connections.

[0189] Parameter optimization:

[0190] Optimization: Parameters in the CBAM module are adjusted via back-propagation to improve the accuracy of feature selection.

[0191] Method: We use optimization algorithms such as gradient descent to update the weights in the CBAM module.

[0192] Improvements:

[0193] Multi-head attention: Increasing the number of attention heads in the CBAM module to better capture the correlations at different positions in the signal sequence.

[0194] Position encoding: Optimize the position encoding strategy to enhance the model's ability to understand the signal change process.

[0195] Residual connection and layer normalization: The stability of network training can be further improved by adjusting the residual connection method or optimizing the parameters of layer normalization.

[0196] In summary, this application designs an adaptive deep neural network for estimating the remaining useful life of intelligent bearings based on Conv-LSTM and CBAM. The advantages of Conv-LSTM in time series data processing and the enhancement of CBAM on the feature attention mechanism can more effectively extract fault features from complex vibration signals, and can automatically adjust the network weights according to actual conditions to better adapt to fault diagnosis needs in different scenarios. Based on this, the network can accurately capture the time dependency and spatial characteristics of the bearing performance degradation process, and realize efficient estimation of the remaining useful life RUL of the bearing.

[0197] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A bearing remaining service life estimation model design method based on Conv-LSTM and CBAM, characterized by Step 101: Collect and prepare a new energy vehicle reducer vibration data set, including a training set and a test set; each sample includes the new energy vehicle reducer vibration data and a corresponding label value, that is, the remaining life of the current state; Step 102: Automatically identify and extract key features in the vibration signal using a feature embedding layer; Step 103: Using dilated convolution to perform deeper feature analysis and extraction on the feature layer after preliminary feature extraction; Step 104: Use the CBAM structure to screen the feature importance of the feature output of the dilated convolutional layer; Step 105: Use CBAM to re-weight the feature map according to the importance of pixels at each position through the spatial attention mechanism; Step 106: CBAM learns the importance of each feature channel through the channel attention mechanism and adjusts the weight of each channel accordingly; Step 107: Design an adaptive activation layer, including learnable parameters: scaling factor and offset parameter, and select a parameterized GELU activation function with adjustable slope and offset.

2. The method according to claim 1, characterized in that The feature embedding layer in step 102 is designed as follows: Step 101: Design of dilated convolution unit: Introduce the dilated convolution layer to more effectively extract time domain and frequency domain features from the vibration signal of the reducer bearing of the new energy vehicle; Use dilated convolution with learnable parameters to extract time and frequency domain features to capture information at different scales; Step 102: multi-scale feature extraction: using dilated convolution kernels with different dilation rates to extract features of different scales to cover a wider frequency range; by setting different dilation rates, using multiple dilated convolution kernels to perform convolution operations on the input signal to extract multi-scale features; Step 103: Feature fusion: combine the time domain features and the frequency domain features to form a comprehensive feature representation; Step 104: Feature standardization: The extracted features are standardized to ensure the scale consistency and numerical stability of the features.

3. The method according to claim 2, characterized in that Also includes: Step 105: Feature enhancement: transform the features through data enhancement technology to improve the generalization ability of the model; Step 106: Adaptive parameter adjustment: by adjusting the learnable parameters in the dilated convolutional layer to adapt to different failure modes and working conditions; Step 107: Feature encoding optimization: Optimize the feature encoding process by adding more convolutional layers or adjusting the convolution kernel size to improve the accuracy of feature extraction.

4. The method according to claim 1, characterized in that: The intermediate feature extraction in step 102 is specifically: Step 101: by setting different dilation rates, a plurality of dilated convolution kernels are used to perform a convolution operation on the frequency domain signal to extract frequency domain features; Step 102: using multiple dilated convolution kernels to perform a convolution operation on the time domain signal to extract time domain features; Step 103: Use global normalization techniques to adjust the distribution of features.

5. The method according to claim 1, characterized in that The Conv-LSTM layer module is designed as follows: Step 101: Conv-LSTM unit design: The Conv-LSTM unit is the basic building block of the Conv-LSTM layer, which combines the convolution operation and the memory unit of LSTM. At each time step, the Conv-LSTM unit receives input and updates its internal state, while producing output. Step 102: Input data preparation: input the time domain and frequency domain features processed by the feature layer into the Conv-LSTM layer; Step 103: Initialize the state: Initialize the hidden state and cell state of the Conv-LSTM layer with a zero matrix; Step 104: Time series processing: The input data of each time step is processed through the Conv-LSTM layer; after each time step, the Conv-LSTM unit updates its hidden state and unit state and generates output; Step 105: Output integration: integrate the outputs of all time steps into a continuous sequence to form the final output of the Conv-LSTM layer; use the last item of the sequence as the final output, or further process the entire sequence.

6. The method according to claim 5, characterized in that The CBAM attention module aims to improve the ability to learn long-distance dependencies in signal processing; the design is as follows: Step 101: Channel attention mechanism: Perform global average pooling and global maximum pooling on the output of the Conv-LSTM layer to obtain the importance score of each channel; generate the weight of each channel through the fully connected layer and activation function, and then multiply these weights with the original feature map; Step 102: Spatial Attention Mechanism: Max pooling and average pooling are performed on the output of the Conv-LSTM layer to capture the importance of different positions; weights of spatial positions are generated through convolutional layers and activation functions, and then these weights are multiplied with the original feature map; Step 103 : Feature fusion: Fusion the results of channel attention and spatial attention; Step 104: feature standardization: standardize the features processed by the CBAM module to ensure the scale consistency and numerical stability of the features; Step 105: Feature propagation: pass the features processed by the CBAM module to the next layer of the network; directly use the features as the input of the next layer, or further enhance the propagation of features through residual connections, etc.; Step 106: Parameter optimization: Adjust the parameters in the CBAM module through back-propagation to improve the accuracy of feature selection, and update the weights in the CBAM module using optimization algorithms such as gradient descent.

7. Application of the bearing remaining service life estimation model based on Conv-LSTM and CBAM in bearing remaining service life estimation.

Images