A bearing remaining life prediction method based on frequency domain degradation sensing

Abstract

This invention discloses a bearing remaining life prediction method based on frequency domain degradation perception. It constructs an integrated prediction system through a time-frequency decomposition dual-branch feature extraction module, a frequency domain degradation perception weight allocation module, and a life prediction loss function optimization module. First, the time-frequency decomposition dual-branch feature extraction module decomposes the original time series into high and low frequencies and inputs it into an sLSTM and mLSTM dual-branch structure for feature extraction. Second, the frequency domain degradation perception weight allocation module performs degradation perception processing on the fused features, generating dynamic weights and modulating the features to obtain the final fused features, which are then mapped to the prediction space. Finally, the life prediction loss function optimization module calculates the loss, optimizes the model output, and obtains the bearing remaining life prediction result. This invention comprehensively covers the full-stage features of bearings from healthy to severely degraded, improves the sensitivity of early fault detection, and still possesses excellent generalization and robustness under complex operating conditions.

Description

Technical Field

[0001] This invention pertains to bearing remaining life prediction methods, specifically employing a bearing remaining life prediction method based on frequency domain degradation sensing. Background Technology

[0002] Rolling bearings, as key supports in rotating machinery, play an indispensable role in fields such as rail transportation, wind power generation, and precision manufacturing. Their failure often leads to overall equipment failure, causing production interruptions, safety hazards, and even economic losses. Therefore, accurate prediction of the remaining useful life (RUL) of bearings has become a core requirement of predictive maintenance. Currently, RUL prediction methods are divided into model-based prediction methods and data-driven prediction methods. Model-based prediction methods construct RUL prediction models using theories such as materials mechanics and tribology. Due to the complex operating conditions of industrial equipment and nonlinear influencing factors such as data noise, and because model-based methods are often specific to particular situations and lack general applicability, they are generally challenging. Data-driven methods utilize machine learning or deep learning techniques to learn bearing degradation patterns from large amounts of historical operating data. Currently, traditional machine learning models combined with manually extracted time-domain and frequency-domain features are widely used, while deep learning models, with their powerful feature extraction capabilities, have become a research hotspot, demonstrating significant advantages in processing high-dimensional, non-stationary signals.

[0003] However, despite the progress made in existing bearing RUL prediction methods, several challenges remain, limiting their application in complex industrial scenarios. First, existing methods lack sensitivity in early degradation detection. Many time-series feature-based models ignore the dynamic changes in the frequency domain of vibration signals, failing to capture subtle signs such as frequency shifts and high-frequency energy enhancements, resulting in prediction delays and biases. Second, feature extraction capabilities are limited. Traditional design features or shallow networks struggle to handle non-stationary degradation processes, especially under load fluctuations or noise interference, exhibiting poor generalization performance. Furthermore, existing deep learning models suffer from high computational complexity, weak robustness to small sample data, and a lack of segmentation of degradation stages, failing to focus on a few critical failure stages. These limitations restrict their application in industrial scenarios.

[0004] Currently, one of the major challenges in bearing remaining life prediction is the lack of comprehensive feature extraction. Specifically, the uneven capture of features across the entire process from the healthy stage to the severe degradation stage leads to insufficient sensitivity of the prediction model to early, subtle degradation features, resulting in limited coverage. In particular, existing methods often neglect the weighting of feature extraction for different scenarios at each stage, weakening the model's adaptability to the entire degradation process. Furthermore, during feature extraction, the model faces complex and variable fault conditions, potentially leading to overfitting to noise or outliers. Two key engineering problems exist: (1) Limitations of feature extraction: During the feature extraction process, the extracted features cannot fully reflect the comprehensive features of the entire degradation state, making the prediction model's response to multi-stage degradation unbalanced. In particular, there is a lack of a weight allocation mechanism for different situations in each stage to dynamically modulate the feature extraction weights.

[0005] (2) Improve the generalization and robustness of the model: During model training, how to ensure that when extracting features, the model avoids overfitting to noise or outliers, generates more reliable predictions, and maintains the generalization and robustness of the model in the face of complex and ever-changing fault conditions. Summary of the Invention

[0006] To address the problems existing in the prior art, the purpose of this invention is to provide a bearing remaining life prediction method based on frequency domain degradation sensing, thereby solving the above problems and improving feature extraction capability and model generalization.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A bearing remaining life prediction method based on frequency domain degradation sensing is implemented using a time-frequency decomposition dual-branch feature extraction module, a frequency domain degradation sensing weight allocation module, and a life prediction loss function optimization module. The prediction method specifically includes the following steps: Step 1: Use the time-frequency decomposition dual-branch feature extraction module to decompose the original time series into high and low frequencies, and input it into the sLSTM and mLSTM dual-branch structure for feature extraction. After obtaining global and local features, feature fusion is performed. Step 2: Use the frequency domain degradation sensing weight allocation module to perform degradation sensing processing on the fused features, generate dynamic weights and modulate the features, and then map the final fused features to the prediction space. Step 3: Calculate the loss using the life prediction loss function optimization module, optimize the model output, and obtain the bearing remaining life prediction result.

[0008] The present invention further defines the technical solution as follows: Furthermore, the feature extraction process of the time-frequency decomposition dual-branch feature extraction module specifically includes: Step 1.1: The time-frequency decomposition dual-branch feature extraction module is used to first decompose the time series data into frequency components. The moving average decomposition method is applied to decompose the input time series data into high-frequency and low-frequency components. By calculating the average value of data points within a window, the fluctuations of the time series are smoothed, and the overall trend of the data is extracted. The low-frequency component represents the slowly changing trend in the original time series, and the high-frequency component is the residual obtained by subtracting the low-frequency component from the original time series. The formula for calculating the trend component is: ; This represents the trend component of the i-th feature at time step t. It is the value of the i-th feature of the original input data at the t-th time step, K is the window size, and M is the window radius; Step 1.2: Introduce learnable weights and fixed mixing coefficients to output the final low-frequency and high-frequency outputs. Apply the sigmoid activation function to the learnable parameters to generate low-frequency weights. The high-frequency weights are complementary to the low-frequency weights, and their sum is 1. The low-frequency and high-frequency outputs are then fused together. The calculation formula is as follows: ; It is the sigmoid activation function. These are learnable parameters. It is a low-frequency weight. It is a high-frequency weight; ; , These are low-frequency output and high-frequency output, respectively. It is the original time series with minimal injection; Step 1.3: Input the obtained low-frequency output into the mLSTM to capture global features with long-range dependencies. The mLSTM's memory state update formula is: ; It is the memory matrix at the current time step t. It is the Gate of Oblivion. It is the memory matrix of the previous time step. It's an input gate. It is a value vector. It is a key vector; Step 1.4: Input the obtained high-frequency output into the sLSTM to capture the local features of local fluctuations. The candidate vector of the sLSTM is calculated as follows: ; in It is the candidate vector of the input; Step 1.5: After passing through mLSTM and sLSTM, two feature vectors, global feature and local feature, are obtained respectively, and then feature fusion is performed.

[0009] Furthermore, the processing procedure of the frequency domain degradation sensing weight allocation module includes: Step 2.1: Perform feature transformation on the input local and global features so that the features of the two branches are in the same transformation space, and then concatenate the transformed local and global features; Step 2.2: Calculate the gating weights using a gating network with a multilayer perceptron structure; Step 2.3: After assigning weights, a degradation perception stage is performed. The fusion features are dynamically modulated according to the degradation stage to make the model more sensitive to different degradation stages. Step 2.4: Gated fusion features obtained By combining the dynamic weight d of stage perception, important feature information in different degradation stages can be given attention. While reducing the amount of computation by replacing the attention mechanism, the modulated information can be recombined to obtain a more stable feature representation. Step 2.5: Process the features through the output layer. Mapping to the prediction space yields the final prediction result.

[0010] Furthermore, in step 2.2, the gating weights are calculated using a gating network with a multilayer perceptron structure, specifically as follows: First, a linear layer is used to reduce the dimensionality of the concatenated feature vector. Then, the ReLU activation function is used to introduce non-linearity, enhancing the model's expressive power. Next, a Dropout layer prevents overfitting. Afterward, a linear layer is used for feature projection, corresponding to the weights of the local and global branches respectively. Finally, a Softmax function is used for normalization to ensure the sum of the weights is 1. After obtaining the weights, gated fusion calculation is performed, outputting the fused feature vector. The calculation formula is as follows: ; , These are the transformed local and global features. , These are the local and global feature weights.

[0011] Furthermore, step 2.3 specifically involves: identifying the probability of degradation stages based on frequency characteristics, calculating five frequency indicators—amplitude variance, frequency shift, spectral kurtosis, harmonic ratio, and high-frequency energy—and generating specific weights for each stage. Then, the final dynamic weights are calculated through a weighted summation operation, specifically a probability-weighted tensor summation, to ensure that the weights reflect the current degradation distribution; wherein: The formula for calculating amplitude variance is: ; For time series calculation The mean of the signal is calculated to quantify the fluctuation range of the signal amplitude and reflect the instability of the signal. The formula for calculating frequency offset is: ; Frequency offset is used to detect the dominant frequency drift of a signal. It is the instantaneous frequency at the current moment. It is the reference initial frequency; The formula for calculating spectral kurtosis is: ; Spectral kurtosis measures the sharpness of peaks in a spectral distribution; high values ​​indicate impulsive characteristics or non-Gaussian noise. It is a time-frequency representation. It is the power spectrum. It is a fourth-order moment; The formula for calculating the harmonic ratio is: ; The harmonic ratio is used to assess the proportion of harmonic components relative to the fundamental frequency, revealing periodic degradation or distortion. It is the fundamental frequency amplitude. It is the amplitude of the kth harmonic; The formula for calculating high-frequency energy is: ; High-frequency energy is used to quantify the energy accumulation in the high-frequency range, which can detect fault symptoms. arrive It is a high-frequency band. It is the spectrum; The five frequency indices are processed as scalars on a one-dimensional basis, and the calculation formula is as follows: ; With corresponding bias These are learnable linear parameters, which are concatenated to obtain a feature vector s. The obtained s is input into the degradation stage classifier to obtain the probability distribution of the normal operation stage, the early degradation stage, and the severe degradation stage. After obtaining the probability distribution, a weight vector is generated for each stage by two layers of perceptrons. The weights are then expected and weighted according to the stage probability to obtain the final dynamic weights.

[0012] Furthermore, in step 2.5, the final fused features are mapped to the prediction space through the output layer. The specific process for calculating the output layer is as follows: , These are the weight parameters and bias parameters of the two linear mappings in the output layer.

[0013] Furthermore, the loss calculation process of the lifetime prediction loss function optimization module includes: Step 3.1: Use MAE to constrain the point-to-point error of the current batch, and use the batch mean as the single-point representative of the training step. The calculation formula is as follows: ; Where B is the batch size and O is the output dimension. It is a prediction; Step 3.2: Smooth the remaining lifetime prediction curve and reduce non-physical inflection points through first-order and second-order variational constraints. The first-order variational constraint suppresses amplitude jumps between adjacent training steps. The calculation formula is as follows: ; The second-order suppression predicts changes in curvature and acceleration, calculated using the following formula: ; It refers to the length of history; Step 3.3: Construct the complete total loss function. The total loss function is a combination of first-order and second-order variational losses. The complete expression for the total variational loss is as follows: ; in, It is the mean absolute error loss. It is the weighting coefficient.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention employs a time-frequency decomposition dual-branch feature extraction module to ensure comprehensive capture of features across all stages from healthy to severely degraded states. Simultaneously, a frequency domain degradation-aware weight allocation module modulates feature weights for each degradation stage through frequency characteristic stage identification and weight allocation mechanisms, mitigating the limitation of insufficient sensitivity to early, subtle signs and ensuring balanced feature extraction across multiple degradation stages. Furthermore, a lifetime prediction loss function optimization module reduces noise interference and overfitting issues related to outliers. The weight allocation in the frequency domain degradation-aware weight allocation module further enhances the model's generalization ability, while also improving its adaptability and stability. This invention effectively alleviates the key engineering problems of imbalanced feature extraction and insufficient model generalization by synergistically employing a time-frequency decomposition dual-branch feature extraction module, a frequency domain degradation perception weight allocation module, and a lifetime prediction loss function optimization module. This method comprehensively captures multi-stage features from healthy to severely degraded states through high- and low-frequency decomposition and stage weight allocation, improving the feature extraction capability of early degradation features. Simultaneously, the lifetime prediction loss function optimization module enhances the model's robustness and generalization ability, providing a new solution for industrial predictive maintenance. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method framework of the present invention; Figure 2 This is a graph showing the RUL prediction results of some bearings in the XJTU-SY bearing dataset of this invention; Figure 3 This is a graph showing the RUL prediction results of a portion of the bearings in the PHM2012 bearing dataset of this invention. Detailed Implementation

[0016] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0017] Based on existing artificial intelligence and deep learning technologies, this invention proposes solutions to two engineering problems faced by existing bearing remaining service life prediction methods: limitations in feature extraction and the need to improve model generalization and robustness. These solutions effectively address the aforementioned issues and enhance feature extraction capabilities and model generalization.

[0018] This invention discloses a bearing remaining life prediction method based on frequency domain degradation sensing. It employs a time-frequency decomposition dual-branch feature extraction module, a frequency domain degradation sensing weight allocation module, and a life prediction loss function optimization module. The time-frequency decomposition dual-branch feature extraction module first separates the input time series signal into low-frequency trend components and high-frequency fluctuation components using the moving average decomposition method. These components are then input into the local branch based on sLSTM and the global branch based on mLSTM, respectively, for feature extraction. This module enables complementary modeling of multi-scale time series data. The low-frequency branch captures long-term trends, while the high-frequency branch captures short-term anomalies, avoiding information confusion and loss caused by traditional single-path processing. The frequency domain degradation perception weight allocation module, based on fused features, calculates indicators such as amplitude variance, frequency shift, spectral kurtosis, harmonic ratio, and high-frequency energy through frequency analysis to identify degradation stages and generate stage weights to modulate the feature vector. This module dynamically adjusts the feature importance according to the degradation level, achieving adaptive weight allocation.

[0019] Example 1 like Figure 1 As shown, this embodiment provides a bearing remaining life prediction method based on frequency domain degradation sensing. The prediction method specifically includes the following steps: Step 1: Use the time-frequency decomposition dual-branch feature extraction module to decompose the original time series into high and low frequencies, and input it into the sLSTM and mLSTM dual-branch structure for feature extraction. After obtaining global and local features, feature fusion is performed. The feature extraction process of the time-frequency decomposition dual-branch feature extraction module specifically includes: Step 1.1: The time-frequency decomposition dual-branch feature extraction module is used to first decompose the time series data into frequency components. The moving average decomposition method is applied to decompose the input time series data into high-frequency and low-frequency components. By calculating the average value of data points within a window, the fluctuations of the time series are smoothed, and the overall trend of the data is extracted. The low-frequency component represents the slowly changing trend in the original time series, and the high-frequency component is the residual obtained by subtracting the low-frequency component from the original time series. The formula for calculating the trend component is: ; This represents the trend component of the i-th feature at time step t. It is the value of the i-th feature of the original input data at the t-th time step, K is the window size, and M is the window radius; Step 1.2: After obtaining the low-frequency components of the low-frequency part, the high-frequency components are obtained by subtracting the low-frequency components from the original time series to obtain the residual. Then, by introducing learnable weights and fixed mixing coefficients, the final low-frequency and high-frequency outputs are output. The sigmoid activation function is applied to the learnable parameters to generate low-frequency weights, and the high-frequency weights are complementary to the low-frequency weights, with the sum of the two being 1. The low-frequency and high-frequency outputs are then fused, and the calculation formula is as follows: ; It is the sigmoid activation function. These are learnable parameters. It is a low-frequency weight. It is a high-frequency weight; ; , These are low-frequency output and high-frequency output, respectively. It is the original time series with minimal injection; Step 1.3: Input the obtained low-frequency output into the mLSTM to capture global features with long-range dependencies. The mLSTM's memory state update formula is: ; It is the memory matrix at the current time step t. It is the Gate of Oblivion. It is the memory matrix of the previous time step. It's an input gate. It is a value vector. It is a key vector; Step 1.4: Input the obtained high-frequency output into the sLSTM to capture the local features of local fluctuations. The candidate vector of the sLSTM is calculated as follows: ; in It is the candidate vector of the input; Step 1.5: After passing through mLSTM and sLSTM, two feature vectors are obtained: global features and local features, respectively. Then, feature fusion is performed. Step 2: The frequency domain degradation-aware weight allocation module performs degradation-aware processing on the fused features, complementing and integrating local and global features, and introducing a degradation-aware mechanism to improve accuracy in time series tasks. The processing steps of the frequency domain degradation-aware weight allocation module include: Step 2.1: Perform feature transformation on the input local and global features so that the features of the two branches are in the same transformation space, and then concatenate the transformed local and global features; Step 2.2: Gating weights are calculated using a gated network with a multilayer perceptron structure. First, a linear layer is used to reduce the dimensionality of the concatenated feature vector. Then, the ReLU activation function is used to introduce non-linearity to enhance the model's expressive power. Next, a Dropout layer is used to prevent overfitting. After that, the linear layer is used for feature projection, corresponding to the weights of the local and global branches respectively. Finally, the Softmax function is used for normalization to ensure that the sum of the weights is 1. After obtaining the weights, gated fusion calculation is performed, and the fused feature vector is output. The calculation formula is as follows: ; , These are the transformed local and global features. , These are the weights of local and global features; Step 2.3: After weight allocation, a degradation perception phase is performed. Based on the dynamic modulation fusion characteristics of the degradation phase, the model becomes more sensitive to different degradation phases. This process first identifies the probability of the degradation phase based on frequency characteristics, calculates five frequency indicators—amplitude variance, frequency shift, spectral kurtosis, harmonic ratio, and high-frequency energy—and generates specific weights for each phase. Then, the final dynamic weights are calculated through a weighted tensor summation operation, specifically a probability-weighted approach, to ensure that the weights reflect the current degradation distribution. The formula for calculating amplitude variance is: ; For time series calculation The mean of the signal is calculated to quantify the fluctuation range of the signal amplitude and reflect the instability of the signal. The formula for calculating frequency offset is: ; Frequency offset is used to detect the dominant frequency drift of a signal. It is the instantaneous frequency at the current moment. It is the reference initial frequency; The formula for calculating spectral kurtosis is: ; Spectral kurtosis measures the sharpness of peaks in a spectral distribution; high values ​​indicate impulsive characteristics or non-Gaussian noise. It is a time-frequency representation. It is the power spectrum. It is a fourth-order moment; The formula for calculating the harmonic ratio is: ; The harmonic ratio is used to assess the proportion of harmonic components relative to the fundamental frequency, revealing periodic degradation or distortion. It is the fundamental frequency amplitude. It is the amplitude of the kth harmonic; The formula for calculating high-frequency energy is: ; High-frequency energy is used to quantify the energy accumulation in the high-frequency range, which can detect fault symptoms. arrive It is a high-frequency band. It is the spectrum; The five frequency indices are processed as scalars on a one-dimensional basis, and the calculation formula is as follows: ; With corresponding bias These are learnable linear parameters, which are concatenated to obtain the feature vector s. Inputting the obtained s into the degradation stage classifier yields the probability distribution of the corresponding degradation stage, as shown in the formula: ; It is a stage classifier with three degradation stages: normal operation stage, early degradation stage, and severe degradation stage. After obtaining the probability distribution, each stage generates a weight vector using a two-layer perceptron. This weight vector is then weighted according to the expected probability of each stage to obtain the final dynamic weights, i.e.: ; It is the dynamic weight vector of the k-th stage. These are the parameters of the two-layer perceptron in the k-th stage;

[0020] d is the final dynamic weight that is a probability-weighted average of the weights at each stage, ensuring that the weights reflect the current degradation distribution; Step 2.4: Gated fusion features obtained By combining the dynamic weights d of stage awareness, important feature information in different degradation stages is given attention. While reducing computational cost by replacing the attention mechanism, the modulated information is recombinated to obtain a more stable feature representation. Specifically: First, stage-aware modulation is performed, and the modulation features are obtained by element-wise multiplication, expressed as: ; By using two fully connected layers, fusion projection is performed to integrate the modulated features, resulting in the final fused feature, expressed as:

[0021] These are the linear mapping weight parameters and bias parameters of the two fully connected layers, respectively; Step 2.5: Process the features through the output layer. Mapping to the prediction space yields the final prediction result. The specific process for calculating the output layer is as follows: ; These are the two linear mapping weights and biases of the output layer; Step 3: Calculate the loss using the life prediction loss function optimization module, optimize the model output, and obtain the bearing's remaining life prediction result. The loss calculation process of this life prediction loss function optimization module includes: Step 3.1: Use MAE to constrain the point-to-point error of the current batch, and use the batch mean. As a single-point representative for the training step, used for the total variational loss, MAE directly constrains the point-to-point error of the current batch. Its advantage is that it is more stable in the face of outliers and is not affected by outliers like MSE. Meanwhile, As a single point representative, this step of the model transforms the overall predicted trend into a scalar, improving stability. The calculation formula is as follows:

[0022] Where B is the batch size and O is the output dimension. It is a prediction; Step 3.2: Smooth the remaining lifetime prediction curve and reduce non-physical inflection points through first-order and second-order variational constraints. The first-order variational constraint suppresses amplitude jumps between adjacent training steps. The calculation formula is as follows: ; The second-order suppression predicts changes in curvature and acceleration, calculated using the following formula: ; It is the length of history. Through the combination of two orders, the curve of remaining service life prediction can be improved from the state under the influence of noise to a smooth and gradual degradation form. Step 3.3: Construct the complete total loss function. The total loss function combines first-order and second-order variational loss, successfully improving the model's stability and robustness in time series prediction. This loss function effectively suppresses unnecessary time series fluctuations, making the model smoother during prediction and avoiding excessive changes and unreasonable jumps. The complete expression for the total variational loss is as follows: ; in, It is the mean absolute error loss. It is the weighting coefficient.

[0023] This embodiment utilizes a time-frequency decomposition dual-branch feature extraction module, which separates high- and low-frequency signals through moving average decomposition. Then, it uses mLSTM to capture long-term trends and sLSTM to capture local fluctuations, comprehensively covering features from healthy to severely degraded stages, avoiding information loss caused by single-path processing. Combined with the stage recognition and dynamic weighting mechanism of the frequency domain degradation perception weight allocation module, it assigns differentiated weights to each stage of normal operation, early degradation, and severe degradation, especially strengthening the capture of early weak degradation signs (such as frequency shift and high-frequency energy enhancement), solving the problems of unbalanced feature extraction and delayed early fault response in traditional methods. The introduction of the Dropout layer in the gated network in this embodiment effectively prevents overfitting; the lifetime prediction loss function optimization module uses MAE loss to deal with outliers, and combines first-order and second-order variational constraints to suppress signal noise and non-physical inflection points, making the prediction curve smoother and more stable; the frequency domain degradation-aware weight allocation module improves the model's generalization ability under complex working conditions such as load fluctuations and noise interference based on the importance of dynamic modulation features of frequency characteristics, and avoids the model being misled by noise or outliers.

[0024] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A method for predicting the remaining life of bearings based on frequency domain degradation sensing, characterized in that: The prediction method employs a time-frequency decomposition dual-branch feature extraction module, a frequency domain degradation-aware weight allocation module, and a lifetime prediction loss function optimization module. The specific steps include: Step 1: Use the time-frequency decomposition dual-branch feature extraction module to decompose the original time series into high and low frequencies, and input it into the sLSTM and mLSTM dual-branch structure for feature extraction. After obtaining global and local features, feature fusion is performed. Step 2: Use the frequency domain degradation sensing weight allocation module to perform degradation sensing processing on the fused features, generate dynamic weights and modulate the features, and then map the final fused features to the prediction space. Step 3: Calculate the loss using the life prediction loss function optimization module, optimize the model output, and obtain the bearing remaining life prediction result.

2. The bearing remaining life prediction method based on frequency domain degradation sensing according to claim 1, characterized in that: The feature extraction process of the time-frequency decomposition dual-branch feature extraction module specifically includes: Step 1.1: The time-frequency decomposition dual-branch feature extraction module is used to first decompose the time series data into frequency components. The moving average decomposition method is applied to decompose the input time series data into high-frequency and low-frequency components. By calculating the average value of data points within a window, the fluctuations of the time series are smoothed, and the overall trend of the data is extracted. The low-frequency component represents the slowly changing trend in the original time series, and the high-frequency component is the residual obtained by subtracting the low-frequency component from the original time series. The formula for calculating the trend component is: ； This represents the trend component of the i-th feature at time step t. It is the value of the i-th feature of the original input data at the t-th time step, K is the window size, and M is the window radius; Step 1.2: Introduce learnable weights and fixed mixing coefficients to output the final low-frequency and high-frequency outputs. Apply the sigmoid activation function to the learnable parameters to generate low-frequency weights. The high-frequency weights are complementary to the low-frequency weights, and their sum is 1. The low-frequency and high-frequency outputs are then fused together. The calculation formula is as follows: ; It is the sigmoid activation function. These are learnable parameters. It is a low-frequency weight. It is a high-frequency weight; ; , These are low-frequency output and high-frequency output, respectively. It is the original time series with minimal injection; Step 1.3: Input the obtained low-frequency output into the mLSTM to capture global features with long-range dependencies. The memory state update formula for the mLSTM is: ； It is the memory matrix at the current time step t. It is the Gate of Oblivion. It is the memory matrix of the previous time step. It's an input gate. It is a value vector. It is a key vector; Step 1.4: Input the obtained high-frequency output into the sLSTM to capture the local features of local fluctuations. The candidate vector of the sLSTM is calculated as follows: ； in It is the candidate vector of the input; Step 1.5: After passing through mLSTM and sLSTM, two feature vectors, global feature and local feature, are obtained respectively, and then feature fusion is performed.

3. The bearing remaining life prediction method based on frequency domain degradation sensing according to claim 1, characterized in that: The processing procedure of the frequency domain degradation sensing weight allocation module includes: Step 2.1: Perform feature transformation on the input local and global features so that the features of the two branches are in the same transformation space, and then concatenate the transformed local and global features; Step 2.2: Calculate the gating weights using a gating network with a multilayer perceptron structure; Step 2.3: After assigning weights, a degradation perception stage is performed. The fusion features are dynamically modulated according to the degradation stage to make the model more sensitive to different degradation stages. Step 2.4: Gated fusion features obtained By combining the dynamic weight d of stage perception, important feature information in different degradation stages can be given attention. While reducing the amount of computation by replacing the attention mechanism, the modulated information can be recombined to obtain a more stable feature representation. Step 2.5: Process the features through the output layer. Mapping to the prediction space yields the final prediction result.

4. The bearing remaining life prediction method based on frequency domain degradation sensing according to claim 3, characterized in that: In step 2.2, the gating weights are calculated using a gating network with a multilayer perceptron structure, specifically as follows: First, a linear layer is used to reduce the dimensionality of the concatenated feature vector. Then, the ReLU activation function is used to introduce non-linearity, enhancing the model's expressive power. Next, a Dropout layer prevents overfitting. Afterward, a linear layer is used for feature projection, corresponding to the weights of the local and global branches respectively. Finally, a Softmax function is used for normalization to ensure the sum of the weights is 1. After obtaining the weights, gated fusion calculation is performed, outputting the fused feature vector. The calculation formula is as follows: ； , These are the transformed local and global features. , These are the local and global feature weights.

5. The bearing remaining life prediction method based on frequency domain degradation sensing according to claim 3, characterized in that: The specific process of step 2.3 is as follows: Based on frequency characteristics, the probability of degradation stages is identified; five frequency indicators are calculated: amplitude variance, frequency shift, spectral kurtosis, harmonic ratio, and high-frequency energy; and specific weights for each stage are generated. Then, the final dynamic weights are calculated through weighted summation, specifically a probability-weighted tensor summation operation, to ensure that the weights reflect the current degradation distribution; wherein: The formula for calculating amplitude variance is: ； For time series calculation The mean of the signal is calculated to quantify the fluctuation range of the signal amplitude and reflect the instability of the signal. The formula for calculating frequency offset is: ; Frequency offset is used to detect the dominant frequency drift of a signal. It is the instantaneous frequency at the current moment. It is the reference initial frequency; The formula for calculating spectral kurtosis is: ; Spectral kurtosis measures the sharpness of peaks in a spectral distribution; high values ​​indicate impulsive characteristics or non-Gaussian noise. It is a time-frequency representation. It is the power spectrum. It is a fourth-order moment; The formula for calculating the harmonic ratio is: ; The harmonic ratio is used to assess the proportion of harmonic components relative to the fundamental frequency, revealing periodic degradation or distortion. It is the fundamental frequency amplitude. It is the amplitude of the kth harmonic; The formula for calculating high-frequency energy is: ; High-frequency energy is used to quantify the energy accumulation in the high-frequency range, which can detect fault symptoms. arrive It is a high-frequency band. It is the spectrum; The five frequency indices are processed as scalars on a one-dimensional basis, and the calculation formula is as follows: ; With corresponding bias These are learnable linear parameters, which are concatenated to obtain a feature vector s. The obtained s is input into the degradation stage classifier to obtain the probability distribution of the normal operation stage, the early degradation stage, and the severe degradation stage. After obtaining the probability distribution, a weight vector is generated for each stage by two layers of perceptrons. The weights are then expected and weighted according to the stage probability to obtain the final dynamic weights.

6. The vibration detection device for rail transit environment according to claim 3, characterized in that: In step 2.5, the final fused features are mapped to the prediction space through the output layer. The specific process for calculating the output layer is as follows: ， These are the weight parameters and bias parameters of the two linear mappings in the output layer.

7. The bearing remaining life prediction method based on frequency domain degradation sensing according to claim 1, characterized in that: The loss calculation process of the lifetime prediction loss function optimization module includes: Step 3.1: Use MAE to constrain the point-to-point error of the current batch, and use the batch mean as the single-point representative of the training step. The calculation formula is as follows: ; Where B is the batch size and O is the output dimension. It is a prediction; Step 3.2: Smooth the remaining lifetime prediction curve and reduce non-physical inflection points through first-order and second-order variational constraints. The first-order variational constraint suppresses amplitude jumps between adjacent training steps. The calculation formula is as follows: ； The second-order suppression predicts changes in curvature and acceleration, calculated using the following formula: ； It refers to the length of history; Step 3.3: Construct the complete total loss function. The total loss function is a combination of first-order and second-order variational losses. The complete expression for the total variational loss is as follows: ; in, It is the mean absolute error loss. It is the weighting coefficient.

Images