Bearing residual life prediction method based on frequency band interpretability guide mechanism

Abstract

The invention discloses a bearing residual life prediction method based on a frequency band interpretability guide mechanism, and the method comprises the following steps: carrying out the continuous wavelet transformation of an original bearing vibration signal, and extracting a multi-scale time-frequency feature with a clear center frequency; training a one-dimensional convolutional neural network model integrating an input channel attention mechanism and a regression prediction structure based on the constructed multi-channel input tensor; outputting a response gradient to each input channel by using a back propagation calculation model, and further obtaining a channel importance score; and taking the channel importance score as static guidance information, embedding the channel importance score into an attention mechanism in a model training process, constructing a regularization constraint term, dynamically regulating and controlling attention distribution by introducing an additional loss function, and guiding the attention mechanism to focus on a high-sensitivity area in combination with the frequency band importance score, so as to achieve the purpose of high sensitivity. Therefore, the physical constraint of the model in the training process is improved.

Description

Technical Field

[0001] This invention relates to the field of bearing remaining life prediction, and specifically to a bearing remaining life prediction method based on a frequency band interpretability guidance mechanism. Background Technology

[0002] As a critical component of rotating machinery, the operating condition of bearings directly affects equipment safety and stability. Bearing failure is one of the main causes of unplanned downtime and production losses in industrial equipment. Accurately predicting the remaining service life (RUL) of bearings is of great significance for achieving predictive maintenance, optimizing maintenance strategies, and reducing operation and maintenance costs.

[0003] Existing Life Utilization (RUL) prediction methods mainly include physical model-based and data-driven methods. Physical model-based methods predict lifespan by establishing mathematical models of degradation mechanisms, which have a certain degree of interpretability, but are highly dependent on operating conditions and lack generalizability. Data-driven methods, especially deep learning techniques, learn from large amounts of monitoring data and automatically extract degradation features, demonstrating good prediction accuracy and generalization ability.

[0004] Deep learning methods, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), can extract complex time-series features from raw vibration signals and achieve end-to-end prediction. However, existing technologies have two main shortcomings: First, the models lack interpretability. As a "black box," deep models are difficult to reveal the basis for predictions and cannot clearly reflect key frequency bands or physical fault characteristic frequencies, thus limiting their trust and optimization in engineering applications.

[0005] Second, the attention mechanism lacks guidance from physical knowledge. Current attention weights rely entirely on data-driven approaches, making them susceptible to noise and spurious correlations. This makes it difficult to stably focus on the true key frequency bands, affecting feature extraction and prediction performance and reducing the model's generalization ability.

[0006] In existing technologies, Huan Wang et al. proposed a multi-layer wavelet attention convolutional neural network (MWA-CNN) in 2023 (a multi-layer wavelet-based interpretable convolutional neural network for noise-resistant mechanical fault diagnosis). By introducing discrete wavelet transform into the convolutional neural network, it was used for mechanical fault diagnosis, achieving improved feature extraction and diagnostic performance in noisy environments. This method progressively filters noise in different frequency bands through multi-layer wavelet decomposition and applies attention to each frequency band to enhance its feature extraction ability, thereby improving the interpretability of the model. However, this method still has shortcomings: on the one hand, although wavelet decomposition is used to process frequency domain features, the overall interpretability of the model is still insufficient, making it difficult to intuitively reflect the degree of attention and prediction basis of the model in different frequency bands; on the other hand, its attention mechanism mainly relies on data-driven approaches, lacks guidance from physical knowledge, and is susceptible to noise or spurious correlation interference, further limiting its interpretability. Summary of the Invention

[0007] The purpose of this invention is to provide a bearing remaining service life prediction method with frequency band-level interpretability analysis capabilities and optimized prediction performance based on a gradient-guided attention mechanism. This method includes performing Morlet wavelet transform on the original vibration signal to extract multi-scale time-frequency features, constructing a multi-channel frequency band tensor, training a one-dimensional convolutional neural network with an attention mechanism based on this tensor, and calculating the sensitivity score for each frequency band using gradient backpropagation; combining bearing structural parameters and fault frequencies to analyze the correlation between the model's focus area and physical faults; embedding the channel scores as guiding signals into the attention regularization term to encourage the model to focus on key frequency bands, improving robustness and generalization ability; and visually verifying the consistency of attention and gradient contribution distributions to enhance the model's interpretability and effectiveness. This method, by combining wavelet transform, gradient contribution analysis, frequency matching analysis, and a guided attention mechanism, enhances the model's interpretability and performance stability, and is particularly suitable for modeling and predicting multi-frequency bearing vibration signals.

[0008] The present invention is achieved by at least one of the following technical solutions.

[0009] A bearing remaining life prediction method based on frequency band interpretability guidance mechanism includes the following steps: inputting the test data into a trained bearing remaining life prediction model and outputting the bearing remaining life prediction result; The training of the bearing remaining life prediction model includes the following steps: S1. Extract multi-scale time-frequency features of the original vibration signal and construct a multi-channel frequency band tensor; train a one-dimensional convolutional neural network model based on the multi-channel frequency band tensor; S2. Evaluate the impact of different frequency band characteristics on the bearing remaining life prediction task from the gradient level, and obtain the channel importance score vector; S3. Match the center frequency of the frequency band corresponding to the top-ranked channel importance score vector with the typical fault characteristic frequency, and use the matched score as attention-guided constraint to retrain the model, finally outputting the bearing remaining life prediction result.

[0010] Furthermore, in step S1, the original bearing vibration signal is subjected to continuous wavelet transform, and the Morlet wavelet is selected as the mother wavelet function to extract multi-scale time-frequency features with a clear center frequency.

[0011] Furthermore, the one-dimensional convolutional neural network model is a one-dimensional convolutional neural network model with a meaning channel attention mechanism and a regression prediction structure.

[0012] Further, in step S2, after the one-dimensional convolutional neural network model is trained, gradient backpropagation is used to calculate the sensitivity score of each frequency band, and the absolute value of the gradient of the prediction loss function with respect to each channel of the input signal is calculated to obtain the response intensity of each channel on each sample. The responses of all samples and channels are normalized to obtain the final channel importance score vector.

[0013] Further, in step S2, the normalized channel gradient weights obtained after training are used as guiding signals and introduced into the channel attention mechanism of the model, and a guiding regularization term is added to the composite loss function; the expression of the composite loss function is as follows: ,in The loss is the mean square error prediction loss. Regularization terms to guide attention is the regularization coefficient.

[0014] Furthermore, the typical fault characteristic frequencies include the inner ring fault frequency, outer ring fault frequency, and rolling element spin frequency, calculated based on the bearing's inner diameter, outer diameter, number of rolling elements, and diameter.

[0015] Furthermore, when the error value is less than or equal to the matching tolerance threshold, it indicates that the frequency band that was successfully matched is faulty.

[0016] The system for implementing the bearing remaining life prediction method based on a frequency band interpretability guidance mechanism includes the following modules: The data preprocessing module is used to extract multi-scale frequency domain features of vibration signals; The channel sensitivity analysis module is used to generate frequency band importance scores and sort them in descending order using a one-dimensional convolutional neural network model; The physical feature matching and verification module is used to compare the center frequency corresponding to the high-ranking frequency band channel with the typical fault characteristic frequency and harmonic frequency calculated from the bearing physical structure parameters, and to establish the mapping relationship between the key frequency band and the failure modes of the inner ring, outer ring, rolling elements and cage. Attention-guided training module, used to train one-dimensional convolutional neural network models.

[0017] A computer device according to the present invention includes a memory and a processor, the memory being electrically connected to the processor, the memory storing a computer program, which, when executed by the processor, causes the processor to implement the method described herein.

[0018] The present invention provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor implements the method described herein.

[0019] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention realizes gradient-based importance analysis and has frequency band-level interpretability. By calculating the gradient response of the prediction loss to the input channel after the model training is completed, and normalizing it as the channel contribution, the model's attention to signals in different frequency bands is clearly quantified. Since each channel in the model input corresponds to a specific frequency band after wavelet decomposition, the gradient contribution corresponds to the importance of each frequency band in the prediction task, realizing frequency band-level interpretability and providing a reliable basis for subsequent model tuning and anomaly analysis.

[0020] (2) This invention combines physical priors to conduct frequency matching analysis, enhancing interpretability and physical consistency. It utilizes the well-defined center frequency characteristics of Morlet wavelets and combines bearing structural parameters (such as rotational speed, number of rolling elements, contact angle, etc.) to calculate the center frequency of each channel. The frequency bands with the highest gradient importance are matched with theoretical fault frequencies. If the error is less than a set threshold, it indicates that the model has focused on potential physical fault characteristic frequencies, thus effectively verifying the model's focus on key frequency responses and enhancing the physical rationality of the model's prediction results.

[0021] (3) This invention proposes a gradient-guided attention mechanism to improve model performance and robustness. The gradient contribution is used as the guiding signal of the Attention module, and an attention regularization loss term is constructed to enable the model to maintain the stability of the attention distribution while focusing on high contribution frequency bands, thereby alleviating the problem of attention drift or over-concentration. While ensuring interpretability, the model's generalization ability and prediction accuracy under complex data are improved. Attached Figure Description

[0022] Figure 1 This is a flowchart of the RUL prediction method based on the frequency band interpretability guidance mechanism, as an example.

[0023] Figure 2The training sample Bearing1_1 is shown as a diagram of the horizontal vibration signal throughout its entire life cycle, as used in the example.

[0024] Figure 3 The example shows the loss descent curves for the training and validation sets.

[0025] Figure 4 The following is a graph showing the changes in four evaluation indicators in the example.

[0026] Figure 5 This is a framework diagram for bearing RUL prediction based on a frequency band interpretability guidance mechanism.

[0027] Figure 6 This is a partial test set of bearing prediction curves. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to examples, but the embodiments of the present invention are not limited to these examples.

[0029] like Figures 1-5 As shown in this embodiment, a bearing remaining life prediction method based on a frequency band interpretability guidance mechanism includes the following steps: inputting the test data into a trained bearing remaining life prediction model and outputting the bearing remaining life prediction result. The training of the bearing remaining life prediction model involves constructing multi-frequency band inputs from bearing vibration signals, combining channel sensitivity analysis, physical fault feature matching, and attention-guided training to achieve focusing on key frequency regions and accurate prediction of remaining life. Specifically, the method includes the following steps: 1. Frequency Band Decoupling Wavelet Preprocessing: The original bearing vibration signal is subjected to continuous wavelet transform using Morlet wavelets to extract time-frequency coefficients for different frequency bands. The wavelet coefficients of each frequency band are mapped to independent channels of the model input, preserving multi-frequency information while avoiding feature aliasing between frequency bands and providing a certain degree of noise suppression. This design provides a foundation for subsequent frequency band-level interpretability analysis.

[0030] This process is based on continuous wavelet transform (CWT) to achieve time-frequency characterization of the signal, and uses the relationship between wavelet scale and frequency to divide frequency band channels.

[0031] Suppose the input original bearing vibration signal is a one-dimensional time-series signal tensor. ,in Indicates the batch size of the signal. The time series length of each sample is represented by the Morlet wavelet function as the mother wavelet. Its definition is as follows:

[0032] in The center angular frequency of the wavelet; This represents the scaling parameter, used to control the scaling of the wavelet function; It is a time variable, representing the first... Each time step is used to describe the position of the signal in the time dimension. For the signal... Applying the CWT transformation yields:

[0033] in, This is a scale parameter, inversely proportional to frequency; For translation parameters, Represent the set of real numbers, and specify the translation parameters. It can take any real value; Indicates complex conjugation. Indicated in scale Translation The time-frequency coefficients below.

[0034] To extract a fixed number of representative frequency bands, C scales are defined. These correspond to C center frequencies respectively. Based on this, importance scores for each frequency band channel are obtained through model pre-training. These frequencies are sampled uniformly or densely near typical bearing failure frequencies to cover key physical response regions. Each scale After transformation, it corresponds to a frequency band channel, and the final time-frequency coefficient tensor is: That is, each sample is mapped to a value after CWT. The two-dimensional signal with multiple frequency channels retains the original time series structure while introducing resolution capabilities in the frequency structure.

[0035] Center frequency of each channel It can be calculated using the Morlet wavelet center frequency formula:

[0036] in, Sampling frequency, The corresponding scale is given. This formula clarifies the central response region of each channel in the frequency domain, enabling subsequent models to match the physical response between the frequency band channels and the bearing's fault characteristic frequencies, thus facilitating the introduction of a physical interpretability mechanism into the model.

[0037] 2. Channel Importance Analysis: After model training, the sum of the absolute values ​​of the gradients of the prediction loss function with respect to each channel of the input signal is calculated, serving as the contribution weights for the corresponding channels. Since each channel corresponds to a specific frequency band, these gradient weights can be used to measure the impact of different frequency bands on the prediction results, achieving frequency band-level model interpretability.

[0038] After completing the initial model training, to obtain the model's attention to different frequency band channels and to explore the sensitivity of the model's prediction results to each input channel at the gradient level, the impact of different frequency band features on the bearing remaining life prediction task is further evaluated. Since each channel corresponds to a specific frequency band, this gradient weight can be used to measure the impact of different frequency bands on the prediction results, achieving frequency band-level model interpretability; the multi-frequency band input tensor contains multiple independent frequency band channels, with the center frequencies of each channel set at equal intervals, the frequency ranges not overlapping and continuously covering the effective bandwidth of the signal. The input tensor... Set to a differentiable state, use backpropagation to target the model's loss function on the training set. Perform one forward and backward propagation, and calculate the gradient of the loss with respect to the input tensor:

[0039] To measure the influence of each channel on the prediction result, we first take the absolute value of the gradient and sum it along the time dimension to obtain the response strength of each channel on each sample:

[0040] in, Indicates the first The sample at the th The frequency band channel and the first The gradient values ​​at each time step are then normalized to obtain the final channel importance score vector. The definition is as follows:

[0041] in, Indicates the first The importance score of each frequency band channel indicates that the higher the value, the more significant the contribution of that channel to the model prediction. Indicates the first The sample at the th The response strength on each frequency band channel.

[0042] Final channel importance score vector This reflects the attention intensity of each frequency band channel automatically extracted from historical training data by the neural network in the task of predicting the remaining life of bearings. Since this score is derived from the gradient of the loss function and has both directionality and causality, it can serve as an important reference for subsequent model attention mechanism design and frequency interpretability analysis.

[0043] 3. Fault Frequency Matching Analysis: Utilizing the well-defined center frequency characteristics of Morlet wavelets and combining them with bearing structural parameters (such as the number of rolling elements, rotational frequency, contact angle, etc.), the theoretical center frequency of each channel is calculated. The channels with the highest contribution are matched with typical fault frequencies (such as outer race fault frequency, inner race fault frequency, etc.). If the frequency difference is within the allowable threshold, it can be determined that the model has successfully focused on the key fault frequency band, thereby improving physical interpretability and verifying the reliability of the obtained contribution.

[0044] To verify the physical reliability of the channel importance score results, a fault frequency matching analysis of the key frequency band was performed by combining wavelet scale and bearing structural parameters.

[0045] Based on the bearing's structural parameters, including the number of rolling elements Rotation speed Contact angle , rolling element diameter Based on the pitch circle diameter D, the typical fault frequency is calculated as follows: Outer ring failure frequency:

[0046] Inner ring failure frequency:

[0047] Rolling element failure frequency:

[0048] Let the channel importance score vector be... Select the top-ranked scores The channel has a center frequency of For any theoretical fault frequency The relative matching error of the channel frequencies is defined as:

[0049] If error ,in If the tolerance threshold is not met, the match is considered successful. If in the previous... If more than half of the high-scoring channels meet the above matching conditions, then the model is considered to have focused on frequency bands with physical fault significance, possessing a certain degree of interpretability and physical consistency. As one example, It is 0.1.

[0050] 4. Attention Guidance Mechanism: The channel gradient weights obtained after training are normalized and used as guidance signals to be introduced into the channel attention mechanism of the model. A guidance regularization term is added to the loss function to effectively promote the model to focus its attention on the key frequency bands to be predicted, improve the stability of the attention distribution, and enhance the robustness and cross-sample generalization ability of the model.

[0051] To enhance the model's ability to focus attention in key frequency bands, an attention-guided regularization term is constructed based on the aforementioned channel importance analysis results and embedded into the model training process, forming an attention regulation mechanism based on gradient sensitivity.

[0052] The channel importance score vector obtained from channel importance analysis is: Normalize it to obtain the guiding vector:

[0053] in The total number of channels and the number of scales The values ​​are the same. Indicates the first The normalized weights of each frequency band channel. Let the attention weight vector generated by the channel attention module in the current training epoch of the model be denoted as . ,in The model represents the first Adaptive attention level for frequency band channels. To incorporate prior sensitivity information reflected in gradient contributions into the attention distribution while maintaining the model's autonomous learning ability, the following guiding regularization term is constructed:

[0054] This measure assesses the deviation between the attention distribution and channel importance assessment during model learning. During training, a flexible constraint is imposed on the attention distribution to make it more responsive to the frequency band effectiveness reflected by gradient sensitivity. The final optimization objective is a weighted combination of the principal loss and the guiding regularization term.

[0055] in The loss is the mean square error prediction loss. Regularization terms to guide attention is the regularization coefficient, a hyperparameter used to adjust the strength of the constraint, ensuring effective fusion of guiding information while maintaining the model's predictive performance.

[0056] This guidance mechanism acts as an intermediary between physical priors and data-driven methods, enhancing the model's ability to focus on highly sensitive frequency bands without significantly interfering with its free learning. This approach can further improve the model's generalization ability and stability under complex conditions, while also providing fundamental support for the physical explanation of the attention mechanism.

[0057] The system for implementing the bearing remaining life prediction method based on a frequency band interpretability guidance mechanism includes the following modules: The data preprocessing module is used to extract multi-scale frequency domain features of vibration signals. In the data preprocessing stage, Morlet continuous wavelet transform is introduced to perform multi-scale transformation on the original time domain signal.

[0058] The channel sensitivity analysis module is used to generate frequency band importance scores and sort them in descending order using a one-dimensional convolutional neural network model.

[0059] The physical feature matching and verification module is used to compare the center frequency corresponding to the high-ranking frequency band channel with the typical fault characteristic frequency and its harmonics calculated from the bearing physical structure parameters, to verify the physical consistency of the frequency response and establish the mapping relationship between the key frequency band and the failure modes of the inner ring, outer ring, rolling elements and cage. The attention-guided training module is used to train a one-dimensional convolutional neural network model. Specifically, it uses the validated frequency band importance score as static guidance information to construct a composite loss function that includes the basic prediction loss and the attention-guided regularization term. Through retraining, it dynamically adjusts the attention weight allocation to enhance the one-dimensional convolutional neural network model's attention to key frequency bands and suppress irrelevant interference, outputting accurate remaining lifetime prediction results.

[0060] This invention obtains the overall importance distribution of each frequency band by performing gradient analysis on the model output, and then selects the frequency bands with high importance rankings to match with known fault characteristic frequencies. Subsequently, the importance distribution of the frequency bands is used as a static guiding signal, and a regularization loss term is introduced to supervise the model's attention output, achieving targeted focus of attention on key frequency bands. While ensuring predictive performance, this invention enhances the interpretability of the model and further proves the consistency between the features of interest and the physical characteristics of mechanical faults through a frequency band verification mechanism.

[0061] As a specific embodiment, this embodiment uses PRONOSTIA bearing life-cycle vibration signal data provided by PHM2012 Challenge as the experimental data source. The platform used is a constant load and constant speed bearing accelerated aging test device. In the experiment, seven sets of bearing operating data under condition 1 (speed 1800 rpm, load 4000 N) were used, including the horizontal and vertical measurement results of the vibration signal. The vibration signal sampling frequency was 25.6 kHz, and it was collected once every 10 seconds, with 2560 sampling points recorded each time, forming typical fixed-length time-series data. The horizontal vibration signal is used as an example for illustration here. Figure 2 The signal variation process of a bearing sample throughout its entire life cycle in the horizontal direction is shown.

[0062] Data Preparation and Preprocessing: To extract multi-scale frequency domain features of vibration signals, Morlet continuous wavelet transform is introduced in the data preprocessing stage to perform multi-scale transformation on the original time-domain signal. This transform decomposes each signal segment into multiple sub-band signals with clear frequency distributions, thereby achieving frequency band decoupling and compression, and providing a certain degree of noise resistance. After the transform, each sample segment is mapped to multiple frequency band response signals, which are used as independent input channels in the subsequent modeling process.

[0063] Bearing1_1 and Bearing1_2 were selected as training samples, Bearing1_3 as validation samples, and Bearing1_4 to Bearing1_7 as test samples. Figure 2 The training sample Bearing1_1 is shown as a diagram of the horizontal vibration signal throughout its entire life cycle, as used in the example. Figure 6 Figures (a) and (b) show the bearing prediction curves for samples Bearing1_3 and Bearing1_6. Nine frequency band response channels were designed based on different scale parameter combinations, forming the tensor shape of the final model input. ,in 9 represents the number of sample batches, 2560 represents the frequency band dimension, and 9 represents the time series length. Each channel frequency band has a defined center frequency, allowing subsequent gradient sensitivity analysis results to be clearly mapped to specific frequency regions, thus enabling physical comparison and verification with known bearing fault characteristic frequencies.

[0064] The wavelet preprocessing described above not only improves the model's ability to perceive changes in the frequency domain of the signal, but also provides a structural foundation for the present invention in the interpretability analysis stage, reflecting the close collaborative design between signal preprocessing and model interpretation.

[0065] 2. Post-hoc calculation and analysis of gradient contribution: To reveal the natural response characteristics of the base model in the frequency dimension and verify its potential ability to focus on key fault frequency bands, this invention introduces a channel sensitivity analysis method based on gradient backpropagation after the initial training is completed. This method, based on the benchmark model, evaluates the importance of different input channels (i.e., frequency bands) to the prediction results, providing a priori physical basis for subsequent attention mechanism-guided training.

[0066] Specifically, while maintaining the network structure, the model parameters that have undergone preliminary training are loaded, and the gradient tracking function of the input tensor is enabled. Samples from the training set are input into the model one by one, and the gradient value of the predicted output relative to the input of each channel is calculated. The absolute values ​​of these gradients are then taken over the time dimension and summed to obtain the gradient contribution vector for each sample across different channels (i.e., different frequency bands). After normalization, a channel weight matrix is ​​formed, reflecting the relative response strength of the model in each frequency band.

[0067] Subsequently, the top K high-weight channels with the highest contribution in each sample were selected, and their corresponding wavelet center frequencies were extracted. Combined with the bearing structural parameters (13 rollers, 3.5 mm roller diameter, 25.6 mm pitch circle diameter, 1800 rpm) provided in the PHM2012 dataset, the following typical fault characteristic frequencies can be calculated: Inner ring fault frequency (BPFI): approximately 221.66 Hz; Outer ring fault frequency (BPFO): approximately 168.34 Hz; Rolling element failure frequency (BSF): approximately 107.66 Hz; Cage failure frequency (FTF): approximately 12.95 Hz.

[0068] Matching analysis was performed on the center frequencies of each channel with typical fault characteristic frequencies such as BPFI and BPFO, and their 1st to 5th harmonics. The results showed that there was a significant correlation between the high-weight channel frequencies and the key fault frequencies, and most matching errors were controlled within 10%. Notably, there were significant differences in the importance scores presented by different channels. This difference was reflected in the matching results as a selective response to different fault frequencies or harmonics, indicating that the model, without guidance, already possessed a certain degree of frequency interpretability and could autonomously focus on key frequency bands closely related to the bearing degradation process. This differentiated focus is not only reasonable but also consistent with the energy distribution of actual fault frequencies, further validating the coupling potential between the model's internal attention mechanism and physical characteristics.

[0069] 3. Model training and visualization analysis The model used includes an input channel attention module, a three-layer one-dimensional convolutional neural network, and a regression prediction module. The input tensor size is... The 18 channels correspond to different frequency bands of the wavelet transform signals from two directions. The attention module outputs channel-level attention weights before training, which describe the frequency band region that the model is currently focusing on.

[0070] Gradient-guided loss is introduced during training, using the aforementioned channel gradient contributions as static guiding signals to supervise the model's attention output. Specifically, as mentioned earlier, a total loss function incorporating gradient-guided constraints is constructed to guide the directional distribution of the model's attention.

[0071] To adjust the strength of this constraint, this embodiment sets corresponding hyperparameters. The value is 0.03, thus achieving effective integration of gradient guidance information and network training without affecting the model's prediction performance.

[0072] Training employs the Huber loss function for main task optimization, with Adam as the optimizer, an initial learning rate of 0.0001, and a batch size of 32. Figure 3 As shown, the loss trends of the training set and validation set were recorded during the experiment, and loss curves were plotted to help determine the stability of the training.

[0073] 4. Results Presentation and Analysis To evaluate the impact of the proposed guided channel attention mechanism on bearing remaining life prediction performance, quantitative and visual analyses were conducted based on the model training logs and prediction results from the experiment.

[0074] In terms of quantitative evaluation, four commonly used evaluation indicators in the current field are selected to comprehensively measure the model performance, namely:

[0075]

[0076]

[0077] in Piecewise exponential functions for different decay rates given in IEEE PHM 2012:

[0078]

[0079] In the calculation formula above, Indicates the actual remaining service life. This represents the predicted remaining useful life. This represents the actual remaining useful life. The smaller the first two indicators, the lower the prediction error. For the score indicator, a higher value indicates better prediction performance and a stronger fit to the lifespan decay trend. The changes in model prediction performance indicators are shown below. Figure 4As shown in (a), (b), (c), and (d).

[0080] For visualization analysis, RUL prediction curves for typical test samples were plotted and compared with actual lifetime curves. The results show that the model exhibits good fit throughout the entire lifetime.

[0081] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for predicting the remaining life of bearings based on a frequency band interpretability guidance mechanism, characterized in that, The process includes the following steps: inputting the data to be tested into the trained bearing remaining life prediction model and outputting the bearing remaining life prediction results; The training of the bearing remaining life prediction model includes the following steps: S1. Extract multi-scale time-frequency features of the original vibration signal and construct a multi-channel frequency band tensor; train a one-dimensional convolutional neural network model based on the multi-channel frequency band tensor; S2. Evaluate the impact of different frequency band characteristics on the bearing remaining life prediction task from the gradient level, and obtain the channel importance score vector; S3. Match the center frequency of the frequency band corresponding to the top-ranked channel importance score vector with the typical fault characteristic frequency, and use the matched score as attention-guided constraint to retrain the model, finally outputting the bearing remaining life prediction result.

2. The bearing remaining life prediction method based on a frequency band interpretability guidance mechanism according to claim 1, characterized in that... In step S1, the original bearing vibration signal is subjected to continuous wavelet transform, and the Morlet wavelet is selected as the mother wavelet function to extract multi-scale time-frequency features with a clear center frequency.

3. The bearing remaining life prediction method based on a frequency band interpretability guidance mechanism according to claim 1, characterized in that... The one-dimensional convolutional neural network model is a one-dimensional convolutional neural network model containing a channel attention mechanism and a regression prediction structure.

4. The bearing remaining life prediction method based on a frequency band interpretability guidance mechanism according to claim 1, characterized in that... In step S2, after the one-dimensional convolutional neural network model is trained, gradient backpropagation is used to calculate the sensitivity score of each frequency band, and the absolute value of the gradient of the prediction loss function with respect to each channel of the input signal is calculated to obtain the response intensity of each channel on each sample. The responses of all samples and channels are normalized to obtain the final channel importance score vector.

5. The bearing remaining life prediction method based on a frequency band interpretability guidance mechanism according to claim 1, characterized in that... In step S2, the channel gradient weights obtained after training the one-dimensional convolutional neural network are normalized and used as guiding signals, which are then introduced into the channel attention mechanism of the one-dimensional convolutional neural network model. A guiding regularization term is added to the composite loss function; the expression for the composite loss function is as follows: ,in The loss is the mean square error prediction loss. Regularization terms to guide attention This is the regularization coefficient.

6. The bearing remaining life prediction method based on a frequency band interpretability guidance mechanism according to claim 1, characterized in that... The typical fault characteristic frequencies include the inner ring fault frequency, outer ring fault frequency, and rolling element spin frequency, which are calculated based on the bearing's inner diameter, outer diameter, number of rolling elements, and diameter.

7. The bearing remaining life prediction method based on a frequency band interpretability guidance mechanism according to claim 1, characterized in that... When the error value is less than or equal to the matching tolerance threshold, it indicates that the frequency band that was successfully matched is faulty.

8. A system for implementing the bearing remaining life prediction method based on a frequency band interpretability guidance mechanism as described in claim 1, characterized in that... It includes the following modules: The data preprocessing module is used to extract multi-scale frequency domain features of vibration signals; The channel sensitivity analysis module is used to generate frequency band importance scores and sort them in descending order using a one-dimensional convolutional neural network model; The physical feature matching and verification module is used to compare the center frequency corresponding to the high-ranking frequency band channel with the typical fault characteristic frequency and harmonic frequency calculated from the bearing physical structure parameters, and to establish the mapping relationship between the key frequency band and the failure modes of the inner ring, outer ring, rolling elements and cage. Attention-guided training module, used to train one-dimensional convolutional neural network models.

9. A computer device comprising a memory and a processor, the memory being electrically connected to the processor, the memory storing a computer program, characterized in that: When the computer program is executed by the processor, it causes the processor to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor implements the method as described in any one of claims 1 to 7.

Images