Method, device and system for predicting residual service life of bearing and medium

Abstract

The invention provides a bearing residual service life prediction method, device and system and a medium, and the method comprises the steps: obtaining a time domain vibration signal collected by a vibration sensor disposed on a bearing, carrying out the preprocessing of the time domain vibration signal, and generating a continuous vibration signal; performing sliding window slicing on the continuous vibration signals to form a vibration signal sequence, extracting multi-dimensional features, and obtaining a vibration feature sequence by combining a working condition code corresponding to the bearing and performing standardized processing according to a bearing instance; inputting the vibration characteristic sequence into a bearing residual service life prediction model, performing convolution operation to extract local characteristics, capturing a forward and backward long-time dependency relationship of the vibration characteristic sequence, and outputting vibration signal time characteristic representation; and processing the vibration signal time characteristic representation through a plurality of radial basis functions in the regression module to obtain a response value of each radial basis function, and mapping to obtain a bearing residual service life prediction result, thereby improving the bearing residual service life prediction precision.

Description

Technical Field

[0001] This application relates to the field of equipment health management, and more specifically, to a method, apparatus, system, and medium for predicting the remaining service life of bearings. Background Technology

[0002] Bearings are among the most ubiquitous and fundamental components of rotating machinery, and their health directly impacts the reliability and safety of critical assets such as motors, compressors, fans, machine tool spindles, rail transit, and wind power. Compared to electrical faults, bearing mechanical failures are often characterized by their insidious onset, rapid propagation, and severe consequences. Early minor spalling or lubrication abnormalities can evolve into serious damage within a short period, leading to downtime or even secondary accidents. Traditional "post-failure repair" or fixed-period preventative maintenance either results in unplanned downtime and high opportunity costs, or leads to over-maintenance and spare parts waste, neither of which can meet the high uptime and lean cost requirements of modern industry. Therefore, the perception, modeling, and assessment of the remaining service life of bearings has become one of the most valuable research and application directions in equipment health management systems.

[0003] Among the existing technologies for bearing remaining service life, traditional thresholding methods and statistical models are not sensitive to early minor faults and are greatly affected by changes in operating conditions; machine learning methods rely on manual feature engineering, have poor generalization ability across operating conditions, and are difficult to characterize complex nonlinear degradation; although deep learning methods can learn end-to-end, the models are like black boxes and lack interpretability, making it difficult to implement in real industrial scenarios, and maintenance personnel find it difficult to trust and adopt them. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method, apparatus, system and medium for predicting the remaining service life of bearings, which improves the accuracy of bearing remaining service life prediction and the interpretability of the prediction results.

[0005] This application provides a method for predicting the remaining service life of a bearing, the method comprising: The time-domain vibration signal is acquired by a vibration sensor deployed on the bearing, and the time-domain vibration signal is preprocessed to generate a continuous vibration signal. The continuous vibration signal is sliced ​​into sliding windows to form a vibration signal sequence. Multi-dimensional features of the vibration signal in each window of the vibration signal sequence are extracted to construct an initial high-dimensional feature vector. The initial high-dimensional feature vector and the working condition code corresponding to the bearing are concatenated, and normalization processing is performed according to the bearing instance to obtain a vibration feature sequence. The multi-dimensional features are extracted from the vibration signal in each window of the vibration signal sequence, namely time-domain statistical features, frequency-domain spectral features, nonlinear features, and time-frequency domain features. The vibration feature sequence is input into a pre-trained bearing remaining service life prediction model. The time-series feature extraction module in the bearing remaining service life prediction model performs a convolution operation on the vibration feature sequence in the time dimension to extract local features of the vibration feature sequence. The extracted local features are then processed to capture the forward and backward long-term dependencies of the vibration feature sequence, and output a vibration signal time feature representation that integrates local and global time-series information. The vibration signal time feature representation is processed by multiple radial basis functions in the regression module of the bearing remaining service prediction model to obtain the response value of each radial basis function, and the bearing remaining service prediction result is obtained by linearly combining the response values ​​of each radial basis function.

[0006] In some embodiments, the bearing remaining service prediction method, wherein the linear combination of the response values ​​of each radial basis function is used to map the bearing remaining service prediction result, includes: By linearly combining the response values ​​of multiple radial basis functions, the predicted remaining service life of the bearing is obtained by mapping multiple quantiles; the quantiles include 0.1, 0.3, 0.5, 0.7 and 0.9 quantiles; The specified quantile prediction value is used as the point estimate of the bearing's remaining service life, and the prediction interval of the bearing's remaining service life is determined based on the bearing's remaining service life prediction values ​​of multiple quantiles. The point estimate and prediction range of the remaining service life of the output bearing are used as the prediction results of the remaining service life of the bearing.

[0007] In some embodiments, the bearing remaining service life prediction method further includes: Based on the bearing remaining service prediction results at continuous time points, a visualization is generated showing the point estimate of the bearing remaining service and the trend of the prediction interval changing over time. Based on the transformation process of the vibration signal time feature representation by the regression module, a model decision visualization diagram is generated. The model decision visualization diagram includes: a radial basis function response curve and a feature contribution time series heatmap. The radial basis function response curve shows the activation values ​​of a specific vibration signal time feature in the regression module on all radial basis functions. The feature contribution time series heatmap shows the changes in the contribution of at least some types of vibration signal time feature representations to the prediction results at corresponding time points at continuous time points. Based on the envelope demodulation and spectrum analysis of the original time-domain vibration signal, an envelope spectrum is obtained; the envelope spectrum includes a frequency band energy variation map and an envelope spectrum feature map; the frequency band energy variation map shows at least one preset frequency energy variation curve related to bearing failure.

[0008] In some embodiments, the bearing remaining service life prediction method includes the time-domain statistical features such as the signal mean, variance, skewness, kurtosis, peak factor, impulse factor, and margin factor. The frequency domain spectral features include: the spectral centroid and spectral flatness obtained by calculating the power spectral density of the vibration signal in the window, and the envelope spectral features extracted using the envelope demodulation method; The nonlinear features include detrended fluctuation analysis fractal dimension, sample entropy, and multi-scale sample entropy; The time-frequency domain features are obtained by performing wavelet transform on the vibration signal of the window.

[0009] In some embodiments, in the bearing remaining service prediction method, the bearing remaining service prediction model is obtained by training based on the composite loss function in the bearing remaining service prediction model; The composite loss function includes a regression main loss term, a monotonicity regularization term, and a quantile regression loss term; The regression main loss term is used to evaluate the difference between the predicted value and the true label; The monotonicity regularization term is used to constrain the predicted values ​​at adjacent time steps, penalizing cases where the predicted values ​​violate the physical law of monotonically decreasing with time. The quantile regression loss term optimizes the ball loss corresponding to multiple quantiles simultaneously.

[0010] In some embodiments, the bearing remaining service life prediction method is described, and the bearing remaining service life prediction model is trained based on a sample dataset. The sample dataset was constructed based on the following method: The complete continuous sample vibration signal from the start of operation of the sample bearing to its failure is sliced ​​into sliding windows to generate a sequence of sample signal windows arranged in chronological order. Each sample signal window constitutes a training sample. For each sample bearing instance, a remaining service life label is assigned to each training sample. The maximum remaining service life of the sample bearing instance is set, and the inflection point is defined as a preset proportion of the maximum remaining service life. The remaining service life label of training samples whose sample bearing operation time is earlier than or equal to the inflection point is assigned the maximum remaining service life value. The remaining service life label of training samples whose sample bearing operation time is later than the inflection point decreases linearly from the maximum value to zero. Each sample is labeled with a corresponding sample operating condition code; the sample operating condition code includes speed code and load code.

[0011] In some embodiments, a bearing remaining service life prediction device is also provided, the device comprising: The acquisition module is used to acquire time-domain vibration signals collected by vibration sensors deployed on the bearing, and to preprocess the time-domain vibration signals to generate continuous vibration signals. The first extraction module is used to slice the continuous vibration signal into a sliding window to form a vibration signal sequence, extract multi-dimensional features of the vibration signal in each window of the vibration signal sequence, construct an initial high-dimensional feature vector, concatenate the initial high-dimensional feature vector with the working condition code corresponding to the bearing, and perform standardization processing according to the bearing instance to obtain a vibration feature sequence; wherein the multi-dimensional features are extracted from the vibration signal in each window of the vibration signal sequence, namely time-domain statistical features, frequency-domain spectral features, nonlinear features and time-frequency domain features. The second extraction module is used to input the vibration feature sequence into the pre-trained bearing remaining service life prediction model. The time-series feature extraction module in the bearing remaining service life prediction model performs a convolution operation on the vibration feature sequence in the time dimension to extract the local features of the vibration feature sequence. The extracted local features are then processed to capture the forward and backward long-term dependencies of the vibration feature sequence, and output a vibration signal time feature representation that integrates local and global time-series information. The processing module is used to process the time feature representation of the vibration signal through multiple radial basis functions in the regression module of the bearing remaining service prediction model, obtain the response value of each radial basis function, and linearly combine the response values ​​of each radial basis function to map and obtain the bearing remaining service prediction result.

[0012] In some embodiments, the processing module includes: By linearly combining the response values ​​of multiple radial basis functions, the predicted remaining service life of the bearing is obtained by mapping multiple quantiles; the quantiles include 0.1, 0.3, 0.5, 0.7 and 0.9 quantiles; The specified quantile prediction value is used as the point estimate of the bearing's remaining service life, and the prediction interval of the bearing's remaining service life is determined based on the bearing's remaining service life prediction values ​​of multiple quantiles. The point estimate and prediction range of the remaining service life of the output bearing are used as the prediction results of the remaining service life of the bearing.

[0013] In some embodiments, a bearing remaining service life prediction system is also provided, the system comprising: a vibration sensor, an edge computing unit, and a cloud computing center; The sensor is used to acquire the original time-domain vibration signal of the bearing; The edge computing unit is used in the steps of the bearing remaining service life prediction method. The cloud computing center is used to train the bearing remaining service life prediction model to obtain the trained bearing remaining service life prediction model.

[0014] In some embodiments, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, performs the steps of the bearing remaining service life prediction method.

[0015] This application provides a method, apparatus, system, and medium for predicting the remaining service life of bearings. Addressing the challenges of low vibration energy and poor signal-to-noise ratio in the early stages of bearing degradation, where traditional threshold-based health indicators are easily overwhelmed by noise and operational disturbances, and deep models are prone to exhibiting "good performance in the later stages and poor performance in the early stages," this method constructs a multi-dimensional vibration feature sequence incorporating operational conditions. It extracts local impacts and bandwidth patterns, capturing forward and backward long-term dependencies across windows to identify bearing degradation trends, thereby improving sensitivity and stability to early weak signals. To address the significant fluctuations in prediction performance caused by data distribution drift across machine models and speeds / loads, this method employs standardization based on bearing instances at the feature and label levels, explicitly introduces operational condition coding, and uses bidirectional time-series modeling to enhance the handling of non-stationary sequences. The method improves the adaptability and generalization capability across operating conditions. By processing the temporal feature representation of the vibration signal through multiple radial basis functions in the regression module of the bearing remaining service life prediction model, a complex and invisible black box mapping is transformed into a process composed of multiple simple and visible basis function responses. By analyzing the basis functions with the largest weights, engineers can directly know which "feature intervals" dominate the current prediction results, enhancing interpretability. The method combines time series models with efficient nonlinear regression methods. Through comprehensive feature extraction, multi-level time series modeling, and the introduction of a regression head, it effectively solves the shortcomings of existing remaining service life prediction methods in terms of multi-operating condition adaptation, nonlinear modeling, physical consistency, and uncertainty quantification. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of the bearing remaining service life prediction method according to an embodiment of this application is shown; Figure 2 A flowchart illustrating the method for constructing a sample dataset according to an embodiment of this application is shown; Figure 3 A schematic diagram of the bearing remaining service life prediction model structure described in an embodiment of this application is shown; Figure 4This paper presents a flowchart illustrating a method for predicting bearing remaining service life by mapping the response values ​​of each radial basis function in a linear combination according to an embodiment of this application. Figure 5 A schematic diagram of the bearing remaining service life prediction device according to an embodiment of this application is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0019] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0021] The main failure modes of bearings include: fatigue spalling (propagation of microcracks on the raceway / rolling element surface), lubrication failure and oil film rupture, contamination and corrosion, local stress concentration caused by assembly or alignment errors, and cage damage. These mechanisms will form a spectral structure in the vibration signal related to the fault characteristic frequency and its harmonics and sidebands, accompanied by changes in energy distribution and modulation characteristics. In the early stage, the signal is weakly non-stationary with a low signal-to-noise ratio, and its characteristics are often masked by background noise and operating condition disturbances (speed, load fluctuations). In the middle and late stages, the impact becomes stronger, and indicators such as envelope spectrum and kurtosis become more sensitive, but the degradation process may also experience a phased "plateau" due to improved lubrication or load changes.

[0022] Research on the remaining service life of bearings has generally evolved from thresholds and empirical indicators, statistical / probabilistic modeling, machine learning to the integration of deep learning and physics.

[0023] Early methods focused on constructing health indicators, such as time-domain statistics (mean, variance, skewness, kurtosis, peak factor, impulse factor, margin factor), frequency-domain energy ratio, envelope spectral peak value, spectral kurtosis, etc., or time-frequency domain indicators (STFT, wavelet energy ratio, instantaneous energy of Hilbert-Huang transform, etc.). Remaining lifetime estimation typically involves trend fitting or threshold extrapolation on these health indicators. These methods are highly interpretable and simple to implement, but they are sensitive to noise and changes in operating conditions, and struggle to characterize complex nonlinear degradation relationships.

[0024] Based on the statistical laws governing failure time, several approaches have emerged, including lifetime distribution fitting (such as Weibull), survival analysis, and degradation process modeling (Wiener / Gamma processes). Another approach involves state-space and Bayesian filtering (Kalman / particle filtering) to perform online state estimation and posterior inference of remaining lifetime under observational noise. The advantages of this approach are that uncertainties can be clearly defined and can be combined with prior physical knowledge; however, in real-world complex degradation scenarios, model assumptions (independence, stationarity, linearity / quasi-linearity) are often difficult to fully satisfy, and feature construction and parameter identification require strong expert experience.

[0025] Based on feature engineering and supervised learning, common models include linear / ridge regression, SVR, random forest, and gradient boosting trees. These methods are highly efficient for small to medium-sized datasets with good feature separability and are also convenient for feature importance analysis and initial screening. However, their end-to-end modeling capability for the original waveform is limited, and they face bottlenecks in multi-condition transfer and strongly nonlinear mapping.

[0026] Deep learning methods have significantly improved end-to-end modeling capabilities and complex pattern extraction capabilities. They are mainly categorized as: CNN / TCN, RNN / LSTM / GRU with attention mechanisms, autoencoders and contrastive learning, and Transformers. The challenges of deep learning include: weak early-stage degradation signals and low feature contrast; severe multi-domain bias; insufficient uncertainty quantification and interpretability; and a lack of intrinsic constraints on the physical consistency of the output's remaining lifetime (non-negativity and non-increasing over time), often requiring external regularization or post-processing.

[0027] Despite significant research progress, the following key challenges remain in achieving stable, interpretable remaining service life engineering applications: (1) The impact of early micro-damage on vibration is often drowned out by noise; fluctuations in rotational speed and load cause time-frequency structure drift, and the envelope spectrum and energy index show "condition correlation". How to suppress condition interference while maintaining sensitivity to early-stage vibration is a problem faced by both features and models.

[0028] (2) The same architecture exhibits significant performance differences across different machine models / operating conditions. If features and models implicitly depend on the statistical characteristics of operating conditions, they will mismatch when encountering unseen operating conditions after deployment. Existing domain alignment methods need to consider interpretability and verification costs in actual industrial applications, and purely adversarial or black-box alignment methods are difficult to pass quality and safety reviews quickly.

[0029] (3) The engineering criteria and laboratory definitions are not entirely consistent, and the "zero point" of the remaining useful life is subjective; the remaining useful life should be non-negative and not increase over time, but many regression models lack structural constraints, and the output may show short-term rebound. Although this can be improved by monotonic regularization or post-processing, how to intrinsically adhere to physical consistency in model structure and learning mechanism remains a challenge.

[0030] (4) Equipment health management involves safety and compliance. Operation and maintenance engineers usually need to see evidence that corresponds to the failure mechanism: such as the energy of a certain characteristic frequency band, sideband changes in the envelope spectrum, and the time window for the occurrence of time domain impacts. End-to-end deep models lack clear visualization and mechanism comparison, which faces resistance in the field promotion and review stages.

[0031] Based on this, this application provides a method, device, system, and medium for predicting the remaining service life of bearings. Addressing the challenges of low vibration energy and poor signal-to-noise ratio in the early stages of bearing degradation, where traditional threshold-based health indicators are easily overwhelmed by noise and operational disturbances, and deep models are prone to exhibiting a "good later, poor early" pattern, this method constructs a multi-dimensional vibration feature sequence combined with operational conditions, extracts local impacts and bandwidth patterns, and captures long-term forward and backward dependencies across windows to capture bearing degradation trends, thereby improving sensitivity and stability to early weak signals. To address the significant fluctuations in prediction performance caused by data distribution drift across machine models and speeds / loads, this method employs standardization based on bearing instances at the feature and label levels, explicitly introduces operational condition coding, and uses bidirectional time-series modeling to enhance the handling of non-stationary conditions. The sequence adaptation enhances the generalization ability across operating conditions. By processing the temporal feature representation of the vibration signal through multiple radial basis functions in the regression module of the bearing remaining service life prediction model, a complex and invisible black box mapping is transformed into a process composed of multiple simple and visible basis function responses. By analyzing the basis functions with the largest weights, engineers can directly know which "feature intervals" dominate the current prediction results, enhancing interpretability. The method combines time series models with efficient nonlinear regression methods. Through comprehensive feature extraction, multi-level time series modeling, and the introduction of a regression head, it effectively solves the shortcomings of existing remaining service life prediction methods in terms of multi-operating condition adaptation, nonlinear modeling, physical consistency, and uncertainty quantification.

[0032] Please refer to Figure 1 , Figure 1A flowchart of the bearing remaining service life prediction method according to an embodiment of this application is shown; as follows: Figure 1 As shown, the method includes the following steps S101-S104: S101. Acquire the time-domain vibration signal collected by the vibration sensor deployed on the bearing, and preprocess the time-domain vibration signal to generate a continuous vibration signal; S102. The continuous vibration signal is sliced ​​into sliding windows to form a vibration signal sequence. Multi-dimensional features of the vibration signal in each window of the vibration signal sequence are extracted to construct an initial high-dimensional feature vector. The initial high-dimensional feature vector and the working condition code corresponding to the bearing are concatenated and standardized according to the bearing instance to obtain a vibration feature sequence. The multi-dimensional features are extracted from the vibration signal in each window of the vibration signal sequence, namely time-domain statistical features, frequency-domain spectral features, nonlinear features and time-frequency domain features. S103. Input the vibration feature sequence into the pre-trained bearing remaining service life prediction model. The time-series feature extraction module in the bearing remaining service life prediction model performs a convolution operation on the vibration feature sequence in the time dimension to extract the local features of the vibration feature sequence. The extracted local features are then processed to capture the forward and backward long-term dependencies of the vibration feature sequence, and the vibration signal time feature representation that integrates local and global time-series information is output. S104. The vibration signal time feature representation is processed by multiple radial basis functions in the regression module of the bearing remaining service prediction model to obtain the response value of each radial basis function, and the response value of each radial basis function is linearly combined to map the bearing remaining service prediction result.

[0033] The technical solution described in this application consists of the following main steps: dataset construction, data slicing and annotation, feature construction, model construction, loss and constraint design, training and validation, and online inference.

[0034] In step S101, the time-domain vibration signal collected by the vibration sensor deployed on the bearing is acquired, and the time-domain vibration signal is preprocessed to generate a continuous vibration signal.

[0035] The bearing remaining service life prediction model is trained based on a sample dataset. Please refer to Figure 2 The sample dataset is constructed based on the following steps S201-S203: S201. The complete continuous sample vibration signal from the start of operation of the sample bearing to its failure is sliced ​​by a sliding window to generate a sequence of sample signal windows arranged in chronological order. Each sample signal window constitutes a training sample. S202. For each sample bearing instance, label the remaining service life of each training sample with a label; wherein, set the maximum remaining service life of the sample bearing instance, and define the inflection point as a preset proportion of the maximum remaining service life; the remaining service life label of training samples whose sample bearing running time is earlier than or equal to the inflection point is assigned the maximum remaining service life value; the remaining service life label of training samples whose sample bearing running time is later than the inflection point decreases linearly from the maximum value to zero. S203. Label each sample with a corresponding sample operating condition code; the sample operating condition code includes speed code and load code.

[0036] To improve the accuracy of predictions, embodiments of this application use an accelerometer to collect data, with the sampling frequency range set to 20kHz to ensure that sufficient signal details are captured.

[0037] To avoid abnormal data caused by sensor failure or data noise, anomaly detection is performed using the statistical characteristics of the signal (such as mean and variance), and noise and environmental changes are filtered out, which means the time-domain vibration signal is preprocessed.

[0038] Data set partitioning is a crucial step in ensuring the model's generalization ability. To avoid information leakage, this embodiment of the application partitions the data into training, validation, and test sets based on bearing instances, with a ratio of 7:1:2. That is, 70% of the data for each bearing instance is used for training, 10% for validation, and 20% for testing. To ensure the model has good generalization ability, the test set data covers different operating conditions and avoids adjacent samples of the same bearing appearing between the training and test sets.

[0039] After acquiring complete and continuous vibration signals from the start of operation to failure of the bearing, data slicing and labeling are performed. In this stage, the continuous vibration signals are first sliced ​​using a sliding window. Based on the bearing's operating characteristics, a window length of 1 second and a step size of 50% of the window length are selected. Each window generates a sample, and the sample's label is defined by its remaining service life.

[0040] For each bearing instance, a maximum remaining service life is defined, and a threshold is set at 20% of this maximum. When the usage time is below the threshold, the remaining service life is marked as the maximum; when the usage time exceeds the threshold, the remaining service life linearly decreases to 0. A Savitzky-Golay filter is used to smooth the remaining service life sequence, reducing label fluctuations and noise.

[0041] In addition, each slice is labeled according to its operating conditions (speed, load). The operating condition information is encoded numerically and used as an auxiliary input to the model.

[0042] In step S102, the continuous vibration signal is sliced ​​into sliding windows to form a vibration signal sequence. Multi-dimensional features of the vibration signal in each window of the vibration signal sequence are extracted to construct an initial high-dimensional feature vector. The initial high-dimensional feature vector and the working condition code corresponding to the bearing are concatenated and standardized according to the bearing instance to obtain a vibration feature sequence. The multi-dimensional features are extracted from the vibration signal in each window of the vibration signal sequence, namely time-domain statistical features, frequency-domain spectral features, nonlinear features and time-frequency domain features.

[0043] After obtaining the continuous vibration signal stream, the primary task is to transform these high-dimensional, unstructured raw signal data into information that can effectively characterize the degradation process of bearing health status. Therefore, sliding window slicing is performed.

[0044] The continuous vibration signals are sliced ​​using a sliding window to form a vibration signal sequence, which is the same as the segmentation method used when constructing the sample dataset, and will not be described again.

[0045] For each sample window, in order to make full use of the multidimensional information in the vibration signal, quantitative indicators are extracted from multiple perspectives, including time domain, frequency domain, time-frequency domain and nonlinear characteristics.

[0046] In some embodiments, the time-domain statistical features include the signal's mean, variance, skewness, kurtosis, peak factor, impulse factor, and margin factor. The frequency domain spectral features include: the spectral centroid and spectral flatness obtained by calculating the power spectral density of the vibration signal in the window, and the envelope spectral features extracted using the envelope demodulation method; The nonlinear features include detrended fluctuation analysis fractal dimension, sample entropy, and multi-scale sample entropy; The time-frequency domain features are obtained by performing wavelet transform on the vibration signal of the window.

[0047] In other words, in order to make full use of the information in the vibration signal, the embodiments of this application extract features from multiple perspectives, including time domain, frequency domain, time-frequency domain and nonlinear features, to ensure that the model can accurately identify the degradation state of the bearing.

[0048] Time-domain features include mean, variance, skewness, kurtosis, impulse factor, crazing factor, and margin factor. These features reflect the basic statistical properties of a signal. As one-dimensional features, time-domain features can reflect the specific time of a fault occurrence. By statistically analyzing features such as mean, variance, skewness, kurtosis, impulse factor, crazing factor, and margin factor, they can reflect signal fluctuations, impulses, and energy changes, thereby inferring whether a fault has occurred.

[0049] Frequency domain features were analyzed by calculating the power spectral density using the Welch method, and extracting the energy ratio and spectral line shape indices (spectral centroid, spectral flatness) of the fault-related frequency band. Furthermore, envelope demodulation was used to extract the envelope spectral features of the signal, thereby analyzing the impulse characteristics present in the signal.

[0050] Frequency domain characteristics are one-dimensional features that can reflect the type of fault that has occurred. By analyzing indicators such as power spectral density, energy ratio, envelope spectrum, and spectral line shape, the frequency distribution pattern is revealed, thus distinguishing different types of faults.

[0051] Time-frequency domain features are obtained through wavelet transform, capturing abrupt changes in non-stationary signals; Time-frequency domain features are two-dimensional features that can integrate information from both the time and frequency domains, reflecting both the time and type of the fault.

[0052] Nonlinear features, including detrended fluctuation analysis fractal dimension, sample entropy, and multi-scale sample entropy, reveal the nonlinear dynamic behavior during bearing degradation. Through joint modeling of multi-domain features, degradation information at different stages in the bearing vibration signal is captured.

[0053] Nonlinear features are one-dimensional features that can reflect the degree of disorder in a signal. By using fractal dimension, sample entropy, and multi-scale sample entropy, the "dynamic stability" and "chaos" of the system state can be distinguished, which helps to detect faults earlier.

[0054] After extracting multi-dimensional features and constructing the features, to ensure the robustness and generalization ability of the model and to eliminate irrelevant systematic biases, two processes are performed on all extracted features: First, the working condition encoding vector representing the working environment is concatenated with the aforementioned initial high-dimensional feature vector to provide the model with a decision-making context; second, to improve the stability of the model and reduce collinearity among features, a standardization method based on bearing instances is used to standardize all features. Statistics from the training set are used to standardize the features of the test set to avoid data leakage.

[0055] The operating condition code is a numerical representation of the external operating conditions (mainly including speed and load) of the bearing during operation. It is used as an auxiliary input to the model to inform the model of the current working background of the signal, so as to eliminate the interference of operating condition changes on the signal characteristics and improve the model's adaptability to different working states.

[0056] The standardization is based on the z-score of the bearing instance, and the inference phase uses training set statistics.

[0057] Here, the data distribution drift across machine types and speeds / loads causes significant fluctuations in prediction performance. Standardized labeling based on bearing instances is adopted on the feature and label side, and operating condition coding is explicitly introduced on the model side. Bidirectional time series modeling is used to enhance the adaptation to non-stationary sequences and improve the generalization ability across operating conditions.

[0058] The vibration feature sequence obtained in step S102 is the direct input for the subsequent bearing remaining service life prediction model to perform time-series pattern learning and remaining service life prediction.

[0059] In steps S103 and S104, the vibration feature sequence is input into a pre-trained bearing remaining service life prediction model. The time-series feature extraction module in the bearing remaining service life prediction model performs a convolution operation on the vibration feature sequence in the time dimension to extract local features of the vibration feature sequence. The extracted local features are then processed to capture the forward and backward long-term dependencies of the vibration feature sequence, and a vibration signal time feature representation that integrates local and global time-series information is output. The vibration signal time feature representation is then processed by multiple radial basis functions in the regression module of the bearing remaining service life prediction model to obtain the response value of each radial basis function. The response values ​​of each radial basis function are then linearly combined to map and obtain the bearing remaining service life prediction result.

[0060] The bearing remaining service life prediction model adopts the CNN-BiLSTM-FastKAN model structure; the temporal feature extraction module includes a convolutional neural network (CNN) and a bidirectional long short-term memory network (BiLSTM), and the regression module uses the FastKAN regression head.

[0061] Please refer to Figure 3 , Figure 3 The diagram shows a schematic of the bearing remaining service life prediction model structure described in this application embodiment; the CNN module can extract key information from the input features, suppress noise in the features, and enable the model to focus on fault-related features, providing a higher-level, structured input for subsequent time series modeling (BiLSTM).

[0062] Bearing life degradation is a dynamic process that evolves slowly over time, with long-term dependencies in the signal. BiLSTM can establish a dynamic relationship between the present and the future, identify the current health status of the bearing, and understand its dynamic evolution trend from health to failure.

[0063] The relationship between bearing life and its characteristic signals is usually nonlinear. FastKAN can fit this complex mapping. By integrating the multidimensional feature vectors output by BiLSTM, potential high-order interaction information is extracted to predict bearing life.

[0064] In this way, by combining the advantages of convolutional neural networks (CNN) in local pattern extraction, bidirectional long short-term memory networks (BiLSTM) in temporal feature modeling, and FastKAN regression head in nonlinear mapping, the efficiency and accuracy of the model under multiple operating conditions and nonlinear conditions are guaranteed.

[0065] Specifically, the CNN module extracts local features from the vibration signal. To better capture local patterns, the CNN uses a kernel size of 5, a stride of 1, and 3 convolutional layers. A BatchNorm layer is added after each convolutional layer for normalization, and ReLU is used as the activation function to ensure the network's non-linear expressive power. Finally, max pooling is used to reduce the data dimensionality by half, avoiding excessive computational complexity.

[0066] BiLSTM module: Captures long-term dependencies in signals. This module has a hidden dimension of 128, two LSTM layers, and a dropout rate of 0.2 to prevent overfitting. The bidirectional structure of BiLSTM better captures the bidirectional dependencies of signals, thus effectively improving the model's temporal prediction capability.

[0067] The FastKAN regression head approximates the nonlinear degradation process using learnable radial basis functions, avoiding the problem of requiring a large number of parameters in traditional regression models. The radial basis functions are initialized with a uniform distribution in the interval [-1, 1] at their center and [0.2, 0.6] at their width, with 16 radial basis functions per channel. This regression head linearly transforms the features output by the BiLSTM and calculates the responses of the radial basis functions, then uses a weighted average to obtain the final predicted remaining lifetime.

[0068] In some embodiments, the bearing remaining service life prediction model is obtained by training based on the composite loss function in the bearing remaining service life prediction model; The composite loss function includes a regression main loss term, a monotonicity regularization term, and a quantile regression loss term; The regression main loss term is used to evaluate the difference between the predicted value and the true label; The monotonicity regularization term is used to constrain the predicted values ​​at adjacent time steps, penalizing cases where the predicted values ​​violate the physical law of monotonically decreasing with time. The quantile regression loss term optimizes the ball loss corresponding to multiple quantiles simultaneously.

[0069] In some embodiments, the main loss is Huber loss and / or mean squared error loss.

[0070] In some embodiments, the loss function and constraint design are as follows.

[0071] First, Huber loss is used as the main regression loss term. This loss function has good robustness to outliers and can effectively reduce the interference of outlier data on the model.

[0072] Secondly, a monotonicity regularization term is designed and introduced to constrain the predicted value of the remaining useful life, ensuring that it decreases monotonically over time and guaranteeing the physical consistency of the remaining useful life.

[0073] In addition to traditional regression loss and monotonicity constraints, this application also employs quantile regression to quantify uncertainty; by outputting multiple quantiles (0.1, 0.3, 0.5, 0.7, 0.9 quantiles), a prediction interval for the remaining useful life is generated, rather than just a single point estimate; this quantile regression uses bouncing loss to independently optimize each quantile, enhancing the model's ability to assess uncertainty and risk.

[0074] To address the issue that many regression model outputs may show short-term increases or negative values, violating the physical principles of maintenance decisions, the bearing remaining service life prediction model described in this application adopts non-negative mapping in the output layer and introduces monotonicity constraints / regularization in the training objective to suppress the increase in predictions at adjacent time steps, ensuring that the remaining service life prediction satisfies the physical consistency of "non-negative and monotonically non-increasing".

[0075] In the training and validation phases of the bearing remaining service life prediction model, this embodiment specifically employs the AdamW optimizer, combined with a single-cycle learning rate scheduler to accelerate training. The batch size used during training is 64, the initial learning rate is set to 1e-3, and the maximum number of training epochs is 200. During training, cross-validation and early stopping mechanisms are used to prevent overfitting, and the optimal model is selected using RMSE, MAE, and PH evaluation metrics on the validation set.

[0076] In some embodiments, please refer to Figure 4 The linear combination of the response values ​​of each radial basis function maps to obtain the bearing remaining service life prediction result, including the following steps S401-S403: S401. By linearly combining the response values ​​of multiple radial basis functions, multiple quantiles of the predicted remaining service life of the bearing are simultaneously mapped to obtain the predicted values; the quantiles include 0.1, 0.3, 0.5, 0.7 and 0.9 quantiles; S402. Use the specified quantile prediction value as the point estimate of the bearing's remaining service life, and determine the prediction interval of the bearing's remaining service life based on the bearing's remaining service life prediction values ​​of multiple quantiles. S403. Output the point estimate and prediction range of the remaining service life of the bearing as the prediction result of the remaining service life of the bearing.

[0077] Here, the response values ​​of each radial basis function are linearly combined to obtain the bearing remaining service life prediction result. Specifically, the FastKAN regression head uses learnable radial basis functions to approximate the edge functions and converges them along the time dimension to output the remaining service life prediction value.

[0078] In some embodiments, the time-dimensional convergence is either mean convergence or attention-based weighted convergence.

[0079] In some embodiments, the output of the bearing remaining service life prediction model is processed by a Softplus function or an equivalent non-negative mapping to ensure that the remaining service life is non-negative.

[0080] In some embodiments, the bearing remaining service life prediction method further includes: Based on the bearing remaining service prediction results at continuous time points, a visualization is generated showing the point estimate of the bearing remaining service and the trend of the prediction interval changing over time. Based on the transformation process of the vibration signal time feature representation by the regression module, a model decision visualization diagram is generated. The model decision visualization diagram includes: a radial basis function response curve and a feature contribution time series heatmap. The radial basis function response curve shows the activation values ​​of a specific vibration signal time feature in the regression module on all radial basis functions. By observing the activation values, it is determined whether the bearing has entered the degradation stage. The feature contribution time series heatmap shows the change in the contribution of at least some types of vibration signal time feature representations to the prediction results at the corresponding time points at continuous time points, establishing a clear correspondence between the model decision process and the actual degradation time axis of the bearing. Envelope spectrum is obtained by performing envelope demodulation and spectral analysis on the original time-domain vibration signal. The envelope spectrum includes a frequency band energy variation map and an envelope spectrum feature map. The frequency band energy variation map displays energy variation curves at least one preset frequency related to bearing failure, reflecting the evolution of frequency band energy related to different failure mechanisms, thereby linking the model prediction results with classic bearing failure mechanisms. In existing technologies, black-box models lack visual evidence consistent with failure mechanisms, affecting practical application. Therefore, this method outputs FastKAN basis response curves and visualizes their temporal importance / feature contribution, linking them to mechanistic features such as envelope spectrum frequency band energy and sideband variations, forming a comparable evidence chain of "data-model-mechanism," meeting engineering interpretability requirements.

[0081] By using point estimates of the remaining bearing life and visualizations of the trend of the prediction interval over time, engineers can intuitively see the overall downward trend of the remaining bearing life, the change in the rate of decay, and the reliability of the prediction (range width).

[0082] The model decision visualization is the core of model interpretability. The radial basis function response curve can reflect the nonlinear piecewise relationship of the contribution of eigenvalues ​​to bearing life prediction in different intervals. The time-series heatmap of feature contribution can reflect how the relative importance of different features to the prediction results evolves over time throughout the entire life cycle of the bearing.

[0083] In some embodiments, the bearing remaining service life prediction method is implemented using a modular design, with each module undertaking different functions, facilitating future expansion and upgrades; for example, it may include a data acquisition module, a feature extraction module, a time series modeling module, a FastKAN regression module, a loss calculation module, an inference and visualization module, etc.

[0084] In some embodiments, the bearing remaining service life prediction method is run in a bearing remaining service life prediction system. The system architecture includes a vibration sensor, an edge computing unit, and a cloud computing center. The sensor is responsible for collecting vibration signals, the edge computing unit performs real-time data processing and model inference, and the cloud computing center performs model training and data analysis to ensure the efficient operation of the system in different scenarios.

[0085] The bearing remaining service life prediction method described in this application has the following advantages and improvements.

[0086] This invention improves the sensitivity of early micro-damage detection: Traditional remaining service life prediction methods typically rely on global features of vibration signals, such as mean and variance; however, in the early stages of bearing degradation, signal changes are often very weak and difficult to detect. By combining time-domain, frequency-domain, time-frequency-domain, and nonlinear features, the method described in this application can capture subtle changes in early micro-damage signals, especially in cases of low signal-to-noise ratio and high background noise, thus enhancing the ability to identify early degradation; by extracting local features using CNN and modeling long-term dependencies using BiLSTM, the shortcomings of traditional methods in early micro-damage detection are effectively overcome.

[0087] It has strong adaptability to multiple working conditions and improved generalization ability across working conditions: Bearings exhibit significant variations in their operating conditions under different operating conditions. Traditional methods for predicting remaining service life often perform well under specific conditions but have poor adaptability across different operating conditions. Therefore, by introducing operating condition coding, the impact of operating condition changes is considered in the model. At the same time, features are standardized according to bearing instances, enabling the model to maintain high prediction accuracy when facing different operating conditions. This design improves the model's generalization ability and can provide more reliable prediction results in practical applications across different operating conditions.

[0088] Enhanced nonlinear modeling capabilities and reduced model complexity: In traditional remaining useful life prediction methods, complex nonlinear relationships often require fitting with a large number of model parameters, which can easily lead to overfitting and incur significant computational overhead. By introducing the FastKAN regression head, the model complexity is effectively reduced while maintaining nonlinear modeling capabilities. The FastKAN regression head replaces traditional polynomial or fully connected layers with learnable radial basis functions, greatly reducing the number of required parameters and improving the computational efficiency of the model. This enables the model to not only approximate complex nonlinear degradation processes but also achieve rapid inference under limited computational resources.

[0089] Physical consistency is ensured, meeting engineering application requirements: Physical consistency is crucial in predicting the remaining service life of bearings, especially the requirements for non-negativity and non-incremental growth over time. Traditional regression models often lack this constraint, leading to potentially negative predictions or short-term rebounds, which is unacceptable for maintenance decisions. Therefore, monotonicity regularization is introduced into the loss function. By limiting the variation in predicted values ​​between adjacent time steps, the non-negativity of the remaining service life and the monotonically decreasing physical consistency over time are ensured. This design effectively avoids unreasonable predictions and makes the model output more aligned with the actual needs of engineering applications.

[0090] The end-to-end model improves the convenience of operation and maintenance: Traditional remaining useful life prediction methods usually require multiple steps of feature engineering and model selection. Based on this, an end-to-end deep learning model is designed to achieve a direct mapping from the original signal to the remaining useful life prediction. This not only reduces the need for manual intervention but also simplifies the model training and deployment process. In practical applications, the end-to-end model can quickly adapt to different data and operating conditions and achieve efficient prediction, greatly improving the convenience and efficiency of operation and maintenance.

[0091] Enhanced interpretability: In industrial applications, especially in equipment health management systems, model interpretability is crucial. Maintenance personnel and decision-makers need to understand how the model makes predictions and which features play a key role in these predictions. Therefore, a FastKAN regression head is introduced into the model, and the interpretability of model decisions is enhanced by visualizing the response curves of the radial basis functions and time-series importance heatmaps. Furthermore, by displaying the frequency band energy and mechanistic characteristics of the envelope spectrum, the interpretability of the model output is further enhanced, enabling maintenance personnel to make more informed decisions based on the prediction results.

[0092] Based on the same inventive concept, this application also provides a bearing remaining service life prediction device corresponding to the bearing remaining service life prediction device method. Since the principle of the device in this application is similar to the bearing remaining service life prediction method described above in this application, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0093] Please refer to Figure 5 , Figure 5 A schematic diagram of the bearing remaining service life prediction device according to an embodiment of this application is shown; the bearing remaining service life prediction device includes: The acquisition module 501 is used to acquire the time-domain vibration signal collected by the vibration sensor deployed on the bearing, and to preprocess the time-domain vibration signal to generate a continuous vibration signal. The first extraction module 502 is used to slice the continuous vibration signal into a sliding window to form a vibration signal sequence, extract multi-dimensional features of the vibration signal in each window of the vibration signal sequence, construct an initial high-dimensional feature vector, concatenate the initial high-dimensional feature vector with the working condition code corresponding to the bearing, and perform standardization processing according to the bearing instance to obtain a vibration feature sequence; wherein the multi-dimensional features are extracted from the vibration signal in each window of the vibration signal sequence, namely time-domain statistical features, frequency-domain spectral features, nonlinear features and time-frequency domain features. The second extraction module 503 is used to input the vibration feature sequence into a pre-trained bearing remaining service life prediction model, and to perform a convolution operation on the vibration feature sequence in the time dimension through the time-series feature extraction module in the bearing remaining service life prediction model to extract the local features of the vibration feature sequence, and to process the extracted local features to capture the forward and backward long-term dependencies of the vibration feature sequence, and output a vibration signal time feature representation that integrates local and global time-series information. The processing module 504 is used to process the time feature representation of the vibration signal through multiple radial basis functions in the regression module of the bearing remaining service life prediction model, obtain the response value of each radial basis function, and linearly combine the response values ​​of each radial basis function to map and obtain the bearing remaining service life prediction result.

[0094] In some embodiments, the processing module includes: By linearly combining the response values ​​of multiple radial basis functions, the predicted remaining service life of the bearing is obtained by mapping multiple quantiles; the quantiles include 0.1, 0.3, 0.5, 0.7 and 0.9 quantiles; The specified quantile prediction value is used as the point estimate of the bearing's remaining service life, and the prediction interval of the bearing's remaining service life is determined based on the bearing's remaining service life prediction values ​​of multiple quantiles. The point estimate and prediction range of the remaining service life of the output bearing are used as the prediction results of the remaining service life of the bearing.

[0095] Based on the same inventive concept, this application also provides a bearing remaining service life prediction system corresponding to the bearing remaining service life prediction device method. Since the principle of the system in this application is similar to the bearing remaining service life prediction method described above in this application, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be described again.

[0096] This application provides a bearing remaining service life prediction system, the system including: a vibration sensor, an edge computing unit, and a cloud computing center; The sensor is used to acquire the original time-domain vibration signal of the bearing; The edge computing unit is used to execute the steps of the bearing remaining service life prediction method. The cloud computing center is used to train the bearing remaining service life prediction model to obtain the trained bearing remaining service life prediction model.

[0097] Based on the same inventive concept, this application also provides a computer-readable storage medium corresponding to the bearing remaining service life prediction device method. Since the principle of the computer-readable storage medium in this application is similar to the bearing remaining service life prediction method described above in this application, the implementation of the computer-readable storage medium can refer to the implementation of the method, and the repeated parts will not be described again.

[0098] A computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the bearing remaining service life prediction method.

[0099] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0100] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0101] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0102] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0103] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the remaining service life of a bearing, characterized in that, The method includes: The time-domain vibration signal is acquired by a vibration sensor deployed on the bearing, and the time-domain vibration signal is preprocessed to generate a continuous vibration signal. The continuous vibration signal is sliced ​​into sliding windows to form a vibration signal sequence. Multi-dimensional features of the vibration signal in each window of the vibration signal sequence are extracted to construct an initial high-dimensional feature vector. The initial high-dimensional feature vector and the working condition code corresponding to the bearing are concatenated, and normalization processing is performed according to the bearing instance to obtain a vibration feature sequence. The multi-dimensional features are extracted from the vibration signal in each window of the vibration signal sequence, namely time-domain statistical features, frequency-domain spectral features, nonlinear features, and time-frequency domain features. The vibration feature sequence is input into a pre-trained bearing remaining service life prediction model. The time-series feature extraction module in the bearing remaining service life prediction model performs a convolution operation on the vibration feature sequence in the time dimension to extract local features of the vibration feature sequence. The extracted local features are then processed to capture the forward and backward long-term dependencies of the vibration feature sequence, and output a vibration signal time feature representation that integrates local and global time-series information. The vibration signal time feature representation is processed by multiple radial basis functions in the regression module of the bearing remaining service prediction model to obtain the response value of each radial basis function, and the bearing remaining service prediction result is obtained by linearly combining the response values ​​of each radial basis function.

2. The bearing remaining service life prediction method according to claim 1, characterized in that, The linear combination of the response values ​​of each radial basis function maps to obtain the bearing remaining service life prediction result, including: By linearly combining the response values ​​of multiple radial basis functions, the predicted remaining service life of the bearing is obtained by mapping multiple quantiles; the quantiles include 0.1, 0.3, 0.5, 0.7 and 0.9 quantiles; The specified quantile prediction value is used as a point estimate of the bearing's remaining service life, and the prediction interval of the bearing's remaining service life is determined based on the bearing's remaining service life prediction values ​​of multiple quantiles. The point estimate and prediction range of the remaining service life of the output bearing are used as the prediction results of the remaining service life of the bearing.

3. The bearing remaining service life prediction method according to claim 2, characterized in that, The method further includes: Based on the bearing remaining service prediction results at continuous time points, a visualization is generated showing the point estimate of the bearing remaining service and the trend of the prediction interval changing over time. Based on the transformation process of the vibration signal time feature representation by the regression module, a model decision visualization diagram is generated. The model decision visualization diagram includes: a radial basis function response curve and a feature contribution time series heatmap. The radial basis function response curve shows the activation values ​​of a specific vibration signal time feature in the regression module on all radial basis functions. The feature contribution time series heatmap shows the changes in the contribution of at least some types of vibration signal time feature representations to the prediction results at corresponding time points at continuous time points. Based on the envelope demodulation and spectrum analysis of the original time-domain vibration signal, an envelope spectrum is obtained; the envelope spectrum includes a frequency band energy variation map and an envelope spectrum feature map; the frequency band energy variation map shows at least one preset frequency energy variation curve related to bearing failure.

4. The method for predicting the remaining service life of a bearing according to claim 1, characterized in that, The time-domain statistical characteristics include the signal's mean, variance, skewness, kurtosis, peak factor, impulse factor, and margin factor. The frequency domain spectral features include: the spectral centroid and spectral flatness obtained by calculating the power spectral density of the vibration signal in the window, and the envelope spectral features extracted using the envelope demodulation method; The nonlinear features include detrended fluctuation analysis fractal dimension, sample entropy, and multi-scale sample entropy; The time-frequency domain features are obtained by performing wavelet transform on the vibration signal of the window.

5. The method for predicting the remaining service life of a bearing according to claim 1 or 2, characterized in that, The bearing remaining service life prediction model is obtained by training based on the composite loss function in the bearing remaining service life prediction model. The composite loss function includes a regression main loss term, a monotonicity regularization term, and a quantile regression loss term; The regression main loss term is used to evaluate the difference between the predicted value and the true label; The monotonicity regularization term is used to constrain the predicted values ​​at adjacent time steps, penalizing cases where the predicted values ​​violate the physical law of monotonically decreasing with time. The quantile regression loss term optimizes the ball loss corresponding to multiple quantiles simultaneously.

6. The method for predicting the remaining service life of a bearing according to claim 1, characterized in that, The bearing remaining service life prediction model is trained based on a sample dataset. The sample dataset was constructed based on the following method: The complete continuous sample vibration signal from the start of operation of the sample bearing to its failure is sliced ​​into sliding windows to generate a sequence of sample signal windows arranged in chronological order. Each sample signal window constitutes a training sample. For each sample bearing instance, a remaining service life label is assigned to each training sample. The maximum remaining service life of the sample bearing instance is set, and the inflection point is defined as a preset proportion of the maximum remaining service life. The remaining service life label of training samples whose sample bearing operation time is earlier than or equal to the inflection point is assigned the maximum remaining service life value. The remaining service life label of training samples whose sample bearing operation time is later than the inflection point decreases linearly from the maximum value to zero. Each sample is labeled with a corresponding sample operating condition code; the sample operating condition code includes speed code and load code.

7. A bearing remaining service life prediction device, characterized in that, The device includes: The acquisition module is used to acquire time-domain vibration signals collected by vibration sensors deployed on the bearing, and to preprocess the time-domain vibration signals to generate continuous vibration signals. The first extraction module is used to slice the continuous vibration signal into a sliding window to form a vibration signal sequence, extract multi-dimensional features of the vibration signal in each window of the vibration signal sequence, construct an initial high-dimensional feature vector, concatenate the initial high-dimensional feature vector with the working condition code corresponding to the bearing, and perform standardization processing according to the bearing instance to obtain a vibration feature sequence; wherein the multi-dimensional features are extracted from the vibration signal in each window of the vibration signal sequence, namely time-domain statistical features, frequency-domain spectral features, nonlinear features and time-frequency domain features. The second extraction module is used to input the vibration feature sequence into the pre-trained bearing remaining service life prediction model. The time-series feature extraction module in the bearing remaining service life prediction model performs a convolution operation on the vibration feature sequence in the time dimension to extract the local features of the vibration feature sequence. The extracted local features are then processed to capture the forward and backward long-term dependencies of the vibration feature sequence, and output a vibration signal time feature representation that integrates local and global time-series information. The processing module is used to process the time feature representation of the vibration signal through multiple radial basis functions in the regression module of the bearing remaining service prediction model, obtain the response value of each radial basis function, and linearly combine the response values ​​of each radial basis function to map and obtain the bearing remaining service prediction result.

8. The bearing remaining service life prediction device according to claim 7, characterized in that, The processing module includes: By linearly combining the response values ​​of multiple radial basis functions, the predicted remaining service life of the bearing is obtained by mapping multiple quantiles; the quantiles include 0.1, 0.3, 0.5, 0.7 and 0.9 quantiles; The specified quantile prediction value is used as a point estimate of the bearing's remaining service life, and the prediction interval of the bearing's remaining service life is determined based on the bearing's remaining service life prediction values ​​of multiple quantiles. The point estimate and prediction range of the remaining service life of the output bearing are used as the prediction results of the remaining service life of the bearing.

9. A bearing remaining service life prediction system, characterized in that, include: Vibration sensors, edge computing units, and cloud computing centers; The sensor is used to acquire the original time-domain vibration signal of the bearing; The edge computing unit is used to perform the steps of the bearing remaining service life prediction method according to any one of claims 1-7; The cloud computing center is used to train the bearing remaining service life prediction model to obtain the trained bearing remaining service life prediction model.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the bearing remaining service life prediction method as described in any one of claims 1 to 7.

Images