Rolling bearing residual life prediction method and system based on wear mechanism constraint

Abstract

The application belongs to the technical field of rolling bearing residual life prediction. A rolling bearing residual life prediction method and system based on wear mechanism constraint are proposed. First, the bearing full life cycle monitoring data is collected, the abrasive particle cumulative concentration is calculated, and the residual life label is constructed. Relying on the abrasive particle morphology data, the oxidation, cutting and fatigue three types of wear indexes are solved, and the wear stage is divided according to the concentration degradation law to obtain the prior label. Taking XGBoost integrated tree as the basis prediction model, a physical guided loss function is constructed by fusing data loss and wear mechanism constraint term, and the loss weight and model parameters are self-adaptively optimized by using the wear stage label and multi-class wear index. Finally, the abrasive particle cumulative concentration is input into the optimized model to realize the residual life prediction. The application combines the wear physical mechanism constraint with the traditional integrated learning, makes up for the defects of poor data model interpretability and weak working condition generalization, and improves the bearing residual life prediction accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of rolling bearing remaining life prediction technology, specifically to a method and system for predicting the remaining life of rolling bearings based on wear mechanism constraints. Background Technology

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Rolling bearings are indispensable basic operating components in rotating machinery, and their operating status directly affects the operational stability and safety of the entire equipment. With the continuous upgrading of equipment operation and maintenance concepts, remaining life prediction based on condition monitoring is gradually becoming the mainstream direction in the field of mechanical equipment health management. The industry has now developed multiple life prediction technology routes, with related technologies continuously expanding around equipment monitoring data, wear conditions, and intelligent prediction models. Various prediction models are also being gradually applied to bearing life assessment scenarios. Using monitoring data to assess bearing degradation status and predict life has become a routine technical means in the industrial field to ensure reliable equipment operation.

[0004] Existing bearing remaining life prediction schemes struggle to effectively integrate actual bearing wear mechanisms into model constraint optimization. Most schemes rely solely on surface monitoring data to construct prediction logic, failing to fully utilize the various wear characteristics corresponding to abrasive grain morphology, and unable to dynamically adjust model constraints based on the state differences at different bearing wear stages. This makes it difficult for prediction models to accurately reflect the actual degradation and evolution of bearings. The matching degree between model outputs and actual wear states has room for improvement, failing to balance data fitting effectiveness with the consistency of wear mechanisms, and hindering the stable output of accurate life prediction results under complex wear conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for predicting the remaining life of rolling bearings based on wear mechanism constraints. Instead of establishing complex analytical physical models, it constructs mechanistic indicators such as cutting wear, fatigue wear, and oxidation wear as derivative constraint terms and embeds them into a machine learning loss function. This achieves deep integration of wear mechanism information and a data-driven model, avoiding complex differential equation modeling and parameter solving processes. While maintaining low computational complexity, it ensures that the prediction results simultaneously meet the requirements of consistent degradation trends, reasonable degradation morphology, and monotonic life evolution, thereby significantly improving the model's prediction accuracy, physical consistency, and engineering interpretability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting the remaining life of rolling bearings based on wear mechanism constraints.

[0007] A method for predicting the remaining life of rolling bearings based on wear mechanism constraints includes the following process: Acquire rolling bearing full life cycle monitoring data, extract sampling interval and abrasive number at each sampling point from the full life cycle monitoring data, calculate abrasive cumulative concentration, determine failure threshold based on abrasive cumulative concentration and generate remaining life label; Based on the macroscopic concentration and microscopic morphology information of abrasive particles in the full life cycle monitoring data, wear rate index, wear degree index, oxidation wear index, cutting wear index and fatigue wear index are calculated respectively. Wear stages are divided according to the change characteristics of each index, and stage prior labels are generated. In the pre-trained ensemble tree prediction model, a loss function containing multiple types of constraints is configured. Combined with the remaining lifetime label, stage prior label, oxidation wear index, cutting wear index, and fatigue wear index, the weight of the loss function is adaptively adjusted to complete the parameter tuning of the ensemble tree prediction model. By optimizing the integrated tree prediction model with abrasive particle cumulative concentration as an input parameter, the remaining life prediction results of rolling bearings are obtained.

[0008] In one implementation of the first aspect of the present invention, determining the failure threshold and generating a remaining life tag based on the abrasive particle accumulation concentration includes: Based on the sampling interval and the number of abrasive particles at each sampling point, the cumulative concentration of abrasive particles is continuously calculated, and the moment when the cumulative concentration of abrasive particles first reaches the failure threshold is determined as the failure moment. The initial remaining lifetime is obtained by calculating the time difference between the current running time and the failure time. The initial remaining lifetime is then truncated by setting a maximum lifetime threshold to obtain the remaining lifetime label.

[0009] In one implementation of the first aspect of the present invention, the oxidation wear index, cutting wear index, and fatigue wear index are calculated based on the microstructure information of the abrasive particles, including: Based on the equivalent circle diameter, roundness, standard deviation of boundary curvature, and aspect ratio, the abrasive particles in the full life cycle monitoring data are classified and the original quantity of each type of abrasive particle is counted. The average proportions of oxidized abrasive particles, cutting abrasive particles, and fatigue abrasive particles were calculated by resampling, and the oxidation wear index, cutting wear index, and fatigue wear index were obtained respectively.

[0010] In one implementation of the first aspect of the present invention, generating a priori stage labels based on the characteristics of changes in abrasive particle cumulative concentration includes: Based on the fluctuation status and growth rate of each indicator, multiple wear stages are divided, the time intervals and stage transition nodes of each wear stage are marked, and stage prior labels are generated for constraint weight allocation and model parameter initialization.

[0011] In one implementation of the first aspect of the present invention, building an integrated tree prediction model includes: constructing an integrated tree prediction model formed by combining multiple regression trees, setting the abrasive particle cumulative concentration as the model input feature, and setting the remaining lifetime label as the model training supervision target. The loss function includes a data fitting loss term, a first derivative consistency constraint term, a second derivative shape constraint term, a monotonicity constraint term, a mechanism proportion constraint term, and a model parameter regularization term. The data fitting loss term is used to fit the observed values ​​of abrasive particle cumulative concentration. The first derivative consistency constraint term, the second derivative shape constraint term, and the mechanism proportion constraint term are all set according to the oxidation wear index, the cutting wear index, and the fatigue wear index.

[0012] As a further limitation of the first aspect of the present invention, the calculation of the first derivative consistency constraint term includes: The physical rate values ​​are calculated by combining oxidation wear index, cutting wear index, fatigue wear index with mechanism sensitivity coefficient and shape function; The instantaneous rate of change of abrasive particle cumulative concentration output by the integrated tree prediction model is calculated, and the overall error between the instantaneous rate of change and the physical rate is obtained to obtain the first derivative consistency constraint term.

[0013] As a further limitation of the first aspect of the present invention, the calculation of the second derivative shape constraint term includes: The physical curvature value is calculated by using oxidation wear index, cutting wear index, fatigue wear index, second-order sensitivity coefficient and second-order shape function. The acceleration of the change in abrasive particle cumulative concentration output by the integrated tree prediction model is calculated, the overall error between the change in acceleration and the physical curvature value is obtained, and the second derivative shape constraint term is obtained.

[0014] As a further limitation of the first aspect of the present invention, the calculation of the monotonicity constraint term includes: By comparing the cumulative abrasive particle concentration output by the model at adjacent time points, the deviation value corresponding to the decrease in value is extracted. The average value of all deviation values ​​is then squared to obtain the monotonicity constraint term.

[0015] In one implementation of the first aspect of the present invention, adaptively adjusting the weights of the loss function and completing model parameter tuning includes: Based on the wear characteristics corresponding to the stage prior labels, initial weights are configured for each constraint term in the loss function; The ensemble tree prediction model was trained using abrasive particle cumulative concentration as input feature and remaining life label as training target. The weight values ​​of each constraint term were gradually adjusted, and the mechanism-related learnable parameters and ensemble tree model hyperparameters were optimized by combining cross-validation to complete the model parameter tuning.

[0016] Secondly, the present invention provides a rolling bearing remaining life prediction system based on wear mechanism constraints.

[0017] A rolling bearing remaining life prediction system based on wear mechanism constraints, comprising: The tag generation unit is configured to: acquire rolling bearing full life cycle monitoring data, extract the sampling interval and the number of abrasive particles at each sampling point from the full life cycle monitoring data, calculate the abrasive particle cumulative concentration, determine the failure threshold based on the abrasive particle cumulative concentration, and generate a remaining life tag. The feature segmentation unit is configured to: calculate oxidation wear index, cutting wear index and fatigue wear index based on the abrasive micromorphology information in the full life cycle monitoring data, divide the wear stage according to the change characteristics of abrasive cumulative concentration, and generate stage prior labels; The model tuning unit is configured to: configure a loss function containing multiple constraint terms in the pre-trained ensemble tree prediction model, and adaptively adjust the weights of the loss function by combining the remaining lifetime label, stage prior label, oxidation wear index, cutting wear index, and fatigue wear index to complete the parameter tuning of the ensemble tree prediction model. The life prediction unit is configured to: use the integrated tree prediction model with the abrasive particle cumulative concentration input parameter optimized to obtain the remaining life prediction result of the rolling bearing.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs corresponding indicators and stage labels based on abrasive particle monitoring data, integrating wear mechanisms into the model constraint process. This improves upon the limitations of traditional methods that struggle to optimize models in conjunction with actual wear mechanisms. The invention calculates the cumulative concentration of abrasive particles by extracting their quantity and generates remaining life labels. Simultaneously, it obtains various wear indicators based on the microscopic morphology of the abrasive particles and divides wear stages using concentration changes, forming stage prior labels. This fully leverages the wear characteristic information contained in the monitoring data. Furthermore, it adaptively adjusts the loss function weights using stage prior labels and various wear indicators to optimize model parameters. This allows the model constraint rules to adapt to the operating states of bearings at different wear stages, making the model's computational logic more closely aligned with the bearing degradation and evolution patterns. This approach enhances the matching degree between model prediction results and the actual wear state of the equipment, balancing data fitting effects with the consistency of wear mechanisms. It helps the ensemble tree prediction model output more stable and reasonable rolling bearing remaining life prediction results under various complex wear conditions.

[0019] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 A schematic diagram of the technical framework of a rolling bearing remaining life prediction method based on wear mechanism derivative constraints, provided as an exemplary embodiment of the present invention; Figure 2 A schematic diagram of the XGBoost algorithm provided as an exemplary embodiment of the present invention; Figure 3 A schematic diagram comparing RUL prediction results is provided as an exemplary embodiment of the present invention, wherein, Figure 3 (a) in the figure is a schematic diagram of the prediction results of cumulative abrasive particle concentration. Figure 3 (b) in the diagram is a schematic representation of the predicted remaining useful life. Figure 3 (c) in the figure is a schematic diagram of the prediction error of the XGBoost model. Figure 3 (d) in the figure is a schematic diagram of the prediction error of the PIML-XGBoost model; Figure 4 A schematic diagram comparing the accuracy of rolling bearing life prediction provided as an exemplary embodiment of the present invention; Figure 5 A schematic diagram of RUL prediction feature importance analysis is provided as an exemplary embodiment of the present invention, wherein, Figure 5 (a) in the diagram shows the ranking of indicator importance. Figure 5 (b) in the diagram is a summary of SHAP; Figure 6 This is a schematic diagram of a rolling bearing remaining life prediction system based on wear mechanism derivative constraints, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] This implementation proposes a method for predicting the remaining life of rolling bearings based on wear mechanism derivative constraints. It eliminates the need for complex analytical physical models by constructing derivative constraint terms from mechanistic indicators such as cutting wear, fatigue wear, and oxidation wear, embedding them into a machine learning loss function. This achieves a deep fusion of wear mechanism information and a data-driven model. Compared to traditional methods, this invention avoids complex differential equation modeling and parameter solving, maintaining low computational complexity while ensuring the prediction results simultaneously meet the requirements of consistent degradation trends, reasonable degradation morphology, and monotonic life evolution. This significantly improves the model's prediction accuracy, physical consistency, and engineering interpretability. Furthermore, this invention integrates wear stage identification results to achieve adaptive adjustment of mechanistic constraint weights under different degradation stages. This allows the model to dynamically optimize the prediction process based on changes in the dominant wear mechanism, enhancing its robustness and generalization ability under complex operating conditions and multi-stage degradation. Combined with SHAP interpretation analysis, it can also reveal the contribution patterns of different wear mechanism indicators to the life prediction results, providing physically meaningful interpretations for equipment operation and maintenance decisions.

[0025] Specifically, the rolling bearing remaining life prediction method based on wear mechanism derivative constraints of the present invention consists of six steps. First, RUL labels are constructed as input data for training the model; then, wear mechanism indices are constructed to dynamically adjust the physical constraint model parameters; next, prior labels based on wear stage identification are constructed to provide prior information for the weight allocation of each constraint in the loss function; a PIML-XGBoost model is established, and a loss function based on wear mechanism constraints is designed; then, the model parameters are selected and tuned; finally, the prediction results are evaluated and SHAP interpretability analysis is performed.

[0026] Technical framework such as Figure 1 As shown, the input on the left contains two types of data: training data and... (i.e., degenerative feature sequence) With corresponding remaining service life label ) as input to the XGBoost model; simultaneously, mechanistic indicators Wear stage label The parameters are input together into the parameter selection and tuning module below to determine the values ​​of the mechanism parameters. (Key parameters in the corresponding wear evolution model) and constraint weight initialization The core component is the XGBoost model, which consists of multiple decision trees (tree 1, tree 2, ..., tree 3). The system integrates and iteratively trains each tree, outputting residual 1, residual 2, ..., residual 3. These residuals, along with the model predictions, are fed into the physical loss function module on the right. This module consists of two parts: data loss. (Empirical risk items such as mean square error) and physical losses (Physical consistency constraints based on wear mechanism) The weighted sum of the two forms the total loss function. This loss function guides the model optimization process, continuously updates the model parameters during iterations, and ultimately outputs the prediction result. This refers to the predicted degradation trajectory and remaining service life. The entire framework achieves a deep integration of data-driven (XGBoost) and physical mechanisms (wear evolution model). By imposing mechanistic constraints on the prediction results through physical loss terms, it improves the rationality, stability, and cross-condition generalization ability of the predictions.

[0027] Step S101: Lifetime prediction tag construction.

[0028] (1) Constructing RUL labels.

[0029] In the supervised learning-based Remaining Useful Life (RUL) prediction problem, the core of label construction lies in transforming the entire lifecycle degradation process of equipment into explicit remaining useful life labels, thereby establishing a quantitative mapping relationship between degradation characteristics and lifespan. Let the rolling bearing, during its operation, be at time... If the failure state is reached, then for any time... Its remaining useful life is defined as the remaining time interval from the current time to the time of failure, that is: (1); in, Indicates the time of equipment failure. Indicates the current running time. This indicates the remaining service life of the rolling bearing at time t. Indicates time The corresponding abrasive cumulative concentration, The failure threshold represents the cumulative concentration of abrasive particles.

[0030] As can be seen from this definition, RUL decreases monotonically with time and satisfies the boundary conditions: (2); in, This represents the remaining service life of the rolling bearing at its initial operating time t=0; This represents the remaining service life of a rolling bearing at the point of failure. This definition is called the absolute time (RUL) definition, and its advantage lies in its clear physical meaning, accurately reflecting the evolution of the equipment's lifespan. However, this method requires training data to contain complete lifecycle data to accurately determine the failure point. Therefore, labels are usually built based on full-life bench tests or long-term online monitoring data.

[0031] (2) Define degradation index.

[0032] In practical engineering, equipment failure is often determined by the degradation index reaching a certain critical threshold, rather than a complete loss of function in the strict sense. For rolling bearing wear systems based on oil monitoring, abrasive particle accumulation concentration is an important physical indicator reflecting the degree of wear, exhibiting a monotonically increasing trend over time. When the accumulation concentration reaches a critical value, it indicates that the wear has progressed to an unacceptable state. Therefore, abrasive particle accumulation concentration is used... As a degradation indicator, it conforms to the actual logic of engineering failure judgment and is defined as follows: (3); in, The sampling interval is... The number of sampling points. This represents the total number of abrasive particles within a single sampling point. Representing the The abrasive volume correlation correction coefficients corresponding to each sampling point.

[0033] (3) Determine the failure threshold.

[0034] The threshold of the degradation index is used as the failure criterion, that is: (4); in, For a moment abrasive particle accumulation concentration, This is the failure threshold. When the cumulative abrasive particle concentration first reaches or exceeds this threshold, the equipment is considered to have entered a failure state.

[0035] threshold The appropriate selection of the threshold has a decisive impact on the accuracy of RUL labels. If the threshold is too low, it will lead to an underestimation of lifespan. If the threshold is too high, it may cause a delay in failure identification. Therefore, this invention combines engineering experience with statistical optimization methods to determine the optimal threshold. First, the problem of predicting abrasive particle accumulation concentration is transformed into a failure discrimination problem by defining a binary classification label: (5); Then, receiver operating characteristic (ROC) curves are constructed based on different thresholds, by maximizing the Youden exponent. Determine the optimal threshold: (6); Where TPR represents the true positive rate, FPR represents the false positive rate, and the optimal threshold is... correspond: (7); After obtaining the failure time, a complete RUL tag sequence can be constructed based on the full lifecycle data: (8); in, The number of sampling points. This represents the label indicating the remaining service life of the rolling bearing corresponding to the i-th sampling point. Representing the The running time corresponding to each sampling point.

[0036] (4) Optimize tag data.

[0037] However, in actual training, the prolonged health phase of the device results in an excessively high proportion of large right-to-weight (RUL) samples, while the proportion of samples nearing failure is relatively small, leading to an imbalanced data distribution and affecting the stability of model training. Therefore, a RUL truncation and normalization strategy is introduced to improve the label distribution characteristics. First, a maximum RUL threshold is set. Define the truncated RUL as: (9); in, Representative moment The corresponding label indicating the remaining service life after truncation. This represents the maximum remaining useful life threshold, used for truncation processing.

[0038] In summary, the failure time is determined by a failure criterion based on the abrasive particle cumulative concentration threshold, and RUL labels are constructed using an absolute time definition. At the same time, the physical consistency and statistical stability of the labels are improved by threshold optimization and label truncation, thereby providing high-quality supervision data for the subsequent training of the RUL prediction model based on XGBoost.

[0039] Step S102: Construction of wear mechanism indicators.

[0040] To quantitatively characterize the wear mechanism of rolling bearings during operation, a mechanistic index system was constructed based on the mapping relationship between abrasive morphology and wear mechanism, including the proportion of three types of abrasive particles: oxidation, cutting, and fatigue. First, by analyzing the microscopic characteristics of abrasive particles (morphology, size, boundary texture, etc.), a correspondence between abrasive particle type and typical wear mechanism was established: normal abrasive particles are thin flakes (0.5~15μm), cutting abrasive particles are slender cutting shapes (25~100μm), fatigue abrasive particles are thin blocks with pits or depressions (10~100μm), and oxidation abrasive particles are smooth spherical shapes.

[0041] Based on this, a three-level discrimination tree is used to classify abrasive particles: the first level distinguishes normal and abnormal abrasive particles by equivalent circular diameter (threshold 15μm); the second level identifies oxidized abrasive particles by combining roundness (0.40) and standard deviation of boundary curvature (0.70); and the third level distinguishes cutting abrasive particles and fatigue abrasive particles by aspect ratio and roundness (threshold 0.38). The initial counts of the four types of abrasive particles are then obtained.

[0042] Considering the significant fluctuations in the total number of abrasive particles in a single sampling, directly using the original count percentage would lead to unstable estimations. Therefore, a nonparametric bootstrap resampling method is introduced. The original sample is resampled multiple times with replacement. The mean of the percentage of each type of abrasive particle after each resampling is calculated as a point estimate, and a 95% confidence interval is constructed using the percentile method. This yields robust percentages of oxidized abrasive particles, cutting abrasive particles, and fatigue abrasive particles, along with their uncertainty ranges. These indicators enable cross-scale quantitative characterization from microscopic morphology to macroscopic wear state, providing a data foundation for subsequent physical constraint parameters.

[0043] Step S103: Construct prior labels based on wear stage identification.

[0044] For RUL prediction, stage identification not only characterizes the wear evolution state but also provides important priors for subsequent mechanism modeling. Therefore, it is necessary to standardize the stage identification results and construct stage labels, stage boundaries, and mechanism labels that can be used in RUL prediction models, providing a basis for loss function design and parameter initialization.

[0045] During the accelerated life cycle of a rolling bearing, the first The stage label of each sample is denoted as ,in , This represents the total number of stages identified within this lifecycle. Furthermore, the transition moments between stages are represented as a set of stage boundaries. ,in, Indicates the first The first stage towards the The time points at which each stage changes are also marked. There are also labels for each stage, such as break-in, stabilization, cutting, oxidation, and acceleration.

[0046] It should be noted that the stage labels are not directly used as the RUL supervision target, but rather as the basis for the weight allocation of each constraint term and the initial parameter settings in the loss function of the RUL prediction model. Specifically, different wear stages correspond to different degradation rates, dominant mechanistic features, and state evolution patterns. Therefore, the constraint strength should be adaptively adjusted according to the stage differences during model training.

[0047] During the break-in and stabilization periods, although the cumulative abrasive particle concentration generally shows an upward trend, local fluctuations and noise disturbances are relatively significant. Therefore, the model should avoid imposing excessively strong monotonicity penalties on short-term fluctuations during this stage. During the cutting, fatigue, or accelerated degradation stages, the wear evolution rate increases significantly, and the growth trend of the cumulative concentration becomes more stable. At this time, the weights of monotonicity constraints and derivative consistency constraints should be appropriately increased to enhance the model's ability to fit the degradation trend. For the near-failure stage, the weights of shape constraints and mechanism consistency constraints can be further increased to make the prediction results more closely match the rapid deterioration characteristics of rolling bearings nearing failure. Thus, stage labeling enables stage-adaptive adjustment of the loss function parameters.

[0048] Step S104: PIML-XGBoost model construction.

[0049] (1) Construction of XGBoost basic prediction model.

[0050] XGBoost is an ensemble learning algorithm based on gradient boosting decision trees. Its core idea is to iteratively build multiple weak learners and then weighted combine them to form a strong learner, thereby approximating the optimal solution of the objective function. Its principle is as follows: Figure 2 As shown, starting from the beginning node, the training dataset is first input, and hyperparameters (number of iterations K and learning rate η) are set; then the initialization phase begins, followed by building the tree structure, and then calculating the loss function; then the core loop begins: for the k-th iteration (k=1,2,…,K), the following steps are executed sequentially: build the k-th iteration... The process involves: traversing a regression tree, visiting features and split points, calculating the gain (information gain or reduction in squared loss), calculating leaf node weights (weighted summation of residuals to obtain the optimal leaf output value), calculating the optimal target value (the prediction correction term for the current tree), updating the model predictions (weighting the predictions of the new tree according to the learning rate η and adding them to the current model output); and then determining... Is it equal to If the result is "No", then return to continue building the next tree; if the result is "Yes", then feed the final ensemble model into the strong learner (i.e., by...). The complete XGBoost model is constructed by weighting weak regression trees, and finally outputs the prediction results and reaches the termination node. The entire process strictly follows the gradient boosting framework of XGBoost, achieving the processing of degenerate features by fitting negative gradients (residuals) round by round and optimizing the splitting criterion and leaf node weights. and remaining lifespan The high-precision fitting provides a data-driven foundation for the subsequent PIML framework coupled with the physical loss function.

[0051] Let the training dataset be: (10); in, The input feature is the cumulative concentration of abrasive particles. The target value is the cumulative concentration of abrasive particles at the next moment. It is the set of real numbers.

[0052] XGBoost model predictions Represented as The summation form of the regression trees: (11); in, For the i-th regression tree, To return to the tree space, Representing the A tree-independent regression model, where K represents the total number of regression trees in the model.

[0053] The regression tree space is defined as: (12); Wherein, mapping function Sample Assign to a leaf node index. The leaf node weight vector, The calculation logic for a single regression tree.

[0054] XGBoost objective function It consists of a loss function and a regularization term: (13); in, The original loss function, For regularization terms, This represents the total number of samples in the training set.

[0055] The commonly used loss function is mean squared error: (14); The regularization term typically takes a double penalty based on the number of leaf nodes and the leaf weights: (15); in, The number of leaf nodes in the tree. and These are the leaf number penalty and leaf weight regularization coefficients, respectively.

[0056] In each iteration, the algorithm uses a second-order Taylor expansion to approximate incremental loss and selects the optimal split using a gain criterion, thereby efficiently constructing a tree structure and controlling the generalization ability through hyperparameters such as learning rate, row / column sampling, and maximum depth.

[0057] However, the purely data-driven XGBoost model still has limitations, namely, its prediction results mainly rely on statistical correlation, making it difficult to guarantee that the output trajectory is consistent with the wear and degradation mechanism. To address this issue, this invention introduces a wear mechanism constraint term into the XGBoost objective function, constructing a PIML-XGBoost model. This algorithm does not simply modify the underlying mathematical form of XGBoost, but rather superimposes a physical constraint term on the original objective function, thereby improving the physical rationality and interpretability of the prediction results. By introducing physical bias into the loss function, the model can achieve a balance between data-driven approaches and physical laws.

[0058] To guide students through physics knowledge, a physics penalty term is introduced based on equation (13). Construct a joint optimization objective for PIML-XGBoost : (16); in, This is the weighting coefficient for the physics penalty, used to balance the contribution of data-driven approaches and physics knowledge.

[0059] (2) Design of mechanism-constrained loss function.

[0060] First, from the perspective of wear mechanism, the degradation process of rolling bearings typically exhibits an evolutionary trend from slow wear to accelerated failure. The corresponding abrasive particle accumulation concentration generally shows a monotonic increase, with different rates of change and curve shapes at different stages. Therefore, the model prediction results should satisfy the following basic physical characteristics: first, the monotonicity of the overall trend, i.e., the degradation index increases irreversibly over time; second, the consistency of the rate of change, i.e., the first derivative of the predicted curve should be consistent with the actual degradation trend; and third, the rationality of the curve shape, i.e., a significant nonlinear enhancement characteristic should be exhibited in the accelerated stage.

[0061] Suppose the model predicts the cumulative abrasive particle concentration as follows: The observed value is The obtained mechanism proportion is Based on the above analysis, the overall loss function takes the following form: (17); in, The loss is the fitting loss for the observed values. and These correspond to the consistency constraints of the first and second derivatives driven by the mechanism, respectively. To constrain the monotonicity of cumulative concentration, For the proportion of mechanisms or phased constraints, For model parameter regularization, The following section provides the mathematical definitions and physical interpretations of each constraint, representing the weight hyperparameters for each physical penalty term.

[0062] (2-1) Data Fitting Term .

[0063] Measure the cumulative abrasive concentration predicted by the model Compared with actual observation Point value fitting error The calculation method is as follows: (18); This item is the fitted data itself, avoiding complete constraint from the wear mechanism.

[0064] (2-2) First derivative mechanism consistency term.

[0065] To reflect the contribution of different mechanisms to the instantaneous generation rate, a mechanism-weighted rate model is used as the prior. : (19); in, The mechanism sensitivity coefficient, , The shape function reflects the effects of load cycles, temperature, and frictional work. The first derivative of the model is obtained through automatic differentiation. Corresponding losses: (20); This approach aligns the instantaneous rate of change predicted by the model with the physical rate derived from the mechanism proportion and the working condition proxy, thereby constraining the growth trend of the model on the time axis with physical information, which can improve robustness, especially when extrapolating to the unobserved interval.

[0066] (2-3) Second derivative shape constraint term.

[0067] The influence of the mechanism on curvature / acceleration is characterized in the form of second-order response priors. express: (twenty one); in, , and All are second-order sensitivity coefficients. , For fatigue and oxidation, the corresponding second-order shape functions are... represent Quantitative characteristic prediction values ​​of cutting wear at any time. represent Quantitative characteristic prediction of fatigue wear over time. represent Quantitative characteristic predictions of oxidative wear over time.

[0068] The corresponding loss is: (twenty two); Second-order terms can capture the differences in acceleration abruptly during the critical stage of fatigue cracking, the slow acceleration during the oxidation stage, and the near-zero acceleration during the cutting stage. This allows the model to distinguish the dynamic characteristics generated by different wear mechanisms, thereby enhancing its ability to perceive changes in the rate of deterioration in RUL prediction.

[0069] (2-4) Monotonicity constraint term for cumulative abrasive particle concentration.

[0070] Since the cumulative concentration cannot be reduced without rinsing or filtration, a soft-constraint penalty is introduced to predict the descent phase: (twenty three); (2-5) Mechanism proportion constraint.

[0071] This ensures that the proportion of mechanisms follows a probability distribution: non-negative and summing to 1. The calculation method is as follows: (twenty four); in, represent Time of the first The predicted proportion of wear mechanisms, where N represents the total number of time steps.

[0072] (2-6) Regularization terms.

[0073] This step helps prevent parameter overfitting, especially by applying a reasonable boundary value to the wear mechanism sensitivity coefficient. The calculation method is as follows: (25); in, Represents the set of all learnable parameters; This represents the L2 regularization weight coefficient.

[0074] After constructing the mechanistic loss function, model performance largely depends on the proper setting of various hyperparameters and constraint weights. Improper parameter configuration can lead to underfitting or overfitting even with the introduction of physical constraints. Therefore, it is necessary to systematically optimize the model parameters by combining mechanistic characteristics and prior information to fully leverage the performance advantages of the PIML-XGBoost model.

[0075] Step S105: Model parameter selection and tuning.

[0076] To ensure a balance between prediction accuracy, physical consistency, and training stability in the PIML-XGBoost model, it is necessary to systematically design and optimize the key learnable parameters and loss function weights. Unlike traditional XGBoost models that rely solely on data-driven learning, the model constructed in this invention incorporates both wear mechanism indices and stage label information. Therefore, the parameter system can be divided into two categories: one is the key learnable parameters driven by wear mechanism indices, and the other is the loss function weights controlled by stage labels. Together, they determine the model's ability to characterize the degradation process and the strength of physical constraints. The physical meaning, acquisition methods, and optimization strategies for each type of parameter are explained below.

[0077] (1) Key learnable parameters.

[0078] First-order sensitivity coefficient This coefficient represents the contribution of each mechanism to the concentration rate per unit time and per unit proportion. Its value range is estimated through piecewise regression of historical complete lifetime curves and optimized within the boundary during training as a learnable parameter.

[0079] Fatigue shape function This function characterizes the effect of rate on cyclic load under fatigue mechanism. Cumulative loop count The amplification relationship.

[0080] Oxidation shape function This function reflects the influence of temperature and frictional work on the rate of oxidation particle generation, and is related to rotational speed and ambient temperature.

[0081] Second-order sensitivity coefficient With second-order shape function , These parameters control the contribution of each mechanism to the acceleration and curvature of the concentration curve. Fatigue often corresponds to a large positive second-order response, such as a sudden increase; cutting corresponds to a near-zero second-order response, such as a steady state; and oxidation may have a moderate to slow second-order response. The values ​​are obtained by second-order difference of historical mechanism index curves.

[0082] (2) Hyperparameter weights , , , and .

[0083] The above parameter weights are used to control the influence of physical priors such as derivatives, shape, and monotonicity on model training, and their values ​​are all between 0 and 1. While introducing mechanistic constraints, how to reasonably allocate the weights of each loss term is a key factor affecting model performance. This invention utilizes the obtained stage labels... Set initial values ​​for it, and the assignment rules are shown in Table 1.

[0084] Table 1: Mapping Relationship Between Stage Labels and Loss Function Weights

[0085] In actual training, a phased training strategy is adopted: 1) Pre-training phase: Based on the phase labels, set the initial weights of each loss term according to the table to enable the model to have preliminary physical constraint capabilities.

[0086] 2) Introduce first-order constraints: gradually increase If linearly increased to Meanwhile, we observed the balance between the first-order physical residual and the point fitting error on the validation set.

[0087] 3) Introduce second-order and monotonicity constraints: then gradually introduce... and The final weights are then adjusted through cross-validation or Bayesian optimization.

[0088] 4) Subsequently, gradient normalization is used to automatically adjust the weights of different loss terms to alleviate optimization conflicts caused by differences in the gradient magnitudes of different terms during training.

[0089] By incorporating wear mechanism indices into the modeling of key learnable parameters and adaptively adjusting the weights of the loss function using stage labels, the model parameters have shifted from empirical setting to data-driven and mechanism-constrained co-optimization. This strategy not only enhances the model's adaptability to different degradation stages but also provides parameter-level support for the physical interpretation of subsequent RUL prediction results.

[0090] Step S106: Evaluation of prediction results and SHAP interpretability analysis.

[0091] (1) Evaluation of prediction results.

[0092] The proposed PIML-XGBoost model was compared with the standard XGBoost model, and the following time series prediction curves and error comparison results were obtained: Figure 3 As shown.

[0093] Figure 3 The cumulative wear particle concentration prediction results shown in (a) indicate that the PIML-XGBoost model has a significant advantage in fitting accuracy compared to the standard XGBoost model, and its prediction curve is highly consistent with the actual wear evolution trajectory. Figure 3The remaining useful life prediction results shown in (b) further demonstrate that PIML-XGBoost suppresses the drastic fluctuations of the traditional model in the early stages of prediction. The standard XGBoost model exhibits significant prediction bias in the early stages of operation (around 500 min), while the prediction robustness of the model is significantly enhanced after introducing physical mechanism constraints. Figure 3 (c) and Figure 3 The comparison results given in (d) show that the standard XGBoost model frequently exhibits significant underestimation and overestimation fluctuations in the mid-to-late stages, and the error distribution is relatively discrete. In contrast, the error curve of PIML-XGBoost fluctuates around the zero mark overall, and the fluctuation amplitude is significantly reduced. This indicates that by introducing wear mechanism-related indicators, the physical mismatch problem of the pure data-driven model can be effectively alleviated, thereby improving the predictive reliability of the model in the later stages of wear evolution and critical degradation stages.

[0094] To compare the timing deviations of the two models in predicting remaining useful life during the experiment, this invention calculated the advance and delay times of the predicted occurrence time of each test sample by XGBoost and PIML-XGBoost relative to the actual failure time, as shown in Table 2.

[0095] Table 2: Comparison of RUL Prediction Performance

[0096] Quantitative analysis of the average delay time and average lead time in the table reveals that the PIML-XGBoost model exhibits a significant advantage in the timeliness of RUL prediction. Compared to the standard XGBoost model, the PIML-XGBoost model shows a significant reduction in average delay time, indicating that physical information constraints effectively suppress the model's lagging assessment of equipment degradation states. Simultaneously, the prediction bias of the PIML-XGBoost model is also significantly reduced in the average lead time dimension, further demonstrating the stability of the model's predictions. Overall, considering all time bias indicators, the method based on physical information constraints effectively improves the consistency between the prediction results and the actual evolution trajectory of the equipment.

[0097] Figure 4 This paper compares the performance of the standard XGBoost model and the PIML-XGBoost model in predicting remaining useful life. Experimental results show that the PIML-XGBoost model significantly outperforms the standard XGBoost model in the remaining useful life prediction task, exhibiting higher accuracy and robustness. Regarding prediction error, this model significantly reduces both the mean absolute error (MAE) and the root mean square error (RMSE), with MAE decreasing by 17% and RMSE decreasing by 12%. Furthermore, the coefficient of determination R0, which characterizes the goodness of fit of the model, is significantly improved. 2It also achieved a substantial improvement, with an increase of 18%. These results indicate that introducing physical prior constraints can correct the prediction bias of purely data-driven models, enhancing the model's fitting ability while improving its accuracy in capturing complex evolutionary patterns.

[0098] (2) Interpretability analysis.

[0099] To elucidate the decision-making mechanism of machine learning models in the process of predicting remaining useful life, this invention introduces the SHAP interpretation tool to perform quantitative interpretability analysis of model features from two dimensions: global importance and local contribution.

[0100] like Figure 5 As shown in (a), the feature importance ranking reveals the weight distribution between data-driven features and physical mechanism indicators in model decision-making. Although there are some differences in the evaluation results based on SHAP values ​​and XGBoost native indicators, the time-series feature C(t-1) ranks first in both methods, indicating that historical degradation status is the core basis for model prediction benchmark construction. Wear rate TGSV and wear degree The weight of the index followed closely, indicating that numerical indicators closely related to abrasive particle concentration are still the dominant factors affecting the predicted value of abrasive particle cumulative concentration. Although the wear mechanism indices OPI, FPI, and CPI have relatively low overall weights, as high-order physical features, they provide key microscopic mechanism constraints in the wear stage transition process, which helps to improve the model's discrimination ability and prediction stability in the nonlinear degradation stage.

[0101] like Figure 5 As shown in (b), the SHAP summary plot is used to analyze the marginal impact of each indicator on the prediction results. For the main feature C(t-1), wear rate TGSV, and wear severity index, the yellow sample points corresponding to high values ​​generally show positive SHAP values. This indicates that as the wear concentration and evolution rate increase, the predicted value output by the model increases accordingly, which is consistent with the physical evolution law of wear damage accumulation. The SHAP value distribution of the time lag features C(t-3) to C(t-5) shows obvious nonlinear diffusion characteristics, reflecting that the PIML-XGBoost model can characterize complex long-term dependencies and has a strong nonlinear mapping ability when dealing with non-stationary wear data. The SHAP distribution of wear mechanism indices OPI, FPI, and CPI is relatively concentrated, but shows significant marginal contributions in specific intervals, reflecting certain threshold characteristics. This indicates that when wear enters the acceleration stage or is close to failure, the mechanism indices can correct the prediction trajectory through physical constraints, thereby suppressing the random bias that may be generated by the pure data-driven model.

[0102] Based on the above interpretability analysis, the characteristic contribution patterns revealed by the SHAP method can provide a quantitative basis for condition-based maintenance of bearings. Specifically, a maintenance triggering mechanism can be constructed with the SHAP value of the time-series characteristic C(t-1) and the wear rate TGSV as its core: when the SHAP values ​​of both are consistently higher than the normal range and exceed a preset threshold, it indicates that wear has entered an accelerated accumulation stage, at which point short-term maintenance or increased monitoring frequency should be arranged; when the SHAP values ​​of wear mechanism indicators OPI, FPI, or CPI show a significant positive jump, it indicates a transition in the wear stage (such as from steady-state wear to severe wear), and an early warning should be triggered immediately, suggesting that the lubricating medium be replaced or the bearing surface condition be checked in the next downtime window. In addition, a linkage rule between the SHAP trend chart and the remaining service life prediction value can be established: if the marginal contribution of high SHAP characteristics (C(t-1), TGSV) increases exponentially, even if the current predicted life is still acceptable, it should be recommended to prepare spare parts in advance and shorten the maintenance cycle, thereby realizing the transformation from "periodic maintenance" to "characteristic contribution-driven dynamic maintenance", effectively avoiding sudden failures and reducing the total life cycle maintenance cost.

[0103] Figure 6 A rolling bearing remaining life prediction system based on wear mechanism constraints is shown, comprising: The tag generation unit 601 is configured to: acquire rolling bearing full life cycle monitoring data, extract the sampling interval and the number of abrasive particles at each sampling point from the full life cycle monitoring data, calculate the abrasive particle cumulative concentration, determine the failure threshold based on the abrasive particle cumulative concentration, and generate a remaining life tag. The feature segmentation unit 602 is configured to: calculate wear rate index, wear degree index, oxidation wear index, cutting wear index and fatigue wear index respectively based on the macroscopic concentration and microscopic morphology information of abrasive particles in the whole life cycle monitoring data, divide the wear stage according to the change characteristics of each index, and generate stage prior labels; The model tuning unit 603 is configured to: configure a loss function containing multiple constraint terms in the pre-trained ensemble tree prediction model, and adaptively adjust the weights of the loss function by combining the remaining lifetime label, stage prior label, oxidation wear index, cutting wear index, and fatigue wear index to complete the parameter tuning of the ensemble tree prediction model. The life prediction unit 604 is configured to: obtain the remaining life prediction result of the rolling bearing by using the integrated tree prediction model after the abrasive particle cumulative concentration input parameter is optimized.

[0104] It is understood that the aforementioned units can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of the present invention. The aforementioned units are based on logical functional division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of the present invention, the system may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0105] According to another embodiment of the present invention, the system of this embodiment can be constructed by running a computer program (including program code) capable of performing the steps involved in the corresponding method of the present invention on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). The computer program can be recorded on, for example, a computer-readable recording medium, loaded into the aforementioned computing device through the computer-readable recording medium, and run therein.

[0106] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can implement the described functions using different methods for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0107] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic cable, digital cable) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of rolling bearings based on wear mechanism constraints, characterized in that, Includes the following processes: Acquire rolling bearing full life cycle monitoring data, extract sampling interval and abrasive number at each sampling point from the full life cycle monitoring data, calculate abrasive cumulative concentration, determine failure threshold based on abrasive cumulative concentration and generate remaining life label; Based on the macroscopic concentration and microscopic morphology information of abrasive particles in the full life cycle monitoring data, wear rate index, wear degree index, oxidation wear index, cutting wear index and fatigue wear index are calculated respectively. Wear stages are divided according to the change characteristics of each index, and stage prior labels are generated. In the pre-trained ensemble tree prediction model, a loss function containing multiple types of constraints is configured. By combining the remaining lifetime label, stage prior label, oxidation wear index, cutting wear index, and fatigue wear index, the weight of the loss function is adaptively adjusted and the parameters of the ensemble tree prediction model are tuned. By optimizing the integrated tree prediction model with abrasive particle cumulative concentration as an input parameter, the remaining life prediction results of rolling bearings are obtained.

2. The method for predicting the remaining life of rolling bearings based on wear mechanism constraints as described in claim 1, characterized in that, The failure threshold is determined based on the cumulative concentration of abrasive particles, and a remaining life label is generated, including: Based on the sampling interval and the number of abrasive particles at each sampling point, the cumulative concentration of abrasive particles is continuously calculated, and the moment when the cumulative concentration of abrasive particles first reaches the failure threshold is determined as the failure moment. The initial remaining lifetime is obtained by calculating the time difference between the current running time and the failure time. The initial remaining lifetime is then truncated by setting a maximum lifetime threshold to obtain the remaining lifetime label.

3. The method for predicting the remaining life of rolling bearings based on wear mechanism constraints as described in claim 1, characterized in that, Based on the microstructure information of abrasive particles, oxidation wear parameters, cutting wear parameters, and fatigue wear parameters are calculated, including: Based on the equivalent circle diameter, roundness, standard deviation of boundary curvature, and aspect ratio, the abrasive particles in the full life cycle monitoring data are classified and the original quantity of each type of abrasive particle is counted. The average proportions of oxidized abrasive particles, cutting abrasive particles, and fatigue abrasive particles were calculated by resampling, and the oxidation wear index, cutting wear index, and fatigue wear index were obtained respectively.

4. The method for predicting the remaining life of rolling bearings based on wear mechanism constraints as described in claim 1, characterized in that, Based on the fluctuation status and growth rate of each indicator, prior labels for each stage are generated, including: Based on the fluctuation status and growth rate of each indicator, multiple wear stages are divided, the time intervals and stage transition nodes of each wear stage are marked, and stage prior labels are generated for constraint weight allocation and model parameter initialization.

5. The method for predicting the remaining life of rolling bearings based on wear mechanism constraints as described in claim 1, characterized in that, Building an integrated tree prediction model includes: constructing an integrated tree prediction model composed of multiple regression trees, setting the cumulative concentration of abrasive particles as the model input feature, and setting the remaining lifetime label as the model training supervision target; The loss function includes a data fitting loss term, a first derivative consistency constraint term, a second derivative shape constraint term, a monotonicity constraint term, a mechanism proportion constraint term, and a model parameter regularization term. The data fitting loss term is used to fit the observed values ​​of abrasive particle cumulative concentration. The first derivative consistency constraint term, the second derivative shape constraint term, and the mechanism proportion constraint term are all set according to the oxidation wear index, the cutting wear index, and the fatigue wear index.

6. The method for predicting the remaining life of rolling bearings based on wear mechanism constraints as described in claim 5, characterized in that, The calculation of the first derivative consistency constraint term includes: The physical rate values ​​are calculated by combining oxidation wear index, cutting wear index, fatigue wear index with mechanism sensitivity coefficient and shape function; The instantaneous rate of change of abrasive particle cumulative concentration output by the integrated tree prediction model is calculated, and the overall error between the instantaneous rate of change and the physical rate is obtained to obtain the first derivative consistency constraint term.

7. The method for predicting the remaining life of rolling bearings based on wear mechanism constraints as described in claim 5, characterized in that, The calculation of the second derivative shape constraint term includes: The physical curvature value is calculated by using oxidation wear index, cutting wear index, fatigue wear index, second-order sensitivity coefficient and second-order shape function. The acceleration of the change in abrasive particle cumulative concentration output by the integrated tree prediction model is calculated, the overall error between the change in acceleration and the physical curvature value is obtained, and the second derivative shape constraint term is obtained.

8. The method for predicting the remaining life of rolling bearings based on wear mechanism constraints as described in claim 5, characterized in that, The calculation of the monotonicity constraint term includes: By comparing the cumulative abrasive particle concentration output by the model at adjacent time points, the deviation value corresponding to the decrease in value is extracted. The average value of all deviation values ​​is then squared to obtain the monotonicity constraint term.

9. The method for predicting the remaining life of rolling bearings based on wear mechanism constraints as described in claim 1, characterized in that, Adaptively adjust the weights of the loss function and complete model parameter tuning, including: Based on the wear characteristics corresponding to the stage prior labels, initial weights are configured for each constraint term in the loss function; The ensemble tree prediction model was trained using abrasive particle cumulative concentration as input feature and remaining life label as training target. The weight values ​​of each constraint term were gradually adjusted, and the mechanism-related learnable parameters and ensemble tree model hyperparameters were optimized by combining cross-validation to complete the model parameter tuning.

10. A rolling bearing remaining life prediction system based on wear mechanism constraints, characterized in that, include: The tag generation unit is configured to: acquire rolling bearing full life cycle monitoring data, extract the sampling interval and the number of abrasive particles at each sampling point from the full life cycle monitoring data, calculate the abrasive particle cumulative concentration, determine the failure threshold based on the abrasive particle cumulative concentration, and generate a remaining life tag. The feature segmentation unit is configured to: calculate wear rate index, wear degree index, oxidation wear index, cutting wear index and fatigue wear index respectively based on the macroscopic concentration and microscopic morphology information of abrasive particles in the whole life cycle monitoring data; divide the wear stage according to the change characteristics of each index; and generate stage prior labels. The model tuning unit is configured to: configure a loss function containing multiple constraint terms in the pre-trained ensemble tree prediction model, and adaptively adjust the weights of the loss function by combining the remaining lifetime label, stage prior label, oxidation wear index, cutting wear index, and fatigue wear index to complete the parameter tuning of the ensemble tree prediction model. The life prediction unit is configured to: use the integrated tree prediction model with the abrasive particle cumulative concentration input parameter optimized to obtain the remaining life prediction result of the rolling bearing.

Images