A method for joint prediction of bearing remaining life range to point using multi-source uncertainty calibration
By using a shared bidirectional long short-term memory network and a multi-source uncertainty calibration method, the problem of the independence between point prediction and interval estimation in bearing remaining life prediction is solved, the consistency and reliability of point prediction and interval estimation are achieved, the prediction time step is optimized, and reliable life point values and risk intervals are provided.
Patent Information
- Application Number
- CN202611354848.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-17
AI Technical Summary
In existing data-driven methods for predicting the remaining life of bearings, point prediction and interval estimation are independent of each other, lack structural consistency, the prediction time step depends on experience, and the sources of uncertainty are not fully characterized, resulting in prediction intervals that are too wide or have insufficient coverage.
A shared bidirectional long short-term memory network encoder is used to extract bearing degradation features. Combining quantile interval prediction and point prediction branches, the network is trained through a joint objective function to optimize time steps and multi-source uncertainties. Calibration is performed using random deactivation, multiquantile regression, and nearest neighbor residual statistics. Finally, the remaining bearing life point prediction and interval are output.
It achieves consistency constraints between point prediction and interval estimation, optimizes the prediction time step, controls the interval width, and provides reliable lifetime point values and risk intervals, thus providing a reliable basis for maintenance decisions.
Smart Images

Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of mechanical equipment failure prediction and health management technology, and in particular to a method for joint prediction of bearing remaining life range to point using multi-source uncertainty calibration. Background Technology
[0002] Rolling bearings are critical components in rotating machinery, and their health status directly affects the safety, stability, and continuity of equipment operation. Predicting the remaining useful life (RUL) of bearings can provide a basis for selecting maintenance timing, spare parts scheduling, and preventive maintenance.
[0003] Existing data-driven methods for predicting the remaining life of bearings mostly utilize recurrent neural networks, convolutional neural networks, or attention networks to extract degradation features from monitoring data such as vibration, and output a deterministic remaining life point value. This type of point prediction can reflect the expected life, but it cannot simultaneously provide the predicted risk range, and it is difficult to judge the reliability of the prediction results when the degradation mode changes, noise increases, or the sample is sparse.
[0004] To provide a reliable range, some methods construct prediction intervals using residual statistics, random sampling, probability distributions, or quantile regression after point prediction. However, traditional methods typically train the point prediction model first and then construct the intervals independently. This lack of structural consistency constraints between the predicted point values and the interval boundaries can easily lead to problems such as point values falling outside the interval, quantile boundaries intersecting, and intervals being too wide to improve coverage.
[0005] Furthermore, the time step in remaining lifetime prediction affects prediction error, early warning lead time, and interval reliability. Fixed time steps typically rely on empirical settings, making it difficult to simultaneously consider point prediction accuracy, prediction interval coverage, and interval width. Uncertainties in the bearing degradation process also originate from model parameters, observational data noise, and the distribution of local degradation samples, making it difficult for a single uncertainty estimation method to maintain stable coverage across different degradation stages.
[0006] Therefore, there is a need for a bearing remaining life prediction method that can first form a credible interval within a unified network structure, then generate point predictions under interval constraints, and jointly optimize the time step and multi-source uncertainties. Summary of the Invention
[0007] This invention proposes a joint prediction method for bearing remaining life intervals to points using multi-source uncertainty calibration, in order to solve the technical problems of existing methods, such as point prediction and interval estimation being independent of each other, prediction time steps relying on experience, incomplete characterization of uncertainty sources, and excessively wide prediction intervals under high coverage conditions.
[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution, comprising the following steps:
[0009] Step 1: Acquire bearing operating status monitoring data, normalize the data and segment it using a sliding time window, construct a bearing degradation feature sequence, and divide it into training set, calibration set and test set;
[0010] Step 2: Input the bearing degradation feature sequence into the shared bidirectional long short-term memory network encoder to extract shared temporal features that simultaneously contain positive and negative degradation dependencies;
[0011] Step 3: Input the shared time series features into the quantile interval prediction branch to obtain the lower quantile boundary and upper quantile boundary with statistical coverage significance. Then input the shared time series features into the point prediction branch and use the quantile boundaries to perform interval consistency projection on the unconstrained point prediction values.
[0012] Step 4: Train the shared bidirectional long short-term memory network encoder, quantile interval prediction branch, and point prediction branch using a joint objective function composed of point prediction loss, quantile loss, and structure regularization term;
[0013] Step 5: Calculate the point prediction error, prediction interval coverage, and average prediction interval width for multiple candidate time steps, and determine the target time step using the comprehensive score;
[0014] Step 6: Determine the model uncertainty, data uncertainty, and neighborhood uncertainty using random deactivation sampling, quantile regression, and nearest neighbor residual statistics, respectively;
[0015] Step 7: Perform conformal calibration on the calibration samples according to the predicted remaining lifetime, and then sequentially perform candidate scale relaxation calibration, neighborhood uncertainty fusion, monotonic regression calibration, degradation stage hierarchical scale adjustment, global compression and safety factor compensation to obtain the final uncertainty scale;
[0016] Step 8: Based on the predicted bearing remaining life point at the target time step and the final uncertainty scale, output the bearing remaining life point prediction result and the corresponding final prediction interval.
[0017] In step 1, the full lifespan monitoring data of the bearing from healthy operation to failure is acquired according to a fixed sampling period. The features of each dimension are normalized and the sliding time window is segmented according to the preset window length and sliding step size. The time interval from the end of the window to the bearing failure time is used as the true remaining life label. The training set is used to update the network parameters, the calibration set is used to determine the conformal threshold, monotonic mapping and scaling parameters, and the test set is used for inference evaluation.
[0018] In step 2, the shared bidirectional long short-term memory network encoder concatenates the forward hidden state and the reverse hidden state. The relationship between the input and the encoding result is as follows:
[0019]
[0020] in, Indicates the size is The bearing degradation characteristic matrix, Indicates the length of the time window. Indicates the dimension of the input features; Indicates the size is Shared temporal feature matrix, This represents the dimension of the hidden state in a unidirectional long short-term memory network.
[0021] In step 3, the quantile interval prediction branch includes a lower quantile prediction head and a boundary increment prediction head. The lower quantile boundary and the upper quantile boundary are calculated as follows:
[0022]
[0023] in, and They represent the first The lower and upper quantile boundaries of each predicted location Indicates the lower quantile prediction head. This represents the output of the boundary increment prediction head; Positive value mapping keeps the boundary increment non-negative, thus ensuring that the upper quantile boundary is not less than the lower quantile boundary.
[0024] The point prediction branch first outputs unconstrained remaining lifetime point prediction values, and then performs interval consistency projection according to the following relationship:
[0025]
[0026] in, This represents the unconstrained remaining lifetime point prediction value. This represents the projected remaining lifetime point prediction value. By first taking the larger value between the unconstrained point prediction value and the lower quantile boundary, and then taking the smaller value between the result and the upper quantile boundary, the projected point prediction value is located within the corresponding quantile interval.
[0027] In step 4, the shared bidirectional long short-term memory network encoder, quantile prediction branch, and point prediction branch are trained using the following joint objective function:
[0028]
[0029] in, and These represent the mean absolute error loss and the mean square error loss, respectively. Indicates quantile loss. Indicates the parameters of the network to be trained. The structure regularization coefficient is represented; the joint objective function is used to collaboratively optimize point prediction accuracy, quantile interval coverage, and model complexity.
[0030] In step 5, the following comprehensive score is constructed for each candidate time step:
[0031]
[0032] in, Indicates the candidate time step. This represents the mean absolute error for the corresponding candidate time step. This indicates a coverage penalty. This represents the interval width penalty term; the coverage penalty term and the interval width penalty term are calculated as follows:
[0033]
[0034]
[0035] in, and Representing candidate time steps The predicted interval coverage and average predicted interval width are as follows. and These represent target coverage and target width, respectively. and These represent the coverage penalty coefficient and the width penalty coefficient, respectively. This represents the maximum remaining lifetime in the dataset; when the coverage reaches the target coverage and the interval width does not exceed the target width, the corresponding penalty term is set to zero. The preferred target coverage is 95%, the target width is 10% of the maximum remaining lifetime, and the candidate time step with the smallest overall score is selected as the target time step.
[0036] In step 6, the network parameters obtained from training are kept unchanged during the inference phase, and the same input is executed. The first time Dropout is enabled for stochastic forward propagation, the mean of point predictions and model uncertainty are calculated as follows:
[0037]
[0038]
[0039] in, Indicates the number of random forward propagations. Indicates the first Point prediction values obtained from random forward propagation express The sample mean of the predicted values for each point. Indicating model uncertainty, This represents a positive constant used to ensure numerical stability.
[0040] In the same step, data uncertainty is determined based on the intervals between quantile prediction results; taking the 0.1 quantile and 0.9 quantile as examples, the data uncertainty is calculated as follows:
[0041]
[0042] in, and These represent the input samples respectively. The lower and upper quantile predicted values are used, and a constant of 2.56 is used to convert the width of the quantile interval into the corresponding scale estimate.
[0043] A predetermined number of nearest-neighbor calibration samples are retrieved within the shared temporal feature space or the normalized input feature space, and the uncertainty of the neighborhood of the historical prediction residuals of the nearest-neighbor samples is calculated as follows:
[0044]
[0045] in, Indicates the number of nearest neighbor samples. Indicates the first The true remaining lifetime of the nearest neighbor samples This represents the corresponding historical remaining life prediction value. Indicates input sample Neighborhood uncertainty.
[0046] In step 7, the calibration samples are divided into multiple remaining lifetime segments according to the predicted remaining lifetime, and within each segment, an empirical conformal threshold for achieving the target coverage is determined according to the following relationship:
[0047]
[0048] in, Indicates the first A set of calibration samples within each remaining lifetime segment This indicates the number of samples in the set. Indicates the target false coverage rate. Indicates the first The empirical conformal threshold corresponding to each remaining lifetime segment.
[0049] The smaller value between model uncertainty and data uncertainty is selected as the candidate scale, and relaxation calibration is performed using the scale factor and bias term determined by the calibration set according to the following relationship:
[0050]
[0051] in, This represents the candidate scale determined by model uncertainty and data uncertainty. and These represent the scale factor and bias term determined using the calibration set, respectively. This represents the candidate scale after relaxation calibration.
[0052] The sample consistency weight is determined based on the degree of consistency between model uncertainty and data uncertainty, and the relaxed and calibrated candidate scale is weighted and fused with the neighborhood uncertainty according to the following relationship:
[0053]
[0054] in, Indicates input sample Consistency weight, This represents the weighted fusion result of multi-source uncertainty; the larger the consistency weight, the more the fusion result focuses on the candidate scale after relaxation calibration, and the smaller the consistency weight, the more the fusion result focuses on neighborhood uncertainty.
[0055] Using the weighted fusion result as the original uncertainty scale for the calibration set samples, a monotonically non-decreasing mapping from the original uncertainty scale to the absolute prediction error is learned on the calibration set through the following optimization relation:
[0056]
[0057] in, Indicates the number of calibration samples. Indicates the first The original uncertainty scale of each calibration sample. and These represent the actual remaining lifetime and the projected remaining lifetime, respectively. This represents a monotonically non-decreasing mapping obtained by minimizing the deviation between the scale mapping value and the absolute prediction error.
[0058] The above multi-source uncertainty weighted fusion result is input into the monotone non-decreasing mapping, and the uncertainty scale after monotone regression calibration is obtained through the following mapping:
[0059]
[0060] Further, an adaptive scaling factor is set according to the degradation stage of the predicted remaining lifetime, and the hierarchical scaling is adjusted according to the following relationship:
[0061]
[0062] in, Indicates input sample The stage of degradation to which it belongs. This represents the adaptive scaling factor corresponding to the degradation stage. This represents the uncertainty scale after adjusting the stratified scale; the scaling factor is increased for degradation stages with larger errors and decreased for degradation stages with smaller errors.
[0063] Under the condition of meeting the target coverage, the hierarchical scale adjustment results are globally compressed and a safety factor is compensated according to the following relationship to obtain the final uncertainty scale:
[0064]
[0065] in, This represents the global compression factor, and its value ranges from greater than 0 to no greater than 1. This represents a safety factor of not less than 1. Indicates input sample The ultimate uncertainty scale.
[0066] In step 8, based on the predicted remaining lifetime at the target time step and the final uncertainty scale, the final prediction interval is constructed as follows:
[0067]
[0068] Among them, ŷ( ) represents the input sample Predicted remaining lifetime at the target time step This represents the final prediction interval corresponding to the predicted point value; thus, the bearing remaining life point prediction result and its uncertainty range can be output simultaneously.
[0069] The beneficial effects of this invention are:
[0070] (1) The joint structure of first interval and then point value is adopted. The point prediction value is directly constrained by the quantile interval. At the same time, the point prediction loss promotes the learning of the interval boundary in the reverse direction, which enhances the consistency between point prediction and interval estimation.
[0071] (2) Positive value mapping ensures that the upper quantile boundary is not less than the lower quantile boundary, avoids quantile crossing, and interval consistency projection avoids the point prediction result from going out of the prediction interval.
[0072] (3) Incorporate point prediction error, interval coverage and interval width into a unified time step comprehensive score so that the selection of prediction lead time no longer depends solely on human experience;
[0073] (4) Simultaneously characterize model uncertainty, data uncertainty and neighborhood uncertainty, and use group conformal, monotonic regression, hierarchical adjustment, compression and security compensation for secondary calibration, which can control the interval width while ensuring coverage reliability;
[0074] (5) It can simultaneously output the remaining life value of the bearing and the corresponding risk range, providing a more intuitive and reliable basis for maintenance decisions.
[0075] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description
[0076] Figure 1 This is the overall flowchart of the bearing remaining life range joint prediction method of the present invention.
[0077] Figure 2 This is a network structure diagram of the interval-to-point joint prediction backbone of the present invention.
[0078] Figure 3 This is a flowchart of the time step optimization strategy of the present invention.
[0079] Figure 4 This is a flowchart of the multi-source uncertainty fusion and secondary calibration process of the present invention.
[0080] Figure 5 This is a graph showing the prediction interval coverage and average prediction interval width at different calibration stages of the present invention.
[0081] Figure 6 This is a comparison chart of the predicted remaining life point of the bearing, the actual remaining life, and the final predicted range of the present invention. Detailed Implementation
[0082] like Figure 1 As shown, the present invention provides a method for joint prediction of bearing remaining life interval to point using multi-source uncertainty calibration, including data processing, interval to point joint prediction, time step optimization, multi-source uncertainty calibration, and result output. The specific implementation process is as follows:
[0083] The entire lifespan of the bearing to be predicted, from healthy operation to failure, is monitored using a fixed sampling period. The monitoring data is preferably a vibration acceleration signal, but may also include at least one of temperature, rotational speed, and load signals. Taking vibration signals as an example, time-domain and frequency-domain statistics are extracted from each sampling file, or a filtered vibration sequence is directly used as input features, and the features in each dimension are normalized.
[0084] Normalized continuous time-series data are segmented into sliding time windows according to a preset window length and sliding step size to generate bearing degradation feature sequences of consistent length. The time interval from the end of the window to the bearing failure time is used as the true remaining lifetime label for the corresponding sample. The resulting samples are divided into training, calibration, and test sets. The training set is used to update network parameters, the calibration set is used to determine conformal thresholds, monotonic mappings, and scaling factors, and the test set is used only for inference evaluation.
[0085] like Figure 2 As shown, each time window sample is input into a shared bidirectional long short-term memory network encoder. The forward long short-term memory network encodes the cumulative degradation information along the time-increasing direction, while the reverse long short-term memory network encodes the overall change trend within the time window along the opposite direction. The hidden states of the two are concatenated to form a shared temporal feature. An attention aggregation layer can also be set after the shared temporal feature to give higher weights to time positions with significant degradation changes.
[0086] The shared temporal features are input into the lower quantile prediction head and boundary increment prediction head of the quantile interval prediction branch, respectively. The lower quantile prediction head outputs the lower quantile boundary, and the boundary increment is added to the lower quantile boundary after being mapped by Softplus positive values to obtain an upper quantile boundary that is not less than the lower quantile boundary. Preferably, the 0.1 quantile and 0.9 quantile are used to construct the initial quantile interval, but other quantile combinations can also be set according to the target coverage. When multiple quantiles are set, monotonic constraints on adjacent quantiles are added during training to avoid quantile boundary intersections.
[0087] After obtaining the upper and lower quantile boundaries, the same shared temporal feature is input into the point prediction branch to obtain unconstrained remaining lifetime point prediction values. Then, based on the upper and lower quantile boundaries, interval consistency projection is performed on the unconstrained point prediction values, ensuring that the projected point prediction values always lie within the initial quantile interval. Since this projection is located in the forward computation path of the network, the point prediction error can be reversed through the shared encoder to the interval prediction branch, while the interval boundaries directly constrain the point prediction results, thereby achieving collaborative optimization of point prediction and interval estimation.
[0088] The shared bidirectional long short-term memory network encoder, quantile interval prediction branch, and point prediction branch are jointly trained. The joint objective function includes mean absolute error loss, mean squared error loss, quantile loss, and L2 regularization term. The mean absolute error loss is used to reduce the impact of outliers, the mean squared error loss is used to enhance the penalty for large prediction errors, the quantile loss is used to ensure statistical coverage of the upper and lower quantile boundaries, and the L2 regularization term is used to suppress overfitting and outlier interval expansion. Preferably, the structure regularization coefficient is set to 1×10⁻⁶. -5 The algorithm is trained using mini-batch gradient descent or the Adam optimizer until it is verified that the loss no longer decreases.
[0089] like Figure 3 As shown, multiple candidate time steps are set, and a prediction model with the same structure is trained or evaluated for each candidate time step. The mean absolute error, prediction interval coverage, and average prediction interval width are calculated respectively. A coverage penalty is applied when the prediction interval coverage is lower than 95%, and a width penalty is applied when the average prediction interval width exceeds 10% of the maximum remaining lifetime in the dataset. The coverage penalty coefficient and the width penalty coefficient are determined by cross-validation and grid search, preferably set to 10 and 2.0 respectively, and the candidate time step with the smallest comprehensive score is determined as the target time step.
[0090] After model training is complete and the target time step is determined, the trained network parameters are kept unchanged. During the inference phase, random deactivation is enabled, and multiple random forward propagations are performed on the same sample. The model uncertainty is determined based on the mean and standard deviation of the multiple point prediction results. The number of random forward propagations can be set from 20 to 200. Increasing the number of propagations can improve the stability of statistical results, but it will also increase the computational cost accordingly.
[0091] Simultaneously, the conditional distribution boundary is obtained using a multiquantile prediction head, and data uncertainty is determined based on the quantile interval width. During training, monotonic constraints are used to avoid quantile crossover, and compact constraints are used to suppress the expansion of invalid intervals under sufficient coverage conditions. Furthermore, a preset number of nearest-neighbor calibration samples are retrieved within the shared temporal feature space or normalized input feature space. The neighborhood uncertainty is determined based on the historical prediction residuals of the nearest-neighbor calibration samples, thus compensating for the increased prediction risk in the near-failure stage or in sparse regions of local samples.
[0092] like Figure 4 As shown, the calibration samples are preferentially divided into 5 remaining lifetime segments according to the predicted remaining lifetime, and the conformal threshold for satisfying the target coverage rate is calculated in each segment. Then, a tighter candidate scale is selected from the model uncertainty and data uncertainty, and relaxation calibration is performed using the scale factor and bias term determined by the calibration set. Then, adaptive fusion is performed based on the sample consistency weight and neighborhood uncertainty. The fusion result is successively subjected to Isotonic monotonic regression, degradation stage hierarchical scaling, global compression under coverage constraints, and safety factor compensation to form the final uncertainty scale and corresponding prediction interval.
[0093] Based on the predicted remaining life point value at the target time step and the final uncertainty scale, the system outputs the remaining life point prediction result and the corresponding final prediction interval for each bearing to be predicted. The maintenance system can use the point prediction value as the expected usable time of the bearing and the lower bound of the prediction interval as a conservative maintenance reference; when the lower bound of the prediction interval is lower than the preset maintenance threshold, or when the width of the prediction interval increases rapidly within consecutive time steps, an inspection or review prompt is triggered.
[0094] In one specific experimental implementation, a set of vibration records from the University of Cincinnati IMS bearing life dataset was selected as operational status monitoring data. The sampling files were arranged sequentially according to the sampling time. Vibration time-domain statistics and frequency-domain statistics were extracted from each sampling file and normalized. Then, a bearing degradation feature sequence was constructed using a sliding time window. The obtained samples were divided into training set, calibration set and test set, and the true remaining life label was generated according to the time interval from the end of the window to the bearing failure time.
[0095] The model was trained and calibrated using the aforementioned interval-to-point joint prediction method. Initial quantile intervals were formed using the 0.1 and 0.9 quantiles, and the target coverage was set to 95%. The calibration samples were divided into 5 segments according to the predicted remaining lifetime. Model uncertainty was determined through multiple random inactivation forward propagations during the inference phase. Subsequently, monotonic regression calibration, nearest neighbor residual compensation, group conformal calibration, adaptive compression, and quantile uncertainty fusion were performed sequentially to obtain the final uncertainty scale.
[0096] Test results at different calibration stages, such as Figure 5 As shown, when random inactivation or monotonic regression is used alone, the prediction interval is narrow but the coverage is insufficient. After introducing nearest neighbor residual compensation and grouped conformal calibration, the coverage is improved and the interval width is increased. After further performing adaptive compression and quantile uncertainty fusion, the final mean absolute error is 12.90 cycles, the prediction interval coverage is about 99.6%, and the average prediction interval width is 120.5 cycles, thus reducing interval redundancy while maintaining reliable coverage.
[0097] Bearing life prediction results are as follows Figure 6 As shown, the predicted remaining lifetime value is consistent with the overall trend of the actual remaining lifetime, and the final prediction interval can encompass most of the actual remaining lifetime value. In the low remaining lifetime stage near failure, the prediction interval widens accordingly as the local error increases, and gradually narrows in the stage where the degradation trend is relatively stable. This indicates that the hierarchical calibration and multi-source uncertainty fusion can adjust the predicted risk range according to the bearing degradation stage.
[0098] The basic principles, main features, and advantages of the present invention have been described above. Those skilled in the art should understand that the above embodiments are only used to illustrate the technical principles of the present invention and do not constitute a limitation on the scope of protection; equivalent substitutions and improvements made to encoder type, bit combination, candidate time step, number of nearest neighbors, number of groups, or calibration parameters without departing from the principles of the present invention should all be included within the scope of protection of the present invention.
Claims
1. A method for joint prediction of bearing remaining life intervals to points using multi-source uncertainty calibration, characterized in that: The operating status monitoring data of the bearing to be predicted is acquired, the operating status monitoring data is normalized and segmented by a sliding time window to obtain the bearing degradation feature sequence, and the bearing degradation feature sequence is input into a shared bidirectional long short-term memory network encoder to obtain shared temporal features. The shared time series features are input into the quantile interval prediction branch and the point prediction branch respectively to obtain the upper and lower quantile boundaries of the remaining lifetime and the unconstrained remaining lifetime point prediction value. Then, based on the upper and lower quantile boundaries, the unconstrained remaining lifetime point prediction value is subjected to interval consistency projection to obtain the remaining lifetime point prediction value located between the upper and lower quantile boundaries. For multiple candidate time steps, a comprehensive score is constructed based on point prediction error, prediction interval coverage, and average prediction interval width, and the candidate time step with the smallest comprehensive score is determined as the target time step. At the target time step, model uncertainty, data uncertainty, and neighborhood uncertainty are determined respectively. The model uncertainty, data uncertainty, and neighborhood uncertainty are fused and calibrated by group conformal calibration to obtain the final uncertainty scale. Based on the remaining lifetime point prediction value and the final uncertainty scale, the bearing remaining lifetime point prediction result and the corresponding final prediction interval are output.
2. The bearing remaining life range joint prediction method as described in claim 1, characterized in that, The joint prediction of intervals to points includes: The shared bidirectional long short-term memory network encoder receives the bearing degradation feature matrix determined by the time window length and the dimension of the input features, and concatenates the forward hidden state and the reverse hidden state to form the shared temporal feature; The quantile interval prediction branch includes a lower quantile prediction head and a boundary increment prediction head. The boundary increment output by the boundary increment prediction head is added to the lower quantile boundary after being mapped by Softplus positive value to obtain an upper quantile boundary that is not less than the lower quantile boundary. The point prediction branch outputs the unconstrained remaining lifetime point prediction value and truncates it between the lower quantile boundary and the upper quantile boundary. The shared bidirectional long short-term memory network encoder, quantile interval prediction branch, and point prediction branch are trained using a joint objective function, which includes mean absolute error loss, mean squared error loss, quantile loss, and structure regularization term.
3. The bearing remaining life range prediction method as described in claim 1, characterized in that, The determination of the target time step includes: For each candidate time step, a prediction model with the same structure is trained or evaluated, and the corresponding mean absolute error, prediction interval coverage, and mean prediction interval width are calculated. When the prediction interval coverage is lower than the target coverage, a coverage penalty is determined based on the difference between the two. When the mean prediction interval width exceeds the target width, a width penalty is determined based on the difference between the two. The weighted sum of the mean absolute error, coverage penalty, and width penalty is used as the comprehensive score.
4. The bearing remaining life range joint prediction method as described in claim 1, characterized in that, The multi-source uncertainty calibration includes: During the inference phase, multiple forward propagations with random deactivation are performed on the same input, and the model uncertainty is determined based on the dispersion of the multiple point prediction results; the data uncertainty is determined based on the spacing between multiple quantile prediction values; nearest neighbor calibration samples are retrieved in the shared temporal feature space or normalized input feature space, and the neighborhood uncertainty is determined based on the historical remaining lifetime prediction residuals of the nearest neighbor calibration samples. The calibration samples are divided into multiple remaining lifetime segments according to the predicted remaining lifetime, and a conformal threshold that satisfies the target coverage is determined in each remaining lifetime segment. The smaller value between the model uncertainty and the data uncertainty is selected as the candidate scale. The candidate scale is relaxed and calibrated using the scale factor and bias term determined by the calibration set. Based on the consistency between the model uncertainty and the data uncertainty, the relaxed and calibrated candidate scale is weighted and fused with the neighborhood uncertainty. The weighted fusion results are sequentially subjected to monotonic regression calibration, degradation stage hierarchical scale adjustment, scale compression under target coverage constraints, and safety factor compensation to obtain the final uncertainty scale.
