A bearing fatigue life prediction method based on physical information embedding

Abstract

This invention provides a bearing fatigue life prediction method based on physical information embedding, belonging to the technical field of rotating machinery condition monitoring and remaining life prediction. It includes: collecting original monitoring segments and arranging them sequentially to construct a continuous full-life segment sequence of the bearing, and constructing labels; encoding the vibration condition state, calculating the basic rated life and cumulative damage for each segment, and combining the cumulative damage with the vibration condition state to form a physical degradation state sequence; constructing and training a life prediction network; collecting the physical degradation state sequence and physical degradation state of the bearing to be predicted as a validation set input to the trained physical information embedding-based life prediction network, obtaining point prediction, lower quantile life, upper quantile life, and damage increment for the next segment, and performing corrections and adding abnormal state markers, determining the reliability level, and obtaining life records. This invention can accurately distinguish between operating condition fluctuations and actual fatigue degradation, improving the accuracy and stability of the prediction results.

Description

Technical Field

[0001] This invention relates to a bearing fatigue life prediction method based on physical information embedding, belonging to the technical field of rotating machinery condition monitoring and remaining life prediction. Background Technology

[0002] Rolling bearings endure cyclic contact loads during continuous rotation, and fatigue damage is influenced by rotational speed, radial load, temperature rise, and individual manufacturing variations. After damage occurs, the root mean square of the bearing housing vibration, the impact intensity, and the energy of the fault characteristic frequency band change over time; load switching also causes similar instantaneous changes. If life prediction relies solely on a single vibration amplitude, it's easy to interpret short-term operating conditions as continuous damage, and it can also delay maintenance and assessment during the accelerated degradation phase. Therefore, life prediction not only needs to identify vibration changes but also needs to incorporate the relationship between operating conditions and fatigue accumulation into the same calculation process.

[0003] Current technical specifications and other standards propose a route for calculating rated life based on the basic rated dynamic load, equivalent dynamic load, and material life index of bearings to estimate the group rated life. This route has a clear physical meaning, is suitable for bearing selection and establishing maintenance benchmarks, and can reflect the basic principle that increased load leads to a shortened rated life. Its limitation is that the calculation parameters mainly correspond to standard operating conditions and group statistics, making it difficult to incorporate impact changes, temperature rise changes, and early local degradation information during the operation of individual bearings. Therefore, it cannot directly provide the individual remaining life updated with monitoring segments.

[0004] Existing vibration characteristic prediction methods calculate the root mean square, kurtosis, peak factor, and fault characteristic frequency band energy from time-domain waveforms and envelope spectra, and then establish a regression relationship between health indicators and remaining life. This approach can describe the condition of individual bearings using online monitoring data, and its computational scale is convenient for engineering deployment. However, the same vibration characteristic is simultaneously affected by rotational speed, load, and damage degree. When the training data is inconsistent with the operating condition distribution of the bearing under test, the model struggles to distinguish between changes in operating conditions and permanent degradation, and the predicted curve may show a life recovery without physical basis after the load decreases.

[0005] Deep time-series prediction methods utilize causal convolution, gated recurrent structures, and attention structures to process continuous vibration sequences. This approach can extract nonlinear degradation trends from multiple adjacent segments and reduce the direct impact of individual anomalous segments on life results. This route typically uses life label error as the primary training objective, and the network readily learns and trains condition-related patterns in bearings. When the load undergoes phased changes, although the time-series characteristics change significantly, there is a lack of traceable relationship between them and the fatigue damage increment, making the model output difficult to interpret.

[0006] Physical information-based prediction methods incorporate rated life relationships, damage evolution residuals, and degradation trend constraints into the data model to reduce deviations from fatigue accumulation patterns in life predictions. Existing methods often only incorporate physical relationships into the training loss, with inference inputs primarily consisting of vibration characteristics. The network lacks a clear understanding of the physical states comprised of segment-by-segment loads, rotational speeds, and cumulative damage. Some methods directly prune abnormal life values ​​at the output, but fail to explain how vibration representation, current operating conditions, and physical damage collectively form point and interval predictions.

[0007] The above approaches emphasize group rated life, vibration statistical characterization, deep temporal relationships, and physical constraints, respectively. Their common shortcoming is that they do not establish a clear data relationship from continuous monitoring segments, segment-by-segment rated life benchmarks, damage increments, fixed field physical states to life output. Summary of the Invention

[0008] The purpose of this invention is to provide a bearing fatigue life prediction method based on physical information embedding, which overcomes the shortcomings of existing bearing life prediction methods such as separation of physical damage and vibration characteristics, difficulty in distinguishing between operating condition changes and permanent degradation, and lack of physical interpretability of life prediction results. This invention enables the prediction model to accurately distinguish between operating condition fluctuations and actual fatigue degradation, thereby improving the accuracy, stability and physical interpretability of the prediction results.

[0009] To achieve the above objectives, the present invention employs the following technical solution: A bearing fatigue life prediction method based on physical information embedding includes: The original monitoring segments were composed of bidirectional vibration components, average rotational speed, average radial load, average bearing temperature, and average ambient temperature collected at the same time stamp. These segments were then arranged sequentially to construct a continuous full-life segment sequence for the bearing. The health stage parameters, degradation start markers, life end markers, and remaining life labels of the continuous full-life segment sequence for the bearing were determined. Vibration condition status is obtained based on the continuous full-life segment sequence of bearings and the parameter encoding of health stages. The basic rated life and damage increment are calculated segment by segment. The cumulative damage is calculated based on the damage increment. The cumulative damage is combined with the vibration condition to obtain the physical degradation state sequence. A lifetime prediction network based on physical information embedding is constructed and trained, including two layers of causal convolution, gated recurrent units, and an output head. The physical degradation state sequence is extracted into a short-term degradation feature sequence through causal convolution. The short-term degradation feature sequence is input into the gated recurrent unit to obtain the historical degradation representation. The historical degradation representation is concatenated with the physical degradation state and processed through a fully connected layer to obtain a shared latent state. The shared latent state is input into the output head to obtain point prediction, lower quantile lifetime, upper quantile lifetime, and next segment damage increment. The physical degradation state and physical degradation state sequence of the bearing to be predicted are collected as the validation set and input into the trained lifetime prediction network to obtain point prediction, lower quantile lifetime, upper quantile lifetime and damage increment of the next segment, and then corrected and abnormal state labels are added. The relative width of the 90% prediction interval is calculated based on the corrected point prediction, lower quantile life, upper quantile life, and abnormal state markers, and the reliability level is determined. The corrected point prediction, lower quantile life, and upper quantile life reliability levels are used as the reliable life results, and the bearing number and cumulative damage are used as the life records.

[0010] Preferably, the specific method for collecting original monitoring segments is as follows: Piezoelectric accelerometers were fixed at the radial horizontal and radial vertical positions in the bearing area of ​​the test bearing housing. The sensitive axes of the piezoelectric accelerometers were aligned with the horizontal and vertical directions, respectively, to collect the horizontal and vertical vibration components. The average rotational speed, average radial load, average bearing temperature, and average ambient temperature were read. The horizontal vibration components, vertical vibration components, average rotational speed, average radial load, average bearing temperature, and average ambient temperature with the same timestamp were combined into one original monitoring segment. The original monitoring segments with complete and continuous time fields are filtered, and a continuous full-life segment sequence of bearings is constructed based on the timestamp order.

[0011] Preferably, the health stage parameters, degradation start markers, life end markers, and remaining life labels of a continuous full-life segment sequence of bearings are determined, specifically through the following methods: The T consecutive complete monitoring segments that appear after the bearing reaches a stable speed and a stable radial load are defined as the healthy reference interval. Within the healthy reference interval, the coefficients of variation of the speed and radial load do not exceed 2%, and the root mean square of the bidirectional vibration does not show an increase of t consecutive monitoring segments. Calculate the health stage parameters, including: median energy of the characteristic frequency band of the health level fault, median energy of the characteristic frequency band of the health vertical fault, absolute deviation of the median energy of the characteristic frequency band of the health level fault, and absolute deviation of the median energy of the characteristic frequency band of the health vertical fault within the health reference interval; and record the median healthy rotational speed, median healthy radial load, root mean square median of healthy horizontal vibration, and root mean square median of healthy vertical vibration. When the horizontal fault characteristic band energy of the subsequent five consecutive monitoring segments is higher than the sum of the median absolute deviation of the healthy horizontal fault characteristic band energy and the median absolute deviation of the healthy horizontal fault characteristic band energy, or when the vertical fault characteristic band energy is higher than the sum of the median absolute deviation of the healthy vertical fault characteristic band energy and the median absolute deviation of the healthy vertical fault characteristic band energy, the first monitoring segment is marked as the degradation start point. The terminating segment of a continuous full-lifetime segment sequence is used as the lifetime end marker; The difference between the sequence number of the terminated segment and the sequence number of the current monitoring segment is multiplied by the original monitoring segment acquisition time interval. As a label for remaining lifespan.

[0012] Preferably, the vibration condition status includes a vibration field and a condition field; The vibration field includes the root mean square of the horizontal vibration component, the kurtosis of the horizontal vibration component, the peak factor of the horizontal vibration component, the fault characteristic frequency band energy of the horizontal vibration component, the root mean square of the vertical vibration component, the kurtosis of the vertical vibration component, the peak factor of the vertical vibration component, and the fault characteristic frequency band energy of the vertical vibration component. The operating condition fields include the speed ratio, radial load ratio, and temperature rise ratio calculated based on the average speed, average radial load, average bearing temperature, and average ambient temperature.

[0013] Preferably, the basic rated life is calculated as follows: , In the formula, For the first The basic rated lifespan corresponding to each monitoring segment. The basic dynamic load rating is given in the bearing model manual. For the first The equivalent dynamic load of the monitoring segment, the first The average radial load of each monitoring segment is taken as the equivalent dynamic load of that monitoring segment. Life expectancy index; The damage increment is obtained by dividing the bearing rotation count within the original monitoring segment acquisition time interval by the basic rated life corresponding to the monitoring segment. The cumulative damage is obtained by accumulating the damage increment based on the monitoring segment.

[0014] Preferably, the cumulative damage and vibration condition are combined to obtain a physical degradation state sequence, as follows: After the vibration condition status is normalized by percentile, the cumulative damage is added as the last field to the vibration condition status to obtain the physical degradation status. The normalized boundaries of each field of the vibration condition status are used to form a boundary matrix. The physical degradation state is normalized using a boundary matrix. The normalized physical degradation state is then arranged in chronological order to form a physical degradation state sequence, where rows correspond to monitoring segments and columns correspond to fields of the physical degradation state.

[0015] Preferably, the first causal convolutional layer adds two all-zero time points at the beginning of the physical degradation state sequence, and only reads the current segment and the two previous segments; The second causal convolutional layer adds four all-zero time points to the front of the physical degradation state sequence and reads the historical position at an expansion interval of 2. After each convolutional layer, layer normalization and linear rectified activation are performed sequentially. During the training phase, random deactivation with a ratio of 0.1 is also performed. The gated loop unit inputs a short-term degradation feature sequence to obtain a time-series state. The time-series state at the last moment is used as a historical degradation representation and is sequentially concatenated with the physical degradation state to form a fusion vector. The output head includes a lifetime head, a damage head, and a lifetime interval head. The lifetime head processes the fused features sequentially through two linear layers to obtain the predicted number of remaining segments. The damage head processes the fused features sequentially through a linear layer and a Softplus layer to obtain the predicted damage increment for the next segment. The lifetime range header processes the fused features sequentially through two linear layers to obtain the number of remaining fragments at the 5% quantile and the number of remaining fragments at the 95% quantile. The hidden layers of the lifespan head and the damage head are activated by linear rectification; The predicted number of remaining segments, the number of remaining segments at the 5th percentile, and the number of remaining segments at the 95th percentile are multiplied by the original monitoring segment acquisition time interval, respectively. We obtain point predictions, lower quantile lifetimes, and upper quantile lifetimes.

[0016] Preferably, the joint training objective of the lifetime prediction network based on physical information embedding is as follows: , in, For remaining life loss, For damage evolution loss, For monotonic loss of lifespan, Interval quantile loss, As a weight for remaining lifetime loss, Weights for damage evolution loss, Weights for monotonic loss of lifetime, The weights for interval quantile loss; The acquisition method is as follows: For training samples whose physical lifetime references terminate normally, calculate the Huber loss between point prediction and remaining lifetime label, and the Huber loss between point prediction and physical lifetime reference, and take the average of the two; for training samples whose physical lifetime references terminate abnormally, only calculate the Huber loss between point prediction and remaining lifetime label. The physical lifetime is calculated recursively based on the damage increment of the next segment and the physical degradation state, and the specific calculation method is as follows: , , In the formula, For the first Physical lifetime reference for each monitoring segment, For indicator functions, when The value is 1 when it is less than 1, and 0 when it reaches 1. For the recursive calculation starting from the current state, the first... Accumulated damage; The maximum allowed segment number for recursion is calculated by dividing the device's specified lifespan by [the specified value]. Sure, For the predicted first Incremental damage to each segment; The recursion terminates normally when the cumulative damage first reaches 1. If the termination condition is not met when the maximum allowable number of lifetime segments of the equipment is reached, the physical lifetime reference is taken as the upper limit and the recursion terminates abnormally. If the recursion outputs a non-numerical value, the physical reference calculation is immediately stopped, and only the lifetime end mark is retained to participate in the lifetime error term. It is the mean square error between the damage head output and the physical damage increment; A positive penalty is applied when the predicted lifetime in the later window is higher than that in the previous window. This includes 5% quantile loss and 95% quantile loss, with constraints that the 5% quantile lifetime should not exceed the reference lifetime and the 95% quantile lifetime should not be lower than the reference lifetime.

[0017] Preferably, the point prediction, lower quantile lifetime, upper quantile lifetime, and next segment damage increment are corrected and abnormal state markers are added, specifically as follows: Negative point predictions are limited to 0. When the current point prediction is higher than the point prediction of the previous valid time step, the point prediction of the previous valid time step is taken. When the single drop of the point prediction exceeds the 99th percentile of the drop in the validation set, the maximum drop corresponding to that percentile is taken. Check if the lower quantile lifetime is not higher than the point prediction and the point prediction is not higher than the upper quantile lifetime. If not, rearrange the points prediction, lower quantile lifetime, and upper quantile lifetime in ascending order, and use them as the lower quantile lifetime correction value, point prediction correction, and upper quantile lifetime in that order.

[0018] Preferably, the relative width of the 90% prediction interval is calculated and the reliability level is determined based on the corrected point prediction, lower quantile lifetime, upper quantile lifetime, and anomalous state markers. The specific method is as follows: Based on the corrected post-point prediction, lower quantile lifetime, upper quantile lifetime, and anomalous state markers, the relative width of the 90% prediction interval and the reliability level are calculated; including: Coverage rate represents the proportion of validation samples whose remaining lifetime labels fall within the 90% prediction range out of the total number of validation samples. The underreporting rate of maintenance indicates the proportion of samples whose remaining life label is lower than the equipment's specified maintenance lead time, but whose point prediction is still higher than the maintenance lead time, out of the total number of samples that should be maintained. The manual review ratio represents the proportion of verification samples classified as low reliability out of the total number of verification samples. Among the candidate reliability boundaries with a coverage rate of not less than 85%, the boundary with the lowest rate of missed maintenance is selected first; if the rates of missed maintenance are the same, the boundary with the lowest rate of manual review is selected. When the relative width is no more than 30% and the abnormal state marker is invalid, the reliability level is marked as high; when it exceeds 30% but does not exceed 50% and the abnormal state marker is invalid, the reliability level is marked as medium; when the relative width exceeds 50% and the abnormal state marker is valid, the reliability level is marked as low.

[0019] The advantages of this invention are as follows: This invention constructs a physical degradation state while retaining bidirectional vibration changes, current operating conditions, and accumulated damage locations. The network can distinguish between short-term responses caused by load switching and continuous fatigue degradation, thereby reducing cross-condition life prediction bias. By calculating the segment-by-segment damage increment from the rated life benchmark and segment rotational speed, and incorporating damage evolution loss into joint training, the prediction network obtains a degradation reference consistent with radial load and rotational speed, which can suppress the lack of basis for positive rebound in the life curve. Through point prediction, the next segment damage increment, and the 90% prediction interval generated from the same 48-dimensional shared latent state, and by implementing consistency constraints at the output, life values, damage trends, and interval widths can change synchronously based on the same degradation state, thereby improving the stability and interpretability of reliability grading. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0021] Figure 1 This is a flowchart of the overall process for predicting bearing fatigue life.

[0022] Figure 2 Build a flowchart for the physical degradation state.

[0023] Figure 3 Flowchart for embedding physical information into a lifetime prediction network.

[0024] Figure 4 Output flowchart for remaining lifetime and reliability.

[0025] Figure 5 This is a schematic diagram of a 12-dimensional physical degradation state sequence.

[0026] Figure 6 A schematic diagram illustrating damage recursion driven by contact load.

[0027] Figure 7 This is a schematic diagram showing the remaining lifespan and the predicted range.

[0028] Figure 8 This diagram illustrates the comparison of lifetime prediction errors using different methods.

[0029] Figure 9 This is a schematic diagram of the remaining lifetime full-lifetime tracking results.

[0030] Figure 10 This is a schematic diagram of the physical constraint ablation results under different load conditions. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1 A bearing fatigue life prediction method based on physical information embedding is proposed. The method takes horizontal vibration component, vertical vibration component, rotational speed, radial load and temperature as raw inputs and sequentially completes the construction of continuous full life segments, the construction of physical degradation state, physical information embedding prediction and life result output.

[0033] like Figure 1 As shown, it specifically includes: S1: Collect bidirectional vibration components, average rotational speed, average radial load, average bearing temperature, and average ambient temperature at the same time stamp to form the original monitoring segment, and arrange them sequentially to construct a continuous full-life segment sequence of the bearing; determine the health stage parameters, degradation start marker, life end marker, and remaining life label of the continuous full-life segment sequence of the bearing; S2: Vibration condition status is obtained based on the continuous full-life segment sequence of the bearing and the parameter encoding of the health stage; S3: Calculate the basic rated life and damage increment segment by segment, calculate the cumulative damage based on the damage increment, and merge the cumulative damage with the vibration condition state into a physical degradation state sequence. S4: Construct and train a lifetime prediction network based on physical information embedding, including two layers of causal convolution, gated recurrent units, and an output head; extract the physical degradation state sequence into a short-term degradation feature sequence through causal convolution, input the short-term degradation feature sequence into the gated recurrent unit to obtain the historical degradation representation, concatenate the historical degradation representation with the physical degradation state and process it through a fully connected layer to obtain a shared latent state, and input the shared latent state into the output head to obtain point prediction, lower quantile lifetime, upper quantile lifetime, and next segment damage increment; S5: Collect the physical degradation state and physical degradation state sequence of the bearing to be predicted as the validation set input to the trained lifetime prediction network to obtain point prediction, lower quantile lifetime, upper quantile lifetime and next segment damage increment, and make corrections and add abnormal state labels. S6: Calculate the relative width of the 90% prediction interval based on the corrected point prediction, lower quantile life, upper quantile life, and abnormal state markers, and determine the reliability level. Use the corrected point prediction, lower quantile life, and upper quantile life reliability level as the reliable life result, and use the bearing number and cumulative damage as the life record.

[0034] As a refinement of the above embodiments, step S1 includes: S1-1 Multi-source monitoring data synchronous acquisition; One piezoelectric accelerometer is fixed at a radially horizontal position and a radially vertical position near the load-bearing area of ​​the test bearing housing, respectively. The sensitive axes of the two sensors are aligned with the horizontal and vertical directions, respectively. Monitoring data is collected synchronously every 10 seconds. Both vibration channels sample continuously for 0.1 seconds at a sampling frequency of 25.6 kHz, obtaining horizontal and vertical vibration components containing 2560 sampling points, respectively. Simultaneously, the average rotational speed, average radial load, average bearing temperature, and average ambient temperature within the corresponding 10-second interval are read. The horizontal vibration component, vertical vibration component, average rotational speed, average radial load, average bearing temperature, and average ambient temperature with the same timestamp are combined into one original monitoring segment. The 10 seconds represents the acquisition interval between adjacent original monitoring segments, and 0.1 seconds represents the sampling duration of a single vibration signal. S1-1 outputs the set of original monitoring segments arranged by timestamp, which is used as the input to S1-2.

[0035] Construction of continuous full-lifetime sequences for S1-2; The system receives the set of raw monitoring segments output from S1-1 and checks the horizontal vibration component, vertical vibration component, average rotational speed, average radial load, average bearing temperature, and average ambient temperature of each raw monitoring segment according to its timestamp. Raw monitoring segments with all six data points intact are retained, and their continuity is checked at fixed time intervals of 10 seconds. The raw monitoring segments with consecutive timestamps are arranged in the order of acquisition to form a continuous monitoring sequence for the corresponding bearing. The first continuous monitoring segment after the bearing has been installed and reached a stable operating state is taken as the start point of the sequence, and the last segment of the full-life test record is taken as the end point, forming a continuous full-life segment sequence for the bearing.

[0036] S1-3 Degradation markers, lifetime labels determination, and dataset isolation; Receive the continuous full-life segment sequence of the bearing output from S1-2. The earliest consecutive T (30 in this embodiment) complete monitoring segments after the bearing reaches a stable speed and stable radial load are determined as the healthy reference interval. Within the healthy reference interval, the coefficients of variation of speed and radial load do not exceed 2%, and the root mean square of bidirectional vibration does not show an increase of t (5 in this embodiment) monitoring segments.

[0037] Calculate the health stage parameters, including: median energy of the horizontal fault characteristic band, median energy of the vertical fault characteristic band, median absolute deviation of the horizontal fault characteristic band energy, and median absolute deviation of the vertical fault characteristic band energy within the health reference interval. Record the median healthy rotational speed, median healthy radial load, root mean square median of healthy horizontal vibration, and root mean square median of healthy vertical vibration. When the horizontal fault characteristic band energy of the subsequent five consecutive monitoring segments is higher than the sum of the median energy of the horizontal fault characteristic band energy and the median absolute deviation of the horizontal fault characteristic band energy, or when the vertical fault characteristic band energy is higher than the sum of the median energy of the vertical fault characteristic band energy and the median absolute deviation of the vertical fault characteristic band energy, the first monitoring segment is marked as the degradation starting point.

[0038] The median of the energy of the characteristic frequency band of the horizontal fault and the median of the energy of the characteristic frequency band of the vertical fault are the median of the energy of the horizontal fault and the energy of the vertical fault in all monitored segments within the healthy reference interval, respectively. The absolute deviation of the two medians is the median of the absolute value of the difference between the energy of each segment and the corresponding median.

[0039] The final segment of the bearing's continuous lifespan segment sequence is used as the lifespan end marker. The remaining lifespan label of the k-th monitoring segment is determined by multiplying the difference between the lifespan end marker number and k by 10s. The bearing's continuous lifespan segment sequence with completed degradation marker and lifespan label assignments is divided into training, validation, and test sets according to the bearing number. All monitoring segments of the same bearing are included in only one dataset. S1-3 outputs the bearing's continuous lifespan segment sequence isolated by bearing number, the median of each healthy stage, the degradation start marker, and the remaining lifespan label. Specifically, the bearing's continuous lifespan segment sequence and the median of each healthy stage are passed to S2, the degradation start marker is passed to S4-3, and the remaining lifespan label is passed to S5-1.

[0040] As a refinement of the above embodiments, such as Figure 2 As shown, step S2 includes: The system receives the continuous full-lifetime segment sequence and health phase parameters output by S1. For each monitoring segment, it calculates the root mean square (RMS) of the horizontal vibration component, the kurtosis of the horizontal vibration component, the peak factor of the horizontal vibration component, and the fault characteristic frequency band energy of the horizontal vibration component, as well as the RMS, kurtosis, peak factor, and fault characteristic frequency band energy of the vertical vibration component, forming eight vibration fields. Then, based on the average rotational speed, average radial load, and average bearing temperature at the same timestamp, it calculates the speed ratio, radial load ratio, and temperature rise ratio, forming three operating condition fields. These eight vibration fields and three operating condition fields are combined in a fixed order to form an 11-dimensional vibration operating condition state.

[0041] The root mean square (RMS) values ​​of the horizontal and vertical vibration components are the root mean square values ​​of the horizontal and vertical vibration components at 2560 sampling points within the monitoring segment, respectively. These values ​​are then normalized by dividing by the median of the healthy horizontal vibration and the median of the healthy vertical vibration, respectively.

[0042] The kurtosis of the horizontal vibration component and the kurtosis of the vertical vibration component are the ratios of the fourth central moment to the square of the variance of the horizontal vibration component and the vertical vibration component, respectively.

[0043] The peak factor of the horizontal vibration component and the peak factor of the vertical vibration component are the ratios of the absolute peak value to the root mean square of the horizontal and vertical vibration components, respectively.

[0044] The fault characteristic frequency band energy of the horizontal vibration component and the fault characteristic frequency band energy of the vertical vibration component are the sum of the energy of the envelope spectrum of the horizontal vibration component and the vertical vibration component within two frequency resolution ranges near the theoretical fault characteristic frequency of the bearing inner ring, outer ring, and rolling element, respectively. They are then normalized by dividing by the median of the healthy horizontal fault characteristic frequency band energy and the median of the healthy vertical fault characteristic frequency band energy, respectively. The theoretical fault characteristic frequency includes the inner ring fault characteristic frequency, the outer ring fault characteristic frequency, and the rolling element fault characteristic frequency.

[0045] The characteristic frequency of inner ring failure is 0.5 times the number of rolling elements multiplied by the shaft rotation frequency, multiplied by 1, plus the ratio of the rolling element diameter to the pitch circle diameter multiplied by the cosine of the contact angle.

[0046] The characteristic frequency of outer ring failure is 0.5 times the number of rolling elements multiplied by the shaft rotation frequency, then multiplied by 1 minus the ratio of the rolling element diameter to the pitch circle diameter multiplied by the difference of the contact angle cosine.

[0047] The characteristic frequency of rolling element failure is the difference between the ratio of the pitch circle diameter to twice the rolling element diameter, multiplied by the shaft rotation frequency, and then multiplied by the square of the ratio of the rolling element diameter to the pitch circle diameter minus the square of the contact angle cosine.

[0048] The shaft rotation frequency is calculated by dividing the current rotation speed by 60. The above-mentioned root mean square and fault characteristic band energy are dimensionless after normalization, and the kurtosis and peak factor are themselves dimensionless.

[0049] The speed ratio is equal to the average speed divided by the median healthy speed; the radial load ratio is equal to the average radial load divided by the median healthy radial load; and the temperature rise ratio is equal to the average bearing temperature minus the average ambient temperature, then divided by the upper limit of the equipment's allowable temperature rise.

[0050] The maximum allowable temperature rise of the equipment is a predefined value obtained from the bearing model manual.

[0051] As a refinement of the above embodiments, such as Figure 2 As shown, step S3 includes S3-1 Segment-by-Segment Rated Life Baseline Calculation; The average radial load output from S1 is received. In this embodiment, only the radial load is applied, thus the average radial load of the k-th monitoring segment is used as the equivalent dynamic load of that monitoring segment. Subsequently, the basic rated life of the monitoring segment is determined based on the basic rated dynamic load corresponding to the bearing model and the rolling element type. The rated dynamic load and rolling element type remain unchanged in the continuous bearing life segment sequence of the same bearing, while the equivalent dynamic load is updated with each monitoring segment, so that each monitoring segment only uses the currently measured operating conditions to form the corresponding life reference. The basic rated life calculation formula is as follows: , In the formula, For the first The basic rated lifespan corresponding to each monitoring segment, in revolutions; The basic rated dynamic load given in the bearing model manual, in N; For the first The equivalent dynamic load of each monitoring segment, in N. The life index is 3, which is used in this embodiment with ball bearings.

[0052] The basic rated lifetime corresponding to the k-th monitoring segment is output and used as input for the S3-2 damage increment calculation. This output is not used as an additional network field, thus avoiding the duplicate inclusion of radial load information in the physical degradation state.

[0053] S3-2 Segment-by-segment damage accumulation; The average rotational speed of the kth monitoring segment in the continuous full-life segment sequence of the bearing, which is based on the basic rated life and output by S1, is received. The number of revolutions and the dimensionless damage increment within the segment interval are calculated, and then accumulated in chronological order to obtain the cumulative damage. , In the formula, This represents the dimensionless damage increment of the k-th monitoring segment; The bearing revolutions within the kth 10s interval are obtained by multiplying the average rotational speed by 10s and converting it to 60s. This is the basic rated life; The cumulative damage at the end of the k-th monitoring segment. This represents the cumulative damage at the end of the previous monitoring segment. The calculation maintains the original time sequence of the segments and does not reorder the damage increments; after the operating condition is restored, it continues to extrapolate from the cumulative damage of the previous segment to preserve historical damage caused by variable load operation.

[0054] 10s is a fixed time interval between adjacent monitoring segments, so it is used both to convert the rotation speed into the number of revolutions per segment and to convert the remaining number of segments output by the network into the remaining time. Always a non-negative quantity. The calculation starts from 0 at the end of the health reference interval and is recursively applied, without regressing due to short-term load decreases.

[0055] The training data is scaled based on the cumulative damage at the bearing's life end marker. This scaling factor is then used to synchronously process the cumulative damage and damage increment of each monitoring segment before the bearing's life end marker, ensuring that the scaled cumulative damage is 1 at the life end marker while maintaining the recursive relationship between cumulative damage and damage increment in adjacent segments. The validation set, test set, and online inference do not read the future life end of the bearing under test; instead, they use the median of the scaling factor for the same bearing model from the training set. This processing ensures that the cumulative damage retains both load and speed variations without prematurely incorporating the predicted future life into the input.

[0056] Output the damage increment and cumulative damage for k monitoring segments. The damage increment is passed to S3-3 and used as the 12th field of the physical degradation state. When constructing the training sample with the k-th monitoring segment as the current time step, the physical damage increment corresponding to the (k+1)-th monitoring segment is read. This is then passed to S5-1. The verification, testing, and inference phases do not read this data. .

[0057] S3-3 Physical State Normalization and Window Organization; The system receives 11-dimensional vibration condition status and cumulative damage, and uses the cumulative damage as the 12th field to obtain the physical degradation state. The first 11 fields are mapped to 0 to 1 using the 1st and 99th percentiles of the training set, with values ​​outside the boundaries being 0 and 1 respectively.

[0058] The two normalized boundaries of each field of the vibration condition state are arranged in a fixed field order to form a 2×12 boundary matrix, which is used to normalize the physical degradation state and map it to 0 to 1. The 2×12 boundary matrix is ​​determined on the training set and saved with the model parameters. During the inference stage, the boundary matrix is ​​loaded to normalize each field of the current monitoring segment. It is not recalculated online to ensure that the normalization standard is consistent between training and inference.

[0059] The normalized physical degradation states are arranged in chronological order as a 32×12 physical degradation state sequence, where the 32 rows correspond to the 32 complete monitoring segments, the 12 columns correspond to the 12 fields of the 12-dimensional physical degradation state, and the last row is the current physical degradation state.

[0060] During the training of the lifetime prediction network, the window length is selected from 16, 24, 32, and 48 segments. The candidate lifetime prediction networks maintain consistent data partitioning, number of channels, and training budget, and select the shortest window with the lowest MAPE on the validation set and a single inference time of no more than 1 second; in this embodiment, 32 segments are used, corresponding to a monitoring range of 320 seconds.

[0061] The 32×12 physical degradation state sequence is passed to S4-1, and the current 12-dimensional physical degradation state is passed to S4-2 and S5-1, with the field order, normalization boundaries, and time direction remaining unchanged. S3-2 passes the physical damage increment to S5-1. S3-3 marks the window as healthy or degraded based on the degradation initiation marker; this marker is only used for sample tissue and experimental analysis and is not used as a network input field.

[0062] As a refinement of the above embodiments, such as Figure 3 As shown, the lifetime prediction network based on physical information embedding includes two layers of causal convolution, a gated recurrent unit, and an output head. Step S4 specifically includes: S4-1 Causal Convolution Short-Time Characterization Extraction; The input is a 32×12 sequence of physical degradation states. The first causal convolution is a causal one-dimensional convolution with a kernel length of 3 and a dilation rate of 1, with an input of 32×12 and an output of 32×64. The second causal convolution is a causal one-dimensional convolution with a kernel length of 3 and a dilation rate of 2, with an input of 32×64 and an output of 32×64. The only output of S3-1 is a 32×64 short-term degradation feature sequence, which is then passed to S4-2.

[0063] Layer 1 adds two zero-level pauses at the beginning of the sequence and reads only the current segment and the two preceding segments; Layer 2 adds four zero-level pauses at the beginning and reads historical positions at an expansion interval of 2. After each convolutional layer, layer normalization, linear rectified activation, and random deactivation with a scale of 0.1 are performed sequentially. Random deactivation is only enabled during the training phase.

[0064] The single output position of the two-layer causal convolution covers the current and the previous six segments, used to extract shock changes and short-term operating condition responses; S4-2 then summarizes the long-term degradation relationships within the complete 32-segment range. The causal structure does not read data after the current time step, consistent with the condition that online lifetime prediction can only use historical records.

[0065] S4-2 gated timing state fusion; A 32×64 short-term degradation feature sequence is received and input into a 96-cell gated recurrent unit to obtain a 32×96 temporal state. The 96-dimensional state at the 32nd time step is taken as the historical degradation representation. Subsequently, the current 12-dimensional physical degradation state is received and concatenated into a 108-dimensional fusion vector in the order of historical representation first and current state last. The 108-dimensional vector is obtained by directly concatenating the 96-dimensional historical degradation representation with the 12-dimensional current physical degradation state. The first 96 dimensions describe the degradation changes within a 320s range, while the last 12 dimensions retain the current vibration, operating condition, and cumulative damage location to prevent the loss of the current state after long sequence compression.

[0066] A 108-dimensional fused vector is input to a 48-unit fully connected layer and linear rectified activation is performed, outputting a 48-dimensional shared latent state. The number of convolutional channels, gate units, and shared layer units are determined from candidate sets of 32 to 96, 64 to 128, and 32 to 64, respectively. Each candidate maintains consistency with the training data and optimization budget, and the combination with the smallest parameter count and a validation set MAPE within 0.2 percentage points of the optimal value is selected; in this embodiment, these are determined to be 64, 96, and 48. The sole output of S4-2 is the 48-dimensional shared latent state, which is simultaneously passed to the three output heads of S4-3.

[0067] S4-3 Multi-task Lifetime Output; The system receives a 48-dimensional shared latent state. The lifetime prediction branch outputs the predicted number of remaining segments sequentially through a fully connected layer of 24 neurons and a linear layer with a single value. The damage prediction branch outputs a non-negative predicted damage increment for the next segment sequentially through a fully connected layer of 16 neurons and a Softplus layer with a single value. The lifetime interval prediction branch outputs the 5th percentile and 95th percentile remaining segment numbers sequentially through a fully connected layer of 16 neurons and a linear layer with a single value. The predicted remaining segment number is in 10-second segments, representing the number of complete monitoring segments the current monitored segment needs to pass before reaching the lifetime endpoint marker.

[0068] The hidden layers of the lifetime head and damage head use linear rectified activation. The two outputs of the interval head correspond to the 5th percentile lifetime and the 95th percentile lifetime. The three output heads share the same 48-dimensional input, but their parameters are independent of each other. S4-3 outputs four values ​​in a fixed order: the number of predicted remaining segments, the predicted damage increment of the next segment, the number of remaining segments at the 5th percentile, and the number of remaining segments at the 95th percentile. The number of predicted remaining segments, the number of remaining segments at the 5th percentile, and the number of remaining segments at the 95th percentile are multiplied by 10s to obtain the point prediction, the lower quantile lifetime, and the upper quantile lifetime. The damage increment of the predicted next segment is used to calculate the damage evolution residual with the physical damage increment output by S3-2 during the training phase, and is used to recursively derive the physical lifetime reference in S5-1. The lower quantile lifetime and the upper quantile lifetime together form the 90% prediction interval. During the training phase, the four outputs of S4-3 are passed to S5-1. After training is completed, S4-1 to S4-3 are re-executed with fixed network parameters, and the four inference values ​​are passed to S5-2.

[0069] The life head describes the remaining operating time of the bearing, the damage head describes the degradation increment for the next monitoring interval, and the interval head describes the uncertainty range of the life estimate. All three tasks share the same 48-dimensional latent state, ensuring that vibration changes, current operating conditions, and accumulated damage consistently affect point prediction, damage prediction, and interval width. The point prediction is a definite remaining life prediction value output by the life prediction branch, distinct from the 90% prediction interval formed by the lower and upper quantile lives.

[0070] As a refinement of the above embodiments, such as Figure 3 As shown, step S5 includes: S5-1 Physical Consistency Joint Training; Receive the remaining lifetime label output by S1, the damage increment training reference output by S3-2, the four training predictions output by S4-3, and the physical lifetime reference, and calculate the joint training objective according to the following formula.

[0071] , In the formula, For the total losses of the joint training, As a weight for remaining lifetime loss, Weights for damage evolution loss, Weights for monotonic loss of lifetime, The weights are the interval quantile loss values.

[0072] The remaining lifetime loss is calculated based on the predicted remaining lifetime, the remaining lifetime label output by S1, and the physical lifetime reference. For training samples where the physical lifetime reference terminates normally, the Huber loss between the predicted remaining lifetime and the remaining lifetime label, and the Huber loss between the predicted remaining lifetime and the physical lifetime reference are calculated separately, and the average of the two is taken as the remaining lifetime loss. For training samples whose physical lifetime references terminate abruptly, only the Huber loss between the predicted remaining lifetime and the remaining lifetime label is calculated.

[0073] Damage evolution loss is the mean square error between the damage head output and the physical damage increment, used to pass the rated lifetime relationship into the shared latent state.

[0074] To compensate for monotonic life loss, a positive increase in the predicted life of the same bearing at adjacent times is penalized.

[0075] The interval quantile loss, namely the 5th percentile and 95th percentile loss, is used to determine the 90% prediction interval.

[0076] All four losses are calculated during the training phase; lifetime labels and training losses are no longer read during the inference phase.

[0077] Compute over all training windows with end-of-life markers; Physical information is embedded in the form of a basic rated life equation and a linear cumulative damage relation. Specifically, the basic rated life equation expresses the basic rated life as the ratio of the basic rated dynamic load to the equivalent dynamic load raised to the power of the life exponent, and the linear cumulative damage relation expresses the physical damage increment as the ratio of the segment revolutions to the basic rated life. These two physical equations are given in S3-1 and S3-2, respectively. Further, the predicted next segment damage increment output by the damage prediction branch is substituted into the two physical equations and subtracted from the physical damage increment determined by the basic rated dynamic load, equivalent dynamic load, and segment revolutions to obtain the damage evolution physical residual. The damage evolution loss is averaged with respect to the damage evolution physical residual, so that the damage head output approaches the physical value determined by the basic rated life equation during training.

[0078] Only two adjacent windows of the same bearing time are compared, and a positive penalty is applied when the predicted lifetime of the later window is higher than that of the earlier window. The 5% quantile lifetime is constrained to be no higher than the reference lifetime, and the 95% quantile lifetime is constrained to be no lower than the reference lifetime. The four losses jointly update the parameters of S4-1, S4-2, and S4-3, so that the short-term vibration characterization, long-term degradation representation, and multi-task output are adjusted based on the same physical degradation relationship.

[0079] The four weights are determined by a fixed validation batch after the first training epoch. The four average losses are calculated separately, normalized using the inverse of each average loss, and the sum of the weights is made equal to 1; if the average loss is below 0.000001, it is calculated as 0.000001. Once determined, the weights remain unchanged for the remaining training epochs, and the optimization direction for the total loss is minimization. Each training batch contains only windows from the training bearings; adjacent windows within the same bearing maintain their temporal order, and hidden states are not passed between bearings. Training terminates when the validation set MAPE does not decrease for 12 consecutive epochs, and the model parameters with the lowest validation set MAPE are retained. Model selection only reads the validation set metrics; test bearings are used for final evaluation only after parameter selection is completed.

[0080] The training samples take a continuous 32×12 physical degradation state sequence of a bearing as input, and use the current remaining lifetime label and the next segment damage increment training reference as supervision information. S5-1 outputs the fixed network parameters corresponding to the lowest MAPE in the validation set, and passes these fixed network parameters to S5-2 for use. After training, lifetime labels, physical lifetime references, loss calculations, and gradient updates are stopped; during the inference phase, only the state construction in S2, and S4-1, S4-2, S4-3, and S5-2 using fixed network parameters are retained.

[0081] The physical lifetime reference recursion method is as follows: Receive the next segment damage increment from the output of S4-3 and the current 12-dimensional physical degradation state from the output of S2, and recursively calculate the physical lifetime reference according to the following formula.

[0082] , In the formula, For the first The physical lifetime reference for each monitoring segment, in seconds; The time interval between adjacent monitoring segments is 10 seconds. For indicator functions, when When the value is less than 1, the value is 1; when the value reaches 1, the value is 0. For the recursive calculation starting from the current state, the first... Accumulated damage; The maximum segment number that can be recursively determined is the maximum lifespan of the device divided by 10 seconds.

[0083] Recursion from the current cumulative damage Begin by predicting the damage increment for the next segment from the damage prediction branch output. Add the current cumulative damage to obtain the predicted cumulative damage for the next segment; the subsequent... The predicted cumulative damage of each segment is based on The process is recursive, and the updated cumulative damage is used as the 12th field of the 12-dimensional physical degradation state, where For the predicted first The damage increment for each segment is calculated. Then, the first row of the original 32×12 sequence is removed, and the updated 12-dimensional state is appended to the 32nd row, forming a new 32×12 sliding window. This sliding window is then sequentially input into S4-1 to S4-3 to calculate the damage increment for the next segment. During the recursion, the vibration condition status remains the value corresponding to the currently monitored segment; only the cumulative damage field is updated.

[0084] The recursion terminates normally when the cumulative damage first reaches 1. If the termination condition is not met when the maximum allowable number of lifetime segments is reached, the physical lifetime reference is taken as the upper limit and the upper limit restriction mark is recorded, which is an abnormal termination. This sample only participates in the interval coverage check. If the recursion outputs a non-numerical value, the physical reference calculation for that sample is immediately stopped, and only the lifetime end mark is retained for participation in the lifetime error term.

[0085] S5-2 Lifetime Output Limitation and Failure Handling; Receive the fixed network parameters output from S5-1, and use the fixed network parameters to execute S4-1 to S4-3 to obtain point prediction, next segment damage increment, lower quantile lifetime, and upper quantile lifetime. First, negative point predictions are restricted to 0; if the current remaining lifetime is higher than the previous valid time step, the previous valid value is taken; if the single drop exceeds the 99th percentile of the drop in the validation set, the maximum drop corresponding to that percentile is taken.

[0086] After completing the point prediction correction, check that the lower quantile lifetime is not higher than the point prediction and the point prediction is not higher than the upper quantile lifetime. If the order is not satisfied, rearrange the three lifetime values ​​in ascending order. The damage increment of the next segment is only used for internal state recursion and numerical validity checks within the model and is not used as input to S6. S5-2 outputs the remaining lifetime prediction correction value, lower quantile lifetime, upper quantile lifetime, and anomalous state marker, and uses this fixed field sequence as input to S6.

[0087] As a refinement of the above embodiments, such as Figure 4 As shown, step S6 includes: S6-1 Reliability boundary determination; The system receives the remaining lifetime prediction correction value, lower quantile lifetime, upper quantile lifetime, and abnormal state markers, and calculates the relative width of the 90% prediction interval. Reliability boundaries are determined solely using the validation set. Coverage represents the proportion of validation samples whose remaining lifetime tags fall within the 90% prediction interval out of the total number of validation samples; the underreporting maintenance ratio represents the proportion of samples whose remaining lifetime tags are below the equipment's specified maintenance lead time, but whose predicted remaining lifetime is still above that lead time out of the total number of samples requiring maintenance; the manual review ratio represents the proportion of validation samples classified as low reliability out of the total number of validation samples. Among candidate boundaries with a coverage rate of not less than 85%, the boundary with the lowest underreporting maintenance ratio is selected first; if the underreporting maintenance ratios are the same, the boundary with the lowest manual review ratio is selected.

[0088] A value less than 30% and with an invalid anomalous state marker is marked as "High"; a value greater than 30% but less than 50% and with an invalid anomalous state marker is marked as "Medium"; and a value greater than 50% and with an valid anomalous state marker is marked as "Low". These boundaries remain unchanged during the testing and inference phases once determined. S6-1 outputs a reliability lifetime result consisting of the corrected point prediction, lower quantile lifetime, upper quantile lifetime, and reliability level, which is then used as input to S6-2.

[0089] S6-2 Life Result Recording and Maintenance Sorting; Receive the reliable life result output from S6-1, and read the bearing number and the current cumulative damage output from S2 to form a life record; then generate a maintenance sequence according to the corrected remaining life in ascending order. Finally, output the bearing number, current cumulative damage, remaining life, 90% prediction interval, reliability level, and maintenance sequence; records with low reliability levels are subject to manual review.

[0090] Example 2 Experimental Design and Effect Verification The comparative validation data generated 18 full-life sequences based on three constant load conditions and one variable load condition. These were further divided into 10 training sequences, 4 validation sequences, and 4 test sequences based on bearing number. Each sequence contained 180 to 420 monitoring segments with 10-second intervals. The bidirectional vibration consisted of a low-amplitude background, an impact component increasing with cumulative damage, fault characteristic frequency band components, and random noise. During the variable load phase, variations in rotational speed (5% to 12%) and radial load (8% to 18%) were added. The random seed was fixed at 808, and all methods used the same data partitioning, 32-segment windows, and training budget.

[0091] The experiment was conducted in a Python 3.12 environment. NumPy was used to generate descriptive monitoring sequences, perform state recursion, and record results using a fixed random seed. PyTorch was used to train the network and perform inference with three output heads. Training was performed using the AdamW optimizer with a batch size of 32 and an initial learning rate of 0.0005. Training was conducted for a maximum of 120 epochs, terminating when the MAPE on the validation set did not decrease for 12 consecutive epochs. The average of the results from three repetitions was used to evaluate random fluctuations.

[0092] The comparison methods include vibration regression, temporal convolution, and physical regularization. Vibration regression only accepts 8 vibration fields; temporal convolution accepts 32×8 vibration sequences; physical regularization adds damage residuals to the training loss of temporal convolution but does not input the 12-dimensional physical state into the network. This method maintains consistency with the three comparison methods in terms of training bearings, validation bearings, test bearings, optimizers, batch size, and stopping rules. When calculating MAPE, test samples with a remaining lifetime label less than 50s are uniformly bounded to 50s as the lower limit of the denominator; normalized RMSE is normalized using the maximum remaining lifetime label of each test sequence.

[0093] The relationship between physical degradation state and lifetime prediction in Experiment 1 is as follows: Figure 5 - Figure 7 As shown, Figure 5 The display shows the 12-dimensional physical degradation state arranged in the fixed field order of S2. The root mean square energy of bidirectional vibration and the frequency band energy of fault characteristics generally increase with the progress of life, the kurtosis preserves local impact fluctuations, the speed ratio and radial load ratio reflect changes in operating conditions, and the cumulative damage maintains a monotonically increasing trend. This result shows that the network input simultaneously preserves vibration degradation characteristics, operating conditions, and physical damage locations, avoiding misinterpreting individual image features and single impact indices as life states.

[0094] Figure 6 The data shows that the incremental damage per segment fluctuates locally with changes in load, while the cumulative damage continues to increase along the lifespan. Figure 7 The predicted remaining lifetime generally follows the reference remaining lifetime decline, with 90% of the prediction range covering the main reference trajectory.

[0095] Experiment 2: Lifetime Prediction Error and Lifetime Tracking Figure 8 , Figure 9 As shown, Figure 8 In the illustrative comparative validation, our method achieved a MAPE of 8.9% and a normalized RMSE of 7.6%. The closest physical regularization methods corresponded to 10.2% and 8.5%, respectively, while our method reduced these figures by 1.3 and 0.9 percentage points, respectively. Temporal convolution methods corresponded to 11.4% and 9.3%, and vibrational regression methods corresponded to 12.1% and 9.9%. The differences remained within the moderate range of similar methods, indicating that the direct participation of the 12-dimensional physical state in the shared representation leads to a stable improvement.

[0096] Figure 9 The results show that the overall point predictions of this method decrease along the reference lifetime, with slight fluctuations remaining at load switching locations, and 90% of the prediction range covers most of the reference trajectory. The physical canonical method shows higher estimates in the early part of the lifetime, with more pronounced fluctuations in the later part.

[0097] Experiment 3 Physical Constraint Ablation and Operating Condition Adaptability Figure 10 As shown, Figure 10 The MAPE values ​​for the complete method under low load, rated load, high load, and variable load conditions were 8.4%, 8.7%, 9.3%, and 10.1%, respectively. After removing the damage residuals, the corresponding values ​​were 10.0%, 10.3%, 11.1%, and 12.0%, with increments of 1.6 to 1.9 percentage points. After removing the lifetime monotonic constraint, the corresponding values ​​were 9.6%, 9.9%, 10.6%, and 11.5%, with increments of 1.2 to 1.4 percentage points. Both ablation methods showed moderate changes, supporting the roles of fragment-by-fraction damage reference and lifetime trend constraint, respectively.

[0098] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A bearing fatigue life prediction method based on physical information embedding, characterized in that, include: The original monitoring segments were composed of bidirectional vibration components, average rotational speed, average radial load, average bearing temperature, and average ambient temperature collected at the same time stamp. These segments were then arranged sequentially to construct a continuous full-life segment sequence for the bearing. The health stage parameters, degradation start markers, life end markers, and remaining life labels of the continuous full-life segment sequence for the bearing were determined. Vibration condition status is obtained based on the continuous full-life segment sequence of bearings and the parameter encoding of health stages. The basic rated life and damage increment are calculated segment by segment. The cumulative damage is calculated based on the damage increment. The cumulative damage is combined with the vibration condition to obtain the physical degradation state sequence. A lifetime prediction network based on physical information embedding is constructed and trained, including two layers of causal convolution, gated recurrent units, and an output head. The physical degradation state sequence is extracted into a short-term degradation feature sequence through causal convolution. The short-term degradation feature sequence is input into the gated recurrent unit to obtain the historical degradation representation. The historical degradation representation is concatenated with the physical degradation state and processed through a fully connected layer to obtain a shared latent state. The shared latent state is input into the output head to obtain point prediction, lower quantile lifetime, upper quantile lifetime, and next segment damage increment. The physical degradation state and physical degradation state sequence of the bearing to be predicted are collected as the validation set and input into the trained lifetime prediction network to obtain point prediction, lower quantile lifetime, upper quantile lifetime and damage increment of the next segment, and then corrected and abnormal state labels are added. The relative width of the 90% prediction interval is calculated based on the corrected point prediction, lower quantile life, upper quantile life, and abnormal state markers, and the reliability level is determined. The corrected point prediction, lower quantile life, and upper quantile life reliability levels are used as the reliable life results, and the bearing number and cumulative damage are used as the life records.

2. The bearing fatigue life prediction method based on physical information embedding according to claim 1, characterized in that, The specific methods for collecting raw monitoring segments are as follows: Piezoelectric accelerometers were fixed at the radial horizontal and radial vertical positions in the bearing area of ​​the test bearing housing. The sensitive axes of the piezoelectric accelerometers were aligned with the horizontal and vertical directions, respectively, to collect the horizontal and vertical vibration components. The average rotational speed, average radial load, average bearing temperature, and average ambient temperature were read. The horizontal vibration components, vertical vibration components, average rotational speed, average radial load, average bearing temperature, and average ambient temperature with the same timestamp were combined into one original monitoring segment. The original monitoring segments with complete and continuous time fields are filtered, and a continuous full-life segment sequence of bearings is constructed based on the timestamp order.

3. The bearing fatigue life prediction method based on physical information embedding according to claim 2, characterized in that, Determining the health stage parameters, degradation start markers, life end markers, and remaining life labels for a continuous full-life sequence of bearings, specifically through the following methods: The T consecutive complete monitoring segments that appear after the bearing reaches a stable speed and a stable radial load are defined as the healthy reference interval. Within the healthy reference interval, the coefficients of variation of the speed and radial load do not exceed 2%, and the root mean square of the bidirectional vibration does not show an increase of t consecutive monitoring segments. Calculate the health stage parameters, including: median energy of the characteristic frequency band of the health level fault, median energy of the characteristic frequency band of the health vertical fault, absolute deviation of the median energy of the characteristic frequency band of the health level fault, and absolute deviation of the median energy of the characteristic frequency band of the health vertical fault within the health reference interval; and record the median healthy rotational speed, median healthy radial load, root mean square median of healthy horizontal vibration, and root mean square median of healthy vertical vibration. When the horizontal fault characteristic band energy of the subsequent five consecutive monitoring segments is higher than the sum of the median absolute deviation of the healthy horizontal fault characteristic band energy and the median absolute deviation of the healthy horizontal fault characteristic band energy, or when the vertical fault characteristic band energy is higher than the sum of the median absolute deviation of the healthy vertical fault characteristic band energy and the median absolute deviation of the healthy vertical fault characteristic band energy, the first monitoring segment is marked as the degradation start point. The terminating segment of a continuous full-lifetime segment sequence is used as the lifetime end marker; The difference between the sequence number of the terminated segment and the sequence number of the current monitoring segment is multiplied by the original monitoring segment acquisition time interval. As a label for remaining lifespan.

4. The bearing fatigue life prediction method based on physical information embedding according to claim 3, characterized in that, The vibration condition status includes a vibration field and a condition field; The vibration field includes the root mean square of the horizontal vibration component, the kurtosis of the horizontal vibration component, the peak factor of the horizontal vibration component, the fault characteristic frequency band energy of the horizontal vibration component, the root mean square of the vertical vibration component, the kurtosis of the vertical vibration component, the peak factor of the vertical vibration component, and the fault characteristic frequency band energy of the vertical vibration component. The operating condition fields include the speed ratio, radial load ratio, and temperature rise ratio calculated based on the average speed, average radial load, average bearing temperature, and average ambient temperature.

5. The bearing fatigue life prediction method based on physical information embedding according to claim 4, characterized in that, The basic rated life is calculated as follows: ， In the formula, For the first The basic rated lifespan corresponding to each monitoring segment. The basic dynamic load rating is given in the bearing model manual. For the first The equivalent dynamic load of the monitoring segment, the first The average radial load of each monitoring segment is taken as the equivalent dynamic load of that monitoring segment. Life expectancy index; The damage increment is obtained by dividing the bearing rotation count within the original monitoring segment acquisition time interval by the basic rated life corresponding to the monitoring segment. The cumulative damage is obtained by accumulating the damage increment based on the monitoring segment.

6. The bearing fatigue life prediction method based on physical information embedding according to claim 5, characterized in that, The cumulative damage and vibration condition are combined to obtain the physical degradation state sequence, as follows: After the vibration condition status is normalized by percentile, the cumulative damage is added as the last field to the vibration condition status to obtain the physical degradation status. The normalized boundaries of each field of the vibration condition status are used to form a boundary matrix. The physical degradation state is normalized using a boundary matrix. The normalized physical degradation state is then arranged in chronological order to form a physical degradation state sequence, where rows correspond to monitoring segments and columns correspond to fields of the physical degradation state.

7. The bearing fatigue life prediction method based on physical information embedding according to claim 6, characterized in that, The first causal convolutional layer adds two all-zero time points at the beginning of the physical degradation state sequence, and only reads the current segment and the two segments before it; The second causal convolutional layer adds four all-zero time points to the front of the physical degradation state sequence and reads the historical position at an expansion interval of 2. After each convolutional layer, layer normalization and linear rectified activation are performed sequentially. During the training phase, random deactivation with a ratio of 0.1 is also performed. The gated loop unit inputs a short-term degradation feature sequence to obtain a time-series state. The time-series state at the last moment is used as a historical degradation representation and is sequentially concatenated with the physical degradation state to form a fusion vector. The output head includes a lifetime head, a damage head, and a lifetime interval head. The lifetime head processes the fused features sequentially through two linear layers to obtain the predicted number of remaining segments. The damage head processes the fused features sequentially through a linear layer and a Softplus layer to obtain the predicted damage increment for the next segment. The lifetime range header processes the fused features sequentially through two linear layers to obtain the number of remaining fragments at the 5% quantile and the number of remaining fragments at the 95% quantile. The hidden layers of the lifespan head and the damage head are activated by linear rectification; The predicted number of remaining segments, the number of remaining segments at the 5th percentile, and the number of remaining segments at the 95th percentile are multiplied by the original monitoring segment acquisition time interval, respectively. We obtain point predictions, lower quantile lifetimes, and upper quantile lifetimes.

8. The bearing fatigue life prediction method based on physical information embedding according to claim 7, characterized in that, The joint training objective of the lifetime prediction network based on physical information embedding is as follows: ， in, For remaining life loss, For damage evolution loss, For monotonic loss of lifespan, Interval quantile loss, As a weight for remaining lifetime loss, Weights for damage evolution loss, Weights for monotonic loss of lifetime, The weights for interval quantile loss; The acquisition method is as follows: For training samples whose physical lifetime references terminate normally, calculate the Huber loss between point prediction and remaining lifetime label, and the Huber loss between point prediction and physical lifetime reference, and take the average of the two; for training samples whose physical lifetime references terminate abnormally, only calculate the Huber loss between point prediction and remaining lifetime label. The physical lifetime is calculated recursively based on the damage increment of the next segment and the physical degradation state, and the specific calculation method is as follows: ， ， In the formula, For the first Physical lifetime reference for each monitoring segment, For indicator functions, when The value is 1 when it is less than 1, and 0 when it reaches 1. For the recursive calculation starting from the current state, the first... Cumulative damage; The maximum allowed segment number for recursion is calculated by dividing the device's specified lifespan by [the specified value]. Sure, For the predicted first Incremental damage to each segment; The recursion terminates normally when the cumulative damage first reaches 1. If the termination condition is not met when the maximum allowable number of lifetime segments of the equipment is reached, the physical lifetime reference is taken as the upper limit and the recursion terminates abnormally. If the recursion outputs a non-numerical value, the physical reference calculation is immediately stopped, and only the lifetime end mark is retained to participate in the lifetime error term. It is the mean square error between the damage head output and the physical damage increment; A positive penalty is applied when the predicted lifetime in the later window is higher than that in the previous window. This includes 5% quantile loss and 95% quantile loss, with constraints that the 5% quantile lifetime should not exceed the reference lifetime and the 95% quantile lifetime should not be lower than the reference lifetime.

9. The bearing fatigue life prediction method based on physical information embedding according to claim 8, characterized in that, The point prediction, lower quantile lifetime, upper quantile lifetime, and next segment damage increment are corrected and abnormal state markers are added, as follows: Negative point predictions are limited to 0. When the current point prediction is higher than the point prediction of the previous valid time step, the point prediction of the previous valid time step is taken. When the single drop of the point prediction exceeds the 99th percentile of the drop in the validation set, the maximum drop corresponding to that percentile is taken. Check if the lower quantile lifetime is not higher than the point prediction and the point prediction is not higher than the upper quantile lifetime. If not, rearrange the points prediction, lower quantile lifetime, and upper quantile lifetime in ascending order, and use them as the lower quantile lifetime correction value, point prediction correction, and upper quantile lifetime in that order.

10. The bearing fatigue life prediction method based on physical information embedding according to claim 9, characterized in that, The relative width of the 90% prediction interval is calculated and the reliability level is determined based on the corrected point prediction, lower quantile lifetime, upper quantile lifetime, and anomalous state markers, as follows: Based on the corrected post-point prediction, lower quantile lifetime, upper quantile lifetime, and anomalous state markers, the relative width of the 90% prediction interval and the reliability level are calculated; including: Coverage rate represents the proportion of validation samples whose remaining lifetime labels fall within the 90% prediction range out of the total number of validation samples. The underreporting rate of maintenance indicates the proportion of samples whose remaining life label is lower than the equipment's specified maintenance lead time, but whose point prediction is still higher than the maintenance lead time, out of the total number of samples that should be maintained. The manual review ratio represents the proportion of verification samples classified as low reliability out of the total number of verification samples. Among the candidate reliability boundaries with a coverage rate of not less than 85%, the boundary with the lowest rate of missed maintenance is selected first; if the rates of missed maintenance are the same, the boundary with the lowest rate of manual review is selected. When the relative width is no more than 30% and the abnormal state marker is invalid, the reliability level is marked as high; when it exceeds 30% but does not exceed 50% and the abnormal state marker is invalid, the reliability level is marked as medium; when the relative width exceeds 50% and the abnormal state marker is valid, the reliability level is marked as low.

Images