Component health prediction method and system fusing physical model and data

Abstract

This invention provides a method and system for predicting the health of components by integrating physical models and data, relating to the field of industrial equipment health management technology. First, it acquires the raw monitoring data stream and a physical degradation mechanism model library of the target mechanical component. The raw monitoring data stream undergoes multi-scale time-frequency feature decomposition and is filtered and matched to generate a set of candidate feature time series. Next, it performs adaptive identification of model parameters, calculates the matching confidence level to select the dominant degradation model, and fuses its parameters with candidate features to generate a fusion characterization parameter for the health status. Finally, it predicts the remaining effective lifespan. This invention integrates the advantages of physical and data analysis, resulting in accurate and reliable predictions.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment health management technology, and more specifically, to a method and system for predicting component health by integrating physical models and data. Background Technology

[0002] In the industrial sector, predicting the health of mechanical components is crucial for ensuring stable equipment operation, reducing maintenance costs, and preventing production interruptions caused by unexpected failures. Currently, methods for predicting the health of mechanical components are mainly divided into physical model-based methods and data-driven methods.

[0003] Physical model-based methods describe the degradation process of mechanical components by establishing accurate physical degradation models. These methods rely on a deep understanding of the internal physical mechanisms of the components, theoretically explaining the causes of failure and degradation patterns. However, the actual operating environment of mechanical components is complex and variable, influenced by various factors such as temperature, load, and vibration. These factors make it extremely difficult to accurately establish a physical model that comprehensively reflects the actual operating conditions. Moreover, physical models typically require extensive professional knowledge and complex mathematical derivations; the model establishment and solution process is cumbersome and computationally costly. For some complex mechanical components, it is even difficult to establish an effective physical model.

[0004] Data-driven methods primarily utilize historical monitoring data, employing algorithms such as machine learning and deep learning to uncover potential patterns within the data, thereby predicting component health. These methods do not require in-depth understanding of the component's physical mechanisms, can handle large amounts of data, and automatically learn features from the data. However, data-driven methods suffer from issues of data quality and representativeness. If the monitoring data contains noise, missing values, or outliers, or if the data does not comprehensively reflect the various operating states and failure modes of the component, the accuracy and reliability of the prediction results will be affected. Furthermore, data-driven methods lack an explanation of the physical mechanisms, making it difficult to fundamentally understand the component degradation process. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for predicting component health by integrating physical models and data, the method comprising: Obtain the raw monitoring data stream of the target mechanical component's operating status during a continuous monitoring period, and obtain the physical degradation mechanism model library corresponding to the target mechanical component. The physical degradation mechanism model library contains multiple component physical degradation sub-models describing different failure modes. The original monitoring data stream of the component's operating status is subjected to multi-scale time-frequency feature decomposition processing to extract the energy distribution evolution features of the target mechanical component at multiple frequency band scales. The energy distribution evolution features are then filtered and matched according to the failure mode feature frequency range corresponding to each of the multiple component physical degradation sub-models to generate a set of candidate feature time series associated with each component physical degradation sub-model. The candidate feature time series set is input into the corresponding component physical degradation sub-model for adaptive identification of model parameters, generating the model parameter evolution sequence of each component physical degradation sub-model in the current monitoring period, and calculating the matching confidence of each component physical degradation sub-model with the actual degradation process of the target mechanical component based on the model parameter evolution sequence; The component physical degradation sub-model with the highest matching confidence is selected as the dominant degradation model, and the model parameter evolution sequence corresponding to the dominant degradation model is deeply fused with the candidate feature time series set corresponding to the dominant degradation model to generate component health status fusion characterization parameters that fuse physical laws and data features. The remaining effective life prediction process is performed on the fusion characterization parameters of the component's health status to generate the remaining effective life prediction result of the target mechanical component.

[0006] Furthermore, embodiments of the present invention also provide a component health prediction system that integrates physical models and data, comprising: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned component health prediction method that integrates physical models and data by executing the machine-executable instructions.

[0007] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of the component health prediction system integrating physical models and data reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the component health prediction system integrating physical models and data to execute the above-described component health prediction method integrating physical models and data.

[0008] Based on the above, by acquiring the raw monitoring data stream of the target mechanical component's operating status and the corresponding physical degradation mechanism model library, the physical model and data are combined, giving full play to the advantages of both. Multi-scale time-frequency feature decomposition processing is performed on the raw monitoring data stream, enabling the extraction of energy distribution evolution characteristics of the target mechanical component at different frequency bands. Based on the failure mode characteristic frequency range corresponding to the component's physical degradation sub-model, the energy distribution evolution characteristics are filtered and matched, generating a set of candidate feature time series associated with each sub-model. This achieves a preliminary association between data features and the physical model. The candidate feature time series set is then input into the corresponding component physical degradation sub-model for adaptive model parameter identification processing, generating model parameters. By analyzing the evolution sequence and calculating the matching confidence, the matching degree between each sub-model and the actual degradation process can be dynamically evaluated, thereby accurately selecting the dominant degradation model. The evolution sequence of the model parameters of the dominant degradation model is then deeply fused with the corresponding candidate feature time series set to generate a component health status fusion characterization parameter that integrates physical laws and data features. This considers both the physical degradation mechanism of the component and makes full use of the information from actual monitoring data. Finally, based on the fusion characterization parameter, the remaining effective life prediction process is performed, which can generate accurate and reliable prediction results of the remaining effective life of the target mechanical component. This helps to arrange maintenance plans in advance, avoid unexpected failures, improve the reliability and availability of equipment, and reduce maintenance costs and production losses. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the execution flow of the component health prediction method that integrates physical models and data provided in the embodiments of the present invention.

[0010] Figure 2 This is a schematic diagram of exemplary hardware and software components of the component health prediction system that integrates physical models and data provided in this embodiment of the invention. Detailed Implementation

[0011] Figure 1 This is a flowchart illustrating a component health prediction method that integrates physical models and data, as provided in one embodiment of the present invention. A detailed description follows.

[0012] Step S110: Obtain the raw monitoring data stream of the target mechanical component's operating status during the continuous monitoring period, and obtain the physical degradation mechanism model library corresponding to the target mechanical component. The physical degradation mechanism model library contains multiple component physical degradation sub-models describing different failure modes.

[0013] In this embodiment, the target mechanical component is specifically a commercial vehicle driveshaft. The continuous monitoring period is set to 30 consecutive days, and the sampling frequency is once per hour. The raw monitoring data stream of the component's operating status includes vibration signal data collected by a vibration sensor installed at the driveshaft bearing housing and acoustic emission signal data collected by an acoustic sensor. The vibration sensor is a piezoelectric accelerometer with a fixed sampling frequency. The collected raw vibration signal data is time-series data, and each data point includes acceleration values ​​in the X, Y, and Z axes. The acoustic sensor is a high-sensitivity microphone with a fixed sampling frequency. The collected raw acoustic emission signal data is single-channel sound pressure level time-series data. The physical degradation mechanism model library is a pre-built database stored in the edge computing module, which contains component physical degradation sub-models for various failure modes that may occur in the driveshaft. For example, there's a physical degradation sub-model for fatigue crack failure mode, based on fracture mechanics theory, describing the relationship between crack propagation rate and stress intensity factor; a physical degradation sub-model for lubrication failure mode, based on fluid lubrication theory, describing the variation of lubricating oil film thickness with parameters such as temperature and rotational speed; and a physical degradation sub-model for bearing wear failure mode, based on wear theory, describing the cumulative process of wear over operating time. Each component's physical degradation sub-model has a unique identifier and corresponding failure mode description information.

[0014] Step S120: Perform multi-scale time-frequency feature decomposition processing on the original monitoring data stream of the component's operating status, extract the energy distribution evolution features of the target mechanical component on multiple frequency band scales, and perform feature screening and matching processing on the energy distribution evolution features according to the failure mode feature frequency range corresponding to each of the multiple component physical degradation sub-models, generating a set of candidate feature time series associated with each component physical degradation sub-model.

[0015] Step S121: Perform data stream segmentation processing on the original monitoring data stream of the component's operating status. Divide the original monitoring data stream of the component's operating status into multiple continuous and non-overlapping monitoring data time window units according to a preset fixed time window length. Each monitoring data time window unit contains the original vibration signal sequence and the original acoustic emission signal sequence of the target mechanical component within the corresponding time interval.

[0016] In this embodiment, the preset fixed time window length is set to 10 minutes, meaning that data collected in 10-minute intervals is divided into a monitoring data time window unit. Since the continuous monitoring period is 30 days, with 24 hours per day and 60 minutes per hour, each monitoring data time window unit is 10 minutes long. Therefore, the number of monitoring data time window units per day is 24 × 6 × 60 ÷ 10 = 864, and the total number of monitoring data time window units over 30 days is 30 × 864 = 25920. Each monitoring data time window unit contains the original vibration signal sequence and the original acoustic emission signal sequence collected within the corresponding 10-minute time interval. For example, the first monitoring data time window unit contains the original vibration signal sequence and the original acoustic emission signal sequence from 00:00 to 00:10 on the first day, the second monitoring data time window unit contains the corresponding signal sequence from 00:10 to 00:20, and so on. Each original vibration signal sequence contains acceleration values ​​in three directions: X-axis, Y-axis, and Z-axis. The number of sampling points in each direction is the sampling frequency multiplied by 10 minutes. Assuming the sampling frequency is A Hz, then the number of sampling points in each direction is A × 600, and the total data volume of the original vibration signal sequence is 3 × A × 600 data points. The original acoustic emission signal sequence is a single channel, and the number of sampling points is also A × 600 data points.

[0017] Step S122: Perform signal preprocessing operations on the original vibration signal sequence and the original acoustic emission signal sequence within each monitoring data time window unit. The signal preprocessing operation includes using a bandpass filter to filter out environmental noise interference components in the original vibration signal sequence and the original acoustic emission signal sequence that are below a first preset frequency threshold and above a second preset frequency threshold, so as to obtain the filtered vibration signal sequence and the filtered acoustic emission signal sequence corresponding to each monitoring data time window unit.

[0018] In this embodiment, for the vibration signal and acoustic emission signal of the commercial vehicle driveshaft, the first preset frequency threshold is set to B Hz, and the second preset frequency threshold is set to C Hz, where B is less than C. A Butterworth filter is used as the bandpass filter, with an order of D. For the original vibration signal sequence within each monitoring data time window unit, bandpass filtering is performed on the X-axis, Y-axis, and Z-axis signals respectively. Specifically, the original vibration signal sequence is input into a pre-designed Butterworth bandpass filter with passband cutoff frequencies of B Hz and C Hz. After passing through the filter, frequency components below B Hz and above C Hz are filtered out, retaining frequency components between B Hz and C Hz, resulting in the X-axis, Y-axis, and Z-axis components of the filtered vibration signal sequence. These components are then combined to form the filtered vibration signal sequence corresponding to the monitoring data time window unit. The original acoustic emission signal sequence is also processed using a Butterworth bandpass filter with the same parameters to filter out environmental noise interference components below B Hz and above C Hz, resulting in the filtered acoustic emission signal sequence. After the above preprocessing operations, low-frequency and high-frequency noise that is not related to the operating state of the drive shaft can be effectively removed, while retaining frequency components that are useful for fault diagnosis and health prediction.

[0019] Step S123: Perform adaptive wavelet packet decomposition layer determination processing on the filtered vibration signal sequence and filtered acoustic emission signal sequence corresponding to each monitoring data time window unit. Dynamically adjust the wavelet packet decomposition layer based on the comparison result between the total energy value of the signal in each monitoring data time window unit and the preset energy threshold, so that the frequency resolution corresponding to the lowest layer frequency band after decomposition meets the preset resolution requirement, and generate the vibration signal wavelet packet decomposition tree structure and acoustic emission signal wavelet packet decomposition tree structure corresponding to each monitoring data time window unit.

[0020] In this embodiment, the preset energy threshold is E, and the preset resolution requirement is that the frequency interval is no greater than F Hz. For the filtered vibration signal sequence corresponding to each monitoring data time window unit, the total energy value of the signal is first calculated. The total energy value is calculated by adding the squares of all sampling points in the filtered vibration signal sequence. Then, the total energy value is compared with the preset energy threshold E. If the total energy value is greater than or equal to E, the initial number of wavelet packet decomposition layers is determined to be G layers; if the total energy value is less than E, the initial number of layers is reduced to G-1 layers. Next, the initial number of layers is verified to check whether the frequency resolution corresponding to the lowest frequency band after decomposition meets the preset resolution requirement. The frequency resolution of wavelet packet decomposition is calculated by dividing (second preset frequency threshold - first preset frequency threshold) by (2 raised to the power of the number of decomposition layers). For example, if the number of decomposition layers is G layers, then the frequency resolution is (CB) / (2^G). If the resolution is less than or equal to F Hz, the number of decomposition layers is accepted; if it is greater than F Hz, the number of decomposition layers is increased until the resolution requirement is met. Following the above method, an appropriate number of wavelet packet decomposition levels is determined for the filtered vibration signal sequence of each monitoring data time window unit, and a corresponding vibration signal wavelet packet decomposition tree structure is generated. Similarly, the same processing is performed on the filtered acoustic emission signal sequence corresponding to each monitoring data time window unit to determine the adaptive number of wavelet packet decomposition levels and generate an acoustic emission signal wavelet packet decomposition tree structure. In the wavelet packet decomposition tree structure, each node represents a frequency band, the root node is the original signal, and each node is decomposed into two child nodes, each corresponding to a different frequency band range.

[0021] Step S124: Extract the vibration signal wavelet packet reconstruction coefficient sequence corresponding to all leaf nodes from the vibration signal wavelet packet decomposition tree structure corresponding to each monitoring data time window unit, and extract the acoustic emission signal wavelet packet reconstruction coefficient sequence corresponding to all leaf nodes from the acoustic emission signal wavelet packet decomposition tree structure corresponding to each monitoring data time window unit, thus forming the original multi-scale decomposition coefficient set for each monitoring data time window unit.

[0022] In this embodiment, after the wavelet packet decomposition tree structure of the vibration signal is determined by adaptive layer number, its leaf nodes are the nodes corresponding to each frequency band at the lowest level. For the wavelet packet decomposition tree structure of the vibration signal in each monitoring data time window unit, all leaf nodes are traversed, and the wavelet packet reconstruction coefficient sequence corresponding to each leaf node is extracted. The length of each reconstruction coefficient sequence is the same as the length of the original filtered vibration signal sequence, which is A×600 data points. The above vibration signal wavelet packet reconstruction coefficient sequences are combined to form the original multi-scale decomposition coefficient set of the vibration signal part of the monitoring data time window unit. Similarly, for the wavelet packet decomposition tree structure of the acoustic emission signal, the wavelet packet reconstruction coefficient sequences of the acoustic emission signal corresponding to all leaf nodes are extracted, and the length of each sequence is also A×600 data points. After combination, the original multi-scale decomposition coefficient set of the acoustic emission signal part is formed. The original multi-scale decomposition coefficient sets of the vibration signal part and the acoustic emission signal part are merged to obtain the complete original multi-scale decomposition coefficient set of each monitoring data time window unit. For example, if the wavelet packet decomposition of the vibration signal has H leaf nodes and the acoustic emission signal has I leaf nodes, then the original multiscale decomposition coefficient set contains H+I reconstructed coefficient sequences.

[0023] Step S125: Perform energy value calculation processing on each vibration signal wavelet packet reconstruction coefficient sequence and each acoustic emission signal wavelet packet reconstruction coefficient sequence in the original multi-scale decomposition coefficient set of each monitoring data time window unit. Calculate the sum of squares of each reconstruction coefficient sequence as the instantaneous energy feature value under that frequency band scale, and obtain the vibration signal energy distribution vector and acoustic emission signal energy distribution vector corresponding to each monitoring data time window unit.

[0024] In this embodiment, for each vibration signal wavelet packet reconstruction coefficient sequence in the original multi-scale decomposition coefficient set, the energy value is calculated by summing the squares of all data points in the sequence. The result is the instantaneous energy feature value at that frequency band scale. For example, if a vibration signal wavelet packet reconstruction coefficient sequence contains A×600 data points, namely a1, a2, ..., a(A×600), then its energy value is a1² + a2² + ... + a(A×600)². The above calculation is performed on all vibration signal wavelet packet reconstruction coefficient sequences to obtain H energy values. These energy values ​​are arranged according to their corresponding frequency band order to form the vibration signal energy distribution vector corresponding to the monitoring data time window unit. The dimension of this vibration signal energy distribution vector is H. Similarly, the sum of squares of the wavelet packet reconstruction coefficient sequence of each acoustic emission signal is calculated as the energy value, resulting in I energy values. These energy values ​​are arranged in frequency band order to form an acoustic emission signal energy distribution vector with a dimension of I. The energy distribution vector reflects the energy distribution at different frequency band scales. Different failure modes will produce energy changes in specific frequency bands. Therefore, these energy values ​​are an important basis for subsequent feature screening and matching.

[0025] Step S126: Arrange the vibration signal energy distribution vectors corresponding to all monitoring data time window units within the continuous monitoring period in chronological order to form a vibration signal energy distribution evolution matrix. Arrange the acoustic emission signal energy distribution vectors corresponding to all monitoring data time window units within the continuous monitoring period in chronological order to form an acoustic emission signal energy distribution evolution matrix. The rows of the vibration signal energy distribution evolution matrix correspond to different frequency band scales, and the columns correspond to different time window units. The rows of the acoustic emission signal energy distribution evolution matrix correspond to different frequency band scales, and the columns correspond to different time window units.

[0026] In this embodiment, there are 25,920 monitoring data time window units within the continuous monitoring period. The vibration signal energy distribution vector corresponding to each monitoring data time window unit is H-dimensional. These 25,920 vectors are arranged in chronological order, with the vector of the first time window unit as the first column of the matrix, the second as the second column, and so on, forming an H-row, 25,920-column vibration signal energy distribution evolution matrix. Each element (h, k) in the matrix represents the energy value of the h-th frequency band scale in the k-th time window unit. Similarly, the acoustic emission signal energy distribution vector is I-dimensional, and the 25,920 vectors are arranged in chronological order to form an I-row, 25,920-column acoustic emission signal energy distribution evolution matrix, where the element (i, k) represents the energy value of the i-th frequency band scale in the k-th time window unit. These two matrices illustrate the energy evolution of different frequency band scales over time.

[0027] Step S127: Analyze the failure mode description information corresponding to each component physical degradation sub-model contained in the physical degradation mechanism model library, extract the lower limit and upper limit of the characteristic frequency range of the failure mode in the vibration signal, and the lower limit and upper limit of the characteristic frequency range in the acoustic emission signal from each failure mode description information, to form the vibration characteristic frequency range and acoustic emission characteristic frequency range corresponding to each component physical degradation sub-model.

[0028] In this embodiment, each component physical degradation sub-model in the physical degradation mechanism model library has detailed failure mode description information. For example, for the component physical degradation sub-model of fatigue crack failure mode, its failure mode description information records that in vibration signals, when fatigue cracks occur, there will be a significant energy change in the frequency range of J Hz to K Hz, so the vibration characteristic frequency range is [J, K]; in acoustic emission signals, the corresponding characteristic frequency range is L Hz to M Hz, and the acoustic emission characteristic frequency range is [L, M]. For the component physical degradation sub-model of lubrication failure mode, the vibration characteristic frequency range may be [N, O], and the acoustic emission characteristic frequency range may be [P, Q]. By parsing these description information, the lower and upper limits of the vibration characteristic frequency range and the lower and upper limits of the acoustic emission characteristic frequency range corresponding to each component physical degradation sub-model are extracted using a text parsing algorithm and stored as attribute information for each sub-model.

[0029] Step S128: Based on the vibration characteristic frequency range corresponding to each component physical degradation sub-model, select the energy distribution row vectors corresponding to all frequency band scales whose center frequency falls within the vibration characteristic frequency range from the vibration signal energy distribution evolution matrix, and combine each selected energy distribution row vector with the corresponding time window unit index to generate a set of vibration source candidate feature time series corresponding to the component physical degradation sub-model.

[0030] In this embodiment, each row of the vibration signal energy distribution evolution matrix corresponds to a frequency band scale, and each frequency band scale has its center frequency. The center frequency is calculated as the average of the lower limit frequency and the upper limit frequency of the frequency band. For example, if the frequency range of a certain frequency band is [R, S], then the center frequency is (R+S) / 2. For the vibration characteristic frequency interval [J, K] of each component physical degradation sub-model, all rows of the vibration signal energy distribution evolution matrix are traversed, and the center frequency of the corresponding frequency band for each row is calculated to determine whether the center frequency falls within the interval [J, K]. If so, the energy distribution row vector corresponding to that row is extracted. This energy distribution row vector contains the energy values ​​of that frequency band scale in all 25920 time window units. Each extracted energy distribution row vector is combined with the corresponding time window unit index (i.e., column index, from 1 to 25920) to form a time series, where each data point contains the time window unit index and the corresponding energy value. Combining all the above time series together constitutes the set of candidate feature time series of vibration sources corresponding to the component physical degradation sub-model. For example, if the center frequencies of T frequency bands fall within the vibration characteristic frequency range, then the set of candidate characteristic time series of vibration sources will contain T time series.

[0031] Step S129: Based on the acoustic emission characteristic frequency range corresponding to each component physical degradation sub-model, select the energy distribution row vectors corresponding to all frequency band scales whose center frequency falls within the acoustic emission characteristic frequency range in the acoustic emission signal energy distribution evolution matrix, and combine each selected energy distribution row vector with the corresponding time window unit index to generate a set of acoustic emission source candidate feature time series corresponding to the component physical degradation sub-model.

[0032] In this embodiment, similar to step S128, each row of the acoustic emission signal energy distribution evolution matrix also corresponds to a frequency band scale and has a corresponding center frequency. For the acoustic emission characteristic frequency interval [L, M] of each component physical degradation sub-model, all rows of the acoustic emission signal energy distribution evolution matrix are traversed, the center frequency of each row's frequency band is calculated, and it is determined whether it falls within the [L, M] interval. For rows that fall within the interval, their energy distribution row vector is extracted and combined with the time window unit index to form a candidate feature time series of the acoustic emission source. For example, assuming that there are U frequency bands that meet the conditions, the set of candidate feature time series of the acoustic emission source contains U time series.

[0033] Step S1210: Normalize the energy values ​​in the vibration source candidate feature time series set and the acoustic emission source candidate feature time series set corresponding to each component physical degradation sub-model. Merge the normalized vibration source candidate feature time series set and the acoustic emission source candidate feature time series set to form a complete candidate feature time series set corresponding to the component physical degradation sub-model. Each candidate feature time series in the complete candidate feature time series set contains the normalized energy evolution information of each time window unit within the continuous monitoring period at the corresponding specific frequency band scale.

[0034] In this embodiment, the normalization process employs the min-max normalization method, mapping each energy value to the interval [0, 1]. For each time series in the candidate feature time series set of vibration sources, the maximum and minimum energy values ​​in the series are first identified. Then, the normalized energy value is calculated as (original energy value - minimum value) / (maximum value - minimum value). If the maximum value equals the minimum value, all energy values ​​in the series are normalized to 0.5. Similarly, the same min-max normalization process is performed on each time series in the candidate feature time series set of acoustic emission sources. After normalization, all time series in the candidate feature time series sets of vibration sources and acoustic emission sources are merged together to form a complete candidate feature time series set. For example, if the vibration source has T time series and the acoustic emission source has U time series, then the complete candidate feature time series set contains T+U time series. Each time series contains 25,920 normalized energy values, corresponding to each time window unit within the continuous monitoring period, reflecting the normalized energy evolution at a specific frequency band scale.

[0035] Step S130: Input the candidate feature time series set into the corresponding component physical degradation sub-model for adaptive identification of model parameters, generate the model parameter evolution sequence of each component physical degradation sub-model in the current monitoring period, and calculate the matching confidence of each component physical degradation sub-model with the actual degradation process of the target mechanical component based on the model parameter evolution sequence.

[0036] Step S131: Obtain the predefined state space expression for each component physical degradation sub-model. The state space expression includes a physical state transition equation describing the evolution of the component degradation state over time and a physical observation equation describing the mapping relationship between the candidate feature time series and the component degradation state. The physical state transition equation contains a physical model parameter vector to be identified, and the physical observation equation contains preset observation noise statistical characteristic parameters.

[0037] In this embodiment, taking the component physical degradation sub-model of fatigue crack failure mode as an example, the physical state transition equation in its state space expression is a nonlinear differential equation, in the form of x(k+1)=f(x(k), θ)+w(k), where x(k) is the component degradation state variable vector at time k, including state components such as crack length and crack propagation rate; θ is the physical model parameter vector to be identified, including parameters such as material constants and stress intensity factor; w(k) is the process noise, which follows a Gaussian distribution with a mean of 0 and a covariance matrix of Q. The physical observation equation is z(k)=h(x(k), θ)+v(k), where z(k) is the observed value of the candidate feature time series at time k; h(·) is the observation function, describing the physical mapping relationship between the component degradation state and the observed features; v(k) is the observation noise, whose statistical characteristic parameters are preset to have a mean of 0 and a covariance matrix of R. These parameters are obtained statistically based on the technical specifications of the sensor and historical data. Different component physical degradation sub-models have different state-space expressions. For example, the state transition equation of the lubrication failure mode sub-model may involve state variables such as lubricating oil film thickness and temperature, while the observation equation is related to the characteristics of acoustic emission signals.

[0038] Step S132: Take the candidate feature time series set corresponding to each component physical degradation sub-model as the observation input sequence of the component physical degradation sub-model, and initialize the prior probability distribution of the physical model parameter vector and the initial probability distribution of the component degradation state variable in the component physical degradation sub-model. The prior probability distribution is preset according to the statistical characteristics of similar historical components.

[0039] In this embodiment, for each component physical degradation sub-model, all time series in its corresponding complete candidate feature time series set are used as the observation input sequence. For example, the candidate feature time series set of the fatigue crack failure mode sub-model contains multiple time series, each corresponding to the energy evolution of different frequency bands. The above time series are combined into a multi-dimensional observation input sequence z(1), z(2), ..., z(25920), where each z(k) is a (T+U) dimensional vector. The prior probability distribution of the physical model parameter vector θ is set to a Gaussian distribution. The mean θ0 and the covariance matrix Pθ0 are obtained statistically from historical data of similar commercial vehicle drive shafts under fatigue crack failure mode. For example, by estimating parameters from a large amount of historical fault data, the mean and variance of each component of θ are obtained, and then the prior covariance matrix is ​​constructed. The initial probability distribution of the component degradation state variable x is also set to a Gaussian distribution. The initial mean x0 and covariance matrix Px0 are determined based on information such as the initial state and manufacturing tolerances of the new component. For example, the initial crack length is set to 0. The covariance matrix reflects the uncertainty of the initial state.

[0040] Step S133: For each component physical degradation sub-model, the sequential importance sampling particle filter algorithm is used to perform joint online recursive estimation processing on the physical model parameter vector and the component degradation state variable. When each time window unit arrives, the importance weight of each particle is updated according to the candidate feature time series value corresponding to the current time window unit, and the particle set is resampled to eliminate the particle degradation phenomenon, so as to obtain the posterior probability distribution sample set of the physical model parameter vector and the posterior probability distribution sample set of the component degradation state variable corresponding to each time window unit.

[0041] In this embodiment, the specific implementation process of the sequential importance sampling particle filter algorithm is as follows: First, N particles are generated for the physical model parameter vector θ and the component degradation state variable x. Each particle is represented as (θ(i), x(i)), i=1,2,...,N, where the value of N is set according to the computational resources and estimation accuracy requirements, for example, N=1000. At the initial moment, the particles are sampled according to the prior probability distribution. For each time window unit k (from 1 to 25920), the state of each particle is first predicted according to the physical state transition equation, i.e., x(i)(k)=f(x(i)(k-1), θ(i))+w(i)(k), where w(i)(k) is the noise sampled from the process noise distribution. Next, the importance weight of each particle is calculated. The weight is based on the observation likelihood function, i.e., w(i)(k)∝p(z(k)|x(i)(k), θ(i)), where p(z(k)|x(i)(k), θ(i)) is the conditional probability density function of the observed value z(k) given the particle's state and parameters, calculated according to the physical observation equation and the observation noise distribution. Then, the weights are normalized so that the sum of the weights of all particles is 1. To eliminate particle degeneracy, when the number of effective particles is lower than a preset threshold (e.g., N / 2), resampling is performed. Particles are selected with replacement according to their weights, with higher-weighted particles having a higher probability of being selected. After resampling, a new particle set is obtained, and the weight of each particle is reset to 1 / N. After the above processing, in each time window unit k, we can obtain the posterior probability distribution sample set {θ(i)(k), i=1, ..., N} of the physical model parameter vector θ and the posterior probability distribution sample set {x(i)(k), i=1, ..., N} of the component degradation state variable x.

[0042] Step S134: Extract the sample mean from the posterior probability distribution sample set of the physical model parameter vector corresponding to each time window unit as the estimated value of the physical model parameter vector point for that time window unit. Arrange the estimated values ​​of the physical model parameter vector points corresponding to all time window units in the continuous monitoring period in chronological order to generate the model parameter evolution sequence of each component physical degradation sub-model.

[0043] In this embodiment, for each time window unit k, the posterior probability distribution sample set of the physical model parameter vector θ contains N samples θ(i)(k). The sample mean is calculated by averaging each parameter component separately. For example, the point estimate of the m-th component of θ is θm(k) = (1 / N)Σ(i=1 to N)θm(i)(k), where θm(i)(k) is the m-th parameter component of the i-th sample. The point estimates of all parameter components are combined to obtain the point estimate of the physical model parameter vector θ(k) for the time window unit k. The above processing is performed on all 25920 time window units within the continuous monitoring period to obtain 25920 point estimates of the physical model parameter vector. After being arranged in chronological order, a model parameter evolution sequence of the physical degradation sub-model of the component is formed. This model parameter evolution sequence reflects the changes of the model parameters over time.

[0044] Step S135: Extract the sample variance of each state component from the posterior probability distribution sample set of the component degradation state variable corresponding to each time window unit, and use it as the component degradation state estimation uncertainty measurement parameter for each state component of the time window unit. Arrange the component degradation state estimation uncertainty measurement parameters of each state component corresponding to all time window units in the continuous monitoring period in chronological order to generate the state estimation uncertainty evolution sequence of each state component of each component physical degradation sub-model.

[0045] In this embodiment, the component degradation state variable x contains multiple state components, such as crack length and crack propagation rate. For each time window unit k and each state component n, N sample values ​​xn(i)(k) of that state component are extracted from the posterior probability distribution sample set {x(i)(k), i=1, ..., N} of the component degradation state variable. The sample variance is calculated as σn²(k)=(1 / (N-1))Σ(i=1toN)(xn(i)(k)-xn_mean(k))², where xn_mean(k) is the sample mean of that state component in time window unit k, i.e., xn_mean(k)=(1 / N)Σ(i=1toN)xn(i)(k). The calculated sample variance σn²(k) is used as the uncertainty metric parameter for estimating the component degradation state of that state component n in time window unit k. For each state component n, the σn²(k) of all time window units within the continuous monitoring period is arranged in chronological order to generate the state estimation uncertainty evolution sequence of that state component. This state estimation uncertainty evolution sequence reflects the change of state estimation uncertainty over time.

[0046] Step S136: Based on the evolution sequence of model parameters of each component's physical degradation sub-model, calculate the absolute value of the first difference of the evolution sequence for each model parameter component. Then, calculate the sum of the absolute values ​​of the first difference of each model parameter component over the entire continuous monitoring period. After normalizing the sum of the absolute values ​​of the first difference of all model parameter components, perform a weighted summation to obtain the index of the severity of change of the component's physical degradation sub-model.

[0047] In this embodiment, the model parameter evolution sequence contains 25920 physical model parameter vector point estimates θ(1), θ(2), ..., θ(25920), each θ(k) has M parameter components. For the m-th parameter component, its evolution sequence is θm(1), θm(2), ..., θm(25920). The absolute value of the first difference is calculated as |θm(k) - θm(k-1)|, where k ranges from 2 to 25920. For this parameter component, the sum of the absolute values ​​of the first difference over the entire continuous monitoring period is S_m = Σ(k = 2 to 25920) |θm(k) - θm(k-1)|. Then, S_m of all M parameter components is normalized using the method S_m' = S_m / max(S_1, S_2, ..., S_M), so that the normalized S_m' is within the range [0, 1]. Next, a weight ωm is assigned to each parameter component. The weights are pre-set based on the degree of influence of the parameter on the model performance. For example, parameters that have a greater impact on the model output are given a larger weight. The sum of all weights is 1. Finally, the drastic change index V = Σ(m=1 to M)ωm*S_m' is defined. The larger the V value, the more drastic the change in the model parameters.

[0048] Step S137: Based on the state estimation uncertainty evolution sequence of each state component in the physical degradation sub-model of each component, calculate the average uncertainty value of each state component during the continuous monitoring period, and then normalize and weight the average uncertainty values ​​of all state components to obtain the average state estimation uncertainty value of the physical degradation sub-model of the component.

[0049] In this embodiment, the state estimation uncertainty evolution sequence for each state component n is σn²(1), σn²(2), ..., σn²(25920). The average uncertainty value is calculated as Un = (1 / 25920)Σ(k=1 to 25920)σn²(k). Un for all state components is normalized as Un' = Un / max(U_1, U_2, ..., U_P), where P is the number of state components, such that Un' is in the range [0, 1]. A weight νn is assigned to each state component, and the weights are set according to the importance of the state component, with the sum of all weights being 1. The average state estimation uncertainty value U = Σ(n=1 to P)νn * Un', where a smaller U value indicates a smaller uncertainty in the state estimation.

[0050] Step S138: Based on the residual sequence between the candidate feature time series set corresponding to each component physical degradation sub-model and the reconstructed feature time series set obtained by reconstructing the state estimate of the component physical degradation sub-model through the physical observation equation, calculate the average sum of squared fitting residuals for each component physical degradation sub-model.

[0051] In this embodiment, for each time window unit k, the mean x_mean(k) of the posterior probability distribution sample set of the component degradation state variables is used as the state estimate and substituted into the physical observation equation z_hat(k)=h(x_mean(k), θ(k)) to obtain the reconstructed feature time series value z_hat(k). The actual observation value of the candidate feature time series set in time window unit k is z(k), then the residual sequence e(k)=z(k)-z_hat(k), where e(k) is a (T+U) dimensional vector. For each time window unit k, the sum of squares of the residual vector SSE(k)=e(k)^T*e(k) is calculated, where ^T denotes transpose. Then, the average fitted residual sum of squares SSE_avg=(1 / 25920)Σ(k=1to25920)SSE(k), the smaller the SSE_avg value, the better the model fits the observed data.

[0052] Step S139: After normalizing the change severity index, the average state estimation uncertainty value, and the average fitting residual square sum, the weighted sum is obtained to obtain the comprehensive inverse index value of each component physical degradation sub-model. The comprehensive inverse index value is then subjected to reciprocal transformation to generate the matching confidence level between each component physical degradation sub-model and the actual degradation process of the target mechanical component. The larger the matching confidence level value, the more accurately the component physical degradation sub-model can describe the current actual degradation process.

[0053] In this embodiment, the following steps are taken: First, the drastic change index V, the average state estimation uncertainty U, and the average sum of squared residuals SSE_avg are normalized. For V, after normalization, V_norm = V / max(V_all), where max(V_all) is the maximum value of V among all component physical degradation sub-models; for U, U_norm = U / max(U_all); and for SSE_avg, SSE_norm = SSE_avg / max(SSE_all). Then, weights α, β, and γ are set for the three indices, with α + β + γ = 1. The weights are determined based on the influence of each index on the model's matching degree. The comprehensive inverse index value C = α * V_norm + β * U_norm + γ * SSE_norm. The matching confidence R = 1 / (1 + C). Thus, the smaller the C value (i.e., when V, U, and SSE_avg are all small), the larger the R value, indicating a higher degree of matching between the component physical degradation sub-model and the actual degradation process of the target mechanical component.

[0054] Step S140: Select the component physical degradation sub-model with the highest matching confidence as the dominant degradation model, and perform deep fusion processing on the model parameter evolution sequence corresponding to the dominant degradation model and the candidate feature time series set corresponding to the dominant degradation model to generate component health status fusion characterization parameters that integrate physical laws and data features.

[0055] Step S141: Compare the matching confidence values ​​of all component physical degradation sub-models, determine the component physical degradation sub-model with the highest matching confidence value as the dominant degradation model, and record the unique identifier of the dominant degradation model in the physical degradation mechanism model library.

[0056] In this embodiment, it is assumed that there are multiple component physical degradation sub-models, such as fatigue cracking, lubrication failure, bearing wear, etc., and each sub-model has a corresponding matching confidence level R1, R2, ..., RQ (Q is the number of sub-models). By comparing the magnitude of these R values, the sub-model corresponding to the largest R value is found. For example, if R3 is the largest, then this sub-model is the sub-model of the fatigue cracking failure mode, which is determined as the dominant degradation model, and its unique identifier in the physical degradation mechanism model library is recorded, such as "Model_Fatigue_Crack_001".

[0057] Step S142: Extract the estimated value of the model parameter vector point in the last time window unit of the model parameter evolution sequence corresponding to the dominant degradation model as the dominant model parameter value at the current moment.

[0058] In this embodiment, the model parameter evolution sequence of the dominant degradation model includes parameter vector point estimates for 25,920 time window units. The last time window unit corresponds to the end of the continuous monitoring period, i.e., the 25,920th time window unit. The model parameter vector point estimate θ(25920) of this time window unit is extracted as the dominant model parameter value at the current moment, denoted as θ_current. This dominant model parameter value contains the estimated values ​​of each parameter of the model at the current moment, reflecting the current physical degradation characteristics.

[0059] Step S143: Extract the energy values ​​of all candidate feature time series in the last time window unit of the candidate feature time series set corresponding to the dominant degradation model, and form the multi-source feature observation vector at the current moment.

[0060] In this embodiment, the candidate feature time series set corresponding to the dominant degradation model contains T+U time series, each with 25920 energy values. The energy value of each time series at the 25920th time window unit is extracted, and these energy values ​​are combined into a (T+U)-dimensional vector, which is the multi-source feature observation vector z_current at the current moment. This multi-source feature observation vector integrates the current energy features of multiple frequency bands of vibration and acoustic emission.

[0061] Step S144: Obtain the physical state transition equation preset by the dominant degradation model, substitute the current dominant model parameter value into the physical state transition equation, and calculate the prior prediction value of the component degradation state for the next time window unit.

[0062] In this embodiment, the physical state transition equation of the dominant degradation model is x(k+1)=f(x(k), θ)+w(k). The current time is the 25920th time window unit, and the corresponding component degradation state estimate is x_mean(25920) (the mean extracted from the state estimate sample set). Substituting θ_current into the state transition equation, the prior prediction value of the component degradation state for the next time window unit (the 25921st time window) is calculated as x_prior=f(x_mean(25920), θ_current). The process noise w(k) is not considered in this prior prediction and will be reserved for subsequent correction.

[0063] Step S145: Obtain the physical observation equation pre-set by the dominant degradation model, substitute the prior prediction value of the component degradation state of the next time window unit into the physical observation equation, and calculate the multi-source feature observation prediction vector of the next time window unit.

[0064] In this embodiment, the physical observation equation is z(k) = h(x(k), θ) + v(k). Substituting x_prior into this equation, the multi-source feature observation prediction vector z_predict = h(x_prior, θ_current) for the next time window unit is calculated. The observation noise v(k) is also temporarily ignored.

[0065] Step S146: Calculate the difference between the multi-source feature observation vector at the current time and the multi-source feature observation prediction vector to obtain the observation residual vector at the current time.

[0066] In this embodiment, the observation residual vector e = z_current - z_predict, where z_current is the actual observation vector at the current time, and z_predict is the observation prediction vector obtained based on prior prediction. e reflects the difference between the actual observation and the model prediction.

[0067] Step S147: Input the observation residual vector at the current moment into a preset Kalman gain matrix for weighted correction processing. The Kalman gain matrix is ​​calculated in real time based on the state estimation covariance matrix and the observation noise covariance matrix of the dominant degradation model, and generates the corrected posterior estimate of the component degradation state.

[0068] In this embodiment, the Kalman gain matrix K is calculated as K = P_prior * H^T * (H * P_prior * H^T + R)^(-1), where P_prior is the covariance matrix of the prior prediction of the component degradation state, H is the Jacobian matrix of the observation equation with respect to the state variables, and R is the observation noise covariance matrix (a preset parameter). After calculating K, the corrected posterior estimate of the component degradation state is x_post = x_prior + K * e. This process uses the observation residuals to correct the prior predicted state, improving the accuracy of the state estimation.

[0069] Step S148: Align the corrected posterior estimate of the component degradation state with the state variable dimension of the dominant degradation model, and extract the state components directly related to the component health as preliminary health state fusion representation parameters.

[0070] In this embodiment, the state variable vector x of the dominant degradation model may contain multiple components, such as crack length, crack propagation rate, and temperature. Among them, crack length is a state component directly related to the component's health. The components of the corrected posterior estimate of the component's degradation state x_post are aligned with the definition of the state variable dimensions to ensure that each component corresponds to the correct physical meaning. Then, the value of the crack length component is extracted as a preliminary health state fusion characterization parameter, denoted as H_preliminary.

[0071] Step S149: Perform time-series smoothing filtering on the preliminary health status fusion characterization parameters, and use the sliding window averaging method to eliminate short-term fluctuations caused by observation noise to generate the final component health status fusion characterization parameters.

[0072] In this embodiment, the sliding window length is set to W time window units, for example, W=5. The preliminary health state fusion characterization parameter H_preliminary and the preliminary health state fusion characterization parameters of the previous W-1 time window units (if any) are combined into a window data sequence. The average value of this sequence is calculated as the final component health state fusion characterization parameter H_final at the current time. For example, if the current time window unit is the 25920th time window unit, and the window data sequence is H_preliminary(25920-W+1), ..., H_preliminary(25920), then H_final=(1 / W)Σ(i=25920-W+1to25920)H_preliminary(i). Through the above smoothing process, short-term noise interference can be effectively eliminated, making the health state characterization parameters more stable.

[0073] Step S1410: Associate and store the component health status fusion characterization parameters with the corresponding monitoring timestamp to form the component health status fusion characterization parameters of the target mechanical component at the end of the continuous monitoring period.

[0074] In this embodiment, the monitoring timestamp at the end of the continuous monitoring period is 24:00 on the 30th day. This timestamp is associated with the final component health status fusion characterization parameter H_final and stored in the local database of the edge computing module. The data record format is (timestamp, H_final) so that it can be called later when predicting the remaining effective lifespan.

[0075] Step S150: Perform remaining effective life prediction processing on the component health status fusion characterization parameters to generate the remaining effective life prediction result of the target mechanical component.

[0076] Step S151: Obtain the failure threshold definition corresponding to the dominant degradation model. The failure threshold definition includes the critical value of the component health status fusion characterization parameter when it reaches the complete failure state, and the warning value of the component health status fusion characterization parameter when it enters the accelerated degradation stage.

[0077] In this embodiment, the dominant degradation model is a sub-model of the fatigue crack failure mode, and the definition of its failure threshold is preset according to the design standards and safety specifications of the drive shaft of commercial vehicles. Among them, the critical value D_critical refers to the value that when the component health state fusion characterization parameter (crack length) reaches this value, the drive shaft will completely fail and cannot work normally. For example, D_critical = a certain length value; the warning value D_warning is a threshold of the crack length. When the crack length reaches this value, it indicates that the drive shaft enters the accelerated degradation stage and it is necessary to start paying attention and predicting the remaining life. For example, D_warning = another length value, and D_warning < D_critical.

[0078] Step S152: Compare the component health state fusion characterization parameter with the warning value in the definition of the failure threshold. If the component health state fusion characterization parameter is lower than the warning value, it is determined that the target mechanical component has entered the accelerated degradation stage, and the remaining useful life prediction process is triggered.

[0079] In this embodiment, the current component health state fusion characterization parameter is H_final, and H_final is compared with D_warning. If H_final >= D_warning, it is determined that the target mechanical component (the drive shaft of commercial vehicles) has entered the accelerated degradation stage, and at this time, the remaining useful life prediction process is triggered; if H_final < D_warning, it means that the component is still in the normal degradation stage, and the remaining useful life prediction is not carried out temporarily, and it will be judged again in the next monitoring cycle.

[0080] Step S153: Extract the point estimates of the model parameter vectors of the nearest preset number of time window units in the model parameter evolution sequence corresponding to the dominant degradation model to form a recent model parameter change trajectory segment.

[0081] In this embodiment, the preset number is set to V time window units, for example, V = 100. Extract the point estimates of the model parameter vectors θ(25920 - V + 1),..., θ(2592) of the last V time window units (that is, from the (25920 - V + 1)th to the 25920th) from the model parameter evolution sequence of the dominant degradation model. These parameter vectors form a recent model parameter change trajectory segment, which reflects the change trend of the model parameters in the recent period.

[0082] Step S154: Perform trend extrapolation fitting on the recent model parameter change trajectory segment, and use the non-linear least squares method to fit the recent model parameter change trajectory segment to a preset exponential decay function or power-law decay function to obtain the future evolution function of the fitted model parameters.

[0083] In this embodiment, it is assumed that the changing trend of the model parameters conforms to an exponential decay function, in the form θ(t) = a*exp(-b*t) + c, where t is the index of the time window unit, and a, b, and c are the parameters to be fitted. A nonlinear least squares method is used, with the parameter values ​​in the recent model parameter change trajectory segment as observed data, and the parameters a, b, and c are estimated by minimizing the sum of squared residuals. Specifically, the objective function is defined as S = Σ(i=1 to V)(θ_obs(i) - (a*exp(-b*t_i) + c))², where θ_obs(i) is the parameter value of the i-th time window unit in the trajectory segment, and t_i is the corresponding time window unit index. The values ​​of a, b, and c are adjusted using an iterative optimization algorithm (such as the Levenberg-Marquardt algorithm) to minimize S, resulting in the fitted model parameter future evolution function θ_future(t) = a_hat*exp(-b_hat*t) + c_hat, where a_hat, b_hat, and c_hat are the estimated parameters. If the exponential decay function does not fit well (e.g., the sum of squared residuals is too large), then the power-law decay function θ(t)=d*t^(-e)+f is used for fitting, and the nonlinear least squares method is used to estimate the parameters d, e, and f.

[0084] Step S155: Substitute the fitted model parameter future evolution function into the physical state transition equation of the dominant degradation model, and use the component health state fusion characterization parameter as the initial state to iteratively calculate the predicted value of the component health state fusion characterization parameter corresponding to each future time window unit until the predicted value of the component health state fusion characterization parameter is lower than the critical value in the definition of the failure threshold for the first time.

[0085] In this embodiment, the initial state is x0 = H_final, and the initial time window unit is 25920. For each future time window unit t (starting from 25921), the model parameter value θ(t) for that time window unit is first obtained according to the model parameter future evolution function θ_future(t). Then, x(t-1) and θ(t) are substituted into the physical state transition equation x(t) = f(x(t-1), θ(t)) to calculate x(t), which is the predicted value of the component health state fusion characterization parameter for that time window unit. This process is repeated continuously to calculate x(25921), x(25922), ..., until a certain time window unit t_fail, such that x(t_fail) <= D_critical. At this point, the iteration stops, and the value of t_fail is recorded.

[0086] Step S156: Record the number of time window units from the current moment to the moment when the value first falls below the critical value, and multiply the number of time window units by the actual time length corresponding to each time window unit to obtain the preliminary prediction value of the remaining effective lifetime based on the physical model extrapolation.

[0087] In this embodiment, the time window unit corresponding to the current moment is 25920, and the time window unit corresponding to the moment when the value first falls below the critical value is t_fail. The number of time window units experienced is N_window = t_fail - 25920. The actual time length corresponding to each time window unit is 10 minutes, i.e., 10 / 60 = 1 / 6 hour. Therefore, the preliminary prediction value of the remaining effective lifetime based on the physical model extrapolation is L_phy = N_window * (1 / 6) hours, which is converted into days (if necessary), for example, L_phy_days = L_phy / 24.

[0088] Step S157: Obtain the evolution trajectory of component health status fusion characterization parameters and the corresponding actual failure time points of multiple historical components of the same type as the dominant degradation model within their respective life cycles, and construct a similar historical trajectory database.

[0089] In this embodiment, the dominant degradation model is fatigue crack failure mode. Therefore, historical data of all commercial vehicle driveshafts that failed due to fatigue cracks are retrieved from the database. The data for each historical component includes the evolution trajectory of its component health status fusion characterization parameters throughout its entire lifespan (i.e., the crack length sequence changing over time) and the actual failure time (i.e., the time when the crack length reaches D_critical). This data is then organized and stored to form a similar historical trajectory database. For example, the database contains records of M historical components, each record containing trajectory data and the failure time.

[0090] Step S158: Select a preset number of historical trajectory segments from the similar historical trajectory database that have the highest similarity to the shape of the evolution trajectory of the fusion characterization parameter of the current health status of the target mechanical component. Calculate the time length that each selected historical trajectory segment takes from its current similar position to the actual failure time point, and use it as the reference remaining life value corresponding to each historical trajectory segment.

[0091] In this embodiment, the preset quantity is set to K, for example, K=5. The evolution trajectory of the current component health status fusion characterization parameter of the target mechanical component refers to the H_final sequence from the start of monitoring to the current time (25920 time window units). For each historical component in the similar historical trajectory database, a segment of its evolution trajectory with the same length as the target trajectory (i.e., 25920 time window units) is extracted, and the shape similarity between the segment and the target trajectory is calculated. The shape similarity is calculated using the Dynamic Time Warping (DTW) algorithm, which can perform flexible matching on the time axis and calculate the similarity distance between two time series. The smaller the distance, the higher the similarity. After calculating the DTW distance for all historical trajectory segments, the K historical trajectory segments with the smallest distance are selected. For each selected historical trajectory segment, its position in the entire life cycle of the historical component (i.e., the current similar position) is found, and then the time length from that position to the actual failure time of the historical component is calculated as the reference remaining life value L_ref(i), i=1,...,K.

[0092] Step S159: Perform a weighted average of multiple reference remaining lifetime values. The weights are determined based on the similarity between each historical trajectory segment and the current trajectory to obtain a data-driven reference prediction value for remaining effective lifetime.

[0093] In this embodiment, the weight w(i) of each reference remaining lifetime value L_ref(i) is inversely proportional to the DTW distance d(i) of that historical trajectory segment, i.e., w(i) = 1 / d(i) / Σ(j=1toK)(1 / d(j)), which makes the historical trajectory segments with higher similarity (smaller distance) have greater weights. Then, the data-driven remaining effective lifetime reference prediction value L_data = Σ(i=1toK)w(i)*L_ref(i).

[0094] Step S1510: The preliminary prediction of remaining effective lifetime based on physical model extrapolation and the reference prediction of remaining effective lifetime based on data-driven model are fused together, and the weighted combination of the two is calculated using the Bayesian model averaging method to generate the final prediction result of remaining effective lifetime.

[0095] In this embodiment, the Bayesian model averaging method treats L_phy and L_data as the results of two different prediction models. First, a prior probability is assigned to each model, assuming that the prior probability of both models is 0.5. Then, the likelihood of each model is calculated based on historical prediction error data. The likelihood reflects the accuracy of the model's prediction; the smaller the prediction error, the higher the likelihood. For example, the likelihood P_phy of the physical model extrapolation method and the likelihood P_data of the data-driven method are calculated through cross-validation. Next, the posterior probability of each model is calculated: the posterior probability of the physical model is P(phy|L_phy, L_data) = P_phy*0.5 / (P_phy*0.5+P_data*0.5), and the posterior probability of the data-driven model is P(data|L_phy, L_data) = 1-P(phy|L_phy, L_data). Finally, the final remaining effective lifetime prediction result L_final = P(phy|L_phy, L_data)*L_phy + P(data|L_phy, L_data)*L_data.

[0096] Step S1511: Output the final remaining effective lifetime prediction result and simultaneously output the confidence interval range corresponding to the prediction result. The confidence interval range is determined comprehensively based on the uncertainty of the model parameters during the physical model extrapolation process and the dispersion of the reference lifetime values ​​in the similar historical trajectory database.

[0097] In this embodiment, the confidence interval range is determined as follows: For the physical model extrapolation part, the parameter estimation uncertainty of the future evolution function of the model parameters is calculated. Multiple parameter samples are generated through Monte Carlo simulation, and multiple remaining lifetime prediction samples are obtained by substituting them into the state transition equation. The standard deviation σ_phy of these samples is calculated. For the data-driven part, the standard deviation σ_data of K reference remaining lifetime values ​​is calculated. The standard deviation of the comprehensive confidence interval σ_final = sqrt((P(phy)*σ_phy)^2 + (P(data)*σ_data)^2). Then, a 95% confidence interval is taken, i.e., L_final ± 1.96*σ_final. The final remaining effective lifetime prediction result L_final and the corresponding confidence interval range (L_final - 1.96*σ_final, L_final + 1.96*σ_final) are output to the user or the host system.

[0098] Step S210: Obtain the real-time operating condition parameter sequence of the target mechanical component, wherein the real-time operating condition parameter sequence includes component rotation speed parameters, component load parameters, and component lubrication status parameters that are synchronously collected with the continuous monitoring period.

[0099] In this embodiment, the target mechanical component is a commercial vehicle driveshaft. Real-time operating parameters are collected by sensors installed at relevant locations on the driveshaft. Component rotation speed parameters are collected by a rotation speed sensor, with a sampling frequency identical to the original monitoring data stream, once per hour, measured in revolutions per minute (rpm). Component load parameters are collected by a torque sensor, measured in Newton-meters (Nm). Component lubrication status parameters are collected by a lubricating oil quality sensor, which detects indicators such as viscosity and water content of the lubricating oil. These indicators are combined into a lubrication status parameter between 0 and 1, where 1 indicates good lubrication and 0 indicates lubrication failure. These parameters are collected synchronously with the continuous monitoring period, forming a real-time operating parameter sequence. Each time window unit (10 minutes) corresponds to one operating parameter record. Therefore, the real-time operating parameter sequence for the entire continuous monitoring period contains 25,920 data points, each containing three parameter values: rotation speed, load, and lubrication status.

[0100] Step S220: Perform condition segmentation identification processing on the real-time operating condition parameter sequence, and divide the continuous monitoring period into multiple different stable operating condition intervals and transient operating condition intervals according to the change patterns of the component speed parameters, component load parameters and component lubrication state parameters.

[0101] In this embodiment, the operating condition segmentation identification process employs a density-based clustering algorithm (DBSCAN). First, each data point in the real-time operating condition parameter sequence is considered a three-dimensional sample (speed, load, lubrication state). The neighborhood radius and minimum sample number parameters of the clustering algorithm are set, and the sample points are clustered into different clusters, each representing a stable operating condition. Sample points not in any cluster are identified as transient operating condition points. Then, based on the temporal order of the sample points, the time window units corresponding to consecutive sample points belonging to the same cluster are merged into a stable operating condition interval, and the time window units corresponding to consecutive transient operating condition points are merged into a transient operating condition interval. For example, high-speed, high-load stable operating condition intervals, low-speed, low-load stable operating condition intervals, starting transient operating condition intervals, braking transient operating condition intervals, etc., may be defined.

[0102] Step S230: For each stable operating condition interval, extract the rate of change of the component health status fusion characterization parameter within the stable operating condition interval, and establish the mapping relationship between the rate of change and the average component rotation speed parameter, average component load parameter and average component lubrication status parameter within the stable operating condition interval, and generate a condition-sensitive degradation rate lookup table.

[0103] In this embodiment, for each stable operating condition interval, the rate of change of the component health status fusion characterization parameters within that interval is first calculated. The rate of change is calculated as (H_final at the end of the interval - H_final at the beginning of the interval) / the number of time window units for the interval's duration. Then, the average component rotational speed parameter (average of rotational speed across all time window units), the average component load parameter (average of load across all time window units), and the average component lubrication status parameter (average of lubrication status across all time window units) within that stable operating condition interval are calculated. These average operating condition parameters and their corresponding degradation rates are used as a sample, and a mapping relationship is established using methods such as multiple linear regression or neural networks, i.e., degradation rate = f(average rotational speed, average load, average lubrication status). The average operating condition parameters and their corresponding degradation rates under different stable operating conditions are stored to generate a condition-sensitive degradation rate lookup table. This table can be used to quickly find the corresponding degradation rate based on the current operating condition parameters.

[0104] Step S240: For each transient operating condition interval, analyze the instantaneous jump amplitude of the component health status fusion characterization parameter within the transient operating condition interval, and correlate the instantaneous jump amplitude with the change in operating condition parameters at the start and end times of the transient operating condition interval to generate a transient operating condition impact coefficient.

[0105] In this embodiment, for each transient operating condition interval, the instantaneous jump amplitude of the component health status fusion characterization parameter within the interval is calculated, i.e., the difference between the maximum and minimum values ​​of H_final within the interval. Simultaneously, the changes in operating parameters at the start and end times of the transient operating condition interval are calculated, including the speed change (end speed - start speed), load change (end load - start load), and lubrication state change (end lubrication state - start lubrication state). Through statistical analysis, a correlation is established between the instantaneous jump amplitude and these operating parameter changes; for example, the larger the change, the larger the instantaneous jump amplitude. This correlation is quantified into a transient operating condition impact coefficient; for example, a larger coefficient value indicates a greater impact of the transient operating condition on component degradation.

[0106] Step S250: When a new raw monitoring data stream of component operating status is subsequently received, the corresponding real-time operating condition parameter sequence is synchronously acquired. Based on the operating condition type in the real-time operating condition parameter sequence, the corresponding operating condition sensitive degradation rate lookup table entry and transient operating condition impact coefficient are found. The remaining effective life prediction result obtained based on the fusion of physical model and data-driven approach is dynamically corrected.

[0107] In this embodiment, when new monitoring data is received, the operating condition type is first identified for the new real-time operating condition parameter sequence to determine whether the current operating condition is stable or transient, and which specific stable operating condition it belongs to. If it is a stable operating condition, the corresponding degradation rate is looked up in the operating condition sensitive degradation rate lookup table and compared with the degradation rate in the remaining effective life prediction result obtained based on the fusion of physical model and data-driven approach. The remaining life prediction result is then corrected based on the difference. If it is a transient operating condition, the remaining life prediction result is reduced or adjusted accordingly based on the transient operating condition impact coefficient to reflect the additional consumption of component lifespan caused by the transient operating condition.

[0108] Step S260: Compare the corrected remaining effective life prediction result with the uncorrected prediction result, generate the operating condition impact correction amount, and output the operating condition impact correction amount as additional information along with the remaining effective life prediction result.

[0109] In this embodiment, the operating condition impact correction amount = corrected remaining useful life prediction result - uncorrected prediction result. If the correction amount is negative, it indicates that the operating condition factor leads to a reduction in remaining useful life; if it is positive, it indicates that the operating condition factor has a positive impact on the remaining useful life (which is relatively rare). The corrected remaining useful life prediction result, the uncorrected result, and the operating condition impact correction amount are output together, so that users can understand the degree of influence of the operating condition factor on the prediction result.

[0110] Step S310: Obtain the design drawing information and material composition test report of the target mechanical component; extract the key geometric dimension parameters and key mating surface roughness parameters of the target mechanical component from the design drawing information; and extract the chemical composition mass percentage parameters and material heat treatment process parameters of the material used in the target mechanical component from the material composition test report.

[0111] In this embodiment, the target mechanical component is a commercial vehicle driveshaft. The design drawings are obtained through CAD files, from which key geometric parameters are extracted, such as the driveshaft's diameter, length, and shoulder fillet radius; and the roughness parameters of key mating surfaces, such as the roughness Ra value of the surface mating with the bearing. The material composition test report is provided by the material supplier, from which the chemical composition mass percentage parameters of the materials used (such as 45# steel) are extracted, such as the mass percentage of carbon, silicon, and manganese; and the material heat treatment process parameters, such as quenching temperature, tempering temperature, and holding time. These parameters reflect the individual manufacturing differences of the target mechanical component.

[0112] Step S320: Correct the stress concentration factor in the physical state transition equation of the dominant degradation model according to the key geometric dimension parameters and roughness parameters, and correct the material fatigue strength coefficient in the physical state transition equation of the dominant degradation model according to the chemical composition mass percentage parameters and material heat treatment process parameters, thereby generating a personalized modified dominant degradation model.

[0113] In this embodiment, the stress concentration factor is an important parameter in the physical state transition equation of the dominant degradation model, and its value is related to key geometric dimensions and surface roughness. Based on the extracted key geometric dimension parameters (such as the shoulder transition fillet radius) and surface roughness parameters (Ra value), the stress concentration factor in the original model is corrected using a stress concentration factor calculation formula (such as based on charts or empirical formulas) to obtain a stress concentration factor that considers individual geometry and surface quality. The material fatigue strength coefficient is closely related to the material composition and heat treatment process. Based on the chemical composition mass percentage parameters (such as carbon content) and heat treatment process parameters (such as quenching temperature), the material fatigue strength coefficient in the original model is corrected using a material fatigue strength database or empirical formulas. Substituting the corrected stress concentration factor and material fatigue strength coefficient into the physical state transition equation of the dominant degradation model, a personalized modified dominant degradation model is obtained.

[0114] Step S330: Replace the original dominant degradation model with the personalized modified dominant degradation model, and repeat the model parameter adaptive identification processing, matching confidence calculation processing, and component health status fusion characterization parameter generation processing to obtain component health status fusion characterization parameters that take into account individual manufacturing differences.

[0115] In this embodiment, the original dominant degradation model is replaced by a personalized modified dominant degradation model. Then, the model parameters are re-estimated according to the model parameter adaptive identification processing method in step S130 to obtain a new model parameter evolution sequence. Next, the matching confidence is recalculated according to the matching confidence calculation method in step S130 (at this time, since the model has been personalized, the matching confidence should be higher). Finally, the component health status fusion characterization parameters considering individual manufacturing differences are generated according to the method in step S140.

[0116] Step S340: Perform remaining effective life prediction processing on the component health status fusion characterization parameters that take into account individual manufacturing differences, and generate individualized remaining effective life prediction results.

[0117] In this embodiment, the component health status fusion characterization parameters, considering individual manufacturing differences, are used as input. Following the remaining effective life prediction processing method in step S150, the prediction is re-performed to obtain an individualized remaining effective life prediction result. This individualized remaining effective life prediction result better reflects the lifespan differences of the target mechanical component due to manufacturing variations, thus improving the accuracy of the prediction.

[0118] Step S410: Obtain the system-level topology information of the mechanical system in which the target mechanical component is located. The system-level topology information includes the identifiers of multiple associated components that have mechanical coupling or motion transmission relationships with the target mechanical component, as well as the connection methods between the associated components.

[0119] In this embodiment, the target mechanical component is a commercial vehicle driveshaft, and the mechanical system it belongs to is a commercial vehicle transmission system. The system-level topology information is obtained through the system design document, including the identifiers of related components that have mechanical coupling or motion transmission relationships with the driveshaft, such as the engine, transmission, drive axle, and wheels; as well as the connection methods between the related components, such as the driveshaft being connected to the transmission via a universal joint, and to the drive axle via another universal joint, etc.

[0120] Step S420: Construct a system-level coupled degenerate graph neural network model based on the system-level topology information. The nodes of the system-level coupled degenerate graph neural network model correspond to each component. The initial features of the nodes include the current component health status fusion representation parameters of the corresponding component, the edges correspond to the connection relationships between components, and the edge weights are preset according to the connection method.

[0121] In this embodiment, the system-level coupled degradation graph neural network model adopts a graph convolutional neural network (GCN) architecture. The nodes of the model correspond to various components in the transmission system, including the driveshaft, engine, and gearbox. The initial feature vector of each node contains the current component health status fusion representation parameters, as well as other relevant state parameters (such as temperature and vibration). Edges correspond to the connection relationships between components; for example, the connection between the driveshaft and the gearbox is represented by an edge. The weights of the edges are pre-set based on factors such as the rigidity of the connection and the magnitude of the transmitted torque; for example, rigid connections have higher weights, while flexible connections have lower weights.

[0122] Step S430: Input the current component health status fusion characterization parameters of the target mechanical component and the current component health status fusion characterization parameters of each associated component into the system-level coupled degradation graph neural network model for information transmission and aggregation processing, and generate the corrected component health status fusion characterization parameters of the target mechanical component under the influence of system-level coupling.

[0123] In this embodiment, the current component health status fusion representation parameters of the target mechanical component (driveshaft) and its associated components are used as the initial feature input to the system-level coupled degradation graph neural network model. The model transmits and aggregates information through graph convolutional layers, and the features of each node are updated in conjunction with the features of its neighboring nodes. For example, the features of the driveshaft node are affected by the features of associated components such as the engine and gearbox. After processing by multiple graph convolutional layers, the final features of the target mechanical component node are output, from which the corrected component health status fusion representation parameters are extracted. These corrected component health status fusion representation parameters take into account the coupling influence of other components in the system on the degradation of the driveshaft.

[0124] Step S440: Re-execute the remaining effective lifetime prediction process based on the corrected component health status fusion characterization parameters to generate the remaining effective lifetime prediction result of the fusion system-level coupling effect.

[0125] In this embodiment, the corrected component health status fusion characterization parameters are used as new inputs, and the remaining effective lifetime prediction is re-predicted according to the remaining effective lifetime prediction processing method in step S150 to obtain the remaining effective lifetime prediction result of the fused system-level coupling effect. This remaining effective lifetime prediction result takes into account the influence of other components in the system, making the prediction more comprehensive and accurate.

[0126] Step S510: Periodically acquire the actual maintenance record data of the target mechanical component. The actual maintenance record data includes the time point of each maintenance operation, the maintenance type code, and the measured values ​​of the component health status fusion characterization parameters of the target mechanical component before and after maintenance.

[0127] In this embodiment, actual maintenance record data is obtained periodically through the commercial vehicle maintenance management system, for example, once a month. The record of each maintenance operation includes the maintenance time (accurate to the minute), maintenance type code (such as "001" indicating bearing replacement, "002" indicating lubrication maintenance, etc.), measured values ​​of component health status fusion characterization parameters before maintenance (obtained by dedicated testing equipment), and measured values ​​after maintenance. The above data is stored in a local database for subsequent analysis.

[0128] Step S520: Calculate the difference between the measured values ​​of the component health status fusion characterization parameters before and after each maintenance operation to obtain the amount of health status recovery brought about by each maintenance operation.

[0129] In this embodiment, the health status recovery amount ΔH = H_after - H_before, where H_after is the measured value of the component health status fusion characterization parameter after repair, and H_before is the measured value before repair. A positive ΔH indicates that the health status has been improved after repair, and the larger the value, the better the improvement effect; a zero or negative ΔH indicates that the repair was ineffective or aggravated the degradation (the above situations are relatively rare and may be caused by improper repair operations).

[0130] Step S530: Perform correlation analysis between the health status recovery amount and the corresponding maintenance type code and the component health status fusion characterization parameter level during maintenance operation to establish an empirical model of maintenance effect under different degradation levels for different maintenance types.

[0131] In this embodiment, the maintenance type code, the component health status fusion parameter level H_before, and the health status recovery amount ΔH for each maintenance operation are used as sample data. An empirical model of maintenance effectiveness is established using statistical analysis or machine learning methods (such as decision trees or neural networks). The input to this empirical model is the maintenance type code and H_before, and the output is the predicted health status recovery amount ΔH_pred. For example, the model can learn the typical value of ΔH_pred for lubrication maintenance type "002" when H_before is within a certain range.

[0132] Step S540: When the component health status fusion characterization parameter of the target mechanical component decreases below the preset maintenance decision threshold, the maintenance effect experience model is invoked to recommend the optimal maintenance type code and the expected health status recovery amount after maintenance based on the current component health status fusion characterization parameter level, and to generate maintenance decision support information containing maintenance suggestions and expected effects.

[0133] In this embodiment, the preset maintenance decision threshold H_decision is set according to the safe operation requirements of the component. When H_final <= H_decision, the maintenance decision support process is triggered. The maintenance effect experience model is invoked, and the current component health status fusion representation parameter level H_final and all possible maintenance type codes are input. The model outputs the expected health status recovery amount ΔH_pred for each maintenance type. The maintenance type code with the largest ΔH_pred is selected as the optimal recommendation, and the expected post-maintenance health status H_after_pred = H_final + ΔH_pred is calculated. The optimal maintenance type code, the expected ΔH_pred, and H_after_pred are integrated into maintenance decision support information and output to the maintenance personnel.

[0134] For example, the method may further include: step S610: packaging the component identifier, continuous monitoring period, remaining effective life prediction result and the evolution trajectory of the corresponding component health status fusion characterization parameters of the target mechanical component to generate a component health prediction report.

[0135] In this embodiment, the component identifier is the unique number of the target mechanical component, such as "Driveshaft_001"; the continuous monitoring period is the start and end time of this monitoring, such as "2023-01-01 00:00:00 to 2023-01-30 24:00:00"; the remaining effective life prediction result is L_final and its confidence interval; the evolution trajectory of the component health status fusion characterization parameter is the change sequence of H_final with time window units. The above information is packaged according to a preset format (such as XML or JSON) to generate a component health prediction report.

[0136] Step S620: Upload the component health prediction report to the cloud health management platform. The cloud health management platform collects health prediction reports of multiple mechanical components of the same type and constructs a big data set of component group health evolution.

[0137] In this embodiment, the edge computing module uploads component health prediction reports to the cloud-based health management platform via a wireless network (such as 4G / 5G). The cloud platform receives health prediction reports from multiple commercial vehicles of the same type of driveshaft (or other components), integrates the data from these reports, and constructs a large dataset of component group health evolution. This large dataset of component group health evolution includes information such as the identification of multiple components, monitoring period, remaining life prediction results, and health status evolution trajectory.

[0138] Step S630: Receive the group health statistical analysis results issued by the cloud health management platform. The group health statistical analysis results include the distribution pattern of the remaining effective life of the same type of components under different working conditions and the statistical characteristics of common failure modes.

[0139] In this embodiment, the cloud-based health management platform performs statistical analysis on the large dataset of component group health evolution, such as analyzing the probability distribution of remaining effective life under high-speed and high-load conditions, and the occurrence ratio of different failure modes (such as fatigue cracks and lubrication failure) in the group. The above analysis results are then sent to the edge computing module for updating the model parameters.

[0140] Step S640: Based on the statistical analysis results of the group health, update and adjust the prior parameter distribution of the component physical degradation sub-model in the physical degradation mechanism model library, and update the matching confidence calculation rules of the dominant degradation model to achieve the synergistic optimization of the individual prediction model of the target mechanical component and the statistical law of the group.

[0141] In this embodiment, based on the statistical analysis results of the population health status, for example, if the actual probability of a certain failure mode is found to be higher than the original prior probability, the prior parameter distribution of the component physical degradation sub-model corresponding to that failure mode is adjusted (e.g., increasing the mean or decreasing the variance). Simultaneously, based on the statistical characteristics of different failure modes in the population data, the weights α, β, and γ in the matching confidence calculation rule are adjusted to make the matching confidence calculation more consistent with reality. Through the above updates and adjustments, the synergistic optimization of the individual prediction model and the population statistical laws is achieved, continuously improving prediction accuracy.

[0142] In one exemplary embodiment, a component health prediction system integrating physical models and data is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2 As shown, the component health prediction system integrating physical models and data includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a component health prediction method integrating physical models and data. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the casing of the component health prediction system that integrates physical models and data, or an external keyboard, touchpad, or mouse, etc.

[0143] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for predicting component health by integrating physical models and data, characterized in that, The method includes: Obtain the raw monitoring data stream of the target mechanical component's operating status during a continuous monitoring period, and obtain the physical degradation mechanism model library corresponding to the target mechanical component. The physical degradation mechanism model library contains multiple component physical degradation sub-models describing different failure modes. The original monitoring data stream of the component's operating status is subjected to multi-scale time-frequency feature decomposition processing to extract the energy distribution evolution features of the target mechanical component at multiple frequency band scales. The energy distribution evolution features are then filtered and matched according to the failure mode feature frequency range corresponding to each of the multiple component physical degradation sub-models to generate a set of candidate feature time series associated with each component physical degradation sub-model. The candidate feature time series set is input into the corresponding component physical degradation sub-model for adaptive identification of model parameters, generating the model parameter evolution sequence of each component physical degradation sub-model in the current monitoring period, and calculating the matching confidence of each component physical degradation sub-model with the actual degradation process of the target mechanical component based on the model parameter evolution sequence; The component physical degradation sub-model with the highest matching confidence is selected as the dominant degradation model, and the model parameter evolution sequence corresponding to the dominant degradation model is deeply fused with the candidate feature time series set corresponding to the dominant degradation model to generate component health status fusion characterization parameters that fuse physical laws and data features. The remaining effective life prediction process is performed on the fusion characterization parameters of the component's health status to generate the remaining effective life prediction result of the target mechanical component.

2. The component health prediction method integrating physical models and data according to claim 1, characterized in that, The process involves multi-scale time-frequency feature decomposition of the raw monitoring data stream of the component's operating status to extract the energy distribution evolution features of the target mechanical component across multiple frequency bands. Furthermore, based on the failure mode feature frequency ranges corresponding to each of the multiple component physical degradation sub-models, the energy distribution evolution features are subjected to feature filtering and matching to generate a set of candidate feature time series associated with each component physical degradation sub-model, including: The raw monitoring data stream of the component's operating status is segmented into multiple continuous and non-overlapping monitoring data time window units according to a preset fixed time window length. Each monitoring data time window unit contains the original vibration signal sequence and the original acoustic emission signal sequence of the target mechanical component within the corresponding time interval. For each monitoring data time window unit, the original vibration signal sequence and the original acoustic emission signal sequence are respectively subjected to signal preprocessing operations. The signal preprocessing operations include using a bandpass filter to filter out environmental noise interference components that are below a first preset frequency threshold and above a second preset frequency threshold from the original vibration signal sequence and the original acoustic emission signal sequence, so as to obtain the filtered vibration signal sequence and the filtered acoustic emission signal sequence corresponding to each monitoring data time window unit. For each monitoring data time window unit, the filtered vibration signal sequence and the filtered acoustic emission signal sequence are subjected to adaptive wavelet packet decomposition layer determination processing. The wavelet packet decomposition layer is dynamically adjusted according to the comparison result between the total energy value of the signal in each monitoring data time window unit and the preset energy threshold, so that the frequency resolution of the lowest layer frequency band after decomposition meets the preset resolution requirement, and the vibration signal wavelet packet decomposition tree structure and the acoustic emission signal wavelet packet decomposition tree structure corresponding to each monitoring data time window unit are generated. Extract the vibration signal wavelet packet reconstruction coefficient sequence corresponding to all leaf nodes from the vibration signal wavelet packet decomposition tree structure corresponding to each monitoring data time window unit, and extract the acoustic emission signal wavelet packet reconstruction coefficient sequence corresponding to all leaf nodes from the acoustic emission signal wavelet packet decomposition tree structure corresponding to each monitoring data time window unit to form the original multi-scale decomposition coefficient set for each monitoring data time window unit. For each vibration signal wavelet packet reconstruction coefficient sequence and each acoustic emission signal wavelet packet reconstruction coefficient sequence in the original multi-scale decomposition coefficient set of each monitoring data time window unit, energy value calculation is performed respectively. The sum of squares of each reconstruction coefficient sequence is calculated as the instantaneous energy feature value under that frequency band scale, and the vibration signal energy distribution vector and acoustic emission signal energy distribution vector corresponding to each monitoring data time window unit are obtained. The vibration signal energy distribution vectors corresponding to all monitoring data time window units within the continuous monitoring period are arranged in chronological order to form a vibration signal energy distribution evolution matrix. The acoustic emission signal energy distribution evolution matrix is ​​also arranged in chronological order to form an acoustic emission signal energy distribution evolution matrix. The rows of the vibration signal energy distribution evolution matrix correspond to different frequency band scales, and the columns correspond to different time window units. The failure mode description information corresponding to each component physical degradation sub-model in the physical degradation mechanism model library is analyzed. The lower limit and upper limit of the characteristic frequency range of each failure mode in the vibration signal, as well as the lower limit and upper limit of the characteristic frequency range in the acoustic emission signal, are extracted from the description information of each failure mode to form the vibration characteristic frequency range and acoustic emission characteristic frequency range corresponding to each component physical degradation sub-model. Based on the vibration characteristic frequency range corresponding to the physical degradation sub-model of each component, energy distribution row vectors corresponding to all frequency band scales whose center frequency falls within the vibration characteristic frequency range are selected from the vibration signal energy distribution evolution matrix. Each selected energy distribution row vector is combined with the corresponding time window unit index to generate a set of vibration source candidate feature time series corresponding to the physical degradation sub-model of the component. Based on the acoustic emission characteristic frequency range corresponding to the physical degradation sub-model of each component, energy distribution row vectors corresponding to all frequency band scales whose center frequency falls within the acoustic emission characteristic frequency range are selected from the acoustic emission signal energy distribution evolution matrix. Each selected energy distribution row vector is combined with the corresponding time window unit index to generate a set of acoustic emission source candidate feature time series corresponding to the physical degradation sub-model of the component. The energy values ​​in the candidate feature time series sets of vibration sources and acoustic emission sources corresponding to each component physical degradation sub-model are normalized respectively. The normalized candidate feature time series sets of vibration sources and acoustic emission sources are then merged to form a complete candidate feature time series set corresponding to the physical degradation sub-model of the component. Each candidate feature time series in the complete candidate feature time series set contains normalized energy evolution information of each time window unit within the continuous monitoring period at the corresponding specific frequency band scale.

3. The component health prediction method integrating physical models and data according to claim 1, characterized in that, The step involves inputting the candidate feature time series set into the corresponding component physical degradation sub-model for adaptive model parameter identification processing, generating a model parameter evolution sequence for each component physical degradation sub-model within the current monitoring period, and calculating the matching confidence level between each component physical degradation sub-model and the actual degradation process of the target mechanical component based on the model parameter evolution sequence, including: Obtain a predefined state space expression for each component physical degradation sub-model. The state space expression includes a physical state transition equation describing the evolution of the component degradation state over time and a physical observation equation describing the mapping relationship between the candidate feature time series and the component degradation state. The physical state transition equation contains a physical model parameter vector to be identified, and the physical observation equation contains preset observation noise statistical characteristic parameters. The candidate feature time series set corresponding to each component physical degradation sub-model is used as the observation input sequence of the component physical degradation sub-model. The prior probability distribution of the physical model parameter vector in the component physical degradation sub-model and the initial probability distribution of the component degradation state variable are initialized. The prior probability distribution is preset according to the statistical characteristics of similar historical components. For each component physical degradation sub-model, the sequential importance sampling particle filtering algorithm is used to perform joint online recursive estimation processing on the physical model parameter vector and the component degradation state variables. When each time window unit arrives, the importance weight of each particle is updated according to the candidate feature time series value corresponding to the current time window unit, and the particle set is resampled to eliminate the particle degradation phenomenon, so as to obtain the posterior probability distribution sample set of the physical model parameter vector and the posterior probability distribution sample set of the component degradation state variables corresponding to each time window unit. The mean of the samples is extracted from the posterior probability distribution sample set of the physical model parameter vector corresponding to each time window unit as the estimated value of the physical model parameter vector point for that time window unit. The estimated values ​​of the physical model parameter vector points corresponding to all time window units in the continuous monitoring period are arranged in chronological order to generate the model parameter evolution sequence of each component physical degradation sub-model. The sample variance of each state component is extracted from the posterior probability distribution sample set of the component degradation state variable corresponding to each time window unit, and used as the component degradation state estimation uncertainty measurement parameter for each state component of the time window unit. The component degradation state estimation uncertainty measurement parameters of each state component corresponding to all time window units in the continuous monitoring period are arranged in chronological order to generate the state estimation uncertainty evolution sequence of each state component of each component physical degradation sub-model. Based on the evolution sequence of model parameters of each component's physical degradation sub-model, the absolute value of the first difference of the evolution sequence of each model parameter component is calculated. Then, the sum of the absolute values ​​of the first difference of each model parameter component over the entire continuous monitoring period is calculated. After normalizing the sum of the absolute values ​​of the first difference of all model parameter components, a weighted sum is obtained to obtain the index of the degree of change of the component's physical degradation sub-model. Based on the state estimation uncertainty evolution sequence of each state component in the physical degradation sub-model of each component, the average uncertainty value of each state component in the continuous monitoring period is calculated. Then, the average uncertainty values ​​of all state components are normalized and weighted to obtain the average state estimation uncertainty value of the physical degradation sub-model of the component. Based on the residual sequence between the candidate feature time series set corresponding to each component physical degradation sub-model and the reconstructed feature time series set obtained by reconstructing the state estimate of the component physical degradation sub-model through the physical observation equation, calculate the average sum of squared fitting residuals for each component physical degradation sub-model; After normalizing the drastic change index, the average state estimation uncertainty value, and the average sum of squared fitted residuals, a weighted sum is obtained to get the comprehensive inverse index value of each component physical degradation sub-model. The comprehensive inverse index value is then subjected to a reciprocal transformation to generate the matching confidence level between each component physical degradation sub-model and the actual degradation process of the target mechanical component. The larger the matching confidence level, the more accurately the component physical degradation sub-model can describe the current actual degradation process.

4. The component health prediction method integrating physical models and data according to claim 1, characterized in that, The component physical degradation sub-model with the highest matching confidence is selected as the dominant degradation model, and the model parameter evolution sequence corresponding to the dominant degradation model is deeply fused with the candidate feature time series set corresponding to the dominant degradation model to generate component health status fusion characterization parameters that integrate physical laws and data features, including: Compare the matching confidence scores of all component physical degradation sub-models, determine the component physical degradation sub-model with the highest matching confidence score as the dominant degradation model, and record the unique identifier of the dominant degradation model in the physical degradation mechanism model library; Extract the model parameter vector point estimate of the last time window unit in the model parameter evolution sequence corresponding to the dominant degradation model as the dominant model parameter value at the current moment; Extract the energy values ​​of all candidate feature time series in the last time window unit from the candidate feature time series set corresponding to the dominant degradation model, and construct the multi-source feature observation vector at the current moment; Obtain the physical state transition equation pre-set by the dominant degradation model, substitute the current dominant model parameter value into the physical state transition equation, and calculate the prior prediction value of the component degradation state for the next time window unit. Obtain the physical observation equation pre-set by the dominant degradation model, substitute the prior prediction value of the component degradation state of the next time window unit into the physical observation equation, and calculate the multi-source feature observation prediction vector of the next time window unit. The difference between the multi-source feature observation vector at the current moment and the multi-source feature observation prediction vector is calculated to obtain the observation residual vector at the current moment. The observation residual vector at the current moment is input into a preset Kalman gain matrix for weighted correction processing. The Kalman gain matrix is ​​calculated in real time based on the state estimation covariance matrix and the observation noise covariance matrix of the dominant degradation model, and a corrected posterior estimate of the component degradation state is generated. Align the corrected posterior estimate of the component degradation state with the state variable dimension of the dominant degradation model, and extract the state components directly related to the component health as preliminary health state fusion representation parameters. The preliminary health status fusion characterization parameters are subjected to time-series smoothing filtering, and the sliding window averaging method is used to eliminate short-term fluctuations caused by observation noise, thereby generating the final component health status fusion characterization parameters. The component health status fusion characterization parameters are associated and stored with the corresponding monitoring timestamps to form the component health status fusion characterization parameters of the target mechanical component at the end of the continuous monitoring period.

5. The component health prediction method integrating physical models and data according to claim 1, characterized in that, The process of performing remaining effective life prediction on the fused characterization parameters of the component's health status to generate the remaining effective life prediction result of the target mechanical component includes: Obtain the failure threshold definition corresponding to the dominant degradation model. The failure threshold definition includes the critical value of the component health status fusion characterization parameter when it reaches the complete failure state, and the warning value of the component health status fusion characterization parameter when it enters the accelerated degradation stage. The component health status fusion characterization parameter is compared with the warning value in the failure threshold definition. If the component health status fusion characterization parameter is lower than the warning value, it is determined that the target mechanical component has entered the accelerated degradation stage, triggering the remaining effective life prediction process. Extract the estimated values ​​of the model parameter vector points of the most recent preset number of time window units in the model parameter evolution sequence corresponding to the dominant degradation model to form a recent model parameter change trajectory segment; The recent model parameter change trajectory segment is subjected to trend extrapolation fitting. The nonlinear least squares method is used to fit the recent model parameter change trajectory segment to a preset exponential decay function or power law decay function to obtain the future evolution function of the fitted model parameters. Substitute the future evolution function of the fitted model parameters into the physical state transition equation of the dominant degradation model, and use the component health status fusion characterization parameter as the initial state to iteratively calculate the predicted value of the component health status fusion characterization parameter corresponding to each future time window unit until the predicted value of the component health status fusion characterization parameter is lower than the critical value in the definition of the failure threshold for the first time. Record the number of time window units from the current moment to the moment when the value first falls below the critical value, and multiply the number of time window units by the actual time length corresponding to each time window unit to obtain the preliminary prediction value of the remaining effective lifetime based on the physical model extrapolation. Acquire the evolution trajectory of component health status fusion characterization parameters and the corresponding actual failure time points of multiple historical components of the same type as the dominant degradation model within their respective life cycles to form a similar historical trajectory database. From the similar historical trajectory database, a preset number of historical trajectory segments with the highest similarity to the shape of the evolution trajectory of the fusion characterization parameter of the current health status of the target mechanical component are selected. The time length experienced by each selected historical trajectory segment from its current similar position to the actual failure time point is calculated as the reference remaining life value corresponding to each historical trajectory segment. Multiple reference remaining lifetime values ​​are weighted and averaged, with the weights determined based on the similarity between each historical trajectory segment and the current trajectory, to obtain a data-driven reference prediction value for remaining effective lifetime. The preliminary remaining useful life prediction based on physical model extrapolation is fused with the reference remaining useful life prediction based on data-driven model, and the weighted combination of the two is calculated using the Bayesian model averaging method to generate the final remaining useful life prediction result. Output the final remaining effective lifetime prediction result, and simultaneously output the confidence interval range corresponding to the prediction result. The confidence interval range is determined comprehensively based on the uncertainty of the model parameters during the physical model extrapolation process and the dispersion of the reference lifetime values ​​in the similar historical trajectory database.

6. The component health prediction method integrating physical models and data according to claim 5, characterized in that, The method further includes: The real-time operating condition parameter sequence of the target mechanical component is obtained, and the real-time operating condition parameter sequence includes component speed parameters, component load parameters and component lubrication status parameters that are collected synchronously with the continuous monitoring period. The real-time operating condition parameter sequence is segmented and identified. Based on the change patterns of the component speed parameters, component load parameters, and component lubrication status parameters, the continuous monitoring period is divided into multiple different stable operating condition intervals and transient operating condition intervals. For each stable operating condition interval, the rate of change of the component health status fusion characterization parameter within the stable operating condition interval is extracted, and a mapping relationship is established between the rate of change and the average component rotation speed parameter, average component load parameter and average component lubrication status parameter within the stable operating condition interval, generating a condition-sensitive degradation rate lookup table. For each transient operating condition interval, the instantaneous jump amplitude of the component health status fusion characterization parameter within the transient operating condition interval is analyzed, and the instantaneous jump amplitude is correlated with the change in operating condition parameters at the start and end times of the transient operating condition interval to generate a transient operating condition impact coefficient. When new raw monitoring data streams of component operating status are received, the corresponding real-time operating condition parameter sequence is obtained synchronously. Based on the operating condition type in the real-time operating condition parameter sequence, the corresponding operating condition sensitive degradation rate lookup table entry and transient operating condition impact coefficient are found. The remaining effective life prediction result obtained based on the fusion of physical model and data-driven approach is dynamically corrected. The revised remaining useful life prediction result is compared with the unrevised prediction result to generate the operating condition impact correction amount, and the operating condition impact correction amount is output as additional information along with the remaining useful life prediction result.

7. The component health prediction method integrating physical models and data according to claim 5, characterized in that, The method further includes: Obtain the design drawings and material composition test reports of the target mechanical component; extract the key geometric dimension parameters and the roughness parameters of the key mating surfaces of the target mechanical component from the design drawings; and extract the chemical composition mass percentage parameters and material heat treatment process parameters of the material used in the target mechanical component from the material composition test reports. The stress concentration factor in the physical state transition equation of the dominant degradation model is corrected based on the key geometric dimension parameters and roughness parameters. The material fatigue strength factor in the physical state transition equation of the dominant degradation model is corrected based on the chemical composition mass percentage parameters and material heat treatment process parameters, thereby generating a personalized modified dominant degradation model. The original dominant degradation model is replaced by the personalized modified dominant degradation model. The adaptive identification of model parameters, the matching confidence calculation, and the generation of component health status fusion characterization parameters are repeated to obtain component health status fusion characterization parameters that take into account individual manufacturing differences. The remaining effective life prediction process is performed on the component health status fusion characterization parameters that take into account individual manufacturing differences, generating individualized remaining effective life prediction results.

8. The component health prediction method integrating physical models and data according to claim 5, characterized in that, The method further includes: Obtain the system-level topology information of the mechanical system in which the target mechanical component is located. The system-level topology information includes the identifiers of multiple associated components that have mechanical coupling or motion transmission relationships with the target mechanical component, as well as the connection methods between each associated component. A system-level coupled degenerate graph neural network model is constructed based on the system-level topology information. The nodes of the system-level coupled degenerate graph neural network model correspond to each component. The initial features of the nodes include the current component health status fusion representation parameters of the corresponding component, the edges correspond to the connection relationships between components, and the weights of the edges are preset according to the connection method. The current component health status fusion characterization parameters of the target mechanical component and the current component health status fusion characterization parameters of each associated component are input into the system-level coupled degradation graph neural network model for information transmission and aggregation processing, thereby generating the corrected component health status fusion characterization parameters of the target mechanical component under the influence of system-level coupling. Based on the corrected component health status fusion characterization parameters, the remaining effective lifetime prediction process is re-executed to generate the remaining effective lifetime prediction result of the fusion system-level coupling effect.

9. The component health prediction method integrating physical models and data according to claim 5, characterized in that, The method further includes: Periodically acquire actual maintenance record data of the target mechanical component. The actual maintenance record data includes the time point of each maintenance operation, maintenance type code, and measured values ​​of the component health status fusion characterization parameters of the target mechanical component before and after maintenance. The difference between the measured values ​​of the component health status fusion characterization parameters before and after each maintenance operation is calculated to obtain the amount of health status recovery brought about by each maintenance operation. The correlation analysis is performed between the health status recovery amount and the corresponding maintenance type code and the component health status fusion characterization parameter level during maintenance operation to establish an empirical model of maintenance effect under different degrees of degradation for different maintenance types. When the component health status fusion characterization parameter of the target mechanical component decreases below the preset maintenance decision threshold, the maintenance effect experience model is invoked to recommend the optimal maintenance type code and the expected health status recovery amount after maintenance based on the current component health status fusion characterization parameter level, and to generate maintenance decision support information containing maintenance suggestions and expected effects.

10. A component health prediction system integrating physical models and data, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the component health prediction method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images