A structure life assessment method and system based on digital twinning

Abstract

The application discloses a kind of based on digital twinning structure life evaluation method and system, it is related to digital twinning technical field, including: preprocessing monitoring and working condition data, generate standardization operating data;Parallel digital twinning reduction model is constructed, and multiple failure twinborn damage state is generated;Real-time damage feature vector is generated by executing cross-time scale normalization alignment;Smooth probability weight vector is generated by probability neural network and time series smoothing;Analysis failure mode competition relationship, generate failure mode competition state;Three-parameter Weibull parameters are corrected, and dynamic reliability is calculated;Multiple source probability likelihood is constructed and inferred, and update parameters are generated;Residual life and reliability are calculated, and evaluation and early warning result are generated.The application realizes multiple failure mode damage online evaluation, dynamic reliability calculation and residual life prediction, improves structure life evaluation and early warning precision.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a method and system for assessing the lifespan of structural components based on digital twins. Background Technology

[0002] Key structural components of large and complex equipment are subjected to alternating multi-physics fields of temperature, pressure, vibration, and contact loads during long-term service. Thermomechanical fatigue, high-cycle fatigue, and localized wear exhibit parallel evolution, nonlinear accumulation, and dynamic competition characteristics. Existing methods for predicting the life and reliability of structural components mostly establish damage models around a single failure mode and use fixed failure mode weights to describe the multi-mode effects. These methods fail to reflect changes in damage rate caused by changes in operating conditions and the migration of dominant failure modes, and lack the ability to identify the real-time contribution weights during the failure mode transition phase.

[0003] Existing hybrid Weibull lifetime analysis methods largely rely on historical failure samples for static parameter fitting. The location parameters of the three-parameter Weibull distribution lack dynamic physical constraints with the real-time cumulative damage of the structural components, making it difficult for lifetime distribution parameters to evolve synchronously with physical degradation. This leads to discrepancies between reliability calculations and actual damage states. Full-life destructive testing of critical structural components in complex equipment is costly and time-consuming, with a limited number of actual failure samples. The online operation phase primarily generates non-failure operational data, digital twin prediction data, and sensor anomaly data. Existing parameter update methods lack a unified probabilistic fusion mechanism for multi-source lifetime information, making it difficult to utilize operational information continuously generated during periods without physical failure to correct lifetime distribution parameters and failure mode weights. This limits the ability to predict remaining lifetime of structural components online, identify competing failures, and update long-term assessment models.

[0004] Therefore, how to provide a method and system for assessing the lifespan of structural components based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a structural component life assessment method and system based on digital twins. This invention integrates parallel digital twins, probabilistic neural networks, temporal smoothing, three-parameter Weibull analysis of physical constraints, and variational Bayesian updates to conduct online assessments of multi-failure-mode damage, competitive states, dynamic reliability, and remaining life. It has the advantages of accurate failure mode identification, high life prediction accuracy, strong physical consistency, and strong online update capability.

[0006] A method for assessing the lifespan of structural components based on digital twins according to an embodiment of the present invention includes:

[0007] Collect and preprocess monitoring data and operating condition data of structural components to generate standardized operating data;

[0008] Based on standardized operating data, a parallel digital twin reduced-order model for thermomechanical fatigue, high-cycle fatigue, and wear is established to calculate the real-time damage and cumulative damage of each failure mode and generate multi-failure twin damage states.

[0009] Based on the multi-failure twin damage state, cross-timescale normalization alignment is performed, and working condition data is associated to generate a real-time damage feature vector.

[0010] A probabilistic neural network is constructed. The real-time damage feature vector is input into the probabilistic neural network to obtain the initial probability weights, and a smoothed probability weight vector is generated through time-series smoothing.

[0011] Based on the smooth probability weight vector analysis of weight trends, the dominant failure modes and migration trends are determined, and the failure mode competition state is generated.

[0012] Read offline test data from the structural component test database, initialize the three-parameter Weibull parameters of each failure mode based on the offline test data, establish nonlinear physical constraints on the location parameters and cumulative damage, correct the location parameters in real time, and calculate the dynamic reliability using a smooth probability weight vector as a time-varying weight.

[0013] Unfailed operational data and sensor anomaly data are extracted from standardized operational data. Virtual accelerated lifetime extrapolation is performed based on multi-failure twin damage states. Multi-source probability likelihood functions are constructed in conjunction with dynamic reliability. Weibull parameters and mode weights are updated through variational Bayesian inference to generate updated parameters.

[0014] Dynamic reliability and remaining lifetime are calculated based on updated parameters, and lifetime assessment results and failure warning results are generated by combining failure mode competition states.

[0015] Optionally, generating standardized operational data includes:

[0016] Monitoring data is collected from temperature sensors, pressure sensors, vibration acceleration sensors, and oil particle sensors on structural components. Load, speed, and working stage data are read from the equipment operation control terminal to generate raw operating data.

[0017] The original operating data is subjected to anomaly removal, missing data marking and time correction. The monitoring data and operating condition data are associated according to the structural component identification and the time of acquisition to generate corrected operating data.

[0018] The operating data is synchronously corrected and standardized in terms of units according to a unified sampling cycle. The monitoring data and operating condition data are arranged according to the collection time sequence to generate standardized operating data.

[0019] Optionally, the generation of multi-failure twin damage states includes:

[0020] Monitoring data and operating condition data are extracted from standardized operating data, and corresponding twin inputs are divided according to thermomechanical fatigue, high cycle fatigue and wear.

[0021] Based on the twin input, digital twin reduced-order models for thermomechanical fatigue, high-cycle fatigue, and wear were established respectively to calculate the thermomechanical response, vibration response, and wear response.

[0022] Real-time damage for each failure mode is calculated based on thermomechanical response, vibration response, and wear response, and cumulative damage is obtained by accumulating the data according to the acquisition time sequence.

[0023] By associating the real-time damage, cumulative damage, and acquisition time of each failure mode, a multi-failure twin damage state is generated.

[0024] Optionally, generating the real-time damage feature vector includes:

[0025] Real-time damage, cumulative damage, and acquisition time of each failure mode are extracted from the multi-failure twin damage state, and the damage sequence is divided according to a unified time window.

[0026] Cycle aggregation is performed on thermomechanical fatigue, high-cycle fatigue, and wear damage respectively to generate multi-scale damage sequences;

[0027] Perform time alignment and normalization on multi-scale damage sequences according to a unified time window to generate aligned damage data;

[0028] The operating condition data corresponding to the time window is extracted from the standardized operating data, the damage data is correlated and aligned and arranged according to the acquisition time sequence to generate a real-time damage feature vector.

[0029] Optionally, the step of generating a smoothed probability weight vector through time-series smoothing includes:

[0030] Historical damage trajectories are generated based on the reduced-order digital twin models of thermomechanical fatigue, high-cycle fatigue, and wear, and associated with corresponding failure mode identifiers to generate probabilistic neural network samples.

[0031] A probabilistic neural network consisting of an input layer, a pattern layer, a summation layer, and an output layer is constructed based on probabilistic neural network samples.

[0032] The real-time damage feature vector is input into the probabilistic neural network to calculate and normalize the matching response with the historical damage trajectory of each failure mode to obtain the initial probability weight.

[0033] Arrange the continuous initial probability weights according to the acquisition time sequence, extract the failure mode identifier corresponding to the maximum probability weight to form an observation sequence, and construct a hidden Markov model based on the historical failure mode transition relationship.

[0034] The observed sequence is input into the Hidden Markov Model to perform state sequence smoothing, and a smoothed probability weight vector is obtained.

[0035] Optionally, the generation of failure mode contention states includes:

[0036] The probability weights corresponding to thermomechanical fatigue, high-cycle fatigue and wear are extracted from the smoothed probability weight vector, and the current dominant failure mode is determined by sorting the probability weights.

[0037] Arrange the probability weights of each failure mode according to the assessment time, calculate the weight difference and weight change rate between adjacent assessment periods, and generate the probability weight change trend.

[0038] Calculate the weight difference between the current dominant failure mode and other failure modes, and combine the direction of change of the weight difference and the trend of change of the probability weight to determine the competition-enhanced failure mode and the migration trend of the dominant failure mode;

[0039] By associating the current dominant failure mode, the competition-enhancing failure mode, and the migration trend of the dominant failure mode, a failure mode competition state is generated.

[0040] Optionally, the real-time correction of the position parameters and the calculation of dynamic reliability using a smoothed probability weight vector as time-varying weights include:

[0041] Offline test data were read from the structural component test database. Life samples and material damage parameters were extracted according to thermomechanical fatigue, high cycle fatigue and wear. The three-parameter Weibull distribution was fitted by maximum likelihood estimation to obtain the initial location parameters, scale parameters and shape parameters of each failure mode.

[0042] The cumulative damage of each failure mode is extracted from the multi-failure twin damage state. The nonlinear physical constraint relationship between the location parameters and the cumulative damage is established based on the material damage parameters. The real-time location parameters of each failure mode are calculated based on the current cumulative damage.

[0043] The real-time position parameters, along with the corresponding scale and shape parameters, are combined to form a modified three-parameter Weibull parameter. Based on the modified three-parameter Weibull parameter, the reliability of each failure mode at the current evaluation time is calculated.

[0044] Extract the time-varying weights corresponding to each failure mode from the smoothed probability weight vector, multiply the reliability of each failure mode by the corresponding time-varying weight and sum them to generate the dynamic reliability of the structural component.

[0045] Optionally, the generation of update parameters includes:

[0046] Extract continuous, non-failed operational segments from standardized operational data to generate non-failed operational data, and construct operational censored likelihood by combining dynamic reliability.

[0047] Using the multi-failure twin damage state as the initial extrapolation state, the thermomechanical fatigue digital twin reduction model, the high-cycle fatigue digital twin reduction model, and the wear digital twin reduction model are driven to perform virtual accelerated life extrapolation, generate virtual accelerated life observation data, and construct virtual life observation likelihood.

[0048] Records that deviate from the normal operating range of the corresponding monitoring quantity are extracted from the standardized operating data to generate sensor anomaly data. The likelihood of physical anomaly symptoms is constructed by combining the cumulative damage of the corresponding failure mode.

[0049] By integrating censored likelihood, virtual lifetime observation likelihood, and physical anomaly symptom likelihood, a multi-source probability likelihood function is constructed. The location parameters, scale parameters, shape parameters, and mode weights of each failure mode are updated through variational Bayesian inference to generate updated parameters.

[0050] Optionally, the step of generating lifetime assessment results and failure warning results by combining failure mode competition states includes:

[0051] The location parameters, scale parameters, shape parameters and mode weights of each failure mode are extracted from the updated parameters, and the update reliability of each failure mode is calculated.

[0052] The updated reliability of each failure mode is weighted according to the mode weight to obtain the updated dynamic reliability, and the remaining life of the structural component is calculated based on the updated parameters.

[0053] Extract the current dominant failure mode and the migration trend of the dominant failure mode from the failure mode competition state, correlate and update the dynamic reliability and the remaining life of the structural component, and generate life assessment results.

[0054] The updated dynamic reliability and remaining life of structural components are compared with preset reliability thresholds and preset remaining life thresholds, respectively. Combined with the current dominant failure mode and the migration trend of the dominant failure mode, a failure warning result is generated.

[0055] A structural component life assessment system based on digital twin according to an embodiment of the present invention includes the following modules:

[0056] The operation data generation module is used to collect and preprocess monitoring data and operating condition data to generate standardized operation data;

[0057] The twin damage calculation module is used to establish digital twin reduced-order models of thermomechanical fatigue, high-cycle fatigue and wear, calculate real-time and cumulative damage, and generate multi-failure twin damage states.

[0058] The damage feature generation module is used to perform cross-timescale alignment and normalization on the multi-failure twin damage state to generate a real-time damage feature vector.

[0059] The probability weight generation module is used to construct a probabilistic neural network to process real-time damage feature vectors, obtain initial probability weights and perform temporal smoothing, and generate a smoothed probability weight vector.

[0060] The competition state generation module is used to analyze the smoothing probability weight vector, determine the dominant failure mode and migration trend, and generate the failure mode competition state.

[0061] The reliability calculation module is used to initialize the three-parameter Weibull parameters, perform nonlinear physical constraint correction on the position parameters, and calculate the dynamic reliability according to the smooth probability weight vector.

[0062] The parameter update module is used to construct a multi-source probability likelihood function by combining standardized operating data, multi-failure twin damage states and dynamic reliability, and generate updated parameters through variational Bayesian inference.

[0063] The assessment and early warning module is used to calculate dynamic reliability and remaining lifetime based on updated parameters, and generate lifetime assessment results and failure early warning results by combining failure mode competition states.

[0064] The beneficial effects of this invention are:

[0065] This invention constructs a parallel digital twin model of thermomechanical fatigue, high-cycle fatigue, and wear to map monitoring data during the operation of structural components in real time, calculates the real-time damage and cumulative damage of each failure mode, and forms a unified damage characterization through cross-timescale normalization. This invention combines probabilistic neural networks and hidden Markov models to calculate smoothing probability weights, and identifies the dominant failure mode and migration trend based on weight changes, reducing the judgment bias of competing failures caused by different damage evolution scales and sudden changes in operating conditions, and improving the accuracy of online decoupling of multiple failure modes and identification of dominant failures.

[0066] This invention initializes three-parameter Weibull parameters based on offline test data, establishes nonlinear physical constraints by combining location parameters with real-time accumulated damage from digital twins, and dynamically weights the reliability of each failure mode according to smoothing probability weights to form the dynamic reliability of structural components. This invention can enhance the physical consistency between statistical lifetime distribution and actual damage evolution, reduce the reliability assessment bias caused by fixed Weibull parameters, and improve the accuracy of dynamic reliability calculation and remaining lifetime prediction under complex working conditions.

[0067] This invention constructs a multi-source probability likelihood function based on non-failure operational data, digital twin virtual accelerated life observation data, and sensor anomaly data, and updates the Weibull parameters and mode weights online through variational Bayesian inference. This invention can make full use of the operational information and virtual life information continuously generated during the period of no physical failure, reduce the limitation of insufficient failure samples on the life assessment model, improve the ability of continuous parameter updates, life status tracking and failure early warning, and enhance the adaptability and stability of the full life cycle assessment of key structural components of complex equipment. Attached Figure Description

[0068] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0069] Figure 1 This is a flowchart of a structural component life assessment method based on digital twins proposed in this invention;

[0070] Figure 2 This is a flowchart of the parallel digital twin reduced-order model structure and damage feature extraction in a structural component life assessment method based on digital twin proposed in this invention;

[0071] Figure 3 This is a flowchart of the physical constraint three-parameter Weibull dynamic reliability calculation in a structural component life assessment method based on digital twin proposed in this invention;

[0072] Figure 4 This is a system structure diagram of a structural component life assessment system based on digital twins proposed in this invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0074] refer to Figure 1 , Figure 2 and Figure 3 A method for assessing the lifespan of structural components based on digital twins, comprising:

[0075] Collect and preprocess monitoring data and operating condition data of structural components to generate standardized operating data;

[0076] Based on standardized operating data, a parallel digital twin reduced-order model for thermomechanical fatigue, high-cycle fatigue, and wear is established to calculate the real-time damage and cumulative damage of each failure mode and generate multi-failure twin damage states.

[0077] Based on the multi-failure twin damage state, cross-timescale normalization alignment is performed, and working condition data is associated to generate a real-time damage feature vector.

[0078] A probabilistic neural network is constructed. The real-time damage feature vector is input into the probabilistic neural network to obtain the initial probability weights, and a smoothed probability weight vector is generated through time-series smoothing.

[0079] Based on the smooth probability weight vector analysis of weight trends, the dominant failure modes and migration trends are determined, and the failure mode competition state is generated.

[0080] Read offline test data from the structural component test database, initialize the three-parameter Weibull parameters of each failure mode based on the offline test data, establish nonlinear physical constraints on the location parameters and cumulative damage, correct the location parameters in real time, and calculate the dynamic reliability using a smooth probability weight vector as a time-varying weight.

[0081] Unfailed operational data and sensor anomaly data are extracted from standardized operational data. Virtual accelerated lifetime extrapolation is performed based on multi-failure twin damage states. Multi-source probability likelihood functions are constructed in conjunction with dynamic reliability. Weibull parameters and mode weights are updated through variational Bayesian inference to generate updated parameters.

[0082] Dynamic reliability and remaining lifetime are calculated based on updated parameters, and lifetime assessment results and failure warning results are generated by combining failure mode competition states.

[0083] In this embodiment, generating standardized operational data includes:

[0084] Monitoring data is collected from temperature sensors, pressure sensors, vibration acceleration sensors, and oil particle sensors on structural components. Load, speed, and working stage data are read from the equipment operation control terminal to generate raw operating data.

[0085] The original operating data is subjected to anomaly removal, missing data marking and time correction. The monitoring data and operating condition data are associated according to the structural component identification and the time of acquisition to generate corrected operating data.

[0086] The operating data is synchronously corrected and standardized in units according to a unified sampling period. The monitoring data and operating condition data are arranged according to the acquisition time sequence to generate standardized operating data, including:

[0087] Standardized operational data includes standardized monitoring data, standardized operating condition data, and operational identification data;

[0088] Standardized monitoring data includes temperature measurements, pressure measurements, vibration acceleration measurements, oil metal abrasive particle concentration, and metal abrasive particle size distribution after time synchronization and unit unification.

[0089] Standardized operating condition data includes load values, speed values, and operating stage identifiers after time synchronization and unit unification;

[0090] The runtime identification data includes structural component identification and acquisition time.

[0091] In this embodiment, generating multiple failure twin damage states includes:

[0092] Monitoring data and operating condition data are extracted from standardized operating data, and corresponding twin inputs are defined according to thermomechanical fatigue, high-cycle fatigue, and wear.

[0093] The corresponding twin inputs are categorized according to thermomechanical fatigue, high-cycle fatigue, and wear, specifically as follows:

[0094] Extract standardized monitoring data and standardized operating condition data from standardized operation data;

[0095] Temperature and pressure measurements are extracted from standardized monitoring data, and load, speed and working stage identifiers are extracted from standardized operating condition data. These are then arranged according to the time of acquisition to generate thermomechanical fatigue twin inputs.

[0096] Vibration acceleration measurements are extracted from standardized monitoring data, and load, rotational speed, and working stage identifiers are extracted from standardized working condition data. These are then arranged according to the acquisition time to generate a high-cycle fatigue twin input.

[0097] The concentration and size distribution of metal abrasive particles in the oil are extracted from standardized monitoring data. The load value, speed value and working stage identifier are extracted from standardized operating condition data. The data are arranged according to the time of collection to generate wear twin input.

[0098] Based on twin inputs, digital twin reduced-order models for thermomechanical fatigue, high-cycle fatigue, and wear were established to calculate the thermomechanical response, vibration response, and wear response.

[0099] The calculations for thermal response, vibration response, and wear response are as follows:

[0100] Temperature, pressure, load, rotational speed, and working stage identifiers are extracted from the thermomechanical fatigue twin input. Geometric and material parameters of the structural components are read from the structural component test database. A transient thermomechanical coupled finite element model of the structural components is established, and the temperature field and stress-strain field are calculated under different working stages.

[0101] The temperature field and stress-strain field at consecutive calculation times are used to form a snapshot matrix of the thermodynamic state. The snapshot matrix of the thermodynamic state is decomposed into eigenvalues ​​corresponding to each mode. The eigenvalue of each mode is divided by the sum of all mode eigenvalues ​​to obtain the corresponding mode energy contribution rate.

[0102] Arrange the modes according to their modal energy contribution rates from largest to smallest and accumulate them sequentially. Retain the top 20 modes with a cumulative modal energy contribution rate of 99.5% as the reduced-order basic modes and establish a thermomechanical fatigue digital twin model for reduced-order mode.

[0103] The temperature and pressure measurements at the current acquisition time are projected onto the reduced-order fundamental mode. The thermodynamic state is updated by combining the load, speed, and operating stage indicators. The reduced-order state equation is solved, and the temperature, equivalent stress, and equivalent strain values ​​corresponding to the critical locations are reconstructed to generate the thermodynamic response, specifically:

[0104] The temperature and pressure measurements at the current acquisition time are mapped to the first 20 reduced-order basic modes according to the corresponding spatial measurement point positions. The measurement values ​​at each measurement point are multiplied by the modal components of the corresponding reduced-order basic modes at the same spatial position and summed to obtain the temperature projection and pressure projection corresponding to each reduced-order basic mode.

[0105] The mechanical load is determined based on the load value at the current acquisition time, the inertial load under the corresponding speed condition is determined based on the speed value, the thermal boundary conditions and mechanical boundary conditions of the corresponding working stage in the transient thermo-mechanical coupled finite element model are selected based on the working stage identifier, and the temperature projection, pressure projection, mechanical load and inertial load are combined to generate the generalized load corresponding to each reduced-order basic mode at the current acquisition time.

[0106] When the current acquisition time is the first acquisition time, the current temperature measurement value, pressure measurement value, load value and speed value are input into the transient thermo-mechanical coupled finite element model to calculate the initial temperature field, initial stress field and initial strain field corresponding to the first acquisition time. The initial temperature field, initial stress field and initial strain field are projected onto the first 20 reduced-order fundamental modes to obtain the initial state coefficients corresponding to the first 20 reduced-order fundamental modes.

[0107] When the current acquisition time is after the first acquisition time, read the state coefficients corresponding to the first 20 reduced-order basic modes of the previous acquisition time. Input the state coefficients of the previous acquisition time, the generalized loads corresponding to each reduced-order basic mode of the current acquisition time, and the time interval between adjacent acquisition times into the reduced-order state equations of the thermomechanical fatigue digital twin reduced-order model, solve the first 20 reduced-order state variables simultaneously, and obtain the state coefficients corresponding to each reduced-order basic mode of the current acquisition time.

[0108] Multiply each state coefficient at the current acquisition moment with the corresponding reduced-order fundamental mode, sum the calculation results of the first 20 orders one by one, and reconstruct the temperature field, stress field and strain field at the current acquisition moment;

[0109] The temperature value, stress components and strain components in each direction corresponding to the critical location of the structural component are extracted from the reconstructed temperature field, stress field and strain field. The equivalent stress value is calculated based on the stress components in each direction and the equivalent strain value is calculated based on the strain components in each direction. The temperature value, equivalent stress value, equivalent strain value and acquisition time are correlated to generate the thermomechanical response.

[0110] The vibration acceleration measurement value, load value, rotational speed value and working stage identifier are extracted from the high-cycle fatigue twin input. The modal test parameters of the structural component are read from the structural component test database to obtain the natural frequencies, damping ratios and modal stress bases of each order, and a high-cycle fatigue digital twin reduced-order model is established.

[0111] Perform a fast Fourier transform on the vibration acceleration measurement values ​​according to the continuous sampling interval, extract the vibration frequency components and corresponding vibration amplitudes, match the vibration frequency components with the natural frequencies of each order stored in the high-cycle fatigue digital twin reduced-order model, and select the corresponding modal order with the smallest frequency difference.

[0112] Determine the rotational frequency and corresponding harmonic components based on the current rotational speed value, and exclude modal responses that are inconsistent with the current operating stage by combining the working stage identifier;

[0113] The vibration amplitude corresponding to the retained frequency components is assigned according to the modal order to obtain the real-time response amplitude corresponding to each mode.

[0114] The modal stress bases corresponding to the critical locations of the structural components are read from the digital twin reduced-order model of high-cycle fatigue. The real-time response amplitude of each modality is multiplied by the corresponding modal stress base to calculate the modal stress components generated by each mode at the critical location.

[0115] The stress components of each modal are superimposed according to the modal order to obtain the dynamic stress sequence of the dangerous location within the current sampling interval;

[0116] The maximum and minimum stress values ​​are extracted from the dynamic stress sequence. The difference between the maximum and minimum stress values ​​is divided by 2 to obtain the dynamic alternating stress at the critical location of the structural component. The modal frequency, modal response amplitude and dynamic alternating stress are then correlated to generate the vibration response.

[0117] The concentration of metal abrasive particles in the oil, the particle size distribution of the metal abrasive particles, the load value, the rotational speed value, and the working stage identifier are extracted from the wear twin input. The contact load is calculated based on the load value, and the sliding displacement is calculated based on the rotational speed value and the time interval between adjacent acquisition times.

[0118] Based on contact load and sliding displacement, the Arcard wear relationship is established, the wear rate is used as the wear state quantity, and the wear observation is formed by the concentration of metal abrasive particles in the oil and the particle size distribution of metal abrasive particles. A reduced-order digital twin model of wear is established.

[0119] Particle filtering is used to perform online inversion correction of wear state variables to obtain the corrected wear rate. The wear depth is obtained by integrating the corrected wear rate according to the acquisition time. The wear response is generated by correlating the corrected wear rate, wear depth and acquisition time.

[0120] Real-time damage for each failure mode was calculated based on thermomechanical response, vibration response, and wear response, and cumulative damage was obtained by accumulating the data according to the acquisition time sequence, where:

[0121] Real-time damage for each failure mode is calculated based on thermomechanical response, vibration response, and wear response, and cumulative damage is obtained by accumulating the data according to the acquisition time sequence, as follows:

[0122] Temperature, equivalent stress, equivalent variation, and acquisition time are extracted from the thermomechanical response. The continuous thermomechanical response is divided into thermomechanical fatigue working cycles according to the working stage identifier. The maximum and minimum equivalent variation values ​​are extracted from each working cycle. The difference between the maximum and minimum equivalent variation values ​​is divided by 2 to obtain the equivalent variation amplitude value. The maximum and minimum equivalent stress values ​​are also extracted. The sum of the two is divided by 2 to obtain the average equivalent stress value.

[0123] The fatigue strength coefficient, fatigue ductility coefficient, fatigue strength index, fatigue ductility index and elastic modulus are read from the structural component test database. The fatigue strength coefficient is subtracted from the average equivalent stress value and then divided by the elastic modulus to obtain the elastic strain coefficient.

[0124] Using the nominal life cycle number of the current working cycle as the unknown quantity, twice the nominal life cycle number is exponentially calculated using the fatigue strength index and the fatigue ductility index respectively. The elastic strain coefficient is multiplied by the exponentially calculated result corresponding to the fatigue strength index, and the fatigue ductility coefficient is multiplied by the exponentially calculated result corresponding to the fatigue ductility index. The two products are summed, and the sum is set to be equal to the equivalent strain amplitude value of the current working cycle. The nominal life cycle number corresponding to the current working cycle is then calculated.

[0125] Divide 1 by the nominal life cycle number corresponding to the current working cycle to obtain the real-time thermomechanical fatigue damage of the current working cycle. Accumulate the real-time thermomechanical fatigue damage according to the acquisition time corresponding to the working cycle to obtain the cumulative thermomechanical fatigue damage.

[0126] Modal frequencies, modal response amplitudes, dynamic alternating stresses, and acquisition times are extracted from the vibration response. High-cycle fatigue analysis windows are divided according to continuous sampling intervals. Each modal frequency is multiplied by the duration of the current analysis window to obtain the actual number of cycles experienced by the corresponding mode within the current analysis window.

[0127] The high-cycle fatigue SN curve parameters are read from the structural component test database. The allowable number of cycles at the corresponding stress level is determined according to the current dynamic alternating stress in the SN curve. The actual number of cycles at each stress level is divided by the corresponding allowable number of cycles to obtain the high-cycle fatigue damage component at the corresponding stress level.

[0128] The high-cycle fatigue damage components corresponding to each stress level within the current analysis window are summed to obtain the real-time high-cycle fatigue damage of the current analysis window. The real-time high-cycle fatigue damage is accumulated sequentially according to the acquisition time corresponding to the analysis window to obtain the cumulative high-cycle fatigue damage.

[0129] The corrected wear rate, wear depth, and acquisition time are extracted from the wear response. The allowable wear failure depth of the contact part of the structural component is read from the structural component test database. The wear depth at the current acquisition time is subtracted from the wear depth at the previous acquisition time to obtain the wear increment of the current acquisition cycle.

[0130] When the current acquisition time is the first acquisition time, the wear rate will be multiplied by the duration of the first acquisition cycle to obtain the wear increment of the first acquisition cycle;

[0131] Divide the wear increment of the current acquisition cycle by the allowable wear failure depth to obtain the real-time wear damage of the current acquisition cycle. Then, accumulate the real-time wear damage of each acquisition cycle according to the acquisition sequence to obtain the cumulative wear damage.

[0132] By associating real-time thermomechanical fatigue damage, cumulative thermomechanical fatigue damage, real-time high-cycle fatigue damage, cumulative high-cycle fatigue damage, real-time wear damage, cumulative wear damage, and the corresponding acquisition time, the real-time damage and cumulative damage of each failure mode are formed.

[0133] By associating the real-time damage, cumulative damage, and acquisition time of each failure mode, a multi-failure twin damage state is generated, where:

[0134] Generate a multi-failure twin damage state, specifically as follows:

[0135] Using the acquisition time as the association index, the real-time thermomechanical fatigue damage, cumulative thermomechanical fatigue damage, real-time high-cycle fatigue damage, cumulative high-cycle fatigue damage, real-time wear damage, and cumulative wear damage corresponding to the same acquisition time are matched.

[0136] The six damage data items corresponding to each acquisition time are arranged sequentially according to the acquisition time sequence, and the structural component identifier and acquisition time are associated to generate a multi-failure twin damage state.

[0137] In this embodiment, generating the real-time damage feature vector includes:

[0138] Real-time damage, cumulative damage, and acquisition time for each failure mode are extracted from the multi-failure twin damage state. The damage sequence is divided according to a unified time window, where:

[0139] Damage sequences are divided according to a unified time window, specifically as follows:

[0140] Using the acquisition time as the time index, the real-time and cumulative damage of thermomechanical fatigue, the real-time and cumulative damage of high-cycle fatigue, and the real-time and cumulative damage of wear are continuously segmented according to a preset unified time window of 60s.

[0141] Damage data for each failure mode falling within the same time window are arranged according to the time of acquisition to generate a damage sequence for the corresponding time window.

[0142] Cycle aggregation was performed on thermomechanical fatigue, high-cycle fatigue, and wear damage respectively to generate multi-scale damage sequences, where:

[0143] Cycle aggregation is performed separately for thermomechanical fatigue, high-cycle fatigue, and wear damage, specifically as follows:

[0144] The working stage identifier is extracted from the standardized operation data and matched with the damage sequence of each failure mode according to the collection time. The continuous time period from one start-up to one shutdown is taken as the thermomechanical fatigue aggregation cycle. The real-time thermomechanical fatigue damage within the aggregation cycle is summed, and the cumulative thermomechanical fatigue damage at the end of the aggregation cycle is extracted to generate the thermomechanical fatigue cycle damage sequence.

[0145] Using a uniform time window of 60s as the high-cycle fatigue aggregation cycle, the real-time high-cycle fatigue damage within each uniform time window is summed, and the cumulative high-cycle fatigue damage at the end of the window is extracted to generate a high-cycle fatigue cycle damage sequence.

[0146] The wear aggregation cycle is divided according to the adjacent acquisition time when the wear real-time damage occurs and updated. The wear real-time damage in each wear aggregation cycle is summed, and the wear cumulative damage at the end of the aggregation cycle is extracted to generate the wear cycle damage sequence.

[0147] Arrange the thermomechanical fatigue cycle damage sequence, high-cycle fatigue cycle damage sequence, and wear cycle damage sequence according to the start and end times corresponding to each polymerization cycle to generate a multi-scale damage sequence.

[0148] Time alignment and normalization are performed on multi-scale damage sequences according to a unified time window to generate aligned damage data, where:

[0149] Generate alignment damage data, specifically:

[0150] A continuous time index is established using a 60s unified time window. The high-cycle fatigue cycle damage sequence is directly mapped to the corresponding time window. The thermomechanical fatigue aggregation cycle and wear aggregation cycle are mapped to the unified time window with time overlap according to the start time and end time, respectively. The real-time cycle damage is allocated according to the ratio of the overlap duration between each unified time window and the corresponding aggregation cycle to the duration of the aggregation cycle.

[0151] For each failure mode, the minimum real-time damage, maximum real-time damage, minimum cumulative damage, and maximum cumulative damage after time alignment are calculated. The current real-time damage is subtracted from the corresponding minimum real-time damage, and then divided by the difference between the maximum real-time damage and the minimum real-time damage to obtain the normalized real-time damage.

[0152] Subtract the corresponding minimum cumulative damage from the current cumulative damage, and then divide by the difference between the maximum and minimum cumulative damage to obtain the normalized cumulative damage. When the maximum and minimum values ​​are the same, set the corresponding normalized damage value to 0.

[0153] Aligned damage data is generated by sequentially associating the normalized real-time damage, normalized cumulative damage, and end time of the time window corresponding to thermomechanical fatigue, high-cycle fatigue, and wear within a 60-second unified time window.

[0154] Operating condition data corresponding to the time window is extracted from standardized operating data, damage data is correlated and aligned, and arranged according to the acquisition time sequence to generate a real-time damage feature vector, where:

[0155] Generate a real-time damage feature vector, specifically as follows:

[0156] Load values, speed values, working stage identifiers, and acquisition times are extracted from standardized operating data, and the operating condition data is divided according to a uniform 60-second time window.

[0157] The average load value within each unified time window is used to obtain the window load value, and the average speed value is used to obtain the window speed value. The working stage identifier corresponding to the end time of the unified time window is extracted to generate window operating condition data.

[0158] According to the end time of the unified time window, the window operating condition data is correlated with the normalized real-time damage of thermomechanical fatigue, the normalized cumulative damage of thermomechanical fatigue, the normalized real-time damage of high-cycle fatigue, the normalized cumulative damage of high-cycle fatigue, the normalized real-time damage of wear, and the normalized cumulative damage of wear within the corresponding time window.

[0159] Six normalized damage data, window load value, window rotation speed value, and working stage identifier are arranged sequentially to form a 9-dimensional damage feature corresponding to a unified time window. These features are then arranged sequentially according to the end time of the unified time window to generate a real-time damage feature vector.

[0160] In this embodiment, the step of generating a smoothing probability weight vector through time-series smoothing includes:

[0161] Historical damage trajectories are generated based on reduced-order digital twin models of thermomechanical fatigue, high-cycle fatigue, and wear, and associated with corresponding failure mode identifiers to generate probabilistic neural network samples, wherein:

[0162] The process of generating probabilistic neural network samples is as follows:

[0163] Historical load values, historical rotational speed values, historical working stage identifiers, and corresponding temperature, pressure, vibration acceleration, oil metal abrasive concentration, and metal abrasive particle size distribution are read from the structural component test database and formed into a historical working condition sequence according to the time of collection.

[0164] The historical operating condition sequence is input into the thermomechanical fatigue digital twin reduced-order model, the high-cycle fatigue digital twin reduced-order model, and the wear digital twin reduced-order model respectively. The real-time damage and cumulative damage corresponding to thermomechanical fatigue, high-cycle fatigue and wear are calculated according to the acquisition time sequence. The historical damage trajectory is generated by aggregating, time aligning and normalizing according to a unified time window of 60 seconds.

[0165] For each unified time window, calculate the proportion of real-time damage from thermomechanical fatigue, real-time damage from high-cycle fatigue, and real-time damage from wear to the sum of the three types of real-time damage. The failure mode corresponding to the largest proportion is determined as the dominant failure mode of the current time window, and associated with thermomechanical fatigue, high-cycle fatigue, or wear indicators respectively.

[0166] The six normalized damage data, window load value, window speed value and working stage identifier within each unified time window are combined to form a nine-dimensional historical damage feature. The nine-dimensional historical damage feature is then associated with the corresponding dominant failure mode identifier to generate a probabilistic neural network sample.

[0167] A probabilistic neural network consisting of an input layer, a pattern layer, a summation layer, and an output layer is constructed based on probabilistic neural network samples, wherein:

[0168] The probabilistic neural network constructed based on probabilistic neural network samples consists of an input layer, a pattern layer, a summation layer, and an output layer, specifically as follows:

[0169] Read the 9-dimensional historical damage features and corresponding dominant failure mode identifiers from the probabilistic neural network samples, and set the 9 feature quantities corresponding to the 9-dimensional historical damage features as input layer nodes to form the input layer;

[0170] Each probabilistic neural network sample is set as a pattern layer neuron. The corresponding 9-dimensional historical damage features are extracted as the neuron center vector. The difference between the input layer output and each center vector is calculated dimension by dimension, the sum is obtained, and then the sum is divided by twice the square of the smoothing factor. The negative value is then taken and the exponential operation is performed to obtain the Gaussian radial basis response of each pattern layer neuron.

[0171] Based on the three dominant failure modes of thermomechanical fatigue, high-cycle fatigue and wear, the neurons in the mode layer were divided into three groups. The Gaussian radial basis responses of the neurons in each group were summed to obtain the response values ​​of the three categories, forming a summation layer.

[0172] The three categories of response values ​​are subjected to exponential operations, and each exponential result is divided by the sum of the three exponential results to obtain the normalized category probabilities corresponding to thermomechanical fatigue, high cycle fatigue and wear, which constitute the output layer and complete the construction of the probabilistic neural network.

[0173] The real-time damage feature vector is input into a probabilistic neural network to calculate and normalize the matching response with the historical damage trajectory of each failure mode, thereby obtaining the initial probability weights, where:

[0174] The matching response to the historical damage trajectory for each failure mode is calculated and normalized, specifically as follows:

[0175] The 9-dimensional real-time damage feature vector corresponding to the current unified time window is input into the input layer of the probabilistic neural network and then passed to the pattern layer.

[0176] Read the 9-dimensional historical damage feature center vectors stored in each neuron of the pattern layer, subtract the real-time damage feature vector from each center vector dimension by dimension, square the 9 differences respectively and sum them to obtain the feature distance between the real-time damage state and the corresponding historical damage trajectory.

[0177] Read the smoothing factor determined during the construction phase of the probabilistic neural network, divide the feature distance by twice the square of the smoothing factor, take the negative value of the result and perform an exponential operation to obtain the matching response corresponding to the neurons of each pattern layer.

[0178] Matching responses of the corresponding neurons are collected according to the three failure modes of thermomechanical fatigue, high-cycle fatigue and wear. Matching responses of the same type are summed to obtain thermomechanical fatigue matching response, high-cycle fatigue matching response and wear matching response.

[0179] Perform exponential operations on the three matched responses respectively, and divide each exponential operation result by the sum of the three exponential operation results to obtain the initial probability weights corresponding to thermomechanical fatigue, high cycle fatigue and wear. The sum of the three initial probability weights is 1.

[0180] The initial probability weights are arranged sequentially according to the acquisition time sequence. The failure mode identifiers corresponding to the highest probability weights are extracted to form an observation sequence. A hidden Markov model is then constructed based on the historical failure mode transition relationships, where:

[0181] The failure mode identifiers corresponding to the highest probability weights are extracted to form an observation sequence, and a hidden Markov model is constructed based on the historical failure mode transition relationships, specifically:

[0182] Arrange the continuous initial probability weights corresponding to thermomechanical fatigue, high-cycle fatigue and wear according to the end time of the unified time window. Compare the three initial probability weights within each unified time window. Use the failure mode identifier corresponding to the largest initial probability weight as the observation identifier of the current time window. Arrange the observation identifiers in chronological order to form an observation sequence.

[0183] The dominant failure mode identifiers of each unified time window are extracted from the historical damage trajectories corresponding to the probabilistic neural network samples. The number of mode transitions between adjacent windows of thermomechanical fatigue, high cycle fatigue and wear are counted in chronological order. The number of times each failure mode transitions to the target failure mode is divided by the total number of transitions of the corresponding failure mode to obtain the state transition probabilities between the three types of failure modes, and a state transition matrix is ​​formed.

[0184] Thermomechanical fatigue, high-cycle fatigue, and wear are set as three implicit states. An initial state probability vector is generated based on the occurrence ratio of the three dominant failure modes in the first unified time window of the historical damage trajectory. The probability of generating three observation labels for each implicit state is calculated according to the corresponding frequency between the actual dominant failure modes in the historical damage trajectory and the failure mode labels corresponding to the maximum probability weight output by the probabilistic neural network, and an observation probability matrix is ​​formed.

[0185] A hidden Markov model is constructed from three hidden states, an initial state probability vector, a state transition matrix, and an observation probability matrix, which is used to perform temporal smoothing of the state changes of failure modes in a continuous observation sequence.

[0186] The Hidden Markov Model is constructed from three hidden states, an initial state probability vector, a state transition matrix, and an observation probability matrix, specifically as follows:

[0187] Thermomechanical fatigue, high-cycle fatigue, and wear are set as the three hidden states of the Hidden Markov Model, and the thermomechanical fatigue label, high-cycle fatigue label, and wear label corresponding to the maximum initial probability weight output by the probabilistic neural network are set as the three observation labels.

[0188] The dominant failure mode identifiers for each unified time window are extracted from the historical damage trajectories corresponding to the probabilistic neural network samples. The occurrence counts of thermomechanical fatigue, high-cycle fatigue, and wear within the first unified time window of the historical damage trajectory are counted as the dominant failure modes. The occurrence counts are divided by the total number of historical damage trajectories to obtain three initial state probabilities, which are then used to form an initial state probability vector.

[0189] Read the dominant failure mode identifiers of adjacent unified time windows in chronological order, and count the number of transitions from thermomechanical fatigue to thermomechanical fatigue, thermomechanical fatigue to high cycle fatigue, thermomechanical fatigue to wear, and other state combinations. Divide the number of times each current failure mode transitions to the three target failure modes by the total number of transitions of the current failure mode to obtain the transition probabilities of nine initial states. Arrange them according to the current failure mode and the target failure mode to form a 3x3 state transition matrix.

[0190] The 9-dimensional historical damage features corresponding to the historical damage trajectory are input into the probabilistic neural network. The observation identifier corresponding to the maximum initial probability weight in each unified time window is extracted and matched with the dominant failure mode identifier in the same time window. The number of times the 3 types of observation identifiers appear in the 3 hidden states is counted respectively. The number of times each type of observation identifier appears is divided by the total number of observations in the corresponding hidden state to obtain 9 initial observation probabilities. They are arranged according to the hidden state and observation identifier to form a 3x3 observation probability matrix.

[0191] The historical observation sequence, initial state probability vector, state transition matrix and observation probability matrix are input into the Baum-Welch training process. Based on the current model parameters, the forward probability of each unified time window being in the three hidden states is calculated along the forward time direction. Then, the backward probability corresponding to each unified time window is calculated from the end of the observation sequence to the beginning.

[0192] Multiply and normalize the forward and backward probabilities of the same unified time window to obtain the state occupancy probability of each hidden state in the corresponding time window. Calculate the joint transition probability between adjacent states based on the forward probability, state transition probability, observation probability and backward probability of two adjacent unified time windows.

[0193] The state occupancy probability and joint transition probability corresponding to all historical observation sequences are summed respectively. The state occupancy probability of the first unified time window is normalized to update the initial state probability vector. The joint transition probability between each state is divided by the sum of the state occupancy probabilities corresponding to the current state to update the state transition matrix. The state occupancy probability corresponding to each observation identifier is divided by the sum of the state occupancy probabilities corresponding to the hidden state to update the observation probability matrix.

[0194] Repeatedly perform forward probability calculation, backward probability calculation and three types of model parameter updates. Training ends when the difference in log likelihood of the observation sequence obtained from two adjacent training sessions is not greater than the preset convergence threshold of 0.0001, or when the number of training sessions reaches the preset maximum number of training sessions of 100. The final initial state probability vector, state transition matrix and observation probability matrix are retained to complete the construction of the hidden Markov model.

[0195] The observed sequence is input into the Hidden Markov Model to perform state sequence smoothing, obtaining a smoothed probability weight vector, where:

[0196] The observed sequence is input into the Hidden Markov Model to perform state sequence smoothing, obtaining a smoothed probability weight vector, specifically:

[0197] Read the initial state probability vector, state transition matrix and observation probability matrix from the hidden Markov model, and match the observation sequence with the continuous initial probability weights corresponding to thermomechanical fatigue, high cycle fatigue and wear according to the end time of the unified time window.

[0198] For the first unified time window, the initial state probabilities corresponding to the three hidden states are multiplied by the observation probability and the corresponding initial probability weight corresponding to the current observation identifier to obtain the initial state path scores corresponding to thermomechanical fatigue, high cycle fatigue and wear.

[0199] For each subsequent unified time window, the state path scores of the three hidden states in the previous unified time window are used as the starting point. Each state path score is multiplied by the state transition probability to the current candidate hidden state. The maximum value of the three products is selected as the preceding path score of the current candidate hidden state. The preceding path score is then multiplied by the observation probability corresponding to the current observation identifier and the initial probability weight corresponding to the current candidate hidden state to obtain the state path score of the current candidate hidden state. The preceding hidden state corresponding to the maximum product is recorded.

[0200] The scores of the three state paths corresponding to thermomechanical fatigue, high-cycle fatigue and wear within each unified time window are summed. Each state path score is divided by the sum of the scores of the three state paths to obtain the smoothing probability weights of thermomechanical fatigue, high-cycle fatigue and wear for the corresponding unified time window.

[0201] Select the hidden state with the highest state path score from the last unified time window, and backtrack window by window according to the previous hidden states to obtain the smooth state sequence corresponding to the continuous unified time windows. Arrange the three smooth probability weights of each unified time window in chronological order to generate a smooth probability weight vector.

[0202] In this embodiment, the generation of failure mode race states includes:

[0203] The probability weights corresponding to thermomechanical fatigue, high-cycle fatigue, and wear are extracted from the smoothed probability weight vector. The current dominant failure mode is determined by sorting the probability weights, where:

[0204] The dominant failure mode is determined by sorting according to probability weights, specifically as follows:

[0205] Extract the thermomechanical fatigue probability weight, high-cycle fatigue probability weight, and wear probability weight from the smoothed probability weight vector of the current unified time window;

[0206] Arrange the three failure modes in descending order of probability weight, determine the failure mode with the highest probability weight as the current dominant failure mode, and record the corresponding probability weight.

[0207] Arrange the probability weights of each failure mode according to the assessment time, calculate the weight difference and weight change rate between adjacent assessment periods, and generate the probability weight change trend, where:

[0208] The trend of changes in generation probability weights is as follows:

[0209] Using the end of each unified time window as the evaluation time, the probability weights of thermomechanical fatigue, high-cycle fatigue and wear are arranged in chronological order to form the probability weight time sequence corresponding to the three types of failure modes.

[0210] For each failure mode, the probability weight at the current evaluation time is subtracted from the probability weight at the previous evaluation time to obtain the weight difference for the current evaluation period.

[0211] Divide the weight difference by the time interval of 60s between the current evaluation time and the previous evaluation time to obtain the weight change rate of the corresponding failure mode.

[0212] When the weight difference is greater than 0, the corresponding failure mode is marked as weight increasing state; when the weight difference is equal to 0, it is marked as weight stable state; when the weight difference is less than 0, it is marked as weight decreasing state; and the absolute value of the weight change rate is used to characterize the corresponding weight change speed.

[0213] By sequentially associating the probability weight, weight difference, weight change rate, and weight change status of each failure mode at the assessment time, a probability weight change trend is generated.

[0214] Calculate the weight difference between the current dominant failure mode and other failure modes. Combine the direction of change in the weight difference with the trend of change in the probability weights to determine the migration trend of competition-enhanced failure modes and the dominant failure mode.

[0215] The competition-enhanced failure modes and the migration trends of the dominant failure modes were identified, specifically as follows:

[0216] Extract the probability weight of the dominant failure mode and the probability weights of the other two failure modes at the current evaluation time. Subtract the probability weights of the other two failure modes from the probability weight of the dominant failure mode at the current evaluation time to obtain the difference between the two mode weights at the current evaluation time.

[0217] Extract the probability weights of the same three failure modes at the previous assessment time, subtract the probability weights of the other two failure modes from the probability weight of the dominant failure mode at the previous assessment time to obtain the difference between the two mode weights at the previous assessment time, and subtract the corresponding difference between the previous mode weights from the current mode weight difference to obtain the change in weight difference.

[0218] When the probability weight difference of a certain non-dominant failure mode is greater than 0, the weight change rate is greater than 0, and the corresponding weight difference change is less than 0, it is determined that the weight interval between the corresponding non-dominant failure mode and the current dominant failure mode continues to shrink, and the corresponding non-dominant failure mode is identified as a competition-enhanced candidate failure mode.

[0219] When there are two competitive enhancement candidate failure modes, compare the absolute values ​​of the changes in the weight difference between the two candidate failure modes. The failure mode with the larger absolute value is identified as the competitive enhancement failure mode. If the absolute values ​​are the same, the failure mode with the larger current probability weight is identified as the competitive enhancement failure mode.

[0220] When a competition-enhanced failure mode exists, the change from the current dominant failure mode to a competition-enhanced failure mode is defined as the dominant failure mode migration trend.

[0221] When there is no competing failure mode, maintaining the current dominant failure mode unchanged is determined as the dominant failure mode migration trend;

[0222] By associating the current dominant failure mode, the competition-enhancing failure mode, and the migration trend of the dominant failure mode, a failure mode competition state is generated.

[0223] In this embodiment, the real-time correction of position parameters and the calculation of dynamic reliability using a smoothed probability weight vector as a time-varying weight include:

[0224] Offline test data were retrieved from the structural component test database. Life samples and material damage parameters were extracted according to thermomechanical fatigue, high-cycle fatigue, and wear. A three-parameter Weibull distribution was fitted using maximum likelihood estimation to obtain the initial location parameters, scale parameters, and shape parameters for each failure mode, where:

[0225] The initial location, scale, and shape parameters of each failure mode are obtained by fitting a three-parameter Weibull distribution using maximum likelihood estimation.

[0226] Thermomechanical fatigue failure test records, high-cycle fatigue failure test records, and wear failure test records are read from the structural component test database. The life values ​​from the start of the test to the time when the corresponding failure criteria are reached are extracted to form thermomechanical fatigue life samples, high-cycle fatigue life samples, and wear life samples. Fatigue strength parameters, fatigue ductility parameters, high-cycle fatigue curve parameters, allowable depth of wear failure, and nonlinear damage accumulation characteristic parameters are extracted to form material damage parameters.

[0227] For each failure mode, the corresponding lifetime samples are arranged in ascending order of lifetime value, the minimum lifetime value is extracted, and candidate values ​​for position parameters are set within the range of values ​​less than the minimum lifetime value. The candidate values ​​for position parameters are subtracted from each lifetime sample to obtain the corresponding lifetime offset, and candidate values ​​for position parameters with lifetime offsets greater than 0 are retained.

[0228] For each set of candidate location parameters, scale parameters, and shape parameters, the shape parameter is divided by the scale parameter to obtain the distribution coefficient. Each lifetime offset is divided by the scale parameter and then exponentially operated on by subtracting 1 from the shape parameter. The result of the exponential operation is multiplied by the distribution coefficient and then multiplied by the negative exponent of the result of the exponential operation on the shape parameter after dividing the lifetime offset by the scale parameter to obtain the three-parameter Weibull probability density corresponding to each lifetime sample.

[0229] Take the natural logarithm of the probability density corresponding to each lifetime sample, and sum the log probability densities of all lifetime samples to obtain the log likelihood value corresponding to the current parameter combination.

[0230] Under the constraints that the position parameter is less than the minimum lifetime value, the scale parameter is greater than 0, and the shape parameter is greater than 0, the three parameters are adjusted, the corresponding log-likelihood values ​​are repeatedly calculated, the log-likelihood values ​​of different parameter combinations are compared, and the position parameter, scale parameter, and shape parameter corresponding to the maximum log-likelihood value are determined as the initial position parameter, initial scale parameter, and initial shape parameter of the corresponding failure mode.

[0231] The same maximum likelihood fitting process was performed on the thermomechanical fatigue life samples, high-cycle fatigue life samples and wear life samples respectively to obtain the initial three-parameter Weibull parameters corresponding to the three types of failure modes.

[0232] The cumulative damage of each failure mode is extracted from the multi-failure twin damage state. A nonlinear physical constraint relationship between the location parameters and the cumulative damage is established based on material damage parameters. The real-time location parameters of each failure mode are calculated based on the current cumulative damage, where:

[0233] The nonlinear physical constraint relationship is as follows:

[0234] ;

[0235] in:

[0236] For the first Failure modes at the evaluation time The corresponding real-time location parameters;

[0237] Indicates the first Initial location parameters for each failure mode;

[0238] Indicates the first The cumulative damage of each failure mode at the current assessment time;

[0239] Indicates the first The nonlinear damage accumulation characteristic index corresponding to each failure mode is calibrated through offline damage mechanics tests.

[0240] This constraint function ensures that the position parameter decreases nonlinearly and monotonically as the cumulative damage increases, thus realizing a real-time hard constraint of the physical damage state on the boundary of the statistical distribution parameter.

[0241] A nonlinear physical constraint relationship between location parameters and cumulative damage is established based on material damage parameters. Real-time location parameters for each failure mode are calculated based on the current cumulative damage. Specifically:

[0242] The cumulative thermomechanical fatigue damage, high-cycle fatigue damage, and wear damage at the current evaluation moment are extracted from the multi-failure twin damage state. The initial position parameters of thermomechanical fatigue, high-cycle fatigue, and wear are extracted from the initial three parameters of Weibull parameters corresponding to each failure mode. The nonlinear damage accumulation characteristic index corresponding to thermomechanical fatigue, high-cycle fatigue, and wear is extracted from the material damage parameters.

[0243] For thermomechanical fatigue, the cumulative thermomechanical fatigue damage is exponentially operated according to the thermomechanical fatigue nonlinear damage accumulation characteristic index to obtain the thermomechanical fatigue nonlinear cumulative damage value. The thermomechanical fatigue nonlinear cumulative damage value is subtracted from 1 to obtain the remaining proportion of thermomechanical fatigue position parameters. The remaining proportion of thermomechanical fatigue position parameters is multiplied by the initial thermomechanical fatigue position parameters to obtain the calculated value of thermomechanical fatigue position parameters. The calculated value of thermomechanical fatigue position parameters is compared with 0 and the larger value is taken to obtain the real-time thermomechanical fatigue position parameters.

[0244] For high-cycle fatigue, the accumulated damage of high-cycle fatigue is exponentially operated according to the high-cycle fatigue nonlinear damage accumulation characteristic index to obtain the high-cycle fatigue nonlinear accumulated damage value. The high-cycle fatigue nonlinear accumulated damage value is subtracted from 1 to obtain the remaining proportion of the high-cycle fatigue position parameter. The remaining proportion of the high-cycle fatigue position parameter is multiplied by the initial high-cycle fatigue position parameter to obtain the calculated value of the high-cycle fatigue position parameter. The calculated value of the high-cycle fatigue position parameter is compared with 0 and the larger value is taken to obtain the real-time position parameter of high-cycle fatigue.

[0245] For wear, the cumulative wear damage is exponentially calculated according to the wear nonlinear damage cumulative characteristic index to obtain the wear nonlinear cumulative damage value. The wear nonlinear cumulative damage value is subtracted from 1 to obtain the remaining proportion of wear position parameters. The remaining proportion of wear position parameters is multiplied by the initial wear position parameters to obtain the calculated value of wear position parameters. The calculated value of wear position parameters is compared with 0 and the larger value is taken to obtain the real-time wear position parameters.

[0246] Based on the real-time location parameters of thermomechanical fatigue, high-cycle fatigue, and wear at the current assessment time, real-time location parameters for each failure mode are generated.

[0247] The real-time position parameters, along with the corresponding scale and shape parameters, form a modified three-parameter Weibull parameter. Based on this modified three-parameter Weibull parameter, the reliability of each failure mode at the current evaluation time is calculated, where:

[0248] The reliability of each failure mode at the current evaluation time is calculated based on the modified three-parameter Weibull parameters, specifically as follows:

[0249] The position parameters corresponding to the current evaluation time are extracted from the real-time position parameters of thermomechanical fatigue, high-cycle fatigue and wear. The corresponding scale parameters and shape parameters are extracted from the initial three-parameter Weibull parameters of each failure mode. The real-time position parameters, scale parameters and shape parameters of the same failure mode are associated to form the corrected three-parameter Weibull parameters of thermomechanical fatigue, high-cycle fatigue and wear respectively.

[0250] The first collection time in the standardized operating data is taken as the starting time of life statistics. The current cumulative running time of the structural component is obtained by subtracting the starting time of life statistics from the current evaluation time.

[0251] Substitute the three-parameter Weibull parameter of thermomechanical fatigue correction and the current cumulative running time into the three-parameter Weibull reliability function corresponding to thermomechanical fatigue. When the current cumulative running time is not greater than the real-time position parameter of thermomechanical fatigue, the reliability of thermomechanical fatigue is determined to be 1.

[0252] When the current cumulative running time is greater than the thermomechanical fatigue real-time position parameter, the current cumulative running time is subtracted from the thermomechanical fatigue real-time position parameter, the difference is divided by the thermomechanical fatigue scale parameter, and then a power operation is performed according to the thermomechanical fatigue shape parameter. After taking the negative value of the power operation result, an exponential operation is performed to obtain the thermomechanical fatigue reliability.

[0253] Substitute the three-parameter Weibull parameter of high-cycle fatigue correction and the current cumulative running time into the three-parameter Weibull reliability function corresponding to high-cycle fatigue. When the current cumulative running time is not greater than the real-time position parameter of high-cycle fatigue, the reliability of high-cycle fatigue is determined to be 1.

[0254] When the current cumulative runtime is greater than the real-time position parameter of high-cycle fatigue, the current cumulative runtime is subtracted from the real-time position parameter of high-cycle fatigue, the difference is divided by the high-cycle fatigue scale parameter, and then a power operation is performed according to the high-cycle fatigue shape parameter. After taking the negative value of the power operation result, an exponential operation is performed to obtain the high-cycle fatigue reliability.

[0255] Substitute the three-parameter Weibull parameter of wear correction and the current cumulative running time into the three-parameter Weibull reliability function corresponding to wear. When the current cumulative running time is not greater than the real-time wear location parameter, the wear reliability is determined to be 1.

[0256] When the current cumulative runtime is greater than the wear real-time location parameter, the current cumulative runtime is subtracted from the wear real-time location parameter, the difference is divided by the wear scale parameter, and then a power operation is performed according to the wear shape parameter. After taking the negative value of the power operation result, an exponential operation is performed to obtain the wear reliability.

[0257] Based on the thermomechanical fatigue reliability, high-cycle fatigue reliability, and wear reliability associated with the current assessment time, the reliability of each failure mode at the current assessment time is determined.

[0258] The three-parameter Weibull reliability function is expressed as:

[0259] ;

[0260] in:

[0261] Indicates the first Failure modes at the evaluation time Corresponding reliability;

[0262] This indicates the running time corresponding to the current evaluation moment;

[0263] Indicates the first The location parameters of each failure mode after real-time correction by nonlinear physical constraint relationships;

[0264] Indicates the first The scale parameters of the failure modes;

[0265] Indicates the first Shape parameters of the failure modes;

[0266] This indicates the failure mode number, which corresponds to thermomechanical fatigue, high-cycle fatigue, and wear, respectively.

[0267] The time-varying weights corresponding to each failure mode are extracted from the smoothed probability weight vector. The reliability of each failure mode is multiplied by its corresponding time-varying weight and summed to generate the dynamic reliability of the structural component, where:

[0268] The dynamic reliability of the structural component is generated by multiplying the reliability of each failure mode by its corresponding time-varying weight and summing the results.

[0269] Extract the time-varying weights of thermomechanical fatigue, high-cycle fatigue, and wear from the smoothed probability weight vector corresponding to the current evaluation time, and read the thermomechanical fatigue reliability, high-cycle fatigue reliability, and wear reliability corresponding to the same evaluation time.

[0270] The thermomechanical fatigue reliability is obtained by multiplying the thermomechanical fatigue time-varying weight by the thermomechanical fatigue weight.

[0271] Multiply the high-cycle fatigue reliability by the high-cycle fatigue time-varying weight to obtain the high-cycle fatigue weighted reliability;

[0272] The wear-weighted reliability is obtained by multiplying the wear reliability by the wear time-varying weight.

[0273] The dynamic reliability of the structural component at the current evaluation time is obtained by summing the thermomechanical fatigue weighted reliability, high-cycle fatigue weighted reliability, and wear weighted reliability, and then correlated with the current evaluation time to form dynamic reliability time series data.

[0274] The dynamic reliability of a structural component is expressed as:

[0275] ;

[0276] in:

[0277] Indicates the structural component at the time of evaluation Corresponding dynamic reliability;

[0278] Indicates the first Failure modes at the evaluation time The corresponding time-varying weights;

[0279] Indicates the first Failure modes at the evaluation time The corresponding three-parameter Weibull reliability;

[0280] This indicates the failure mode number, which corresponds to thermomechanical fatigue, high-cycle fatigue, and wear, respectively.

[0281] The sum of the three time-varying weights is 1.

[0282] In this embodiment, the generation of update parameters includes:

[0283] Extracting continuous, non-failure-prone operational segments from standardized operational data generates non-failure operational data, and combining this with dynamic reliability to construct operational censored likelihood, where:

[0284] The runtime censored likelihood is constructed by combining dynamic reliability, specifically as follows:

[0285] Extract the start time and current evaluation time of the current continuous non-failure running segment from the non-failure running data, and calculate the current continuous running time based on the two.

[0286] Extract the corresponding structural component dynamic reliability from the dynamic reliability time series data at the current evaluation time, take the structural component that is still in a non-failure state at the current evaluation time as the right-censored observation, and take the current structural component dynamic reliability as the running censored likelihood of the current right-censored observation;

[0287] Runtime censoring likelihood is represented as:

[0288] ;

[0289] This represents the runtime censored likelihood at the current evaluation time.

[0290] This indicates the dynamic reliability of the structural component at the current evaluation moment;

[0291] This indicates the cumulative running time corresponding to the current evaluation moment;

[0292] Using the multi-failure twin damage state as the initial extrapolation state, the thermomechanical fatigue digital twin reduction model, the high-cycle fatigue digital twin reduction model, and the wear digital twin reduction model are driven to perform virtual accelerated life extrapolation, generating virtual accelerated life observation data and constructing virtual life observation likelihood, where:

[0293] Perform virtual accelerated lifetime extrapolation to generate virtual accelerated lifetime observation data and construct virtual lifetime observation likelihood, specifically as follows:

[0294] Extract the thermomechanical fatigue cumulative damage, high-cycle fatigue cumulative damage, and wear cumulative damage corresponding to the current evaluation time from the multi-failure twin damage state, and use the three types of cumulative damage as the extrapolation initial damage state of the thermomechanical fatigue digital twin reduced-order model, the high-cycle fatigue digital twin reduced-order model, and the wear digital twin reduced-order model, respectively.

[0295] Extract temperature, pressure, vibration acceleration, oil metal abrasive concentration, metal abrasive particle size distribution, load, and rotational speed values ​​from the standardized operating data within the most recent 60-second unified time window before the current evaluation time. Calculate the average value of the same type of continuous numerical data within the unified time window, and extract the working stage identifier corresponding to the end time of the unified time window to generate the average extrapolated working condition.

[0296] The average extrapolation conditions are input into the thermomechanical fatigue digital twin reduced-order model, the high-cycle fatigue digital twin reduced-order model, and the wear digital twin reduced-order model, respectively. The current evaluation time is used as the virtual extrapolation start time. The average extrapolation conditions are continuously extrapolated to the thermomechanical response, vibration response, and wear response. The thermomechanical fatigue cumulative damage, high-cycle fatigue cumulative damage, and wear cumulative damage are updated sequentially according to the corresponding damage calculation relationship.

[0297] The three types of cumulative damage are checked sequentially according to the virtual extrapolation time. The virtual extrapolation stops when any one of the following—thermomechanical fatigue cumulative damage, high-cycle fatigue cumulative damage, and wear cumulative damage—reaches 1. The failure mode corresponding to the first cumulative damage to reach 1 is determined as the dominant failure mode in the virtual extrapolation. The time difference between the virtual extrapolation termination time and the current assessment time is used as the remaining life prediction value obtained from the digital twin virtual extrapolation, denoted as . ;

[0298] The standard deviation of the simulation uncertainty corresponding to the remaining lifetime prediction value is obtained based on the virtual extrapolation process, denoted as […]. By associating the current assessment time, three types of extrapolated initial cumulative damage, virtual extrapolated dominant failure mode, remaining lifetime prediction value, and simulation uncertainty standard deviation, virtual accelerated lifetime observation data is generated.

[0299] The location, scale, and shape parameters are extracted from the current modified three-parameter Weibull parameters to form the current Weibull distribution parameter set. The theoretical remaining life of the structural component at the current evaluation time is calculated based on the current Weibull distribution parameter set, denoted as . ;

[0300] remaining life prediction Subtract the theoretical remaining lifespan The remaining lifetime deviation is obtained, and the square of the remaining lifetime deviation is divided by twice the standard deviation of the simulation uncertainty. The square of the result is negative and an exponential operation is performed. The result of the exponential operation is then divided by the standard deviation of the simulation uncertainty. The product of square roots is used to obtain the virtual lifetime observation likelihood;

[0301] in:

[0302] The virtual lifetime observation likelihood is expressed as:

[0303] ;

[0304] in:

[0305] This represents the observed likelihood of virtual lifetime;

[0306] This represents the remaining lifetime prediction value obtained by performing virtual accelerated lifetime extrapolation with the current multi-failure twin damage state as the initial state.

[0307] Represents the set of parameters based on the current Weibull distribution. The calculated theoretical remaining lifetime;

[0308] This represents the standard deviation of the simulation uncertainty corresponding to the virtual accelerated lifetime extrapolation;

[0309] Represents the current set of Weibull distribution parameters;

[0310] Records deviating from the normal operating range of the corresponding monitored quantities are extracted from standardized operational data to generate sensor anomaly data. This data, combined with the cumulative damage of the corresponding failure modes, is used to construct a likelihood of physical anomaly symptoms, where:

[0311] Generate sensor anomaly data, specifically:

[0312] Temperature measurement, vibration acceleration measurement, oil metal abrasive concentration and acquisition time are extracted from standardized operating data. The working temperature gradient, vibration signal kurtosis and oil metal abrasive concentration are calculated according to a uniform 60s time window.

[0313] Read the anomaly judgment threshold values ​​corresponding to the working temperature gradient, vibration signal kurtosis, and oil metal abrasive concentration from the structural component test database, and compare the current working temperature gradient, vibration signal kurtosis, and oil metal abrasive concentration with the corresponding anomaly judgment threshold values ​​respectively.

[0314] When the operating temperature gradient exceeds the corresponding anomaly judgment threshold, the corresponding record will be marked as a thermomechanical fatigue anomaly record;

[0315] When the kurtosis of the vibration signal exceeds the corresponding anomaly judgment threshold, the corresponding record will be marked as a high-cycle fatigue anomaly record;

[0316] When the concentration of metal abrasive particles in the oil exceeds the corresponding anomaly judgment threshold, the corresponding record will be marked as a wear anomaly record;

[0317] Associate anomaly monitoring features, acquisition time, and corresponding failure mode identifiers to generate sensor anomaly data;

[0318] The likelihood of physical anomalies is constructed by combining the cumulative damage of the corresponding failure modes, specifically as follows:

[0319] Extract abnormal monitoring features, acquisition time and corresponding failure mode identifiers from sensor anomaly data, and extract the cumulative damage of the corresponding failure mode from the multi-failure twin damage state according to the acquisition time;

[0320] Based on the degree to which the anomaly monitoring characteristics exceed the corresponding anomaly judgment threshold, and combined with the cumulative damage of the corresponding failure mode, physical damage inversion is performed to obtain the local equivalent wasted life ratio caused by the current anomaly state for the k-th failure mode, denoted as . ;

[0321] Read the first from the structural component test database The nominal total lifetime corresponding to each failure mode at the current stress level The proportion of local equivalent lifespan consumption With nominal total life Multiply to obtain the equivalent consumed lifetime corresponding to the abnormal state;

[0322] Nominal total life Subtract the first Current real-time location parameters of each failure mode , to obtain the lifetime consumption corresponding to the current real-time location parameters;

[0323] Subtract the lifetime consumption corresponding to the current real-time position parameter from the equivalent lifetime consumed corresponding to the abnormal state to obtain the abnormal lifetime deviation.

[0324] Divide the square of the abnormal lifetime deviation by twice the square of the inversion uncertainty standard deviation, take the negative of the result and perform an exponential operation, then divide the result of the exponential operation by the standard deviation of the inversion uncertainty. The product of square roots yields the likelihood of physical anomalies;

[0325] in:

[0326] The likelihood representation of physical anomalies is as follows:

[0327] ;

[0328] in:

[0329] Indicates the likelihood of physical anomalies;

[0330] Indicates the first The proportion of local equivalent wasted lifetime obtained by physical damage inversion of sensor abnormality characteristics for each failure mode;

[0331] Indicates the first The nominal total lifetime corresponding to each failure mode at the current stress level;

[0332] Indicates the first The real-time position parameters of each failure mode after nonlinear physical constraint correction at the current evaluation moment;

[0333] This represents the standard deviation of the uncertainty corresponding to the physical damage inversion.

[0334] This indicates the failure mode number, which corresponds to thermomechanical fatigue, high-cycle fatigue, and wear, respectively.

[0335] Indicates the current evaluation time;

[0336] By integrating censored likelihood, virtual lifetime observation likelihood, and physical anomaly symptom likelihood, a multi-source probability likelihood function is constructed. The location parameters, scale parameters, shape parameters, and mode weights of each failure mode are updated using variational Bayesian inference to generate updated parameters, where:

[0337] By integrating censored likelihood, virtual lifetime observation likelihood, and physical anomaly symptom likelihood, a multi-source probability likelihood function is constructed. The location parameters, scale parameters, shape parameters, and mode weights of each failure mode are then updated via variational Bayesian inference. Specifically:

[0338] Read the runtime censoring likelihood, virtual lifetime observation likelihood, and physical anomaly symptom likelihood corresponding to the current evaluation period, match the three likelihoods according to the same evaluation time, and multiply the three likelihoods to generate the multi-source probability likelihood value for the current evaluation period.

[0339] Extract the location, scale, and shape parameters corresponding to thermomechanical fatigue, high-cycle fatigue, and wear from the current modified three-parameter Weibull parameters; extract the mode weights corresponding to the three types of failure modes from the smooth probability weight vector; and combine all parameters to be updated into a parameter set.

[0340] The parameter set is represented as:

[0341] ;

[0342] in:

[0343] This represents the set of parameters to be updated.

[0344] Indicates the first Location parameters of the failure modes;

[0345] Indicates the first The scale parameters of the failure modes;

[0346] Indicates the first Shape parameters of the failure modes;

[0347] Indicates the first Mode weights for each failure mode;

[0348] Indicates the total number of failure modes;

[0349] For each failure mode, the variational posterior corresponding to the location parameter and the scale parameter is set to a log-normal distribution, the variational posterior corresponding to the shape parameter is set to a truncated Gaussian distribution, and the joint variational posterior of the mode weights corresponding to thermomechanical fatigue, high cycle fatigue and wear is set to a Dirichlet distribution, forming a total factorial variational distribution.

[0350] The total factorization variational distribution corresponding to the set of parameters to be updated is represented as:

[0351] ;

[0352] in:

[0353] Denotes the joint variational posterior distribution of the set of parameters to be updated;

[0354] Let represent the variational posterior distribution corresponding to the location parameter of the k-th failure mode, and set it to a log-normal distribution;

[0355] Let represent the variational posterior distribution corresponding to the scale parameter of the k-th failure mode, and set it to a log-normal distribution;

[0356] Let represent the variational posterior distribution corresponding to the shape parameter of the k-th failure mode, and set it as a truncated Gaussian distribution;

[0357] Let represent the variational posterior distribution corresponding to the mode weight vector, and set it as a Dirichlet distribution;

[0358] Read the runtime censoring likelihood corresponding to each period from the continuous historical evaluation period, take the natural logarithm of each runtime censoring likelihood and take the opposite number to obtain the corresponding censoring information;

[0359] Divide each censored information amount by the sum of censored information amounts for all historical evaluation periods to obtain normalized censored information weights. Multiply each normalized censored information weight by its corresponding natural logarithm. Sum all products, take the opposite number, and then divide by the natural logarithm of the number of historical evaluation periods to obtain dynamic censored information entropy.

[0360] After taking the negative of the dynamic censored information entropy, an exponential operation is performed to obtain the dynamic information entropy forgetting factor. The number of periods between the current evaluation period and each historical evaluation period is used as the decay order. The dynamic information entropy forgetting factor is then exponentially operated according to the corresponding decay order to obtain the decay weight of each historical evaluation period.

[0361] Take the natural logarithm of the censored likelihood of each historical run, multiply it by the corresponding decay weight, and sum all the products to obtain the decayed censored log-likelihood of the historical run.

[0362] The log-values ​​of the censored historical running likelihood, the current virtual lifetime observation likelihood, and the current physical anomaly sign likelihood are combined. The parameter posterior distribution formed in the previous evaluation period is used as the current parameter prior distribution. The variational evidence lower bound for the current evaluation period is constructed by combining the current variational posterior distribution.

[0363] The variational lower bound of evidence, which includes the dynamic information entropy forgetting factor, is expressed as:

[0364] ;

[0365] in:

[0366] This indicates the lower bound of variational evidence corresponding to the current assessment period;

[0367] Indicates according to the variational posterior distribution Calculate the expectation;

[0368] Indicates the likelihood of runtime censoring;

[0369] This represents the observed likelihood of virtual lifetime;

[0370] Indicates the likelihood of physical anomalies;

[0371] Represents the dynamic information entropy forgetting factor, with a value range of [value range missing]. This is used to reduce the contribution of historical censorship information to current parameter updates;

[0372] This represents the Kullback-Leibler divergence between the variational posterior distribution and the prior distribution;

[0373] This represents the variational posterior distribution for the current evaluation period;

[0374] This represents the prior distribution of parameters formed in the previous evaluation period;

[0375] Reparameterization is performed on the log-normal variational distributions corresponding to the position and scale parameters, the truncated Gaussian variational distributions corresponding to the shape parameters, and the Dirichlet variational distributions corresponding to the mode weights, transforming the random sampling process into a differentiable computational process with respect to the variational parameters.

[0376] The gradient of the variational evidence lower bound with respect to each variational parameter is calculated. The Adam optimization algorithm is used to update the variational parameters with a learning rate of 0.005. The variational evidence lower bound is recalculated after each parameter update. The iteration continues until the change between two adjacent evidence lower bounds is lower than the preset convergence threshold of 0.0001, or the number of iterations reaches the preset maximum number of iterations of 1500, at which point the variational inference ends.

[0377] The posterior means of position parameters, scale parameters and shape parameters corresponding to thermomechanical fatigue, high cycle fatigue and wear are extracted from the converged variational posterior distribution, and the posterior means of mode weights corresponding to the three failure modes are extracted from the Dirichlet posterior distribution.

[0378] The posterior mean of the weights of the three types of modes is normalized so that the sum of the weights of the thermomechanical fatigue update mode, the high-cycle fatigue update mode, and the wear update mode is 1. The update parameters are generated by associating the update location parameters, update scale parameters, update shape parameters, and update mode weights corresponding to each failure mode.

[0379] In this embodiment, the step of generating lifetime assessment results and failure warning results by combining failure mode competition states includes:

[0380] The location parameters, scale parameters, shape parameters, and mode weights of each failure mode are extracted from the updated parameters, and the update reliability of each failure mode is calculated, where:

[0381] The update reliability of each failure mode is calculated as follows:

[0382] The update location parameters, update scale parameters, update shape parameters and update mode weights corresponding to thermomechanical fatigue, high cycle fatigue and wear are extracted from the update parameters respectively, and the parameter correspondence is established according to the failure mode identifier.

[0383] Extract the first data collection time and the current evaluation time of the structural component from the standardized operation data, and calculate the cumulative operation time of the structural component up to the current evaluation time;

[0384] For thermomechanical fatigue, high-cycle fatigue and wear, the effective life offset of the corresponding failure mode is obtained by subtracting the corresponding update position parameter from the cumulative running time.

[0385] Divide each effective life offset by the corresponding update scale parameter, perform a power operation on the obtained ratio according to the corresponding update shape parameter, take the opposite number of the power operation result and perform an exponential operation to obtain the thermomechanical fatigue update reliability, high cycle fatigue update reliability and wear update reliability.

[0386] The updated reliability of each failure mode is weighted according to its mode weight to obtain the updated dynamic reliability. The remaining life of the structural component is then calculated based on the updated parameters, where:

[0387] To obtain updated dynamic reliability, specifically:

[0388] The update mode weights corresponding to thermomechanical fatigue, high-cycle fatigue and wear are extracted from the update parameters. The update reliability of thermomechanical fatigue, high-cycle fatigue and wear corresponding to the current evaluation time are extracted from the update reliability data of each failure mode.

[0389] Multiply the thermomechanical fatigue renewal reliability by the corresponding renewal mode weight, multiply the high-cycle fatigue renewal reliability by the corresponding renewal mode weight, multiply the wear renewal reliability by the corresponding renewal mode weight, and sum the three products to obtain the renewal dynamic reliability at the current evaluation time.

[0390] The remaining life of structural components is calculated based on the updated parameters, specifically as follows:

[0391] Extract the update location parameter, update scale parameter, update shape parameter and update mode weight corresponding to each failure mode from the update parameters, and read the current cumulative running time;

[0392] Using the current cumulative operating time as the starting point for calculating the remaining life, the predicted time is incremented sequentially in a 60s time step. Each predicted time is substituted into the three-parameter Weibull reliability function corresponding to the failure mode to calculate the future reliability of thermomechanical fatigue, high cycle fatigue and wear at each predicted time.

[0393] The future reliability of the three failure modes at each prediction time is multiplied by the corresponding update mode weights and summed to obtain the updated dynamic reliability prediction sequence starting from the current evaluation time.

[0394] Divide the updated dynamic reliability at each future prediction time by the updated dynamic reliability at the current evaluation time to obtain the conditional survival probability sequence based on the current survival status.

[0395] Numerical integration is performed on the conditional survival probability sequence with a time step of 60s to calculate the theoretical expected survival time after the current evaluation time, and the theoretical expected survival time is taken as the remaining life of the structural component.

[0396] By associating the current assessment time with the updated dynamic reliability and remaining structural lifespan, updated lifespan status data is generated.

[0397] Extract the current dominant failure mode and its migration trend from the failure mode competition state, correlate and update the dynamic reliability and remaining structural life, and generate a life assessment result, where:

[0398] The lifetime assessment results are generated as follows:

[0399] Extract the current dominant failure mode and the migration trend of the dominant failure mode from the failure mode competition state, and read the updated dynamic reliability and remaining life of the structural component at the current evaluation time.

[0400] Establish correlations based on the current assessment time to form a life status record that includes the current dominant failure mode, the migration trend of the dominant failure mode, updated dynamic reliability, and the remaining life of the structural components.

[0401] Arrange the life status records according to the assessment time sequence to generate structural component life assessment results;

[0402] The updated dynamic reliability and remaining structural life are compared with preset reliability thresholds and preset remaining life thresholds, respectively. Combined with the current dominant failure mode and the migration trend of the dominant failure mode, a failure warning result is generated, where:

[0403] The failure warning results are generated as follows:

[0404] The safe service criteria for structural components of the same model were read from the structural component test database, and the reliability safety limit of 0.90 was used as the reliability threshold.

[0405] The nominal total lifetime and three-parameter Weibull initial parameters of each failure mode are read from the structural component test database. With a reliability of 0.90 as the safe service boundary, the safe service boundary running time corresponding to each failure mode is calculated. The nominal total lifetime is subtracted from the corresponding safe service boundary running time, and the minimum value of the obtained safe remaining lifetime is taken as the remaining lifetime threshold.

[0406] The updated dynamic reliability and the remaining life of the structural component are compared with the corresponding thresholds. When the updated dynamic reliability is less than 0.90, a reliability failure warning sign is generated. When the remaining life of the structural component is less than the remaining life threshold, a remaining life failure warning sign is generated. When both are lower than the corresponding thresholds, a composite failure warning sign is generated.

[0407] The system associates failure warning indicators, the current dominant failure mode, the corresponding update mode weight, the migration trend of the dominant failure mode, updates the dynamic reliability and the remaining life of the structural components, and generates failure warning results.

[0408] refer to Figure 4 A structural component life assessment system based on digital twins includes the following modules:

[0409] The operation data generation module is used to collect and preprocess monitoring data and operating condition data to generate standardized operation data;

[0410] The twin damage calculation module is used to establish digital twin reduced-order models of thermomechanical fatigue, high-cycle fatigue and wear, calculate real-time and cumulative damage, and generate multi-failure twin damage states.

[0411] The damage feature generation module is used to perform cross-timescale alignment and normalization on the multi-failure twin damage state to generate a real-time damage feature vector.

[0412] The probability weight generation module is used to construct a probabilistic neural network to process real-time damage feature vectors, obtain initial probability weights and perform temporal smoothing, and generate a smoothed probability weight vector.

[0413] The competition state generation module is used to analyze the smoothing probability weight vector, determine the dominant failure mode and migration trend, and generate the failure mode competition state.

[0414] The reliability calculation module is used to initialize the three-parameter Weibull parameters, perform nonlinear physical constraint correction on the position parameters, and calculate the dynamic reliability according to the smooth probability weight vector.

[0415] The parameter update module is used to construct a multi-source probability likelihood function by combining standardized operating data, multi-failure twin damage states and dynamic reliability, and generate updated parameters through variational Bayesian inference.

[0416] The assessment and early warning module is used to calculate dynamic reliability and remaining lifetime based on updated parameters, and generate lifetime assessment results and failure early warning results by combining failure mode competition states.

[0417] Example 1: To verify the feasibility of this invention in practice, this example selects a high-strength, high-speed, heavy-duty diesel engine cylinder head as the simulation evaluation object, using vermicular graphite cast iron as the material. The system continuously simulated for 960 hours, with a unified evaluation window of 60 seconds, resulting in a total of 57,600 evaluation windows. Data on temperature, pressure, vibration acceleration, oil metal abrasive particle concentration, particle size distribution, load, rotational speed, and operating stage were collected, with a 0.42% abrupt change anomaly and a 0.61% missing record setting. After anomaly removal, missing record marking, and time synchronization, a certain window corresponds to a temperature of 462.8℃, a pressure of 15.4MPa, and a root mean square vibration acceleration of 6.31m / s². 2 The abrasive particle concentration was 31.7 mg / L, the load rate was 88.6%, and the rotation speed was 1824 r / min.

[0418] A digital twin model for thermomechanical fatigue, employing approximately 550,000 finite elements, performs transient thermomechanical coupling analysis on start-up, shutdown, loading, and unloading conditions. After intrinsic orthogonal decomposition, the top 20 modes with a cumulative modal energy contribution rate reaching 99.5% are retained. The high-cycle fatigue model uses the main natural frequencies of 145Hz and 280Hz, with a vibration sampling frequency of 2kHz. The wear model calculates the wear rate based on the Archard relation and uses oil particle data corrected by particle filtering. After 720 hours of cumulative operation, the model yields the following results: critical location temperature 468.2℃, equivalent stress 286.4MPa, equivalent amplitude 0.00318, principal vibration frequency 279.6Hz, dynamic alternating stress 92.7MPa, and cumulative wear depth 18.6 mm. m corresponds to 0.417, 0.463, and 0.184 for thermomechanical fatigue, high-cycle fatigue, and wear cumulative damage, respectively.

[0419] The three types of damage were aligned and normalized according to a 60-second window completion time, and combined with the window load, window rotation speed, and working stage to form a 9-dimensional feature. The feature vector formed for the corresponding 720-hour evaluation window is [0.642, 0.517, 0.711, 0.563, 0.286, 0.244, 0.886, 0.829, 3]. In five adjacent windows, the normalized real-time damage for high-cycle fatigue increased from 0.671 to 0.711, thermomechanical fatigue decreased from 0.655 to 0.642, and wear increased from 0.271 to 0.286.

[0420] In the offline phase, 2000 PNN training samples were constructed, including 667 samples dominated by thermomechanical fatigue, 667 samples dominated by high-cycle fatigue, and 666 samples dominated by wear. 500 independent test samples were also included. The training conditions covered load rates of 0.35–1.00, speeds of 1200–2100 r / min, and temperatures of 340–510℃. Three representative training samples were [0.781, 0.694, 0.326, 0.251, 0.118, 0.104, 0.920, 0.840, 3], [0.318, 0.287, 0.844, 0.736, 0.141, 0.126, 0.870, 0.960, 3], and [0.226, 0.198, 0.294, 0.263, 0.812, 0.748, 0.630, 0.580, 2], corresponding to the labels thermomechanical fatigue, high-cycle fatigue, and wear, respectively. The single-window recognition accuracy of PNN was 93.6%, which was improved to 95.1% after HMM smoothing.

[0421] The current window input to the PNN yields three initial probability weights of 0.41, 0.48, and 0.11. When affected by vibration spikes, the single-window weights become 0.33, 0.58, and 0.09, which are then adjusted to 0.44, 0.47, and 0.09 after HMM smoothing to prevent premature switching of the dominant mode due to transient noise. After 18 consecutive windows, the smoothing weights become 0.39, 0.50, and 0.11, confirming the system's shift from thermomechanical fatigue to high-cycle fatigue as the dominant failure mode. The high-cycle fatigue weight increases from 0.462, 0.471, and 0.486 to 0.500, while the thermomechanical fatigue weight decreases from 0.423 to 0.390.

[0422] The structural component test database included 42 sets of thermomechanical fatigue life samples, 45 sets of high-cycle fatigue life samples, and 38 sets of wear life samples. After maximum likelihood fitting, the initial location parameters for the three failure modes were 900h, 780h, and 1400h, respectively. After nonlinear correction incorporating the cumulative damage at 720h, the location parameters decreased to 584.9h, 445.6h, and 1269.1h, respectively. The dynamic reliability was corrected from 0.9978 calculated by the fixed-parameter model to 0.9903, reducing the overestimation of reliability caused by the fixed parameters.

[0423] After 960 hours of operation, no physical failures occurred in the last 100 hours. The concentration of metal abrasive particles in the oil increased from 24.9 mg / L to 42.6 mg / L, and the local equivalent wasted life ratio of wear reached 0.317. Digital twin virtual extrapolation yielded a remaining life of 928 hours, with a simulation uncertainty standard deviation of 71 hours. Multi-source likelihood was used for variational Bayesian updates with an Adam learning rate of 0.005, a maximum iteration count of 1500 steps, and convergence at step 1218. The updated weights for the three modes were 0.32, 0.43, and 0.25, respectively, with an update dynamic reliability of 0.8868 and a remaining life of 864 hours. Since the update dynamic reliability was below the 0.90 safety threshold and the remaining life was below the 910-hour safe remaining life threshold, the system generated a composite failure warning, identifying high-cycle fatigue as the current dominant mode and wear as a competitive enhancement mode.

[0424] The comparative experiment used the same 2000 modeling samples and 500 test samples, including 260 single dominant mode samples, 120 dominant mode migration sequences, and 120 multi-mode competing samples. The Miner linear cumulative damage method had an average lifetime prediction error of 18.6% and a dominant failure mode identification accuracy of 62.3%; the 2-parameter Weibull method had 15.4% and 68.7%; the LSTM method had 9.8% and 86.5%; the digital twin combined with the fixed 3-parameter Weibull method had 7.1% and 90.4%; and the method of this invention had 4.3% and 96.2%. In the 120 dominant mode migration sequences, the present invention correctly identified 116, achieving a migration identification accuracy of 96.7%. The results show that the present invention can improve the accuracy of competing failure mode identification and reduce the remaining lifetime prediction error through physical constraints and multi-source variational Bayesian updates.

[0425] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing the lifespan of structural components based on digital twins, characterized in that, include: Collect and preprocess monitoring data and operating condition data of structural components to generate standardized operating data; Based on standardized operating data, a parallel digital twin reduced-order model for thermomechanical fatigue, high-cycle fatigue, and wear is established to calculate the real-time damage and cumulative damage of each failure mode and generate multi-failure twin damage states. Based on the multi-failure twin damage state, cross-timescale normalization alignment is performed, and working condition data is associated to generate a real-time damage feature vector. A probabilistic neural network is constructed. The real-time damage feature vector is input into the probabilistic neural network to obtain the initial probability weights, and a smoothed probability weight vector is generated through time-series smoothing. Based on the smooth probability weight vector analysis of weight trends, the dominant failure modes and migration trends are determined, and the failure mode competition state is generated. Read offline test data from the structural component test database, initialize the three-parameter Weibull parameters of each failure mode based on the offline test data, establish nonlinear physical constraints on the location parameters and cumulative damage, correct the location parameters in real time, and calculate the dynamic reliability using a smooth probability weight vector as a time-varying weight. Unfailed operational data and sensor anomaly data are extracted from standardized operational data. Virtual accelerated lifetime extrapolation is performed based on multi-failure twin damage states. Multi-source probability likelihood functions are constructed in conjunction with dynamic reliability. Weibull parameters and mode weights are updated through variational Bayesian inference to generate updated parameters. Dynamic reliability and remaining lifetime are calculated based on updated parameters, and lifetime assessment results and failure warning results are generated by combining failure mode competition states.

2. The method for assessing the lifespan of structural components based on digital twins according to claim 1, characterized in that, The generation of standardized operational data includes: Monitoring data is collected from temperature sensors, pressure sensors, vibration acceleration sensors, and oil particle sensors on structural components. Load, speed, and working stage data are read from the equipment operation control terminal to generate raw operating data. The original operating data is subjected to anomaly removal, missing data marking and time correction. The monitoring data and operating condition data are associated according to the structural component identification and the time of acquisition to generate corrected operating data. The operating data is synchronously corrected and standardized in terms of units according to a unified sampling cycle. The monitoring data and operating condition data are arranged according to the collection time sequence to generate standardized operating data.

3. The method for assessing the lifespan of structural components based on digital twins according to claim 1, characterized in that, The generation of multi-failure twin damage states includes: Monitoring data and operating condition data are extracted from standardized operating data, and corresponding twin inputs are divided according to thermomechanical fatigue, high cycle fatigue and wear. Based on the twin input, digital twin reduced-order models for thermomechanical fatigue, high-cycle fatigue, and wear were established respectively to calculate the thermomechanical response, vibration response, and wear response. Real-time damage for each failure mode is calculated based on thermomechanical response, vibration response, and wear response, and cumulative damage is obtained by accumulating the data according to the acquisition time sequence. By associating the real-time damage, cumulative damage, and acquisition time of each failure mode, a multi-failure twin damage state is generated.

4. The method for assessing the lifespan of structural components based on digital twins according to claim 1, characterized in that, The generation of the real-time damage feature vector includes: Real-time damage, cumulative damage, and acquisition time of each failure mode are extracted from the multi-failure twin damage state, and the damage sequence is divided according to a unified time window. Cycle aggregation is performed on thermomechanical fatigue, high-cycle fatigue, and wear damage respectively to generate multi-scale damage sequences; Perform time alignment and normalization on multi-scale damage sequences according to a unified time window to generate aligned damage data; The operating condition data corresponding to the time window is extracted from the standardized operating data, the damage data is correlated and aligned and arranged according to the acquisition time sequence to generate a real-time damage feature vector.

5. The method for assessing the lifespan of structural components based on digital twins according to claim 1, characterized in that, The generation of the smoothed probability weight vector through time-series smoothing includes: Historical damage trajectories are generated based on the reduced-order digital twin models of thermomechanical fatigue, high-cycle fatigue, and wear, and associated with corresponding failure mode identifiers to generate probabilistic neural network samples. A probabilistic neural network consisting of an input layer, a pattern layer, a summation layer, and an output layer is constructed based on probabilistic neural network samples. The real-time damage feature vector is input into the probabilistic neural network to calculate and normalize the matching response with the historical damage trajectory of each failure mode to obtain the initial probability weight. Arrange the continuous initial probability weights according to the acquisition time sequence, extract the failure mode identifier corresponding to the maximum probability weight to form an observation sequence, and construct a hidden Markov model based on the historical failure mode transition relationship. The observed sequence is input into the Hidden Markov Model to perform state sequence smoothing, and a smoothed probability weight vector is obtained.

6. The method for assessing the lifespan of structural components based on digital twins according to claim 1, characterized in that, The failure mode race state includes: The probability weights corresponding to thermomechanical fatigue, high-cycle fatigue and wear are extracted from the smoothed probability weight vector, and the current dominant failure mode is determined by sorting the probability weights. Arrange the probability weights of each failure mode according to the assessment time, calculate the weight difference and weight change rate between adjacent assessment periods, and generate the probability weight change trend. Calculate the weight difference between the current dominant failure mode and other failure modes, and combine the direction of change of the weight difference and the trend of change of the probability weight to determine the competition-enhanced failure mode and the migration trend of the dominant failure mode; By associating the current dominant failure mode, the competition-enhancing failure mode, and the migration trend of the dominant failure mode, a failure mode competition state is generated.

7. The method for assessing the lifespan of structural components based on digital twins according to claim 1, characterized in that, The real-time correction of position parameters and the calculation of dynamic reliability using a smoothed probability weight vector as time-varying weights include: Offline test data were read from the structural component test database. Life samples and material damage parameters were extracted according to thermomechanical fatigue, high cycle fatigue and wear. The three-parameter Weibull distribution was fitted by maximum likelihood estimation to obtain the initial location parameters, scale parameters and shape parameters of each failure mode. The cumulative damage of each failure mode is extracted from the multi-failure twin damage state. The nonlinear physical constraint relationship between the location parameters and the cumulative damage is established based on the material damage parameters. The real-time location parameters of each failure mode are calculated based on the current cumulative damage. The real-time position parameters, along with the corresponding scale and shape parameters, are combined to form a modified three-parameter Weibull parameter. Based on the modified three-parameter Weibull parameter, the reliability of each failure mode at the current evaluation time is calculated. Extract the time-varying weights corresponding to each failure mode from the smoothed probability weight vector, multiply the reliability of each failure mode by the corresponding time-varying weight and sum them to generate the dynamic reliability of the structural component.

8. The method for assessing the lifespan of structural components based on digital twins according to claim 1, characterized in that, The generated update parameters include: Extract continuous, non-failed operational segments from standardized operational data to generate non-failed operational data, and construct operational censored likelihood by combining dynamic reliability. Using the multi-failure twin damage state as the initial extrapolation state, the thermomechanical fatigue digital twin reduction model, the high-cycle fatigue digital twin reduction model, and the wear digital twin reduction model are driven to perform virtual accelerated life extrapolation, generate virtual accelerated life observation data, and construct virtual life observation likelihood. Records that deviate from the normal operating range of the corresponding monitoring quantity are extracted from the standardized operating data to generate sensor anomaly data. The likelihood of physical anomaly symptoms is constructed by combining the cumulative damage of the corresponding failure mode. By integrating censored likelihood, virtual lifetime observation likelihood, and physical anomaly symptom likelihood, a multi-source probability likelihood function is constructed. The location parameters, scale parameters, shape parameters, and mode weights of each failure mode are updated through variational Bayesian inference to generate updated parameters.

9. The method for assessing the lifespan of structural components based on digital twins according to claim 1, characterized in that, The generation of lifetime assessment results and failure warning results by combining failure mode competition states includes: The location parameters, scale parameters, shape parameters and mode weights of each failure mode are extracted from the updated parameters, and the update reliability of each failure mode is calculated. The updated reliability of each failure mode is weighted according to the mode weight to obtain the updated dynamic reliability, and the remaining life of the structural component is calculated based on the updated parameters. Extract the current dominant failure mode and the migration trend of the dominant failure mode from the failure mode competition state, correlate and update the dynamic reliability and the remaining life of the structural component, and generate life assessment results. The updated dynamic reliability and remaining life of structural components are compared with preset reliability thresholds and preset remaining life thresholds, respectively. Combined with the current dominant failure mode and the migration trend of the dominant failure mode, a failure warning result is generated.

10. A structural component life assessment system based on digital twins, comprising the structural component life assessment method based on digital twins as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The operation data generation module is used to collect and preprocess monitoring data and operating condition data to generate standardized operation data; The twin damage calculation module is used to establish digital twin reduced-order models of thermomechanical fatigue, high-cycle fatigue and wear, calculate real-time and cumulative damage, and generate multi-failure twin damage states. The damage feature generation module is used to perform cross-timescale alignment and normalization on the multi-failure twin damage state to generate a real-time damage feature vector. The probability weight generation module is used to construct a probabilistic neural network to process real-time damage feature vectors, obtain initial probability weights and perform temporal smoothing, and generate a smoothed probability weight vector. The competition state generation module is used to analyze the smoothing probability weight vector, determine the dominant failure mode and migration trend, and generate the failure mode competition state. The reliability calculation module is used to initialize the three-parameter Weibull parameters, perform nonlinear physical constraint correction on the position parameters, and calculate the dynamic reliability according to the smooth probability weight vector. The parameter update module is used to construct a multi-source probability likelihood function by combining standardized operating data, multi-failure twin damage states and dynamic reliability, and generate updated parameters through variational Bayesian inference. The assessment and early warning module is used to calculate dynamic reliability and remaining lifetime based on updated parameters, and generate lifetime assessment results and failure early warning results by combining failure mode competition states.

Images