A digital twin-based method for engine crankshaft stress analysis

Abstract

This invention discloses a digital twin-based method for engine crankshaft stress analysis, comprising the following steps: S1, acquiring multi-source observation data and preprocessing it; S2, parsing the state offset trajectory of the current analysis scenario, performing time-series registration and joint encoding; S3, constructing a twin calibration model, embedding latent variables, constraining the parameter candidate space, and establishing correlations; S4, performing variance contribution analysis on the parameter candidate space, screening for parameter combinations and performing constrained reparameterization, driving the twin calibration model to perform virtual simulation; S5, constructing a feedback evaluation quantity, iteratively deciding on the reparameterization direction, and updating the adjustable parameter terms of the twin calibration model; S6, performing state correction on the adjustable parameter terms, determining the calibration parameter state, and generating stress analysis results. This invention enables digital twin stress simulation of engine crankshafts under multiple analysis scenarios, improving the stability of crankshaft stress analysis results.

Description

Technical Field

[0001] This invention relates to the field of stress analysis technology, and in particular to a method for stress analysis of engine crankshafts based on digital twins. Background Technology

[0002] The engine crankshaft is a critical component in an engine that withstands complex alternating forces. Its stress state analysis is typically used for structural strength verification, fatigue risk assessment, and design parameter correction. In existing technologies, crankshaft stress analysis mainly relies on a combination of finite element simulation and experimental testing. This involves first establishing a finite element model of the crankshaft, then setting calculation parameters based on experimental or operating conditions to obtain the crankshaft response results, and finally verifying the simulation results using experimental data. This method can meet conventional analysis needs under single-scenario, fixed conditions.

[0003] However, in actual R&D and testing, crankshaft stress analysis often needs to be adapted to different analysis scenarios, such as component testing, bench testing, and assembly conditions. The data acquisition methods, measurement point layout, sampling period, and response scale differ under different scenarios, making it difficult to directly correlate experimental data, simulation data, and scenario records. Existing methods largely rely on manual data processing and experience-based judgment of scenario differences, lacking an effective mechanism for uniformly organizing multi-source observation data and expressing scenario states, which easily leads to subsequent analyses being based on inconsistent data.

[0004] Meanwhile, existing finite element simulation models lack reusability when switching between multiple scenarios. When the analysis scenario changes, it is usually necessary to readjust model parameters, boundary conditions, or solution settings, or even rebuild parts of the model. This process is cumbersome, and different personnel may have different understandings of parameter adjustment methods, which can easily introduce human error. Especially when there are many model parameters, it is often difficult to accurately determine which parameters have a greater impact on the response of the current scenario through conventional experience.

[0005] Furthermore, existing simulation and experimental calibration processes often employ a "calculation results - experimental results - manual correction" approach. This typically allows only coarse adjustments based on response deviations, lacking a continuous decision-making mechanism for the direction and magnitude of parameter adjustments, as well as stable correction methods for the updated parameter states. Therefore, during multiple iterations, issues such as unstable parameter correction, slow convergence speed, or difficulty in solidifying calibration results can easily arise, affecting the consistency of crankshaft stress analysis results under different scenarios.

[0006] Therefore, how to provide a method for engine crankshaft stress analysis based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a digital twin-based method for engine crankshaft stress analysis. This invention fully utilizes finite element parametric simulation, latent variable proxy modeling, Sobol sensitivity analysis, DDPG reinforcement learning, and unscented Kalman filtering techniques. It details the process of scene state encoding, twin calibration model construction, virtual deduction, parameter feedback correction, and calibration state solidification for multi-scenario crankshaft stress analysis. It has the advantages of strong model reusability, high parameter calibration stability, and high accuracy in multi-scenario stress analysis.

[0008] An engine crankshaft stress analysis method based on digital twin according to an embodiment of the present invention includes the following steps: S1. Acquire multi-source observation data of the engine crankshaft and preprocess it to form a scene observation sequence; S2. Based on the scene observation sequence, analyze the state offset trajectory of the current analysis scene relative to the benchmark analysis scene, perform temporal registration and joint encoding on the state offset trajectory, and form a scene state vector. S3. A twin calibration model is constructed by adopting finite element parameterization and latent variable proxy joint modeling, latent variable embedding is performed on the scene state vector, and the parameter candidate space corresponding to the current analysis scene is constrained in the latent variable space to establish the correlation between the parameter candidate space and the twin response. S4. Using the Sobol sensitivity analysis method, variance contribution analysis is performed on the parameter candidate space to screen out the sensitive parameter combinations corresponding to the current analysis scenario, and the sensitive parameter combinations are reparameterized under constraints to drive the twin calibration model to perform virtual inference. S5. Based on the deviation between the virtual simulation results and the experimental response data, a feedback evaluation quantity is constructed. A reinforcement learning method based on DDPG is used to iteratively decide the reparameterization direction of the sensitive parameter combination, and the adjustable parameter terms of the Siamese calibration model are updated according to the iterative decision results. S6. Using the unscented Kalman filter method, the adjustable parameter terms are corrected for their state, and the calibration parameter state of the current analysis scenario is determined based on the feedback evaluation quantity, generating stress analysis results.

[0009] Optionally, the preprocessing specifically includes: A unified time-series index is constructed based on the acquisition time of multi-source observation data. A sliding window consistency check method is used to identify abrupt and missing segments in the observation data. The multi-source observation data includes experimental acquisition data, simulation operation data, and scene record data. Neighborhood trend constraint correction is applied to mutated fragments, and segmented interpolation is used to complete missing fragments. Based on the correlation between the measurement points corresponding to the analysis scenario, the measurement point numbers are rearranged and the sampling periods are aligned in the corrected observation data. A robust scaling transformation method is used to unify the scale of observation amplitudes from different sources, forming a scene observation sequence with a unified time reference, a unified measurement point order, and a unified numerical scale.

[0010] Optionally, the twin calibration model is a computational model for parameter calibration and response extrapolation, wherein: The twin calibration model includes parametric finite element calculation relationships, latent variable surrogate functions, and feedback correction relationships; The parameterized finite element calculation relationship is used to represent the finite element calculation process in the benchmark analysis scenario as a calculation process driven by adjustable parameter terms. The adjustable parameter items are: model parameters that are adjusted according to changes in the analysis scenario during the finite element calculation process; The latent variable surrogate function is used to express the scene state vector using latent variables, and to limit the range of values ​​of adjustable parameter items based on the latent variable expression results, thereby forming a parameter candidate space corresponding to the current analysis scene. The feedback correction relationship is used to iteratively update the adjustable parameter items based on the deviation between the virtual simulation results and the experimental response data, so that the twin calibration model can participate in the response simulation and calibration solidification in the current analysis scenario according to the updated adjustable parameter items.

[0011] Optionally, S2 specifically includes: S21. Based on the scene observation sequence corresponding to the current analysis scenario, divide the sampling segments with the scene observation sequence corresponding to the benchmark analysis scenario, and calculate the state difference value between the two sets of scene observation sequences within the same sampling segment, where: The current analysis scenario is the scenario in which the engine crankshaft stress analysis is performed. The baseline analysis scenario is a pre-selected reference scenario used to compare differences with the current analysis scenario. The state difference value is the numerical offset between the two sets of scenario observation sequences within the same sampling segment. S22. Connect the state difference values ​​in sequence according to the sampling segment order, and mark the continuous offset segments according to the change direction and change magnitude of adjacent state difference values ​​to obtain the state offset trajectory of the current analysis scene relative to the benchmark analysis scene. S23. Using the dynamic time warping method, calculate the minimum cumulative distance path between the state offset trajectory and the corresponding change trajectory, and perform time alignment on the sampling segments in the state offset trajectory based on the minimum cumulative distance path to obtain the registered state offset trajectory. S24. A bidirectional gated cyclic unit is used to jointly encode the registered state offset trajectory, and the joint encoding result is bottleneck compressed and scaled to form a scene state vector.

[0012] Optionally, S24 specifically includes: S241. Based on the registered state offset trajectory, a bidirectional gated loop unit is used to perform recursive coding along the forward sequence of the sampling segments. The offset memory of the previous sampling segment is retained by the update gate, and the participation of the current sampling segment in the candidate hidden state is adjusted by the reset gate to obtain the forward recursive coding result. S242. A bidirectional gated loop unit is used to perform recursive encoding in reverse order of the sampling segments. The offset memory of the next sampling segment is retained by the update gate, and the participation of the current sampling segment in the candidate hidden state is adjusted by the reset gate to obtain the reverse recursive encoding result. S243. According to the same sampling segment position, the forward recursive coding result and the reverse recursive coding result are fused accordingly, and the coding components with conflicting directions are weakened based on the offset consistency between the two to obtain the joint coding result. S244. Bottleneck compression and scale constraint are applied to the joint coding results to retain the dominant change components in the state offset trajectory and to limit the compressed coding results to a preset state range, forming a scene state vector.

[0013] Optionally, S3 specifically includes: S31. In the benchmark analysis scenario, the finite element calculation process is decomposed into parameters. The model conditions that change with the analysis scenario are set as adjustable parameter items. Multiple sets of parameter samples are constructed around the adjustable parameter items, so that each set of parameter samples corresponds to a response solution record in the finite element calculation process. The parameter samples are formed in a way that each adjustable parameter item has no less than 10 sets of values. S32. Based on multiple sets of parameter samples and corresponding response solution records, train a radial basis function surrogate model so that the radial basis function surrogate model represents the computational correlation between the adjustable parameter terms and the twin response, wherein: The radial basis function surrogate model is a nonlinear approximation model that uses parameter samples as input variables and response solution records as target variables. The radial basis function surrogate model determines the contribution weight of each known parameter sample to the estimated twin response by calculating the radial distance between the sample of the parameter to be estimated and each known parameter sample, and obtains the estimated twin response based on the combination of contribution weights. When the average response error of the radial basis function surrogate model on the validation samples is greater than 8%, increase the number of parameter samples and retrain the radial basis function surrogate model. S33. Using the variational autoencoder method, latent variables are embedded into the scene state vector to obtain the distribution representation of the current analysis scene in the latent variable space. The latent variable values ​​corresponding to the current analysis scene are determined by reparameterized sampling. The dimension of the latent variable space is set according to one-quarter to one-half of the dimension of the scene state vector. S34. Based on the values ​​of the latent variables, the range of adjustable parameter terms is narrowed to form the parameter candidate space corresponding to the current analysis scenario. The parameter candidate space is then used as the parameter domain of the radial basis function surrogate model to obtain the twin calibration model for the current analysis scenario.

[0014] Optionally, S4 specifically includes: S41. Construct a Sobol sampling matrix in the parameter candidate space, such that each set of sampling results consists of the values ​​of adjustable parameter terms, and configure each set of sampling results into the Siamese calibration model to obtain the corresponding Siamese response values, wherein: The twin response value is: the numerical result obtained by the twin calibration model to perform response calculation under the corresponding sampling result, and the twin response value is associated with the corresponding sampling result one by one, and is used to calculate the variance contribution of the adjustable parameter term in subsequent calculations. S42. Based on the sampling results of each group and the corresponding twin response values, perform variance decomposition, calculate the first-order sensitivity index and the total effect sensitivity index corresponding to each adjustable parameter, and obtain the variance contribution results of each adjustable parameter to the twin response change. S43. Based on the first-order sensitivity index and the total effect sensitivity index, the adjustable parameter items that meet the preset screening conditions for variance contribution results are combined to obtain the sensitivity parameter combination corresponding to the current analysis scenario. S44. Based on the parameter candidate space, the boundary constraints of the sensitive parameter combination are reconstructed, and the reconstructed sensitive parameter combination is embedded into the twin calibration model to perform virtual inference in the current analysis scenario.

[0015] Optionally, S42 specifically includes: S421. Split the Sobol sampling matrix into a first sampling matrix and a second sampling matrix. Based on the first sampling matrix and the second sampling matrix, reconstruct the adjustable parameter terms of the twin calibration model and calculate the corresponding first response sequence and second response sequence. S422. For each adjustable parameter item, replace the sampling column corresponding to the adjustable parameter item in the first sampling matrix with the same sampling column in the second sampling matrix to obtain the cross-sampling matrix corresponding to the adjustable parameter item, and calculate the cross-response sequence based on the cross-sampling matrix. S423. Based on the response differences between the first response sequence, the second response sequence, and the cross-response sequence, calculate the independent variance contribution and total effect variance contribution of the adjustable parameter term to the twin response change. S424. Based on the independent variance contribution and the total effect variance contribution, the values ​​are proportionalized according to the total variance of the twinned response to obtain the first-order sensitivity index and the total effect sensitivity index corresponding to the adjustable parameter term.

[0016] Optionally, S5 specifically includes: S51. Based on the virtual simulation results, align the experimental response data with the same sampling positions, calculate the response difference at the same sampling position, and statistically analyze the absolute amplitude change and continuous variation amplitude of the response difference within 3 to 5 adjacent sampling positions to construct a feedback evaluation quantity, wherein: The test response data consists of the measured response sequence obtained through physical experiments under the current analysis scenario, and the test response data and the virtual simulation results are established in a one-to-one correspondence according to the sampling location. The response difference is the numerical difference between the virtual simulation result and the experimental response data at the same sampling location. S52. Based on the feedback evaluation, construct the reward value of the DDPG reinforcement learning method. Set the reward value as the sum of the decrease in the absolute magnitude of the response difference and the decrease in the magnitude of continuous change, so that the reward value represents the degree of reduction of the response bias by the current round of reparameterization adjustment. S53. The current value of the feedback evaluation quantity and the sensitive parameter combination is used as the state quantity of the DDPG reinforcement learning method. The policy network outputs the continuous action quantity, and the continuous action quantity is parsed into the reparameterization direction and adjustment magnitude of the sensitive parameter combination. S54. Based on the parameter candidate space, the adjustment range is subject to boundary constraints, so that the adjustment range of the sensitive parameter combination in each iteration is controlled within 10% to 20% of the corresponding value range. The sensitive parameter combination is then corrected according to the reparameterization direction and the adjustment range after boundary constraints, and the corrected sensitive parameter combination is updated to the adjustable parameter terms of the twin calibration model.

[0017] Optionally, S6 specifically includes: S61. Take the updated adjustable parameter as the current parameter state and the previously corrected adjustable parameter as the prior parameter state. Based on the offset between the current parameter state and the prior parameter state, construct the state estimation object of the unscented Kalman filter method. S62. Construct Sigma sampling points around the state estimation object, and substitute the adjustable parameter terms corresponding to each Sigma sampling point into the Siamese calibration model to calculate the prediction feedback evaluation quantity corresponding to each Sigma sampling point. S63. Based on the deviation between each predicted feedback evaluation quantity and the feedback evaluation quantity, calculate the state correction gain, and correct the current parameter state through the state correction gain to obtain the corrected adjustable parameter item. S64. When the corrected feedback evaluation quantity meets the preset convergence condition, the corrected adjustable parameter item is determined as the calibration parameter state of the current analysis scenario, and the engine crankshaft stress analysis result is obtained based on the calibration parameter state.

[0018] The beneficial effects of this invention are: First, this invention performs unified preprocessing on multi-source observation data and analyzes the state offset trajectory of the current analysis scenario relative to the benchmark analysis scenario, so that experimental data, simulation operation data and scenario recording data can participate in the analysis under a unified time reference, measurement point sequence and numerical scale, reducing the interference caused by data inconsistency to subsequent stress analysis.

[0019] Secondly, this invention constructs a twin calibration model through joint modeling of finite element parameterization and latent variable proxy, and uses scene state vectors to constrain the parameter candidate space, enabling the model to adapt to different analysis scenarios and reducing the process of repeated modeling and manual parameter tuning; at the same time, it screens sensitive parameter combinations through Sobol sensitivity analysis, making parameter adjustment more targeted.

[0020] Finally, this invention constructs a feedback evaluation quantity based on the deviation between the virtual simulation results and the experimental response data, uses the DDPG reinforcement learning method for iterative decision-making, and combines the unscented Kalman filtering method to perform state correction on the adjustable parameter terms, making the parameter update process more stable and improving the reliability and consistency of the engine crankshaft stress analysis results. Attached Figure Description

[0021] 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: Figure 1 This is an overall flowchart of an engine crankshaft stress analysis method based on digital twin proposed in this invention; Figure 2 This is a schematic diagram of the construction of a twin calibration model for an engine crankshaft stress analysis method based on digital twin proposed in this invention; Figure 3 This is a flowchart illustrating the parameter candidate space constraints and sensitive parameter screening process for an engine crankshaft stress analysis method based on digital twins proposed in this invention. Detailed Implementation

[0022] 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.

[0023] refer to Figures 1-3 A method for stress analysis of engine crankshafts based on digital twins includes the following steps: S1. Acquire multi-source observation data of the engine crankshaft and preprocess it to form a scene observation sequence; S2. Based on the scene observation sequence, analyze the state offset trajectory of the current analysis scene relative to the benchmark analysis scene, perform temporal registration and joint encoding on the state offset trajectory, and form a scene state vector. S3. A twin calibration model is constructed by adopting finite element parameterization and latent variable proxy joint modeling, latent variable embedding is performed on the scene state vector, and the parameter candidate space corresponding to the current analysis scene is constrained in the latent variable space to establish the correlation between the parameter candidate space and the twin response. S4. Using the Sobol sensitivity analysis method, variance contribution analysis is performed on the parameter candidate space to screen out the sensitive parameter combinations corresponding to the current analysis scenario, and the sensitive parameter combinations are reparameterized under constraints to drive the twin calibration model to perform virtual inference. S5. Based on the deviation between the virtual simulation results and the experimental response data, a feedback evaluation quantity is constructed. A reinforcement learning method based on DDPG is used to iteratively decide the reparameterization direction of the sensitive parameter combination, and the adjustable parameter terms of the Siamese calibration model are updated according to the iterative decision results. S6. Using the unscented Kalman filter method, the adjustable parameter terms are corrected for their state, and the calibration parameter state of the current analysis scenario is determined based on the feedback evaluation quantity, generating stress analysis results.

[0024] In this embodiment, the preprocessing specifically includes: A unified time-series index is constructed based on the acquisition time of multi-source observation data, and a sliding window consistency check method is used to identify abrupt and missing segments in the observation data. Neighborhood trend constraint correction is applied to mutated fragments, and segmented interpolation is used to complete missing fragments. Based on the correlation between the measurement points corresponding to the analysis scenario, the measurement point numbers are rearranged and the sampling periods are aligned in the corrected observation data. A robust scaling transformation method is used to unify the scale of observation amplitudes from different sources, forming a scene observation sequence with a unified time reference, a unified measurement point order, and a unified numerical scale.

[0025] In one implementation, the current engine crankshaft analysis scenario to be analyzed is denoted as Scenario A, and the pre-selected reference analysis scenario is denoted as Scenario B. Multi-source observation data includes experimental acquisition data D1, simulation operation data D2, and scenario recording data D3. D1 is the measured response data acquired at the experimental end, D2 is the simulation response data obtained from the twin calibration model at the same analysis location, and D3 is a data table recording the scenario number, sampling time, measurement point number, and test batch.

[0026] During preprocessing, a unified time-series index T is established based on the sampling time fields in D1, D2, and D3, and D1 and D2 are time-linked according to T. A fixed-length sliding window W is used to check the consistency of D1 and D2. When a sampling point deviates significantly from the trend of the previous and subsequent sampling points, it is marked as a sudden change segment; when a sampling time in T does not have a corresponding response value in D1 or D2, it is marked as a missing segment.

[0027] For abrupt changes, corrections are made based on the neighborhood change trends of the same measurement point before and after sampling times. For missing measurement segments, segmented interpolation is performed to complete the missing segments based on the adjacent valid response data before and after them. Subsequently, based on the measurement point number field in D3, the order of measurement points in D1 and D2 is uniformly adjusted to the arrangement order of P1, P2, P3, and data with inconsistent sampling periods are resampled at equal intervals according to T.

[0028] A robust scaling transformation method was used to unify the scale of the observation amplitudes of D1 and D2, resulting in a scene observation sequence with a unified time reference, a unified measurement point order, and a unified numerical scale.

[0029] In this embodiment, the twin calibration model is a computational model for parameter calibration and response extrapolation, wherein: The twin calibration model includes parametric finite element calculation relationships, latent variable surrogate functions, and feedback correction relationships; Among them, the parameterized finite element calculation relationship is used to represent the finite element calculation process in the benchmark analysis scenario as a calculation process driven by adjustable parameter terms; The adjustable parameters are: model parameters that are adjusted as the analysis scenario changes during the finite element calculation process; The latent variable surrogate function is used to express the scene state vector using latent variables, and to limit the range of values ​​of adjustable parameter terms based on the latent variable expression results, thereby forming the parameter candidate space corresponding to the current analysis scene. The feedback correction relationship is used to iteratively update the adjustable parameter items based on the deviation between the virtual simulation results and the experimental response data, so that the twin calibration model can participate in the response simulation and calibration solidification in the current analysis scenario according to the updated adjustable parameter items.

[0030] In one implementation, the baseline analysis scenario is denoted as Scenario B, and the current analysis scenario is denoted as Scenario A. First, a finite element calculation process is established under Scenario B, and the model parameters that change with scenario switching are denoted as adjustable parameter terms K. K may include at least one of support equivalent parameters, contact equivalent parameters, or computational constraint parameters. Multiple sets of parameter values ​​are set around K, and finite element response solutions are performed for each set, obtaining the twin response R corresponding to each set of K, thereby establishing the parameterized finite element calculation relationship between K and R.

[0031] In the current analysis scenario A, the scenario state vector is input into the latent variable surrogate function to obtain the latent variable Z. The latent variable Z is not directly used as the stress analysis result, but rather to limit the range of values ​​for the adjustable parameter K, shrinking K from the global range corresponding to the benchmark analysis scenario B to a parameter candidate space suitable for the current analysis scenario A. Subsequently, the twin calibration model selects candidate values ​​for the adjustable parameter K within the parameter candidate space and performs response deduction based on the calculation relationship between K and R.

[0032] After obtaining the test response data under scenario A, the deviation between the response simulation results and the test response data is calculated, and the adjustable parameter K is corrected based on the deviation results. The corrected K is then re-involved in the response simulation of the twin calibration model until the deviation meets the preset conditions. At this point, the current K is determined as the calibration parameter state corresponding to scenario A.

[0033] In this embodiment, S2 specifically includes: S21. Based on the scene observation sequence corresponding to the current analysis scenario, divide the sampling segments with the scene observation sequence corresponding to the benchmark analysis scenario, and calculate the state difference value between the two sets of scene observation sequences within the same sampling segment, where: The current analysis scenario is the one used for this engine crankshaft stress analysis. Among them, the benchmark analysis scenario is a pre-selected reference scenario used to compare the differences with the current analysis scenario, and the state difference value is the numerical offset between the two sets of scenario observation sequences within the same sampling segment; S22. Connect the state difference values ​​in sequence according to the sampling segment order, and mark the continuous offset segments according to the change direction and change magnitude of adjacent state difference values ​​to obtain the state offset trajectory of the current analysis scene relative to the benchmark analysis scene. S23. Using the dynamic time warping method, calculate the minimum cumulative distance path between the state offset trajectory and the corresponding change trajectory, and perform time alignment on the sampling segments in the state offset trajectory based on the minimum cumulative distance path to obtain the registered state offset trajectory. S24. A bidirectional gated cyclic unit is used to jointly encode the registered state offset trajectory, and the joint encoding result is bottleneck compressed and scaled to form a scene state vector.

[0034] In this embodiment, S24 specifically includes: S241. Based on the registered state offset trajectory, a bidirectional gated loop unit is used to perform recursive coding along the forward sequence of the sampling segments. The offset memory of the previous sampling segment is retained by the update gate, and the participation of the current sampling segment in the candidate hidden state is adjusted by the reset gate to obtain the forward recursive coding result. S242. A bidirectional gated loop unit is used to perform recursive encoding in reverse order of the sampling segments. The offset memory of the next sampling segment is retained by the update gate, and the participation of the current sampling segment in the candidate hidden state is adjusted by the reset gate to obtain the reverse recursive encoding result. S243. According to the same sampling segment position, the forward recursive coding result and the reverse recursive coding result are fused accordingly, and the coding components with conflicting directions are weakened based on the offset consistency between the two to obtain the joint coding result. S244. Bottleneck compression and scale constraint are applied to the joint coding results to retain the dominant change components in the state offset trajectory and to limit the compressed coding results to a preset state range, forming a scene state vector.

[0035] In this embodiment, S3 specifically includes: S31. In the benchmark analysis scenario, the finite element calculation process is decomposed into parameters. The model conditions that change with the analysis scenario are set as adjustable parameter items. Multiple sets of parameter samples are constructed around the adjustable parameter items, so that each set of parameter samples corresponds to a response solution record in the finite element calculation process. The parameter samples are formed in such a way that each adjustable parameter item has no less than 10 sets of values. S32. Based on multiple sets of parameter samples and corresponding response solution records, train a radial basis function surrogate model so that the radial basis function surrogate model represents the computational correlation between the adjustable parameter terms and the twin response, wherein: The radial basis function surrogate model is a nonlinear approximation model that uses parameter samples as input variables and response solution records as target variables. The radial basis function surrogate model determines the contribution weight of each known parameter sample to the estimated twin response by calculating the radial distance between the sample of the parameter to be estimated and each known parameter sample, and obtains the estimated twin response based on the combination of contribution weights. When the average response error of the radial basis function surrogate model on the validation samples is greater than 8%, increase the number of parameter samples and retrain the radial basis function surrogate model. S33. Using a variational autoencoder method, latent variables are embedded into the scene state vector to obtain the distribution representation of the current analysis scene in the latent variable space. Then, through reparameterized sampling, the latent variable values ​​corresponding to the current analysis scene are determined. Specifically, this includes: An encoder is used to perform nonlinear encoding on the scene state vector. In the scene state vector, the latent components that represent the magnitude and direction of the state offset are extracted, and the latent components are converted into the latent variable mean and latent variable variance. Based on the mean and variance of the latent variables, construct the distribution of latent variables corresponding to the current analysis scenario, and limit the distribution of latent variables to the latent variable space corresponding to the range of values ​​of the adjustable parameter terms; Reparameterized sampling is performed within the defined latent variable space to obtain multiple latent variable sample values, and the multiple latent variable sample values ​​are sorted by center offset according to the latent variable mean. Select the latent variable sample value with the smallest center offset as the latent variable value corresponding to the current analysis scenario, so that the latent variable value can be used for the value center and value range of the subsequent constraint adjustable parameter items; The dimension of the latent variable space is set to one-quarter to one-half of the dimension of the scene state vector; S34. Based on the values ​​of the latent variables, narrow the range of adjustable parameter terms to form a parameter candidate space corresponding to the current analysis scenario. Use this parameter candidate space as the parameter domain of the radial basis function surrogate model to obtain a twin calibration model for the current analysis scenario. In this embodiment, S4 specifically includes: S41. Construct a Sobol sampling matrix in the parameter candidate space, such that each set of sampling results consists of the values ​​of adjustable parameter terms, and configure each set of sampling results into the Siamese calibration model to obtain the corresponding Siamese response values, wherein: The twin response value is the numerical result obtained by the twin calibration model from the response calculation under the corresponding sampling result. The twin response value is associated with the corresponding sampling result and is used to calculate the variance contribution of the adjustable parameter term in the subsequent calculation. S42. Based on the sampling results of each group and the corresponding twin response values, perform variance decomposition, calculate the first-order sensitivity index and the total effect sensitivity index corresponding to each adjustable parameter, and obtain the variance contribution results of each adjustable parameter to the twin response change. S43. Based on the first-order sensitivity index and the total effect sensitivity index, the adjustable parameter items that meet the preset screening conditions for variance contribution results are combined to obtain the sensitivity parameter combination corresponding to the current analysis scenario. S44. Based on the parameter candidate space, reconstruct the boundary constraints of the sensitive parameter combination, embed the reconstructed sensitive parameter combination into the Siamese calibration model, and perform virtual deduction under the current analysis scenario, specifically including: In the combination of sensitive parameters, the value center and value boundary of each sensitive parameter are determined, and the value center and value boundary are used as the constraint conditions for the constrained reparameterization. The Sigmoid boundary transformation method is used to perform a bounded transformation on the unconstrained adjustment of each sensitive parameter, converting the unconstrained adjustment into reparameterized parameter values ​​located within the value boundaries; Based on the reparameterized parameter values, replace the corresponding adjustable parameter terms in the twin calibration model, while maintaining the offset direction and offset magnitude of the reparameterized parameter values ​​relative to the value center; Based on the replaced adjustable parameters, the response calculation link of the twin calibration model is started, enabling the twin calibration model to perform virtual simulation under the current analysis scenario according to the reparameterized sensitive parameter combination.

[0036] In this embodiment, S42 specifically includes: S421. Split the Sobol sampling matrix into a first sampling matrix and a second sampling matrix. Based on the first sampling matrix and the second sampling matrix, reconstruct the adjustable parameter terms of the twin calibration model and calculate the corresponding first response sequence and second response sequence. S422. For each adjustable parameter item, replace the sampling column corresponding to the adjustable parameter item in the first sampling matrix with the same sampling column in the second sampling matrix to obtain the cross-sampling matrix corresponding to the adjustable parameter item, and calculate the cross-response sequence based on the cross-sampling matrix. S423. Based on the response differences between the first response sequence, the second response sequence, and the cross-response sequence, calculate the independent variance contribution and total effect variance contribution of the adjustable parameter term to the twin response change. S424. Based on the independent variance contribution and the total effect variance contribution, the values ​​are proportionalized according to the total variance of the twinned response to obtain the first-order sensitivity index and the total effect sensitivity index corresponding to the adjustable parameter term.

[0037] In this embodiment, S5 specifically includes: S51. Based on the virtual simulation results, align the experimental response data with the same sampling positions, calculate the response difference at the same sampling position, and statistically analyze the absolute amplitude change and continuous variation amplitude of the response difference within 3 to 5 adjacent sampling positions to construct a feedback evaluation quantity, wherein: The test response data consists of the measured response sequence obtained through physical experiments under the current analysis scenario, and the test response data and the virtual simulation results are established in a one-to-one correspondence according to the sampling location. The response difference is the numerical difference between the virtual simulation result and the experimental response data at the same sampling location. S52. Based on the feedback evaluation, construct the reward value of the DDPG reinforcement learning method. Set the reward value as the sum of the decrease in the absolute magnitude of the response difference and the decrease in the magnitude of continuous change, so that the reward value represents the degree of reduction of the response bias by the current round of reparameterization adjustment. S53. The current value of the feedback evaluation quantity and the sensitive parameter combination is used as the state quantity of the DDPG reinforcement learning method. The policy network outputs continuous action quantities, and the continuous action quantities are parsed into the reparameterization direction and adjustment magnitude of the sensitive parameter combination, specifically including: Based on the changing trend of the feedback evaluation quantity, the current value of the sensitive parameter combination is embedded into a state to form a reinforcement learning state quantity that includes the degree of bias convergence and the current position of the parameter. The policy network in the DDPG reinforcement learning method is used to perform deterministic policy calculation on the reinforcement learning state variables, and obtain the continuous action vector corresponding to the combination of sensitive parameters. The continuous motion vector is decomposed into directional and amplitude components. The directional component is used to determine the reparameterization direction of the sensitive parameter combination, and the amplitude component is used to determine the adjustment amplitude of the sensitive parameter combination. Based on the parameter candidate space, the adjustment range is constrained by action boundaries, and the reparameterized direction and adjustment range after action boundary constraints are used as the iterative decision result of the sensitive parameter combination. S54. Based on the parameter candidate space, the adjustment range is subject to boundary constraints, so that the adjustment range of the sensitive parameter combination in each iteration is controlled within 10% to 20% of the corresponding value range. The sensitive parameter combination is then corrected according to the reparameterization direction and the adjustment range after boundary constraints, and the corrected sensitive parameter combination is updated to the adjustable parameter terms of the twin calibration model.

[0038] In this embodiment, S6 specifically includes: S61. Take the updated adjustable parameter as the current parameter state and the previously corrected adjustable parameter as the prior parameter state. Based on the offset between the current parameter state and the prior parameter state, construct the state estimation object of the unscented Kalman filter method. S62. Construct Sigma sampling points around the state estimation object, and substitute the adjustable parameter terms corresponding to each Sigma sampling point into the Siamese calibration model to calculate the prediction feedback evaluation quantity corresponding to each Sigma sampling point. S63. Based on the deviation between each predicted feedback evaluation quantity and the feedback evaluation quantity, calculate the state correction gain, and correct the current parameter state through the state correction gain to obtain the corrected adjustable parameter item. S64. When the corrected feedback evaluation quantity meets the preset convergence condition, the corrected adjustable parameter item is determined as the calibration parameter state of the current analysis scenario, and the engine crankshaft stress analysis result is obtained based on the calibration parameter state.

[0039] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-scenario stress analysis of a crankshaft for a certain type of engine. This crankshaft requires adaptation to component-level support analysis, bench constraint analysis, and assembly state analysis during R&D verification. Existing methods typically require repeated adjustments to the finite element model parameters for different scenarios and rely on manual comparison of the experimental and simulation responses for correction. Due to differences in sampling periods, measurement point sequences, and response amplitude scales under different scenarios, some experimental data also exhibit short-term abrupt changes and missing measurements, making it difficult to stably reuse the model after switching scenarios, and the parameter calibration process is time-consuming.

[0040] In this scenario, experimental data D1, simulation data D2, and scenario record data D3 are first acquired, and a unified time-series index T is established based on the sampling time field. For abrupt changes in D1 and D2, a sliding window consistency check method is used for identification, and corrections are made based on the neighborhood change trend of the same measurement point before and after sampling times. For missing measurement segments, segmented interpolation is performed to complete the missing segments based on adjacent valid response data. Subsequently, based on the measurement point number field in D3, D1 and D2 are uniformly adjusted to the measurement point order of P1, P2, P3, P4, P5, and data with inconsistent sampling periods are resampled at equal intervals. Finally, a robust scaling transformation method is used to unify the observation amplitudes, forming a scenario observation sequence.

[0041] Subsequently, using the calibrated component-level support analysis as the baseline analysis scenario, and the bench constraint analysis and assembly state analysis as the current analysis scenarios, the state offset trajectory is parsed based on the scenario observation sequence, and a dynamic time warping method is used for time registration to align the misaligned response change segments under different scenarios. Then, a bidirectional gated cyclic unit is used to jointly encode the registered state offset trajectory to form a scenario state vector. This scenario state vector is used to characterize the state difference between the current scenario and the baseline scenario, rather than being directly used as the stress analysis result.

[0042] Next, a twin calibration model is formed through joint modeling using finite element parameterization and latent variable proxy. In the baseline analysis scenario, model parameters that change with the scenario are defined as adjustable parameter terms. Multiple sets of parameter samples are constructed around these adjustable parameter terms, and a radial basis function surrogate model is trained using response solution records to establish the computational correlation between the adjustable parameter terms and the twin response. For the current analysis scenario, a variational autoencoder method is used to embed latent variables into the scenario state vector, and the range of adjustable parameter terms is narrowed based on the latent variable values ​​to form a parameter candidate space.

[0043] Within the parameter candidate space, the Sobol sensitivity analysis method is used to screen sensitive parameter combinations, and the Sigmoid boundary transformation method is used to perform constrained reparameterization of these sensitive parameter combinations, driving the Siamese calibration model to perform virtual simulation. Subsequently, the virtual simulation results are aligned with the experimental response data to construct feedback evaluation quantities. A reinforcement learning method based on DDPG is used to determine the reparameterization direction and adjustment magnitude, and then the state of the adjustable parameter terms is corrected using an unscented Kalman filter to obtain the calibration parameter state corresponding to the current analysis scenario.

[0044] Verification showed that in bench constraint analysis scenarios, traditional manual calculations required eight rounds of adjustments, with an average response error of 14.8% and a maximum measurement point error of 22.6%. After adopting this invention, the average response error decreased to 7.8% after four iterations, and the maximum measurement point error decreased to 10.4%. In assembly state analysis scenarios, the traditional method had an average response error of 18.9% and a parameter fluctuation range of 0.31. After adopting this invention, the average response error decreased to 8.6%, and the parameter fluctuation range decreased to 0.12. In bench constraint analysis scenarios, for stress analysis results at five measurement points, the correlation coefficient between the analyzed values ​​obtained by this invention and the converted experimental response values ​​reached 0.94, indicating that this invention can improve the stability and consistency of crankshaft stress analysis under conditions of multi-scenario switching and data inconsistency.

[0045] Table 1 Comparison of Crankshaft Stress Calibration Effects in Multiple Scenarios

[0046] As can be seen from the data in Table 1, compared with the traditional manual calculation method, the method of the present invention shows better calibration stability and response consistency in different analysis scenarios. It can reduce the repeated calculation process, reduce response error and parameter fluctuation, and make the crankshaft stress analysis results closer to the experimental response data.

[0047] 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. An engine crankshaft stress analysis method based on digital twinning, characterized by, Includes the following steps: S1. Acquire multi-source observation data of the engine crankshaft and preprocess it to form a scene observation sequence. The multi-source observation data includes test acquisition data, simulation operation data and scene recording data. S2. Based on the scene observation sequence, analyze the state offset trajectory of the current analysis scene relative to the benchmark analysis scene, perform temporal registration and joint encoding on the state offset trajectory, and form a scene state vector. S3. A twin calibration model is constructed by adopting finite element parameterization and latent variable proxy joint modeling, latent variable embedding is performed on the scene state vector, and the parameter candidate space corresponding to the current analysis scene is constrained in the latent variable space to establish the correlation between the parameter candidate space and the twin response. S4. Using the Sobol sensitivity analysis method, variance contribution analysis is performed on the parameter candidate space to screen out the sensitive parameter combinations corresponding to the current analysis scenario, and the sensitive parameter combinations are reparameterized under constraints to drive the twin calibration model to perform virtual inference. S5. Based on the deviation between the virtual simulation results and the experimental response data, a feedback evaluation quantity is constructed. A reinforcement learning method based on DDPG is used to iteratively decide the reparameterization direction of the sensitive parameter combination, and the adjustable parameter terms of the Siamese calibration model are updated according to the iterative decision results. S6. Using the unscented Kalman filter method, the adjustable parameter terms are corrected for their state, and the calibration parameter state of the current analysis scenario is determined based on the feedback evaluation quantity, generating stress analysis results.

2. The engine crankshaft stress analysis method based on digital twin according to claim 1, characterized in that, The preprocessing specifically includes: A unified time-series index is constructed based on the acquisition time of multi-source observation data, and a sliding window consistency check method is used to identify abrupt and missing segments in the observation data. Neighborhood trend constraint correction is applied to mutated fragments, and segmented interpolation is used to complete missing fragments. Based on the correlation between the measurement points corresponding to the analysis scenario, the measurement point numbers are rearranged and the sampling periods are aligned in the corrected observation data. A robust scaling transformation method is used to unify the scale of observation amplitudes from different sources, forming a scene observation sequence with a unified time reference, a unified measurement point order, and a unified numerical scale.

3. The engine crankshaft stress analysis method based on digital twin according to claim 1, characterized in that, The twin calibration model is a computational model for parameter calibration and response extrapolation, wherein: The twin calibration model includes parametric finite element calculation relationships, latent variable surrogate functions, and feedback correction relationships; The parameterized finite element calculation relationship is used to represent the finite element calculation process in the benchmark analysis scenario as a calculation process driven by adjustable parameter terms. The adjustable parameter items are: model parameters that are adjusted according to changes in the analysis scenario during the finite element calculation process; The latent variable surrogate function is used to express the scene state vector using latent variables, and to limit the range of values ​​of adjustable parameter items based on the latent variable expression results, thereby forming a parameter candidate space corresponding to the current analysis scene. The feedback correction relationship is used to iteratively update the adjustable parameter items based on the deviation between the virtual simulation results and the experimental response data, so that the twin calibration model can participate in the response simulation and calibration solidification in the current analysis scenario according to the updated adjustable parameter items.

4. The engine crankshaft stress analysis method based on digital twin according to claim 1, characterized in that, S2 specifically includes: S21. Based on the scene observation sequence corresponding to the current analysis scenario, divide the sampling segments with the scene observation sequence corresponding to the benchmark analysis scenario, and calculate the state difference value between the two sets of scene observation sequences within the same sampling segment, where: The current analysis scenario is the scenario in which the engine crankshaft stress analysis is performed. The baseline analysis scenario is a pre-selected reference scenario used to compare differences with the current analysis scenario. The state difference value is the numerical offset between the two sets of scenario observation sequences within the same sampling segment. S22. Connect the state difference values ​​in sequence according to the sampling segment order, and mark the continuous offset segments according to the change direction and change magnitude of adjacent state difference values ​​to obtain the state offset trajectory of the current analysis scene relative to the benchmark analysis scene. S23. Using the dynamic time warping method, calculate the minimum cumulative distance path between the state offset trajectory and the corresponding change trajectory, and perform time alignment on the sampling segments in the state offset trajectory based on the minimum cumulative distance path to obtain the registered state offset trajectory. S24. A bidirectional gated cyclic unit is used to jointly encode the registered state offset trajectory, and the joint encoding result is bottlenecked and scaled to form a scene state vector.

5. The engine crankshaft stress analysis method based on digital twin according to claim 4, characterized in that, S24 specifically includes: S241. Based on the registered state offset trajectory, a bidirectional gated loop unit is used to perform recursive coding along the forward sequence of the sampling segments. The offset memory of the previous sampling segment is retained by the update gate, and the participation of the current sampling segment in the candidate hidden state is adjusted by the reset gate to obtain the forward recursive coding result. S242. A bidirectional gated loop unit is used to perform recursive encoding in reverse order of the sampling segments. The offset memory of the next sampling segment is retained by the update gate, and the participation of the current sampling segment in the candidate hidden state is adjusted by the reset gate to obtain the reverse recursive encoding result. S243. According to the same sampling segment position, the forward recursive coding result and the reverse recursive coding result are fused accordingly, and the coding components with conflicting directions are weakened based on the offset consistency between the two to obtain the joint coding result. S244. Bottleneck compression and scale constraint are applied to the joint coding results to retain the dominant change components in the state offset trajectory and to limit the compressed coding results to a preset state range, forming a scene state vector.

6. The engine crankshaft stress analysis method based on digital twin according to claim 1, characterized in that, S3 specifically includes: S31. In the benchmark analysis scenario, the finite element calculation process is decomposed into parameters. The model conditions that change with the analysis scenario are set as adjustable parameter items. Multiple sets of parameter samples are constructed around the adjustable parameter items, so that each set of parameter samples corresponds to a response solution record in the finite element calculation process. The parameter samples are formed in a way that each adjustable parameter item has no less than 10 sets of values. S32. Based on multiple sets of parameter samples and corresponding response solution records, train a radial basis function surrogate model so that the radial basis function surrogate model represents the computational correlation between the adjustable parameter terms and the twin response, wherein: The radial basis function surrogate model is a nonlinear approximation model that uses parameter samples as input variables and response solution records as target variables. The radial basis function surrogate model determines the contribution weight of each known parameter sample to the estimated twin response by calculating the radial distance between the sample of the parameter to be estimated and each known parameter sample, and obtains the estimated twin response based on the combination of contribution weights. When the average response error of the radial basis function surrogate model on the validation samples is greater than 8%, increase the number of parameter samples and retrain the radial basis function surrogate model. S33. Using the variational autoencoder method, latent variables are embedded into the scene state vector to obtain the distribution representation of the current analysis scene in the latent variable space. The latent variable values ​​corresponding to the current analysis scene are determined by reparameterized sampling. The dimension of the latent variable space is set according to one-quarter to one-half of the dimension of the scene state vector. S34. Based on the values ​​of the latent variables, the range of adjustable parameter terms is narrowed to form the parameter candidate space corresponding to the current analysis scenario. The parameter candidate space is then used as the parameter domain of the radial basis function surrogate model to obtain the twin calibration model for the current analysis scenario.

7. The engine crankshaft stress analysis method based on digital twin according to claim 1, characterized in that, S4 specifically includes: S41. Construct a Sobol sampling matrix in the parameter candidate space, such that each set of sampling results consists of the values ​​of adjustable parameter terms, and configure each set of sampling results into the Siamese calibration model to obtain the corresponding Siamese response values, wherein: The twin response value is: the numerical result obtained by the twin calibration model to perform response calculation under the corresponding sampling result, and the twin response value is associated with the corresponding sampling result one by one, and is used to calculate the variance contribution of the adjustable parameter term in subsequent calculations. S42. Based on the sampling results of each group and the corresponding twin response values, perform variance decomposition, calculate the first-order sensitivity index and the total effect sensitivity index corresponding to each adjustable parameter, and obtain the variance contribution results of each adjustable parameter to the twin response change. S43. Based on the first-order sensitivity index and the total effect sensitivity index, the adjustable parameter items that meet the preset screening conditions for variance contribution results are combined to obtain the sensitivity parameter combination corresponding to the current analysis scenario. S44. Based on the parameter candidate space, the boundary constraints of the sensitive parameter combination are reconstructed, and the reconstructed sensitive parameter combination is embedded into the twin calibration model to perform virtual inference in the current analysis scenario.

8. The engine crankshaft stress analysis method based on digital twin according to claim 7, characterized in that, S42 specifically includes: S421. Split the Sobol sampling matrix into a first sampling matrix and a second sampling matrix. Based on the first sampling matrix and the second sampling matrix, reconstruct the adjustable parameter terms of the twin calibration model and calculate the corresponding first response sequence and second response sequence. S422. For each adjustable parameter item, replace the sampling column corresponding to the adjustable parameter item in the first sampling matrix with the same sampling column in the second sampling matrix to obtain the cross-sampling matrix corresponding to the adjustable parameter item, and calculate the cross-response sequence based on the cross-sampling matrix. S423. Based on the response differences between the first response sequence, the second response sequence, and the cross-response sequence, calculate the independent variance contribution and total effect variance contribution of the adjustable parameter term to the twin response change. S424. Based on the independent variance contribution and the total effect variance contribution, the values ​​are proportionalized according to the total variance of the twinned response to obtain the first-order sensitivity index and the total effect sensitivity index corresponding to the adjustable parameter term.

9. The engine crankshaft stress analysis method based on digital twin according to claim 1, characterized in that, S5 specifically includes: S51. Based on the virtual simulation results, align the experimental response data with the same sampling positions, calculate the response difference at the same sampling position, and statistically analyze the absolute amplitude change and continuous variation amplitude of the response difference within 3 to 5 adjacent sampling positions to construct a feedback evaluation quantity, wherein: The test response data consists of the measured response sequence obtained through physical experiments under the current analysis scenario, and the test response data and the virtual simulation results are established in a one-to-one correspondence according to the sampling location. The response difference is the numerical difference between the virtual simulation result and the experimental response data at the same sampling location. S52. Based on the feedback evaluation, construct the reward value of the DDPG reinforcement learning method. Set the reward value as the sum of the decrease in the absolute magnitude of the response difference and the decrease in the magnitude of continuous change, so that the reward value represents the degree of reduction of the response bias by the current round of reparameterization adjustment. S53. The current value of the feedback evaluation quantity and the sensitive parameter combination is used as the state quantity of the DDPG reinforcement learning method. The policy network outputs the continuous action quantity, and the continuous action quantity is parsed into the reparameterization direction and adjustment magnitude of the sensitive parameter combination. S54. Based on the parameter candidate space, the adjustment range is subject to boundary constraints, so that the adjustment range of the sensitive parameter combination in each iteration is controlled within 10% to 20% of the corresponding value range. The sensitive parameter combination is then corrected according to the reparameterization direction and the adjustment range after boundary constraints, and the corrected sensitive parameter combination is updated to the adjustable parameter terms of the twin calibration model.

10. The engine crankshaft stress analysis method based on digital twin according to claim 1, characterized in that, S6 specifically includes: S61. Take the updated adjustable parameter as the current parameter state and the previously corrected adjustable parameter as the prior parameter state. Based on the offset between the current parameter state and the prior parameter state, construct the state estimation object of the unscented Kalman filter method. S62. Construct Sigma sampling points around the state estimation object, and substitute the adjustable parameter terms corresponding to each Sigma sampling point into the Siamese calibration model to calculate the prediction feedback evaluation quantity corresponding to each Sigma sampling point. S63. Based on the deviation between each predicted feedback evaluation quantity and the feedback evaluation quantity, calculate the state correction gain, and correct the current parameter state through the state correction gain to obtain the corrected adjustable parameter item. S64. When the corrected feedback evaluation quantity meets the preset convergence condition, the corrected adjustable parameter item is determined as the calibration parameter state of the current analysis scenario, and the engine crankshaft stress analysis result is obtained based on the calibration parameter state.

Images