An ai-enabled data-driven simulation correction and failure prediction method

Abstract

The application relates to an AI-enabled data-driven simulation correction and fault prediction method, belonging to the technical field of vehicle engineering. The method comprises the following steps: obtaining computer-aided engineering historical simulation data, constructing a multi-task artificial intelligence prediction architecture, combining a transfer learning mechanism, adjusting new design variant simulation data to form an adaptive durability performance prediction model; establishing a closed-loop correction mechanism, locating the error source through historical simulation and measured data deviation analysis, dynamically adjusting the correction parameters to optimize the simulation results; constructing a cross-modal attention fusion network, combining a time-space alignment mechanism to fuse sensor time series data and simulation stress spectrum; finally, based on the fusion data, an end-to-end neural network model is constructed to simultaneously complete fault type determination, position positioning and severity evaluation, and is associated with design data. The method realizes rapid adaptation of new design variants, improves simulation accuracy and fault diagnosis accuracy, and provides accurate basis for design optimization.

Description

Technical Field

[0001] This invention belongs to the field of vehicle engineering technology, specifically relating to an AI-enabled data-driven simulation correction and fault prediction method. Background Technology

[0002] In the field of vehicle engineering technology, vehicle durability directly determines the vehicle's driving safety, reliability, and service life, and is one of the core evaluation indicators in the vehicle design phase. At the same time, fault diagnosis and timely design optimization during vehicle operation are crucial for reducing maintenance costs and enhancing product competitiveness. As the automotive industry develops towards intelligence and customization, vehicle design variations are increasing. Traditional R&D models based on experience and physical testing are no longer sufficient to meet the needs of efficient and accurate design. Computer-aided engineering (CAE) simulation technology, with its advantages of low cost and high efficiency, is widely used in vehicle durability performance prediction and fault-related simulation analysis.

[0003] Currently, CAE simulation technology has become a core supporting tool for vehicle design and development. By constructing simulation models of vehicle components or the entire vehicle and inputting design parameters, load conditions, boundary conditions, and other information, it is possible to predict vehicle durability performance and conduct preliminary analysis of failure mechanisms. However, traditional CAE simulation technology and existing related improvement technologies still have many technical pain points that urgently need to be addressed, specifically in the following aspects: First, traditional CAE simulation models have extremely poor adaptability to new design variants. During the vehicle design process, a large number of new design variants will emerge for different market demands, power types (such as fuel vehicles and electric vehicles), or configuration parameters. Existing CAE simulation models are mostly built for specific design schemes. When faced with new design variants, engineers need to readjust model parameters, verify simulation logic, and even rebuild the simulation model, resulting in a lengthy simulation cycle and low R&D efficiency. Although some technologies have introduced simple parametric modeling methods, they still cannot break through the "one-to-one" design-simulation correspondence mode, making it difficult to quickly adapt to the needs of multiple types and batches of new design variants. Secondly, the inaccuracy of cross-modal data fusion makes it difficult to support efficient fault diagnosis. Vehicle fault diagnosis requires the combination of two types of core data: one is the time-series data collected by sensors during vehicle operation (such as vibration, temperature, pressure, etc.), and the other is the spatial dimension data obtained from CAE simulation (such as stress maps, displacement distribution, etc.). Existing technologies have obvious problems with inaccurate time-space alignment when processing these two types of cross-modal data. They mostly use a fixed window alignment method, which cannot be dynamically adjusted according to the complexity of the working conditions. At the same time, the data fusion process only performs simple feature stitching and does not establish a deep correlation between time-series features and spatial features. As a result, the fused data cannot fully extract key fault-related information, thus affecting the accuracy of fault diagnosis. Third, existing AI prediction models suffer from poor generalization ability and low training efficiency. Although some technologies have introduced AI into CAE simulation and fault diagnosis, existing models mostly adopt a single-task architecture, which cannot simultaneously meet the needs of multiple objectives such as durability performance prediction and fault diagnosis. The uniform sampling strategy used in the model training process does not pay enough attention to data from niche or extreme working conditions, resulting in low prediction accuracy under special working conditions. At the same time, when migrating to new design variants, catastrophic forgetting is prone to occur, and the effective parameters of the historical model cannot be effectively inherited. A large amount of new data is required for retraining, which further reduces R&D efficiency.

[0004] In summary, current technologies in the field of vehicle engineering, such as CAE simulation correction, durability performance prediction, fault diagnosis, and design optimization, still suffer from many shortcomings, including poor adaptability, insufficient accuracy, low efficiency, and disconnected links, making it difficult to meet the needs of rapid vehicle development and high-quality assurance in the context of intelligent and customized vehicles. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides an AI-enabled data-driven simulation correction and fault prediction method. The objective of this invention can be achieved through the following technical solutions: include: S1: Obtain historical simulation data of computer-aided engineering, and construct a multi-task artificial intelligence prediction architecture based on the extracted design parameters, load conditions, boundary conditions and corresponding simulation results; use the sorted historical data to train the basic prediction model, introduce the transfer learning mechanism and adjust the simulation data of the new design variant to form a durability performance prediction model that matches the new design variant. S2: Establish a closed-loop correction mechanism to locate the source of error by analyzing the deviation between historical simulation results and measured data; input measured data of the new design variant to perform preliminary correction on the simulation results; and dynamically adjust the correction parameters by comparing the preliminary correction results with the measured data a second time. S3: Acquire sensor time-series data and corrected simulated stress maps during vehicle operation, and construct a cross-modal attention fusion network; extract the temporal dimension features of the sensor time-series data and the spatial dimension features of the simulated stress maps respectively, and synchronously match the temporal dimension features and spatial dimension features through a spatiotemporal alignment mechanism; establish a deep association between the temporal dimension features and spatial dimension features based on the cross-attention mechanism, and finally generate fused data; S4: Based on the fused data, construct an end-to-end neural network model; determine the fault type, locate the fault location, and assess the severity of the fault through the end-to-end neural network model; at the same time, associate the computer-aided engineering simulation data from the design phase, and optimize the new design variant based on the analysis results.

[0006] Specifically, the process of constructing the multi-task artificial intelligence prediction architecture is as follows: Based on the organized design parameters, load conditions, boundary conditions, and corresponding simulation results, record the common correlations between different data dimensions; The logic of the correspondence between preset design parameters and durability performance, load conditions and simulation results; An architecture is constructed that includes a shared feature extraction layer and a task-specific layer. The shared feature extraction layer extracts general correlation patterns from various types of data, while the task-specific layer refines feature processing based on the characteristics of different components.

[0007] Specifically, the process of training the basic prediction model using the processed historical data is as follows: The historical simulation data is preprocessed to remove abnormal data and redundant information. The processed data is divided into training data and validation data according to a preset ratio. The training data is input into the multi-task artificial intelligence prediction architecture, and the network weights and thresholds are adjusted through iterative calculation. After each iteration, the prediction error is calculated, and the training step size is dynamically adjusted based on the prediction error feedback. The accuracy of the prediction is monitored in real time using validation data until the deviation between the prediction result and the actual simulation result is within a preset range.

[0008] Specifically, the process of forming a durability performance prediction model that matches the new design variant is as follows: The inputs are design parameters, load conditions and boundary conditions, and the output is the predicted durability performance of the component. It includes: an input layer, a shared feature extraction layer, a task-specific layer, and an output layer; The input layer receives and standardizes data, converting unstructured data into a standard processing format; the shared feature extraction layer captures the general correlation patterns between data; the task-specific layer specifically enhances key features related to component durability; and the output layer maps the refined features output by the task-specific layer into quantitative indicators of durability.

[0009] Specifically, the process of locating the source of error through deviation analysis of historical simulation results and measured data is as follows: historical simulation results and corresponding measured data are compared one by one according to the same working conditions and time nodes, the deviation value of each group of data is calculated, and a deviation set is formed; the load range, boundary condition type and data dimension of the deviation set are preset; and the core factors causing the deviation are traced by combining the assumptions, simplified logic and material parameter settings of the durability performance prediction model.

[0010] Specifically, the process of dynamically adjusting the correction parameters is as follows: set the allowable range of deviation between the preliminary correction result and the measured data, compare the simulation result after preliminary correction with the measured data, if the deviation exceeds the allowable range, determine the parameter adjustment range according to the magnitude and direction of the deviation, prioritize the adjustment of correction parameters directly related to the source of error, perform simulation correction again after adjustment, compare the results again, record the parameter value and the corresponding deviation change for each adjustment, and iterate repeatedly until the deviation is within the allowable range.

[0011] Specifically, the process of constructing the cross-modal attention fusion network is as follows: It includes a feature extraction layer, an attention association layer, and a fusion output layer; The feature extraction layer captures the temporal dimension change features of the sensor time series data and extracts the spatial distribution features of the simulated stress spectrum; the attention association layer calculates the association weight between the temporal dimension features and the spatial dimension features, strengthens the key information related to fault prediction and weakens the secondary interference; the fusion output layer logically integrates the associated temporal features and spatial features to generate fused data that has both temporal change patterns and spatial distribution information.

[0012] Specifically, the process of synchronously matching time-dimensional features and spatial-dimensional features through the spatiotemporal alignment mechanism is as follows: Based on vehicle operating condition labels, establish the correspondence between time-dimensional features and spatial-dimensional features; The time dimension features are segmented according to the working condition cycle, and the spatial dimension features are marked according to the corresponding working condition. The time segment and spatial feature region corresponding to each working condition are marked. For feature segments with asynchronous time, linear interpolation is used to supplement missing feature points or truncate redundant feature segments, so that the time dimension features and spatial dimension features under the same working condition are matched one by one.

[0013] Specifically, the process of establishing a deep association between time-dimensional features and spatial-dimensional features is as follows: based on the time-dimensional features and spatial-dimensional features after synchronous matching, calculate the association weight between the time-dimensional features and spatial-dimensional features; calculate the mutual information value of each feature dimension, allocate association weights according to the mutual information value, strengthen the association of key features related to fault prediction, and weaken the interference of secondary features; combine the temporal change information in the time-dimensional features with the location distribution information in the spatial-dimensional features through weighted fusion.

[0014] Specifically, the process of constructing the end-to-end neural network model is as follows: Using fused data as input, the output targets are fault type determination, fault location, and fault severity assessment. It includes an input matching layer, a deep mining layer, a cross-task association layer, and a multi-branch output layer; The input matching layer performs modality unification and dimension normalization on the fused data. The deep mining layer extracts multi-scale deep fault features through dilated convolution and residual connections. The cross-task association layer establishes feature interaction links for the prediction target. The multi-branch output layer presets activation functions and loss terms according to task characteristics and outputs prediction results synchronously.

[0015] Specifically, the process of determining the fault type, locating the fault location, and assessing the severity of the fault through the end-to-end neural network model is as follows: inputting the fused data into the end-to-end neural network model and pre-setting typical feature templates for various types of faults; using the first branch network to combine the fault features and compare them with the pre-set fault feature templates to determine the fault type; the second branch network generates and fuses multi-scale fault heat maps, and combines them with the spatial information of the simulated stress spectrum to locate specific areas; the third regression branch uses a hierarchical regression strategy to output a quantitative value of the fault severity.

[0016] Specifically, the process of optimizing the new design variant based on the analysis results is as follows: based on the analysis results of fault type, fault location and severity, correlate the computer-aided engineering simulation data of the design stage, trace the design parameters and process settings related to the fault; formulate targeted adjustment schemes for different problems, input the optimized design parameters into the durability performance prediction model for verification, confirm that the fault risk has been reduced to the preset level, and verify the stability of the optimization effect.

[0017] The beneficial effects of this invention are as follows: (1) By setting up a multi-task artificial intelligence prediction architecture and transfer learning mechanism, combined with a closed-loop correction mechanism, on the one hand, the multi-task architecture can be used to explore the common correlation and unique characteristics between data such as design parameters and load conditions. Then, by reusing historical model parameters through transfer learning, only a small amount of simulation data of the new design variant needs to be collected for fine-tuning to form an adapted durability performance prediction model. There is no need to rebuild the simulation model for the new design variant, which greatly reduces the time cost of model building and training and significantly shortens the vehicle development cycle. On the other hand, the closed-loop correction mechanism forms a deviation set by comparing historical simulation and measured data under each working condition, accurately traces the core cause of the deviation, and then dynamically adjusts the correction parameters according to the magnitude and direction of the deviation. Through repeated iterations, it ensures that the simulation deviation is within the allowable range, effectively solving the problem of insufficient accuracy caused by the idealized assumptions of traditional simulation models. (2) By setting up a cross-modal attention fusion network and a spatiotemporal alignment mechanism, combined with an end-to-end neural network model containing an input matching layer, a deep mining layer, a cross-task association layer and a multi-branch output layer, the spatiotemporal alignment mechanism is first used to achieve accurate synchronous matching between sensor time series data and simulated stress maps based on vehicle operating condition labels. Through segmentation, labeling and interpolation / truncation, etc., the mutual information values ​​between features are calculated and association weights are assigned through a cross-attention mechanism to strengthen key fault-related information and weaken secondary interference, generating high-quality fusion data that combines temporal variation patterns and spatial distribution information. Secondly, the deep mining layer of the end-to-end model extracts multi-scale deep fault features through dilated convolution and residual connections. The cross-task association layer establishes a feature interaction link for fault type determination, location positioning and severity assessment, avoiding the limitations of independent training for a single task and achieving synchronous and accurate output of the three types of targets. Finally, the CAE simulation data in the association design stage is used to trace the cause of the fault, formulate a targeted adjustment plan and verify it through the durability performance prediction model to form a closed loop of the whole link. Attached Figure Description

[0018] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart illustrating an AI-enabled data-driven simulation correction and fault prediction method according to the present invention. Figure 2 This is a data flow diagram of an AI-enabled data-driven simulation correction and fault prediction method according to the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0021] Please see Figure 1-2 A data-driven simulation correction and fault prediction method empowered by AI; include: S1: Obtain historical simulation data of computer-aided engineering, and construct a multi-task artificial intelligence prediction architecture based on the extracted design parameters, load conditions, boundary conditions and corresponding simulation results; use the sorted historical data to train the basic prediction model, introduce the transfer learning mechanism and adjust the simulation data of the new design variant to form a durability performance prediction model that matches the new design variant. S2: Establish a closed-loop correction mechanism to locate the source of error by analyzing the deviation between historical simulation results and measured data; input measured data of the new design variant to perform preliminary correction on the simulation results; and dynamically adjust the correction parameters by comparing the preliminary correction results with the measured data a second time. S3: Acquire sensor time-series data and corrected simulated stress maps during vehicle operation, and construct a cross-modal attention fusion network; extract the temporal dimension features of the sensor time-series data and the spatial dimension features of the simulated stress maps respectively, and synchronously match the temporal dimension features and spatial dimension features through a spatiotemporal alignment mechanism; establish a deep association between the temporal dimension features and spatial dimension features based on the cross-attention mechanism, and finally generate fused data; S4: Based on the fused data, construct an end-to-end neural network model; determine the fault type, locate the fault location, and assess the severity of the fault through the end-to-end neural network model; at the same time, associate the computer-aided engineering simulation data from the design phase, and optimize the new design variant based on the analysis results.

[0022] In this embodiment, the computer-aided engineering historical simulation data specifically includes: simulation model files of different vehicle design variants, simulation calculation process data, simulation result output data, as well as supporting working condition setting records, parameter adjustment logs, etc., covering historical data of various simulation scenarios such as durability performance simulation and mechanical performance simulation of key components such as body, chassis, and power system.

[0023] In this embodiment, design parameters refer to the key technical parameters that affect durability and failure risk, determined during the design phase of the vehicle and its core components. They are one of the core inputs for building simulation models and predictive architectures. Specifically, they include: vehicle level: wheelbase, track width, vehicle weight, center of gravity height, etc.; component level: dimensional parameters of the body frame, material grade and mechanical performance parameters (such as tensile strength and modulus of elasticity), spring stiffness / damping coefficient of the chassis suspension system, gear module / number of teeth of transmission components, structural dimensions and fixing method of battery pack (electric vehicle), etc.; process-related parameters: strength parameters of welding / bolted connections, component assembly clearances, etc.

[0024] In this embodiment, load conditions refer to the types of loads and corresponding operating scenarios that the core components of a vehicle experience during actual operation or simulation analysis. These directly determine the relevance of the simulation analysis and the effectiveness of fault prediction. Specifically, these include: normal operating conditions: loads corresponding to constant speed driving (different vehicle speeds), acceleration / deceleration, and normal turning (such as engine output torque, tire ground load, and suspension stress); special operating conditions: extreme load scenarios such as climbing (different slopes), bumpy roads (different road grades), emergency braking / emergency steering, and full load / overload; and load parameter details: the magnitude, direction, point of application, frequency of change, and duration of the load (such as the frequency range of periodic vibration loads and the peak value of impact loads).

[0025] In this embodiment, boundary conditions refer to the preconditions set during CAE simulation to clarify the constraint state and environmental influences of the simulation model, which are the basis for ensuring the rationality of the simulation results. Specifically, they include: constraint conditions: the fixing method of the components (such as the connection constraint between the vehicle body and the frame, the rolling constraint of the tires), and degree of freedom restrictions (such as the rotational degree of freedom of rotating components, and the displacement restriction of fixed components); environmental conditions: the ambient temperature (such as low temperature -30℃, high temperature 60℃), humidity, air pressure, and the influence of special environments (such as salt spray, dust) on the components; contact conditions: the contact type between components (such as rigid contact, elastic contact), and the coefficient of friction (such as the coefficient of friction between the brake disc and the brake pad).

[0026] In this embodiment, the simulation results specifically include: durability-related results: predicted fatigue life, cumulative damage, maximum stress / strain distribution data, and fatigue failure risk level of core components; mechanical performance-related results: deformation, stiffness / strength calculation, and vibration modal parameters (such as natural frequency and mode shape) of components; and fault-related results: predicted fault risk areas, potential fault types (such as cracks and excessive deformation), and critical loads / times for fault occurrence during the simulation.

[0027] Specifically, the process of constructing the multi-task artificial intelligence prediction architecture is as follows: Based on the organized design parameters, load conditions, boundary conditions, and corresponding simulation results, record the common correlations between different data dimensions; The logic of the correspondence between preset design parameters and durability performance, load conditions and simulation results; An architecture is constructed that includes a shared feature extraction layer and a task-specific layer. The shared feature extraction layer extracts general correlation patterns from various types of data, while the task-specific layer refines feature processing based on the characteristics of different components.

[0028] Specifically, the process of training the basic prediction model using the processed historical data is as follows: The historical simulation data is preprocessed to remove abnormal data and redundant information. The processed data is divided into training data and validation data according to a preset ratio. The training data is input into the multi-task artificial intelligence prediction architecture, and the network weights and thresholds are adjusted through iterative calculation. After each iteration, the prediction error is calculated, and the training step size is dynamically adjusted based on the prediction error feedback. The accuracy of the prediction is monitored in real time using validation data until the deviation between the prediction result and the actual simulation result is within a preset range.

[0029] Specifically, the process of forming a durability performance prediction model that matches the new design variant is as follows: The inputs are design parameters, load conditions and boundary conditions, and the output is the predicted durability performance of the component. It includes: an input layer, a shared feature extraction layer, a task-specific layer, and an output layer; The input layer receives and standardizes data, converting unstructured data into a standard processing format; the shared feature extraction layer captures the general correlation patterns between data; the task-specific layer specifically enhances key features related to component durability; and the output layer maps the refined features output by the task-specific layer into quantitative indicators of durability.

[0030] Specifically, the process of locating the source of error through deviation analysis of historical simulation results and measured data is as follows: historical simulation results and corresponding measured data are compared one by one according to the same working conditions and time nodes, the deviation value of each group of data is calculated, and a deviation set is formed; the load range, boundary condition type and data dimension of the deviation set are preset; and the core factors causing the deviation are traced by combining the assumptions, simplified logic and material parameter settings of the durability performance prediction model.

[0031] Specifically, the process of dynamically adjusting the correction parameters is as follows: set the allowable range of deviation between the preliminary correction result and the measured data, compare the simulation result after preliminary correction with the measured data, if the deviation exceeds the allowable range, determine the parameter adjustment range according to the magnitude and direction of the deviation, prioritize the adjustment of correction parameters directly related to the source of error, perform simulation correction again after adjustment, compare the results again, record the parameter value and the corresponding deviation change for each adjustment, and iterate repeatedly until the deviation is within the allowable range.

[0032] Specifically, the process of constructing the cross-modal attention fusion network is as follows: It includes a feature extraction layer, an attention association layer, and a fusion output layer; The feature extraction layer captures the temporal dimension change features of the sensor time series data and extracts the spatial distribution features of the simulated stress spectrum; the attention association layer calculates the association weight between the temporal dimension features and the spatial dimension features, strengthens the key information related to fault prediction and weakens the secondary interference; the fusion output layer logically integrates the associated temporal features and spatial features to generate fused data that has both temporal change patterns and spatial distribution information.

[0033] Specifically, the process of synchronously matching time-dimensional features and spatial-dimensional features through the spatiotemporal alignment mechanism is as follows: Based on vehicle operating condition labels, establish the correspondence between time-dimensional features and spatial-dimensional features; The time dimension features are segmented according to the working condition cycle, and the spatial dimension features are marked according to the corresponding working condition. The time segment and spatial feature region corresponding to each working condition are marked. For feature segments with asynchronous time, linear interpolation is used to supplement missing feature points or truncate redundant feature segments, so that the time dimension features and spatial dimension features under the same working condition are matched one by one.

[0034] Specifically, the process of establishing a deep association between time-dimensional features and spatial-dimensional features is as follows: based on the time-dimensional features and spatial-dimensional features after synchronous matching, calculate the association weight between the time-dimensional features and spatial-dimensional features; calculate the mutual information value of each feature dimension, allocate association weights according to the mutual information value, strengthen the association of key features related to fault prediction, and weaken the interference of secondary features; combine the temporal change information in the time-dimensional features with the location distribution information in the spatial-dimensional features through weighted fusion.

[0035] Specifically, the process of constructing the end-to-end neural network model is as follows: Using fused data as input, the output targets are fault type determination, fault location, and fault severity assessment. It includes an input matching layer, a deep mining layer, a cross-task association layer, and a multi-branch output layer; The input matching layer performs modality unification and dimension normalization on the fused data. The deep mining layer extracts multi-scale deep fault features through dilated convolution and residual connections. The cross-task association layer establishes feature interaction links for the prediction target. The multi-branch output layer presets activation functions and loss terms according to task characteristics and outputs prediction results synchronously.

[0036] Specifically, the process of determining the fault type, locating the fault location, and assessing the severity of the fault through the end-to-end neural network model is as follows: inputting the fused data into the end-to-end neural network model and pre-setting typical feature templates for various types of faults; using the first branch network to combine the fault features and compare them with the pre-set fault feature templates to determine the fault type; the second branch network generates and fuses multi-scale fault heat maps, and combines them with the spatial information of the simulated stress spectrum to locate specific areas; the third regression branch uses a hierarchical regression strategy to output a quantitative value of the fault severity.

[0037] Specifically, the process of optimizing the new design variant based on the analysis results is as follows: based on the analysis results of fault type, fault location and severity, correlate the computer-aided engineering simulation data of the design stage, trace the design parameters and process settings related to the fault; formulate targeted adjustment schemes for different problems, input the optimized design parameters into the durability performance prediction model for verification, confirm that the fault risk has been reduced to the preset level, and verify the stability of the optimization effect.

[0038] In this embodiment, the formation process of the durability performance prediction model is as follows: Historical data preprocessing: First, CAE historical simulation data from the vehicle development process over many years are collected, covering various vehicle models and core components (such as body longitudinal beams, front suspension control arms, drive half shafts, etc.), with sufficient data volume; the data is standardized to normalize design parameters, load conditions and other data to a reasonable range, a small number of abnormal data are removed using the 3σ criterion, and missing data points are supplemented by linear interpolation. Basic prediction model training: Based on preprocessed historical data, a multi-task architecture is built, including an input layer (neurons adapted to data dimensions), a shared feature extraction layer (multi-level decreasing fully connected layers), a task-specific layer (divided into multiple branches according to component type), and an output layer. The Adam optimizer is used, with the learning rate initially set to a normal initial value and gradually decaying according to preset rules. The training set and validation set are divided in a normal ratio. L2 regularization is introduced during training (setting an appropriate regularization coefficient) to avoid overfitting. Training stops when the prediction deviation on the validation set is less than a preset threshold for several consecutive rounds, thus obtaining the basic prediction model. New Design Variant Model Adaptation: For a core component design variant of a newly developed vehicle model, model parameters of corresponding components from several historical models with high structural similarity to the variant are selected as initial transfer parameters; a small amount of key operating condition simulation data (including multiple typical operating conditions) of the new design variant is collected, and an incremental fine-tuning strategy is adopted to reduce the learning rate and perform a small number of rounds of parameter optimization, ultimately forming a durability performance prediction model adapted to the new design variant. The model adaptation time is significantly shortened compared to the traditional remodeling method. In this embodiment, the specific construction steps of the cross-modal attention fusion network are as follows: Network architecture construction: The network input is defined as two types of cross-modal data (sensor time series data and simulation stress spectrum), and the output is fused data. A three-layer architecture including feature extraction layer, attention association layer and fusion output layer is constructed. Feature extraction layer construction: The feature extraction layer adopts a parallel dual-branch structure. The temporal feature branch uses a 1D convolution kernel to extract features from sensor time-series data (including various signals such as vibration, temperature, and pressure) with a conventional sampling frequency and duration, and outputs a temporal feature map with an appropriate dimension. The spatial feature branch uses a 2D convolution kernel to perform multiple rounds of convolution operations on simulated stress maps with a conventional resolution, and outputs a spatial feature map with an appropriate dimension. Attention association layer construction: A bidirectional cross-attention mechanism is adopted to calculate the association weight matrix between the two types of features, set a reasonable mutual information value threshold, retain feature association pairs with high mutual information values, strengthen fault-related key features (such as vibration peak time sequence features and stress concentration spatial features), and weaken the interference of low-weight secondary features. Construction of the fusion output layer: An adaptive weighted fusion strategy is adopted, and the fusion ratio is dynamically allocated according to the feature association weight. The feature pairs with high association weight and medium association weight are respectively given an appropriate fusion ratio. Finally, the fused data with the appropriate dimension is output. The fault feature identification of the fused data is significantly improved compared with the single modality data.

[0039] In this embodiment, the specific steps for constructing the end-to-end neural network model are as follows: The input is clearly defined as cross-modal fusion data, and the output is three types of targets: fault type (multiple common faults), fault location (accuracy meets engineering requirements), and fault severity (multi-level quantization). A four-layer architecture is built, including an input matching layer, a deep mining layer, a cross-task association layer, and a multi-branch output layer. Input matching layer construction: Modality unification (converting temporal-spatial hybrid features into tensor format) and dimensionality normalization are performed on the input fusion data, and a data cache queue is set up to improve data input efficiency; Deep mining layer construction: A structure combining dilated convolution and residual connections is adopted, with multiple dilated convolution blocks (the dilation rate increases according to a preset rule), and the convolution kernels are of a standard size. Each convolution block is followed by a BatchNorm layer and a residual skip connection to extract multi-scale deep fault features. Cross-task association layer construction: build a feature interaction matrix, establish feature sharing links for three types of prediction targets, and adopt a reasonable attention weight allocation mechanism; Multi-branch output layer construction: Dedicated modules are configured according to task characteristics. The fault type branch uses the Softmax activation function and the cross-entropy loss function; the fault location branch uses the Sigmoid activation function and the Dice loss function; the severity branch uses the Linear activation function and the mean squared error loss function. By reasonably setting the weights of each loss term, the three types of prediction results are output synchronously, and the model inference time meets the real-time requirements.

[0040] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An AI-enabled data-driven simulation correction and fault prediction method, characterized in that, include: S1: Obtain historical simulation data of computer-aided engineering, and construct a multi-task artificial intelligence prediction architecture based on the extracted design parameters, load conditions, boundary conditions and corresponding simulation results; A basic prediction model is trained using the sorted historical data. A transfer learning mechanism is introduced and a durability performance prediction model matching the new design variant is formed by adjusting the simulation data of the new design variant. S2: Establish a closed-loop correction mechanism to locate the source of error by analyzing the deviation between historical simulation results and measured data; input measured data of the new design variant to perform preliminary correction on the simulation results; and dynamically adjust the correction parameters by comparing the preliminary correction results with the measured data a second time. S3: Acquire sensor time-series data and corrected simulated stress maps during vehicle operation, and construct a cross-modal attention fusion network; extract the temporal dimension features of the sensor time-series data and the spatial dimension features of the simulated stress maps respectively, and synchronously match the temporal dimension features and spatial dimension features through a spatiotemporal alignment mechanism; establish a deep association between the temporal dimension features and spatial dimension features based on the cross-attention mechanism, and finally generate fused data; S4: Based on the fused data, construct an end-to-end neural network model; determine the fault type, locate the fault location, and assess the severity of the fault through the end-to-end neural network model; at the same time, associate the computer-aided engineering simulation data from the design phase, and optimize the new design variant based on the analysis results.

2. The method according to claim 1, characterized in that, The specific process of constructing the multi-task artificial intelligence prediction architecture is as follows: Based on the organized design parameters, load conditions, boundary conditions, and corresponding simulation results, record the common correlations between different data dimensions; The logic of the correspondence between preset design parameters and durability performance, load conditions and simulation results; An architecture is constructed that includes a shared feature extraction layer and a task-specific layer. The shared feature extraction layer extracts general correlation patterns from various types of data, while the task-specific layer refines feature processing based on the characteristics of different components.

3. The method according to claim 1, characterized in that, The specific process of training the basic prediction model using the processed historical data is as follows: The historical simulation data is preprocessed to remove abnormal data and redundant information. The processed data is divided into training data and validation data according to a preset ratio. The training data is input into the multi-task artificial intelligence prediction architecture, and the network weights and thresholds are adjusted through iterative calculation. After each iteration, the prediction error is calculated, and the training step size is dynamically adjusted based on the prediction error feedback. The accuracy of the prediction is monitored in real time using validation data until the deviation between the prediction result and the actual simulation result is within a preset range.

4. The method according to claim 1, characterized in that, The specific process for forming a durability performance prediction model that matches the new design variant is as follows: The inputs are design parameters, load conditions and boundary conditions, and the output is the predicted durability performance of the component. It includes: an input layer, a shared feature extraction layer, a task-specific layer, and an output layer; The input layer receives data and standardizes it, converting unstructured data into a standard processing format; A shared feature extraction layer captures general correlation patterns between data; a task-specific layer specifically enhances key features related to component durability. The output layer maps the refined features output from the task-specific layer into a durability performance quantification index.

5. The method according to claim 1, characterized in that, The specific process of locating the source of error through deviation analysis of historical simulation results and measured data is as follows: compare the historical simulation results with the corresponding measured data one by one according to the same working conditions and time nodes, calculate the deviation value of each group of data, and form a deviation set; preset the load range, boundary condition type and data dimension of the deviation set; and trace the core factors that cause the deviation by combining the assumptions, simplified logic and material parameter settings of the durability performance prediction model.

6. The method according to claim 1, characterized in that, The specific process of dynamically adjusting the correction parameters is as follows: set the allowable range of deviation between the preliminary correction result and the measured data, compare the simulation result after preliminary correction with the measured data, if the deviation exceeds the allowable range, determine the parameter adjustment range according to the magnitude and direction of the deviation, prioritize the adjustment of correction parameters directly related to the source of error, perform simulation correction again after adjustment, compare the results again, record the parameter value and the corresponding deviation change for each adjustment, and iterate repeatedly until the deviation is within the allowable range.

7. The method according to claim 1, characterized in that, The specific process of constructing the cross-modal attention fusion network is as follows: It includes a feature extraction layer, an attention association layer, and a fusion output layer; The feature extraction layer captures temporal variation features from sensor time-series data and extracts spatial distribution features from simulated stress maps. The attention association layer calculates the association weights between temporal and spatial features, strengthening key information related to fault prediction and weakening secondary interferences; the fusion output layer logically integrates the associated temporal and spatial features to generate fused data that combines temporal variation patterns and spatial distribution information.

8. The method according to claim 1, characterized in that, The specific process of synchronously matching time-dimensional features and spatial-dimensional features through the spatiotemporal alignment mechanism is as follows: Based on vehicle operating condition labels, establish the correspondence between time-dimensional features and spatial-dimensional features; The time dimension features are segmented according to the working condition cycle, and the spatial dimension features are marked according to the corresponding working condition. The time segment and spatial feature region corresponding to each working condition are marked. For feature segments with asynchronous time, linear interpolation is used to supplement missing feature points or truncate redundant feature segments, so that the time dimension features and spatial dimension features under the same working condition are matched one by one.

9. The method according to claim 1, characterized in that, The specific process for establishing a deep association between time-dimensional features and spatial-dimensional features is as follows: based on the time-dimensional features and spatial-dimensional features after synchronous matching, calculate the association weight between the time-dimensional features and spatial-dimensional features; calculate the mutual information value of each feature dimension, allocate association weights according to the mutual information value, strengthen the association of key features related to fault prediction, and weaken the interference of secondary features; combine the temporal change information in the time-dimensional features with the location distribution information in the spatial-dimensional features through weighted fusion.

10. The method according to claim 1, characterized in that, The specific process for constructing the end-to-end neural network model is as follows: Using fused data as input, the output targets are fault type determination, fault location, and fault severity assessment. It includes an input matching layer, a deep mining layer, a cross-task association layer, and a multi-branch output layer; The input matching layer performs modality unification and dimension normalization on the fused data. The deep mining layer extracts multi-scale deep fault features through dilated convolution and residual connections. The cross-task association layer establishes feature interaction links for the prediction target. The multi-branch output layer presets activation functions and loss terms according to task characteristics and outputs prediction results synchronously.

11. The method according to claim 1, characterized in that, The specific process of determining the fault type, locating the fault location, and assessing the severity of the fault through the end-to-end neural network model is as follows: inputting the fused data into the end-to-end neural network model and pre-setting typical feature templates for various types of faults; using the first branch network to combine the fault features and compare them with the pre-set fault feature templates to determine the fault type; the second branch network generates and fuses multi-scale fault heat maps, and combines them with the spatial information of the simulated stress spectrum to locate specific areas; the third regression branch uses a hierarchical regression strategy to output a quantitative value of the fault severity.

12. The method according to claim 1, characterized in that, The specific process of optimizing the new design variant based on the analysis results is as follows: based on the analysis results of fault type, fault location and severity, correlate the computer-aided engineering simulation data in the design phase, and trace the design parameters and process settings related to the fault. For different problems, targeted adjustment plans are formulated, and the optimized design parameters are input into the durability performance prediction model for verification. The failure risk is confirmed to be reduced to the preset level, and the stability of the optimization effect is verified.

Images