A Lightweight Design Method for New Energy Vehicle Body Frame Based on Digital Twin

Abstract

This invention discloses a lightweight design method for new energy vehicle body frames based on digital twins. The invention relates to the field of new energy vehicle body structure design and computer-aided engineering simulation technology, and includes the following steps: Step S1: Establishing a digital twin connected to the physical vehicle body sensing system; Step S2: A dynamic precision topology network module, used to dynamically divide the simulation precision region of the body frame according to the mechanical energy transfer path and real-time sensing data. This lightweight design method for new energy vehicle body frames based on digital twins achieves intelligent allocation of computing resources through a dynamic precision topology network. While ensuring accurate prediction of key structures, it effectively reduces the computational burden of multi-physics coupled simulation; it effectively resolves the sharp contradiction between model accuracy, real-time response, and optimization efficiency, enabling lightweight design to achieve effective closed-loop iteration under complex constraints.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle body structure design and computer-aided engineering simulation technology, specifically a lightweight design method for new energy vehicle body frames based on digital twins. Background Technology

[0002] In the field of new energy vehicles, lightweight design of the vehicle body frame is crucial for improving driving range and reducing energy consumption. Digital twin technology, due to its ability to construct virtual mappings of physical entities and achieve virtual-real interaction, provides a powerful tool for optimizing lightweight vehicle body frame design. Ideally, design methods based on high-fidelity digital twins should be able to continuously integrate real-time sensor data from the physical world, such as stress and temperature, and utilize accurate multi-physics coupled simulation models to efficiently perform multi-objective optimization iterations under numerous stringent performance constraints to find the optimal lightweight solution. However, existing digital twin-based lightweight vehicle body design methods face a fundamental bottleneck at the data processing level. Constructing and maintaining a high-precision, multi-physics coupled digital twin model itself requires enormous computational resources, and the simulation process is significantly time-consuming. At the same time, lightweight design itself is a complex multi-objective, multi-constraint optimization problem, and each design iteration requires calling these high-precision models for evaluation, further exacerbating the computational burden.

[0003] More importantly, to realize the core value of digital twins—virtual-real synchronization—the system also needs the ability to process and fuse real-time or near-real-time data streams from the physical vehicle body to dynamically update the twin's state and drive the optimization process. These three elements—maintaining high model fidelity for accurate prediction, rapid response and fusion of real-time data to maintain twin synchronization, and efficient execution of complex multi-objective optimization iterations—create a sharp contradiction within limited computing resources and time windows. Existing technologies often struggle to effectively coordinate these three aspects, forcing compromises on model accuracy, data update timeliness, or optimization efficiency. This results in delayed digital twin updates, lengthy design optimization cycles, and ultimately limits the speed, accuracy, and engineering practicality of lightweight design. Therefore, effectively resolving the contradiction between high-precision multiphysics simulation, real-time data fusion efficiency, and complex multi-objective lightweight optimization computation in digital twin models has become a critical technical challenge that this method urgently needs to overcome. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides a lightweight design method for a new energy vehicle body frame based on digital twins, comprising the following steps:

[0005] Step S1: Establish a digital twin that connects to the physical vehicle body sensing system;

[0006] Step S2: Dynamic precision topology network module, used to dynamically divide the simulation precision area of ​​the vehicle frame according to the mechanical energy transfer path and real-time sensor data;

[0007] Step S3: Dual-track drive optimization engine module, including real-time track updates and offline track optimization;

[0008] Step S4: The dual validation module performs residual screening and cross-model consistency checks on the output scheme.

[0009] Preferably, the execution of the dynamic precision topology network module includes:

[0010] The mechanical energy distribution of the vehicle frame is calculated using a stress flow density algorithm, and continuous regions with energy intensity exceeding a preset threshold are identified as topology-sensitive regions.

[0011] Cluster analysis of real-time strain sensing data identifies dynamic high-sensitivity regions where stress fluctuations exceed a set tolerance.

[0012] By overlaying the topology-sensitive domain with the dynamic high-sensitivity region, a key precision domain map covering the entire vehicle frame is generated.

[0013] The atlas divides the vehicle body frame into a full-precision zone, a medium-precision zone, and a proxy model zone.

[0014] Preferably, the full-precision region uses a complete finite element model, the medium-precision region uses a simplified shell element model, and the surrogate model region describes the structural response through parametric equations.

[0015] Preferably, in the dual-track drive optimization engine module:

[0016] The track is deployed on the edge computing node in real time, receives the physical vehicle body sensor data stream, predicts the mechanical state change of the full-precision area through a pre-trained neural network, and directly modifies the displacement of the finite element mesh node in that area.

[0017] The offline optimized trajectory is deployed on a cloud server, runs a multi-objective genetic algorithm to generate lightweight candidate solutions, and calls a Bayesian optimizer to adjust the geometric parameters of the candidate solutions based on the real-time updated full-precision region state of the trajectory output.

[0018] Preferably, the neural network is a hybrid architecture of long short-term memory network and three-dimensional convolutional network, with the input being time-series strain data and spatial temperature distribution, and the output being the displacement increment of six degrees of freedom in the full precision region.

[0019] Preferably, the dual verification module includes:

[0020] First residual screening: At fixed intervals, three sub-regions of the surrogate model area are randomly selected and switched to full-precision model calculation performance indicators. If the deviation from the surrogate model prediction exceeds the allowable tolerance, the surrogate model parameter self-learning is triggered and the optimization process is paused.

[0021] The second cross-model verification: For the final lightweight solution, full-precision virtual collision simulation and reinforcement learning-based surrogate model inference are performed simultaneously, and the differences between the two results in the maximum deformation of the vehicle body and the first-order modal frequency are compared.

[0022] Preferably, the reinforcement learning-based agent model is trained independently using historical high-precision simulation data and has no shared parameters with the main optimization system.

[0023] Preferably, the virtual collision simulation includes multiphysics coupling calculations under frontal collision, side collision, and torsional conditions.

[0024] Preferably, the method outputs a lightweight solution when the state update delay of the digital twin is lower than a set threshold and the time taken for a single optimization iteration is shorter than the baseline time.

[0025] Preferably, the key precision domain map is dynamically updated every two hours. The update triggering conditions include: the rate of change of statistical features of real-time sensing data exceeding a threshold value, or the optimization engine triggering the identification of new topology sensitive domains.

[0026] This invention provides a lightweight design method for the body frame of new energy vehicles based on digital twins. It has the following beneficial effects:

[0027] This lightweight design method for new energy vehicle body frames based on digital twins achieves intelligent allocation of computing resources through dynamic precision topology networks. It strictly limits high-fidelity simulation to the mechanical energy-dominated path and real-time high-sensitivity region, ensuring accurate prediction of key structures while effectively reducing the computational burden of multi-physics coupled simulation. Combined with the decoupled design of the dual-track drive optimization engine, the real-time track achieves twin synchronization through neural network prediction and local correction mechanisms, while the offline track relies on cloud computing power to perform deep multi-objective optimization. This effectively resolves the sharp contradiction between model accuracy, real-time response, and optimization efficiency, enabling lightweight design to achieve effective closed-loop iteration under complex constraints.

[0028] This lightweight design method for new energy vehicle body frames based on digital twins constructs a full-process quality defense line through a dual verification system: residual screening dynamically maintains the credibility of the proxy model through a periodic self-correction mechanism, avoiding the risk of model drift caused by long-term operation; cross-model verification forces parallel execution of high-precision physical simulation and independent data-driven inference, with a dual-channel arbitration mechanism to prevent safety misjudgments caused by single model failure; and intelligent circuit breaker rules automatically intercept dangerous solutions when update delay exceeds limits, optimization convergence is abnormal, or verification indicators conflict, ensuring that the output results have both lightweight benefits and engineering feasibility, providing full-chain technical protection for the safety of new energy vehicle bodies. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the module interaction of a lightweight design method for a new energy vehicle body frame based on digital twins according to the present invention.

[0030] Figure 2 This is a flowchart illustrating a lightweight design method for a new energy vehicle body frame based on digital twins according to the present invention. Detailed Implementation

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

[0032] Please see Figure 1 and Figure 2 This invention provides a technical solution: a lightweight design method for a new energy vehicle body frame based on digital twins, comprising the following steps:

[0033] Step S1: Establish a digital twin that connects to the physical vehicle body sensing system;

[0034] Step S2: Dynamic precision topology network module, used to dynamically divide the simulation precision area of ​​the vehicle frame according to the mechanical energy transfer path and real-time sensor data;

[0035] Step S3: Dual-track drive optimization engine module, including real-time track updates and offline track optimization;

[0036] Step S4: The dual validation module performs residual screening and cross-model consistency checks on the output scheme.

[0037] It should be further explained that, in the specific implementation process, after establishing a digital twin connected to the physical vehicle body sensing system, the following procedure is followed:

[0038] S01. Construction and execution of dynamic precision topology networks, including the following:

[0039] Sensitive domain identification based on mechanical energy transfer path: The energy distribution intensity of each region of the vehicle frame is calculated by stress flow density algorithm, and continuous structural regions with energy transfer intensity exceeding a preset threshold are automatically extracted and marked as topological sensitive domains;

[0040] Dynamic region labeling that integrates real-time data: Spatiotemporal clustering analysis is performed on the accessed real-time strain sensing data stream. If the stress fluctuation amplitude of a certain region continues to exceed the set tolerance range, it is added as a dynamic high-sensitivity region.

[0041] Precision domain map generation: The topology-sensitive domain and the dynamic high-sensitivity region are superimposed and fused to form a key precision domain map covering the entire vehicle frame. This map divides the structure into three types of regions:

[0042] Full-precision region, namely: topology sensitive region + dynamic high-sensitivity region: must fully preserve geometric details and material nonlinear properties;

[0043] Medium precision region, i.e., secondary energy transfer paths: allows for simplified cell types and mesh densities;

[0044] The proxy model region, namely the low-energy fluctuation region, replaces the finite element model with parametric equations.

[0045] S02. The coordinated operation of the dual-track drive optimization engine includes the following:

[0046] Real-time track updates: When the physical vehicle body sensors transmit new strain and temperature data, the edge computing nodes immediately call the pre-trained hybrid neural network. The long short-term memory network processes the temporal data, and the three-dimensional convolutional network processes the spatial distribution, directly predicting the displacement and stress increment in the full-precision area. The global finite element solver is skipped, and the predicted increment values ​​are mapped to the corresponding mesh nodes in the full-precision area. Only the local state update is performed in this area, keeping the model of the other areas unchanged.

[0047] Offline track optimization: A multi-objective genetic algorithm is run in the background to generate candidate design schemes with lightweight coefficient, structural stiffness, and modal frequency as optimization objectives; when the track output is updated in real time with a new full-precision zone state, the Bayesian optimizer is immediately activated to fine-tune the cross-sectional thickness and material distribution parameters of the candidate schemes with this state as boundary conditions.

[0048] S03. Triggering and execution of dual authentication, including the following:

[0049] First residual screening: At fixed intervals, the system automatically selects three sub-regions from the proxy model area and switches to the full-precision finite element model to recalculate the performance indicators; if the calculation result of any sub-region deviates from the predicted value of the proxy model by more than the allowable tolerance, the proxy model parameter self-learning process is triggered and all optimization iterations are frozen until the deviation is corrected.

[0050] The second step is cross-model verification: For the final lightweight solution, full-precision virtual collision simulation and independent agent model inference are launched simultaneously: the virtual collision simulation needs to cover three working conditions: frontal impact, lateral compression and torsional load; the independent agent model adopts a reinforcement learning architecture, is trained based on historical high-precision simulation data and is isolated from the main system; the solution can only be verified and output when the difference between the two models in key indicators such as the maximum deformation of the vehicle body and the first-order modal frequency is less than the set threshold.

[0051] The execution of the dynamic precision topology network module includes:

[0052] The mechanical energy distribution of the vehicle frame is calculated using a stress flow density algorithm, and continuous regions with energy intensity exceeding a preset threshold are identified as topology-sensitive regions.

[0053] Cluster analysis of real-time strain sensing data identifies dynamic high-sensitivity regions where stress fluctuations exceed a set tolerance.

[0054] By overlaying the topology-sensitive domain with the dynamic high-sensitivity region, a key precision domain map covering the entire vehicle frame is generated.

[0055] The diagram divides the vehicle body frame into a full-precision area, a medium-precision area, and a proxy model area.

[0056] It should be further explained that, in the specific implementation process, the construction of the dynamic precision topology network is carried out according to the following logic, including:

[0057] S04. Topology Sensitive Domain Identification Based on Physical Mechanisms: Standard working condition loads, such as bending and torsion, are applied to the initial finite element model of the vehicle body frame. The mechanical energy transfer intensity is calculated element by element using a stress flow density algorithm. This intensity is characterized by the product of the element stress tensor and the strain energy density. The largest connected region with energy intensity exceeding a preset threshold is automatically extracted. If the region contains key structural nodes, such as welding points or mounting holes, its boundary is extended outward by one element width to ensure integrity. During the identification process, regions with local stress concentration but low energy transfer contribution, such as small fillets and process holes, are excluded. Only continuous structures with dominant load transfer paths are retained.

[0058] S05. Dynamic high-sensitivity region labeling based on real-time data: Continuously receive strain time-series data uploaded by the physical vehicle sensor network, group it by spatial location and perform sliding window clustering analysis; if the strain fluctuation amplitude of a certain sensor group continuously exceeds the set tolerance range, and the fluctuation frequency matches the current driving state of the vehicle, which includes acceleration and steering, then the area covered by the sensor is labeled as a dynamic high-sensitivity region; when multiple sensor groups in the same area trigger the labeling condition at the same time, they are automatically merged into a single high-sensitivity region and boundary smoothing is performed.

[0059] S06. Generation and Conflict Resolution of Key Precision Domain Maps: Spatial overlay of topologically sensitive domains and dynamically high-sensitivity regions, handling overlap or separation cases according to the following rules:

[0060] Overlapping areas: upgraded to the highest priority full-precision area;

[0061] Separate regions: Retain their respective independent region types;

[0062] When the boundary gap is smaller than the element size: forced connection forms a continuous region;

[0063] The final map output includes three types of partitioning instructions, including:

[0064] Command executed in full-precision region: Load the complete nonlinear material model and fine mesh;

[0065] Instructions executed in the medium-precision region: equivalent homogeneous shell elements are used and mesh coarsening is allowed;

[0066] Execution instructions for the proxy model area: Activate the parameterized response surface equations and disable finite element calculations.

[0067] The full-precision region uses a complete finite element model, the medium-precision region uses a simplified shell element model, and the surrogate model region uses parametric equations to describe the structural response.

[0068] It should be further explained that, in the specific implementation process, for the three types of regions divided by the key precision domain map, the model construction is carried out according to the following mandatory rules:

[0069] S07. Model loading mechanism for full-precision region: When a region is marked as full-precision region by the map, a nonlinear finite element model containing geometric details is forcibly loaded. The material constitutive relation must be based on the experimental calibration curve. Among them, geometric details include stiffeners and welds.

[0070] If there are multi-physics coupling effects in the region, such as thermal fatigue, the coupled solver is activated synchronously and the complete boundary condition propagation path is preserved; the mesh density follows the preset curvature adaptive rule: the mesh is automatically refined when the surface curvature radius is less than the threshold, and the mesh can be moderately relaxed in flat regions.

[0071] S08. Simplified execution logic for the medium-precision region: For secondary energy transfer paths marked as being in the medium-precision region, uniformly replace them with equivalent shell element models. Specific rules include:

[0072] If the thickness variation rate of the original solid structure is lower than the set tolerance, it is converted into a shell element with uniform thickness; if there are local features, such as small holes or shallow grooves, geometric filling is performed directly when their size is smaller than the average side length of the element; the mesh density is subject to hierarchical control: the boundary layer adjacent to the full-precision region maintains a fine mesh; the internal region is allowed to be coarsened to an integer multiple of the original mesh size, but the element aspect ratio is strictly controlled within a reasonable range.

[0073] S09. Parametric implementation of the proxy model region: For the low-sensitivity proxy model region, the finite element calculation engine is completely shut down and replaced by a predefined parametric response surface equation; the equation input variables only include the boundary displacement and temperature field of the adjacent full-precision / medium-precision regions; when real-time sensing data detects abnormal fluctuations in this region, such as a sudden change in single-point strain, the equation is immediately frozen and the spectrum update process is triggered.

[0074] In the dual-track drive optimization engine module:

[0075] The track is deployed on the edge computing node in real time, receives the physical vehicle body sensor data stream, predicts the mechanical state change of the full-precision area through a pre-trained neural network, and directly modifies the displacement of the finite element mesh node in that area.

[0076] The offline optimized trajectory is deployed on a cloud server, runs a multi-objective genetic algorithm to generate lightweight candidate solutions, and calls a Bayesian optimizer to adjust the geometric parameters of the candidate solutions based on the real-time updated full-precision region state of the trajectory output.

[0077] It should be further explained that, in the specific implementation process, the dual-track drive optimization engine operates collaboratively according to the following logic, including the following steps:

[0078] S010. Real-time update of the track edge response: When the physical body sensor transmits new strain and temperature data to the edge computing node, the pre-trained hybrid neural network is immediately invoked. This network synchronously processes the spatiotemporal data stream and outputs the displacement increment prediction value of six degrees of freedom in the full accuracy region.

[0079] Skip the global equilibrium iteration process of the finite element method and directly map the prediction increment to the grid node coordinates corresponding to the full-precision region. Only modify the node displacement field in this region while keeping the state of the rest of the region frozen. If the confidence of the neural network prediction is lower than the set threshold, such as if the input data exceeds the training range, switch to the simplified finite element solver to perform local recalculation and send a data anomaly alarm to the cloud.

[0080] S011. Deep Iteration of Offline Optimized Track in the Cloud: A multi-objective genetic algorithm is continuously run in the cloud to generate a candidate solution library with vehicle body mass, first-order torsional stiffness, and key point fatigue life as optimization objectives; when a new full-precision displacement field is pushed to the track in real time, a three-stage response is immediately triggered, including the following three stages:

[0081] Phase 1: Using the real-time updated displacement field as the forced boundary condition, verify whether the stress in the full-precision region of the candidate scheme exceeds the limit;

[0082] Phase 2: If the number of out-of-limit schemes exceeds the total percentage threshold, the Bayesian optimizer is activated to adjust the plate thickness distribution of the candidate schemes.

[0083] Phase 3: The adjusted scheme is reinjected into the genetic algorithm population, replacing the individual with the lowest fitness.

[0084] S012. State synchronization rules between the two tracks: Each time the real-time track completes an update, it sends a full-precision compressed displacement field containing a timestamp to the cloud; the cloud-optimized track only activates the Bayesian fine-tuning process when it detects that the displacement field change amplitude is greater than the noise tolerance; if the edge continuously sends fault alarms, the cloud pauses optimization and rolls back to the previous stable version twin state.

[0085] The neural network is a hybrid architecture of long short-term memory network and three-dimensional convolutional network. The input is time-series strain data and spatial temperature distribution, and the output is the displacement increment of six degrees of freedom in the full accuracy region.

[0086] It should be further explained that, in the specific implementation process, the hybrid neural network performs spatiotemporal data processing and prediction according to the following rules, including the following steps:

[0087] S013. Dynamic filtering and structuring of input data: Real-time reception of raw data streams uploaded by edge sensors, automatic shielding of sensor nodes with signal strength below the noise threshold, retaining only valid measurement points; effective strain data are reorganized into a grid according to the spatial location of the vehicle body, and missing points are filled in by linear interpolation of adjacent nodes; temperature field data are used to independently construct a three-dimensional thermal distribution map, which is spatially aligned with the strain grid but with separate channels.

[0088] S014. Dual-channel collaborative processing mechanism:

[0089] Temporal feature extraction channel: For the strain time series data of each sensor node, a long short-term memory network is used to extract trend features according to a fixed time window; if the feature fluctuation direction is consistent and the amplitude increases within three consecutive time windows, it is marked as a significant change pattern and its weight is increased.

[0090] Spatial feature extraction channel: The reconstructed strain mesh and temperature distribution map are input into a three-dimensional convolutional network, and local stress concentration areas and thermo-coupling hot spots are identified through multi-scale convolutional kernels; when the spatial gradient exceeds the material yield threshold, high-resolution feature map calculation is forcibly started.

[0091] S015. Cross-channel feature fusion and displacement prediction: The trend feature vector output by the time-series channel and the thermally coupled feature map output by the spatial channel are concatenated in the channel dimension; cross-domain feature fusion is performed through a fully connected layer. If conflicting signals are detected in the fused features, such as time-series compression but spatial stretching, the spatial channel results are adopted first and data verification is triggered.

[0092] The final output layer generates displacement increment values ​​for all six degrees of freedom in the full-precision region, and adds a confidence score: the confidence score is dynamically calculated based on the Mahalanobis distance between the input data and the training set; when the confidence score is lower than a set threshold, the prediction results are frozen and the simplified solver is activated.

[0093] The two-factor authentication module includes:

[0094] First residual screening: At fixed intervals, three sub-regions of the surrogate model area are randomly selected and switched to full-precision model calculation performance indicators. If the deviation from the surrogate model prediction exceeds the allowable tolerance, the surrogate model parameter self-learning is triggered and the optimization process is paused.

[0095] The second cross-model verification: For the final lightweight solution, full-precision virtual collision simulation and reinforcement learning-based surrogate model inference are performed simultaneously, and the differences between the two results in the maximum deformation of the vehicle body and the first-order modal frequency are compared.

[0096] It should be further explained that, in the specific implementation process, the dual verification module performs reliability assurance according to the following rules, including the following steps:

[0097] S016. Periodic triggering of the first residual screening: At fixed intervals, the system automatically selects three spatially discrete sub-regions from the proxy model area and immediately switches to the full-precision finite element model to recalculate the equivalent stress and displacement values ​​under the current working condition.

[0098] If the absolute deviation between the full-precision calculation result of any sub-region and the predicted value of the surrogate model exceeds the allowable tolerance, the surrogate model is deemed to have failed and a three-stage self-correction is performed, including the following three stages:

[0099] Phase 1: Freeze all optimization iteration processes and retain a snapshot of the current design state;

[0100] Phase 2: Using the full-precision results as a benchmark, backfit the surrogate model parameters until the residuals converge to the tolerance range;

[0101] Phase 3: Perform adjacent region expansion verification on the corrected model, and unfreeze the optimization process after confirming that there are no errors;

[0102] If three consecutive screenings do not exceed the tolerance, the interval between the next screening will be automatically extended.

[0103] S017. Second Cross-Model Validation Scheme Output Interception: When the optimized trajectory generates the final lightweight scheme, force the parallel initiation of two independent evaluations:

[0104] Full-precision virtual collision simulation: running on a high-performance cluster in the cloud, fully loading geometric nonlinearity and material plasticity models, covering three standard working conditions: frontal 100% overlap rigid wall collision, side column collision, and torsional load.

[0105] Reinforcement learning agent model inference: Call an independently trained deep Q-network model, input the scheme design parameters, and directly output the predicted values ​​of key performance indicators;

[0106] The dual-channel result comparison implements strict arbitration: for core indicators such as the maximum deformation of the vehicle body and the first-order torsional mode frequency, if the difference between the results of the two models is less than the set threshold, the solution is immediately released to the manufacturing system;

[0107] If the differences in key indicators exceed the limits, conflict items will be automatically marked and root cause analysis will be performed. If the full-precision simulation results do not meet the safety standards, the scheme will be discarded and optimization will be restarted. If the surrogate model's prediction deviation is too large, its retraining process will be triggered and the scheme will retain its eligibility for review.

[0108] The agent model based on reinforcement learning is trained independently using historical high-precision simulation data and has no shared parameters with the main optimization system.

[0109] It should be further noted that, in the specific implementation process, the construction and operation of the reinforcement learning agent model strictly adhere to the following independent principles:

[0110] S018. Completely isolated acquisition of training data: Only historical high-precision simulation data is used as training samples; the use of any intermediate results or real-time sensor data generated during real-time optimization is prohibited; sample selection is subject to dual filtering, including the following two layers of filtering:

[0111] First-level filtering: Exclude historical simulation cases with confidence scores below a set threshold;

[0112] Second-level filtering: Automatically discard samples when their distribution deviates from the current vehicle model parameters by more than the allowable deviation;

[0113] The final training set covers a balanced distribution of positive and negative samples for frontal collision, side collision, and torsion conditions, and the total number of samples remains constant to avoid data drift.

[0114] S019. Special design of model architecture and training: A deep Q-network architecture is adopted, but the number of hidden layer nodes and the type of activation function are forced to be different from the hybrid neural network of the main system;

[0115] The training process is carried out in three phases:

[0116] Basic training: Minimize prediction error on historical datasets;

[0117] Adversarial enhancement: Injecting noisy samples and forcing the model to maintain a stable output;

[0118] Safety boundary reinforcement: Increase the weight of critical samples that are close to the material failure threshold by ten times;

[0119] All parameters are frozen immediately after training is complete, and online updates or fine-tuning are prohibited.

[0120] S020. Runtime hard isolation measures: Deployed on a separate physical server, with design parameters and return results transmitted only through encrypted APIs to the main digital twin system; input and output variables are subject to formatting constraints, including:

[0121] Input: Only accepts the geometric parameters and material grade of the vehicle body frame;

[0122] Output: Strictly limited to five core indicators, including maximum vehicle body deformation and first-order modal frequency;

[0123] If the input parameters exceed the historical training range, an error code will be returned immediately and the calculation will be refused.

[0124] Virtual collision simulation includes multiphysics coupling calculations under frontal collision, side collision, and torsional conditions.

[0125] It should be further explained that, in the specific implementation process, the construction and execution of virtual collision simulation conditions follow the following intelligent selection and simplification rules, including the following steps:

[0126] S021. Forced coverage of basic working conditions: Unconditionally execute full-precision simulation of three standard collision scenarios: frontal 100% overlap rigid wall collision: adopt the initial velocity and boundary conditions specified by regulations, but the vehicle body attitude dynamically adjusts the pitch angle according to the real-time twin mass distribution;

[0127] Side pole impact: The collision point is intelligently selected based on historical damage data from vehicle sensors, choosing the most vulnerable area, and the pole diameter is set according to statistical values ​​of actual road conditions;

[0128] Pure torsional condition: The applied load value is a fixed multiple of the current model's maximum design torque, but the direction is automatically aligned according to the weak axis of the real-time twin stiffness.

[0129] All operating conditions must retain local failure effects such as weld breakage and material tearing, and simplification or neglect is not allowed.

[0130] S022. Dynamic screening of derivative working conditions: Extract high-frequency non-standard working conditions for this vehicle model from the historical accident database, such as offset collision and slope collision, and selectively activate them according to the following rules: If the safety margin of the current lightweighting solution is lower than the threshold under standard working conditions, the corresponding high-risk derivative working conditions will be automatically added; If the number of optimization iterations reaches the critical point and still has not converged, then derivative working conditions that are strongly related to the structural characteristics of the candidate solution (such as special collisions for the weight reduction area) will be activated.

[0131] Derivative working conditions allow for moderate simplification: only the complete solution for the full-precision region is retained, while the medium-precision model is used for non-critical regions; the time step of the collision process can be scaled up to an integer multiple of the standard working condition.

[0132] S023. On-demand coupling of multiphysics, including:

[0133] Basic coupled field: All working cases must include a joint structural-dynamic solution;

[0134] Extended physical field condition triggering: When the material temperature sensing data continuously exceeds the threshold: activate the thermo-mechanical coupling field, but only in the full accuracy region; when the collision velocity exceeds the critical value: activate the fluid-structure coupling, and use the parameterized wind pressure model to replace the complete CFD; if there are composite material areas in the vehicle body: force activation of interlayer delamination effect simulation.

[0135] The method outputs a lightweight solution when the state update delay of the digital twin is lower than a set threshold and the time taken for a single optimization iteration is shorter than the baseline time.

[0136] It should be further explained that, in the specific implementation process, the output of the lightweight solution follows a strict three-level decision-making process, including the following steps:

[0137] S024. Real-time threshold dynamic calibration: Continuously monitor the status update delay of the digital twin. When the update cycle is shorter than the real-time set threshold for five consecutive times, trigger the automatic threshold reduction mechanism to improve the response standard.

[0138] If the resource utilization rate of edge computing nodes continues to fall below the warning line, the upward floating threshold is allowed to be delayed to a certain extent, but the floating range is constrained by the safety factor; when any single update timeout reaches the critical multiple, the scheme output is immediately suspended and system diagnosis is initiated.

[0139] S025. Composite determination of efficiency gain optimization: Calculate the compression ratio of the average time of the current iteration round relative to the baseline iteration time. When the compression ratio continuously exceeds the efficiency gain set threshold, activate the accelerated output channel.

[0140] Synchronous verification of the convergence curve of the optimization target: If the key target, such as mass and stiffness, has entered a stable convergence stage, the efficiency gain is recognized as effective; if the target value still fluctuates drastically, it is judged as a pseudo gain and the current compression ratio is ignored; when both efficiency gain and convergence stability meet the target, the candidate scheme is marked to enter the final verification queue.

[0141] S026. Circuit Breaker Arbitration of Verification Results: Force dual verification for schemes entering the queue; regardless of how excellent the real-time performance and efficiency indicators are, the scheme shall be discarded immediately if any of the following situations occur: any standard working condition in the full-precision virtual collision simulation exceeds the safety tolerance; the difference in the core indicators of cross-model verification continues to widen; residual screening is not completed or is in a correction state; the scheme shall be released to the manufacturing system and the current optimization state shall be frozen only when the real-time performance meets the standard, the efficiency gain is effective and all verifications pass.

[0142] The critical precision domain map is dynamically updated every two hours. Update trigger conditions include: the rate of change of statistical features in real-time sensor data exceeding a threshold, or the optimization engine triggering the identification of new topologically sensitive domains. It should be further noted that in the specific implementation process, the dynamic update of the critical precision domain map follows intelligent triggering and arbitration rules, including the following steps:

[0143] S027. Real-time data-driven emergency update: Continuously monitor the data stream of the entire vehicle sensor network. When the rate of change of the statistical characteristics of any sensor group exceeds the threshold, such as a sudden change in the strain mean or spectral energy migration, immediately start local spectrum redrawing.

[0144] The redrawing scope is limited to the three-layer unit radiation area centered on the mutation point, and a spatial correlation test is performed: if the synchronization of three adjacent sensors is abnormal, it is expanded to the entire subsystem; if an isolated point mutates, only that point is marked for observation and the response is delayed; the update process adopts incremental loading, only replacing the accuracy instructions of the affected area, keeping the rest of the area frozen to reduce disturbance.

[0145] S028. Preventive update triggered by optimized path: When the structural modification scheme generated by offline optimized track involves adjustment of topology sensitive domain, such as adding weight reduction holes or removing reinforcing ribs, the map pre-update is automatically triggered before the next iteration cycle.

[0146] The pre-update execution difference comparison priority strategy is as follows: if the overlap between the new sensitive area and the original map is higher than the threshold, the boundary is refreshed locally; if a completely new load path appears, the stress flow recalculation of the entire model is initiated; the update results need to be verified by a lightweight virtual working condition, such as static torsion, and can only be injected into the optimization process after the verification is passed.

[0147] S029. Safety fallback through periodic updates: Global map review is forcibly initiated at fixed intervals, but with flexible simplification based on system load: When edge computing resources are sufficient: the stress flow algorithm and data clustering are fully executed; when resources are scarce: only historically high-sensitivity areas are verified, and the surrogate model confidence score is used for non-critical areas; when low-risk deviations are found during the review, such as a slight increase in stress in the surrogate model area, an asynchronous correction task queue is generated for delayed processing to ensure uninterrupted real-time optimization.

[0148] S030. Conflict Arbitration for Multi-Source Updates: When urgent updates, preventative updates, and periodic updates are triggered simultaneously, a three-level priority response is executed, including the following:

[0149] Level 1 Priority: Urgent updates involving security-critical areas;

[0150] Second-level priority: Optimize preventative updates triggered by path changes;

[0151] Level 3 Delay: Global Periodic Review;

[0152] After arbitration is completed, a synchronization lock command is sent to the dual-track engine to ensure the consistency of the optimization state during the update.

[0153] It should be further explained that, in the specific implementation process, after establishing a digital twin of the new energy vehicle body frame and connecting it to the physical sensor network, the system automatically divides the body structure regions through a unique dynamic precision topology network. Specifically, firstly, based on the mechanical energy transfer path, the continuous region where the dominant load is transferred is identified as the topological sensitive region. At the same time, the fluctuation characteristics of real-time strain data are analyzed, and regions where stress changes continuously exceed the normal range are marked as dynamic high-sensitivity regions. The two are superimposed to generate a key precision domain map, dividing the body into a full-precision region, a medium-precision region, and a surrogate model region. The full-precision region forcibly retains geometric details and material nonlinear characteristics, the medium-precision region adopts an equivalent simplified model, and the surrogate model region completely shuts down finite element calculations and replaces them with parametric equations. This map is dynamically updated at fixed intervals or when sensor data changes abruptly, with the update process prioritizing safety-critical regions.

[0154] A dual-track optimization engine operates synchronously for efficient design. The real-time updated track is deployed on edge computing nodes. When new data arrives from sensors, a pre-trained hybrid neural network directly predicts the displacement increment across the full-precision region, skipping the global solver and only modifying the coordinates of the mesh nodes in that region. If the prediction confidence is insufficient, a simplified solver is switched to and an alert is issued. The offline optimized track runs a multi-objective genetic algorithm in the cloud to generate candidate solutions. When the real-time track pushes a new full-precision region state, a Bayesian optimizer is immediately activated using that state as boundary conditions to fine-tune the design parameters. Synchronization between the two tracks is achieved on demand through compressed displacement field packages with excessively large variations.

[0155] All output solutions undergo dual verification. The first residual screening is periodically initiated: three sub-regions of the surrogate model area are randomly selected and switched to full-precision model recalculation. If any result deviates beyond the limit from the surrogate model's prediction, the optimization process is frozen and the parameters are corrected by backfitting. The second verification is enforced before solution output: full-precision virtual collision simulation and independently trained reinforcement learning surrogate model inference are run simultaneously. The collision simulation must include frontal rigid wall collision, side pole collision, and torsional conditions, and retain local failure effects such as weld point fracture. The solution can only pass if the difference between the two models in core indicators such as maximum vehicle body deformation and first-order modal frequency is less than a threshold.

[0156] The final release of the solution requires meeting three conditions: the update latency of the digital twin must consistently remain below the real-time threshold of dynamic calibration; the efficiency gain of the optimized iteration must be verified as effective through convergence and stability; and all dual verifications must pass. If any step triggers a circuit breaker condition (such as exceeding collision safety limits or incomplete verification), the solution will be immediately discarded.

[0157] By employing a dynamic precision allocation mechanism, full-precision calculations are focused on critical areas that account for less than 20% of the vehicle's volume, resolving the conflict between high-fidelity simulation and computational resources. Through a dual-track decoupling architecture, parallel twin updates and efficient deep optimization are achieved, resolving the conflict between real-time performance and design quality. A closed-loop verification system ensures the reliability of the solution through periodic self-correction and pre-output dual-model arbitration.

[0158] It should be further explained that, in the specific implementation process, a lightweight design method for a new energy vehicle body frame based on digital twins includes the following steps:

[0159] Step S1: Construct a dynamic precision topology network: Identify the topological sensitive domain based on the mechanical energy transfer path of the vehicle frame, and simultaneously analyze the fluctuation characteristics of real-time strain data to mark the dynamic high-sensitivity area; superimpose the two to generate a key precision domain map, and divide the vehicle body into a full precision area, a medium precision area, and a proxy model area; when the real-time sensing data changes abruptly or the structure is optimized, the map is incrementally updated.

[0160] Step S2: Dual-track drive collaborative optimization, including:

[0161] Real-time track edge response: Sensor data is input into a pre-trained neural network, which directly outputs the displacement increment of the full-precision area and modifies the corresponding node coordinates. If the prediction confidence is insufficient, a simplified solver is switched.

[0162] Offline track cloud iteration: Run a multi-objective genetic algorithm to generate candidate schemes. When a new state in the full-precision range is received from the real-time track push, the Bayesian optimizer is immediately started with the state as the boundary condition to fine-tune the design parameters.

[0163] Step S3: Perform two-factor authentication, including:

[0164] Residual screening: Periodically switch random sub-regions of the surrogate model area to full-precision model recalculation; if the deviation exceeds the limit, freeze optimization and backfit to correct the surrogate model.

[0165] Cross-model validation: Before the solution is output, full-precision virtual collision simulation and independent reinforcement learning agent model inference are run simultaneously. Validation is passed only when the difference in core indicators is less than the threshold.

[0166] Step S4: Intelligent decision output, including releasing the lightweight solution if and only if the following conditions are met: the digital twin update delay is continuously lower than the dynamic calibration threshold, the optimized efficiency gain is valid after convergence and stability verification and all dual verifications pass; if any step triggers the circuit breaker condition, such as collision safety exceeding the limit or verification not being completed, the solution is immediately abandoned.

[0167] By using a dynamic precision topology network to achieve intelligent allocation of computing resources, high-fidelity simulation is strictly limited to the mechanical energy-dominated path and the real-time high-sensitivity region. While ensuring accurate prediction of key structures, it effectively reduces the computational burden of multi-physics coupled simulation. Combined with the decoupled design of the dual-track driven optimization engine, the real-time track achieves twin synchronization through neural network prediction and local correction mechanisms, while the offline track relies on cloud computing power to perform deep multi-objective optimization. This effectively resolves the sharp contradiction between model accuracy, real-time response and optimization efficiency, enabling lightweight design to achieve effective closed-loop iteration under complex constraints.

[0168] A dual verification system constructs a full-process quality defense line: residual screening dynamically maintains the credibility of the proxy model through a periodic self-correction mechanism, avoiding the risk of model drift caused by long-term operation; cross-model verification forces parallel execution of high-precision physical simulation and independent data-driven inference, and uses a dual-channel arbitration mechanism to prevent safety misjudgments caused by single model failure; in conjunction with intelligent circuit breaker rules, dangerous solutions are automatically intercepted when update delay exceeds limits, optimization convergence is abnormal, or verification indicators conflict, ensuring that the output results have both lightweight benefits and engineering feasibility, providing full-chain technical protection for the safety of new energy vehicle bodies.

[0169] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0170] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital-twin-based lightweight design method for a new energy vehicle body frame, characterized in that, Includes the following steps: Step S1: Establish a digital twin that connects to the physical vehicle body sensing system; Step S2: Dynamic precision topology network module, used to dynamically divide the simulation precision area of ​​the vehicle frame according to the mechanical energy transfer path and real-time sensor data; Step S3: Dual-track drive optimization engine module, including real-time track updates and offline track optimization; Step S4: The dual validation module performs residual screening and cross-model consistency checks on the output scheme; The execution of the dynamic precision topology network module includes: The mechanical energy distribution of the vehicle frame is calculated using a stress flow density algorithm, and continuous regions with energy intensity exceeding a preset threshold are identified as topology-sensitive regions. Cluster analysis of real-time strain sensing data identifies dynamic high-sensitivity regions where stress fluctuations exceed a set tolerance. By overlaying the topology-sensitive domain with the dynamic high-sensitivity region, a key precision domain map covering the entire vehicle frame is generated. The atlas divides the vehicle body frame into a full-precision zone, a medium-precision zone, and a proxy model zone; In the dual-track drive optimization engine module: The track is deployed on the edge computing node in real time, receives the physical vehicle body sensor data stream, predicts the mechanical state change of the full-precision area through a pre-trained neural network, and directly modifies the displacement of the finite element mesh node in that area. The offline optimized track is deployed on a cloud server, runs a multi-objective genetic algorithm to generate lightweight candidate solutions, and calls a Bayesian optimizer to adjust the geometric parameters of the candidate solutions based on the real-time updated full-precision region state of the track output. The dual verification module includes: First residual screening: At fixed intervals, three sub-regions of the surrogate model area are randomly selected and switched to full-precision model calculation performance indicators. If the deviation from the surrogate model prediction exceeds the allowable tolerance, the surrogate model parameter self-learning is triggered and the optimization process is paused. The second cross-model verification: For the final lightweight solution, full-precision virtual collision simulation and reinforcement learning-based surrogate model inference are performed simultaneously, and the differences between the two results in the maximum deformation of the vehicle body and the first-order modal frequency are compared.

2. The new energy vehicle body frame lightweight design method based on digital twinning according to claim 1, characterized in that: The full-precision region uses a complete finite element model, the medium-precision region uses a simplified shell element model, and the surrogate model region describes the structural response through parameterized equations.

3. The lightweight design method for a new energy vehicle body frame based on digital twins according to claim 1, characterized in that: The neural network is a hybrid architecture of long short-term memory network and three-dimensional convolutional network. The input is time-series strain data and spatial temperature distribution, and the output is the displacement increment of six degrees of freedom in the full accuracy region.

4. The lightweight design method for a new energy vehicle body frame based on digital twins according to claim 1, characterized in that: The reinforcement learning-based agent model is trained independently using historical high-precision simulation data and has no shared parameters with the main optimization system.

5. The lightweight design method for a new energy vehicle body frame based on digital twins according to claim 1, characterized in that: The virtual collision simulation includes multiphysics coupling calculations under frontal collision, side collision, and torsional conditions.

6. The lightweight design method for a new energy vehicle body frame based on digital twins according to claim 1, characterized in that: The method outputs a lightweight solution when the state update delay of the digital twin is lower than a set threshold and the time taken for a single optimization iteration is shorter than the baseline time.

7. The lightweight design method for a new energy vehicle body frame based on digital twins according to claim 1, characterized in that: The key precision domain map is dynamically updated every two hours. The update trigger conditions include: the rate of change of statistical features of real-time sensing data exceeds the threshold value, or the optimization engine triggers the identification of new topology sensitive domains.

Images