A method for constructing fine-tuning data set of a multi-modal large model of a chassis tuning
By employing a deep jump traversal mechanism and parameter broadcast fusion technology, the problem of automated traversal of underlying parameters and data mapping in the construction of the training dataset for the large chassis tuning model was solved, generating a high-quality multimodal instruction fine-tuning dataset that meets the training requirements of the large chassis tuning model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI UNIV OF TECH
- Filing Date
- 2026-01-06
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies struggle to automate the batch traversal of underlying parameters when constructing training data for large chassis calibration models. The original discrete time-series data output from simulations lacks explicit semantic mapping with expert evaluation rules, making it impossible to generate high-quality multimodal instruction fine-tuning data.
By employing a deep jump traversal mechanism, a file system polling mechanism, and parameter broadcast fusion technology, combined with expert evaluation rules, a communication mechanism between an automated script program and the CarSim simulation software is constructed to achieve precise control of underlying parameters and structured transformation of data, generating a multimodal instruction fine-tuning dataset.
It achieves precise control of parameters of deep subsystems such as suspension and steering, resolves asynchronous timing conflicts between the simulation solver and external scripts, and generates a high-quality multimodal instruction fine-tuning dataset to meet the training requirements of large chassis tuning models.
Smart Images
Figure CN122153429A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of methods for constructing large model datasets for chassis tuning, specifically a method for constructing fine-tuning datasets for a multimodal large model of chassis tuning. Background Technology
[0002] Currently, when constructing large-scale training data for chassis tuning AI models, the methods generally involve manually collecting data or using simple batch processing scripts to replace parameters.
[0003] Existing methods for constructing training datasets for large chassis tuning models suffer from the following problems: In batch simulation, traditional scripts struggle to automatically modify parameters at the underlying levels of subsystems such as suspension and steering, and lack a stable process synchronization mechanism, resulting in the inability to efficiently generate high-fidelity samples covering the entire parameter domain. Regarding fine-tuning dataset construction, the raw discrete time-series data output from simulations lacks an explicit semantic mapping to the input operating parameters and expert evaluation rules, making it difficult to automatically transform into structured data containing a complete causal logic chain of "phenomenon-diagnosis-suggestion," thus failing to meet the demand for high-quality multimodal command fine-tuning data for training large chassis tuning models. Summary of the Invention
[0004] This invention provides a method for constructing a fine-tuning dataset for a large multimodal chassis tuning model, in order to solve the problems in existing technologies when constructing a large chassis tuning model using large-scale training data. These problems include difficulty in achieving automated batch traversal of underlying parameters, asynchronous timing conflicts between external control scripts and simulation solvers, and the lack of explicit semantic mapping between the original simulation data and expert evaluation rules, which makes it impossible to meet the fine-tuning requirements of multimodal commands.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for constructing a fine-tuning dataset for a multimodal large-scale chassis tuning model is described below: Step 1: Construct a multi-dimensional sampling space for vehicle chassis parameters, write an automated script program, and establish a communication connection between the script program and the CarSim simulation software through the COM interface; the script program samples from the multi-dimensional sampling space to generate a batch simulation task queue containing multiple sets of different configuration parameters, and the script program transmits the batch simulation task queue to the CarSim simulation software through the COM interface; Step 2: The automated script program, based on the storage hierarchy of parameters in the CarSim simulation software, uses a deep jump traversal mechanism to control the CarSim simulation software interface to jump from the top-level entry point to the bottom-level parameter definition interface in order to locate the parameter definition interface where the bottom-level parameters to be modified are located. Step 3: After locking the underlying parameter definition interface, the automated script program accurately locates the numerical input control based on the preset control identifier, directly injects the scalar values in the batch simulation task queue generated in Step 1 into the numerical input control, and sends a save command to the CarSim simulation software to update the database file of the CarSim simulation software. Step 4: Use an automated script program to control the CarSim simulation software to fall back to the top-level interface and trigger the closed-loop simulation control process. The asynchronous timing conflict between script execution and solver calculation is resolved through the file system polling mechanism of the automated script program, and the CarSim simulation software generates the original result file. Step 5: The automated script program reads the original result file generated by the CarSim simulation software and uses parameter broadcast fusion technology to align the static input parameters with the dynamic timing results, generating flat structured data containing complete operating condition information. Step 6: Use an automated script program to process the flattened data generated in Step 5. Generate a serialized time-series feature file through working condition slicing and resampling. Perform performance diagnosis based on the expert evaluation rule system, thereby generating structured dialogue data containing user query text and expert response text. Step 7: Use a script program to perform multimodal alignment and encapsulation of the temporal feature file path generated in Step 6 and the structured dialogue data, and construct a JSON dataset that meets the requirements of the instruction fine-tuning format, which is the multimodal instruction fine-tuning dataset.
[0006] Furthermore, in step 2, the execution logic of the deep jump traversal mechanism is as follows: the script program parses the parent-child hierarchical relationship of the underlying parameters to be modified in the CarSim simulation software database, constructs a jump path from the main control level to the subsystem level and then to the component level, and controls the CarSim simulation software interface to enter each underlying sub-interface in sequence to complete the addressing and activation of the target parameter interface.
[0007] Furthermore, in step 3, the process of directly injecting the parameter scalar values is as follows: the script program establishes a mapping relationship table between the vehicle physical parameters and the control IDs of the CarSim simulation software interface. After the interface jump is completed in step 2, the target value input box is retrieved and locked according to the mapping relationship. The randomly generated parameters are directly written into the value input box using the automated interface function, and the parameter configuration after the modification is solidified by calling the saved interface.
[0008] Furthermore, when executing the closed-loop simulation control process in step 4, the historical result files and log files under the target output path are forcibly retrieved and deleted before the simulation run command is triggered; then, after the simulation command is triggered, the daemon process is started to poll and scan the target path, detect the existence and size of newly generated files in real time, and only when a new file is detected and the file size exceeds the effective threshold is the current simulation determined to be valid and enter the data processing process.
[0009] Furthermore, the alignment process using the parameter broadcasting fusion technology in step 5 is as follows: a basic data frame indexed by the simulation time step is established, vehicle motion state quantities are extracted from the original result file and filled into the data frame, and then the set of static input parameters generated in the current simulation is traversed, each parameter is copied as a constant vector and concatenated into each row of the data frame, thereby generating a self-describing data table containing complete causal relationships.
[0010] Furthermore, the process of processing the flattened data in step 6 includes two parallel processes: data feature processing and text generation. The data feature processing process is as follows: based on the vehicle mileage information, the data is divided into a first mileage interval corresponding to the comfort evaluation and a second mileage interval corresponding to the handling evaluation. The multi-channel heterogeneous time-series data after being sliced is mapped into a feature matrix of fixed time length using an interpolation algorithm and saved as a serialized file as the visual modal input of the large model. The text generation process is as follows: Calculate the statistical characteristics of physical state quantities in each working condition segment and compare them with preset performance boundary thresholds and physical hard limits. Based on the comparison results, use a template engine to generate two parts of text. The first part is user query text containing the current working condition description and parameter setting values, and the second part is expert response text containing performance problem diagnosis and adjustment suggestions.
[0011] Furthermore, the process of constructing the multimodal instruction fine-tuning dataset in step 7 is as follows: establish a sample structure containing three elements: visual modality, text instruction, and target response; wherein, the serialized feature matrix generated in step 6 is configured as the visual modality used to characterize the objective dynamic response of the vehicle, the user query text generated in step 6 is configured as the text instruction used to guide the large model to focus on a specific tuning problem, and the expert response text generated in step 6 is configured as the generation target used to train the large model to output the correct tuning strategy. Finally, the three are encapsulated into a JSON format list to obtain the multimodal instruction fine-tuning dataset.
[0012] This invention, based on a deep jump traversal mechanism, a file system polling mechanism, and a parameter broadcasting fusion and expert rule annotation method for large models, has the advantages of breaking through the limitations of traditional scripts to achieve precise control and full coverage of parameters of deep subsystems such as suspension, steering, and anti-roll bars. At the same time, it effectively solves the asynchronous timing conflict between the simulation solver and the external script, ensuring the accuracy and stability of batch data generation, and can transform discrete simulation data into high-quality multimodal instruction fine-tuning data with explicit causal logic of "input-output", providing a solid data foundation for training AI models with chassis tuning capabilities. Attached Figure Description
[0013] Figure 1 This is a flowchart of the method according to an embodiment of the present invention. Detailed Implementation
[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0015] like Figure 1 As shown in the figure, this embodiment discloses a method for constructing a fine-tuning dataset for a multimodal large model of C-Class Hatchback chassis tuning, wherein the multimodal large model is the controlled object, and the road environment is configured as a mixed road surface including speed bump excitation and flat road lane change. The process is as follows: Step 1: Construct a multi-dimensional sampling space for vehicle chassis parameters, write an automated script program, and establish a communication connection between the script program and the CarSim simulation software through the COM interface; the script program samples from the multi-dimensional sampling space to generate a batch simulation task queue containing multiple sets of different configuration parameters, and the script program transmits the batch simulation task queue to the CarSim simulation software through the COM interface.
[0016] In this embodiment, the script program first defines a dictionary structure for key vehicle chassis parameters, setting minimum and maximum values for each parameter. The sampling range of the constructed multi-dimensional sampling space covers the spring stiffness of the front / rear suspension, the damping coefficient of the shock absorber, the anti-roll bar stiffness, and the transmission ratio and torsion bar stiffness of the steering system.
[0017] The script uses Python's win32com.client library to call the Dispatch method to instantiate a CarSim.Application object. This object serves as the COM interface for communication between the external script and the CarSim software. The script obtains a handle to the CarSim simulation software through this object, thereby establishing an instruction transmission channel between the external code and the simulation kernel, and realizing the communication connection between the script and the CarSim simulation software.
[0018] After the script program initializes the batch simulation task queue, it samples from the multidimensional sampling space to generate a batch simulation task queue containing multiple sets of different configuration parameters, which is then transmitted to the CarSim simulation software.
[0019] Step 2: The automated script program uses a deep jump traversal mechanism to control the CarSim simulation software interface to jump from the top-level entry point to the bottom-level parameter definition interface based on the storage level of the parameters configured in the current simulation task in the CarSim simulation software, so as to lock the bottom-level parameter definition interface where the bottom-level parameters to be modified are located.
[0020] In this embodiment, taking advantage of the dispersed storage of parameters in the CarSim simulation software, for each set of parameters in the task queue, the script calls the Gotolibrary function to control the software interface to recursively jump to different interfaces. The specific path is as follows: first, it jumps to the "Procedures" interface; then to the "Steering" interface; subsequently, it jumps to the "Suspension" interface; and finally, it jumps to the "Auxiliary Roll Moment" interface nested within the suspension system. This interface jump order reflects the parent-child hierarchical relationship from the whole vehicle (Procedures) to subsystems (Steering, Suspension) and then to components (Auxiliary Roll Moment).
[0021] After each navigation, the target interface at that level is addressed and activated. The script immediately performs an unlock operation to ensure the current dataset is editable. In this embodiment, after the unlock operation is completed, the script keeps the interface in an unlocked, editable state for subsequent parameter injection, without needing to unlock the interface again.
[0022] Step 3: After locating the underlying parameter definition interface, the script program accurately locates the numerical input control based on the preset control identifier, directly injects the scalar values in the batch simulation task queue generated in Step 1 into the numerical input control, and sends a save command to the CarSim simulation software to update the CarSim simulation software's database file.
[0023] In the parameter injection process of this embodiment, the script program pre-establishes a mapping table between vehicle physical parameters and CarSim simulation software interface control IDs, and uses this mapping table to achieve precise positioning of the target control. The script program traverses all controls in the currently active window through the application programming interface, and locks the numerical input box ID corresponding to the physical parameters according to the mapping relationship.
[0024] Then, using the cs.Yellow function provided by the CarSim automation interface (which is used to write values to controls with specified IDs), the following write operations are performed: the randomly generated suspension spring values are directly written to the KSPRING_L value input control (i.e., the value input box mentioned above), the damping values are written to the DAMPER_L value input control, and the anti-roll bar stiffness is written to the MX_AUX_COEFFICIENT value input control; for the steering system, the steering ratio values are written to the C_FACTOR value input control, and the steering torsion bar stiffness values are written to the K_TBAR value input control.
[0025] This process does not require switching control modes. It directly injects parameters by overwriting the original values and immediately calls the Save function to perform a save operation, ensuring that the modified chassis parameters are written to the underlying simulation database file.
[0026] Step 4: Use an automated script program to control the CarSim simulation software to fall back to the top-level interface and trigger the closed-loop simulation control process. The asynchronous timing conflict between script execution and solver calculation is resolved through the file system polling mechanism of the automated script program, and the CarSim simulation software generates the original result file.
[0027] In this embodiment, to resolve the asynchronous timing conflict between script execution and solver computation, a closed-loop control logic based on file status is designed. Specifically, before triggering the simulation run, the script program uses the os.path module to search the output directory and calls the os.remove function to forcibly delete historical result files and log files with the same name, clearing the historical cache to prevent data confusion. Subsequently, the script program calls the Run function in the top-level interface to activate the background solver for dynamic calculation, and then enters the monitoring program of the while loop structure.
[0028] The monitoring program polls the target path to detect whether the target result file exists and whether the file size meets the requirements. It uses time.sleep to set the sampling interval and continuously calls the os.path.exists and os.path.getsize functions to detect the target result file. Only when a new file is detected (i.e. os.path.exists returns True) and the file size exceeds the preset byte threshold (i.e. os.path.getsize returns a value greater than the preset value) is the simulation considered successful and the loop is exited. If the preset timeout period is exceeded, an exception is thrown and the current task is skipped.
[0029] Step 5: The automated script program reads the original result file generated by the CarSim simulation software and uses parameter broadcast fusion technology to align the static input parameters with the dynamic timing results, generating flat structured data containing complete operating condition information.
[0030] In this embodiment, after the simulation is completed, the original result file generated by the CarSim simulation software in step 4 is read using the Pandas data analysis library. Key time-series data, including time, vehicle mileage, longitudinal speed, lateral acceleration, roll angle, yaw rate, slip angle, steering wheel angle, steering wheel torque, vertical acceleration, pitch angle, and four-wheel suspension compression stroke, are extracted according to a preset mapping table.
[0031] To construct a self-describing dataset, this embodiment employs parameter broadcasting fusion technology. The static input parameters corresponding to this simulation (i.e., the configuration parameters corresponding to the current task in the batch simulation task queue generated in step 1, such as "front spring = 50, front anti-roll bar = 400") are treated as a constant column, copied, and concatenated to each row of the time-series data table. The resulting CSV file contains both the current vehicle dynamic response and the vehicle configuration parameters that caused that response in each row, achieving row-level alignment between input and output and forming a flattened data structure.
[0032] Step 6: Use an automated script program to process the flattened data generated in Step 5. Generate a serialized time-series feature file through working condition slicing and resampling. Perform performance diagnosis based on the expert evaluation rule system, thereby generating structured dialogue data containing user query text and expert response text.
[0033] In this embodiment, a script program is used to perform deep post-processing on the flattened data generated in step 5, which includes two parallel processes: data feature generation and diagnostic text generation.
[0034] In terms of data feature generation, the script slices the data based on vehicle mileage (Station): data between 10 and 60 meters is used as a comfort evaluation segment, and data between 140 and 260 meters is used as a handling evaluation segment. Then, the `interp1d` function from the SciPy library is used to perform linear interpolation resampling on the sliced multi-channel heterogeneous time-series data, uniformly mapping it to a fixed-length (600 steps in this embodiment) feature matrix and saving it as a .pkl serialized file, which serves as the visual modal input for the large model.
[0035] In terms of diagnostic text generation, an expert evaluation rule system is constructed to diagnose the slice data: (1) Physical limit detection: Real-time monitoring of the four-wheel suspension travel, comparing the compression travel with the maximum compression limit of the suspension design (in this embodiment, it is set to a threshold close to 60mm), if it exceeds the limit, it is determined that the suspension is bottoming out; comparing the rebound travel with the maximum rebound limit of the suspension design (in this embodiment, it is set to a threshold close to -40mm), if it is lower than the value, it is determined that the wheel is off the ground. (2) Comfort diagnosis: In the comfort segment, if the peak vertical acceleration exceeds the preset human comfort tolerance limit (e.g., 0.55g), it is determined that the bump is strong; if the pitch angle exceeds the preset vehicle posture stability threshold (e.g., 1.2deg), it is determined that the vehicle body is swaying. (3) Handling diagnosis: In the handling segment, if the roll angle exceeds the preset roll stability threshold (e.g., 3.0deg), it is determined that the roll support is insufficient; if the slip angle exceeds the preset tire linear zone boundary (e.g., 2.5deg), it is determined that the tail is fishtailing and unstable. (4) Steering feel diagnosis: Calculate the ratio of lateral acceleration to steering wheel torque to obtain the force gradient. Compare this gradient value with a preset steering feel target range (e.g., 2.5 Nm / g to 5.5 Nm / g): if it is higher than the upper limit of the range, it is determined that the steering feel is too heavy; if it is lower than the lower limit of the range, it is determined that the steering feel is floaty. Based on the above judgment results, generate the corresponding user query text and expert response text in combination with the current parameters.
[0036] Step 7: Use a script program to perform multimodal alignment and encapsulation of the temporal feature file path generated in Step 6 and the structured dialogue data, and construct a JSON dataset that meets the requirements of the instruction fine-tuning format, which is the multimodal instruction fine-tuning dataset.
[0037] In this embodiment, after feature extraction and text generation are completed, the data is encapsulated into JSON format that meets the requirements for training multimodal large models.
[0038] Specifically, a sample structure is established that includes three elements: visual modality, text instruction, and target response. Specifically, the path to the .pkl serialized file generated in step 6 is configured as the visual modality (image) representing the vehicle's objective dynamic response; the user query text generated in step 6 is configured as the text instruction (human prompt) guiding the large model to focus on a specific tuning issue; and the expert response text generated in step 6 is configured as the generated target (gpt response) for training the large model to output the correct tuning strategy.
[0039] Finally, the script encapsulates the above three elements into a JSON object and writes it to the total dataset file, forming a standard multimodal instruction fine-tuning dataset.
[0040] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. These embodiments are merely descriptions of preferred embodiments and are not intended to limit the scope or concept of the invention. The specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. Such combinations, as long as they do not violate the spirit of the present invention, should also be considered as part of this disclosure. To avoid unnecessary repetition, the present invention will not further describe the various possible combinations.
[0041] This invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this invention and without departing from the design idea of this invention, all modifications and improvements made by those skilled in the art to the technical solutions of this invention should fall within the protection scope of this invention. The technical content for which protection is sought in this invention has been fully described in the claims.
Claims
1. A method for constructing a fine-tuning dataset for a multimodal large-scale chassis tuning model, characterized in that, The process is as follows: Step 1: Construct a multi-dimensional sampling space for vehicle chassis parameters, write an automated script program, and establish a communication connection between the script program and the CarSim simulation software through the COM interface; the script program samples from the multi-dimensional sampling space to generate a batch simulation task queue containing multiple sets of different configuration parameters, and the script program transmits the batch simulation task queue to the CarSim simulation software through the COM interface; Step 2: The automated script program, based on the storage hierarchy of parameters in the CarSim simulation software, uses a deep jump traversal mechanism to control the CarSim simulation software interface to jump from the top-level entry point to the bottom-level parameter definition interface in order to locate the parameter definition interface where the bottom-level parameters to be modified are located. Step 3: After locking the underlying parameter definition interface, the automated script program accurately locates the numerical input control based on the preset control identifier, directly injects the scalar values in the batch simulation task queue generated in Step 1 into the numerical input control, and sends a save command to the CarSim simulation software to update the database file of the CarSim simulation software. Step 4: Use an automated script program to control the CarSim simulation software to fall back to the top-level interface and trigger the closed-loop simulation control process. The asynchronous timing conflict between script execution and solver calculation is resolved through the file system polling mechanism of the automated script program, and the CarSim simulation software generates the original result file. Step 5: The automated script program reads the original result file generated by the CarSim simulation software and uses parameter broadcast fusion technology to align the static input parameters with the dynamic timing results, generating flat structured data containing complete operating condition information. Step 6: Use an automated script program to process the flattened data generated in Step 5. Generate a serialized time-series feature file through working condition slicing and resampling. Perform performance diagnosis based on the expert evaluation rule system, thereby generating structured dialogue data containing user query text and expert response text. Step 7: Use a script program to perform multimodal alignment and encapsulation of the temporal feature file path generated in Step 6 and the structured dialogue data, and construct a JSON dataset that meets the requirements of the instruction fine-tuning format, which is the multimodal instruction fine-tuning dataset.
2. The method for constructing a fine-tuning dataset for a multimodal large-scale chassis calibration model according to claim 1, characterized in that, In step 2, the execution logic of the deep jump traversal mechanism is as follows: the script program parses the parent-child hierarchical relationship of the underlying parameters to be modified in the CarSim simulation software database, constructs a jump path from the main control level to the subsystem level and then to the component level, and controls the CarSim simulation software interface to enter each underlying sub-interface in sequence to complete the addressing and activation of the target parameter interface.
3. The method for constructing a fine-tuning dataset for a multimodal large-scale chassis calibration model according to claim 1, characterized in that, In step 3, the process of directly injecting the parameter scalar values is as follows: the script program establishes a mapping relationship table between the vehicle physical parameters and the control IDs of the CarSim simulation software interface. After the interface jump is completed in step 2, the target value input box is retrieved and locked according to the mapping relationship. The randomly generated parameters are directly written into the value input box using the automated interface function, and the parameter configuration after the interface is saved and solidified is called.
4. The method for constructing a fine-tuning dataset for a multimodal large-scale chassis tuning model according to claim 1, characterized in that, When executing the closed-loop simulation control process in step 4, the historical result files and log files under the target output path are forcibly retrieved and deleted before the simulation run command is triggered. Then, after the simulation command is triggered, the daemon process is started to poll and scan the target path, detect the existence and size of newly generated files in real time, and only when a new file is detected and the file size exceeds the effective threshold is the current simulation determined to be valid and enter the data processing process.
5. The method for constructing a fine-tuning dataset for a multimodal large-scale chassis calibration model according to claim 1, characterized in that, The alignment process using parameter broadcasting fusion technology in step 5 is as follows: a basic data frame indexed by the simulation time step is established, vehicle motion state quantities are extracted from the original result file and filled into the data frame, and then the set of static input parameters generated in the current simulation is traversed, each parameter is copied as a constant vector and concatenated into each row of the data frame, thereby generating a self-describing data table containing complete causal relationships.
6. The method for constructing a fine-tuning dataset for a multimodal large-scale chassis calibration model according to claim 1, characterized in that, Step 6 involves processing the flattened data in two parallel processes: data feature processing and text generation. The data feature processing process is as follows: based on the vehicle mileage information, the data is divided into a first mileage interval corresponding to the comfort evaluation and a second mileage interval corresponding to the handling evaluation. The multi-channel heterogeneous time-series data after being sliced is mapped into a feature matrix of fixed time length using an interpolation algorithm and saved as a serialized file as the visual modal input of the large model. The text generation process is as follows: Calculate the statistical characteristics of physical state quantities in each working condition segment and compare them with preset performance boundary thresholds and physical hard limits. Based on the comparison results, use a template engine to generate two parts of text. The first part is user query text containing the current working condition description and parameter setting values, and the second part is expert response text containing performance problem diagnosis and adjustment suggestions.
7. The method for constructing a fine-tuning dataset for a multimodal large-scale chassis tuning model according to claim 1, characterized in that, The process of constructing the multimodal instruction fine-tuning dataset in step 7 is as follows: establish a sample structure containing three elements: visual modality, text instruction, and target response; wherein, the serialized feature matrix generated in step 6 is configured as the visual modality to characterize the objective dynamic response of the vehicle, the user query text generated in step 6 is configured as the text instruction to guide the large model to focus on a specific tuning problem, and the expert response text generated in step 6 is configured as the generation target to train the large model to output the correct tuning strategy. Finally, the three are encapsulated into a JSON format list to obtain the multimodal instruction fine-tuning dataset.