A large model-based vil interface level autonomous driving sensor fault intelligent injection method and system
Patent Information
- Application Number
- CN202611194442.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-07
- Publication Date
- 2026-10-27
AI Technical Summary
[0003]现有技术存在如下缺陷：1）、故障注入的智能化与适配性不足故障类型单一：人工配置的故障脚本仅能模拟传感器无数据、数据漂移等基础故障，无法覆盖传感器延迟、噪声突变、目标框丢失等复杂长尾故障，故障类型覆盖度不足
[0088]1、智能化水平显著提升，适配研发复杂测试需求：本发明通过大模型实现自然语言故障需求解析与工况自适应故障参数决策，突破传统人工脚本配置的局限，可模拟各类复杂长尾故障与多传感器协同故障，故障类型覆盖更全面，能够充分匹配自动驾驶系统研发阶段的高复杂度测试需求。
Smart Images
Figure CN122985631A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of autonomous driving testing, vehicle-in-the-loop (VIL) testing, large model application and sensor fault simulation, and specifically relates to a method and system for intelligent injection of VIL interface-level autonomous driving sensor faults based on a large model. Background Technology
[0002] In the field of autonomous driving sensor testing, vehicle-in-the-loop (VIL) testing is the core means of verifying the robustness of the system under sensor failure. Existing technologies mainly use VIL benches to perform sensor fault injection testing. The mainstream methods are to manually configure fault parameters, trigger faults with fixed scripts, or simulate a limited number of fault types based on simple algorithms. These methods are widely used in the research and development and mass production verification stages of autonomous driving sensors.
[0003] Existing technologies have the following shortcomings: 1) Insufficient intelligence and adaptability of fault injection: Limited fault types: Manually configured fault scripts can only simulate basic faults such as sensor data loss and data drift, failing to cover complex long-tail faults such as sensor delay, noise mutation, and target box loss, resulting in insufficient fault type coverage. Low operating condition matching: Fault injection timing and parameters are entirely manually set, unable to be dynamically adjusted according to real-time vehicle driving conditions (such as vehicle speed, lighting, and road conditions) on the VIL test bench, easily leading to unreasonable situations such as "injecting low-speed fault parameters under high-speed conditions," resulting in significant deviations between test results and real-world scenarios. 2) Insufficient accuracy and flexibility of VIL interface-level fault injection: Injection level deviation: Existing faults are mostly implemented at the simulation layer, not reaching the physical interface layer (CAN / Ethernet / FAKRA) between the VIL test bench and sensors, failing to simulate real interface communication faults, resulting in test results that do not reflect the actual fault response of sensor hardware. Difficulty in simulating multi-sensor collaborative faults: It is impossible to achieve multi-sensor coordinated fault injection (such as simultaneous delay faults in cameras and millimeter-wave radar), making it difficult to verify the fault tolerance capability of autonomous driving systems for multi-sensor collaborative faults.
[0004] 3) Low fault testing efficiency and closed-loop capability; long test script development cycle: adding a new sensor fault type requires manually writing a dedicated test script, with an average development cycle of 3-5 days, which cannot quickly adapt to the fault testing needs of new sensors; lack of automatic analysis and optimization: the test results after fault injection need to be manually analyzed, and it is impossible to automatically associate fault parameters with system response, making it difficult to achieve iterative optimization of fault injection strategy, and the testing efficiency is only 1 / 50 of the actual requirements.
[0005] The reasons for the above-mentioned defects are as follows: 1) Insufficient integration of fault injection with the large model. The massive fault data mining and scenario matching capabilities of the large model are not utilized. Fault types and injection strategies still rely on human experience, and it is impossible to achieve generalization and dynamic adjustment based on real fault cases. The natural language parsing capabilities of the large model are not applied to fault requirement input, making it impossible to achieve the convenient operation of "natural language description of fault → automatic generation of injection instructions", which increases the threshold for testing operations.
[0006] 2) Weak fault control capabilities at the VIL test bench interface layer: Existing VIL test benches lack a unified interface-level fault injection module, making it impossible to achieve standardized fault triggering and parameter adjustment for different interface types such as CAN / Ethernet / FAKRA. Poor coordination between the test bench and the fault injection module prevents real-time interaction between test bench operating data and the fault injection module, resulting in a disconnect between fault injection and actual operating conditions.
[0007] 3) The entire process of fault testing lacks automation. An automated closed loop of "fault requirement input - instruction generation - interface-level injection - result analysis - strategy optimization" has not been established, requiring manual intervention at each stage, resulting in low testing efficiency. Furthermore, a standardized fault knowledge base is lacking; there is no accumulation of sensor fault cases, fault characteristics, and system responses from real-world roads, making it impossible to provide a scientific basis for fault injection. Summary of the Invention
[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for intelligent fault injection of VIL interface-level autonomous driving sensors based on a large model. This method and system achieves fault injection driven by natural language, adaptive to operating conditions, at the physical interface level, with multi-sensor collaboration and automated closed loop, significantly improving the authenticity, comprehensiveness and efficiency of testing.
[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0010] A method for intelligent fault injection of VIL interface-level autonomous driving sensors based on a large model includes the following steps:
[0011] Step S1: Construct a sensor fault research and development database, and perform fine-tuning training on the large model based on the database;
[0012] Step S2: Receive natural language fault test requests and simultaneously collect VIL bench condition data in real time;
[0013] Step S3: The large model parses the natural language requirements and combines them with the real-time operating data of the VIL test bench to generate standardized fault injection instructions at the VIL interface level.
[0014] Step S4: According to the standardized fault injection instruction, drive the VIL interface-level fault injection hardware module to perform fault injection at the physical interface layer between the sensor and the VIL test bench.
[0015] Step S5: Collect full-link response data of the autonomous driving system to complete the verification of fault validity and automatic analysis of system robustness defects;
[0016] Step S6: Generate R&D optimization suggestions, feed all test data back to the sensor fault R&D database, and iteratively update the large model to form a test closed loop.
[0017] Further optimization, step S1 specifically includes the following sub-steps:
[0018] Step S11: Collect real sensor failure cases from actual vehicles and VIL test benches, and extract failure type, failure characteristics, interface type, and system response data;
[0019] Step S12: Establish a precise mapping table for VIL bench hardware and interfaces, clarifying the correspondence between sensors and CAN, Ethernet, and FAKRA physical interfaces. The specific implementation involves: compiling a sensor list, mapping hardware connections on-site, establishing a one-to-one correspondence, standardizing data entry into the database, and verifying and locking the mapping relationship. The mapping table includes precise correspondences for sensor identity information, physical interface type, communication link and channel, electrical and protocol attributes, and fault injection points.
[0020] Through this mapping table, the fault commands generated by the large model can be directly located to the physical interface corresponding to the target sensor. The fault scheduling unit can quickly drive the interface adaptation unit to complete the docking, realize the accurate injection of faults into the specified interface, and avoid problems such as erroneous triggering and cross-sensor interference.
[0021] Step S13: Statistically analyze the distribution patterns of fault parameters under different operating conditions to form operating condition-fault association rules.
[0022] Operating condition-fault association rules are criteria for establishing standardized correspondences and constraints between vehicle operating condition parameters and sensor interface fault parameters. These rules provide a basis for adaptive fault parameter decisions for large models, ensuring that fault injection conforms to the characteristics of real-world road fault distribution. Specifically, the rules include the following four categories:
[0023] 1) Association rules between operating condition type and fault type: Based on the vehicle's driving operating condition, match the most common and representative sensor interface fault type under that condition. Under high-speed cruising conditions, match fault types such as interface noise, data transmission delay, and intermittent data packet loss; under congested urban driving conditions, match fault types such as short burst noise, target frame jitter, and data discontinuity; under conditions of drastic changes in lighting, match fault types such as image sensor interface disturbance and abnormal fluctuations in exposure correlation; under vehicle acceleration and deceleration conditions, match fault types such as IMU and CAN interface data drift and timing deviation.
[0024] 2) Correlation rules between operating condition intensity and fault parameter amplitude: Based on the operating condition intensity level, the amplitude, range, and distribution characteristics of fault parameters are constrained to ensure consistency with actual faults. The higher the vehicle speed, the smaller the allowable fault delay and the higher the fault trigger frequency, resulting in noise amplitudes that more closely resemble the actual distribution characteristics of a real vehicle. The more drastic the change in light intensity, the greater the fluctuation range of noise at the image sensor interface. The greater the vehicle acceleration, the higher the allowable upper limit for data drift at the IMU and CAN interfaces.
[0025] 3) Association rules between operating condition timing and fault triggering timing: Based on the timing of changes in vehicle driving status, the fault triggering timing, duration range, and synchronization conditions are determined. During the high-speed driving phase when entering a curve, the fault triggering timing is synchronized with the vehicle's steering signal; during the dynamic acceleration and deceleration phase of the vehicle, the fault is triggered within the operating condition change range; during the vehicle's constant speed and stable driving phase, the fault is continuously injected with constant parameters.
[0026] 4) Association rules between operating conditions and multi-sensor collaborative faults: Configure single-sensor fault or multi-sensor collaborative fault modes according to the characteristics of the operating scenario. Under high-speed braking conditions, configure the camera and millimeter-wave radar to experience data delay faults simultaneously; under conditions where the vehicle passes over speed bumps, configure the multi-sensor interfaces to experience intermittent contact failure faults simultaneously; under conditions of sudden changes in light, configure only the camera interface to experience noise faults, while the radar sensors remain in normal working condition.
[0027] Step S14: Integrate fault cases, interface mapping tables, and operating condition-fault association rules into the database to form a standardized sensor fault R&D database;
[0028] Step S15: Using the sensor fault R&D database as training samples, fine-tune the large model with small samples; the large model fine-tuning adopts the small sample fine-tuning algorithm based on LoRA (Low-Rank Adaptation) to reduce training costs and retain the original capabilities of the large model. The specific formula is as follows:
[0029] (1) Fine-tuning the objective function of the large model: used to minimize the error in fault requirement analysis, the error in matching operating parameters, and the error in generating fault commands, to ensure that the fine-tuned model is adapted to the fault injection scenario:
[0030] ;
[0031] in: For natural language requirement parsing loss, the cross-entropy loss function is used to measure the deviation between the core elements (target sensor, fault type, etc.) parsed by the model and the true labels; For the operating condition-fault parameter matching loss, mean squared error loss is used to measure the deviation between the fault parameters output by the model and the standard parameters in the operating condition-fault association rules; To generate a loss for faulty instructions, perplexity loss is used to ensure that the generated standardized instructions conform to the protocol specifications. These are weighting coefficients, all ranging from [0.2, 0.5]. They are dynamically adjusted based on the training convergence performance, prioritizing the accuracy of the working condition matching. (The value is slightly higher).
[0032] (2) LoRA fine-tuning parameter update formula:
[0033] ;
[0034] in: This is the original weight matrix of the attention layer in the large model; A low-rank matrix (dimension: ), A low-rank matrix (dimension: ), Low-rank dimension (values range from 16 to 64, adjusted according to model size); To fine-tune the attention layer weight matrix and ensure the model accurately captures the correlation between fault requirements and operating conditions, only the attention layer weights of the large model are updated using low-rank adaptation, without changing the original model parameters, thus reducing training overhead.
[0035] Step S16: Optimize the natural language fault requirement parsing capability, fault type matching capability, working condition adaptive parameter decision-making capability, and system defect analysis capability of the large model;
[0036] Step S17: Establish a communication link between the trained large model and the VIL test bench and fault injection hardware module, and enter the ready-to-trigger state.
[0037] Further optimization is made to the natural language fault test requirement received in step S2, specifically as follows:
[0038] The large model natural language parsing unit allows input of text descriptions containing the target sensor, interface type, fault type, operating conditions, test target, and fault parameter range, eliminating the need to write test scripts.
[0039] The VIL benchtop data acquisition unit collects real-time data on vehicle speed, acceleration, light intensity, road type, and sensor operating status, and transmits the data synchronously to the large model core processor via a real-time communication link.
[0040] Further optimization, step S3 specifically includes:
[0041] Step S31: The large model core processor receives the natural language fault test requirements and the real-time operating condition data of the VIL test bench; it parses the natural language fault test requirements and extracts the core elements of the requirements, such as the target sensor, interface type, fault type, test conditions, fault parameters, and duration.
[0042] Step S32: Call the operating condition-fault association rules in the sensor fault R&D database, and adaptively optimize the fault parameters, triggering timing, and triggering frequency in combination with real-time operating condition data;
[0043] Step S33: Based on the optimized fault configuration, generate a standardized fault injection instruction that includes sensor identifier, interface type, fault type, precise fault parameters, trigger timing, duration, and collaborative fault configuration.
[0044] The adaptive optimization of fault parameters employs a weighted fusion algorithm based on operating condition features. The specific formula is as follows, which is used to dynamically adjust fault parameters according to real-time operating conditions to ensure that the fault matches the real-world scenario:
[0045] ;
[0046] in: For the optimized final fault parameters (such as noise amplitude, delay time, packet loss rate); This represents the number of similar operating conditions matched in the operating condition-fault association rules. Let be the weight of the i-th type of similar working conditions, satisfying The higher the similarity between the current operating condition and the real-time operating condition, the greater the weight. These are the standard fault parameters corresponding to the i-th type of similar operating conditions; This is a condition adaptation function used to correct standard parameters to adapt to the current real-time operating conditions. , For real-time operating parameters (such as vehicle speed and lighting conditions). For the standard parameters of the i-th type of similar working conditions, A correction factor (with a value of 0.01–0.05) is used to ensure that the parameter correction range conforms to the actual fault patterns.
[0047] Step S34: Send the standardized fault injection command to the VIL benchtop fault scheduling unit through the communication link and wait for execution.
[0048] Further optimization, step S4 specifically includes:
[0049] Step S41: The fault scheduling unit receives the standardized fault injection instruction, parses and obtains the target sensor, interface type, fault type, fault parameters, triggering time, duration and collaborative fault configuration information;
[0050] Step S42: The interface adapter unit establishes a precise electrical connection and communication match with the CAN, Ethernet or FAKRA physical interface corresponding to the target sensor according to the interface type;
[0051] Step S43: According to the triggering timing specified by the instruction, the fault triggering unit performs fault triggering at the physical interface layer and generates an interface fault signal consistent with the instruction;
[0052] Step S44: The parameter adjustment unit monitors the fault parameters in real time, dynamically adjusting and maintaining the fault intensity, trigger frequency, and duration consistent with the command requirements. Parameter adjustment employs a PID closed-loop control algorithm to ensure fault parameter stability. The specific formula is as follows:
[0053] ;
[0054] in: The output control quantity of the parameter adjustment unit (such as voltage adjustment quantity, frequency adjustment quantity); For fault parameter deviation, These are the fault parameters monitored in real time at time t; This is the proportionality coefficient. The integral coefficient is... These are differential coefficients, dynamically adjusted according to different fault types (e.g., noise faults). Delay fault: ).
[0055] By using PID closed-loop control, the deviation of fault parameters is kept within ±5%, ensuring the accuracy and reproducibility of fault injection.
[0056] Step S45: Support independent fault injection for a single sensor, or perform synchronous / asynchronous fault linkage injection for multiple sensors according to the collaborative fault configuration;
[0057] Step S46: Fault injection is completed entirely at the physical interface layer between the sensor and the VIL test bench, without going through the simulation layer, thus realistically reproducing the characteristics of hardware interface faults.
[0058] Further optimization, step S5 specifically includes:
[0059] Step S51: During the entire fault injection period, continuously collect raw sensor output data, domain controller perception result data, decision planning layer control command data, and vehicle actuator status data to form full-link response data of the autonomous driving system;
[0060] Step S52: Perform preprocessing on the collected end-to-end response data, including timestamp alignment, outlier removal, noise filtering, and format normalization.
[0061] Step S53: Compare the preprocessed data with the fault injection command parameters in real time to verify the validity of the fault type, fault parameters, triggering timing, and duration; the formula is as follows:
[0062] ;
[0063] in: To verify the total number of data entries; The number of data points where the parameter deviation is within the allowable range (±5%); when When the fault injection is valid, it is determined that the fault injection is effective.
[0064] Step S54: Based on the fault effectiveness verification results, perform correlation analysis between the fault injection parameters and system response indicators to locate robustness defects such as perception degradation, target loss, decision delay, and control anomalies. Defect localization employs a correlation analysis algorithm to quantify the degree of correlation between fault parameters and system response, as shown in the following formula:
[0065] ;
[0066] in: The correlation coefficient has a value range of [-1, 1]. The larger the value, the higher the correlation between the fault parameters and the system response; The total number of samples taken during the fault injection duration; This represents the average value of the real-time monitored fault parameters. This represents the average value of the system response index.
[0067] when When the system response is abnormal, it is determined that the abnormality is caused by the corresponding fault parameters, and the defect is identified as a robustness flaw.
[0068] Step S55: Output defect location and quantitative assessment results including fault injection compliance rate, perceived performance degradation, decision response delay, system fault tolerance level, and security risk level.
[0069] Further optimization, step S6 specifically includes the following sub-steps:
[0070] Step S61: Based on the robustness defect analysis results of step S5, and combined with similar fault optimization cases in the sensor fault R&D database, automatically generate feasible R&D optimization suggestions including algorithm improvement, hardware adaptation, and strategy adjustment.
[0071] Step S62: Standardize and organize the fault parameters, operating condition data, system response results, defect analysis conclusions, and R&D optimization suggestions from this test;
[0072] Step S63: Feed back the standardized and organized full test data and update it to the sensor fault R&D database;
[0073] Step S64: Using the updated database as new samples, perform a new round of small-sample fine-tuning on the large model to optimize fault decision-making and defect analysis capabilities; the new round of fine-tuning adopts an incremental training method, and the objective function is improved by adding defect analysis loss to the original function, as shown in the following formula:
[0074] ;
[0075] in: For defect analysis loss, cross-entropy loss is used to measure the deviation between the defect type and cause located by the model and the actual analysis results. The weighting coefficient is 0.3–0.4, ensuring that the model optimization focuses on both instruction generation and defect analysis capabilities.
[0076] Step S65: Complete database iteration and large model iteration to form a complete closed loop of "fault injection - test analysis - R&D optimization - strategy iteration".
[0077] A VIL interface-level intelligent fault injection system for autonomous driving based on a large model includes:
[0078] The system includes a large model decision-making unit, a VIL benchtop condition acquisition unit, a fault scheduling unit, a VIL interface-level fault injection hardware module, a system response acquisition unit, an automated analysis unit, and a closed-loop iterative optimization unit.
[0079] The large model decision unit is used to parse natural language fault requirements and generate standardized fault injection instructions by combining operating condition data.
[0080] The VIL benchtop condition acquisition unit is used to collect vehicle speed, light intensity, road conditions, and sensor operating status in real time and upload them to the large model decision unit.
[0081] The fault scheduling unit is used to receive and forward fault injection instructions and to coordinate and control the execution sequence of fault injection.
[0082] The VIL interface-level fault injection hardware module is used to perform single-sensor or multi-sensor collaborative fault injection at the CAN, Ethernet, and FAKRA physical interface layers.
[0083] The system response acquisition unit is used to acquire raw sensor data, domain controller perception results, decision commands, and vehicle execution status.
[0084] The automated analysis unit is used to verify the validity of the fault and locate the robustness defects of the system.
[0085] The closed-loop iterative optimization unit is used to generate R&D optimization suggestions, feed test data back to the sensor fault R&D database, and iteratively update the large model to achieve a closed-loop process.
[0086] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above-described method.
[0087] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0088] 1. Significantly improved intelligence level, adapting to complex testing needs in R&D: This invention realizes natural language fault requirement analysis and condition-adaptive fault parameter decision-making through a large model, breaking through the limitations of traditional manual script configuration. It can simulate various complex long-tail faults and multi-sensor collaborative faults, with more comprehensive fault type coverage, and can fully match the high-complexity testing needs in the R&D stage of autonomous driving systems.
[0089] 2. High accuracy of interface-level injection, realistically reproducing hardware fault characteristics: Fault injection directly acts on the CAN / Ethernet / FAKRA physical interface layer between the sensor and the VIL test bench, rather than the simulation layer. It can realistically reproduce the actual fault characteristics of the sensor hardware and communication link, and the test results are closer to the real vehicle scenario. It can provide a real and reliable test basis for hardware matching and algorithm development of autonomous driving systems.
[0090] 3. Significantly improved R&D efficiency and shortened iteration cycle: This invention automates the entire fault testing process, eliminating the need for manual test script writing and significantly reducing the test preparation time for new fault types. At the same time, it automatically outputs R&D optimization suggestions based on large models, greatly improving the efficiency of fault and defect location and analysis, and effectively shortening the R&D iteration cycle of autonomous driving systems.
[0091] 4. Forming a complete development closed loop to support continuous optimization: Through the database feedback of test data and optimization suggestions, continuous iterative optimization of the large model and fault injection strategy is achieved. As R&D and testing continue to advance, the accuracy and adaptability of fault injection are continuously improved, which can stably support the robustness verification of the entire process of autonomous driving system from R&D to mass production.
[0092] 5. Strong compatibility, good applicability, and industrialization value: This invention can be expanded based on existing VIL test benches without large-scale hardware modifications. It supports testing of different brands and types of autonomous driving sensors and can be quickly applied to R&D and testing stages in autonomous driving R&D centers, vehicle manufacturers, and sensor suppliers. It has strong versatility and promising prospects for industrial application. Attached Figure Description
[0093] Figure 1 This is a flowchart of the intelligent fault injection method for VIL interface-level autonomous driving based on a large model, according to the present invention. Detailed Implementation
[0094] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0095] Example 1:
[0096] This embodiment verifies the impact of intermittent voltage noise faults at the FAKRA interface of the forward-looking camera on the target vehicle tracking function of the domain controller perception algorithm under high-speed cruising conditions of 100km / h, identifies system robustness defects, and provides optimization suggestions that can be directly implemented for perception algorithm development.
[0097] like Figure 1 As shown, the intelligent fault injection method for VIL interface-level autonomous driving based on a large model specifically includes:
[0098] Step S1: Construction of R&D Database and Training of Large Model: Collect more than 2,000 real-world fault cases of front-view cameras during the R&D phase of autonomous driving systems, focusing on the characteristic parameters of intermittent noise faults caused by poor contact of the FAKRA interface; compile a precise mapping table of VIL bench hardware and interface to clarify the corresponding connection relationship between the front-view camera and the FAKRA physical interface; statistically analyze the triggering frequency and intensity distribution of interface noise in the high-speed operating range of 80-120km / h, and establish operating condition-fault association rules; uniformly input the above fault cases, interface mapping relationships, and operating condition-fault association rules into and construct a sensor fault R&D database.
[0099] Using this database as training samples, the LoRA algorithm is used to fine-tune the large model with small samples, setting the low-rank dimension r=32 and the weight coefficients... The large-scale model was optimized to improve its correlation decision-making, parameter matching, and system defect analysis capabilities for intermittent noise faults at the FAKRA interface under high-speed operating conditions. After training, the large-scale model was integrated with the VIL test bench and fault execution module, and entered a test-ready state. Using this database as training samples, the large-scale model underwent small-sample fine-tuning training, focusing on enhancing its correlation decision-making, parameter matching, and system defect analysis capabilities for intermittent noise faults at the FAKRA interface under high-speed operating conditions. After training, the large-scale model was integrated with the VIL test bench and fault execution module, and entered a test-ready state.
[0100] Step S2: Natural Language Requirement Input and Real-Time Bench Condition Data Acquisition: R&D testers directly input test requirements in text form through the large model's natural language requirement parsing unit: a high-speed 100km / h cruise scenario, simulating intermittent voltage noise faults at the FAKRA interface of the forward-looking camera, with noise intensity fluctuating randomly between 0.25-0.35V, a trigger frequency of 5 times / second, lasting 15 seconds, to verify the robustness of the target vehicle tracking function. No test scripts need to be written throughout the process. The VIL bench starts the high-speed cruise simulation scenario, and the condition data acquisition unit collects real-time data on the current conditions: vehicle speed 100km / h, light intensity 1000lux, and the forward-looking camera functioning normally. The condition data is synchronously uploaded to the large model's core processor via a real-time communication link.
[0101] Step S3, Large-Scale Intelligent Decision-Making and Standardized Fault Instruction Generation: After receiving natural language requirements and real-time operating condition data, the large-scale model core processor parses and extracts core elements such as the target sensor, interface type, fault type, fault parameters, and test target. It then calls the operating condition-fault association rules in the sensor fault R&D database, and, combined with the 100km / h high-speed operating condition, calculates the optimized fault parameters using an adaptive optimization formula. The specific calculation process is as follows: First, determine the number of similar operating conditions n=3 groups, with the corresponding standard operating conditions and parameters as follows: , , The operating condition weights are set to The correction factor k = 0.03. Based on the working condition adaptation function... Calculate the fit coefficients for the three sets of working conditions respectively: , , Then, the fault parameters are substituted into the adaptive optimization formula. The breakdown is as follows: , , The optimal baseline parameters are obtained by summing them. Centered on this benchmark parameter and combined with an engineering fluctuation margin of ±0.05V, the optimized noise intensity range was finally determined to be 0.26-0.34V, with the trigger frequency maintained at 5 times / second.
[0102] The large model generates standardized fault injection instructions, which include: target sensor - forward-looking camera, interface type - FAKRA, fault type - intermittent voltage noise, fault parameters - random fluctuation of 0.26-0.34V, trigger frequency 5 times / second, trigger timing - immediate, duration - 15s; the instructions are sent to the VIL bench fault scheduling unit through the fault instruction generation unit.
[0103] Step S4, VIL Interface-Level Fault Precision Injection Execution: Under the coordination of the benchtop collaborative controller, the fault scheduling unit issues execution commands to the VIL interface-level fault execution module; the interface adaptation unit completes precise electrical connection with the physical interface of the forward-looking camera FAKRA; the fault triggering unit triggers intermittent voltage noise at the physical interface layer according to the command sequence; the parameter adjustment unit adopts PID closed-loop control and sets... The system monitors and maintains the noise level within the random fluctuation range of 0.26-0.34V in real time. The specific control logic is as follows: based on the PID control formula... Among them, the deviation , The noise voltage is monitored in real time at time t. The voltage deviation is quickly corrected by the proportional term, the steady-state error is eliminated by the integral term, and the high-frequency voltage jitter is suppressed by the derivative term, so as to ensure that the fault parameter deviation is controlled within ±5% and to ensure the accuracy and reproducibility of fault injection.
[0104] This fault injection was completed at the physical interface layer, without going through the simulation layer, thus realistically reproducing the characteristics of hardware communication faults.
[0105] Step S5, System Response Acquisition and R&D-oriented Automated Analysis: During the 15-second fault injection period, the system response acquisition unit continuously collects raw output data from the forward-looking camera, perception result data from the domain controller, control command data from the decision-making and planning layer, and status data from the vehicle's actuators, forming full-link response data for the autonomous driving system. The collected data undergoes preprocessing: first, timestamp alignment is performed to ensure consistency in the timing of data across all stages; then, the 3σ criterion is used to remove outliers, i.e., if... (where x is a single response data, This is the mean of this type of data. If the standard deviation of this type of data is used, it is identified as an outlier and removed; then, the min-max normalization formula is applied. All data are mapped to the [0,1] interval to facilitate subsequent defect analysis.
[0106] After preprocessing, the fault validity verification formula is followed. The preprocessed data was compared with the fault injection command parameters in real time. A total of 200 verification data entries were collected in this test. Of these, 192 data points had parameter deviations within the allowable range of ±5%. ), calculated ,because The FAKRA interface noise fault injection was deemed effective.
[0107] Based on the effectiveness verification results, a correlation analysis algorithm was used to locate the system's robustness defects, and the results were substituted into the correlation coefficient formula. (where T is the total number of samples, This represents the average value of the real-time monitored fault parameters. The correlation coefficient is calculated from the mean of the system response index. .because The system determined that there was a strong correlation between the voltage noise of the FAKRA interface of the forward-looking camera and the abnormal target tracking. This abnormality was a system robustness defect. The final output defect quantitative assessment results were as follows: fault injection compliance rate 96%, perception performance degradation 18%, decision response delay increased by 30ms, system fault tolerance level decreased to level 2, and safety risk level was moderate.
[0108] Step S6, R&D Optimization Suggestion Generation and Closed-Loop Iteration: Based on the above defect analysis results and combined with similar fault optimization cases in the sensor fault R&D database, implementable R&D optimization suggestions are automatically generated: First, optimize the target tracking filtering strategy of the perception algorithm, add a voltage noise adaptive filtering module to filter intermittent noise in the range of 0.2-0.4V; second, optimize the FAKRA interface contact design, add anti-loosening buckles to reduce the probability of poor interface contact; third, add target tracking abnormal redundancy judgment logic to the decision layer, and automatically switch to backup sensor data when the perceived target is lost for more than 50ms.
[0109] After standardizing and organizing the fault parameters, operating condition data, system response results, defect analysis conclusions, and optimization suggestions from this test, the data was updated back to the sensor fault R&D database. Using this new case as a sample, a new round of incremental fine-tuning was performed on the large model, setting... This optimizes the defect localization accuracy of the model; after the large model iteration is completed, it can more accurately adapt to interface noise fault testing under high-speed operating conditions, forming a complete R&D closed loop of "fault injection - test analysis - R&D optimization - strategy iteration".
[0110] This embodiment injects intermittent noise faults at the FAKRA interface level of the forward-looking camera under high-speed cruise conditions, accurately locating robustness defects in the domain controller perception algorithm of the development version, and automatically outputting directly engineering-applicable optimization suggestions from the large model. Compared with traditional methods, this method eliminates the need to write test scripts, reduces fault injection preparation time from 3 days to 5 minutes, and reduces system defect analysis from 1 day to 10 minutes, significantly improving the efficiency and authenticity of autonomous driving system R&D testing. This method can be extended to various interface-level fault testing scenarios such as millimeter-wave radar Ethernet interface latency, lidar CAN interface packet loss, and multi-sensor collaborative faults; all modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention fall within the protection scope of this invention.
[0111] Example 2:
[0112] A VIL interface-level intelligent fault injection system for autonomous driving based on a large model includes:
[0113] The system includes a large model decision-making unit, a VIL benchtop condition acquisition unit, a fault scheduling unit, a VIL interface-level fault injection hardware module, a system response acquisition unit, an automated analysis unit, and a closed-loop iterative optimization unit.
[0114] The large model decision unit is used to parse natural language fault requirements and generate standardized fault injection instructions by combining operating condition data.
[0115] The VIL benchtop condition acquisition unit is used to collect vehicle speed, light intensity, road conditions, and sensor operating status in real time and upload them to the large model decision unit.
[0116] The fault scheduling unit is used to receive and forward fault injection instructions and to coordinate and control the execution sequence of fault injection.
[0117] The VIL interface-level fault injection hardware module is used to perform single-sensor or multi-sensor collaborative fault injection at the CAN, Ethernet, and FAKRA physical interface layers.
[0118] The system response acquisition unit is used to acquire raw sensor data, domain controller perception results, decision commands, and vehicle execution status.
[0119] The automated analysis unit is used to verify the validity of the fault and locate the robustness defects of the system.
[0120] The closed-loop iterative optimization unit is used to generate R&D optimization suggestions, feed test data back to the sensor fault R&D database, and iteratively update the large model to achieve a closed-loop process.
[0121] Example 3:
[0122] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above-described method.
[0123] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for intelligent fault injection of VIL interface-level autonomous driving sensors based on a large model, characterized in that, Includes the following steps: Step S1: Construct a sensor fault research and development database, and perform fine-tuning training on the large model based on the database; Step S2: Receive natural language fault test requests and simultaneously collect VIL bench condition data in real time; Step S3: The large model parses the natural language requirements and combines them with the real-time operating data of the VIL test bench to generate standardized fault injection instructions at the VIL interface level. Step S4: According to the standardized fault injection instruction, drive the VIL interface-level fault injection hardware module to perform fault injection at the physical interface layer between the sensor and the VIL test bench. Step S5: Collect full-link response data of the autonomous driving system to complete the verification of fault validity and automatic analysis of system robustness defects; Step S6: Generate R&D optimization suggestions, feed all test data back to the sensor fault R&D database, and iteratively update the large model to form a test closed loop.
2. The intelligent fault injection method for VIL interface-level autonomous driving sensors based on a large model according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: Step S11: Collect real sensor failure cases from actual vehicles and VIL test benches, and extract failure type, failure characteristics, interface type, and system response data; Step S12: Establish a precise mapping table for VIL bench hardware and interfaces, and clarify the correspondence between sensors and CAN, Ethernet, and FAKRA physical interfaces; Step S13: Statistically analyze the distribution patterns of fault parameters under different operating conditions to form operating condition-fault association rules; Step S14: Integrate fault cases, interface mapping tables, and operating condition-fault association rules into the database to form a standardized sensor fault R&D database; Step S15: Using the sensor fault R&D database as training samples, fine-tune the large model with small samples; Step S16: Optimize the natural language fault requirement parsing capability, fault type matching capability, working condition adaptive parameter decision-making capability, and system defect analysis capability of the large model; Step S17: Establish a communication link between the trained large model and the VIL test bench and fault injection hardware module, and enter the ready-to-trigger state.
3. The intelligent fault injection method for VIL interface-level autonomous driving sensors based on a large model according to claim 1, characterized in that, The specific requirements for receiving natural language fault tests in step S2 are as follows: Through the large model natural language parsing unit, the input text description includes the target sensor, interface type, fault type, operating conditions, test target, and fault parameter range, without the need to write test scripts; The VIL benchtop data acquisition unit collects vehicle speed, acceleration, light intensity, road type, and sensor operating status data in real time, and transmits the data synchronously to the large model core processor via a real-time communication link.
4. The intelligent fault injection method for VIL interface-level autonomous driving sensors based on a large model according to claim 1, characterized in that, Step S3 specifically includes: Step S31: The large model core processor receives the natural language fault test requirements and the real-time operating condition data of the VIL test bench; it parses the natural language fault test requirements and extracts the core elements of the requirements, such as the target sensor, interface type, fault type, test conditions, fault parameters, and duration. Step S32: Call the operating condition-fault association rules in the sensor fault R&D database, and adaptively optimize the fault parameters, triggering timing, and triggering frequency in combination with real-time operating condition data; Step S33: Based on the optimized fault configuration, generate a standardized fault injection instruction that includes sensor identifier, interface type, fault type, precise fault parameters, trigger timing, duration, and collaborative fault configuration; Step S34: Send the standardized fault injection command to the VIL benchtop fault scheduling unit through the communication link and wait for execution.
5. The intelligent fault injection method for VIL interface-level autonomous driving sensors based on a large model according to claim 1, characterized in that, Step S4 specifically includes: Step S41: The fault scheduling unit receives the standardized fault injection instruction, parses and obtains the target sensor, interface type, fault type, fault parameters, triggering time, duration and collaborative fault configuration information; Step S42: The interface adapter unit establishes a precise electrical connection and communication match with the CAN, Ethernet or FAKRA physical interface corresponding to the target sensor according to the interface type; Step S43: According to the triggering timing specified by the instruction, the fault triggering unit performs fault triggering at the physical interface layer and generates an interface fault signal consistent with the instruction; Step S44: The parameter adjustment unit monitors the fault parameters in real time, dynamically adjusts and maintains the fault intensity, trigger frequency, and duration in accordance with the command requirements; Step S45: Support independent fault injection for a single sensor, or perform synchronous / asynchronous fault linkage injection for multiple sensors according to the collaborative fault configuration; Step S46: Fault injection is completed entirely at the physical interface layer between the sensor and the VIL test bench, without going through the simulation layer, thus realistically reproducing the characteristics of hardware interface faults.
6. The intelligent fault injection method for VIL interface-level autonomous driving sensors based on a large model according to claim 1, characterized in that, Step S5 specifically includes: Step S51: During the entire fault injection period, continuously collect raw sensor output data, domain controller perception result data, decision planning layer control command data, and vehicle actuator status data to form full-link response data of the autonomous driving system; Step S52: Perform preprocessing on the collected end-to-end response data, including timestamp alignment, outlier removal, noise filtering, and format normalization. Step S53: Compare the preprocessed data with the fault injection command parameters in real time to verify the validity of the fault type, fault parameters, triggering time, and duration; Step S54: Based on the results of the fault effectiveness verification, analyze the correlation between the fault injection parameters and the system response indicators to locate the robustness defects of perception degradation, target loss, decision delay, and control anomaly. Step S55: Output defect location and quantitative assessment results including fault injection compliance rate, perceived performance degradation, decision response delay, system fault tolerance level, and security risk level.
7. The intelligent fault injection method for VIL interface-level autonomous driving sensors based on a large model according to claim 1, characterized in that, Step S6 specifically includes the following sub-steps: Step S61: Based on the robustness defect analysis results of step S5, and combined with similar fault optimization cases in the sensor fault R&D database, automatically generate feasible R&D optimization suggestions including algorithm improvement, hardware adaptation, and strategy adjustment. Step S62: Standardize and organize the fault parameters, operating condition data, system response results, defect analysis conclusions, and R&D optimization suggestions from this test; Step S63: Feed back the standardized and organized full test data and update it to the sensor fault R&D database; Step S64: Using the updated database as new samples, perform a new round of small-sample fine-tuning on the large model to optimize fault decision-making and defect analysis capabilities; Step S65: Complete database iteration and large model iteration to form a complete closed loop of "fault injection - test analysis - R&D optimization - strategy iteration".
8. A VIL interface-level intelligent fault injection system for autonomous driving based on a large model, characterized in that, include: The system includes a large model decision-making unit, a VIL benchtop condition acquisition unit, a fault scheduling unit, a VIL interface-level fault injection hardware module, a system response acquisition unit, an automated analysis unit, and a closed-loop iterative optimization unit. The large model decision unit is used to parse natural language fault requirements and generate standardized fault injection instructions by combining operating condition data. The VIL benchtop condition acquisition unit is used to collect vehicle speed, light intensity, road conditions, and sensor operating status in real time and upload them to the large model decision unit. The fault scheduling unit is used to receive and forward fault injection instructions and to coordinate and control the execution sequence of fault injection. The VIL interface-level fault injection hardware module is used to perform single-sensor or multi-sensor collaborative fault injection at the CAN, Ethernet, and FAKRA physical interface layers. The system response acquisition unit is used to acquire raw sensor data, domain controller perception results, decision commands, and vehicle execution status. The automated analysis unit is used to verify the validity of the fault and locate the robustness defects of the system. The closed-loop iterative optimization unit is used to generate R&D optimization suggestions, feed test data back to the sensor fault R&D database, and iteratively update the large model to achieve a closed-loop process.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method described in any one of claims 1-7.