A Method and System for Self-Update of Vehicle-Mounted Model Based on User Behavior Feedback

Abstract

This application discloses a method and system for self-updating an in-vehicle edge model based on user behavior feedback. The method may include: acquiring real-time data during vehicle operation, including vehicle status data, environmental perception data, and user interaction behavior data; constructing a prediction model with a frozen base and pluggable skill plugins, the prediction model outputting predicted actions in real-time during daily driving or cockpit interaction; real-time monitoring of whether the driver engages in takeover or overriding intervention behaviors, and if so, extracting and saving preference data for a preset time period; adjusting the prediction model based on the preference data using the edge GPU during idle periods; and updating the adjusted prediction model through weight merging. This invention, through edge-side fine-tuning combining the Direct Preference Optimization algorithm and LoRA, eliminates the need for cloud-based training, consumes no cloud computing power, and completely solves the pain points of data compliance and privacy leakage.

Description

Technical Field

[0001] This invention relates to the field of vehicle control, and more specifically, to a method and system for self-updating an on-board model based on user behavior feedback. Background Technology

[0002] Traditional in-vehicle AI or control systems rely on fixed rule engines or static models trained at the factory. Their main drawbacks are: the inability to optimize based on different drivers' unique personalities, specific routes (such as bumpy sections of daily commuting), or driving preferences; traditional skill updates rely on full or incremental OTA updates from the OEM, failing to achieve "daily updates" or "instant evolution" for individual vehicles based on their specific environments; and the computing power of in-vehicle chips (such as high-performance intelligent cockpit or intelligent driving chips) is not fully utilized when idle (such as during parking and charging, or overnight parking), and the uploading of users' private driving data to the cloud poses privacy and compliance risks.

[0003] Therefore, it is necessary to develop a self-updating method and system for vehicle-mounted edge models based on user behavior feedback.

[0004] The information disclosed in the background section of this invention is intended only to enhance the understanding of the general background of this invention, and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] This invention proposes a self-updating method and system for vehicle-mounted edge models based on user behavior feedback. It can perform edge-side fine-tuning by combining the Direct Preference Optimization algorithm with LoRA, without needing to transmit back to the cloud for training, thus consuming no cloud computing power and completely solving the pain points of data compliance and privacy leakage.

[0006] In a first aspect, embodiments of this disclosure provide a self-updating method for an in-vehicle terminal model based on user behavior feedback, including: Real-time data is acquired during vehicle operation, including vehicle status data, environmental perception data, and user interaction behavior data. Construct a prediction model for a frozen base and pluggable skill plugins. The prediction model outputs predicted actions in real time during daily driving or cockpit interaction. Real-time monitoring of whether the driver takes over or overrides intervention behavior; if so, the preference data for a preset time period is captured and saved. When the vehicle is in an idle state, the prediction model is adjusted by the edge GPU based on preference data; The adjusted prediction model is updated by merging weights.

[0007] Preferably, the predictive model for the frozen base and pluggable skill plugins includes: A basic large model is built on a high-computing chip in the vehicle, and the network weights of the basic large model are completely frozen on the vehicle. Multiple LoRA modules are connected in parallel to the basic large model. Each LoRA module corresponds to a driving or interaction action and is used to fine-tune the weights.

[0008] Preferably, the weights used in the prediction model inference are: W = W0 + ΔW Where W0 is the frozen weight, and ΔW is the fine-tuning weight of the LoRA module.

[0009] Preferably, the takeover action includes: the system plans to brake late, and the driver depresses the brake pedal in advance; Intervention behaviors covered include: when the system-set air conditioning or ambient lighting parameters do not meet the user's needs, the user manually adjusts them via the panel.

[0010] Preferably, the preference data extracted within a preset time period includes: Using the detected takeover or coverage intervention behavior as the time point, environmental perception data, the suboptimal trajectory predicted by the prediction model, and the user's actual preference trajectory are extracted within a time window from T seconds before the time point to T seconds after the time point.

[0011] Preferably, determining that the vehicle is in an idle state includes: The vehicle is in Park and charging mode; The vehicle enters a deep sleep state at night.

[0012] Preferably, adjusting the prediction model based on preference data using an edge GPU includes: The adjustment goal is to increase the probability of the prediction model generating user preference trajectories, while decreasing the probability of generating predicted trajectories. The ΔW of the LoRA module in the prediction model is fine-tuned by minimizing the loss function.

[0013] Preferably, before updating the adjusted prediction model by weight merging, the adjusted prediction model is subjected to gray-scale verification.

[0014] Preferably, gray-scale validation of the adjusted prediction model includes: The goodness-of-fit of the adjusted prediction model's output with the user's actual operation is verified by running the algorithm multiple times. If the goodness-of-fit exceeds the preset safety threshold, the adjusted prediction model is deemed to be ready for updating. Otherwise, the prediction model is readjusted.

[0015] Secondly, embodiments of this disclosure also provide an in-vehicle terminal-side model self-updating system based on user behavior feedback, comprising: The data acquisition module is used to acquire real-time data during vehicle operation. Real-time data includes vehicle status data, environmental perception data, and user interaction behavior data. The large model base on the end side is used to construct a prediction model of the frozen base and pluggable skill plug-ins. The prediction model outputs predicted actions in real time during daily driving or cockpit interaction. The differential performance evaluation module is used to monitor in real time whether the driver has taken over or covered the intervention behavior. If so, it will extract and save the preference data for a preset time period. The local evolutionary fine-tuning module is used to adjust the prediction model based on preference data via the edge GPU during idle periods; The hot deployment control module is used to update the adjusted prediction model by merging weights.

[0016] Preferably, the predictive model for the frozen base and pluggable skill plugins includes: A basic large model is built on a high-computing chip in the vehicle, and the network weights of the basic large model are completely frozen on the vehicle. Multiple LoRA modules are connected in parallel to the basic large model. Each LoRA module corresponds to a driving or interaction action and is used to fine-tune the weights.

[0017] Preferably, the weights used in the prediction model inference are: W = W0 + ΔW Where W0 is the frozen weight, and ΔW is the fine-tuning weight of the LoRA module.

[0018] Preferably, the takeover action includes: the system plans to brake late, and the driver depresses the brake pedal in advance; Intervention behaviors covered include: when the system-set air conditioning or ambient lighting parameters do not meet the user's needs, the user manually adjusts them via the panel.

[0019] Preferably, the preference data extracted within a preset time period includes: Using the detected takeover or coverage intervention behavior as the time point, environmental perception data, the suboptimal trajectory predicted by the prediction model, and the user's actual preference trajectory are extracted within a time window from T seconds before the time point to T seconds after the time point.

[0020] Preferably, determining that the vehicle is in an idle state includes: The vehicle is in Park and charging mode; The vehicle enters a deep sleep state at night.

[0021] Preferably, adjusting the prediction model based on preference data using an edge GPU includes: The adjustment goal is to increase the probability of the prediction model generating user preference trajectories, while decreasing the probability of generating predicted trajectories. The ΔW of the LoRA module in the prediction model is fine-tuned by minimizing the loss function.

[0022] Preferably, before updating the adjusted prediction model by weight merging, the adjusted prediction model is subjected to gray-scale verification.

[0023] Preferably, gray-scale validation of the adjusted prediction model includes: The goodness-of-fit of the adjusted prediction model's output with the user's actual operation is verified by running the algorithm multiple times. If the goodness-of-fit exceeds the preset safety threshold, the adjusted prediction model is deemed to be ready for updating. Otherwise, the prediction model is readjusted.

[0024] The methods and systems of the present invention have other features and advantages that will be apparent from or will be set forth in detail in the accompanying drawings and following detailed description, which together serve to explain the particular principles of the invention. Attached Figure Description

[0025] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same parts.

[0026] Figure 1 A flowchart illustrating the steps of a vehicle-mounted edge model self-updating method based on user behavior feedback according to an embodiment of the present invention is shown.

[0027] Figure 2 A block diagram of an in-vehicle end-side model self-updating system based on user behavior feedback according to an embodiment of the present invention is shown.

[0028] Explanation of reference numerals in the attached figures: 201. Data acquisition module; 202. End-side large model base; 203. Differential performance evaluation module; 204. Local evolution fine-tuning module; 205. Hot deployment control module. Detailed Implementation

[0029] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0030] To facilitate understanding of the solutions and effects of the embodiments of the present invention, two specific application examples are given below. Those skilled in the art should understand that these examples are merely for the purpose of understanding the present invention, and any specific details therein are not intended to limit the present invention in any way.

[0031] Example 1

[0032] Figure 1 A flowchart illustrating the steps of a vehicle-mounted edge model self-updating method based on user behavior feedback according to an embodiment of the present invention is shown.

[0033] like Figure 1 As shown, the in-vehicle terminal model self-update method based on user behavior feedback includes: Step 101: Acquire real-time data during vehicle operation. Real-time data includes vehicle status data, environmental perception data, and user interaction behavior data. Step 102: Construct a prediction model for the frozen base and pluggable skill plugins. The prediction model outputs predicted actions in real time during daily driving or cockpit interaction. Step 103: Monitor in real time whether the driver has taken over or overridden intervention behavior. If so, extract and save the preference data for a preset time period. Step 104: During idle periods, the prediction model is adjusted based on preference data using the edge GPU; Step 105: Update the adjusted prediction model by merging weights.

[0034] In one example, the predictive model for building a frozen base and pluggable skill plugins includes: A basic large model is built on a high-computing chip in the vehicle, and the network weights of the basic large model are completely frozen on the vehicle. Multiple LoRA modules are connected in parallel to the basic large model. Each LoRA module corresponds to a driving or interaction action and is used to fine-tune the weights.

[0035] In one example, the weights used in the predictive model inference are: W = W0 + ΔW Where W0 is the frozen weight, and ΔW is the fine-tuning weight of the LoRA module.

[0036] In one example, the takeover action includes: the system plans to brake late, but the driver presses the brake pedal earlier; Intervention behaviors covered include: when the system-set air conditioning or ambient lighting parameters do not meet the user's needs, the user manually adjusts them via the panel.

[0037] In one example, the preference data extracted for a preset time period includes: Using the detected takeover or coverage intervention behavior as the time point, environmental perception data, the suboptimal trajectory predicted by the prediction model, and the user's actual preference trajectory are extracted within a time window from T seconds before the time point to T seconds after the time point.

[0038] In one example, determining whether a vehicle is in an idle state includes: The vehicle is in Park and charging mode; The vehicle enters a deep sleep state at night.

[0039] In one example, adjusting the prediction model based on preference data using an edge GPU includes: The adjustment goal is to increase the probability of the prediction model generating user preference trajectories, while decreasing the probability of generating predicted trajectories. The ΔW of the LoRA module in the prediction model is fine-tuned by minimizing the loss function.

[0040] In one example, before updating the adjusted prediction model by weight merging, a grayscale validation is performed on the adjusted prediction model.

[0041] In one example, gray-scale validation of the adjusted prediction model includes: The goodness-of-fit of the adjusted prediction model's output with the user's actual operation is verified by running the algorithm multiple times. If the goodness-of-fit exceeds the preset safety threshold, the adjusted prediction model is deemed to be ready for updating. Otherwise, the prediction model is readjusted.

[0042] Specifically, vehicle status data, environmental perception data, and user interaction behavior data are acquired in real time during vehicle operation.

[0043] Constructing predictive models for frozen bases and pluggable skill plugins, including: A frozen foundational model is deployed on a high-performance computing chip in the vehicle (such as a lightweight UniAD or Sparse4D architecture) to handle basic visual perception, BEV space construction, and basic trajectory planning. The network weights of this foundational model are completely frozen on the vehicle to ensure that the vehicle has essential safe driving and basic interaction capabilities, preventing catastrophic forgetting due to subsequent learning.

[0044] A skill expert plugin library is coupled in parallel, with multiple low-rank matrices (LoRA modules) with minimal parameters connected alongside specific network layers (such as the Attention module or FFN module of the Transformer). Each independent LoRA module corresponds to a specific driving or interaction skill, such as "aggressive lane-changing preference," "anti-cutting in congested traffic," or "dynamic response to specific ambient lighting." The weight calculation formula during actual model inference is as follows: W = W0 + ΔW Where W0 is the basic large model weight (frozen), and ΔW is the fine-tuning weight of the LoRA module.

[0045] Deploy a preference reward and evaluation model, and deploy a lightweight edge-side discrimination model to quantify the goodness of fit between "user's actual operation" and "system predicted operation", providing an evaluation basis for subsequent preference learning.

[0046] During daily driving or cockpit interaction, the prediction model and the currently mounted skill plugins run continuously in the background in shadow mode, outputting predicted actions in real time.

[0047] The system monitors in real time whether the driver engages in takeover or overriding intervention behaviors. Takeover behaviors include: the system plans to brake too late, but the driver presses the brake pedal prematurely. Overriding intervention behaviors include: the system sets air conditioning or ambient lighting parameters that do not meet the user's needs, and the user manually adjusts them via the panel.

[0048] When the trigger condition is met, the system automatically extracts the following data within a time window from T seconds before takeover to T seconds after takeover: environmental perception data (camera, radar, etc.); the trajectory predicted by the prediction model (labeled Y_reject); and the user's actual preferred trajectory (labeled Y_chosen). Y_chosen and Y_reject are encapsulated into a pair of preference training data and stored in the vehicle's local storage or a high-priority cache.

[0049] Determining whether a vehicle is in an idle state includes, but is not limited to: the vehicle being in Park (P) and charging; or the vehicle entering a deep sleep state at night.

[0050] When the above conditions are met, the system wakes up the edge GPU (such as the Tensor Core of the RTX architecture) and enters the "nighttime self-learning mode".

[0051] The Direct Preference Optimization (DPO) algorithm is used to directly use the preference data pair (Y_chosen, Y_reject) without the need for additional training of the reward model.

[0052] The adjustment objective is to increase the probability of the prediction model generating the user preference trajectory Y_chosen, while decreasing the probability of generating the predicted trajectory Y_reject. This is achieved by minimizing the DPO loss function, fine-tuning the parameters ΔW of the LoRA module corresponding to the skill to be updated, while keeping the frozen parameters W0 of the prediction model unchanged.

[0053] After fine-tuning, the adjusted LoRA module is first run in shadow mode, without directly taking over vehicle control. The prediction output of the adjusted prediction module is then compared with the user's actual operation to verify the fit, and this is done multiple times. It is determined whether the fit exceeds a preset safety threshold. If it does, the adjusted prediction model is deemed ready for updating; otherwise, the prediction model is readjusted.

[0054] Without restarting the prediction model or affecting normal vehicle operation, the old version of the skill LoRA module can be hot-swapped using weight merging. This enables the skill to self-evolve and achieves continuous personalized optimization of autonomous driving and cockpit interaction.

[0055] Example 2

[0056] Figure 2 A block diagram of an in-vehicle end-side model self-updating system based on user behavior feedback according to an embodiment of the present invention is shown.

[0057] like Figure 2 As shown, the in-vehicle end-side model self-updating system based on user behavior feedback includes: The data acquisition module is used to acquire real-time data during vehicle operation. Real-time data includes vehicle status data, environmental perception data, and user interaction behavior data. The large model base on the end side is used to construct a prediction model of the frozen base and pluggable skill plug-ins. The prediction model outputs predicted actions in real time during daily driving or cockpit interaction. The differential performance evaluation module is used to monitor in real time whether the driver has taken over or covered the intervention behavior. If so, it will extract and save the preference data for a preset time period. The local evolutionary fine-tuning module is used to adjust the prediction model based on preference data via the edge GPU during idle periods; The hot deployment control module is used to update the adjusted prediction model by merging weights.

[0058] In one example, the predictive model for building a frozen base and pluggable skill plugins includes: A basic large model is built on a high-computing chip in the vehicle, and the network weights of the basic large model are completely frozen on the vehicle. Multiple LoRA modules are connected in parallel to the basic large model. Each LoRA module corresponds to a driving or interaction action and is used to fine-tune the weights.

[0059] In one example, the weights used in the predictive model inference are: W = W0 + ΔW Where W0 is the frozen weight, and ΔW is the fine-tuning weight of the LoRA module.

[0060] In one example, the takeover action includes: the system plans to brake late, but the driver presses the brake pedal earlier; Intervention behaviors covered include: when the system-set air conditioning or ambient lighting parameters do not meet the user's needs, the user manually adjusts them via the panel.

[0061] In one example, the preference data extracted for a preset time period includes: Using the detected takeover or coverage intervention behavior as the time point, environmental perception data, the suboptimal trajectory predicted by the prediction model, and the user's actual preference trajectory are extracted within a time window from T seconds before the time point to T seconds after the time point.

[0062] In one example, determining whether a vehicle is in an idle state includes: The vehicle is in Park and charging mode; The vehicle enters a deep sleep state at night.

[0063] In one example, adjusting the prediction model based on preference data using an edge GPU includes: The adjustment goal is to increase the probability of the prediction model generating user preference trajectories, while decreasing the probability of generating predicted trajectories. The ΔW of the LoRA module in the prediction model is fine-tuned by minimizing the loss function.

[0064] In one example, before updating the adjusted prediction model by weight merging, a grayscale validation is performed on the adjusted prediction model.

[0065] In one example, gray-scale validation of the adjusted prediction model includes: The goodness-of-fit of the adjusted prediction model's output with the user's actual operation is verified by running the algorithm multiple times. If the goodness-of-fit exceeds the preset safety threshold, the adjusted prediction model is deemed to be ready for updating. Otherwise, the prediction model is readjusted.

[0066] Those skilled in the art should understand that the above description of the embodiments of the present invention is only intended to illustrate the beneficial effects of the embodiments of the present invention, and is not intended to limit the embodiments of the present invention to any of the examples given.

[0067] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A self-updating method for an in-vehicle terminal model based on user behavior feedback, characterized in that, include: Real-time data is acquired during vehicle operation, including vehicle status data, environmental perception data, and user interaction behavior data. Construct a prediction model for a frozen base and pluggable skill plugins. The prediction model outputs predicted actions in real time during daily driving or cockpit interaction. Real-time monitoring of whether the driver takes over or overrides intervention behavior; if so, the preference data for a preset time period is captured and saved. When the vehicle is in an idle state, the prediction model is adjusted by the edge GPU based on preference data; The adjusted prediction model is updated by merging weights.

2. The self-updating method for the vehicle-mounted terminal model based on user behavior feedback according to claim 1, wherein, The predictive model for building a frozen base and pluggable skill plugins includes: A basic large model is built on a high-computing chip in the vehicle, and the network weights of the basic large model are completely frozen on the vehicle. Multiple LoRA modules are connected in parallel to the basic large model. Each LoRA module corresponds to a driving or interaction action and is used to fine-tune the weights.

3. The self-updating method for the vehicle-mounted terminal model based on user behavior feedback according to claim 2, wherein, The weights used in predictive model inference are: W = W0 + ΔW Where W0 is the frozen weight, and ΔW is the fine-tuning weight of the LoRA module.

4. The in-vehicle terminal model self-updating method based on user behavior feedback according to claim 1, wherein, Takeover actions include: the system plans to brake late, but the driver presses the brake pedal prematurely; Intervention behaviors covered include: when the system-set air conditioning or ambient lighting parameters do not meet the user's needs, the user manually adjusts them via the panel.

5. The self-updating method for the vehicle-mounted terminal model based on user behavior feedback according to claim 1, wherein, The preference data extracted for a preset time period includes: Using the detected takeover or coverage intervention behavior as the time point, environmental perception data, the suboptimal trajectory predicted by the prediction model, and the user's actual preference trajectory are extracted within a time window from T seconds before the time point to T seconds after the time point.

6. The self-updating method for vehicle-side models based on user behavior feedback according to claim 1, wherein, Determining whether a vehicle is in an idle state includes: The vehicle is in Park and charging mode; The vehicle enters a deep sleep state at night.

7. The self-updating method for the vehicle-mounted terminal model based on user behavior feedback according to claim 3, wherein, Adjusting the prediction model based on preference data using edge GPUs includes: The adjustment goal is to increase the probability of the prediction model generating user preference trajectories, while decreasing the probability of generating predicted trajectories. The ΔW of the LoRA module in the prediction model is fine-tuned by minimizing the loss function.

8. The self-updating method for vehicle-side models based on user behavior feedback according to claim 1, wherein, Before updating the adjusted prediction model by weight merging, a gray-scale validation is performed on the adjusted prediction model.

9. The self-updating method for vehicle-side models based on user behavior feedback according to claim 8, wherein, The gray-scale validation of the adjusted prediction model includes: The goodness-of-fit of the adjusted prediction model's output with the user's actual operation is verified by running the algorithm multiple times. If the goodness-of-fit exceeds the preset safety threshold, the adjusted prediction model is deemed to be ready for updating. Otherwise, the prediction model is readjusted.

10. A vehicle-mounted end-side model self-updating system based on user behavior feedback, characterized in that, include: The data acquisition module is used to acquire real-time data during vehicle operation. Real-time data includes vehicle status data, environmental perception data, and user interaction behavior data. The large model base on the end side is used to construct a prediction model of the frozen base and pluggable skill plug-ins. The prediction model outputs predicted actions in real time during daily driving or cockpit interaction. The differential performance evaluation module is used to monitor in real time whether the driver has taken over or covered the intervention behavior. If so, it extracts and saves the preference data for a preset time period. The local evolutionary fine-tuning module is used to adjust the prediction model based on preference data via the edge GPU during idle periods; The hot deployment control module is used to update the adjusted prediction model by merging weights.

Images