Self-learning method of intelligent cockpit operating system and intelligent cockpit operating system

Abstract

The application provides a self-learning method of an intelligent cockpit operating system and the intelligent cockpit operating system, and the method comprises the following steps: obtaining interaction data generated in the running process of the intelligent cockpit operating system; generating interaction track data for representing interaction results based on the interaction data; determining a target self-learning strategy based on the interaction track data; generating target update data based on the target self-learning strategy and the interaction track data in the case that vehicle state data meets a preset self-learning access condition corresponding to the target self-learning strategy; and loading the target update data to the intelligent cockpit operating system, so that the intelligent cockpit operating system responds to a user interaction request based on the target update data. In this way, by generating the interaction track data based on the interaction data and generating and loading the target update data when the vehicle state meets the access condition, the intelligent cockpit operating system can realize safe and controllable vehicle-end self-learning update, thereby improving the individual adaptation capability and the interaction optimization efficiency.

Description

Technical Field

[0001] This application relates to the field of intelligent cockpit operating system technology, and in particular to a self-learning method for an intelligent cockpit operating system and an intelligent cockpit operating system. Background Technology

[0002] With the evolution of automotive electronic and electrical architecture, intelligent cockpits have entered a period of explosive growth in the application of large-scale language models in vehicles, and are widely used in diverse scenarios such as voice interaction, route planning, and vehicle control. In long-term practical use, users have placed higher demands on the personalization of cockpit system interaction, the accuracy of response, and the continuous growth of capabilities.

[0003] Existing technologies primarily employ a static inference architecture based on cloud-based training and OTA (Over-The-Air) distribution to update cockpit system capabilities. However, this centralized and fixed processing model struggles to provide localized optimization for the specific habits of different users, resulting in a lack of endogenous evolutionary mechanisms for the intelligent experience. This easily leads to a one-size-fits-all approach and a peak performance upon delivery. Furthermore, the optimization process under this model heavily relies on manual intervention to collect failure cases and fix them in the cloud. This not only results in iteration cycles lasting months but also prevents the system from promptly correcting user feedback due to a non-closed-loop feedback mechanism. In addition, existing solutions often require service restarts or even system image reloads when issuing updates, exhibiting high system intrusion, easily disrupting the continuity of cockpit services, and failing to effectively utilize the vehicle's vast idle computing power.

[0004] In summary, existing technologies cannot achieve vehicle-side self-learning and updates based on the interactive data generated during the operation of the intelligent cockpit operating system while ensuring vehicle safety and cockpit service continuity. This results in low personalization adaptability and interaction optimization efficiency of the intelligent cockpit operating system. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a self-learning method for an intelligent cockpit operating system and an intelligent cockpit operating system. By generating interactive trajectory data based on interactive data and generating and loading target update data when the vehicle status meets the access conditions, the intelligent cockpit operating system can achieve safe and controllable vehicle-side self-learning updates, thereby improving personalized adaptability and interaction optimization efficiency.

[0006] In a first aspect, the present invention provides a self-learning method for an intelligent cockpit operating system, comprising: Acquire interactive data generated during the operation of the intelligent cockpit operating system.

[0007] Interaction trajectory data is generated based on the interaction data to characterize the interaction results.

[0008] The target self-learning strategy is determined based on the interaction trajectory data; wherein, the target self-learning strategy includes a first self-learning strategy for adjusting the prompt information, and / or a second self-learning strategy for adjusting the model adaptation information.

[0009] Acquire vehicle status data, and if the vehicle status data meets the preset self-learning admission conditions corresponding to the target self-learning strategy, generate target update data based on the target self-learning strategy and interaction trajectory data.

[0010] The target update data is loaded into the intelligent cockpit operating system so that the intelligent cockpit operating system can respond to user interaction requests based on the target update data.

[0011] In an optional implementation, the interaction data includes interaction context data and interaction feedback data; the interaction feedback data includes explicit feedback data and / or implicit feedback data.

[0012] Acquire interactive data generated during the operation of the intelligent cockpit operating system, including: Acquire the interaction context data generated by the intelligent cockpit operating system in response to user interaction requests.

[0013] Explicit feedback data is generated when user input of error correction information, evaluation information, negative semantic information, or text feedback information is detected; and / or implicit feedback data is generated when user interruption of broadcast, skipping of output content, manual modification of system execution results, closure of executed functions, or failure to perform subsequent operations within a preset time period is detected.

[0014] The interaction context data and interaction feedback data are associated to determine the interaction data.

[0015] In an optional implementation, interaction trajectory data for characterizing the interaction result is generated based on the interaction data, including: The feedback type and evaluation information corresponding to the user's interaction request are determined based on the interaction feedback data; the feedback evaluation information includes the feedback score and / or anomaly label.

[0016] The system associates user request information, system response information, action execution information, vehicle environment information, feedback type, and feedback evaluation information.

[0017] Based on the association results, interaction trajectory data corresponding to user interaction requests is generated.

[0018] In an optional implementation, the target self-learning strategy is determined based on the interaction trajectory data, including: Based on the feedback type, feedback score, and anomaly label in the interaction trajectory data, the interaction anomaly type corresponding to the interaction trajectory data is determined.

[0019] Based on the pre-defined correspondence between interaction exception types and self-learning strategies, the self-learning strategy corresponding to each interaction exception type is determined.

[0020] If the self-learning strategy corresponding to the interaction exception type is the first self-learning strategy, then the first self-learning strategy will be determined as the target self-learning strategy.

[0021] If the self-learning strategy corresponding to the interaction exception type is the second self-learning strategy, then the second self-learning strategy will be determined as the target self-learning strategy.

[0022] When the self-learning strategy corresponding to the interaction exception type includes a first self-learning strategy and a second self-learning strategy, the first self-learning strategy and the second self-learning strategy are determined as the target self-learning strategy.

[0023] In an optional implementation, when the target self-learning strategy includes a first self-learning strategy, target update data is generated based on the target self-learning strategy and the interaction trajectory data, including: Obtain target interaction trajectory data with the same abnormal label or belonging to the same interaction abnormal type from the interaction trajectory data.

[0024] The content to be adjusted in the prompt message is determined based on the target interaction trajectory data.

[0025] The current prompt message is adjusted based on the content to be adjusted, resulting in an updated prompt message.

[0026] The updated message is identified as the target update data.

[0027] In an optional implementation, when the target self-learning strategy includes a second self-learning strategy, generating target update data based on the target self-learning strategy and interaction trajectory data includes: Filter target interaction trajectory data that meet preset quality conditions from the interaction trajectory data.

[0028] Supervised training samples are constructed based on target interaction trajectory data; wherein, the supervised training samples include input feature data and target output label data.

[0029] The model adaptation information in the intelligent cockpit operating system is trained based on supervised training samples to obtain updated model adaptation information.

[0030] The updated model adaptation information is determined as the target update data.

[0031] In an optional implementation, if the vehicle state data meets the preset self-learning admission conditions corresponding to the target self-learning strategy, target update data is generated based on the target self-learning strategy and the interaction trajectory data, including: The corresponding preset self-learning admission conditions are determined based on the target self-learning strategy.

[0032] If the vehicle status data indicates that the vehicle is in a non-driving state, the power supply status meets the preset power supply conditions, the processor temperature meets the preset temperature conditions, and the computing power utilization rate meets the preset computing power conditions, then the vehicle status data is determined to meet the preset self-learning admission conditions.

[0033] If the vehicle status data meets the preset self-learning admission conditions, the self-learning process corresponding to the target self-learning strategy is initiated to generate target update data.

[0034] In an optional implementation, the method further includes the following steps in generating target update data based on the target self-learning strategy and interaction trajectory data: Monitor vehicle operating status and hardware resource status.

[0035] If the vehicle's operating status or hardware resource status meets the preset interruption conditions, the current self-learning process is suspended or terminated, and the breakpoint status information corresponding to the current self-learning process is saved.

[0036] If the preset self-learning admission conditions are met again, the current self-learning process will resume based on the breakpoint status information.

[0037] In an optional implementation, loading the target update data into the smart cockpit operating system includes: Load the target update data into the alternate loading area.

[0038] If the inference call chain of the intelligent cockpit operating system meets the preset switching conditions, the inference call chain will be switched from the current loading area to the backup loading area so that the intelligent cockpit operating system can respond to user interaction requests based on target update data.

[0039] In a second aspect, the present invention provides an intelligent cockpit operating system, comprising: The data acquisition module is used to acquire interactive data generated during the operation of the intelligent cockpit operating system.

[0040] The trajectory generation module is used to generate interactive trajectory data that represents the interaction results based on the interactive data.

[0041] The strategy determination module is used to determine the target self-learning strategy based on the interaction trajectory data; wherein, the target self-learning strategy includes a first self-learning strategy for adjusting the prompt information, and / or a second self-learning strategy for adjusting the model adaptation information.

[0042] The update generation module is used to acquire vehicle status data and, if the vehicle status data meets the preset self-learning admission conditions corresponding to the target self-learning strategy, generate target update data based on the target self-learning strategy and interaction trajectory data.

[0043] The loading module is used to load target update data into the smart cockpit operating system, so that the smart cockpit operating system can respond to user interaction requests based on the target update data.

[0044] This application provides a self-learning method and an intelligent cockpit operating system. By acquiring interaction data during the operation of the intelligent cockpit operating system and generating interaction trajectory data, the requests, responses, and feedback during user interaction can be transformed into a data foundation that can be used for self-learning analysis. By determining a first self-learning strategy for adjusting prompt information and / or a second self-learning strategy for adjusting model adaptation information based on the interaction trajectory data, lightweight prompt optimization or model adaptation optimization can be performed for different types of interaction problems, thereby improving the adaptability of the intelligent cockpit operating system to user habits, interaction preferences, and abnormal feedback. By generating target update data when vehicle status data meets preset self-learning admission conditions, the self-learning process can be ensured to be executed under safe vehicle status and suitable resource conditions, thereby reducing the impact of the self-learning process on vehicle operation safety and system stability. By loading the target update data into the intelligent cockpit operating system, the intelligent cockpit operating system can respond to subsequent user interaction requests based on the update results, thereby improving the personalized response capability, interaction optimization efficiency, and long-term user experience of the intelligent cockpit operating system.

[0045] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application are realized and obtained through the structures particularly pointed out in the description, claims, and drawings.

[0046] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0048] Figure 1A flowchart illustrating the self-learning method of the intelligent cockpit operating system provided in this application embodiment; Figure 2 A flowchart of the interactive data acquisition method provided in the embodiments of this application; Figure 3 A flowchart of the interaction trajectory data generation method provided in the embodiments of this application; Figure 4 A flowchart of the target self-learning strategy determination method provided in the embodiments of this application; Figure 5 A schematic diagram of the intelligent cockpit operating system provided in an embodiment of this application.

[0049] Icons: 1-Data acquisition module; 2-Trajectory generation module; 3-Strategy determination module; 4-Update generation module; 5-Loading module. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] To help those skilled in the art better understand this application, a brief introduction to its application scenarios and design concepts is provided.

[0052] As intelligent cockpit operating systems gradually integrate functions such as voice interaction, navigation planning, vehicle control, entertainment recommendations, and large-scale model inference, users continuously generate a large amount of interaction data related to the cockpit system during vehicle use. For example, users may interrupt the system's response during voice interaction, manually modify the route planned by the system during navigation, close the function just executed by the system during vehicle control, or express dissatisfaction with the system's response through error correction, evaluation, or text feedback. These interactive behaviors reflect users' real usage habits, interaction preferences, and problems in the system's response results, forming an important data foundation for the continuous optimization of intelligent cockpit operating systems.

[0053] However, existing smart cockpit operating systems typically respond to user interaction requests using pre-trained models, fixed prompt word templates, or large cloud-based model interfaces. When optimizing cockpit system capabilities, they usually rely on developers collecting abnormal interaction cases, adjusting prompt words, training models, or updating versions in the cloud, and then distributing the changes to vehicles remotely. While this approach can achieve basic human-machine interaction functions, user feedback generated during actual vehicle use is often difficult to convert into data needed for system self-learning in a timely manner. This results in the system being unable to quickly correct interaction errors based on user feedback and struggling to develop personalized adaptation capabilities for different users' long-term usage habits.

[0054] Meanwhile, the self-learning processing of the intelligent cockpit operating system is also constrained by the vehicle's status. Performing self-learning processing while the vehicle is in motion, the power supply is unstable, the processor temperature is high, or computing resources are heavily consumed may affect vehicle safety or the stability of the cockpit system. Uploading large amounts of interaction data to the cloud for learning can easily increase communication costs and create pressure on user privacy protection. Furthermore, existing systems often require restarting the application, reloading services, or performing a full vehicle software upgrade when loading updated prompts or model parameters, which can easily disrupt the intelligent cockpit operating system's continuous response to user interaction requests.

[0055] Based on this, embodiments of this application provide a self-learning method and an intelligent cockpit operating system. The method transforms the interaction data generated during the operation of the intelligent cockpit operating system into interaction trajectory data that can characterize the interaction results, and determines the corresponding self-learning strategy based on the interaction trajectory data. Specifically, for problems requiring adjustments to interaction expression, response style, or prompt constraints, a first self-learning strategy for adjusting prompt information is adopted; for problems requiring learning long-term user preferences, task execution patterns, or model adaptation capabilities, a second self-learning strategy for adjusting model adaptation information is adopted. Thus, the intelligent cockpit operating system can select the corresponding self-learning processing method according to different types of interaction problems, improving the targeted nature of self-learning updates.

[0056] Furthermore, this application introduces vehicle state data and preset self-learning admission conditions before generating target update data. Target update data is only generated based on the target self-learning strategy and interaction trajectory data when the vehicle state data meets the corresponding admission conditions. Based on this, self-learning processing can be executed under conditions of vehicle safety, stable power supply, and adequate processor temperature and computing resources, thereby reducing the impact of self-learning processing on vehicle operation safety and cockpit system stability.

[0057] Based on this, this application loads the generated target update data into the intelligent cockpit operating system, enabling the intelligent cockpit operating system to respond to subsequent user interaction requests based on the target update data. Therefore, this application can achieve vehicle-side self-learning updates based on user interaction data, while ensuring vehicle safety and cockpit service continuity. This allows the intelligent cockpit operating system to gradually adapt to user interaction habits and usage preferences, thereby improving the personalized responsiveness, interaction optimization efficiency, and long-term user experience of the intelligent cockpit operating system.

[0058] To facilitate understanding of this embodiment, the embodiments of this application will be described in detail below.

[0059] This application provides a self-learning method for an intelligent cockpit operating system. The intelligent cockpit operating system may include at least one of the following: voice interaction function, navigation function, vehicle control function, entertainment media function, information recommendation function, and large model inference function.

[0060] Reference Figure 1 The self-learning method for the intelligent cockpit operating system provided in this application includes: Step S101: Obtain the interaction data generated during the operation of the intelligent cockpit operating system.

[0061] Here, interaction data refers to data generated by the intelligent cockpit operating system during operation, reflecting the user interaction process, system response process, task execution process, and feedback results. Interaction data can include interaction context data and interaction feedback data. Interaction context data characterizes the context in which the intelligent cockpit operating system processes user interaction requests. Interaction feedback data characterizes the user's feedback to the intelligent cockpit operating system's response, or characterizes changes in user behavior after the system's response.

[0062] After receiving a user interaction request, the intelligent cockpit operating system records the corresponding user request information. This user request information can include at least one of the following: voice-recognized text, touch input, button input, gesture input, application interface input, and multimodal input. For example, if a user inputs "navigate home and buy coffee along the way" via voice, the intelligent cockpit operating system uses the voice recognition result as the user request information. As another example, if a user taps the air conditioning control on the central control screen, the intelligent cockpit operating system uses the touch location, control icon, and adjustment target as the user request information.

[0063] When responding to user interaction requests, the intelligent cockpit operating system records system response information. This information may include at least one of the following: model input information, prompts, model output information, voice announcements, interface displays, navigation planning results, recommendations, vehicle control commands, and task processing results. The system response information reflects the content of the intelligent cockpit operating system's response to user interaction requests.

[0064] When executing a response, the intelligent cockpit operating system records execution action information. This information may include at least one of the following: target functional module, called interface, tool call result, vehicle control message, navigation route execution status, media playback status, task execution success status, task execution failure status, response time, and error code. This execution action information reflects whether the intelligent cockpit operating system completed the corresponding task according to the user's interaction request, and whether any abnormalities occurred during task execution.

[0065] The intelligent cockpit operating system also acquires vehicle environmental information. This information may include at least one of the following: vehicle gear status, vehicle speed, vehicle position, vehicle driving status, screen status, occupant status, door status, window status, air conditioning status, charging status, low-voltage battery voltage, power battery charge, processor temperature, computing power utilization, available storage space, and network connection status. This vehicle environmental information provides the basis for subsequent interaction result judgments and self-learning access decisions.

[0066] Interactive feedback data includes explicit feedback data and implicit feedback data. Explicit feedback data is generated when users actively express their feedback intentions. The intelligent cockpit operating system generates explicit feedback data when it detects user input of error correction information, evaluation information, negative semantic information, or text feedback information. For example, when a user clicks the "correction" control, submits a "dissatisfied" rating, enters "this route is incorrect," or verbally says "incorrect," "wrong," or "answer again," the intelligent cockpit operating system generates explicit feedback data. Explicit feedback data directly reflects the user's subjective evaluation of the system's response results.

[0067] Implicit feedback data is data generated when users do not directly input evaluation content, but their behavior reflects problems with the system's response or that the system's response matches their preferences. The intelligent cockpit operating system generates implicit feedback data when it detects that a user interrupts the broadcast, skips output content, manually modifies the system's execution result, closes an executed function, or fails to perform a subsequent operation within a preset time. For example, if the user interrupts the broadcast shortly after it begins, the intelligent cockpit operating system will treat the interruption as implicit feedback data. Another example is when the intelligent cockpit operating system generates a navigation route, and the user manually drags waypoints or adds avoidable areas; the intelligent cockpit operating system will treat this route modification as implicit feedback data. Yet another example is when the intelligent cockpit operating system activates the seat heating function based on a voice request, and the user immediately deactivates the seat heating function; the intelligent cockpit operating system will treat the deactivation of the executed function as implicit feedback data.

[0068] In an optional implementation, the intelligent cockpit operating system acquires interaction data through a background listening thread. This background listening thread is activated when a user interaction request enters the inference call chain of the intelligent cockpit operating system, and continuously monitors user feedback behavior, application status changes, and vehicle status changes within a preset time window. The background listening thread binds user request information, system response information, execution action information, vehicle environment information, and interaction feedback data corresponding to the same user interaction request to ensure that data from the same interaction process is completely recorded.

[0069] In an optional implementation, the intelligent cockpit operating system acquires interaction data through a gateway proxy service. The gateway proxy service resides between the business function module and the model inference module. When a business function module sends an inference request to the model inference module, the gateway proxy service intercepts the inference request and records the corresponding input, prompts, model output, tool call results, and feedback behavior. In this way, the intelligent cockpit operating system can collect interaction data without altering the calling logic of the business function modules.

[0070] In an optional implementation, the intelligent cockpit operating system preprocesses the interaction data. Preprocessing may include at least one of the following: data cleaning, data anonymization, timestamp alignment, user identifier mapping, scene identifier addition, outlier removal, duplicate data merging, and format conversion. Through preprocessing, the interaction data can meet the input requirements for subsequent interaction trajectory generation and self-learning processing.

[0071] Step S102: Generate interaction trajectory data to characterize the interaction results based on the interaction data.

[0072] Here, interaction trajectory data refers to structured representations of one or more user interaction processes. Interaction trajectory data is used to characterize the relationships between user requests, system responses, task execution, vehicle environment, and feedback results. The intelligent cockpit operating system generates interaction trajectory data based on interaction data, transforming fragmented user behavior and system operation data into data objects that can be used for self-learning analysis.

[0073] The intelligent cockpit operating system determines the feedback type and feedback evaluation information corresponding to user interaction requests based on interactive feedback data. The feedback type indicates the feedback source or method. Feedback types can include explicit and implicit feedback types. Feedback evaluation information indicates the user's evaluation of the system's response. Feedback evaluation information can include at least one of the following: feedback score, anomaly label, success label, number of feedback occurrences, time of feedback occurrence, and feedback description.

[0074] The intelligent cockpit operating system can determine feedback scores and anomaly labels based on explicit feedback data. For example, when a user clicks the "Dissatisfied" control, the intelligent cockpit operating system sets the feedback score to a negative value and labels the response as unsatisfactory. When a user inputs "I misspoke," the intelligent cockpit operating system labels the response as semantic understanding error or incorrect answer. When a user inputs "Re-answer," the intelligent cockpit operating system labels the response as output content not meeting expectations.

[0075] The intelligent cockpit operating system can also determine feedback scores and anomaly labels based on implicit feedback data. For example, if a user interrupts the voice broadcast within a preset time after it begins, the intelligent cockpit operating system will label it as either "output length error" or "broadcast content error." If a user manually modifies the route after it has been generated, the intelligent cockpit operating system will label it as "path planning error." If a user immediately cancels the control result after the system has executed vehicle control, the intelligent cockpit operating system will label it as either "control intent understanding error" or "execution result error." If the user does not perform any subsequent operation within a preset time after the system responds, the intelligent cockpit operating system can set the feedback score to a neutral score, a negative score, or a success score, depending on the specific scenario.

[0076] The intelligent cockpit operating system correlates user request information, system response information, action execution information, vehicle environment information, feedback type, and feedback evaluation information, and generates interaction trajectory data corresponding to the user's interaction request based on the correlation results. Interaction trajectory data may include at least one of the following: trajectory identifier, timestamp, user identifier, vehicle identifier, scene identifier, user request content, system response content, action execution information, execution status, vehicle status snapshot, feedback type, feedback score, and anomaly label. The trajectory identifier uniquely identifies an interaction process. The scene identifier characterizes the business scenario corresponding to the interaction trajectory data, such as a navigation scenario, voice interaction scenario, vehicle control scenario, entertainment media scenario, or information recommendation scenario. The vehicle status snapshot records the vehicle environment information at the time of the interaction.

[0077] In an optional implementation, the interaction trajectory data is stored in a structured data format. This structured data format can be key-value pair, table, object, or message format. For example, the interaction trajectory data may include trajectory identifiers, timestamps, scene identifiers, user request content, prompts, model output, tool call results, execution status, feedback scores, and anomaly labels. The structured data format improves the efficiency of subsequent querying, filtering, statistics, and training sample construction.

[0078] In an optional implementation, the interaction trajectory data can also be stored in association with corresponding unstructured data. Unstructured data may include at least one of the following: voice clips, image clips, screenshots, log files, route geometry data, and model inference logs. The intelligent cockpit operating system associates structured and unstructured data through trajectory identifiers to enable retrospective analysis of abnormal interaction processes when needed.

[0079] In an optional implementation, the intelligent cockpit operating system categorizes and labels the interaction trajectory data based on feedback evaluation information. The categorization labels may include at least one of the following: successful interaction labels, failed interaction labels, low-confidence interaction labels, high-quality sample labels, samples requiring optimization labels, and abnormal sample labels. Through categorization and labeling, the intelligent cockpit operating system can quickly filter data suitable for either a first self-learning strategy or a second self-learning strategy during subsequent self-learning processing.

[0080] In an optional implementation, the intelligent cockpit operating system indexes the interaction trajectory data. The index may include at least one of the following: a time index, a scene index, a user index, an anomaly label index, a feedback score index, and a self-learning strategy index. Through this indexing, the intelligent cockpit operating system can count the number of times the same scene or the same anomaly label appears within a preset time window and determine whether the conditions for triggering self-learning processing have been met.

[0081] Step S103: Determine the target self-learning strategy based on the interaction trajectory data; wherein, the target self-learning strategy includes a first self-learning strategy for adjusting the prompt information, and / or a second self-learning strategy for adjusting the model adaptation information.

[0082] Here, the target self-learning strategy refers to the self-learning processing method determined by the intelligent cockpit operating system based on interaction trajectory data, used to generate target update data. The target self-learning strategy includes at least one of a first self-learning strategy and a second self-learning strategy. The first self-learning strategy is used to adjust prompt information. The second self-learning strategy is used to adjust model adaptation information.

[0083] Prompt information refers to at least one of the following: instructions, constraints, contextual supplementary information, role description information, task boundary information, output format information, and style control information used by the intelligent cockpit operating system when invoking models or business function modules. Adjusting prompt information can change the intelligent cockpit operating system's response style, output length, tone, task priority, business boundaries, and security constraints. Model adaptation information refers to at least one of the following: model configuration data, model incremental parameters, adapter parameters, routing configuration, feature mapping relationships, and personalized preference parameters used to adapt the model to vehicle business scenarios, user preferences, or task execution logic. Adjusting model adaptation information can change the intelligent cockpit operating system's response to complex scenario rules, long-term user preferences, and task execution logic.

[0084] The intelligent cockpit operating system determines the interaction anomaly type corresponding to the interaction trajectory data based on the feedback type, feedback score, and anomaly label. Interaction anomaly types can include at least one of the following: verbose output anomaly, interaction style anomaly, semantic understanding anomaly, path planning anomaly, task execution anomaly, vehicle control anomaly, recommendation result anomaly, response timeout anomaly, and user preference mismatch anomaly. Interaction anomaly types characterize the categories of problems that the intelligent cockpit operating system needs to optimize.

[0085] The intelligent cockpit operating system determines the corresponding self-learning strategy based on a preset correspondence between interaction anomaly types and self-learning strategies. When the self-learning strategy corresponding to an interaction anomaly type is the first self-learning strategy, the intelligent cockpit operating system designates the first self-learning strategy as the target self-learning strategy. When the self-learning strategy corresponding to an interaction anomaly type is the second self-learning strategy, the intelligent cockpit operating system designates the second self-learning strategy as the target self-learning strategy. When the self-learning strategy corresponding to an interaction anomaly type includes both the first and second self-learning strategies, the intelligent cockpit operating system designates both the first and second self-learning strategies as the target self-learning strategies.

[0086] For example, when the interaction anomaly type is a verbose output anomaly, the intelligent cockpit operating system determines that the length or expression of the response content needs adjustment, and the target self-learning strategy may include a first self-learning strategy. The intelligent cockpit operating system subsequently limits the output length or changes the response style by adjusting the prompt information. As another example, when the interaction anomaly type is a route planning anomaly, and the same user manually modifies the route multiple times in the same commuting scenario, the intelligent cockpit operating system determines that the model adaptation information needs to learn the user's long-term route preferences, and the target self-learning strategy may include a second self-learning strategy. Furthermore, when the interaction anomaly type involves both response expression issues and long-term user preference issues, the target self-learning strategy may include both the first and second self-learning strategies simultaneously.

[0087] In an optional implementation, the intelligent cockpit operating system determines the target self-learning strategy based on the frequency of anomaly occurrences. The intelligent cockpit operating system counts the number of times the same anomaly label or the same interaction anomaly type occurs within a preset time window. When the number of occurrences reaches a first threshold, the intelligent cockpit operating system determines the first self-learning strategy as the target self-learning strategy. When the number of occurrences reaches a second threshold, and there is interaction trajectory data that meets the training quality requirements, the intelligent cockpit operating system determines the second self-learning strategy as the target self-learning strategy. The second threshold can be greater than the first threshold to ensure that the update of model adaptation information is based on more stable data.

[0088] In an optional implementation, the intelligent cockpit operating system determines the target self-learning strategy based on the feedback score. When the feedback score is lower than a first score threshold, the intelligent cockpit operating system identifies the corresponding interaction trajectory data as data requiring optimization. If the corresponding problem can be solved by changing the output constraints, expression method, or task boundary, the intelligent cockpit operating system determines that the target self-learning strategy includes a first self-learning strategy. If the corresponding problem requires changing the model's adaptability to user preferences, scene rules, or task logic, the intelligent cockpit operating system determines that the target self-learning strategy includes a second self-learning strategy.

[0089] In optional implementations, the intelligent cockpit operating system determines the target self-learning strategy by combining scene identifiers. For scenarios involving voice broadcasting, question-and-answer responses, and information recommendation, the intelligent cockpit operating system may prioritize the first self-learning strategy. For scenarios involving navigation planning, complex task orchestration, multi-turn dialogue, and cross-functional linkage, the intelligent cockpit operating system may prioritize the second self-learning strategy. For scenarios involving both expression methods and complex task logic, the intelligent cockpit operating system may employ both the first and second self-learning strategies simultaneously.

[0090] Step S104: Obtain vehicle status data, and if the vehicle status data meets the preset self-learning admission conditions corresponding to the target self-learning strategy, generate target update data based on the target self-learning strategy and interaction trajectory data.

[0091] Here, vehicle status data refers to data used to determine whether the intelligent cockpit operating system is suitable for performing self-learning processing. Vehicle status data may include at least one of the following: vehicle gear position, vehicle speed, power supply status, charging status, occupant status, screen status, processor temperature, computing power utilization, available storage space, low-voltage battery voltage, power battery charge, vehicle ignition status, vehicle door and window status, and network connectivity status. By acquiring vehicle status data, the intelligent cockpit operating system determines whether self-learning processing will affect vehicle operational safety, power supply safety, thermal safety, storage safety, and cockpit service continuity.

[0092] The intelligent cockpit operating system determines corresponding preset self-learning admission conditions based on the target self-learning strategy. These preset self-learning admission conditions refer to the vehicle and resource state conditions that must be met before initiating the self-learning process corresponding to the target self-learning strategy. The preset self-learning admission conditions for different target self-learning strategies can be the same or different. Since the computational load corresponding to the first self-learning strategy is usually less than that corresponding to the second self-learning strategy, the preset self-learning admission conditions for the first self-learning strategy can be lower than those for the second self-learning strategy.

[0093] When the target self-learning strategy includes a first self-learning strategy, the intelligent cockpit operating system can determine that the vehicle state data meets the preset self-learning admission conditions corresponding to the first self-learning strategy when the vehicle is in a non-driving state, the power supply status meets preset power supply conditions, the processor temperature meets preset temperature conditions, and the computing power utilization rate meets preset computing power conditions. A non-driving state can include the vehicle being parked or at zero speed. Meeting the preset power supply conditions can include the low-voltage battery voltage being greater than a first voltage threshold, or the power battery charge being greater than a first charge threshold. Meeting the preset temperature conditions can include the processor temperature being lower than a first temperature threshold. Meeting the preset computing power utilization rate can include the computing power utilization rate being lower than a first utilization rate threshold.

[0094] When the target self-learning strategy includes a second self-learning strategy, the intelligent cockpit operating system can determine that the vehicle state data meets the preset self-learning admission conditions corresponding to the second self-learning strategy when the vehicle is parked, charging, unoccupied, the screen is off, the power supply meets higher preset power supply conditions, the processor temperature meets higher preset temperature conditions, the computing power utilization meets higher preset computing power conditions, and the available storage space is greater than a preset storage threshold. By setting stricter admission conditions, the training process corresponding to the second self-learning strategy can be executed in a safer and more stable vehicle environment.

[0095] When vehicle status data does not meet the preset self-learning admission conditions corresponding to the target self-learning strategy, the intelligent cockpit operating system will not initiate the corresponding self-learning process. The intelligent cockpit operating system can delay the self-learning process, lower its priority, continue monitoring vehicle status data, or record the reason for not initiating the process. In this way, the intelligent cockpit operating system avoids performing self-learning processing under conditions of driving, insufficient power supply, excessive temperature, or limited computing resources.

[0096] When the vehicle status data meets the preset self-learning admission conditions corresponding to the target self-learning strategy, the intelligent cockpit operating system generates target update data based on the target self-learning strategy and interaction trajectory data. Target update data refers to data used to update the responsiveness of the intelligent cockpit operating system. Target update data may include at least one of updated prompt information and updated model adaptation information.

[0097] When the target self-learning strategy includes the first self-learning strategy, the intelligent cockpit operating system obtains target interaction trajectory data with the same anomaly label or belonging to the same interaction anomaly type from the interaction trajectory data. Based on the target interaction trajectory data, the intelligent cockpit operating system determines the content to be adjusted for the prompt information. The content to be adjusted may include at least one of the following: output length constraints, response format constraints, tone constraints, task execution priority, prohibition conditions, supplementary explanation conditions, safety boundary conditions, and tool call constraints. Based on the content to be adjusted, the intelligent cockpit operating system adjusts the current prompt information to obtain the updated prompt information, and determines the updated prompt information as the target update data.

[0098] For example, multiple target interaction trajectory data indicate that users frequently interrupt long announcements. The intelligent cockpit operating system determines that the content to be adjusted in the prompts should be shortening the response length and prioritizing the output of conclusions. The intelligent cockpit operating system adds a concise response constraint to the current prompts, resulting in updated prompts. As another example, multiple target interaction trajectory data indicate that users do not require lengthy explanations of the route during navigation tasks. The intelligent cockpit operating system adds a constraint to prioritize the execution of navigation tasks to the prompts.

[0099] When the target self-learning strategy includes a second self-learning strategy, the intelligent cockpit operating system filters target interaction trajectory data that meets preset quality conditions from the interaction trajectory data. Preset quality conditions may include at least one of the following: feedback score reaches a preset score, execution status is successful, user correction result can be used as target output, data fields are complete, scene identification is clear, anomaly labels are valid, and manual confirmation is passed. The intelligent cockpit operating system constructs supervised training samples based on the target interaction trajectory data. Supervised training samples include input feature data and target output label data. Input feature data may include at least one of the following: user request content, scene identification, vehicle environment information, and historical preference features. Target output label data may include at least one of the following: corrected system response content, user-confirmed execution result, corrected navigation route, target control action, or high-quality response content.

[0100] The intelligent cockpit operating system trains the model adaptation information based on supervised training samples to obtain updated model adaptation information. Model adaptation information may include at least one of the following: model adaptation parameters, model incremental weights, adapter weights, personalized preference parameters, scene rule parameters, and routing configuration parameters. The training process may employ at least one of the following methods: low-rank adaptation training, incremental training, preference learning, supervised fine-tuning, efficient parameter fine-tuning, or rule parameter updating. The intelligent cockpit operating system determines the updated model adaptation information as the target update data.

[0101] In an optional implementation, the intelligent cockpit operating system performs version verification during the generation of target update data. Version verification may include at least one of the following: format verification, compatibility verification, security policy verification, size verification, inference latency verification, and effect evaluation verification. If the target update data passes version verification, the intelligent cockpit operating system marks it as loadable. If the target update data fails version verification, the intelligent cockpit operating system either discards the target update data or regenerates it.

[0102] In an optional implementation, the intelligent cockpit operating system monitors the vehicle's operating status and hardware resource status during the self-learning process. When the vehicle's operating status or hardware resource status meets preset interruption conditions, the intelligent cockpit operating system suspends or terminates the current self-learning process and saves the breakpoint status information corresponding to the current self-learning process. Preset interruption conditions may include at least one of the following: the vehicle switches from a non-driving state to a driving state, a door is opened, the brake pedal is depressed, the low-voltage battery voltage is lower than an interruption threshold, the power battery charge is lower than an interruption threshold, the processor temperature exceeds an interruption threshold, the computing power utilization rate exceeds an interruption threshold, and the status of the occupants changes. Breakpoint status information may include at least one of the following: current training step size, task progress, temporary parameters, data processing location, verification information, and task status identifier. When the preset self-learning admission conditions are met again, the intelligent cockpit operating system resumes the execution of the current self-learning process based on the breakpoint status information. In this way, the intelligent cockpit operating system can reduce data corruption and redundant calculations caused by abnormal interruptions.

[0103] Step S105: Load the target update data into the intelligent cockpit operating system so that the intelligent cockpit operating system can respond to user interaction requests based on the target update data.

[0104] Here, after generating the target update data, the intelligent cockpit operating system loads the target update data into itself. Loading the target update data means enabling the intelligent cockpit operating system to use the target update data when responding to user interaction requests. The target update data can be loaded into at least one of the following modules: prompt information management module, model adaptation module, inference call chain, routing configuration module, business function module, or gateway proxy module.

[0105] When the target updated data includes updated prompts, the intelligent cockpit operating system loads the updated prompts into the prompt management module or the inference call chain. When processing subsequent user interaction requests, the intelligent cockpit operating system injects the updated prompts into the model input or business request, enabling it to generate response results based on the updated prompts. For example, if the updated prompts include a concise answer constraint, the intelligent cockpit operating system will output a more concise answer in subsequent voice question-and-answer scenarios.

[0106] When the target update data includes updated model adaptation information, the intelligent cockpit operating system loads the updated model adaptation information into the model adaptation module or inference call chain. When processing user interaction requests subsequently, the intelligent cockpit operating system adjusts the model inference results or task execution logic based on the updated model adaptation information. For example, if the updated model adaptation information includes the user's commuting route preferences, the intelligent cockpit operating system generates route results that better match the user's historical preferences in subsequent navigation planning scenarios.

[0107] In an optional implementation, the intelligent cockpit operating system loads the target update data into a backup loading area, and switches the inference call chain from the current loading area to the backup loading area when it detects that the inference call chain of the intelligent cockpit operating system meets preset switching conditions. The backup loading area is used to store the target update data. The current loading area is used to provide the currently effective prompt information or model adaptation information for the current inference call chain. The preset switching conditions may include at least one of the following: the inference call chain is in an idle state, the current request processing is completed, and the target update data meets the requirements of the verification and switching time window. By switching between the backup loading area and the current loading area, the intelligent cockpit operating system can reduce the impact of the target update data loading process on the user interaction requests being executed.

[0108] In an optional implementation, the intelligent cockpit operating system performs signature verification, integrity verification, and compatibility verification on the target update data before loading it. Signature verification confirms the source of the target update data is trustworthy. Integrity verification confirms the target update data is not corrupted. Compatibility verification confirms the target update data can be recognized and accessed by the current intelligent cockpit operating system. If the target update data passes verification, the intelligent cockpit operating system loads it. If the target update data fails verification, the intelligent cockpit operating system does not load it.

[0109] In an optional implementation, the intelligent cockpit operating system continuously monitors the operational effectiveness of the target update data after loading it. The operational effectiveness may include at least one of the following: response success rate, user feedback score, frequency of anomaly tags, response time, resource utilization, and system stability indicators. When the target update data leads to an increase in anomaly feedback, an increase in response time, or a decrease in system stability, the intelligent cockpit operating system can stop using the target update data or revert to the prompts or model adaptation information prior to loading the target update data. In this way, the intelligent cockpit operating system can improve the operational reliability after self-learning updates.

[0110] The self-learning method for the intelligent cockpit operating system provided in this application converts interactive data generated during operation into interactive trajectory data, and determines a first self-learning strategy and / or a second self-learning strategy based on the interactive trajectory data. When vehicle status data meets preset self-learning admission conditions, the intelligent cockpit operating system generates target update data and loads it into the operating system, enabling it to use the target update data in subsequent user interaction requests. Therefore, the intelligent cockpit operating system can achieve self-learning updates based on real user interaction data while ensuring vehicle status safety and cockpit service continuity, improving the intelligent cockpit operating system's personalized adaptability, interaction optimization efficiency, and long-term user experience.

[0111] In an optional implementation, the interaction data includes interaction context data and interaction feedback data; the interaction feedback data includes explicit feedback data and / or implicit feedback data.

[0112] Here, interaction context data describes the input content, system processing, vehicle environment status, and execution results of the intelligent cockpit operating system in response to user interaction requests. Interaction feedback data describes the user's proactive evaluation or behavioral feedback on the intelligent cockpit operating system's response results.

[0113] Reference Figure 2 Step S101 includes the following steps S201-S203.

[0114] Step S201: Obtain the interaction context data generated by the intelligent cockpit operating system in response to user interaction requests.

[0115] Here, user interaction requests can originate from voice channels, touch channels, physical button channels, gesture channels, in-vehicle application interfaces, or a combination of multiple channels. User interaction requests corresponding to the voice channel can include structured text output by ASR (Automatic Speech Recognition), emotion tags corresponding to the user's voice, speech recognition confidence, and user identifiers corresponding to voiceprint features. User interaction requests corresponding to the touch channel can include click coordinates on the central control screen, swipe trajectory, long press duration, drag direction, drag step length, commands to return to the previous menu, cancel commands, or skip commands. User interaction requests corresponding to the physical button channel can include steering wheel button status, central control knob signals, and custom shortcut key trigger signals. In multimodal input scenarios, the intelligent cockpit operating system can use voice text, touch actions, vehicle status, and application interface information as input sources for user interaction requests.

[0116] Interaction context data includes at least one of user request information, system response information, executed action information, and vehicle environment information. User request information records the original request content made by the user to the intelligent cockpit operating system. User request information may include the user's original voice text, the target touch control, input content, the target application, the target function, scene identifiers, and user identifiers. For example, when a user says "Navigate home and buy coffee on the way," the intelligent cockpit operating system records the voice-recognized text, the navigation scene identifier, and the user identifier. When the user drags navigation waypoints on the central control screen, the intelligent cockpit operating system records the drag coordinates, the route position before and after the drag, and the navigation application identifier.

[0117] System response information records the response content generated by the intelligent cockpit operating system in response to user interaction requests. System response information may include currently active prompts, model input information, model output information, voice broadcast text, TTS (text-to-speech) audio control parameters, in-vehicle interface rendering commands, navigation route results, POI (point of interest) lists, recommended content, vehicle control objectives, and task processing results. For interaction scenarios involving large model inference, system response information may also include prompts injected during model inference, invoked tool channels, inference output text, and inference confidence. For interaction scenarios involving navigation, media, or vehicle control, system response information may also include map interface call results, media resource retrieval results, or vehicle control command generation results.

[0118] Action execution information records the actions performed by the intelligent cockpit operating system based on system response information. This information may include the target functional module, target intelligent agent identifier, execution method, calling interface, tool call result, vehicle control message, navigation route distribution status, media playback status, task execution status, response time, failure reason, and system error code. For example, in a navigation scenario, action execution information may include the route planning interface, POI retrieval interface, route geometry data, route distribution result, and path deviation error code. In a vehicle control scenario, action execution information may include control messages and execution status corresponding to actuators such as air conditioning, windows, seats, lights, or the trunk.

[0119] Vehicle environmental information is used to record the vehicle environment and hardware resource status at the time of user interaction requests. This information may include vehicle gear status, vehicle speed, vehicle location, current charging status, low-voltage battery voltage, power battery charge, screen status, occupant status, driver's seat status, door status, window status, lock status, vehicle ignition status, cabin ambient temperature, processor temperature, computing power utilization, available storage space, and network connection status. Vehicle environmental information can be obtained through CAN (Controller Area Network), LIN (Low-Cost Serial Communication Network), in-vehicle Ethernet, BMS (Battery Management System), body controller, cabin domain controller, temperature sensors, low-voltage power sensors, and system resource monitoring interfaces.

[0120] In one specific implementation, the intelligent cockpit operating system encapsulates application-layer business requests using a runtime software development kit (SDK) and forwards the encapsulated requests to the inference call chain. The proxy service in the inference call chain generates a trajectory identifier and starts a background listening thread before the business request enters the model inference module or business tool module. The background listening thread collects user request information, system response information, action execution information, and vehicle environment information without blocking the foreground response process. In this way, the intelligent cockpit operating system can collect complete interaction context data without altering the original processing logic of business modules such as navigation, voice interaction, vehicle control, and entertainment media.

[0121] In another specific implementation, the intelligent cockpit operating system can set up a data acquisition interface within each functional intelligent agent, allowing the navigation intelligent agent, voice interaction intelligent agent, vehicle control intelligent agent, and entertainment media intelligent agent to report interaction context data respectively. The data acquisition interface uniformly converts the data reported by different functional intelligent agents into a preset format. Based on this, the intelligent cockpit operating system can adapt to the interaction data acquisition needs of different business scenarios.

[0122] Step S202: Generate explicit feedback data when user input of error correction information, evaluation information, negative semantic information, or text feedback information is detected; and / or generate implicit feedback data when user interruption of broadcast, skipping of output content, manual modification of system execution results, closure of executed functions, or failure to perform subsequent operations within a preset time period is detected.

[0123] Here, explicit feedback data refers to data generated when users actively express their feedback intentions. The intelligent cockpit operating system generates explicit feedback data when it detects that a user is expressing feedback through the interface, voice, text input, or physical buttons. Error correction information may include the user clicking an error correction control, selecting the correct result, modifying the system's recognized content, or submitting a corrected execution result. Evaluation information may include likes, dislikes, star ratings, satisfaction ratings, or regenerating trigger events. Negative semantic information may include voice or text with negative, corrective, or dissatisfying meanings such as "incorrect," "mistakenly said," "stop talking nonsense," "answer again," or "this is not the way." Text feedback information may include complaints, notes, preference descriptions, or problem descriptions submitted by the user through input boxes.

[0124] When generating explicit feedback data, the intelligent cockpit operating system can record the feedback source, time of occurrence, content, recipient, intensity, and user identifier. The feedback source can be voice, touch, text input, or physical buttons. The feedback recipient can be the current response, the current navigation route, the current vehicle control result, the current recommended content, or the current application interface. The feedback intensity can be determined based on the evaluation level, the intensity of negative words, the number of corrections, or the number of times the user provides repeated feedback. For example, if the user continuously inputs "I misspoke" and "answer again," the intelligent cockpit operating system can mark the explicit feedback data as strong negative feedback.

[0125] Implicit feedback data is data generated when users do not directly input evaluation content, but their behavior reflects whether the system's response meets their expectations. The intelligent cockpit operating system generates implicit feedback data by monitoring subsequent user behavior, application state changes, and task execution results. User interruption of broadcast refers to the user stopping the broadcast by voice, button, or touch before it is completed after the intelligent cockpit operating system has started broadcasting. If the broadcast duration is less than a preset broadcast threshold, such as less than 3 seconds, the intelligent cockpit operating system can identify the interruption as implicit feedback data and associate the interruption with the current broadcast content. Skipping output content refers to the user ending the current output content by clicking skip, closing the card, returning to the previous level, or switching tasks after the system has output text, voice, cards, or recommended content.

[0126] Manually modifying system execution results refers to users manually correcting the results already generated or executed by the intelligent cockpit operating system. In navigation scenarios, manually modifying system execution results can include users dragging waypoints, adding waypoints, selecting avoidable sections, reselecting routes, or canceling system-recommended routes. In vehicle control scenarios, manually modifying system execution results can include users immediately adjusting or turning off corresponding functions after the system turns on the air conditioning, windows, seat heating, or ambient lighting. In entertainment and media scenarios, manually modifying system execution results can include users skipping recommended songs, switching media sources, or canceling playlists.

[0127] Closing an executed function refers to the user closing or canceling a function operation within a short period of time after the intelligent cockpit operating system completes the operation based on user interaction requests or system recommendations. The intelligent cockpit operating system can identify the act of closing an executed function as the current execution result not conforming to the user's true intention. No subsequent operation executed within a preset time frame means that after the intelligent cockpit operating system outputs a response result, no user confirmation, continued operation, or task completion behavior is detected within a preset time window. The intelligent cockpit operating system can determine the feedback meaning corresponding to the failure to execute subsequent operations within a preset time frame based on the scenario type. For example, in a multi-turn interaction scenario requiring further user confirmation, failure to execute subsequent operations within a preset time frame may indicate that the system's understanding is inaccurate or the output guidance is unclear; in an information broadcast scenario where no further user operation is required, failure to execute subsequent operations within a preset time frame can be considered neutral or successful feedback.

[0128] In one specific implementation, the intelligent cockpit operating system sets feedback recognition rules for different feedback behaviors. These rules include triggering conditions, feedback types, feedback scores, and anomaly labels. When a user clicks an error correction control or inputs negative semantic information, the feedback type is marked as explicit feedback, the feedback score is set to a negative score, and the anomaly label is set to semantic understanding anomaly, incorrect answer, or task execution anomaly. When a user interrupts the broadcast shortly after it begins, the feedback type is marked as implicit feedback, the feedback score is set to a negative score, and the anomaly label is set to output redundancy anomaly. When a user modifies a waypoint shortly after navigation route guidance begins, the feedback type is marked as implicit feedback, the feedback score is set to a negative score, and the anomaly label is set to path planning anomaly.

[0129] In another specific implementation, the intelligent cockpit operating system can also convert positive behaviors into interactive feedback data. Positive behaviors may include the user listening to a broadcast in its entirety, the user confirming the recommendation result, the user continuing to drive according to the system's planned route, the user maintaining the vehicle control state executed by the system, or the user giving a thumbs-up to the system's response. The intelligent cockpit operating system can set the feedback score corresponding to the positive behavior as a positive score and use the corresponding interactive data as candidate data for high-quality samples.

[0130] Step S203: Associate the interaction context data and the interaction feedback data to determine the interaction data.

[0131] Here, after obtaining the interaction context data and interaction feedback data, the intelligent cockpit operating system correlates the interaction context data and interaction feedback data. The purpose of this correlation is to clarify the user interaction request, system response content, executed action, and vehicle environmental status corresponding to each piece of interaction feedback data.

[0132] In one specific implementation, the intelligent cockpit operating system associates data based on trajectory identifiers. Upon receiving a user interaction request, the intelligent cockpit operating system generates a unique trajectory identifier and writes this identifier into the user request information, system response information, executed action information, and vehicle environment information. When the intelligent cockpit operating system detects interaction feedback data within a preset time window, it binds the interaction feedback data to the same trajectory identifier. The preset time window can be set according to the business scenario. The preset time window can be shorter for voice question-and-answer scenarios and longer for navigation planning scenarios.

[0133] In another specific implementation, the intelligent cockpit operating system associates data based on timestamps and scene identifiers. The intelligent cockpit operating system records the timestamps of user interaction requests, system responses, executed actions, and interactive feedback, and records the corresponding scene identifiers. The intelligent cockpit operating system associates data that are closest in time, have the same scene identifier, the same user identifier, or the same business function module. Associating data using timestamps and scene identifiers is applicable to scenarios where trajectory identifiers are not pre-assigned.

[0134] In another specific implementation, the intelligent cockpit operating system associates data based on execution objects. Execution objects can be navigation routes, vehicle control functions, media content, voice prompts, or recommendation cards. The intelligent cockpit operating system associates the user's feedback behavior to the execution object with the corresponding interaction context data. For example, when a user turns off the seat heating function that was just turned on, the intelligent cockpit operating system associates the turning-off behavior with the interaction context data of turning on the seat heating function. When a user manually modifies navigation waypoints, the intelligent cockpit operating system associates the modification behavior with the original navigation route planning result.

[0135] The intelligent cockpit operating system can perform data cleaning, data anonymization, format conversion, and integrity verification on the interaction data after correlation processing. Data cleaning includes removing invalid, duplicate, and obviously erroneous fields. Data anonymization includes anonymizing or localizing user identifiers, vehicle identifiers, location data, voice and text, and personal preference information. Format conversion involves converting data from different sources into a unified data structure. Integrity verification includes checking the integrity of the binding relationships between user request information, system response information, action information, vehicle environment information, and interaction feedback data. Through the above processing, the interaction data can serve as the data foundation for subsequently generating interaction trajectory data.

[0136] In an optional implementation, the intelligent cockpit operating system writes interaction data to local storage. This interaction data can be stored using a relational database, file system, log queue, or message cache. The intelligent cockpit operating system can configure data retention periods, scenario blacklists / whitelists, and user privacy authorization switches. For sensitive data not authorized by the user, the intelligent cockpit operating system does not persist it, or only stores anonymized feature data. Based on this, the intelligent cockpit operating system can retain data for self-learning processing while ensuring user privacy.

[0137] In an optional implementation, the intelligent cockpit operating system records the currently effective version of the prompt information, the version of the model adaptation information, and the version of the dynamic routing configuration in the interaction data. By recording the version information, the intelligent cockpit operating system can determine the system configuration status corresponding to the abnormal results when analyzing interaction anomalies, thereby improving the targeting of prompt information adjustments and model adaptation information training.

[0138] In an optional implementation, refer to Figure 3 Step S102 includes the following steps S301-S303.

[0139] Step S301: Determine the feedback type and feedback evaluation information corresponding to the user interaction request based on the interaction feedback data; wherein, the feedback evaluation information includes feedback score and / or anomaly label.

[0140] Here, feedback type is used to characterize the source or form of feedback data. Feedback types can include at least one of the following: explicit feedback, implicit feedback, positive feedback, negative feedback, neutral feedback, manually confirmed feedback, and automatically determined feedback by the system. The intelligent cockpit operating system determines the feedback type based on the triggering method of the interactive feedback data. For example, when a user clicks an evaluation control or enters correction text, the feedback type is explicit. When a user interrupts the broadcast or modifies the route, the feedback type is implicit. When a user confirms the system's recommendation result, the feedback type is positive.

[0141] Feedback evaluation information is used to characterize the evaluation of the interaction results by the interaction feedback data. Feedback evaluation information includes feedback scores and / or anomaly labels. Feedback scores can be discrete, continuous, graded, or binary. Discrete scores can include positive, zero, and negative scores. Graded scores can include levels such as satisfactory, neutral, and unsatisfactory. Anomaly labels are used to identify the categories of problems present in the interaction results. Anomaly labels can include at least one of the following: verbose output anomaly, incorrect answer, semantic understanding anomaly, path planning anomaly, task execution anomaly, vehicle control anomaly, recommendation mismatch, response timeout, tool call failure, and user preference mismatch.

[0142] In one specific implementation, the intelligent cockpit operating system determines feedback scores and anomaly labels based on explicit feedback data. When a user inputs error correction information or negative semantic information, the intelligent cockpit operating system sets the feedback score to a negative score and determines an anomaly label based on the content of the negative semantic information. For example, when a user says "Stop talking nonsense," the intelligent cockpit operating system sets the anomaly label to "output redundancy anomaly." When a user says "This isn't the way," the intelligent cockpit operating system sets the anomaly label to "path planning anomaly." When a user says "I don't understand what I'm saying," the intelligent cockpit operating system sets the anomaly label to "semantic understanding anomaly." When a user clicks "like" or inputs positive feedback, the intelligent cockpit operating system sets the feedback score to a positive score.

[0143] In another specific implementation, the intelligent cockpit operating system determines feedback scores and anomaly labels based on implicit feedback data. If a user interrupts a broadcast before it is completed, the intelligent cockpit operating system sets the feedback score to a negative value and labels it as either "output length error" or "broadcast content error." If a user modifies waypoints after route generation, the intelligent cockpit operating system sets the feedback score to a negative value and labels it as "path planning error." If a user immediately disables a function after the system performs vehicle control, the intelligent cockpit operating system sets the feedback score to a negative value and labels it as either "task execution error" or "vehicle control error." If a user immediately skips recommended content after it is displayed, the intelligent cockpit operating system sets the feedback score to a negative value and labels it as "recommendation mismatch."

[0144] In another specific implementation, the intelligent cockpit operating system modifies feedback evaluation information based on the frequency and duration of feedback. For example, if the same user generates the same anomalous label for the same scene multiple times within a preset time window, the intelligent cockpit operating system increases the confidence level of the corresponding anomalous label. If a user repeatedly interrupts the same type of broadcast, the intelligent cockpit operating system prioritizes outputting lengthy anomalous messages. If a user maintains the system's recommended results without generating negative feedback for an extended period, the intelligent cockpit operating system marks the corresponding interaction trajectory data as high-quality data. Through frequency and duration correction, feedback evaluation information can more accurately reflect the user's true preferences.

[0145] Step S302: Associate the user request information, system response information, execution action information, vehicle environment information, feedback type, and feedback evaluation information.

[0146] Here, after determining the feedback type and feedback evaluation information, the intelligent cockpit operating system correlates user request information, system response information, executed action information, vehicle environment information, feedback type, and feedback evaluation information. This correlation process forms a complete data link corresponding to the same user interaction request. The complete data link indicates what request the user made, what response the intelligent cockpit operating system gave, what action the intelligent cockpit operating system executed, the vehicle's current state, the user's feedback, and the feedback evaluation result.

[0147] In one specific implementation, the intelligent cockpit operating system uses user request information as the starting point of the data link, system response information as the inference result node, executed action information as the execution result node, vehicle environment information as the environment snapshot node, and feedback type and feedback evaluation information as the feedback result node. The intelligent cockpit operating system connects the above nodes through trajectory identifiers, timestamps, user identifiers, vehicle identifiers, and scene identifiers to generate a complete association result.

[0148] In another specific implementation, the intelligent cockpit operating system incorporates model inference link information during the association process. This model inference link information may include current prompts, runtime-injected prompts, model output text, tool call channels, inference steps, inference time, and model version information.

[0149] In another specific implementation, the intelligent cockpit operating system incorporates a post-execution feedback observation window during the association process. This window represents a period of time after the system response or execution action is completed. Within this window, the intelligent cockpit operating system collects data on whether the user interrupted, skipped, modified, undone, confirmed, or performed no action. The operating system then associates the feedback behavior generated within the post-execution feedback observation window with the corresponding user interaction request. By setting up this window, the intelligent cockpit operating system can avoid incorrectly associating subsequent irrelevant operations with the current user interaction request.

[0150] Step S303: Generate interaction trajectory data corresponding to the user's interaction request based on the association results.

[0151] Here, the intelligent cockpit operating system generates interaction trajectory data corresponding to user interaction requests based on the correlation results. This interaction trajectory data is a structured record of the complete processing of user interaction requests. It is used to subsequently determine the target self-learning strategy, generate target update data, and evaluate the update effect.

[0152] Interaction trajectory data may include at least one of the following: trajectory identifier, timestamp, user identifier, vehicle identifier, scene identifier, user request content, system response content, current prompt information, runtime prompt information, model output content, tool call results, executed action information, execution status, vehicle status snapshot, feedback type, feedback score, exception label, number of feedback occurrences, response time, and version information. The trajectory identifier uniquely identifies each interaction process. The timestamp records the time the interaction occurred. The user identifier distinguishes different users. The vehicle identifier distinguishes different vehicles. The scene identifier distinguishes business scenarios such as navigation, voice interaction, vehicle control, entertainment media, or information recommendation. Version information may include the prompt information version, model adaptation information version, and routing configuration version.

[0153] In one specific implementation, the interaction trajectory data is stored in the form of structured objects. These structured objects include basic fields, vehicle status fields, inference link fields, execution action fields, and feedback evaluation fields. The basic fields record the trajectory identifier, timestamp, user identifier, vehicle identifier, and scene identifier. The vehicle status fields record the vehicle's gear position, speed, low-voltage battery voltage, power battery charge, charging status, screen status, occupant status, processor temperature, and computing power utilization. The inference link fields record the user request content, prompts, model output content, tool call channels, and inference process. The execution action fields record the target functional module, execution method, execution status, response time, and error code. The feedback evaluation fields record the feedback type, feedback score, exception label, feedback description, and the number of times feedback occurred.

[0154] In another specific implementation, the intelligent cockpit operating system stores the interaction trajectory data in a local trajectory storage space. This trajectory storage space can employ a combination of relational databases and file systems. Structured fields can be stored in the relational database, while unstructured data such as voice clips, image clips, log files, and route geometry data can be stored in the file system. Structured fields and unstructured data are associated through trajectory identifiers. This storage method allows the intelligent cockpit operating system to quickly query and analyze interaction trajectory data, and also to revisit abnormal interaction processes when needed.

[0155] In another specific implementation, the intelligent cockpit operating system indexes the interaction trajectory data. The index may include a time index, a scene index, a user index, an anomaly label index, a feedback score index, a prompt message version index, and a model adaptation information version index. Through these indexes, the intelligent cockpit operating system can count the number of times the same user, the same scene, or the same anomaly label appears within a preset time window.

[0156] In another specific implementation, the intelligent cockpit operating system samples and labels the interaction trajectory data. Sample labeling can include failure sample labels, success sample labels, high-quality sample labels, low-confidence sample labels, and samples awaiting review labels. Failure sample labels correspond to interaction trajectory data with negative feedback or anomaly labels. Success sample labels correspond to interaction trajectory data that has been confirmed by the user or has not generated negative feedback. High-quality sample labels correspond to interaction trajectory data with complete fields, clear feedback, reliable execution results, and suitability for model adaptation information training.

[0157] In an optional implementation, refer to Figure 4 Step S103 includes the following steps S401-S405.

[0158] Step S401: Based on the feedback type, feedback score, and anomaly label in the interaction trajectory data, determine the interaction anomaly type corresponding to the interaction trajectory data.

[0159] Here, the interaction anomaly type indicates the category of problems encountered by the intelligent cockpit operating system during the response to user interaction requests. The interaction anomaly type can be determined based on a combination of feedback type, feedback score, and anomaly label from the interaction trajectory data. Feedback type indicates the source and method of feedback, such as explicit feedback, implicit feedback, positive feedback, or negative feedback. Feedback score indicates the user's evaluation of the interaction result. Anomaly label indicates the specific problem identifier corresponding to the interaction result.

[0160] When determining the type of interaction anomaly, the intelligent cockpit operating system can first determine the feedback type. If the feedback type is explicit and the feedback score is negative, the intelligent cockpit operating system determines the interaction anomaly type based on the user's error correction information, negative semantic information, or text feedback information. For example, when the user says "incorrect," "wrong answer," or "answer again," the intelligent cockpit operating system can determine the interaction anomaly type as semantic understanding anomaly, incorrect answer, or output content not meeting expectations. When the user says "stop talking nonsense" or clicks to skip lengthy content, the intelligent cockpit operating system can determine the interaction anomaly type as lengthy output anomaly or interaction style anomaly.

[0161] If the feedback type is implicit, the intelligent cockpit operating system determines the type of interaction anomaly based on the relationship between user behavior and system response. For example, if the user interrupts the voice broadcast before it is completed after the intelligent cockpit operating system starts, the system can classify the interaction anomaly as output redundancy, broadcast content anomaly, or interaction style anomaly. If the user manually drags waypoints, adds avoidance points, or reselects a route shortly after the system generates a navigation route, the system can classify the interaction anomaly as path planning anomaly or user preference mismatch anomaly. If the user immediately closes or cancels the same function after the system executes vehicle control, the system can classify the interaction anomaly as control intent comprehension anomaly or task execution anomaly.

[0162] In one implementation, the intelligent cockpit operating system can directly determine the type of interaction anomaly based on anomaly tags. When the anomaly tag indicates lengthy output, long text, frequent interruptions, or skipping of output content, the intelligent cockpit operating system identifies the interaction anomaly as a lengthy output anomaly. When the anomaly tag indicates path deviation, detour, user dragging waypoints, or avoiding routes, the intelligent cockpit operating system identifies the interaction anomaly as a path planning anomaly. When the anomaly tag indicates comprehension failure, incorrect intent recognition, or incorrect response, the intelligent cockpit operating system identifies the interaction anomaly as a semantic understanding anomaly.

[0163] In another implementation, the intelligent cockpit operating system can determine the priority of interaction exception types by combining feedback scores and the frequency of exception occurrences. When the same exception label appears multiple times within a preset time window, or when the feedback score remains below a preset threshold, the intelligent cockpit operating system increases the processing priority of the corresponding interaction exception type. Based on this, the intelligent cockpit operating system can prioritize handling exceptions that have a greater impact on user experience, reducing the interference of occasional operations on the self-learning process.

[0164] Step S402: Based on the preset correspondence between interaction exception types and self-learning strategies, determine the self-learning strategy corresponding to the interaction exception type.

[0165] Here, the intelligent cockpit operating system pre-sets a correspondence between interaction anomaly types and self-learning strategies. This correspondence indicates the appropriate self-learning handling method for different interaction anomaly types. The self-learning strategies include a first self-learning strategy for adjusting prompts and a second self-learning strategy for adjusting model adaptation information. The first self-learning strategy primarily handles problems that can be improved by adjusting prompts, output constraints, interaction styles, or task boundaries. The second self-learning strategy primarily handles problems that require learning long-term user preferences, complex task patterns, or business decision-making logic through training model adaptation information.

[0166] When the interaction anomaly type is characterized by excessively long output, abnormal interaction style, abnormal response format, abnormal task boundary, or abnormal output constraint, the intelligent cockpit operating system can determine the first self-learning strategy as the self-learning strategy corresponding to the interaction anomaly type. These types of problems are usually related to the role setting, output length, tone requirements, response format, task priority, or business constraints in the prompt information; adjusting the prompt information can quickly improve the interaction results.

[0167] When the interaction anomaly type is path planning anomaly, user preference mismatch anomaly, complex task execution anomaly, multi-turn dialogue logic anomaly, or cross-functional linkage decision anomaly, the intelligent cockpit operating system can determine the second self-learning strategy as the self-learning strategy corresponding to the interaction anomaly type. These types of problems are usually related to the model's adaptability to long-term user behavior, vehicle usage scenarios, time and location preferences, or task execution patterns. Simply adjusting the prompt information may not reliably solve the problem; therefore, the intelligent cockpit operating system needs to train the model adaptation information based on interaction trajectory data.

[0168] When an interaction anomaly involves both expression and deep preferences, the intelligent cockpit operating system can use both a first self-learning strategy and a second self-learning strategy to determine the corresponding self-learning strategy for the interaction anomaly type. For example, if a user's response to "navigate home and buy coffee on the way" repeatedly interrupts a lengthy explanation and requires manual route modification, the intelligent cockpit operating system can reduce unnecessary explanations using the first self-learning strategy and learn the user's preferences for coffee shop brands, route detours, and commuting time windows using the second self-learning strategy.

[0169] Step S403: If the self-learning strategy corresponding to the interaction exception type is the first self-learning strategy, then the first self-learning strategy is determined as the target self-learning strategy.

[0170] Here, the first self-learning strategy is used to adjust the prompts. Prompts are the instructions used by the intelligent cockpit operating system when invoking models or business function modules. Prompts may include at least one of the following: role settings, output format, output length, tone style, task boundaries, tool call constraints, security constraints, business priorities, and contextual supplementary information.

[0171] When interaction anomalies can be resolved through adjustments to the prompts, the intelligent cockpit operating system will adopt the first self-learning strategy as the target self-learning strategy. For example, if a user repeatedly interrupts a broadcast or skips long text content, it indicates that the current prompts do not adequately constrain the output length and response style; in this case, the intelligent cockpit operating system will adopt the first self-learning strategy as the target self-learning strategy. As another example, if a user repeatedly requests "No explanation, just execute," it indicates that the task priority in the current prompts needs adjustment; in this case, the intelligent cockpit operating system will adopt the first self-learning strategy as the target self-learning strategy.

[0172] The first self-learning strategy corresponds to APO (Auto-Prompt Optimization) processing. APO processing is used to automatically analyze, correct, and verify the current prompt information based on failed or low-scoring interaction trajectories. APO processing has a relatively low computational cost and is suitable for quickly updating output style, output length, and task constraints.

[0173] Step S404: If the self-learning strategy corresponding to the interaction exception type is the second self-learning strategy, then the second self-learning strategy is determined as the target self-learning strategy.

[0174] Here, the second self-learning strategy is used to adjust the model adaptation information. Model adaptation information is data used to adapt the model to user preferences, vehicle scenarios, and business tasks. Model adaptation information may include at least one of the following: incremental model parameters, adapter parameters, personalized preference parameters, business rule parameters, scenario routing configuration, or feature mapping relationships.

[0175] When interaction anomalies require improvements in model adaptability to resolve, the intelligent cockpit operating system identifies the second self-learning strategy as the target self-learning strategy. For example, if a user manually corrects the navigation route multiple times during the same commute, it indicates that the intelligent cockpit operating system needs to learn the user's long-term preferences for roads, distances, times, brands, or geofences; in this case, the intelligent cockpit operating system identifies the second self-learning strategy as the target self-learning strategy. As another example, if a user repeatedly corrects the system's execution order in multiple task rounds, it indicates that the model's adaptability to complex task logic is insufficient; in this case, the intelligent cockpit operating system identifies the second self-learning strategy as the target self-learning strategy.

[0176] The second self-learning strategy can correspond to LoRA (low-rank adaptation) training, supervised fine-tuning, efficient parameter fine-tuning, preference learning, or personalized rule parameter training. Compared to the first self-learning strategy, the second self-learning strategy requires more training samples and more computing resources, and is suitable for execution when the vehicle is in a safe state, the power supply is stable, and computing resources are idle.

[0177] Step S405: If the self-learning strategy corresponding to the interaction exception type includes a first self-learning strategy and a second self-learning strategy, then the first self-learning strategy and the second self-learning strategy are determined as the target self-learning strategy.

[0178] Here, in some scenarios, a single self-learning strategy is insufficient to fully resolve interaction anomalies. The intelligent cockpit operating system can combine a first self-learning strategy and a second self-learning strategy as the target self-learning strategy. The first self-learning strategy is used to quickly adjust prompts, enabling subsequent responses to improve in terms of expression and task constraints as rapidly as possible. The second self-learning strategy is used to train the model to adapt to information, allowing the model to gradually internalize long-term user preferences and complex task patterns.

[0179] For example, in a navigation scenario, after a user inputs "navigate home and buy coffee on the way," the intelligent cockpit operating system might output a lengthy route explanation, causing the user to interrupt the broadcast; furthermore, the planned route might not fit the user's habits, leading the user to manually drag waypoints. To address this, the intelligent cockpit operating system can first employ a first self-learning strategy, adjusting the prompts to prioritize concise route results and reduce explanatory text; simultaneously, it can employ a second self-learning strategy, using the user's corrected route, coffee brand preferences, and commuting scenario data as training data to update the model's adaptation information.

[0180] In an optional implementation, the intelligent cockpit operating system can set the execution order of the first self-learning strategy and the second self-learning strategy. The intelligent cockpit operating system can execute the first self-learning strategy first, followed by the second self-learning strategy. Alternatively, the intelligent cockpit operating system can execute the first and second self-learning strategies separately when different access conditions are met. For example, the first self-learning strategy can be executed within a short silent time window, while the second self-learning strategy can be executed when the vehicle is parked and charging and no one is inside the vehicle.

[0181] In an optional implementation, step S403 includes the following steps S501-S504.

[0182] Step S501: Obtain target interaction trajectory data with the same abnormal label or belonging to the same interaction abnormal type from the interaction trajectory data.

[0183] Here, the target interaction trajectory data refers to the interaction trajectory data used to adjust the prompts. The intelligent cockpit operating system filters data from the interaction trajectory data that has the same anomaly label or belongs to the same interaction anomaly type, using this as the target interaction trajectory data. By aggregating data of the same problem type, the intelligent cockpit operating system can identify consistently existing prompt problems, reducing the impact of occasional, isolated feedback on prompt updates.

[0184] For example, if multiple interaction trajectory data points all have lengthy output anomaly labels, or if multiple interaction trajectory data points all belong to interaction style anomalies, the intelligent cockpit operating system will use these interaction trajectory data points as target interaction trajectory data. As another example, if multiple interaction trajectory data points all contain feedback behaviors such as "Stop talking nonsense," "Skip," or "Interrupt broadcast," the intelligent cockpit operating system can classify these interaction trajectory data points into the same interaction anomaly type and use them as target interaction trajectory data.

[0185] In one implementation, the intelligent cockpit operating system can filter target interaction trajectory data based on a time window. For example, the intelligent cockpit operating system can filter similar abnormal data that has appeared within the past 24 hours, 72 hours, or one week. If the number of similar abnormal data reaches a preset threshold, the intelligent cockpit operating system activates a first self-learning strategy. By setting a time window and a quantity threshold, the intelligent cockpit operating system can avoid the adverse effects of expired historical data on the adjustment of current prompts.

[0186] In another implementation, the intelligent cockpit operating system can filter target interaction trajectory data based on user identifiers, scene identifiers, and anomaly tags. Filtering by user identifiers enables personalized prompts for individual users. Filtering by scene identifiers enables differentiated prompts for different scenarios such as navigation, voice Q&A, vehicle control, or entertainment media.

[0187] Step S502: Determine the content to be adjusted in the prompt information based on the target interaction trajectory data.

[0188] Here, the intelligent cockpit operating system analyzes the user request content, current prompt information, system response content, execution status, feedback type, feedback score, and anomaly label in the target interaction trajectory data to determine the content to be adjusted in the prompt information. The content to be adjusted represents the constraints that need to be added, deleted, or modified in the current prompt information.

[0189] The content to be adjusted may include at least one of the following: output length limit, answer format limit, tone style limit, task priority adjustment, prohibition of explanation conditions, tool call priority, business boundary constraints, scenario supplementary information, security prompt constraints, and multi-turn dialogue follow-up questioning rules.

[0190] For example, if the target interaction trajectory data indicates that the user frequently interrupts the broadcast, the intelligent cockpit operating system determines that the adjustments to be made are to reduce the length of responses, prioritize outputting conclusions, and avoid repetitive explanations. If the target interaction trajectory data indicates that the user repeatedly requests a response, the intelligent cockpit operating system determines that the adjustments to be made are to enhance intent confirmation, add necessary follow-up questions, or adjust the response format. If the target interaction trajectory data indicates that the user prefers direct control actions rather than lengthy explanations in vehicle control scenarios, the intelligent cockpit operating system determines that the adjustments to be made are to prioritize executing control actions and reduce explanatory text.

[0191] In one implementation, the intelligent cockpit operating system can analyze target interaction trajectory data through an evaluation model. The evaluation model reads the user's request content, current prompts, and system responses, and outputs the cause of the interaction anomaly and suggestions for modifying the prompts. For example, the evaluation model can determine that the anomaly is caused by excessively long output, incorrect task priority, missing scenario constraints, or unclear tool call boundaries. Based on the evaluation model's output, the intelligent cockpit operating system determines the content of the prompts that needs adjustment.

[0192] In another implementation, the intelligent cockpit operating system can determine the content to be adjusted using rule templates. Different interaction anomaly types correspond to different rule templates. Lengthy output anomalies correspond to shortened response templates. Path planning anomalies correspond to priority navigation task templates. Task boundary anomalies correspond to templates that add prohibition or supplementary conditions. Determining the content to be adjusted using rule templates can improve the stability of prompt message adjustments.

[0193] Step S503: Adjust the current prompt information based on the content to be adjusted to obtain the updated prompt information.

[0194] Here, the intelligent cockpit operating system adjusts the current prompt message based on the content to be adjusted, resulting in an updated prompt message. Adjustment methods may include adding prompt constraints, deleting conflict constraints, modifying output format, adjusting task priority, adding scenario conditions, adding prohibition conditions, adjusting tone style, or adjusting tool call descriptions.

[0195] For example, if the current prompt requires "detailed explanation of the route planning reasons," and the target interaction trajectory data shows that the user frequently interrupts lengthy announcements, the intelligent cockpit operating system can adjust the current prompt to "prioritize outputting route results, with the response not exceeding a preset word count, and only explaining the reasons when the user asks follow-up questions." As another example, if the current prompt requires "explain before execution," and the target interaction trajectory data shows that the user has repeatedly indicated that no explanation is needed, the intelligent cockpit operating system can adjust the current prompt to "prioritize task execution in vehicle control and navigation scenarios, and then output a brief confirmation."

[0196] In one implementation, the intelligent cockpit operating system validates the updated prompts. Validation may include checking if the number of characters or tokens exceeds limits, whether the prompt contains conflicting instructions, whether it violates security constraints, whether the prompt can be correctly parsed by the model, and whether it contains sensitive words or non-compliant content. A token represents a word segmentation unit used by the model when processing text. If the validation passes, the intelligent cockpit operating system uses the updated prompt as available target update data. If the validation fails, the intelligent cockpit operating system readjusts the prompt or abandons the update.

[0197] In another implementation, the intelligent cockpit operating system generates a version number for the updated prompts. The version number distinguishes different prompt versions. The intelligent cockpit operating system records the reason for the update, the corresponding anomaly tag, the source interaction trajectory data, and the generation time. Using the version number and update history, the intelligent cockpit operating system can backtrack or roll back in case of subsequent operational anomalies.

[0198] Step S504: The updated prompt information is identified as the target update data.

[0199] Here, after the updated notification message passes verification, the intelligent cockpit operating system identifies the updated notification message as the target update data. The target update data may include at least one of the following: the updated notification message text, the notification message version number, the effective scenario, the effective users, the effective time, the corresponding anomaly tag, the expected improvement target, and the rollback indicator.

[0200] In an optional implementation, the intelligent cockpit operating system can restrict the updated prompts to specific users, vehicles, or scenarios. For example, to address a user's preference for concise answers, the updated prompts can only be effective when that user is using the vehicle. Similarly, to address task priority issues in navigation scenarios, the updated prompts can only be effective within navigation scenarios. By setting the scope of effectiveness, the intelligent cockpit operating system can prevent localized optimizations from impacting other users or other business scenarios.

[0201] In an optional implementation, the intelligent cockpit operating system can set the updated prompt information to a pending loading state. When the loading conditions are met, the intelligent cockpit operating system loads the updated model adaptation information, enabling it to respond to user interaction requests based on the updated model adaptation information.

[0202] In an optional implementation, step S404 includes the following steps S601-S604.

[0203] Step S601: Filter target interaction trajectory data that meets preset quality conditions from the interaction trajectory data.

[0204] Here, the second self-learning strategy is used to train the model to adapt information; therefore, the intelligent cockpit operating system needs to select suitable training data from the interaction trajectory data. Target interaction trajectory data that meets preset quality conditions is data that can be used as a source of training samples.

[0205] Preset quality conditions may include complete data fields, clear user request content, parsable system response content, valid execution action information, clear feedback evaluation information, available vehicle environment information, high confidence of anomaly labels, feedback scores meeting preset score conditions, user correction results being usable as target output, successful execution results being usable as positive samples, data not expired, data being protected by privacy authorization, and data being de-identified.

[0206] For example, in navigation scenarios, the route manually corrected by the user can be used as target output label data. When filtering target interaction trajectory data, the intelligent cockpit operating system requires that the original route, corrected route, user request content, vehicle location, and feedback evaluation information be complete. As another example, in voice interaction scenarios, user-confirmed satisfactory responses can be used as high-quality positive samples. When filtering target interaction trajectory data, the intelligent cockpit operating system requires that the user request content, model output content, and positive feedback data be complete.

[0207] In one implementation, the intelligent cockpit operating system can filter successful interaction trajectory data as a source of training samples. Successful interaction trajectory data consists of data that indicates user confirmation, high feedback scores, or no negative feedback. Through this successful interaction trajectory data, the model adaptation information can learn response methods that are acceptable to the user.

[0208] In another implementation, the intelligent cockpit operating system can filter user-corrected failed interaction trajectory data as a source of training samples. This user-corrected failed interaction trajectory data includes both the system's original response and the user's correction results. The user's correction results can serve as the target output label data. Based on this, the model adaptation information can learn the direction of improvement from incorrect responses to correct responses.

[0209] Step S602: Construct supervised training samples based on target interaction trajectory data; wherein, supervised training samples include input feature data and target output label data.

[0210] Here, the intelligent cockpit operating system constructs supervised training samples based on target interaction trajectory data. These supervised training samples are used to train the model's adaptation information. The supervised training samples include input feature data and target output label data. The input feature data represents the input content received by the model during training. The target output label data represents the output result that the model should learn.

[0211] Input feature data may include at least one of the following: user request content, scene identifier, user preference features, vehicle environment information, historical interaction summary, current prompt information, tool invocation context, and task objective information. Target output label data may include at least one of the following: user-confirmed system response content, user-corrected navigation route, user-corrected control actions, satisfactory answer content, correct task execution order, target tool invocation result, or corrected recommendation result.

[0212] For example, in a navigation scenario, input feature data could include "navigate home and buy coffee on the way," current vehicle location, destination, historical commuting routes, candidate coffee shop information, and the current time period. The target output label data could include the user-corrected route, the coffee shop actually selected by the user, and the final confirmed navigation result. As another example, in a vehicle control scenario, input feature data could include the user's voice request, current in-vehicle temperature, seat status, and air conditioning status. The target output label data could include the control parameters ultimately retained by the user.

[0213] In one implementation, the intelligent cockpit operating system uses a tokenizer to convert input feature data and target output label data into vector data that the model can process. The tokenizer segments the text content into tokens and converts the tokens into numerical representations. For non-text data such as navigation routes, vehicle status, and control parameters, the intelligent cockpit operating system can first perform structured encoding and then concatenate or map it with text vectors.

[0214] In another implementation, the intelligent cockpit operating system performs data anonymization before building supervised training samples. Data anonymization may include removing users' real names, hiding precise addresses, obfuscating location coordinates, replacing vehicle identifiers, and anonymizing user identifiers. Through data anonymization, the intelligent cockpit operating system can reduce the risk of sensitive information leakage during local training.

[0215] Step S603: Train the model adaptation information in the intelligent cockpit operating system based on supervised training samples to obtain updated model adaptation information.

[0216] Here, the intelligent cockpit operating system trains the model adaptation information based on supervised training samples. This model adaptation information can include incremental model parameters, LoRA adapter parameters, personalized preference parameters, scene rule parameters, task routing parameters, or model invocation configuration parameters. This information is used to enable the intelligent cockpit operating system to better adapt to user preferences and vehicle business scenarios without completely retraining the base model.

[0217] In one implementation, the intelligent cockpit operating system uses LoRA to train model adaptation information. LoRA training involves inserting low-rank matrices as bypasses in parts of the network structure of the base model and training only the incremental parameters corresponding to these low-rank matrices. In this way, the intelligent cockpit operating system can complete model adaptation information training under limited on-vehicle computing power, reducing training resource consumption and storage space usage.

[0218] In another implementation, the intelligent cockpit operating system employs a parameter-efficient fine-tuning method to train model adaptation information. This method can include adapter training, cue parameter training, preference parameter training, or small-scale incremental training. Through this efficient fine-tuning, the intelligent cockpit operating system can learn user preferences and scenario rules without updating all parameters of the base model.

[0219] In another implementation, the intelligent cockpit operating system can generate model adaptation information using rule-based parameter training. For example, for navigation preferences, the intelligent cockpit operating system can update road preference weights, location preference weights, time period preference weights, and detour distance weights based on target interaction trajectory data. For entertainment media preferences, the intelligent cockpit operating system can update content type preference weights, playback time period preference weights, and skip behavior penalty weights.

[0220] During training, the intelligent cockpit operating system can break down the training task into multiple computational segments. Each computational segment processes a portion of the supervised training samples or executes a portion of the training steps. By splitting the computational segments, the intelligent cockpit operating system can suspend the training process in a timely manner when the vehicle state changes and resume the training process when the vehicle state once again meets the admission criteria.

[0221] After training, the intelligent cockpit operating system receives updated model adaptation information. This updated information can be represented as an adapter weight file, a personalized preference parameter table, a business rule parameter table, or a routing configuration file. The intelligent cockpit operating system can quantize and compress this updated model adaptation information to reduce storage space and inference overhead. Quantization and compression convert the model adaptation information into a low-precision data format.

[0222] In one implementation, the intelligent cockpit operating system verifies the updated model adaptation information. Verification may include file integrity verification, size verification, inference latency verification, basic capability impact verification, target scene effect verification, and security policy verification. If the updated model adaptation information passes verification, the intelligent cockpit operating system uses it as loadable data. If the updated model adaptation information fails verification, the intelligent cockpit operating system stops loading the updated model adaptation information or reverts to the model adaptation information before training.

[0223] Step S604: Determine the updated model adaptation information as the target update data.

[0224] Here, after training and verification, the intelligent cockpit operating system determines the updated model adaptation information as the target update data. The target update data may include at least one of the following: updated model adaptation information file, version number, effective users, effective scenarios, effective time, training sample summary, performance evaluation report, and rollback flag.

[0225] In an optional implementation, the intelligent cockpit operating system can restrict the updated model adaptation information to specific users or scenarios. For example, model adaptation information trained based on a user's commuting route can only be effective in that user's navigation scenario. Model adaptation information trained based on a user's voice interaction preferences can only be effective after that user logs into their vehicle account. By setting the effective scope, the intelligent cockpit operating system can prevent personalized training results from affecting other users.

[0226] In an optional implementation, the intelligent cockpit operating system can set the updated model adaptation information to a pending loading state. When the loading conditions are met, the intelligent cockpit operating system loads the updated model adaptation information, enabling it to respond to user interaction requests based on the updated model adaptation information.

[0227] In an optional implementation, step S104 involves generating target update data based on the target self-learning strategy and interaction trajectory data when the vehicle state data meets the preset self-learning admission conditions corresponding to the target self-learning strategy, including the following steps S701-S703.

[0228] Step S701: Determine the corresponding preset self-learning admission conditions based on the target self-learning strategy.

[0229] Here, preset self-learning admission conditions are used to determine whether the intelligent cockpit operating system is currently suitable for initiating self-learning processing. Since different target self-learning strategies correspond to different computational loads, durations, resource consumption, and security risks, the intelligent cockpit operating system needs to determine the corresponding preset self-learning admission conditions based on the target self-learning strategy. The target self-learning strategies include a first self-learning strategy and a second self-learning strategy. The first self-learning strategy is used to adjust prompt information, mainly handling issues such as output length, response style, task boundaries, output format, and task priority; the computational load corresponding to the first self-learning strategy is relatively small. The second self-learning strategy is used to adjust model adaptation information, mainly handling issues such as long-term preferences, complex task rules, scenario decision logic, and user habit internalization; the computational load corresponding to the second self-learning strategy is larger, and the second self-learning strategy has higher requirements for power supply status, temperature status, computing resources, and storage space.

[0230] When the target self-learning strategy includes a first self-learning strategy, the intelligent cockpit operating system determines the preset self-learning access conditions corresponding to the first self-learning strategy. The preset self-learning access conditions corresponding to the first self-learning strategy may include: the vehicle is in a non-driving state, the vehicle speed is zero, the screen is in an off state or a low-load state, the low-voltage battery voltage is greater than a first voltage threshold, the power battery charge is greater than a first charge threshold, the processor temperature is within a first temperature range, the computing power utilization rate is less than a first utilization rate threshold, and the vehicle is not currently in a special scenario that prohibits self-learning. Special scenarios that prohibit self-learning may include car wash mode, maintenance mode, extreme low-temperature environments, the initial stage of high-voltage charging, vehicle malfunction, or a state where the user has disabled the self-learning function.

[0231] When the target self-learning strategy includes a second self-learning strategy, the intelligent cockpit operating system determines the preset self-learning admission conditions corresponding to the second self-learning strategy. The preset self-learning admission conditions for the second self-learning strategy can be more stringent than those for the first self-learning strategy. These preset self-learning admission conditions may include: the vehicle is parked, the vehicle speed is zero, the vehicle is connected to charging, the screen is off, the occupant status indicates no one is in the vehicle, the low-voltage battery voltage is greater than a second voltage threshold, the power battery charge is greater than a second charge threshold, the processor temperature is within a second temperature range, the computing power utilization rate is less than a second utilization rate threshold, the available storage space is greater than a preset storage threshold, and the system time is within a preset silent time window. The second voltage threshold can be greater than the first voltage threshold, the second charge threshold can be greater than the first charge threshold, and the second utilization rate threshold can be less than the first utilization rate threshold. By setting more stringent conditions, the intelligent cockpit operating system can avoid the impact of model adaptation information training on vehicle power supply, thermal management, and front-end operations.

[0232] When the target self-learning strategy includes both a first self-learning strategy and a second self-learning strategy, the intelligent cockpit operating system can determine the preset self-learning admission conditions corresponding to the first and second self-learning strategies, respectively. The intelligent cockpit operating system can first execute the self-learning strategy that meets the admission conditions, or it can determine the execution order according to task priority, anomaly severity, computational resource usage, or user authorization status. For example, if the first self-learning strategy meets the admission conditions but the second self-learning strategy does not, the intelligent cockpit operating system first performs a prompt message adjustment and adds the model adaptation information training task to the execution queue. Subsequently, when the vehicle is parked and charging, unoccupied, and computational resources are idle, the intelligent cockpit operating system will then execute the training task corresponding to the second self-learning strategy.

[0233] In an optional implementation, the intelligent cockpit operating system can also determine preset self-learning admission conditions based on the data conditions of the interaction trajectory data. For a first self-learning strategy, the intelligent cockpit operating system can require the same anomaly label or the same interaction anomaly type to reach a first quantity threshold within a preset time window. For a second self-learning strategy, the intelligent cockpit operating system can require target interaction trajectory data that meets preset quality conditions to reach a second quantity threshold. By limiting the initiation timing of self-learning processing together with data conditions and vehicle state conditions, the intelligent cockpit operating system can reduce unnecessary updates caused by sporadic feedback.

[0234] Step S702: If the vehicle status data indicates that the vehicle is in a non-driving state, the power supply status meets the preset power supply conditions, the processor temperature meets the preset temperature conditions, and the computing power utilization rate meets the preset computing power conditions, then the vehicle status data meets the preset self-learning admission conditions.

[0235] Here, vehicle status data is used to reflect whether the vehicle is currently suitable for performing self-learning processing. Vehicle status data may include vehicle gear position, vehicle speed, low-voltage battery voltage, power battery charge, charging status, screen status, occupant status, door status, window status, vehicle ignition status, processor temperature, computing power utilization, available storage space, and network connectivity status. The intelligent cockpit operating system can acquire vehicle status data through the vehicle bus, body controller, BMS, cockpit domain controller, system resource monitoring interface, temperature sensors, and power management services.

[0236] A non-driving state is a crucial condition for initiating self-learning processing. This non-driving state can be determined by both the vehicle's gear position and speed. For example, if the vehicle is in park and its speed is zero, the intelligent cockpit operating system determines that the vehicle is in a non-driving state. By identifying the non-driving state, the intelligent cockpit operating system avoids consuming driving-related resources or affecting cockpit responsiveness during driving.

[0237] Meeting preset power supply conditions is fundamental to ensuring stable operation of the self-learning process. The power supply status can be determined based on the low-voltage battery voltage, the power battery charge level, and the charging status. For example, the preset power supply conditions for the first self-learning strategy might include a low-voltage battery voltage consistently exceeding a first voltage threshold and a power battery charge level exceeding a first charge threshold. The preset power supply conditions for the second self-learning strategy might include a low-voltage battery voltage consistently exceeding a second voltage threshold, a power battery charge level exceeding a second charge threshold, and the vehicle being in a charging connection state. By determining the power supply status, the intelligent cockpit operating system can prevent the self-learning process from causing a drop in low-voltage battery charge or affecting the vehicle's next start.

[0238] Maintaining a processor temperature that meets preset conditions is crucial for ensuring hardware thermal safety. Processor temperature can include the cockpit chip temperature, NPU temperature, or processor core junction temperature. The intelligent cockpit operating system determines that the processor temperature meets preset conditions when it falls within the preset range. For example, if the processor temperature is below a first temperature threshold, the intelligent cockpit operating system allows the execution of a first self-learning strategy. If the processor temperature is below a second temperature threshold, the intelligent cockpit operating system allows the execution of a second self-learning strategy. The temperature conditions for the second self-learning strategy can be more stringent to reduce the risk of overheating during prolonged training.

[0239] Meeting preset computing power utilization conditions is essential for ensuring the continuity of cockpit services. Computing power utilization can include CPU (graphics processor) utilization, GPU utilization, NPU (neural network processor) utilization, or model inference queue utilization. The intelligent cockpit operating system determines that the computing power utilization meets the preset conditions when it falls below a preset threshold. The preset computing power conditions corresponding to the first self-learning strategy allow for higher resource utilization. The preset computing power conditions corresponding to the second self-learning strategy require the NPU utilization to be below an even lower threshold. By determining computing power utilization, the intelligent cockpit operating system can prevent self-learning processing from affecting front-end services such as voice interaction, navigation planning, vehicle control, and media playback.

[0240] In optional implementations, the preset self-learning admission conditions may further include the occupant status, screen status, and available storage space. When the occupant status indicates that the vehicle is unoccupied, the intelligent cockpit operating system can execute the training task corresponding to the second self-learning strategy. When the screen is off or in a low refresh rate state, the intelligent cockpit operating system can reduce the resource consumption of the foreground display. When the available storage space is greater than a preset storage threshold, the intelligent cockpit operating system can save training samples, temporary files, breakpoint status information, and target update data.

[0241] In an optional implementation, the intelligent cockpit operating system can continuously monitor vehicle status data, rather than only making a judgment before the self-learning process begins. For example, the intelligent cockpit operating system reads vehicle status data every preset time interval and determines whether the vehicle status data continuously meets preset self-learning admission conditions. If the vehicle status data meets the preset self-learning admission conditions for multiple consecutive monitoring cycles, the intelligent cockpit operating system determines that the vehicle status data meets the preset self-learning admission conditions. Through continuous monitoring, the intelligent cockpit operating system can reduce the probability of false starts caused by instantaneous status fluctuations.

[0242] Step S703: If the vehicle status data meets the preset self-learning admission conditions, initiate the self-learning process corresponding to the target self-learning strategy to generate target update data.

[0243] Here, when the vehicle status data meets the preset self-learning admission conditions, the intelligent cockpit operating system initiates the self-learning process corresponding to the target self-learning strategy. The self-learning process corresponding to the first self-learning strategy includes adjusting the prompt information. The self-learning process corresponding to the second self-learning strategy includes training the model adaptation information. The intelligent cockpit operating system selects corresponding data from the interaction trajectory data according to the target self-learning strategy and generates target update data.

[0244] When the target self-learning strategy includes the first self-learning strategy, the intelligent cockpit operating system filters target interaction trajectory data from the interaction trajectory data that have the same anomaly label or belong to the same interaction anomaly type. The intelligent cockpit operating system analyzes the user request content, system response content, feedback score, anomaly label, and current prompt information in the target interaction trajectory data to determine the content to be adjusted in the prompt information. Based on the content to be adjusted, the intelligent cockpit operating system adjusts the current prompt information to obtain the updated prompt information, and uses the updated prompt information as the target update data. This process is suitable for execution within a short silent window.

[0245] When the target self-learning strategy includes a second self-learning strategy, the intelligent cockpit operating system selects target interaction trajectory data that meets preset quality conditions from the interaction trajectory data and constructs supervised training samples. Based on the supervised training samples, the intelligent cockpit operating system trains the model adaptation information to obtain updated model adaptation information, which is then used as the target update data. During training, the intelligent cockpit operating system can divide the training task into multiple computational segments. Each computational segment corresponds to a training step size or a mini-batch of samples. By splitting computational segments, the intelligent cockpit operating system can quickly suspend the training task when the vehicle state changes.

[0246] In an optional implementation, the intelligent cockpit operating system enters a silent self-learning mode before initiating self-learning processing. Once in silent self-learning mode, the operating system can suspend unnecessary background applications, limit resource consumption of non-critical tasks, allocate some computing resources to self-learning processing, reduce the screen refresh rate or turn off the backlight, disable unnecessary audio output channels, and activate high-power consumption protection logic. Through silent self-learning mode, the intelligent cockpit operating system can convert vehicle idle time into self-learning time while reducing the impact of self-learning processing on cockpit front-end services.

[0247] In an optional implementation, the intelligent cockpit operating system verifies the target update data after completing the self-learning process. Verification may include data integrity verification, format verification, version compatibility verification, security policy verification, resource usage verification, and performance evaluation verification. If the target update data passes verification, the intelligent cockpit operating system marks it as pending loading. If the target update data fails verification, the intelligent cockpit operating system discards the target update data or re-executes the corresponding self-learning process.

[0248] In an optional implementation, in step S104, during the process of generating target update data based on the target self-learning strategy and interaction trajectory data, the method further includes the following steps S801-S803.

[0249] Step S801: Monitor the vehicle's operating status and hardware resource status.

[0250] Here, the intelligent cockpit operating system continuously monitors the vehicle's operating status and hardware resource status during the self-learning process. The vehicle's operating status reflects whether the vehicle is still suitable for continuing self-learning. The hardware resource status reflects whether self-learning will lead to risks related to power supply, temperature, computing power, or storage.

[0251] Vehicle operating status may include at least one of the following: vehicle gear status, vehicle speed, vehicle ignition status, brake pedal status, door status, occupant status, driver's seat status, charging status, vehicle fault status, and user operation status. Hardware resource status may include at least one of the following: low-voltage battery voltage, power battery charge, processor temperature, NPU temperature, computing power utilization, memory utilization, available storage space, storage write status, and power bus fluctuation status.

[0252] In one implementation, the intelligent cockpit operating system monitors the low-voltage battery voltage, power battery charge, and charging status through a power management service. The operating system also monitors processor temperature, NPU temperature, and power bus fluctuations through an environmental monitoring module. Furthermore, it monitors computing power utilization, training task progress, and storage write status through a task scheduler. Through multi-source status monitoring, the intelligent cockpit operating system can promptly detect security risks during the self-learning process.

[0253] In another implementation, the intelligent cockpit operating system sets different monitoring frequencies for different target self-learning strategies. The first self-learning strategy requires less computation, so the monitoring frequency can be lower. The second self-learning strategy requires more computation, so the monitoring frequency can be higher. By dynamically adjusting the monitoring frequency, the intelligent cockpit operating system can achieve a balance between resource consumption and safety monitoring.

[0254] Step S802: If the vehicle operating status or hardware resource status meets the preset interruption conditions, suspend or terminate the current self-learning process and save the breakpoint status information corresponding to the current self-learning process.

[0255] Here, preset interrupt conditions are used to characterize the triggering conditions that require the current self-learning process to stop or pause. Preset interrupt conditions may include: the vehicle switching from a non-driving state to a driving state, the vehicle shifting out of parking gear, the vehicle speed being greater than zero, the user pressing the brake pedal, a change in the vehicle's ignition status, the driver's side door being opened, the occupant status changing from empty to occupied, the low-voltage battery voltage falling below the interrupt voltage threshold, the power battery charge falling below the interrupt charge threshold, the processor temperature exceeding the interrupt temperature threshold, the computing power utilization rate exceeding the interrupt utilization rate threshold, the available storage space falling below the interrupt storage threshold, or abnormal fluctuations in the power bus.

[0256] When the vehicle's operating status or hardware resource status meets preset interrupt conditions, the intelligent cockpit operating system suspends or terminates the current self-learning process. Suspension pauses the current self-learning process while preserving the conditions for continued execution. Termination ends the current self-learning process and releases occupied resources. The intelligent cockpit operating system can choose to suspend or terminate based on the severity of the interrupt condition. For example, when a door is opened, the user presses the brake pedal, or the vehicle's ignition status changes, the intelligent cockpit operating system can suspend the current self-learning process. When the processor temperature exceeds the fuse threshold, the low-voltage battery voltage remains below the hard critical threshold, or the power bus experiences severe fluctuations, the intelligent cockpit operating system can terminate the current self-learning process and enter a locked state that prevents the self-learning process from being restarted for a preset time.

[0257] When suspending the current self-learning process, the intelligent cockpit operating system saves the breakpoint state information corresponding to the current self-learning process. The breakpoint state information is used to resume the execution of the current self-learning process later. The breakpoint state information may include at least one of the following: task identifier, target self-learning strategy, current processing stage, current training step size, processed sample position, gradient increment, temporary parameters, intermediate results of target update data, storage write location, checksum, task queue status, and generation time.

[0258] In one specific implementation, the intelligent cockpit operating system uses a Write-Ahead Log (WAL) to store breakpoint state information. The WAL is used to write log data before writing data to the target storage area. After each training step, the intelligent cockpit operating system writes the training breakpoint pointer, gradient increment, temporary parameters, and checksum to the log storage area, and updates the snapshot storage area after the log writing is complete. Through the WAL, the intelligent cockpit operating system can recover the most recent valid training state after a sudden power outage, system crash, or task interruption, reducing the risk of model adaptation information corruption.

[0259] In another specific implementation, the intelligent cockpit operating system breaks down the self-learning process into multiple atomic computation segments. After each atomic computation segment is completed, the intelligent cockpit operating system saves the breakpoint state information once. If the vehicle state changes, the intelligent cockpit operating system suspends the training task after the current atomic computation segment ends. By using atomic computation segments and breakpoint state information, the intelligent cockpit operating system can reduce redundant calculations and reduce the risk of data inconsistency during interruptions.

[0260] Step S803: If the preset self-learning admission conditions are met again, resume the current self-learning process based on the breakpoint status information.

[0261] Here, after the current self-learning process is suspended, the intelligent cockpit operating system continues to monitor vehicle status data. When the vehicle status data once again meets the preset self-learning access conditions, the intelligent cockpit operating system resumes the current self-learning process based on the breakpoint status information. Meeting the preset self-learning access conditions again may include the vehicle being in a non-driving state again, the power supply status stabilizing, the processor temperature dropping to a safe range, the computing power utilization rate decreasing, the available storage space meeting requirements, or the occupant status in the vehicle once again meeting the conditions.

[0262] When resuming execution of the current self-learning process, the intelligent cockpit operating system reads the breakpoint status information and performs integrity verification on it. Integrity verification may include checksum comparison, log tail validity check, task identifier consistency check, and version consistency check. If the breakpoint status information passes the integrity verification, the intelligent cockpit operating system continues execution of the current self-learning process from the position recorded in the breakpoint status information. If the breakpoint status information fails the integrity verification, the intelligent cockpit operating system can either roll back to the previous valid breakpoint or restart the current self-learning process.

[0263] In one specific implementation, after the vehicle re-enters a silent state, the intelligent cockpit operating system reads the WAL log, scans the last records in the log storage area, and calculates the checksum of the log file. If the checksum matches, the intelligent cockpit operating system loads the most recent valid training state and restores the training step counter to the step size recorded in the breakpoint state information. If the checksum does not match, the intelligent cockpit operating system rolls back to the previous valid log record and then loads the previous valid training state. Through this recovery method, the intelligent cockpit operating system can avoid repeatedly executing completed computational segments and also avoid using corrupted model adaptation information.

[0264] In another specific implementation, the intelligent cockpit operating system re-executes the access control decision before resuming the current self-learning process. Re-executing the access control decision ensures that the vehicle's state remains safe during recovery. If the vehicle's state again meets the preset interruption conditions during the recovery process, the intelligent cockpit operating system suspends or terminates the current self-learning process again. Through repeated access control decisions and breakpoint recovery, the intelligent cockpit operating system can adapt to the characteristics of the in-vehicle environment, such as unstable vehicle usage scenarios, users getting into the vehicle at any time, or changes in the vehicle's power supply status.

[0265] In an optional implementation, step S105 loads the target update data into the smart cockpit operating system, including the following steps S901-S902.

[0266] Step S901: Load the target update data into the standby loading area.

[0267] Here, the target update data is used to update the processing content used by the intelligent cockpit operating system when responding to user interaction requests. The target update data may include at least one of the following: updated prompt information, updated model adaptation information, and updated routing configuration information. The updated prompt information is used to adjust the intelligent cockpit operating system's role settings, output format, output length, tone style, task boundaries, tool call constraints, and business priorities during the model inference process. The updated model adaptation information is used to adjust the intelligent cockpit operating system's adaptability to user preferences, scenario rules, task execution logic, and complex business decisions. The updated routing configuration information is used to adjust the distribution method of user interaction requests among different models, different tool channels, or different business function modules.

[0268] The standby loading area is a region in the intelligent cockpit operating system used for temporarily loading target update data. The standby loading area can be a memory area, cache area, configuration loading area, model adaptation information loading area, or file mapping area. The current loading area is the loading area currently being used by the intelligent cockpit operating system. The current loading area stores currently effective prompts, model adaptation information, or routing configuration information. While the intelligent cockpit operating system loads target update data into the standby loading area, the current loading area continues to serve ongoing user interaction requests. By setting up a standby loading area, the intelligent cockpit operating system can complete the preparation of target update data without interrupting the current inference call chain.

[0269] In one implementation, the intelligent cockpit operating system loads the target update data via an AI Proxy (intelligent proxy gateway). The AI ​​Proxy is positioned between the business function module and the model inference module. When a business function module initiates an inference request to the model inference module, the AI ​​Proxy forwards the request and injects prompt information or model adaptation information into the inference request according to the currently effective loading region. After generating the target update data, the AI ​​Proxy first loads the target update data into a backup loading region, rather than directly replacing the current loading region. The currently processed inference request continues to respond based on the current loading region.

[0270] In one implementation, the intelligent cockpit operating system performs static verification on the target update data before loading it. Static verification may include at least one of the following: format verification, integrity verification, signature verification, version compatibility verification, size verification, resource usage verification, and security policy verification. Format verification confirms that the target update data can be recognized by the intelligent cockpit operating system. Integrity verification confirms that the target update data is not corrupted. Signature verification confirms that the source of the target update data is trustworthy. Version compatibility verification confirms that the target update data matches the current intelligent cockpit operating system version, model version, or business function module version. After the target update data passes static verification, the intelligent cockpit operating system loads the target update data into a backup loading area. If the target update data fails static verification, the intelligent cockpit operating system does not load the target update data and records the reason for the verification failure.

[0271] In one implementation, the target update data includes the updated prompt information. The intelligent cockpit operating system writes the updated prompt information into the prompt information cache in the backup loading area, and records the version number, effective scenario, effective user, generation time, source anomaly tag, and rollback identifier corresponding to the updated prompt information. The intelligent cockpit operating system performs syntax checks, token quantity checks, and security constraint checks on the updated prompt information in the backup loading area. Tokens represent word segmentation units used by the model when processing text.

[0272] In another implementation, the target update data includes updated model adaptation information. The intelligent cockpit operating system loads the updated model adaptation information into the model adaptation information cache in the backup loading area. The model adaptation information can be an adapter weight file, a personalized preference parameter table, a scene rule parameter table, or a model call configuration file. The intelligent cockpit operating system performs file verification, memory alignment, quantization format checking, and inference compatibility checking of the model adaptation information in the backup loading area.

[0273] In another implementation, the target update data includes updated routing configuration information. The intelligent cockpit operating system loads the updated routing configuration information into the routing table cache in the standby loading area. The routing configuration information may include the mapping between business scenarios and model services, tool channel selection rules, end-to-end cloud collaborative invocation rules, and canary deployment ratios. By loading the routing configuration information, the intelligent cockpit operating system can use the new request distribution strategy in subsequent user interaction requests.

[0274] In an optional implementation, the backup loading area uses a shadow buffer. The shadow buffer is a memory area independent of the current loading area. The intelligent cockpit operating system silently loads the target update data into the shadow buffer and completes initialization within it. The current loading area does not participate in the target update data writing process. Through the shadow buffer, the loading process of the target update data will not block the current inference call chain, nor will it change the processing result of ongoing user interaction requests.

[0275] Step S902: If the inference call link of the intelligent cockpit operating system meets the preset switching conditions, the inference call link is switched from the current loading area to the backup loading area so that the intelligent cockpit operating system can respond to user interaction requests based on target update data.

[0276] Here, the preset switching conditions are used to determine whether the intelligent cockpit operating system is suitable for switching the inference call chain from the current loading area to the backup loading area. The preset switching conditions may include at least one of the following: the target update data in the backup loading area has passed verification; the inference call chain is in an idle state; the current inference request has been completed; the current business function module is not in a critical control process; the effective time corresponding to the target update data has arrived; the effective scenario corresponding to the target update data matches the current business scenario; and the user identifier corresponding to the target update data matches the current user.

[0277] The inference call chain being idle means that the intelligent cockpit operating system has not detected any new requests entering the model inference module within a preset time window, or that any inference requests that have already been initiated have been responded to. The current inference request being completed means that the model inference, tool invocation, response output, and business execution processes corresponding to the user interaction request have ended. The current business function module not being in a critical control process means that the intelligent cockpit operating system is not executing any critical tasks related to driving safety, vehicle control, or strong real-time response. By judging these conditions, the intelligent cockpit operating system can reduce the impact of the handover process on foreground user interaction and vehicle control tasks.

[0278] In one implementation, the AI ​​Proxy continuously monitors the communication link status between the business function module and the model inference module. When the AI ​​Proxy detects an idle gap in the communication link, and the duration of the idle gap exceeds a preset switching time threshold, the AI ​​Proxy triggers a switching operation. The switching operation can be completed through pointer switching, routing table switching, configuration reference switching, or service handle switching. Pointer switching refers to the intelligent cockpit operating system switching the active pointer in the inference call link that points to the currently loaded region to an active pointer that points to an alternative loaded region. Through pointer switching, subsequent user interaction requests entering the inference call link will read the target update data in the alternative loaded region.

[0279] In one specific implementation, the intelligent cockpit operating system employs CAS (Compare-and-Swap) atomic operations to perform load region switching. CAS atomic operations are used to modify the active pointer without causing thread conflicts. When the intelligent cockpit operating system detects that a preset switching condition is met, it switches the active pointer from the current load region to a backup load region via a CAS atomic operation. After the switch is complete, new user interaction requests are inferred based on the target update data in the backup load region. User interaction requests already in the inference call chain continue to use the current load region before the switch until the processing of user interaction requests already in the inference call chain is complete. In this way, the intelligent cockpit operating system can avoid the forced interruption of requests that are being processed.

[0280] In another specific implementation, the intelligent cockpit operating system ensures the integrity of request processing through reference counting. The currently loaded region is not immediately released after the switchover is complete. The intelligent cockpit operating system sets a reference count for the currently loaded region. Each inference request using the currently loaded region corresponds to a reference count. When an inference request is completed, the intelligent cockpit operating system decrements the corresponding reference count. After the reference count reaches zero, the intelligent cockpit operating system releases the currently loaded region or retains it as a backup fallback region. Through reference counting, the intelligent cockpit operating system can ensure that user interaction requests entering the inference call chain before the switchover occurs are fully executed.

[0281] In an optional implementation, the intelligent cockpit operating system retains the replaced loading region as a fallback loading region after the switch is completed. The fallback loading region stores the prompts, model adaptation information, or routing configuration information that took effect before the switch. After the switch, the intelligent cockpit operating system monitors the performance of the target update data. Performance metrics may include user feedback scores, frequency of anomaly tags, inference success rate, response time, resource utilization, and system stability indicators. If the target update data leads to an increase in anomaly feedback, a higher inference failure rate, response time exceeding a preset threshold, or a decrease in system stability, the intelligent cockpit operating system switches the inference call chain from the backup loading region back to the fallback loading region. By using the fallback loading region, the intelligent cockpit operating system can mitigate the persistent impact of anomalies in the target update data.

[0282] In an optional implementation, the intelligent cockpit operating system can set gray-scale switching rules. These rules control which user interaction requests use the target updated data first, while other user interaction requests continue to use the data in the currently loaded area. For example, the intelligent cockpit operating system can control the scope of the target updated data based on user ID, business scenario, request ratio, or time window. During the gray-scale switching period, the intelligent cockpit operating system continuously compares the response effect corresponding to the target updated data with the response effect corresponding to the currently loaded area. When the response effect corresponding to the target updated data meets preset conditions, the intelligent cockpit operating system increases the usage ratio of the target updated data until the entire inference call chain switches to the backup loading area. Through gray-scale switching rules, the intelligent cockpit operating system can reduce the risk of the target updated data taking effect all at once.

[0283] In an optional implementation, the intelligent cockpit operating system records loading audit information after the handover is complete. Loading audit information may include the version number of the target updated data, loading time, handover time, handover duration, affected users, affected scenarios, loading area identifier before handover, loading area identifier after handover, verification results, rollback identifier, and running status. Loading audit information is used for subsequent troubleshooting, version management, and update effectiveness evaluation.

[0284] In one specific embodiment, the self-learning process of the intelligent cockpit operating system in navigation and voice interaction scenarios is used as an example to illustrate the self-learning method of the intelligent cockpit operating system provided in this application.

[0285] Vehicle user A frequently makes voice interaction requests such as "Navigate home and buy coffee on the way" when leaving get off work on weekdays. Upon receiving this request, the intelligent cockpit operating system converts the user's speech into text and sends the text to the navigation module and the inference call link. Based on the current vehicle location, the user's frequently visited destinations, map data, and the POI list, the intelligent cockpit operating system generates a navigation route and outputs the route instructions to user A via voice broadcast.

[0286] During the initial interaction, the intelligent cockpit operating system generated a lengthy route description, and the recommended coffee shop was far from User A's usual commute route. After the voice prompt began, User A said "Stop talking nonsense" and manually dragged waypoints on the navigation interface to adjust the route to the vicinity of another coffee shop. In responding to this voice interaction request, the intelligent cockpit operating system acquired user request information, system response information, action execution information, and vehicle environment information. The user request information included the voice-recognized text corresponding to "Navigate home and buy coffee on the way." The system response information included the original navigation route generated by the intelligent cockpit operating system, the recommended coffee shop information, and the voice prompt text. The action execution information included the navigation route distribution status, route planning interface call results, and voice prompt status. The vehicle environment information included the vehicle's current location, gear status, speed, low-voltage battery voltage, power battery charge, screen status, processor temperature, and computing power utilization.

[0287] Simultaneously, the intelligent cockpit operating system detected that User A uttered the phrase "Stop talking nonsense" and identified this speech as negative semantic information, thereby generating explicit feedback data. The intelligent cockpit operating system also detected that User A manually dragged waypoints after the navigation route was generated, thereby generating implicit feedback data. The intelligent cockpit operating system correlates the interaction context data with the interaction feedback data to form interaction data.

[0288] The intelligent cockpit operating system determines the feedback type and evaluation information based on interactive feedback data. For explicit feedback data corresponding to "Stop talking nonsense," the intelligent cockpit operating system classifies the feedback type as explicit negative feedback, sets the feedback score to a negative score, and labels the anomaly as "output verbose anomaly." For implicit feedback data corresponding to user A manually dragging waypoints, the intelligent cockpit operating system classifies the feedback type as implicit negative feedback, sets the feedback score to a negative score, and labels the anomaly as "path planning anomaly." Subsequently, the intelligent cockpit operating system correlates user request information, system response information, action execution information, vehicle environment information, feedback type, and feedback evaluation information to generate interactive trajectory data. The interactive trajectory data records the trajectory identifier, timestamp, user identifier, scene identifier, user request content, original navigation route, user-corrected navigation route, original broadcast text, feedback type, feedback score, and anomaly label.

[0289] Over the next three days, User A repeatedly made similar "navigate home and buy coffee on the way" requests during the same off-get off work hours, interrupting lengthy announcements multiple times and repeatedly adjusting the navigation route to the vicinity of the same brand of coffee shop. The intelligent cockpit operating system continuously generated corresponding interaction trajectory data and statistically analyzed the interaction trajectory data of the same user, the same scenario, and the same anomaly label. When the number of occurrences of the lengthy anomaly output within a preset time window reached a preset number, the intelligent cockpit operating system determined that the interaction anomaly type corresponded to the first self-learning strategy. When the path planning anomaly and the user-corrected route data met preset quality conditions, the intelligent cockpit operating system determined that the interaction anomaly type corresponded to the second self-learning strategy. Therefore, the intelligent cockpit operating system jointly determined the first and second self-learning strategies as the target self-learning strategy.

[0290] When the vehicle is parked at night, at zero speed, connected to a charging terminal, unoccupied, with the screen off, low-voltage battery voltage stable, power battery charge above a preset threshold, processor temperature below a preset threshold, and computing power utilization below a preset threshold, the intelligent cockpit operating system determines that the vehicle status data meets the preset self-learning admission conditions corresponding to the target self-learning strategy. The intelligent cockpit operating system then initiates the self-learning process corresponding to the target self-learning strategy.

[0291] For the first self-learning strategy, the intelligent cockpit operating system acquires target interaction trajectory data with abnormally long output labels from the interaction trajectory data, and analyzes the original broadcast text, user interruption locations, and negative semantic information of the user in the target interaction trajectory data. The intelligent cockpit operating system determines that the content to be adjusted in the prompt information is to reduce the length of navigation broadcasts, prioritize the output of navigation results, and reduce the explanation of route reasons. The intelligent cockpit operating system adjusts the current prompt information based on the content to be adjusted, resulting in updated prompt information. The updated prompt information is used to constrain the length of voice output in navigation scenarios and requires the intelligent cockpit operating system to prioritize providing brief navigation results when the user does not ask for reasons.

[0292] For the second self-learning strategy, the intelligent cockpit operating system filters target interaction trajectory data that meets preset quality conditions from the interaction trajectory data. The target interaction trajectory data includes user A's original voice request, the vehicle's current location, the time period before and after get off work, the original navigation route, the user's corrected navigation route, and the coffee shop location finally confirmed by the user. The intelligent cockpit operating system constructs supervised training samples based on the target interaction trajectory data. The input feature data in the supervised training samples includes the user's request content, scene identifier, time period, vehicle's current location, and candidate coffee shop information. The target output label data in the supervised training samples includes the user's corrected navigation route and the coffee shop location finally confirmed by the user. The intelligent cockpit operating system trains the model adaptation information based on the supervised training samples to obtain updated model adaptation information. The updated model adaptation information is used to improve the intelligent cockpit operating system's ability to adapt to user A's commuting route preferences and coffee shop preferences.

[0293] During the self-learning process, the intelligent cockpit operating system continuously monitors the vehicle's operating status and hardware resource status. When the intelligent cockpit operating system detects that the vehicle status still meets the preset self-learning entry conditions, the self-learning process continues. When the intelligent cockpit operating system detects that a door is open, someone has entered the vehicle, the low-voltage battery voltage is below a preset interrupt threshold, or the processor temperature exceeds a preset interrupt threshold, the intelligent cockpit operating system suspends the current self-learning process and saves the breakpoint state information. The breakpoint state information includes the current processing stage, the location of processed samples, the training step size, temporary parameters, and verification information. After the vehicle re-enters a state that meets the preset self-learning entry conditions, the intelligent cockpit operating system resumes the current self-learning process based on the breakpoint state information and continues to generate target update data.

[0294] After self-learning processing is complete, the intelligent cockpit operating system uses the updated prompts and model adaptation information as the target update data. The operating system performs format, integrity, and compatibility checks on the target update data. Once the target update data passes these checks, the operating system loads it into the backup loading area. The current loading area continues to provide services for ongoing user interaction requests. When the operating system detects that the inference call chain is idle and the target update data meets preset switching conditions, it switches the inference call chain from the current loading area to the backup loading area.

[0295] After the switch is complete, User A sends another voice interaction request: "Navigate home and buy coffee on the way." The intelligent cockpit operating system generates a shorter voice prompt based on the updated prompt information and prioritizes coffee shops and commuting routes that align with User A's historical habits based on the updated model adaptation information. Thus, the intelligent cockpit operating system can self-learn and update based on real user interaction data, while ensuring vehicle safety and continuous cockpit service, and the updated effects are reflected in subsequent user interaction requests.

[0296] Based on the above embodiments, this application provides an intelligent cockpit operating system, referring to... Figure 5 The intelligent cockpit operating system provided in this application embodiment includes: Data acquisition module 1 is used to acquire interactive data generated during the operation of the intelligent cockpit operating system.

[0297] The trajectory generation module 2 is used to generate interactive trajectory data that represents the interaction results based on the interactive data.

[0298] The strategy determination module 3 is used to determine the target self-learning strategy based on the interaction trajectory data; wherein, the target self-learning strategy includes a first self-learning strategy for adjusting the prompt information, and / or a second self-learning strategy for adjusting the model adaptation information.

[0299] The update generation module 4 is used to acquire vehicle status data and generate target update data based on the target self-learning strategy and interaction trajectory data, provided that the vehicle status data meets the preset self-learning admission conditions corresponding to the target self-learning strategy.

[0300] Loading module 5 is used to load target update data into the intelligent cockpit operating system, so that the intelligent cockpit operating system can respond to user interaction requests based on the target update data.

[0301] The intelligent cockpit operating system provided in this application embodiment, through the coordinated cooperation of the data acquisition module, trajectory generation module, strategy determination module, update generation module, and loading module, can convert the interactive data generated during the operation of the intelligent cockpit operating system into interactive trajectory data. Based on the interactive trajectory data, it determines the target self-learning strategy for adjusting prompt information and / or model adaptation information. When the vehicle status data meets the preset self-learning admission conditions, it generates target update data and loads it into the intelligent cockpit operating system. This enables the intelligent cockpit operating system to achieve self-learning updates based on user interaction feedback under the premise of vehicle status safety and continuous cockpit service, thereby improving the personalized adaptability, response accuracy, and interaction optimization efficiency of the intelligent cockpit operating system.

[0302] In an optional implementation, the interaction data includes interaction context data and interaction feedback data; the interaction feedback data includes explicit feedback data and / or implicit feedback data. The data acquisition module 1 is further configured to: Acquire the interaction context data generated by the intelligent cockpit operating system during the response to user interaction requests. Generate explicit feedback data when user input of error correction information, evaluation information, negative semantic information, or text feedback information is detected; and / or generate implicit feedback data when user interruption of broadcast, skipping of output content, manual modification of system execution results, closure of executed functions, or failure to perform subsequent operations within a preset time period is detected. Associate the interaction context data and interaction feedback data to determine the interaction data.

[0303] In an optional implementation, trajectory generation module 2 is further configured to: Based on interactive feedback data, the feedback type and evaluation information corresponding to the user's interaction request are determined; the feedback evaluation information includes a feedback score and / or anomaly label. User request information, system response information, executed action information, vehicle environment information, feedback type, and feedback evaluation information are correlated. Based on the correlation results, interaction trajectory data corresponding to the user's interaction request is generated.

[0304] In an optional implementation, the strategy determination module 3 is further configured to: Based on the feedback type, feedback score, and anomaly label in the interaction trajectory data, the interaction anomaly type corresponding to the interaction trajectory data is determined. Based on the preset correspondence between interaction anomaly types and self-learning strategies, the self-learning strategy corresponding to each interaction anomaly type is determined. If the self-learning strategy corresponding to the interaction anomaly type is the first self-learning strategy, it is determined as the target self-learning strategy. If the self-learning strategy corresponding to the interaction anomaly type is the second self-learning strategy, it is determined as the target self-learning strategy. If the self-learning strategy corresponding to the interaction anomaly type includes both the first and second self-learning strategies, both are determined as the target self-learning strategies.

[0305] In an optional implementation, the strategy determination module 3 is further configured to: Obtain target interaction trajectory data with the same anomaly label or belonging to the same interaction anomaly type from the interaction trajectory data. Determine the content to be adjusted in the prompt message based on the target interaction trajectory data. Adjust the current prompt message based on the content to be adjusted to obtain the updated prompt message. Set the updated prompt message as the target update data.

[0306] In an optional implementation, the strategy determination module 3 is further configured to: Target interaction trajectory data that meets preset quality conditions is selected from the interaction trajectory data. Supervised training samples are constructed based on the target interaction trajectory data; these samples include input feature data and target output label data. The model adaptation information in the intelligent cockpit operating system is trained based on the supervised training samples to obtain updated model adaptation information. The updated model adaptation information is then determined as the target update data.

[0307] In an optional implementation, the update generation module 4 is further configured to: Based on the target self-learning strategy, corresponding preset self-learning admission conditions are determined. If the vehicle state data indicates that the vehicle is in a non-driving state, the power supply status meets preset power supply conditions, the processor temperature meets preset temperature conditions, and the computing power utilization rate meets preset computing power conditions, then the vehicle state data is determined to meet the preset self-learning admission conditions. If the vehicle state data meets the preset self-learning admission conditions, the self-learning process corresponding to the target self-learning strategy is initiated to generate target update data.

[0308] In an optional implementation, the update generation module 4 is further configured to: Monitor vehicle operating status and hardware resource status. If the vehicle operating status or hardware resource status meets the preset interruption conditions, suspend or terminate the current self-learning process and save the breakpoint status information corresponding to the current self-learning process. If the preset self-learning admission conditions are met again, resume execution of the current self-learning process based on the breakpoint status information.

[0309] In an optional implementation, loading module 5 is further configured to: The target update data is loaded into the backup loading area. If the inference call chain of the intelligent cockpit operating system meets the preset switching conditions, the inference call chain is switched from the current loading area to the backup loading area, enabling the intelligent cockpit operating system to respond to user interaction requests based on the target update data.

[0310] The computer program product provided in this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0311] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0312] Furthermore, in the description of the embodiments of this application, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0313] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0314] In the description of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0315] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

Claims

1. A self-learning method for an intelligent cockpit operating system, characterized in that, include: Acquire interactive data generated during the operation of the intelligent cockpit operating system; Based on the interaction data, interaction trajectory data is generated to characterize the interaction results; A target self-learning strategy is determined based on the interaction trajectory data; wherein, the target self-learning strategy includes a first self-learning strategy for adjusting prompt information, and / or a second self-learning strategy for adjusting model adaptation information; Acquire vehicle status data, and if the vehicle status data meets the preset self-learning admission conditions corresponding to the target self-learning strategy, generate target update data based on the target self-learning strategy and the interaction trajectory data; The target update data is loaded into the smart cockpit operating system so that the smart cockpit operating system can respond to user interaction requests based on the target update data.

2. The self-learning method for the intelligent cockpit operating system according to claim 1, characterized in that, The interaction data includes interaction context data and interaction feedback data; the interaction feedback data includes explicit feedback data and / or implicit feedback data. The acquisition of interactive data generated during the operation of the intelligent cockpit operating system includes: Obtain the interaction context data generated by the intelligent cockpit operating system during the process of responding to user interaction requests; The explicit feedback data is generated when user input of error correction information, evaluation information, negative semantic information, or text feedback information is detected; and / or the implicit feedback data is generated when user interruption of broadcasting, skipping of output content, manual modification of system execution results, closure of executed functions, or failure to perform subsequent operations within a preset time period is detected. The interaction context data and the interaction feedback data are associated to determine the interaction data.

3. The self-learning method for the intelligent cockpit operating system according to claim 2, characterized in that, The generation of interaction trajectory data to characterize the interaction result based on the interaction data includes: Based on the interactive feedback data, the feedback type and feedback evaluation information corresponding to the user interaction request are determined; wherein, the feedback evaluation information includes a feedback score and / or anomaly label; The user request information, system response information, action execution information, vehicle environment information, feedback type, and feedback evaluation information are associated. Based on the association results, interaction trajectory data corresponding to the user's interaction request is generated.

4. The self-learning method for the intelligent cockpit operating system according to claim 3, characterized in that, The step of determining the target self-learning strategy based on the interaction trajectory data includes: Based on the feedback type, feedback score, and anomaly label in the interaction trajectory data, determine the interaction anomaly type corresponding to the interaction trajectory data; Based on the preset correspondence between interaction exception types and self-learning strategies, the self-learning strategy corresponding to the interaction exception type is determined. If the self-learning strategy corresponding to the interaction exception type is the first self-learning strategy, then the first self-learning strategy is determined as the target self-learning strategy. If the self-learning strategy corresponding to the interaction exception type is the second self-learning strategy, then the second self-learning strategy is determined as the target self-learning strategy. When the self-learning strategy corresponding to the interaction exception type includes the first self-learning strategy and the second self-learning strategy, the first self-learning strategy and the second self-learning strategy are determined as the target self-learning strategy.

5. The self-learning method for the intelligent cockpit operating system according to claim 4, characterized in that, When the target self-learning strategy includes the first self-learning strategy, generating target update data based on the target self-learning strategy and the interaction trajectory data includes: Obtain target interaction trajectory data with the same abnormal label or belonging to the same interaction abnormal type from the interaction trajectory data; Based on the target interaction trajectory data, determine the content to be adjusted in the prompt information; The current prompt message is adjusted based on the content to be adjusted, resulting in an updated prompt message; The updated prompt message is identified as the target update data.

6. The self-learning method for the intelligent cockpit operating system according to claim 4, characterized in that, When the target self-learning strategy includes the second self-learning strategy, generating target update data based on the target self-learning strategy and the interaction trajectory data includes: Target interaction trajectory data that meets preset quality conditions are selected from the interaction trajectory data; Supervised training samples are constructed based on the target interaction trajectory data; wherein, the supervised training samples include input feature data and target output label data; The model adaptation information in the intelligent cockpit operating system is trained based on the supervised training samples to obtain updated model adaptation information. The updated model adaptation information is determined as the target update data.

7. The self-learning method for the intelligent cockpit operating system according to claim 1, characterized in that, When the vehicle state data meets the preset self-learning admission conditions corresponding to the target self-learning strategy, generating target update data based on the target self-learning strategy and the interaction trajectory data includes: Based on the target self-learning strategy, determine the corresponding preset self-learning admission conditions; If the vehicle status data indicates that the vehicle is in a non-driving state, the power supply status meets the preset power supply conditions, the processor temperature meets the preset temperature conditions, and the computing power utilization rate meets the preset computing power conditions, then the vehicle status data is determined to meet the preset self-learning admission conditions. If the vehicle status data meets the preset self-learning admission conditions, the self-learning process corresponding to the target self-learning strategy is initiated to generate the target update data.

8. The self-learning method for the intelligent cockpit operating system according to claim 1, characterized in that, In the process of generating target update data based on the target self-learning strategy and the interaction trajectory data, the method further includes: Monitor vehicle operating status and hardware resource status; If the vehicle operating state or the hardware resource state meets the preset interruption conditions, the current self-learning process is suspended or terminated, and the breakpoint state information corresponding to the current self-learning process is saved. If the preset self-learning admission conditions are met again, the current self-learning process is resumed based on the breakpoint status information.

9. The self-learning method for the intelligent cockpit operating system according to claim 1, characterized in that, The step of loading the target update data into the smart cockpit operating system includes: Load the target update data into the backup loading area; If the inference call link of the intelligent cockpit operating system is detected to meet the preset switching conditions, the inference call link is switched from the current loading area to the backup loading area, so that the intelligent cockpit operating system can respond to user interaction requests based on the target update data.

10. An intelligent cockpit operating system, characterized in that, include: The data acquisition module is used to acquire interactive data generated during the operation of the intelligent cockpit operating system; The trajectory generation module is used to generate interaction trajectory data representing the interaction result based on the interaction data; The strategy determination module is used to determine a target self-learning strategy based on the interaction trajectory data; wherein the target self-learning strategy includes a first self-learning strategy for adjusting prompt information, and / or a second self-learning strategy for adjusting model adaptation information; An update generation module is used to acquire vehicle status data and, if the vehicle status data meets the preset self-learning admission conditions corresponding to the target self-learning strategy, generate target update data based on the target self-learning strategy and the interaction trajectory data. A loading module is used to load the target update data into the smart cockpit operating system, so that the smart cockpit operating system can respond to user interaction requests based on the target update data.

Images