Vehicle-mounted user intent recognition and task generation system based on large model

Abstract

This application relates to a vehicle-mounted user intent recognition and task generation system based on a large model, belonging to the technical field of automotive interaction. It includes: a user intent recognition module, used for: acquiring voice input from an in-vehicle user and parsing the voice input to obtain multiple user intents; a context awareness module, used for: collecting current environmental data of the vehicle and obtaining the current context based on the current environmental data using a large model; a task generation module, used for: searching the vehicle's task capability interfaces to generate task combinations based on the user intents and the current context; the task combinations involve calling different intelligent agents; a reflection module, used for: evaluating whether the generated task combinations are consistent with the user intents and the current context and adjusting inconsistent task combinations; and a task sequence orchestration module, used for: confirming the control order of different serial and parallel tasks based on the task combinations and calling the corresponding intelligent agents.

Description

Technical Field

[0001] This application relates to the technical field of automotive interaction, and in particular to an in-vehicle user intent recognition and task generation system based on a large model. Background Technology

[0002] Currently, vehicle cockpit dialogue systems all use command-based semantics. Users speak explicit commands, the in-vehicle dialogue system performs semantic recognition, and takes the next action based on the command. Current solutions generally suffer from insufficient generalization capabilities. For example, they can recognize commands like "turn on the air conditioner," but cannot recognize generalized commands like "adjust the air conditioner to a suitable temperature."

[0003] To address the generalization problem, large model capabilities are widely used in dialogue systems. Leveraging the reasoning and contextual understanding abilities of large models, they can significantly improve the ability of traditional predictive agents to recognize the user's specific intent, yielding excellent results. The introduction of large model capabilities has enhanced the generalization ability of speech, making human-vehicle dialogue increasingly natural. However, it has also brought new challenges. The introduction of large model technology has led users to place higher demands on dialogue systems, expecting them to better understand the intent within the dialogue. Because human speech naturally contains multiple intent scenarios, the current accurate classification methods based on commands in in-vehicle dialogue systems are no longer suitable for the broadened, generalized intents enhanced by large models, thus impacting the user experience.

[0004] Therefore, how to achieve smarter human-vehicle interaction that better meets user needs has become an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as insufficient generalization ability, inability to accurately identify multiple user intentions, and insufficient intelligence in task generation logic, this invention provides an in-vehicle user intention recognition and task generation system based on a large model.

[0006] The in-vehicle user intent recognition and task generation system based on a large model provided in this application adopts the following technical solution.

[0007] A vehicle-mounted user intent recognition and task generation system based on a large model includes:

[0008] User intent recognition module; used for: acquiring voice input from users inside the vehicle, and parsing the voice input to obtain multiple user intents;

[0009] The context awareness module is used to: collect the vehicle's current environmental data and obtain the current context based on the current environmental data using a large model;

[0010] The task generation module is used to: generate a task combination based on the user intent and the task capability interface for searching vehicles in the current context; the task combination involves calling different intelligent agents;

[0011] The reflection module is used to: evaluate whether the generated task combinations are consistent with the user intent and the current context, and adjust inconsistent task combinations;

[0012] The task sequence orchestration module is used to: determine the control order of different serial and parallel tasks based on task combinations and call the corresponding agents.

[0013] Optionally, the user intent recognition module includes:

[0014] Input data processing unit; receives and processes user voice input;

[0015] Large model understanding unit; uses large models to analyze user voice input and identify user subjective intentions;

[0016] Intent classification unit; transforming user intent into different dimensions of demands; the different dimensions of demands include: rest and relaxation, entertainment and leisure, soothing and calming, environmental control, navigation and driving, communication and social interaction, safety reminders, and work productivity.

[0017] Optionally, the context-aware module includes:

[0018] Environmental data acquisition unit; collects current environmental data of the vehicle, including: vehicle operating status, in-vehicle environment, external environment, safety status, and user status;

[0019] Large-scale model context analysis unit; based on the analysis of current environmental data, the current context is obtained.

[0020] Optionally, the system further includes a self-learning module; the self-learning module is used to: collect user feedback data and perform intent relevance analysis; generate standard positive and negative databases based on the feedback data and adjust the large model.

[0021] Optionally, standard positive and negative databases are generated based on the feedback data, including:

[0022] After each task is executed, user feedback data will be recorded; the feedback data includes: task completion status, user satisfaction evaluation, behavioral patterns, and task execution results; when the user manually adjusts the tasks generated by the system, the relevant task generation data and task arrangement order will be used as negative feedback.

[0023] Optionally, the task generation module includes:

[0024] Task search unit; searches for the vehicle's task capability interfaces based on user intent and current context;

[0025] Large model task generation unit; generates task combinations that conform to user intent and context based on pre-trained large models.

[0026] Optionally, the self-learning module is also used to: monitor the completion status of tasks and changes in current environmental data in real time during the execution of task combinations, and re-evaluate the rationality of task combinations based on the current context;

[0027] The self-learning module is also used to record and analyze users' long-term behavioral data and extract users' habits and preferences.

[0028] Optionally, the reflection module includes:

[0029] Reflection and evaluation unit; evaluate whether the generated task combination is consistent with the user's intent and the current context;

[0030] Safety assessment unit; ensuring the rationality and safety of task combinations;

[0031] Task correction unit; adjusts task combinations that do not meet expectations.

[0032] This invention provides a vehicle-mounted user intent recognition and task generation system based on a large model. By introducing the semantic reasoning and contextual understanding capabilities of the large model, it significantly improves the accuracy of user intent recognition, enabling the parsing of multiple intents in complex contexts and dynamically generating task combinations that meet actual needs based on the current vehicle situation. The system optimizes the execution order of tasks through a task sequence orchestration module, ensuring efficient coordination between serial and parallel tasks. Simultaneously, a reflection module evaluates and adjusts the rationality and safety of tasks, enhancing the system's safety and reliability. A self-learning module continuously optimizes the weights of the large model and the task generation logic based on user behavior and feedback, achieving personalized customization and gradually meeting the unique needs of users. A context-aware module combines data from the vehicle's internal and external environment with large model reasoning analysis to accurately adapt to diverse driving scenarios, providing more precise reference for task generation. This system effectively solves the problems of insufficient generalization ability, inaccurate recognition of multiple intents, and unintelligent task generation in existing technologies, significantly improving the intelligence level of in-vehicle systems and user interaction experience, providing users with safer, more comfortable, and more efficient driving services. Attached Figure Description

[0033] Figure 1 This is a system block diagram of the in-vehicle user intent recognition and task generation system based on a large model, according to an embodiment of this application. Detailed Implementation

[0034] Large model capabilities are widely used in dialogue systems. Leveraging the reasoning and contextual understanding abilities of large models, they can significantly improve the ability of traditional predictive agents to recognize the user's specific intent, achieving excellent results. The introduction of large model capabilities has enhanced the generalization ability of speech, making human-vehicle dialogue increasingly natural. However, it has also brought new challenges. The introduction of large model technology has led users to place higher demands on dialogue systems, expecting the system to better understand the intent within the dialogue. Because human speech naturally contains multiple intent scenarios, the current accurate classification methods of command-based dialogue systems are no longer suitable for the broadened, generalized intents enhanced by large models, impacting user experience. For example, a user's request in a dialogue system might be: "What is the weather like in Beijing right now?" Follow-up user surveys reveal the following possible intents:

[0035] Today's weather forecast for Beijing, the corresponding intelligent agent is the weather broadcast;

[0036] The overall climate change over a relatively long period of time is represented by a weather query from the intelligent agent.

[0037] From a geographical and historical perspective, the overall weather conditions in the Beijing area during this period correspond to encyclopedic knowledge.

[0038] Current in-vehicle dialogue systems understand human language through a single classification, specifically selecting a single agent and slot to achieve a functional loop. However, a user's specific intentions are related to vehicle status, geographical location, and mood; a single agent classification cannot meet the user's contextual needs.

[0039] To address the aforementioned issues, this application proposes a vehicle-mounted user intent recognition and task generation system based on a large model.

[0040] The following is in conjunction with the appendix Figure 1 The present application will be further described with reference to specific embodiments:

[0041] First, it should be noted that in the description of this application, the use of directional terms such as "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer" indicates the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for descriptive purposes and do not indicate or imply that the device or component 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 use of numerical quantifiers such as "first," "second," and "third" is for descriptive purposes only and should not be construed as indicating or implying relative importance. Additionally, in this application, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, interference fits, transition fits, or integral connections; they can refer to direct connections or indirect connections through an intermediate medium. Therefore, those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0042] This application discloses an in-vehicle user intent recognition and task generation system based on a large model. (Refer to...) Figure 1 As one implementation of a large-model-based in-vehicle user intent recognition and task generation system, the large-model-based in-vehicle user intent recognition and task generation system includes:

[0043] User intent recognition module 101; used for: acquiring voice input from users inside the vehicle, and parsing the voice input to obtain multiple user intents;

[0044] Context perception module 102 is used to: collect the current environmental data of the vehicle and obtain the current context based on the current environmental data through a large model;

[0045] Task generation module 103 is used to: generate a task combination based on the user intent and the task capability interface for searching vehicles in the current context; the task combination involves calling different intelligent agents;

[0046] The reflection module 104 is used to: evaluate whether the generated task combination is consistent with the user intent and the current context, and adjust the inconsistent task combination;

[0047] The task sequence orchestration module 105 is used to: determine the operation order of different serial and parallel tasks according to the task combination and call the corresponding intelligent agents.

[0048] Specifically, the user intent recognition module 101 parses the user's voice input to understand multiple intent expressions in complex contexts. The context awareness module 102 collects data from the vehicle's internal and external environment and analyzes this data using a large model to obtain current contextual information. The task generation module 103 combines user intent and the current context to determine the most suitable task combination and completes these tasks by calling different intelligent agents, ensuring the effectiveness and efficiency of task execution. For example, if the user simultaneously requests to play music and set a destination, the system will start navigation first and then play music to avoid the navigation voice prompts being overridden by the music. The reflection module 104 is responsible for evaluating whether the generated task combination matches the user's actual needs. If they do not match, it will automatically adjust the task combination to ensure that the final execution result better meets the user's needs. The task sequence orchestration module 105 is responsible for determining the execution order between tasks and how to coordinate the simultaneous operation of multiple tasks. The vehicle-mounted user intent recognition and task generation system based on a large model significantly improves the intelligence level of the vehicle system through the application of deep learning and artificial intelligence technologies, providing users with a safer, more convenient, and more comfortable driving environment.

[0049] The following example illustrates this point:

[0050] In a car scenario, a user inputs "I'm a little tired" via voice. The system first uses the user intent recognition module to identify the user's request as "rest and relax." Next, the context awareness module collects current environmental data, such as the vehicle speed being 0, the gear being in park, the air conditioning being on, and the user being alone in the car. The large-scale model task generation module generates task combinations based on this context and user intent, such as adjusting the seat to a comfortable position, playing soft music, or activating a nap mode. The reflection mechanism module checks whether these tasks are consistent with the current context and whether there are any safety hazards. If a task is found to be undesirable, such as a conflict between adjusting the window and the air conditioning setting, the task is corrected and regenerated. The task sequence orchestration module orchestrates tasks in parallel or sequentially to ensure logical execution. After task execution, if the user switches the music to another softer track, the self-learning module collects this user data as a bad case for fine-tuning the task generation model. After training with a sufficient amount of data, the model will be able to provide personalized responses to the user. Through this method, the present invention can intelligently identify user intentions and generate appropriate task combinations according to specific situations, which greatly improves the user's driving experience and the intelligence level of the system.

[0051] As a specific implementation of a vehicle-mounted user intent recognition and task generation system based on a large model, the system acquires the voice input of the user inside the vehicle, parses the voice input to obtain multiple user intents, including:

[0052] The user's voice input in the vehicle is parsed to obtain user input x;

[0053] The user input x is inferred using a large model to obtain the output probability corresponding to each agent; wherein, the large model is configured with a set of agents and agent weights; each element in the set of agents corresponds to an intent; the agent weights include the preset weights of each agent; the weights represent the importance of the agent in the inference process;

[0054] For each agent, determine whether its output probability is greater than a preset probability threshold; if so, it is determined to meet the condition.

[0055] Output all agents that meet the conditions as multiple user intents obtained from parsing.

[0056] Specifically, the system parses the user's voice input into the vehicle to obtain user input x, accurately converting the initial voice data into text format suitable for subsequent processing. A large model is used to infer the user input x, obtaining the output probability for each agent, and evaluating the relevance of each agent to the user's intent, thereby improving the accuracy of intent recognition. Each agent has a specific weight, which helps the model better understand which agents are more important for a specific task, thus giving them higher priority when recognizing user intent. For each agent, its output probability is checked against a preset probability threshold, ensuring that only agents with high-probability matches are selected, reducing false matches and improving system response quality and user experience. The system outputs all agents that meet the conditions as multiple parsed user intents, filtering out the group of agents most likely to match the user's intent, enabling the system to respond to user requests more accurately.

[0057] As a specific implementation of a vehicle-mounted user intent recognition and task generation system based on a large model, the system determines the control sequence of different serial and parallel tasks according to the task combination and invokes the corresponding intelligent agents, including:

[0058] Determine the response delay weights of the agent: ; where each intelligent agent S i The response delay is t i The lower the latency, the higher the weight.

[0059] Determine agent priority weights ; among them, each intelligent agent Corresponding to a static priority weight ;

[0060] Determine the weight of user personalized preferences ; Used to characterize user preferences for different intelligent agents;

[0061] Determine the weights of vehicle restriction states ;

[0062] The arbitration function A is determined based on the response delay weight, agent priority weight, user personalized preference weight, and vehicle restriction state weight; the formula for the arbitration function A is as follows: ; It is an intelligent agent The output probability distribution represents the probability that the user's intention belongs to that agent, where i = 1, 2, ..., n, and n is the total number of agents;

[0063] The agent with the largest value in arbitration function A is selected. The agent that responds first responds to the other agents in descending order of the value of the arbitration function A.

[0064] Specifically, by considering the response time of agents, those with faster response times can be prioritized, thereby accelerating the overall system response speed. Each agent has a static priority weight, and some key or frequently used functions are given priority consideration, ensuring that important functions are processed first, improving user experience. Considering users' personalized needs, the system can adjust the selection of agents based on users' historical behavior and preferences, making services more aligned with user habits and enhancing user satisfaction. The vehicle's state (such as driving status, safety requirements, etc.) affects the available agents and services. By considering these constraints, the system can make more reasonable and safer service decisions. By comprehensively considering the above factors, the arbitration function A can fully evaluate the applicability of each agent, thereby selecting the agent most suitable for the current situation. Selecting the agent with the largest value of the arbitration function A as the agent's output ensures that the best agent is selected under the current circumstances, maximizing user satisfaction while also considering factors such as system response speed and security.

[0065] As a specific implementation of a vehicle-mounted user intent recognition and task generation system based on a large model, the system further includes: a self-learning module; the self-learning module is used to: collect user feedback data and perform intent correlation analysis; generate standard positive and negative databases based on the feedback data and adjust the large model.

[0066] Specifically, after each task is executed, the system records user feedback data, such as task completion status, user satisfaction rating, behavioral patterns, and task execution results. If the user frequently manually adjusts system-generated tasks (such as canceling a task or changing the task order), the system will treat this data as negative feedback, indicating that the current task generation logic may be inadequate.

[0067] By analyzing this feedback data, the system gradually learns users' preferences and habits. For example, if a user habitually plays soft music and turns off the car lights while resting, the system will automatically generate a combination of these two tasks when it recognizes the user's "rest" intention in the future, thus making the task generation more in line with the user's expectations.

[0068] As a specific implementation of a vehicle-mounted user intent recognition and task generation system based on a large model, standard positive and negative databases are generated based on feedback data, including:

[0069] After each task is executed, user feedback data will be recorded; the feedback data includes: task completion status, user satisfaction evaluation, behavioral patterns, and task execution results; when the user manually adjusts the tasks generated by the system, the relevant task generation data and task arrangement order will be used as negative feedback.

[0070] As a specific implementation of a large-model-based in-vehicle user intent recognition and task generation system, the self-learning module is also used for:

[0071] The weights of the intelligent agent are dynamically updated based on user feedback and preferences.

[0072] The dynamic updating of the agent's weights based on user feedback and preferences specifically includes:

[0073] Set a weight vector w and an update function U; where the weight vector w contains the weights of each agent. The weights; the update function U is used to dynamically adjust the weights;

[0074] Set the feedback vector r; the feedback vector r contains the user's feedback information to each agent's output;

[0075] The update function U takes the feedback vector r and the current weight vector w as input and outputs a new weight vector. ;in ;in The learning rate is used to control the update speed. ;

[0076] Updated weight vector This will be used in the next agent arbitration process, in which each agent... Output probability distribution It will be weighted according to the corresponding weights; n represents the total number of agents.

[0077] Specifically, by collecting user feedback and preferences, the system can continuously adjust the weights of the agents to better align with user preferences. The weight vector *w* contains the weights of each agent, while the update function *U* dynamically adjusts these weights based on user feedback. The feedback vector *r* contains user feedback on each agent's output, guiding the weight updates. By utilizing user feedback, the system can better understand user needs and make more accurate service decisions. By controlling the speed of weight updates using a learning rate α, the system can quickly adapt to user needs while avoiding instability caused by over-adjustment. During arbitration, the output probability distribution of each agent is weighted according to its corresponding weight. The system tends to select agents that have demonstrated better performance based on user feedback, thus providing higher-quality service.

[0078] As one implementation of a vehicle-mounted user intent recognition and task generation system based on a large model, the self-learning module is also used to: monitor the completion status of tasks and changes in current environmental data in real time during the execution of task combinations, and re-evaluate the rationality of task combinations according to the current context.

[0079] The self-learning module is also used to record and analyze users' long-term behavioral data and extract users' habits and preferences.

[0080] Specifically, during task execution, the system can monitor task completion and environmental changes in real time, and reassess the task's rationality based on the current context. For example, if the system generates a task to play music after recognizing the user's intention to "rest," but detects a sudden increase in noise outside the car (such as traffic noise), the system will adjust the music volume or switch to noise-canceling music based on this contextual change. Furthermore, if task conflicts occur during execution (such as opening the car window and starting the air conditioning simultaneously), a reflection mechanism will intervene, adjusting the task generation strategy to ensure the continuity and rationality of task execution in the new context.

[0081] The system continuously records and analyzes users' long-term behavioral data to extract their habits and preferences. For example, if a user frequently chooses to play specific types of music or adjust the color of the car's interior lights while driving at night, the system will learn this pattern and prioritize generating tasks that match the user's habits in similar situations.

[0082] The system not only considers the user's current needs but also their past behavioral patterns to predict potential future needs, thereby generating relevant task combinations in advance. For example, if a user is accustomed to taking a break every hour during long-distance night driving, the system will automatically suggest regular breaks or arrange appropriate rest stops when it recognizes a long-distance driving scenario.

[0083] For example, a user often takes a break after driving for two hours during long drives, and during these breaks, they like to play soft music and adjust the seat back angle. By recording these behaviors, the system gradually learns the user's rest preferences. Next, when the user says "I'm a little tired" via voice during a long drive, the system will generate a task combination suggesting stopping for a rest, automatically playing soft music and adjusting the seat angle. If the user also manually adjusts the interior lighting brightness, the system will record this and add the lighting adjustment task to the next rest session.

[0084] The system will directly generate such personalized task combinations when performing similar tasks again, thereby reducing the user's operation steps and providing services that better meet individual needs.

[0085] As one implementation of a vehicle-mounted user intent recognition and task generation system based on a large model, the system collects current environmental data of the vehicle and uses a large model to obtain the current context based on the current environmental data. Specific steps include:

[0086] A preliminary analysis of the vehicle status information identifies at least one key feature for judging the current situation.

[0087] Set an initial inference path and set feature weights based on the branching logic of the decision tree;

[0088] Inference is performed based on the current inference path, and the path is adjusted based on the output feedback. During the path adjustment process, backpropagation is used to adjust the feature weights.

[0089] After multiple inference paths, a converged inference path is obtained, and the current situation is obtained based on the converged inference path.

[0090] Specifically, by conducting preliminary analysis of key features of vehicle status information, the system can quickly capture important clues about the current situation. Through the branching logic of the decision tree, the system can select an appropriate inference path based on the current key features and set corresponding weights according to the importance of the features. The system then performs inference based on the current path and adjusts the path according to the output feedback, while also adjusting the feature weights through backpropagation during the path adjustment process. The system dynamically adjusts the inference path based on the feedback from the inference results and continuously optimizes the situation recognition process through backpropagation, thereby improving accuracy. Through multiple iterative inferences until the path converges, the system ensures the stability of the situation recognition results and reduces errors caused by a single inference path.

[0091] As one implementation of a large-model-based in-vehicle user intent recognition and task generation system, the user intent recognition module includes:

[0092] Input data processing unit; receives and processes user voice input;

[0093] Large model understanding unit; uses large models to analyze user voice input and identify user subjective intentions;

[0094] Intent classification unit; transforming user intent into different dimensions of demands; the different dimensions of demands include: rest and relaxation, entertainment and leisure, soothing and calming, environmental control, navigation and driving, communication and social interaction, safety reminders, and work productivity;

[0095] The context-aware module includes:

[0096] Environmental data acquisition unit; collects current environmental data of the vehicle, including: vehicle operating status, in-vehicle environment, external environment, safety status, and user status;

[0097] Large-scale model context analysis unit; Based on the analysis of current environmental data using a large-scale model, the current context is obtained.

[0098] The task generation module includes:

[0099] Task search unit; searches for the vehicle's task capability interfaces based on user intent and current context;

[0100] Large model task generation unit; generates task combinations that conform to user intent and context based on pre-trained large models;

[0101] The reflection module includes;

[0102] Reflection and evaluation unit; evaluate whether the generated task combination is consistent with the user's intent and the current context;

[0103] Safety assessment unit; ensuring the rationality and safety of task combinations;

[0104] Task correction unit; adjusts task combinations that do not meet expectations.

[0105] It should be noted that the above embodiments are only used to illustrate this application and are not intended to limit the technical solutions described in this application. Although this specification has described this application in detail with reference to the above embodiments, those skilled in the art should understand that they can still make modifications or equivalent substitutions to this application. All technical solutions and improvements that do not depart from the spirit and scope of this application should be covered within the scope of the claims of this application.

Claims

1. A vehicle-mounted user intent recognition and task generation system based on a large model, characterized in that, include: User intent recognition module; Used for: acquiring voice input from users inside the vehicle, and parsing the voice input to obtain multiple user intentions; The context awareness module is used to: collect the vehicle's current environmental data and obtain the current context based on the current environmental data using a large model; The task generation module is used to: generate a task combination based on the user intent and the task capability interface for searching vehicles in the current context; the task combination involves calling different intelligent agents; The reflection module is used to: evaluate whether the generated task combinations are consistent with the user intent and the current context, and adjust inconsistent task combinations; The task sequence orchestration module is used to: determine the control order of different serial and parallel tasks based on task combinations and call the corresponding intelligent agents; Based on the task combination, determine the control order of different serial and parallel tasks and invoke the corresponding intelligent agents, including: Determine the response delay weights of the agent: ; Determine agent priority weights ; among them, each intelligent agent Corresponding to a static priority weight ; Determine the weight of user personalized preferences ; Used to characterize user preferences for different intelligent agents; Determine the weights of vehicle restriction states The arbitration function A is determined based on the response delay weight, agent priority weight, user personalized preference weight, and vehicle restriction state weight; the formula for the arbitration function A is as follows: ; It is an intelligent agent The output probability distribution; The agent with the largest value in arbitration function A is selected. The agent that responds first responds to the other agents in descending order of the value of the arbitration function A. The system also includes a self-learning module; the self-learning module is used to: collect user feedback data and perform intent relevance analysis; generate standard positive and negative databases based on the feedback data and adjust the large model; the self-learning module is also used to: The weights of the intelligent agent are dynamically updated based on user feedback and preferences. The dynamic updating of the agent's weights based on user feedback and preferences specifically includes: Set a weight vector w and an update function U; where the weight vector w contains the weights of each agent. The weights; the update function U is used to dynamically adjust the weights; Set the feedback vector r; the feedback vector r contains the user's feedback information to each agent's output; The update function U takes the feedback vector r and the current weight vector w as input and outputs a new weight vector; where ;in The learning rate is used to control the update speed. ; Updated weight vector This will be used in the next agent arbitration process, in which each agent... Output probability distribution It will be weighted according to the corresponding weights; n represents the total number of agents.

2. The in-vehicle user intent recognition and task generation system based on a large model according to claim 1, characterized in that, The user intent recognition module includes: Input data processing unit; receives and processes user voice input; Large model understanding unit; uses large models to analyze user voice input and identify user subjective intentions; Intent classification unit; transforming user intent into different dimensions of demands; the different dimensions of demands include: rest and relaxation, entertainment and leisure, soothing and calming, environmental control, navigation and driving, communication and social interaction, safety reminders, and work productivity.

3. The in-vehicle user intent recognition and task generation system based on a large model according to claim 2, characterized in that, The context-aware module includes: Environmental data acquisition unit; collects current environmental data of the vehicle, including: vehicle operating status, in-vehicle environment, external environment, safety status, and user status; Large-scale model context analysis unit; based on the analysis of current environmental data, the current context is obtained.

4. The in-vehicle user intent recognition and task generation system based on a large model according to claim 3, characterized in that, Based on the feedback data, standard positive and negative databases are generated, including: After each task is executed, user feedback data will be recorded; the feedback data includes: task completion status, user satisfaction evaluation, behavioral patterns, and task execution results; when the user manually adjusts the tasks generated by the system, the relevant task generation data and task arrangement order will be used as negative feedback.

5. The in-vehicle user intent recognition and task generation system based on a large model according to claim 4, characterized in that, The task generation module includes: Task search unit; searches for the vehicle's task capability interfaces based on user intent and current context; Large model task generation unit; Generate task combinations that match user intent and context based on pre-trained large models.

6. The in-vehicle user intent recognition and task generation system based on a large model according to claim 5, characterized in that, The self-learning module is also used to: monitor the completion status of tasks and changes in current environmental data in real time during the execution of task combinations, and re-evaluate the rationality of task combinations based on the current situation; The self-learning module is also used to record and analyze users' long-term behavioral data and extract users' habits and preferences.

7. The in-vehicle user intent recognition and task generation system based on a large model according to claim 5, characterized in that, The reflection module includes: Reflection and evaluation unit; evaluate whether the generated task combination is consistent with the user's intent and the current context; Safety assessment unit; ensuring the rationality and safety of task combinations; Task correction unit; adjusts task combinations that do not meet expectations.

Images