Data processing method and system for intelligent automobile agent

Abstract

The invention relates to the technical field of agents, in particular to a data processing method and system for an intelligent automobile agent, and the method comprises the steps: collecting an MCP process for a vehicle, and extracting the current semantics of original input information when a user intention is received; searching the knowledge network to obtain matched memory information, and inputting the matched memory information into an inference engine to generate an execution plan; and calling the corresponding MCP processes to execute the execution plans respectively, and taking the execution process as new user interaction data. In order to solve the problem that a vehicle-mounted intelligent agent in the prior art is inaccurate in judging user intention and calling tools, the intelligent agent and a corresponding knowledge network are constructed in a vehicle-mounted system, interaction data generated by a user are processed and stored, and in the actual prediction process, the user intention is predicted accurately. According to the invention, prediction is carried out based on the memory entries stored in the knowledge network in combination with the MCP process of the vehicle operation application, the corresponding execution plan is generated, and the accuracy of judgment of user intentions and tool calling is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent technology, and specifically to a data processing method and system for intelligent vehicle intelligent agents. Background Technology

[0002] An intelligent in-vehicle infotainment system (IVI) is an intelligent control system configured in intelligent vehicles to provide users with functions such as voice control, adaptive adjustment of cabin parameters, and operation of entertainment applications.

[0003] For example, patent application CN202211058831.5 discloses an intelligent cockpit domain controller and its control method, as well as a vehicle. This cockpit domain controller integrates an instrument panel, infotainment controller, networking module, remote management unit, low-speed pedestrian alarm, fatigue monitoring, driving recorder, head-up display, etc., enabling not only a reconstruction of the automotive electronic and electrical architecture but also conforming to the development trend of electronic and electrical architecture. It also optimizes problems such as low bandwidth, network flattening, and the need for a separate ECU unit for each function in distributed electronic and electrical architectures. This reduces the overall vehicle electrical cost, creates value for customers, and improves customer experience. By further centralizing the control logic, development efficiency is improved, and software iteration and updates are easier. Simultaneously, the cockpit domain controller, through a high-computing-power SOC, can enhance the customer experience and solve bottlenecks such as the large number of controllers and the dispersed multi-functional logic in distributed architectures.

[0004] For example, patent application CN202311098248.1 discloses an in-vehicle intelligent device interaction method, dashboard, intelligent cockpit, and vehicle thereof. The method includes the following steps: acquiring gesture information or touch operation commands from occupants to open a movable flip dashboard; detecting occupants on the movable flip dashboard and moving it to a position accessible to the occupants; and intelligently interacting with in-vehicle devices through the movable flip dashboard. This method enables the movable flip dashboard to be opened based on occupants' gestures or touch operations, and allows it to be moved to a position accessible to the occupants. Furthermore, it facilitates intelligent interaction with in-vehicle devices through the movable flip dashboard, achieving intelligent and automated operation of in-vehicle devices and enhancing the user experience.

[0005] However, in actual implementation, the inventors found that most existing vehicle-mounted agent frameworks use hard-coded tool lists, lack dynamic discovery capabilities, have imperfect tool information discovery mechanisms, and the tool descriptions do not meet the understanding capabilities of LLM (such as lack of natural language descriptions), leading to inaccurate inference decisions by the agent and ultimately affecting the accuracy of user experience. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, a data processing method for intelligent vehicle intelligent agents is provided. On the other hand, it also provides a corresponding system.

[0007] The specific technical solution is as follows: A data processing method for an intelligent vehicle agent includes a memory construction process and an execution process; The memory construction process includes: Step A1: Interact with the user and obtain user interaction data, and deduce atomic knowledge units based on the user interaction data; Step A2: Link the atomic knowledge units to form a knowledge network; The execution process includes: Step B1: When raw input information containing user intent is received, the vehicle's MCP process is collected, and the current semantics are extracted from the raw input information; Step B2: Search the knowledge network based on the current semantics to obtain matching memory information; Step B3: Input the original input information, the matching memory information, and the MCP process into the inference engine to generate an execution plan; Step B4: Invoke the corresponding MCP process to execute the execution plan, and use the current execution process as the new user interaction data.

[0008] On the other hand, step A1 includes: Step A11: Extract the original vehicle input data when the user interacts with the in-vehicle device, and perform text conversion based on the original vehicle input data to generate descriptive text; Step A12: Segment the description text to obtain multiple unit entities; Step A13: Perform logical deduction on the description text based on the unit entity to obtain the atomic knowledge unit.

[0009] On the other hand, step A2 includes: Step A21: Search for associated nodes of the atomic knowledge unit relative to the knowledge graph to establish multi-dimensional links; Step A22: Add the atomic knowledge unit to the knowledge graph according to the multi-dimensional links.

[0010] On the other hand, step A21 includes: Step A211: Match the atomic knowledge units with associated nodes in the knowledge graph based on semantic features; Step A212: Derive the correlation between the atomic knowledge unit and the associated node to obtain the association category; Step A213: Generate the multi-dimensional link based on the associated category.

[0011] On the other hand, step A22 includes: Step A221: Determine the relative relationship between the atomic knowledge unit and the associated node according to the association category; Step A222: Select the corresponding fusion rule based on the relative relationship, and then update the atomic knowledge unit to the knowledge graph based on the fusion rule.

[0012] On the other hand, step B1 includes: Step B11: When the user intent is received, the vehicle status data is collected and used together with the user intent as the raw input; Step B12: Extract the currently running MCP process from the MCP host; Step B13: Extract the semantic information from the original input.

[0013] On the other hand, step B2 includes: Step B21: Based on the current semantic input, the knowledge network searches for matching associated nodes; Step B22: Extract knowledge slices based on the associated nodes; Step B23: Assemble all the knowledge slices to obtain the matching memory information output.

[0014] On the other hand, step B3 includes: Step B31: Input the original input information and the matching memory information into the inference engine to obtain the planning node; Step B32: The inference engine matches the MCP process according to the planning node to determine the matching process that needs to be executed for the planning node; Step B33: Generate the execution plan according to the matching process and the planning node.

[0015] On the other hand, prior to performing step B1, the procedure also includes: Step B01: When the MCP host calls the corresponding in-vehicle application, it creates the corresponding MCP process.

[0016] A data processing system for an intelligent vehicle's intelligent agent, used in the aforementioned data processing method.

[0017] The above technical solution has the following advantages or beneficial effects: To address the inaccuracies in judging user intentions and calling tools in existing in-vehicle intelligent agents, this embodiment constructs an intelligent agent and a corresponding knowledge network in the in-vehicle system. It processes and stores the user-generated interaction data. During actual prediction, prediction is performed based on the memory entries stored in the knowledge network combined with the MCP process of the application running in the vehicle, generating a corresponding execution plan, thereby improving the accuracy of judging user intentions and calling tools.

[0018] Furthermore, based on the above settings, the problem of low integration efficiency between application developers and OEM automakers can be solved. The current point-to-point integration model leads to a typical "N×M" integration problem. That is, if N AI applications or services need to be integrated into M different models or brands of vehicles, N×M custom developments are required. With the above settings, the model can automatically identify the corresponding tools integrated into the system and schedule the corresponding processes to generate the corresponding execution plans, reducing the number of custom developments. Attached Figure Description

[0019] Embodiments of the invention will be described more fully with reference to the accompanying drawings. However, the drawings are for illustration and explanation only and do not constitute a limitation on the scope of the invention.

[0020] Figure 1 This is a schematic diagram of the memory construction process according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the execution process in an embodiment of the present invention; Figure 3 This is a schematic diagram of step A1 in an embodiment of the present invention; Figure 4 This is a schematic diagram of step A2 in an embodiment of the present invention; Figure 5 This is a schematic diagram of step A21 in an embodiment of the present invention; Figure 6 This is a schematic diagram of step A22 in an embodiment of the present invention; Figure 7 This is a schematic diagram of step B1 in an embodiment of the present invention; Figure 8 This is a schematic diagram of step B2 in an embodiment of the present invention; Figure 9 This is a schematic diagram of step B3 in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0024] This invention includes: A data processing method for an intelligent vehicle agent includes a memory construction process and an execution process; like Figure 1 As shown, the memory construction process includes: Step A1: Interact with the user and obtain user interaction data, and deduce atomic knowledge units based on the user interaction data; Step A2: Link atomic knowledge units to form a knowledge network; like Figure 2 As shown, the execution process includes: Step B1: When raw input information containing user intent is received, the vehicle's MCP process is collected, and the current semantics are extracted from the raw input information; Step B2: Search the knowledge network based on the current semantics to obtain matching memory information; Step B3: Input the original input information, matching memory information, and MCP process into the inference engine to generate an execution plan; Step B4: Call the corresponding MCP process to execute the execution plan, and use the execution process as new user interaction data.

[0025] Specifically, in response to the problem of inaccurate judgment of user intent and tool invocation by in-vehicle intelligent agents in existing technologies, this embodiment constructs an intelligent agent and a corresponding knowledge network in the in-vehicle system, processes and stores the user-generated interaction data, and performs prediction based on the memory entries stored in the knowledge network combined with the MCP process of the application running in the vehicle to generate a corresponding execution plan, thereby improving the accuracy of judging user intent and invoking tools.

[0026] Specifically, the above data processing method is mainly configured as a software implementation in the vehicle system during practical application. The vehicle system includes a three-layer core architecture, including the AAOS host layer, the central agent service layer, and the federated MCP server network.

[0027] The first layer, the AAOS host layer, corresponds to the MCP host environment. Android Automotive OS, as the system's MCP host, is responsible for the following core duties: Lifecycle management: Start, stop, and monitor the central agent service and all MCP server processes; Service discovery and registration: During application installation / update, automatically parse the `AgentManifest.xml` file in the application package to identify whether it provides MCP service and register the tool capabilities to the system-level tool registry; Communication Management: Establishes an efficient communication channel between the MCP client and server using the Android Binder IPC mechanism; provides secure network connection management for cloud services; Security Context: Provides an isolated security sandbox for each running component, strictly controlling resource access permissions.

[0028] The second layer is the central agent service layer, which builds an AI agent based on a large language model (LLM). This agent is capable of executing the aforementioned data processing methods and handling all interactions based on the MCP protocol using the MCP client component, including: Tool discovery: Query available tools from the MCP server; Request serialization: Converts advanced tasks into tool call parameters; Response parsing: The processing tool returns results and converts them into a form that the Agent can understand; Error handling: Manage timeouts, exceptions, and other abnormal situations.

[0029] The third layer consists of a federated network of MCP servers, comprising numerous independent MCP servers, each exposing its unique functionality to the AI ​​Agent via the standard MCP protocol. The servers are divided into three categories: 1. Vehicle Abstraction Layer (VAL) Server Ownership: Developed by the OEM and possessing system signature-level permissions; Role: The sole secure gateway for accessing the underlying vehicle hardware; accomplish: Encapsulates the underlying Android Car API (such as CarPropertyManager, CarPowerManager, etc.); Expose a carefully curated, secure, read-only or controlled set of tools; High-risk operations (such as directly setting vehicle speed) are prohibited.

[0030] 2. OEM cloud server: Role: A bridge between the in-vehicle system and the OEM backend platform; Functionality: Provides functionality that requires a network connection and user account access.

[0031] 3. Third-party application MCP server Requirement: Any third-party application that exposes functionality to the Agent must implement and register an MCP server. mechanism: The application declares AgentManifest.xml within its package; When the system starts, AAOS parses the manifest file and registers the application's MCP server; Agents can invoke application functions through the unified MCP protocol.

[0032] Based on the above settings, after the vehicle infotainment system is powered on, each application will typically start its corresponding MCP process based on the corresponding MCP server, and declare the capabilities and input / output formats of the corresponding MCP process according to a specific format protocol.

[0033] At the same time, the vehicle system begins to continuously receive user input as user interaction data. Based on the user interaction data, it inputs corresponding content into the intelligent agent for reasoning. For example, if the user inputs a command multiple times near a fixed time point, it infers that the command is related to time, and selectively updates the knowledge network with the reasoned content as atomic knowledge units.

[0034] In the MCP host, the improved Zettelkasten (card box note-taking method) principle is adopted to store each memory (user preferences, travel destination, important facts, etc.) as an independent atomic knowledge unit. Then the system dynamically establishes links between atomic knowledge units to form a knowledge network.

[0035] When the original input information containing the user's intent is received, the vehicle's MCP process is collected to provide a basis for the agent to infer schedulable tools. The current semantics are extracted from the original input information, and the knowledge network is searched based on the current semantics to obtain matching memory information.

[0036] Using the MCP process and matching memory information as references, the LLM inference engine in the agent can perform ReAct loop thinking on the original input information, determine the current user intent, find available tools or extract evidence from memory information, loop this thinking process, thereby understanding the user intent and planning according to available tools, generating a directed acyclic graph of multi-step tasks as an execution plan, and then returning to the host system, where the MCP component sequentially distributes the task steps to each MCP process for execution, thereby handling the user's needs.

[0037] Similarly, this processing flow also extracts atomic knowledge units from user interaction data and updates them to the knowledge network. The aforementioned memory generation-update process and reasoning process can be performed simultaneously or in stages.

[0038] In one embodiment, such as Figure 3 As shown, step A1 includes: Step A11: Extract the original vehicle input data when the user interacts with the in-vehicle device, and perform text conversion based on the original vehicle input data to generate descriptive text; Step A12: Segment the description text to obtain multiple unit entities; Step A13: Perform logical deduction on the description text based on the unit entity to obtain atomic knowledge units.

[0039] Specifically, in order to achieve a better generation effect of atomic knowledge units, in this embodiment, the original vehicle input data is extracted when the user interacts with the in-vehicle device. The original vehicle input data includes relevant content input by the user through voice or commands, such as accessing applications, searching for content, inputting voice, etc., as well as vehicle-related status information collected by the domain controller, such as driving status, cabin temperature, time, weather, etc., as the background of the user's input behavior.

[0040] Considering that the user input is a relatively vague instruction, a corresponding text template is pre-prepared to provide a better description of the behavior. By filling in the corresponding parameter fields in the vehicle's relevant status into the text template and concatenating it with the user input, a textualized descriptive text is obtained.

[0041] Subsequently, the descriptive text was segmented using dictionary matching, and then the segmentation results were tagged with parts of speech to extract nouns and modifiers to obtain multiple unit entities.

[0042] Using the aforementioned unit entities as the main keywords for derivation, logical derivation is performed on the description text to obtain atomic knowledge units.

[0043] Each memory is represented as an atomic unit of knowledge, containing the following core fields: "noteId": "note_20250101_001", "noteType": "user_preference|location|habit|event|relationship", "content": "User preference: Working hours are 9:00-18:00, Monday to Friday", "timestamp": "2025-01-01T09:00:00Z", "confidence": 0.95, "source": "calendar_app|voice_input|derived|system", "tags": ["work_schedule", "routine", "monday-friday"], In one embodiment, such as Figure 4 As shown, step A2 includes: Step A21: Search for related nodes of atomic knowledge units relative to the knowledge graph to establish multi-dimensional links; Step A22: Add atomic knowledge units to the knowledge graph according to multi-dimensional links.

[0044] Specifically, in order to achieve better update results, in this embodiment, the atomic knowledge units are first searched for related nodes based on keywords relative to multiple pre-stored nodes in the knowledge graph. Related nodes with similar semantic content are obtained, and the relationship between the updated content and the related nodes is determined, such as whether it is a subset of the related nodes, mutually corroborates the related nodes, or conflicts with the related nodes, thereby forming multi-dimensional links.

[0045] Subsequently, based on this multi-dimensional link, the atomic knowledge units are compared with the knowledge graph to determine whether the knowledge graph needs to be updated, and then selectively modified.

[0046] In one embodiment, such as Figure 5 As shown, step A21 includes: Step A211: Match associated nodes for atomic knowledge units in the knowledge graph based on semantic features; Step A212: Derive the correlation between atomic knowledge units and associated nodes to obtain the association category; Step A213: Generate multi-dimensional links based on the associated categories.

[0047] Specifically, in order to achieve better generation results for multi-dimensional links, in this embodiment, atomic knowledge units are first matched with related nodes in the knowledge graph based on semantic features, thereby obtaining related nodes with similar semantic content.

[0048] Subsequently, the atomic knowledge units and associated nodes are input into the inference engine to deduce the correlation between the atomic knowledge units and associated nodes, and the correlation categories include: is_instance_of: general to specific relationship ("Going to work" → "Leaving for the Longyao Road office at 9:00 on Monday"); Contradictions: Memory conflict detection ("Users prefer to depart at 9:00 AM on Mondays" vs. "Actual data shows that departures are usually at 8:30 AM on Mondays"). related_to: indirect association ("User likes to listen to music" ↔ "Listens to classical music every day at work"); supports: Evidence supports the relationship (facts support the inference); Triggers: Causal triggering relationships ("Rainy day" → "Users tend to book ride-hailing services instead of walking"); temporal_sequence: time sequence relationship ("after get off work" → "usually go to the gym").

[0049] Based on the above classification categories, atomic knowledge units and related nodes are assembled to obtain multi-dimensional links.

[0050] In one embodiment, such as Figure 6 As shown, step A22 includes: Step A221: Determine the relative relationship between atomic knowledge units and associated nodes according to the association category; Step A222: Select the corresponding fusion rule based on the relative relationship, and then update the atomic knowledge unit to the knowledge graph based on the fusion rule.

[0051] Specifically, in order to achieve a better update process, in this embodiment, the relative relationship between atomic knowledge units and associated nodes is first determined according to the association category, and then the splicing process of atomic knowledge nodes of different categories is realized by determining the corresponding fusion rules.

[0052] For example, for subsets, tree-like node identifiers can be added to add new atomic knowledge units to the subsets of their parent nodes. For conflicting nodes, the content is adjusted using a time-weighted approach, with closer units receiving higher weights. Furthermore, lateral relationships, supporting evidence, causal relationships, and time-series relationships are typically labeled as parallel nodes, thus achieving better maintenance of the knowledge network.

[0053] In one embodiment, such as Figure 7 As shown, step B1 includes: Step B11: When the user's input intent is received, the vehicle status data is collected and used together with the user intent as the raw input; Step B12: Extract the currently running MCP process from the MCP host; Step B13: Extract semantic information from the original input.

[0054] Specifically, in order to achieve better extraction of user input, in this embodiment, when the vehicle system receives the user's input intent during operation, it collects vehicle status data, including driving status, vehicle speed, temperature, etc., and concatenates it with the user intent to serve as the original input.

[0055] Subsequently, the currently running MCP process is extracted from the MCP host to indicate the available tools to the inference engine. At the same time, semantic information is extracted from the original input through the semantic recognition model.

[0056] In one embodiment, such as Figure 8 As shown, step B2 includes: Step B21: Find matching related nodes based on the current semantic input knowledge network; Step B22: Extract knowledge slices based on associated nodes; Step B23: Assemble all the knowledge slices to obtain the matching memory information output.

[0057] Specifically, to achieve better extraction of matching memory information, in this embodiment, when the current semantics is extracted, it is first input into the knowledge network to search for matching related nodes. Simultaneously, the specific processing content within the related nodes is matched to check for memories of similar historical requests. Based on the search results, the knowledge network creates knowledge slices from the extracted related nodes and their corresponding links. Finally, all knowledge slices are assembled to output the matching memory information.

[0058] In one embodiment, such as Figure 9 As shown, step B3 includes: Step B31: Input the original input information and matching memory information into the inference engine to obtain the planned nodes; Step B32: The inference engine determines the matching process of the planning node to be executed based on the planning node matching MCP process; Step B33: Generate an execution plan based on the matching process and planning nodes.

[0059] Specifically, to achieve better planning results, in this embodiment, the original input information and matching memory information are first input into the inference engine. The inference engine, based on the ReAct mechanism, sequentially understands the user's intent combined with the matching memory information to infer the optimal strategy in the current context, thus forming planning nodes. Based on this, the inference engine matches the actual MCP process according to the planning nodes to determine the matching process that needs to be executed for the planning nodes, and finally generates an execution plan based on the matching process and the planning nodes.

[0060] Subsequently, the tool is executed and the results are obtained. These results are compared with the expected information in Zettelkasten, and new anomalous patterns or optimization opportunities are discovered, thereby updating the knowledge network in real time. In one embodiment, the method further includes the following steps before performing step B1: Step B01: When the MCP host calls the corresponding in-vehicle application, it creates the corresponding MCP process.

[0061] A data processing system for an intelligent vehicle's intelligent agent, used in the aforementioned data processing method.

[0062] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A data processing method for an intelligent vehicle's intelligent agent, characterized in that, This includes the memory construction process and the execution process; The memory construction process includes: Step A1: Interact with the user and obtain user interaction data, and deduce atomic knowledge units based on the user interaction data; Step A2: Link the atomic knowledge units to form a knowledge network; The execution process includes: Step B1: When raw input information containing user intent is received, the vehicle's MCP process is collected, and the current semantics are extracted from the raw input information; Step B2: Search the knowledge network based on the current semantics to obtain matching memory information; Step B3: Input the original input information, the matching memory information, and the MCP process into the inference engine to generate an execution plan; Step B4: Invoke the corresponding MCP process to execute the execution plan, and use the current execution process as the new user interaction data.

2. The data processing method according to claim 1, characterized in that, Step A1 includes: Step A11: Extract the original vehicle input data when the user interacts with the in-vehicle device, and perform text conversion based on the original vehicle input data to generate descriptive text; Step A12: Segment the description text to obtain multiple unit entities; Step A13: Perform logical deduction on the description text based on the unit entity to obtain the atomic knowledge unit.

3. The data processing method according to claim 1, characterized in that, Step A2 includes: Step A21: Search for associated nodes of the atomic knowledge unit relative to the knowledge graph to establish multi-dimensional links; Step A22: Add the atomic knowledge unit to the knowledge graph according to the multi-dimensional links.

4. The data processing method according to claim 3, characterized in that, Step A21 includes: Step A211: Match the atomic knowledge units with associated nodes in the knowledge graph based on semantic features; Step A212: Derive the correlation between the atomic knowledge unit and the associated node to obtain the association category; Step A213: Generate the multi-dimensional link based on the associated category.

5. The data processing method according to claim 4, characterized in that, Step A22 includes: Step A221: Determine the relative relationship between the atomic knowledge unit and the associated node according to the association category; Step A222: Select the corresponding fusion rule based on the relative relationship, and then update the atomic knowledge unit to the knowledge graph based on the fusion rule.

6. The data processing method according to claim 1, characterized in that, Step B1 includes: Step B11: When the user intent is received, the vehicle status data is collected and used together with the user intent as the raw input; Step B12: Extract the currently running MCP process from the MCP host; Step B13: Extract the semantic information from the original input.

7. The data processing method according to claim 4, characterized in that, Step B2 includes: Step B21: Based on the current semantic input, the knowledge network searches for matching associated nodes; Step B22: Extract knowledge slices based on the associated nodes; Step B23: Assemble all the knowledge slices to obtain the matching memory information output.

8. The data processing method according to claim 1, characterized in that, Step B3 includes: Step B31: Input the original input information and the matching memory information into the inference engine to obtain the planning node; Step B32: The inference engine matches the MCP process according to the planning node to determine the matching process that needs to be executed for the planning node; Step B33: Generate the execution plan according to the matching process and the planning node.

9. The data processing method according to claim 6, characterized in that, The procedure before performing step B1 also includes: Step B01: When the MCP host calls the corresponding in-vehicle application, it creates the corresponding MCP process.

10. A data processing system for an intelligent vehicle intelligent agent, characterized in that, Used to implement the data processing method as described in any one of claims 1-9.

Images