An AI-based smart park response method and an AI agent system

Abstract

The application aims to provide an AI-based smart park response method and an AI agent system, wherein the response method comprises: S10: obtaining the input of a current user and matching it with a preset high-frequency question library; if the input is not matched, step S20 is executed; otherwise, step S30 is executed according to the matched business classification and key parameters; S20: according to the input of the current user, context information associated with the current input is retrieved from a preset memory module and input into an AI model for intent recognition to obtain corresponding business classification and key parameters; then step S30 is executed; S30: according to the corresponding business classification and key parameters, a skill plug-in library is called to control the AI model to complete the execution of the business corresponding to the user input and output the execution as a response result to the user; the input of the user and the response result are written into the memory module. Through high-frequency question combination rule matching, the progressive disclosure of the skill is skipped, the Token is saved, the interaction times are reduced, the response speed is effectively improved, and the system loss is reduced.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) platform technology, and in particular to an AI-based smart park response method and AI intelligent agent system. Background Technology

[0002] With the continuous expansion of smart park and smart building construction, the key to realizing intelligent operation and precise management of people, things, and events through IoT platforms has become the core value of smart parks. By building a unified IoT management platform, not only can remote monitoring and intelligent early warning of equipment and facilities be achieved, but data barriers between various business systems can also be broken down, forming a closed-loop optimization of park operation and management, thereby comprehensively improving the park's operational efficiency, service quality, and management level.

[0003] However, in the operation and maintenance management of traditional IoT platforms, whether it's a webpage on a computer or a mini-program on a mobile device, the functional modules and interface layout are usually pre-set by the developer, and users cannot flexibly adjust them according to actual needs. Although the rise of low-code platforms in recent years has alleviated this rigidity to some extent, it still requires professionals to configure functions and design interfaces through the backend before it can be finally delivered to end users. This model still has limitations in responding to the diverse and dynamic management needs of industrial parks, including slow response times and insufficient flexibility.

[0004] With the rapid development of artificial intelligence technology, AI models possessing autonomous understanding, reasoning, and interaction capabilities have become a new path to solve the problems of poor flexibility and limited interaction methods in traditional systems. Existing smart park systems combine traditional IoT architectures (including the perception layer, network layer, platform layer, and application layer) with gateway-channel architectures from OpenClaw intelligent agent frameworks. This supports multi-terminal instant messaging (such as WeChat / DingTalk / Lark), access, and semantic recognition of user input. A trained AI model then matches user input with business categories, providing corresponding skill call solutions. Subsequently, the park's dedicated skill system executes the corresponding instructions and returns the results to the user through an interactive interface, thus completing the user's response. However, most of these response processes require multiple queries from the user and multiple interactions with the AI ​​model (LLM) to obtain an accurate answer. Summary of the Invention

[0005] Based on this, the purpose of this invention is to provide an AI-based smart park response method and AI agent. This method combines historical user behavior dynamic data from the park to generate high-frequency questions that correlate different services with user behavior. Through rule matching, it enables the AI ​​agent to record processing prompts for high-frequency questions during user interactions, skipping the gradual disclosure of skills. This saves tokens, reduces the number of interactions, and effectively improves response speed and reduces system overhead. This invention achieves the above effects through the following technical means.

[0006] On the one hand, the present invention provides an AI-based smart park response method, which includes: S10: Obtain the current user's input and match it with a preset high-frequency question database. If there is no match, proceed to step S20; otherwise, proceed to step S30 based on the matched business category and key parameters. Here, the high-frequency question database is generated by obtaining dynamic data of all user behaviors within a certain time period and combining it with a time decay model. S20: Retrieve the context information associated with the current input from a preset memory module and input it into the AI ​​model for intent recognition to obtain the corresponding business category and key parameters; then execute step S30; wherein, the memory module stores the business usage data of each user and the context data of each user's session; S30: Based on the corresponding business category and key parameters, call the first skill plugin library, control the AI ​​model through the skill plugin library to complete the business corresponding to the user input, and output it as the response result to the user; among them, the skill plugin library encapsulates multiple SKILLs to complete the corresponding business scenarios of the smart park; S40: Write the user's input and response results into the memory module.

[0007] This invention first introduces an AI agent to respond to user behavior in smart parks by optimizing and improving OpenClaw and Hermes Agent technologies. It uses dynamic user behavior data within a period to avoid intent bias from static historical behavior. Combined with a time decay model, it adjusts the impact of the number of user behaviors and the time difference of the same type of user behavior on high-frequency issues in smart parks. This allows for more accurate response matching to the diverse and dynamic management needs of smart park systems. Combined with rule matching before subsequent intent recognition, it can skip the gradual disclosure that requires users to ask questions and call different skills multiple times when high-frequency issues occur, reducing the token and interaction number consumed by the agent and improving the response speed.

[0008] Furthermore, the high-frequency question database is generated through the following steps: The dynamic data of all user behaviors within the specified time period are classified, and then the basic weight of each category of user behavior is calculated based on the frequency of user behavior in each category and the total frequency of user behavior. Based on the time difference between the occurrence time and the current time of each category of user behavior and the basic weight, an exponential decay model is used to calculate the popularity score of each category of user behavior. The popularity scores are sorted, and high-frequency questions are extracted based on the popularity scores to generate a high-frequency question library.

[0009] Furthermore, the heat score is calculated using the following formula:

[0010] Where n is the frequency of user behavior of this category being triggered within the period; i is the temporal order of user behavior triggering; is the attenuation coefficient, is a constant greater than 0; w is the base weight of user behavior in this category.

[0011] Further, step S40 includes: S401: Determine whether to end the dialogue based on the current user's input. If yes, proceed to step S402. Otherwise, cache and write the complete user input and response results into the memory module. S402: Map the user input and response results to a set of corresponding triples according to a preset memory retrieval function; S403: Construct a long-term memory graph based on the set of triples for the user service usage data in the memory module, and complete the cross-session writing of the memory module; wherein the long-term memory graph structure includes three node types: TASK, SKILL, and EVENT, and four node relationships: PERFORM, INVOKE, SOLVER_BY, and AT_TIME.

[0012] Furthermore, the rule matching in step S10 is performed using regular expression rules.

[0013] Further, step S20 includes: S201.A: Based on the context information of the current user in the memory module, perform individual behavior modeling to obtain the personalized recommendation score for each service of the user; S201.B: Calculate the basic recommendation score for each service under the user's profile based on the user's profile; S202: Calculate the dynamic weight score for each service based on the personalized recommendation score and the basic recommendation score; S203: Input the current user's input and the dynamic weight score into the AI ​​model for intent recognition and recommendation to obtain the corresponding business category and key parameters.

[0014] Furthermore, the dynamic weight score is calculated using the following formula:

[0015]

[0016]

[0017] in, This is a dynamic adjustment coefficient, negatively correlated with user activity or registration duration t. Based on the recommended score, For personalized recommendation scores, n represents the total number of service categories, and k represents the user's k-th interaction in history. Let be the intensity of the k-th action. For target business i and user history usage business The similarity.

[0018] On the other hand, the present invention provides an AI agent for smart park response, comprising: a high-frequency question library: for acquiring dynamic data of user behavior within a certain time period, generating high-frequency questions based on the dynamic data of user behavior and combining it with a time decay model, and storing them; The fast matching unit is used to obtain the current user's input and match it with the high-frequency question database. If no match is found, the business parameter retrieval unit is called; otherwise, the SKILL call unit is called based on the matched business category and key parameters. Business parameter acquisition unit: used to retrieve context information associated with the current input from a preset memory module based on the current user's input, and input it into the AI ​​model for intent recognition to obtain the corresponding business category and key parameters; Memory module: Used to store each user's business usage data and the context data of each user's session. SKILL Invocation Unit: Used to invoke the skill plugin library according to the corresponding business category and key parameters, and return the response result; Skill plugin library: used to encapsulate multiple skills to complete the corresponding business scenarios of smart parks; Memory module update unit: used to write the user's input and response results into the memory module.

[0019] Furthermore, the high-frequency problem library unit includes: Category weight calculation subunit: Classifies the dynamic data of all user behaviors within the specified time period, and then calculates the basic weight of each category of user behavior based on the frequency of user behavior in each category and the total frequency of user behavior. Behavior popularity score calculation subunit: Based on the time difference between the occurrence time of each category of user behavior and the current time and the basic weight, the popularity score of each category of user behavior is calculated by the time decay model. High-frequency question generation subunit: Sort the popularity scores, extract high-frequency questions based on the popularity scores, and generate a high-frequency question library.

[0020] Furthermore, the business parameter acquisition unit includes: Personalized recommendation score calculation subunit: used to model individual behavior based on the context information of the current user in the memory module, and obtain the personalized recommendation score for each service of the user; Basic recommendation score calculation subunit: used to calculate the basic recommendation score for each service under the user's corresponding profile; Dynamic weight calculation subunit: used to calculate the dynamic weight score for each service based on the personalized recommendation score and the basic recommendation score; AI Intent Recognition Model Call Subunit: This unit is used to input the current user's input and the dynamic weights into the AI ​​model for intent recognition and recommendation, thereby obtaining the corresponding business category and key parameters.

[0021] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0022] Figure 1 This is a structural block diagram of an AI intelligent agent system for smart park response; Figure 2 To execute Figure 1 The flowchart of the AI ​​intelligent agent system shown is shown. Figure 3 The flowchart for generating the high-frequency problem library is shown below. Figure 4 A structural block diagram of the business parameter acquisition unit; Figure 5 To execute Figure 4 The flowchart for the business parameter acquisition unit is shown below; Figure 6 This is a flowchart of the update process for the hybrid memory module in a preferred embodiment.

[0023] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings of the embodiments of this application. Detailed Implementation

[0024] Besides the aforementioned requirement for multiple interactions between the AI ​​model and the user to provide an accurate response, the way users ask questions and the content of those questions can also affect the AI ​​model's response. If the question content deviates from the expected outcome, the resulting response will also be different. Therefore, while existing AI-based smart parks offer enhanced flexibility and scalability, their accuracy and speed of response still need improvement.

[0025] Therefore, based on the aforementioned problems, this invention generates AI agent systems by generating data based on users' historical interactions and preferences, and by adding a memory module to store users' interaction history. This system can remember each user's operating habits and historical queries. Furthermore, addressing the rigidity of traditional context management and the problem of traditional systems simply discarding old content when dialogues become too long, this invention persistently stores old dialogues, generates summaries for retrieval, and combines this with a time decay model to construct a high-frequency question library. This allows user input to be combined with high-frequency questions for fast and accurate responses, avoiding situations where the initial query direction is correct, but the response shifts with the increase in the number of interactions.

[0026] Please see Figure 1 The AI ​​agent used in the smart park includes: a preset high-frequency question library, a memory module, a skill plugin library, a fast matching unit 10, a business parameter acquisition unit 20, a SKILL call unit 30, and a memory module update unit 40.

[0027] Please see Figure 2 The specific working process of this AI intelligent agent system is as follows: The fast matching unit 10 is used to execute S10: obtain the current user's input and match it with a preset high-frequency question library. If there is no match, the business parameter acquisition module is called. Otherwise, the SKILL call unit is called according to the matched business category and key parameters. The high-frequency question library is generated by obtaining dynamic data of all user behaviors within a certain time period and combining it with a time decay model.

[0028] Frequently Asked Questions (FAQs) Caching. The AI ​​agent performs statistical analysis on dynamic user behavior data (equivalent to user questions) to identify high-frequency, patterned queries (such as "number of visitors to the park this week"). For such questions, the gateway session system directly matches and returns pre-cached processing results or simplified prompts, skipping subsequent intent recognition and skill-based gradual disclosure processes, thereby achieving millisecond-level response and significantly reducing the token consumption generated by large model calls.

[0029] Please see Figure 3 The high-frequency question database is generated through the following steps: The category weight calculation subunit is used to perform the following steps: classify the dynamic data of all user behaviors within the specified time period, and then calculate the basic weight of each category of user behavior based on the frequency of user behavior in each category and the total frequency of user behavior.

[0030] The behavior popularity score calculation subunit performs the following steps: Based on the time difference between the occurrence time and the current time of each category of user behavior and the basic weight, a popularity score for each category of user behavior is calculated using an exponential decay model. The popularity score is calculated using the following formula:

[0031] Where n is the frequency of user behavior of this category being triggered within the period; i is the temporal order of user behavior triggering; λ is the decay coefficient, a constant greater than 0, which determines the rate of "forgetting". The larger λ is, the faster old data becomes invalid (the system reacts more sensitively, but is easily affected by random factors). The smaller λ is, the longer historical data is retained (the recommendations are more stable, but the response to new demands is slower); w is the basic weight of user behavior for this category. (Time Delta): The time difference (usually in days or hours) between the time of the i-th action and the current time. Each category of user action here refers to the final command invoked by the user after a dialogue, such as "Query the number of visitors to the park this week" or "Filter vehicle information with a parking duration of more than 3 days".

[0032] The final popularity score for each business intent (such as "querying energy consumption") is calculated using the popularity score calculation formula.

[0033] The high-frequency question generation subunit is used to perform the following steps: sorting the popularity scores, extracting high-frequency questions based on the popularity scores, and generating a high-frequency question library.

[0034] Finally, the popularity scores are sorted, and the top N business intentions are selected as high-frequency questions to form a high-frequency question library. These high-frequency questions are then recommended on the user's page.

[0035] This invention uses a gateway session system as the unified message entry point for AI agents. It borrows from and optimizes the gateway-channel architecture of the OpenClaw agent framework. OpenClaw is an open-source, local-first AI agent runtime whose core architecture adopts a star topology of "single gateway + multiple channels": the Gateway, acting as the control plane, is a long-running Node.js process that manages all inbound and outbound communication; the Channels are responsible for abstracting the protocol differences of various messaging platforms (WhatsApp, Telegram, Slack, Discord, Signal, etc.) and normalizing input messages from different platforms into a unified internal message format for Agent processing. Simultaneously, this architecture also supports CLI, macOS / iOS / Android nodes coordinating with the Gateway via RPC to achieve multi-device collaboration.

[0036] First, rule matching is performed, especially for high-frequency, patterned user questions. Regular expression rules are configured to directly match the intent, skipping intent recognition and reasoning, thus achieving a rapid response. For example: Regular expression ^ query. Electricity cost $ → Intent "Energy consumption data query", generates key parameters for calling the Energy Consumption Data Query SKILL.

[0037] Regular expression ^(on|off). Air Conditioner → Intent "Device Control", generates key parameters for calling the Device Control SKILL.

[0038] Generating regular expressions. The report $→ intent "Generate Report" generates the key parameters for calling the Generate Report SKILL.

[0039] If the rule matching fails, proceed to business parameter acquisition unit 20.

[0040] The business parameter acquisition unit 20 is used to perform step S20: based on the input of the current user, retrieve the context information associated with the current input from a preset memory module, and input it into the AI ​​model for intent recognition to obtain the corresponding business classification and key parameters.

[0041] The business parameter acquisition unit 20 performs intent recognition by inputting data into a domain-specific intent classifier (AI large model). Its core responsibility is to accurately determine whether the user's question falls within the scope of park management and map it to specific business skills. For domain-specific questions, it further identifies the intent type, such as: query (device status / data), control (on / off / adjustment), analysis (energy consumption trend / alarm root cause), and report (generating daily / weekly reports). However, during the interaction between user input and the AI ​​agent, it is not always possible to accurately match the internal configuration of the AI ​​agent. Therefore, in a preferred embodiment, the business parameter acquisition unit 20 adopts a hybrid recommendation strategy, dynamically fusing "role-based group profiles" and "individual-based historical behavior sequences" to generate a final recommendation list, thereby improving the quality and efficiency of the interaction. Please refer to [link to relevant documentation]. Figure 4 and Figure 5 At this time, the preferred components of the business parameter acquisition unit 20 include: The personalized recommendation score calculation subunit 201 is used to execute step S201.A: to perform individual behavior modeling based on the context information corresponding to the current user in the memory module, and to obtain the personalized recommendation score corresponding to each service of the user.

[0042] This invention pre-constructs standard skill heatmaps for different functional roles in a smart park. Utilizing user-based collaborative filtering, user groups with the same role tags (e.g., "Security Supervisor," "Facility Maintenance") are considered as similar neighbor sets. The global popularity of specific business skills within this role group is then calculated. As a basic recommendation score in cold start or weak interaction scenarios Specifically, it is calculated using the following formula:

[0043] This means counting the frequency with which a certain function is used within that group of characters.

[0044] The basic recommendation score calculation subunit 201.B is used to execute step S201.B: calculate the basic recommendation score corresponding to each service under the user's corresponding profile.

[0045] The second is individual behavior sequence modeling. The system tracks users' interaction logs in real time and constructs behavior sequences that include timestamps, operation objects, and results.

[0046] Short-term interest capture: Utilize the attention mechanism to analyze the user's most recent N interactions and identify the current contextual intent (e.g., if a user continuously queries a specific device, then recommend related control commands for that device).

[0047] Long-term preference mining: Combining time decay algorithms, the periodicity and stability of users' historical operation habits are calculated to generate personalized recommendation scores. Specifically, it is calculated using the following formula:

[0048] The summation formula is used to calculate the impact of the user's historical behavior on the current recommendation, where k is the user's kth interaction in history. The weight represents the intensity of the k-th action. A higher weight indicates that the action better reflects the true intent. The time decay factor, Using cosine similarity, the relationship between "target function i" and "user history operation function" is calculated. The similarity between them. For example, if a user frequently operates "Chiller A", and "Chiller B" is highly similar to A in attributes (same type, same location), then the Sim value is close to 1, and B's recommendation score will be higher.

[0049] The dynamic weight calculation subunit 202 is used to perform step S202: calculate the dynamic weight score of each service based on the personalized recommendation score and the basic recommendation score.

[0050] Finally, the dynamic weight score is calculated using the following formula:

[0051] The variable is a dynamic adjustment coefficient, negatively correlated with user activity or registration duration *t*. This means that new users primarily rely on role-based recommendations, gradually transitioning to recommendations based on historical behavior as usage time increases. Subsequently, the top N users are displayed, sorted from highest to lowest according to *S*. *N* is configurable.

[0052] S203: Based on the dynamic weight score and the current user's input, perform intent recognition and recommendation to obtain the corresponding business category and key parameters.

[0053] The system quickly interacts with the recommended content until the user's needs are fully met, and the corresponding business category and key parameters are obtained.

[0054] The SKILL invocation unit 30 is used to execute step S30: invoking the skill plugin library according to the corresponding business category and key parameters, and returning the response result.

[0055] The SKILL invocation unit 30 accurately routes the task to the corresponding skill plugin library in the park's dedicated Skill system based on the identified intent and key parameters (such as device ID and time range). The skill plugin library is used to encapsulate multiple Skills that complete the corresponding business scenarios of the smart park. Each Skill corresponds to a specific business scenario (such as smart energy, smart security, smart access, and device control) and follows a unified interface specification (input: intent parameters; output: structured results).

[0056] Skill example: QueryEnergySkill: Receives {building: "A", start_date, end_date}, calls the energy management API, and returns energy consumption data.

[0057] ControlACSkill: Receives {device_id, target_temp, mode}, calls the device control API, and returns the execution result.

[0058] AlarmAnalysisSkill: Receives {alarm_id}, calls the data platform API, analyzes the root cause of the alarm, and generates handling suggestions.

[0059] For complex skills, the system supports "progressive disclosure," which guides the user to supplement necessary parameters through multiple rounds of dialogue. However, for high-frequency questions identified by the fast matching unit 20, the guidance process can be skipped directly, and cached or default parameters can be used for quick execution, achieving "one-click access."

[0060] The memory module update unit 40 is used to perform step S40: writing the user's input and response results into the memory module.

[0061] The data is written into the memory module to provide a source of historical data for interaction and subsequent dynamic updates of frequently asked questions.

[0062] This invention first introduces an AI agent to respond to user behavior in a smart park by optimizing and improving OpenClaw and Hermes Agent technologies. It uses dynamic user behavior data within a period to avoid intent bias from static historical behavior. Combined with a time decay model, it adjusts the impact of the number of user behaviors and the time difference of the same type of user behavior on high-frequency issues in the smart park. This allows for more accurate response matching to the diverse and dynamic management needs of the smart park system. Combined with rule matching before subsequent intent recognition, it skips the gradual disclosure process requiring users to ask questions and call different SKIs multiple times when high-frequency issues occur, reducing the token and interaction count overhead of the agent and improving response speed. Furthermore, it improves the existing business parameter acquisition unit 20 by introducing contextual information for role profiling and individual behavior sequence modeling. It dynamically establishes user input and user interaction recommendations from user, behavior, and time dimensions to assist users in interacting with the agent and achieve rapid and accurate responses to non-high-frequency issues.

[0063] Furthermore, in the aforementioned AI agents, traditional smart park systems lack memory. With the increasing application of large language models, context bloat has become a key bottleneck restricting agent performance. Current AI agents use memory modules for context management, and the mainstream solutions fall into two categories: Legacy engines and Lossless-claw plugins. The former employs a brute-force truncation strategy, while the latter pursues lossless storage; however, both have significant drawbacks in practical applications. For example: One approach is force truncation, while the other aims for lossless storage; however, both have significant drawbacks in practical applications.

[0064] The Legacy default engine is a patient suffering from "amnesia" due to brute-force truncation. Core logic: When the number of dialogue tokens exceeds 92% of the model window, a brute-force compression mechanism is triggered.

[0065] Processing method: Only the most recent three rounds of dialogue are retained, and the rest of the content is given a summary generated by LLM. The original messages are permanently replaced.

[0066] Main drawbacks: loss of original data, forced agent to undergo periodic "brainwashing", resulting in decreased intelligence as the conversation goes on for longer.

[0067] The OpenClaw solution, a benchmark in the Lossless-claw community, offers lossless DAG storage, but it is computationally intensive. Core logic: Introduce a DAG (Directed Acyclic Graph) to permanently retain the original message in SQLite.

[0068] Hierarchical structure: A multi-level summarization mechanism is adopted, consisting of Depth0 (fragment summary) and Depth1 (complete summary).

[0069] Key drawbacks: Tokens continue to grow; backtracking costs are extremely high; experience cannot be reused across sessions.

[0070] To address the aforementioned issues, in another embodiment of this invention, the memory module updating unit 50 updates the memory module as follows: Figure 6 The memory module update unit shown in the figure is used to build and update the memory module. Step S40 specifically includes: S401: Determine whether to end the dialogue based on the current user's input. If yes, proceed to step S402. Otherwise, cache and write the complete user input and response results into the memory module. Short-term memory (in-session): Maintains the complete dialogue flow of the current session to ensure the continuity of real-time interactions. S402: Map the user input and response results to a set of corresponding triples according to a preset memory retrieval function; Long-term memory (cross-session): Instead of storing lengthy original dialogue text, the system periodically transforms the dialogue stream into a long-term memory graph, constructing a memory network through logical relationships. The graph is defined as G=(V,E). Three core node types are defined: TASK (task, e.g., "Query energy consumption of building A"), SKILL (skill, e.g., "QueryEnergySkill"), and EVENT (event, e.g., "Air conditioner temperature alarm"). The relationship edges between nodes are also strictly defined.

[0071] S403: Construct a long-term memory graph based on the set of triples for the user service usage data in the memory module, and complete the cross-session writing of the memory module; wherein the long-term memory graph structure includes three node types: TASK, SKILL, and EVENT, and four node relationships: PERFORM, INVOKE, SOLVER_BY, and AT_TIME.

[0072] A memory retrieval function f maps dialogue fragments d∈D to a set of triples T.

[0073] The set of relations Strictly defined as: r1=PERFORM: (User, PERFORM, TASK) r2=INVOKE: (TASK, INVOKE, SKILL) r3=SOLVED_BY: (EVENT, SOLVED_BY, SKILL) r4=AT_TIME: (TASK, AT_TIME, Timestamp) The system periodically converts old short-term memories into triples and updates long-term behavioral summaries at the user level (e.g., "User Manager Zhang has a habit of checking electricity bills on the morning of the 5th of each month"). When a new conversation begins, the system, based on the current user question and context, retrieves the most relevant historical memories from the triple network and summary library through vector similarity or logical rules, and injects them into the prompts for the current conversation.

[0074] Through the above steps, a structured and hierarchical hybrid memory module is gradually updated in the interaction between users and the smart park to solve the memory loss caused by traditional "brute force truncation" and the cumbersome retrieval problem caused by "lossless but inefficient".

[0075] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A smart park response method based on AI, characterized in that, include: S10: Obtain the current user's input and match it with a preset high-frequency question database. If there is no match, proceed to step S20; otherwise, proceed to step S30 based on the matched business category and key parameters. The high-frequency question database is generated by acquiring dynamic data of all user behaviors within a certain time period and combining it with a time decay model. S20: Based on the current user's input, retrieve the context information associated with the current input from a preset memory module, and input it into the AI ​​model for intent recognition to obtain the corresponding business category and key parameters; then execute step S30. The memory module stores the business usage data of each user and the context data of each user's session. S30: Based on the corresponding business category and key parameters, call the first skill plugin library, control the AI ​​model to complete the business execution corresponding to the user input through the skill plugin library, and output it as the response result to the user; This skill plugin library encapsulates multiple skills to complete business scenarios corresponding to smart parks; S40: Write the user's input and response results into the memory module.

2. The AI-based smart park response method according to claim 1, characterized in that, The high-frequency problem database is generated through the following steps: The dynamic data of all user behaviors within the specified time period are classified, and then the basic weight of each category of user behavior is calculated based on the frequency of user behavior in each category and the total frequency of user behavior. Based on the time difference between the occurrence time and the current time of each category of user behavior and the basic weight, an exponential decay model is used to calculate the popularity score of each category of user behavior. The popularity scores are sorted, and high-frequency questions are extracted based on the popularity scores to generate a high-frequency question library.

3. The AI-based smart park response method according to claim 2, characterized in that, The heat score is calculated using the following formula: Where n is the frequency of user behavior of this category being triggered within the period; i is the temporal order of user behavior triggering; is the attenuation coefficient, is a constant greater than 0; w is the base weight of user behavior in this category.

4. The AI-based smart park response method according to claim 3, characterized in that, Step S40 includes: S401: Determine whether to end the dialogue based on the current user's input. If yes, proceed to step S402. Otherwise, cache and write the complete user input and response results into the memory module. S402: Map the user input and response results to a set of corresponding triples according to a preset memory retrieval function; S403: Construct a long-term memory graph based on the set of triples for the user service usage data in the memory module, and complete the cross-session writing of the memory module; wherein the long-term memory graph structure includes three node types: TASK, SKILL, and EVENT, and four node relationships: PERFORM, INVOKE, SOLVER_BY, and AT_TIME.

5. The AI-based smart park response method according to claim 4, characterized in that, The rule matching in step S10 is performed using regular expression rules.

6. The AI-based smart park response method according to any one of claims 1-5, characterized in that, Step S20 includes: S201.A: Based on the context information of the current user in the memory module, perform individual behavior modeling to obtain the personalized recommendation score for each service of the user; S201.B: Calculate the basic recommendation score for each service under the user's profile based on the user's profile; S202: Calculate the dynamic weight score for each service based on the personalized recommendation score and the basic recommendation score; S203: Input the current user's input and the dynamic weight score into the AI ​​model for intent recognition and recommendation to obtain the corresponding business category and key parameters.

7. The AI-based smart park response method according to claim 6, characterized in that, The dynamic weight score is calculated using the following formula: in, This is a dynamic adjustment coefficient, negatively correlated with user activity or registration duration t. Based on the recommended score, For personalized recommendation scores, n represents the total number of service categories, and k represents the user's k-th interaction in history. Let be the intensity of the k-th action. For target business i and user history usage business The similarity.

8. An AI intelligent agent system for smart park response, characterized in that, include: The high-frequency question library is generated by acquiring dynamic data of all user behaviors within a certain time period and combining it with a time decay model. The memory module stores each user's business usage data and the context data of each user's session. The skill plugin library encapsulates multiple skills to complete corresponding business scenarios in smart parks; The fast matching unit is used to obtain the current user's input and match it with the high-frequency question database. If no match is found, the business parameter acquisition unit is called; otherwise, the SKILL call unit is called based on the matched business category and key parameters. Business parameter acquisition unit: used to retrieve context information associated with the current input from a preset memory module based on the current user's input, and input it into the AI ​​model for intent recognition to obtain the corresponding business classification and key parameters; SKILL Invocation Unit: Used to invoke the skill plugin library according to the corresponding business category and key parameters, control the AI ​​model to complete the business corresponding to the user input through the skill plugin library, and output it as the response result to the user; Memory module update unit: used to write the user's input and response results into the memory module.

9. The AI ​​agent according to claim 8, characterized in that, The high-frequency problem library is generated through the following units: Category weight calculation subunit: Classifies the dynamic data of all user behaviors within the specified time period, and then calculates the basic weight of each category of user behavior based on the frequency of user behavior in each category and the total frequency of user behavior. Behavioral popularity score calculation subunit: Based on the time difference between the occurrence time of each category of user behavior and the current time and the basic weight, the popularity score of each category of user behavior is calculated through an exponential decay model. High-frequency question generation subunit: Sort the popularity scores, extract high-frequency questions based on the popularity scores, and generate a high-frequency question library.

10. The AI ​​agent according to claim 9, characterized in that, The business parameter acquisition unit includes: Personalized recommendation score calculation subunit: used to model individual behavior based on the context information of the current user in the memory module, and obtain the personalized recommendation score for each service of the user; Basic recommendation score calculation subunit: used to calculate the basic recommendation score for each service under the user's corresponding profile; Dynamic weight calculation subunit: used to calculate the dynamic weight score for each service based on the personalized recommendation score and the basic recommendation score; AI Intent Recognition Model Call Subunit: This unit is used to input the current user's input and the dynamic weights into the AI ​​model for intent recognition and recommendation, thereby obtaining the corresponding business category and key parameters.

Images