Intelligent tower crane-oriented natural language interaction AI agent system and method thereof

Abstract

The invention provides a natural language interaction AI agent system for an intelligent tower crane and a method thereof. Wherein the agent system comprises a natural language understanding module, a task planning module, a tool scheduling engine, a large language model interface, a data adaptation layer, a result generation module and a learning optimization module; the proxy method comprises the following steps: S1, natural language input processing; s2, intention understanding and task planning; s3, data acquisition and preprocessing; s4, intelligent analysis and reasoning; s5, verifying and optimizing a result; s6, generating a natural language answer; and S7, feedback learning and optimization are carried out. According to the method, natural language interaction between the user and the intelligent tower crane system is achieved by constructing a special AI agent architecture, the large language model is automatically called to intelligently analyze the running state of the tower crane, and analysis results and decision suggestions which are easy to understand are provided for the user.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, specifically to an AI agent system based on a large language model, used for intelligent analysis and decision support of smart tower cranes through natural language interaction. Background Technology

[0002] With the rapid development of artificial intelligence technology, AI agents, as intelligent systems capable of autonomously perceiving their environment, formulating plans, and executing actions, have been widely applied in various fields. However, in the field of intelligent industrial equipment, especially in the intelligent interaction of construction machinery, existing technologies have the following problems: 1. Limited and outdated interaction methods: The human-machine interaction of traditional industrial equipment mainly relies on professional operating interfaces, instrument panels and alarm systems. Users need to have professional knowledge to understand complex technical parameters, resulting in high learning costs and high operating thresholds.

[0003] 2. Lack of intelligent understanding capabilities: Existing construction machinery systems cannot understand users' natural language queries and cannot translate users' everyday language into specific equipment analysis tasks, which limits the ease of use and widespread adoption of the system.

[0004] 3. Fragmented analytical capabilities: Although industrial equipment generates a large amount of data, it lacks a unified intelligent analysis framework. Various analytical functions are independent of each other and cannot provide users with comprehensive decision support.

[0005] 4. Lack of adaptive learning mechanism: Most existing construction machinery systems are driven by static rules and cannot optimize themselves based on user feedback and usage habits, making it difficult to adapt to the personalized needs of different users.

[0006] Therefore, there is an urgent need for a natural language interactive AI agent system specifically designed for industrial equipment such as smart tower cranes, capable of understanding users' natural language queries, automatically calling large language models for intelligent analysis, and returning valuable decision-making information. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a natural language interactive AI agent system and method for smart tower cranes. By constructing a dedicated AI agent architecture, it enables natural language interaction between users and the smart tower crane system, automatically invokes a large language model to intelligently analyze the tower crane's operating status, and provides users with easily understandable analysis results and decision suggestions.

[0008] This invention first proposes a natural language interactive AI agent system for smart tower cranes, comprising: The natural language understanding module is used to receive users' natural language input, perform semantic parsing, and identify user intent. The task planning module is used to formulate corresponding task plans based on the identified user intent; The tool scheduling engine is used to automatically select and invoke the appropriate analysis tools according to the task plan; The large language model interface is used to call large language models for deep reasoning and analysis; The data adaptation layer is used to convert the raw data of the tower crane system into a format that can be understood by the large language model; The results generation module is used to convert the analysis results into natural language responses and return them to the user. The learning optimization module is used to continuously optimize system performance based on user feedback.

[0009] Furthermore, the natural language understanding module employs a pre-trained language model based on a deep learning model (Transformer) architecture. This pre-trained language model has the following characteristics: An integrated professional glossary for the field of building engineering, including tower crane terminology, component names, and operating specifications; Employing few-shot learning techniques, it can quickly adapt to the specific expressions used in the tower crane field; It supports multi-turn dialogue context understanding and can handle consecutive related queries.

[0010] Furthermore, the task planning module adopts a hierarchical task decomposition strategy, which includes the following characteristics: Intent Classification Layer: User queries are classified into four categories: status query, risk assessment, efficiency analysis, and knowledge consultation. Task decomposition layer: Decomposes complex queries into a sequence of executable atomic tasks related to tower crane sensor data, historical records, and environmental factors; Dependency Analysis Layer: Analyzes the dependencies between tasks, especially the dependencies between real-time data and historical trends, to determine the execution order; Prioritization layer: Prioritize tasks based on their importance and urgency.

[0011] Furthermore, the tool scheduling engine includes the following tool categories: Data query toolset: including sensor data query tools, historical data retrieval tools, and real-time status acquisition tools; Analysis and calculation toolset: including security assessment tools, efficiency analysis tools, and trend prediction tools; Knowledge retrieval toolset: including a standards and specifications search tool, an operation manual search tool, and a case study database search tool; Visualization generation toolset: including chart generation tools, report generation tools, and 3D visualization tools.

[0012] Furthermore, the large language model interface adopts a Model-as-a-Service (MAS) architecture, which has the following characteristics: It supports unified calling of various large language models, including the GPT series, Claude series, and domestic large models; Implement model load balancing and failover to ensure high service availability; Integrated model performance monitoring and cost control mechanisms; Supports hot-swapping and version management of models.

[0013] Furthermore, the data adaptation layer employs multimodal data fusion technology, which has the following characteristics: Structured data processing: converting structured data such as sensor data and control parameters into natural language descriptions; Time series data processing: converting time series data into trend descriptions and anomaly annotations; Image data processing: converting video surveillance data into scene descriptions and object recognition results; Contextual enhancement: Adding domain knowledge background and explanatory information to the data.

[0014] Furthermore, the result generation module employs a combination of template-based and generative methods, and has the following characteristics: Safety-critical information uses predefined templates to ensure accuracy and consistency; The analysis and interpretation of information are generated using a large language model, providing personalized expression methods; Confidence ratings provide a confidence score for each analysis result; Diverse output options support various output formats such as text, charts, and voice.

[0015] The present invention further proposes a natural language interaction AI agent method for smart tower cranes, which adopts the aforementioned natural language interaction AI agent system for smart tower cranes and includes the following steps: S1: Natural Language Input Processing Receive natural language query input from users, perform speech recognition (if it is voice input) and text preprocessing, and extract key information and context of the query; S2: Intent Understanding and Task Planning Based on the natural language understanding module, intent recognition and semantic parsing are performed to map user queries to specific analysis task types, and detailed task execution plans and tool call sequences are formulated. S3: Data Acquisition and Preprocessing According to the task requirements, call the corresponding data query tools to obtain relevant data of the tower crane system, and clean, format and enhance the context of the raw data. S4: Intelligent Analysis and Reasoning The preprocessed data is input into a large language model, and combined with domain knowledge, it is used for in-depth analysis and reasoning to generate preliminary analysis results and conclusions. S5: Result Validation and Optimization The analysis results are tested for reasonableness and compared with the preset safety rules and constraints. The analysis conclusions are then adjusted and optimized based on the verification results. S6: Natural Language Response Generation The analysis results are converted into natural language expressions that are easy for users to understand, with necessary explanations and suggestions added, and the final answer is generated and returned to the user. S7: Feedback Learning and Optimization Collect user feedback on the quality of responses, analyze system performance and user satisfaction, and update model parameters and optimization strategies.

[0016] Furthermore, in step S2, the intent recognition employs a multi-level classification strategy, as follows: Primary category: Identify the basic types of queries (inquiry, request, instruction, etc.); Secondary classification: Identify the specific domain of the query (security, efficiency, status, knowledge, etc.); Third-level classification: Identify the precise intent of the query (real-time status, historical trends, risk warnings, operational suggestions, etc.); In step S4, the large language model invocation employs the thought chain reasoning method, which includes the following characteristics: Problem decomposition: Breaking down a complex problem into multiple subproblems; Step-by-step reasoning: Analyze each sub-problem step by step in a logical order; Results Synthesis: The analysis results of each sub-problem are synthesized and summarized; Confidence assessment: Provides a credibility score for the final conclusion.

[0017] Furthermore, the method also includes a security mechanism, which has the following features: Input security check: Perform security checks on user input and filter out malicious queries; Permission verification: Verify the user's query permissions to ensure data access security; Output content review: Conduct security reviews of AI-generated answers to avoid misleading information; Operational boundary restrictions: Ensure that the AI ​​agent does not perform any device control operations, but only provides analysis suggestions.

[0018] The present invention has the following beneficial effects: 1. Significantly lower barrier to entry: Users do not need to learn complex professional interfaces and can obtain professional analysis results through natural language, reducing the barrier to entry by more than 80%.

[0019] 2. Improve interaction efficiency: Shift from the traditional "search-understand-analyze" model to the "ask-get answer" model, improving interaction efficiency by more than 5 times.

[0020] 3. Enable intelligent analysis: Through the powerful reasoning capabilities of large language models, complex correlation analysis and trend prediction can be performed, significantly improving the depth and accuracy of analysis.

[0021] 4. Supports personalized services: The system can learn users' query habits and preferences to provide personalized analysis results and suggestions.

[0022] 5. Excellent scalability: The modular architecture design allows the system to be easily expanded to other types of industrial equipment.

[0023] 6. Ensure safety and reliability: Through multiple security mechanisms, ensure that the system will not have a negative impact on equipment security. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is an architecture diagram of the natural language interaction AI agent system for smart tower cranes according to the present invention; Figure 2 This is an architecture diagram of the natural language understanding module in this invention; Figure 3 This is a flowchart of the task planning module in this invention; Figure 4 This is an organizational structure diagram of the tool scheduling engine in this invention; Figure 5 This is a flowchart illustrating the calling process of the large language model in this invention; Figure 6 This is a flowchart of the data adaptation layer processing in this invention; Figure 7 This is an example diagram of a security scenario in this invention; Figure 8 This is an example diagram of an efficiency analysis scenario in this invention; Figure 9 This is an example diagram of a risk warning scenario in this invention. Detailed Implementation

[0026] 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. Example 1

[0027] Reference Figure 1-9 A natural language interactive AI agent system for smart tower cranes includes: The natural language understanding module is used to receive users' natural language input, perform semantic parsing, and identify user intent. The task planning module is used to formulate corresponding task plans based on the identified user intent; The tool scheduling engine is used to automatically select and invoke the appropriate analysis tools according to the task plan; The large language model interface is used to call large language models for deep reasoning and analysis; The data adaptation layer is used to convert the raw data of the tower crane system into a format that can be understood by the large language model; The results generation module is used to convert the analysis results into natural language responses and return them to the user. The learning optimization module is used to continuously optimize system performance based on user feedback.

[0028] like Figure 1 As shown in the figure, the AI ​​agent system of this invention adopts a seven-layer architecture design, from top to bottom: Natural Language Understanding Module (including NLP processing, intent recognition, and semantic parsing), Task Planning Module (including task decomposition, dependency analysis, and priority ranking), Tool Scheduling Engine (including four tool groups: data query, analysis and calculation, knowledge retrieval, and visualization), Large Language Model Interface (supporting GPT, Claude, and domestic models), Data Adaptation Layer (processing structured data, time series, and image processing), Result Generation Module (template generation, generative output, credibility labeling, and diversified output), and Learning Optimization Module (user feedback collection, performance analysis, parameter optimization, and strategy update). External systems are marked with dashed boxes: User Interface Layer (mobile app, web interface, voice input), and Intelligent Tower Crane System (sensor system, camera system).

[0029] As a further implementation, the natural language understanding module employs a pre-trained language model based on a deep learning model (Transformer) architecture, and the pre-trained language model has the following characteristics: An integrated professional glossary for the field of building engineering, including tower crane terminology, component names, and operating specifications; Employing few-shot learning techniques, it can quickly adapt to the specific expressions used in the tower crane field; It supports multi-turn dialogue context understanding and can handle consecutive related queries.

[0030] like Figure 2 The text processing flow of the Natural Language Understanding module includes: input text → text preprocessing (word segmentation, cleaning) → domain vocabulary integration (construction terminology, crane terminology) → intent classification (using BERT / RoBERTa models) → slot filling (entity extraction) → contextual understanding (multi-turn dialogue support) → output structured intents and entities. This module integrates a specialized dictionary and a few-shot learning adapter (marked with dashed boxes).

[0031] As a further implementation scheme, the task planning module adopts a hierarchical task decomposition strategy specifically optimized for the safety and efficiency analysis of smart tower cranes. The hierarchical task decomposition strategy includes the following features: Intent Classification Layer: User queries are categorized into four main types: tower crane status query, risk assessment, efficiency analysis, and knowledge consultation; Task decomposition layer: Decomposes complex queries into a sequence of executable atomic tasks related to tower crane sensor data, historical records, and environmental factors; Dependency Analysis Layer: Analyzes the dependencies between tasks, especially the dependencies between real-time data and historical trends, to determine the execution order; Prioritization layer: Prioritize tower crane operations based on their importance and urgency to enable rapid risk warning and decision support.

[0032] The workflow diagram for the task planning module is shown below. Figure 3 .

[0033] As a further implementation plan, such as Figure 4 The tool scheduling engine includes tool categories specifically designed for smart tower crane data processing and analysis, including: Data query toolkit: includes a tower crane sensor data query tool, a historical data retrieval tool, and a real-time status acquisition tool; Analysis and calculation toolset: including tower crane safety assessment tools (such as tilt angle and load over-limit assessment), efficiency analysis tools (such as lifting times and time calculation), and trend prediction tools (such as wind speed and operation risk prediction). Knowledge retrieval toolset: including a standards and specifications search tool, a tower crane operation manual search tool, and a case study database search tool; Visualization generation toolset: including chart generation tools, report generation tools, and 3D visualization tools.

[0034] As a further implementation, the large language model interface adopts a Model-as-a-Service (MAS) architecture, which has the following characteristics: It supports unified calling of various large language models, including the GPT series, Claude series, and domestic large models; Implement model load balancing and failover to ensure high service availability; Integrated model performance monitoring and cost control mechanisms; Supports hot-swapping and version management of models.

[0035] like Figure 5 The large language model call process includes: user query input → prompt word engineering (template construction, context injection, domain knowledge integration) → model selection (GPT-4, Claude, load balancing) → API call management (request queuing, rate limiting, error handling) → response processing (result parsing, confidence scoring, security checks) → thought chain reasoning (problem decomposition, step-by-step analysis, result synthesis) → final output.

[0036] As a further implementation, the data adaptation layer employs multimodal data fusion technology, which has the following characteristics: Structured data processing: converting structured data such as sensor data and control parameters into natural language descriptions; Time series data processing: converting time series data into trend descriptions and anomaly annotations; Image data processing: converting video surveillance data into scene descriptions and object recognition results; Contextual enhancement: Adding domain knowledge background and explanatory information to the data.

[0037] like Figure 6 The data adaptation layer employs multimodal data fusion technology. The processing flow includes: raw data input (sensors, images, time series) → data type classification (structured, unstructured, temporal) → parallel processing path: structured data transformation, object detection scene description, anomaly detection trend analysis → narrative summary → context enhancement (domain knowledge injection, background information) → format standardization → LLM-ready output. The entire process includes a quality check mechanism.

[0038] As a further implementation, the result generation module adopts a combination of template-based and generative methods, and has the following characteristics: Safety-critical information uses predefined templates to ensure accuracy and consistency; The analysis and interpretation of information are generated using a large language model, providing personalized expression methods; Confidence ratings provide a confidence score for each analysis result; Diverse output options support various output formats such as text, charts, and voice.

[0039] like Figure 7-9 As shown, three typical interaction scenarios are illustrated: Safety query scenario: User asks "Is the tower crane safe now?" → System calls safety assessment tool → LLM analysis → Answer "The current tower crane status is safe, tilt angle of 0.2° is normal...".

[0040] Efficiency analysis scenario: User asks "How was the work efficiency today?" → System retrieves historical data → Efficiency calculation → Trend analysis → Answer "15 hoisting operations completed today, efficiency improved by 8%...".

[0041] Risk warning scenario: User asks "What are the risks?" → Environmental monitoring → Risk assessment → Predictive analysis → Answer "Main risk: Wind speed will increase in 1 hour, it is recommended to complete the operation in advance...".

[0042] Specific application instructions are as follows: The modules in this embodiment communicate through a standard API interface. The natural language understanding module is based on a pre-trained model and fine-tuned for the construction engineering field. This module can recognize typical queries such as: - "Is the tower crane safe now?" - "How efficient is today's work?" - "How much longer is expected to take to complete?" - "What risks should be noted?"

[0043] The task planning module employs a combination of finite state machine (FSM) and rule engine. For the query "Is the tower crane safe now?", the system will formulate the following task plan: 1. Obtain the current sensor status; 2. Check environmental conditions; 3. Analyze historical anomaly records; 4. Assess the overall safety level; 5. Generate a safety assessment report.

[0044] The tool scheduling engine maintains a tool registry, and each tool has a standard input / output interface. For example: Def get_sensor_status(): """Tool for obtaining sensor status""" return{ "tilt_angle":0.2, "wind_speed":2.1, "load_weight":8.5, "vibration_level":"normal" } The specific implementation of the natural language understanding module includes: Text cleaning: removing irrelevant characters and noise; word segmentation: using tools such as jieba for Chinese word segmentation; Entity recognition: Identifying technical terms related to tower cranes; Intent recognition stage: A multi-label classification model is used, which can identify multiple intents simultaneously. For example, for the query "What is the current status and risk of the tower crane?", the system can identify two intents: status query and risk assessment; Slot filling stage: Extract key information from the query, such as time range, equipment number, and target of interest.

[0045] The large language model uses a unified interface design: Class LLMInterface: Def init(self,model_type="gpt-4"): self.model_type = model_type Def analyze(self,prompt,context_data): """Analyze using a large language model""" #Build prompt words full_prompt = self.build_prompt(prompt,context_data) #Calling the model Response = self.call_model(full_prompt) #Analysis Results Return self.parse_response(response) Prompt word engineering is one of the key technologies. The system uses structured prompt word templates: You are a professional tower crane safety analysis expert. Please analyze the current safety status of the tower crane based on the following data: Sensor data: - Tilt angle: {tilt_angle} degrees - Wind speed: {wind_speed} m / s - Load weight: {load_weight} tons - Vibration level: {vibration_level} Please analyze from the following aspects: 1. Overall security assessment 2. Potential Risk Identification 3. Operational Recommendations Please answer in plain and easy-to-understand language, avoiding overly technical expressions.

[0046] Three typical interaction scenarios Scenario 1: Security Status Inquiry User typed: "Is the tower crane safe now?" System Processing Flow: 1. Intent Recognition: Safety Status Query; 2. Task Planning: Acquire Sensor Data → Safety Assessment → Report Generation; 3. Tool Invocation: Invoke Sensor Status Acquisition Tool; 4. Data Adaptation: Convert Numerical Data into Descriptive Text; 5. LLM Analysis: Perform Safety Assessment Based on Data; 6. Result Generation: "Current tower crane status is safe. Tilt angle of 0.2° is within the normal range, wind speed of 2.1 m / s is suitable for operation, load of 8.5 tons is within limits, and all indicators are normal." Scenario 2: Efficiency Analysis Query User input: "How was my homework efficiency today?" System Processing Flow: 1. Intent Recognition: Efficiency Analysis and Query; 2. Task Planning: Acquire Historical Data → Efficiency Calculation → Comparative Analysis; 3. Tool Invocation: Invoke Historical Data Retrieval Tool; 4. LLM Analysis: Perform Efficiency Comparison and Trend Analysis; 5. Result Generation: "Today, 15 hoisting operations were completed, with an average time of 12 minutes per operation, an 8% improvement over yesterday. The main reason is the lower wind speed, resulting in smoother operations." Scenario 3: Risk Warning Inquiry User input: "What risks should I be aware of?" System Processing Flow: 1. Intent Recognition: Risk Warning Query; 2. Task Planning: Environmental Monitoring → Equipment Inspection → Risk Assessment; 3. Tool Invocation: Invoking Multiple Monitoring Tools; 4. LLM Analysis: Comprehensive Risk Assessment; 5. Result Generation: "Current Main Risks: 1) Wind speed is expected to increase to 4 m / s in 1 hour; it is recommended to complete high-altitude operations in advance; 2) Thunderclouds are gathering in the southeast direction; weather changes need to be monitored." The system employs a multi-layered learning optimization strategy: User feedback learning: Collect user satisfaction ratings for answers, analyze users' subsequent query behavior, and adjust answer strategies based on feedback; Model performance optimization: Monitor model response time and accuracy, periodically evaluate the performance of different models, and automatically select the optimal model configuration; Knowledge base updates: Extract new knowledge points from user interactions, update the professional vocabulary and rule base, and continuously expand the case library.

[0047] Security measures are in place; input security checks are performed. Def check_input_safety(user_input): """Check the security of user input""" #Check if control commands are included control_keywords=["Start","Stop","Control","Operation"] If any keyword in user_input is used for keyword in control_keywords, return False, "Device control operations are not supported." #Check for sensitive information sensitive_keywords=["password","permissions","administrator"] if any(keyword in user_input for keyword in sensitive_keywords): `return False,"Sensitive information query not supported"` `return True,"Input security"` Access control mechanism: The system uses role-based access control (RBAC), and users with different roles have different query permissions. Operator: Basic status query; Safety officer: Safety assessment and risk analysis; Project Manager: Efficiency Analysis and Comprehensive Report; Maintenance personnel: Equipment health status and maintenance recommendations Output content review: Automatically review AI-generated answers to ensure they do not contain misleading security advice, do not disclose sensitive system information, and comply with industry norms and standards. Through the above technical solution, this invention successfully realizes the natural language interaction function of smart tower cranes, providing a new technical path for the intelligent interaction of industrial equipment.

[0048] Example 2

[0049] A natural language interaction AI agent method for smart tower cranes, employing the natural language interaction AI agent system for smart tower cranes described in Example 1, includes the following steps: S1: Natural Language Input Processing Receive natural language query input from users, perform speech recognition (if it is voice input) and text preprocessing, and extract key information and context of the query; S2: Intent Understanding and Task Planning Based on the natural language understanding module, intent recognition and semantic parsing are performed to map user queries to specific analysis task types, and detailed task execution plans and tool call sequences are formulated. S3: Data Acquisition and Preprocessing According to the task requirements, call the corresponding data query tools to obtain relevant data of the tower crane system, and clean, format and enhance the context of the raw data. S4: Intelligent Analysis and Reasoning The preprocessed data is input into a large language model, and combined with domain knowledge, it is used for in-depth analysis and reasoning to generate preliminary analysis results and conclusions. S5: Result Validation and Optimization The analysis results are tested for reasonableness and compared with the preset safety rules and constraints. The analysis conclusions are then adjusted and optimized based on the verification results. S6: Natural Language Response Generation The analysis results are converted into natural language expressions that are easy for users to understand, with necessary explanations and suggestions added, and the final answer is generated and returned to the user. S7: Feedback Learning and Optimization Collect user feedback on the quality of responses, analyze system performance and user satisfaction, and update model parameters and optimization strategies.

[0050] As a further implementation scheme, in step S2, the intent recognition employs a multi-level classification strategy, as follows: Primary category: Identify the basic types of queries (inquiry, request, instruction, etc.); Secondary classification: Identify the specific domain of the query (security, efficiency, status, knowledge, etc.); Third-level classification: Identify the precise intent of the query (real-time status, historical trends, risk warnings, operational suggestions, etc.).

[0051] As a further implementation, in step S4, the large language model invocation employs a thought chain reasoning method, which includes the following features: Problem decomposition: Breaking down a complex problem into multiple subproblems; Step-by-step reasoning: Analyze each sub-problem step by step in a logical order; Results Synthesis: The analysis results of each sub-problem are synthesized and summarized; Confidence assessment: Provides a credibility score for the final conclusion.

[0052] As a further implementation, the method also includes a security mechanism, which includes the following features: Input security check: Perform security checks on user input and filter out malicious queries; Permission verification: Verify the user's query permissions to ensure data access security; Output content review: Conduct security reviews of AI-generated answers to avoid misleading information; Operational boundary restrictions: Ensure that the AI ​​agent does not perform any device control operations, but only provides analysis suggestions.

[0053] The present invention has the following advantages: 1. Significantly lower barrier to entry: Users do not need to learn complex professional interfaces and can obtain professional analysis results through natural language, reducing the barrier to entry by more than 80%.

[0054] 2. Improve interaction efficiency: Shift from the traditional "search-understand-analyze" model to the "ask-get answer" model, improving interaction efficiency by more than 5 times.

[0055] 3. Enable intelligent analysis: Through the powerful reasoning capabilities of large language models, complex correlation analysis and trend prediction can be performed, significantly improving the depth and accuracy of analysis.

[0056] 4. Supports personalized services: The system can learn users' query habits and preferences to provide personalized analysis results and suggestions.

[0057] 5. Excellent scalability: The modular architecture design allows the system to be easily expanded to other types of industrial equipment.

[0058] 6. Ensure safety and reliability: Through multiple security mechanisms, ensure that the system will not negatively impact equipment security.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A natural language interactive AI agent system for smart tower cranes, characterized in that: include: The natural language understanding module is used to receive users' natural language input, perform semantic parsing, and identify user intent. The task planning module is used to formulate corresponding task plans based on the identified user intent; The tool scheduling engine is used to automatically select and invoke the appropriate analysis tools according to the task plan; The large language model interface is used to call large language models for deep reasoning and analysis; The data adaptation layer is used to convert the raw data of the tower crane system into a format that can be understood by the large language model; The results generation module is used to convert the analysis results into natural language responses and return them to the user. The learning optimization module is used to continuously optimize system performance based on user feedback.

2. The natural language interaction AI agent system for smart tower cranes as described in claim 1, characterized in that: The natural language understanding module employs a pre-trained language model based on a deep learning model architecture. The pre-trained language model has the following characteristics: A professional terminology database for the field of integrated building engineering; Employing few-shot learning techniques, it can quickly adapt to the specific expressions used in the tower crane field; It supports multi-turn dialogue context understanding and can handle consecutive related queries.

3. The natural language interaction AI agent system for smart tower cranes as described in claim 1, characterized in that: The task planning module adopts a hierarchical task decomposition strategy, which includes: Intent Classification Layer: User queries are classified into four categories: status query, risk assessment, efficiency analysis, and knowledge consultation. Task decomposition layer: Decomposes complex queries into a sequence of executable atomic tasks; Dependency analysis layer: Analyzes the dependencies between tasks and determines the execution order; Prioritization layer: Prioritize tasks based on their importance and urgency.

4. The natural language interaction AI agent system for smart tower cranes as described in claim 1, characterized in that: The tool scheduling engine includes the following tool categories: Data query toolset: including sensor data query tools, historical data retrieval tools, and real-time status acquisition tools; Analysis and calculation toolset: including security assessment tools, efficiency analysis tools, and trend prediction tools; Knowledge retrieval toolset: including a standards and specifications search tool, an operation manual search tool, and a case study database search tool; Visualization generation toolset: including chart generation tools, report generation tools, and 3D visualization tools.

5. The natural language interaction AI agent system for smart tower cranes as described in claim 1, characterized in that: The large language model interface adopts a model-as-a-service architecture, which has the following characteristics: Supports unified invocation of multiple large language models; Implement model load balancing and failover to ensure high service availability; Integrated model performance monitoring and cost control mechanisms; Supports hot-swapping and version management of models.

6. The natural language interaction AI agent system for smart tower cranes as described in claim 1, characterized in that: The data adaptation layer employs multimodal data fusion technology, which has the following characteristics: Structured Data processing: Converting structured data such as sensor data and control parameters into natural language descriptions; Time series data processing: converting time series data into trend descriptions and anomaly annotations; Image data processing: converting video surveillance data into scene descriptions and object recognition results; Contextual enhancement: Adding domain knowledge background and explanatory information to the data.

7. The natural language interaction AI agent system for smart tower cranes as described in claim 1, characterized in that: The result generation module employs a combination of template-based and generative methods, and has the following characteristics: Safety-critical information uses predefined templates to ensure accuracy and consistency; The analysis and interpretation of information are generated using a large language model, providing personalized expression methods; Confidence ratings provide a confidence score for each analysis result; Diverse output options support various output formats such as text, charts, and voice.

8. A natural language interaction AI agent method for smart tower cranes, characterized in that: The natural language interaction AI agent system for smart tower cranes, as described in any one of claims 1-7, includes the following steps: S1: Natural Language Input Processing It receives natural language query input from users, performs speech recognition and text preprocessing, and extracts key information and context of the query. S2: Intent Understanding and Task Planning Based on the natural language understanding module, intent recognition and semantic parsing are performed to map user queries to specific analysis task types, and detailed task execution plans and tool call sequences are formulated. S3: Data Acquisition and Preprocessing According to the task requirements, call the corresponding data query tools to obtain relevant data of the tower crane system, and clean, format and enhance the context of the raw data. S4: Intelligent Analysis and Reasoning The preprocessed data is input into a large language model, and combined with domain knowledge, it is used for in-depth analysis and reasoning to generate preliminary analysis results and conclusions. S5: Result Validation and Optimization The analysis results are tested for reasonableness and compared with the preset safety rules and constraints. The analysis conclusions are then adjusted and optimized based on the verification results. S6: Natural Language Response Generation The analysis results are converted into natural language expressions that are easy for users to understand, with necessary explanations and suggestions added, and the final answer is generated and returned to the user. S7: Feedback Learning and Optimization Collect user feedback on the quality of responses, analyze system performance and user satisfaction, and update model parameters and optimization strategies.

9. The natural language interaction AI agent method for smart tower cranes as described in claim 1, characterized in that: In step S2, the intent recognition employs a multi-level classification strategy, as follows: Primary category: Identifies the basic type of query; Secondary classification: Identifying the specific domain of the query; Three-level classification: Identifying the precise intent of the query; In step S4, the large language model invocation employs the thought chain reasoning method, which includes the following characteristics: Problem decomposition: Breaking down a complex problem into multiple subproblems; Step-by-step reasoning: Analyze each sub-problem step by step in a logical order; Results Synthesis: The analysis results of each sub-problem are synthesized and summarized; Confidence assessment: Provides a credibility score for the final conclusion.

10. The natural language interaction AI agent method for smart tower cranes as described in claim 1, characterized in that: The method also includes a security mechanism, which has the following features: Input security check: Perform security checks on user input and filter out malicious queries; Permission verification: Verify the user's query permissions to ensure data access security; Output content review: Conduct security reviews of AI-generated answers to avoid misleading information; Operational boundary restrictions: Ensure that the AI ​​agent does not perform any device control operations, but only provides analysis suggestions.

Images