A construction AI risk analysis and decision-making method based on agent security management

Abstract

This invention provides a construction AI-based risk analysis and decision-making method for safety management. Through a knowledge base construction mechanism involving intelligent parsing, vectorized storage, and incremental updates, it transforms scattered safety knowledge from projects into digital assets, enabling the continuous accumulation of enterprise safety knowledge assets. A RAG-enhanced large language model question-answering system equips each on-site personnel with a personal AI expert, providing near-instantaneous responses to on-site issues and improving efficiency. By employing a performance quantitative analysis model, an Attention-LSTM time-series risk prediction model, and a hazard clustering analysis algorithm, it uncovers management shortcomings and predicts risk trends from massive amounts of data, making safety management decisions more data-driven and moving away from experience-based, extensive management. AI time-series prediction identifies high-risk areas and hazard trends seven days in advance, generating actionable control recommendations, shifting safety management from post-event remediation to pre-event prevention, significantly reducing the probability of accidents.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology for construction safety, and in particular to a construction AI risk analysis and decision-making system and method, electronic device and computer-readable storage medium based on intelligent agent safety management. Background Technology

[0002] As a high-risk industry, the construction sector prioritizes safety management as the core of project management. Currently, five long-standing pain points in construction safety management severely hinder management efficiency and safety control: First, the accumulation of safety knowledge is difficult, leading to significant asset loss. Core knowledge such as project safety regulations, management experience, and accident handling plans are scattered across paper documents, electronic files, and the personal experience of managers. This knowledge is lost with project completion and staff turnover. New employee training is lengthy and slow, creating a vicious cycle of "project completion, knowledge wiped out," preventing the formation of reusable digital knowledge assets for the company. Second, slow on-site management response and high reliance on experts. Handling on-site safety issues, accessing regulations, and developing solutions heavily depend on project safety experts. However, expert resources are scarce and their time is limited. When encountering problems, frontline staff often face the dilemma of "inaccurate searches on Baidu, incomplete regulations, and difficulty finding experts," resulting in delayed responses and non-standard handling, which can easily lead to safety hazards. Third, weak data value mining and experience-driven decision-making. The project operation generates massive amounts of data on personnel performance, equipment operation, hazard identification, and risk control. However, most of this data remains dormant in the form of reports, relying solely on manual statistical analysis. This fails to uncover underlying patterns and trends in hazards and risks, leaving safety management stuck in a passive "post-event review and remedial rectification" mode, unable to achieve pre-event warnings and precise policy implementation. Fourth, a closed-loop management system is difficult to form, resulting in poor implementation. In traditional safety management models, the transmission of hazard warnings, rectification instructions, and management requirements relies on manual reporting and offline notifications. This leads to delayed information transmission, high omission rates, and an inability to achieve closed-loop control of "risk identification - instruction delivery - rectification execution - effect verification," resulting in a management vacuum of "requirements from above, but non-compliance from below." Fifth, management costs remain high, and capability coverage is uneven. To ensure effective safety management, companies need to assign dedicated safety experts to each project, resulting in high labor costs. Furthermore, the experts' capabilities cannot cover all times and all areas of the construction site. The safety management capabilities of grassroots personnel vary greatly, making them highly susceptible to accidents due to insufficient understanding and improper operation.

[0003] Most existing construction safety management information systems only implement basic functions such as data entry, process approval, and report statistics, without introducing AI technologies such as large language models and multi-agent architecture to achieve intelligent upgrades in management capabilities. The few systems that have introduced AI technology only implement single functions such as hazard identification and voice Q&A. The modules are isolated from each other and have not formed a closed-loop management system that covers the entire process from knowledge accumulation, interactive Q&A, analysis and decision-making to message push. They cannot adapt to the safety management needs of all scenarios in construction projects, let alone solve the aforementioned core pain points in the industry. Summary of the Invention

[0004] To address the technical problems existing in the prior art, the present invention provides the following technical solution: On the one hand, a construction AI risk analysis and decision-making system based on multi-agent collaboration is provided, including a basic data layer, a data knowledge layer, an AI capability layer, a multi-agent collaborative scheduling layer, and an application interaction layer, wherein: The basic data layer is used for the collection, cleaning and standardized storage of multi-source construction safety-related data, providing raw data support for the upper layer. It has a built-in data preprocessing module to complete the removal of outliers, filling of missing values ​​and standardization of raw data. The data knowledge layer is used to transform unstructured documents and structured data into reusable digital knowledge assets. It includes four core libraries: a standard library, a project system library, a construction plan library, and a risk case library, as well as a vector database and a real-time data lake. It also includes a document intelligent parsing unit, a text segmentation and vectorization unit, a knowledge association modeling unit, and an incremental training and update unit. The AI ​​capability layer provides intelligent algorithms and model support for the system, including a large language model reasoning unit, a RAG enhanced retrieval and re-ranking unit, a construction risk time-series prediction unit, a safety performance quantitative analysis unit, a knowledge graph construction engine, a speech recognition and semantic understanding unit, and an intelligent recommendation engine. The multi-agent collaborative scheduling layer is the central scheduling core of the system, including a knowledge base management agent, a security question-and-answer interaction agent, a performance analysis and decision-making agent, a message push agent, and an agent collaborative scheduling center. The agent collaborative scheduling center is used to realize the task scheduling, message communication and collaborative operation of the four agents, and complete the closed-loop management of the entire process from knowledge accumulation, interactive question and answer, analysis and decision-making to message push. The application interaction layer serves as the entry point for interaction between the system and users, including mobile apps, WeChat mini-programs, PC management terminals, WeChat robots, and on-site smart terminals.

[0005] On the other hand, a construction AI risk analysis and decision-making method based on multi-agent collaboration is provided, which is implemented based on the above system and includes the following steps: Initialization and construction of project-specific knowledge base: Create project space, upload project-related documents and industry standards, complete intelligent document parsing, text segmentation and vectorization processing through knowledge base management intelligent agent, build project-specific knowledge base and knowledge graph, and configure knowledge base incremental update rules; Multi-agent collaborative system and system parameter initialization: Configure user permissions, operating parameters of each agent, connect with the project's existing management system and IoT devices, and complete the adaptation of the system to the project management scenario; Real-time acquisition and preprocessing of multi-source construction safety data: Multi-source construction data is acquired in real time through the basic data layer, and preprocessing such as outlier removal, missing value filling and data normalization is completed. The standardized data is then written into the real-time data lake. RAG-enhanced intelligent question-and-answer interaction for security management: It receives text / voice questions from users, completes speech recognition, intent recognition, RAG-enhanced retrieval and reordering, and generates professional answers that match the knowledge base and real-time project data through a large language model; AI-based quantitative analysis of safe performance based on multi-dimensional data: The intelligent decision-making agent automatically calculates the performance completion rate and comprehensive score of the grid and personnel, completes the assignment of security codes, and identifies weak links in performance. Dynamic prediction of construction risks and intelligent decision generation: The construction risk time series prediction unit predicts the risk evolution trend in the next 7 days, performs cluster analysis on the hidden danger data, and generates practical AI decision suggestions and safety management reports by combining the knowledge base; Automated push and closed-loop control of management information: The daily safety briefing is pushed on a scheduled basis and real-time warning of emergency risks is given through a message push intelligent agent, and the execution status is tracked to form a closed-loop control. Continuous iteration of the knowledge base and incremental optimization of the model: Daily incremental updates to the knowledge base, and incremental fine-tuning of the model based on user feedback and new project data, to achieve continuous evolution of system capabilities.

[0006] On the other hand, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions, which, when executed by the processor, implement the method described above.

[0007] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement the above method.

[0008] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. A project-specific knowledge system that can be accumulated and reused has been built, solving the pain point of knowledge loss in traditional management: Through the knowledge base construction mechanism of intelligent parsing, vectorized storage and incremental updates, the scattered security knowledge of the project is transformed into digital assets. After the project is completed, it can be fully accumulated and reused. New employees can quickly learn project management requirements through AI Q&A, shortening the training cycle by more than 80%, and realizing the continuous accumulation of enterprise security knowledge assets.

[0009] 2. It has enabled the universal reuse of expert capabilities, solving the pain points of scarcity and high dependence on traditional management expert resources: Through the RAG-enhanced large language model question-and-answer system, each on-site personnel is equipped with a 24 / 7 personal AI expert, which responds to on-site issues in seconds and improves the efficiency of standardized queries and solution writing by more than 10 times, completely breaking the limitation of expert resources and realizing the sharing of expert capabilities by all employees.

[0010] 3. It has enabled in-depth value mining of security data, solving the pain points of traditional data being dormant and unusable: Through the performance quantitative analysis model, the Attention-LSTM time series risk prediction model, and the hidden danger clustering analysis algorithm, it has mined management shortcomings and predicted risk trends from massive amounts of data, realizing the upgrade from "manual statistical reports" to "AI deep analysis", making security management decisions based on evidence and bidding farewell to the experience-driven extensive management model.

[0011] 4. It has achieved pre-event risk warning and closed-loop management, solving the pain points of traditional management's delayed response and post-event remediation: by using AI time-series prediction, high-risk areas and potential hazard trends can be identified 7 days in advance, generating actionable management suggestions. At the same time, the message push intelligent agent enables real-time information delivery and closed-loop management, transforming safety management from post-event remediation to pre-event prevention, and significantly reducing the probability of accidents.

[0012] 5. A multi-agent collaborative closed-loop management system was constructed, solving the pain points of isolated and uncoordinated modules in traditional systems: through the collaborative scheduling of four intelligent agents—knowledge base management, question-and-answer interaction, analysis and decision-making, and message push—a fully automated closed-loop management system was achieved, from knowledge accumulation, interactive question-and-answer, analysis and decision-making to execution push. This significantly reduced the manual cost of safety management, improved management efficiency and standardization, and met the safety management needs of the entire construction industry. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0014] Figure 1 This is a schematic diagram of a system architecture provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a data processing flow for a basic data layer provided in an embodiment of the present invention; Figure 3 This is a diagram illustrating a mechanism for RAG-enhanced retrieval and reordering provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a three-layer communication architecture for cloud-edge-device collaboration provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the implementation process of a method provided in an embodiment of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0016] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0017] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0018] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] This invention proposes a construction AI risk analysis and decision-making system based on intelligent agent safety management. Through a multi-agent collaborative architecture, it achieves an intelligent closed loop throughout the entire construction safety management process, constructing a project-specific safety management AI think tank and thoroughly solving the core pain points of existing technologies. The construction AI risk analysis and decision-making method is implemented based on this system. The following will describe in detail the system architecture, application, and implementation principles of the method.

[0021] The application principles of this system and method will be described in detail below.

[0022] I. Construction AI Risk Analysis and Decision-Making System Based on Intelligent Agent Safety Management like Figure 1 As shown, this system adopts a five-layer distributed cloud-native architecture, consisting of a basic data layer, a data knowledge layer, an AI capability layer, a multi-agent collaborative scheduling layer, and an application interaction layer, from bottom to top. Each layer achieves data communication and command interaction through standardized interfaces. The entire chain employs encrypted transmission and hierarchical access control to meet the data security and compliance requirements of the construction industry. The system adopts a "cloud-edge-device" collaborative deployment model, supporting both public cloud deployment and local private deployment, adapting to the security and compliance requirements of different projects.

[0023] The system strictly adheres to the Level 3 requirements of the Cybersecurity Classified Protection System 2.0, constructing a full-link data security assurance system: 1. Project-level data isolation: The knowledge base and database of each project are deployed independently, and the role-based access control model of RBAC is adopted to achieve least privilege control and all operations are fully audited. 2. End-to-end encryption: Data transmission uses HTTPS / TLS 1.3 encryption, data storage uses AES-256 symmetric encryption, and sensitive information is anonymized to prevent data leakage; 3. Private Deployment Support: For classified projects, it supports full-system local private deployment, ensuring that data does not leave the project domain, thus meeting the compliance requirements of state-owned enterprises and government projects; 4. Compliance Audit: The system design fully complies with the Level 3 requirements of Cybersecurity Classified Protection 2.0, and regular security vulnerability scans and penetration tests are conducted to ensure the safe and stable operation of the system.

[0024] The architecture of this system will be described in detail below.

[0025] (I) Basic Data Layer The basic data layer serves as the system's data foundation, responsible for the collection, cleaning, standardized storage, and data lifecycle management of multi-source construction safety-related data, providing high-quality raw data support for upper-layer applications. This layer comprises six functional units, each achieving interface-based data synchronization with external systems through standardized interfaces: 1. Personnel Data Collection Unit: Collects basic information, attendance data, job performance completion data, safety training records, violation records, and personnel qualification information of project management personnel, construction teams, and special operation personnel. It connects to the project personnel management system and real-name attendance system through API interface to complete real-time data synchronization and standardized processing. 2. Equipment Data Acquisition Unit: Collects basic information, maintenance records, annual inspection information, and operating status parameters of special equipment and construction machinery; connects with IoT sensors to collect real-time data such as equipment operating current, vibration, and temperature; and constructs a health record for the entire life cycle of the equipment. 3. Hazard and Risk Data Collection Unit: Collects project hazard investigation records, rectification closed-loop data, risk classification and control list, high-risk operation approval records, and historical accident cases; connects to the on-site video monitoring system; and extracts on-site violations and hazard data through target detection models. 4. Safety Performance Data Collection Unit: Collects performance task lists, task completion records, inspection routes, and inspection results data from grid-based management to achieve full-process traceability of performance behavior; 5. IoT Sensing Data Acquisition Unit: Connects to IoT terminals such as on-site environmental sensors, high-risk operation monitoring equipment, and slope settlement monitoring equipment to collect on-site environmental parameters and hazard source monitoring data; 6. External Industry Data Integration Unit: Connects with the latest safety standards, policy documents, accident reports, and industry warnings issued by the state, industry, and local governments, providing authoritative external knowledge support for the system.

[0026] like Figure 2 As shown, the basic data layer has a built-in data preprocessing module that standardizes the collected raw data. The core processing logic includes: removing outliers using the 3σ principle, filling missing values ​​using linear interpolation, and mapping data of different dimensions to the [0,1] interval using a min-max normalization algorithm to eliminate dimensional differences. The normalization calculation formula is: in, For normalized standardized data, The original data, This is the minimum value for this data dimension. This represents the maximum value for this data dimension.

[0027] The preprocessed standardized data is uniformly written into the Delta Lake real-time data lake, and ACID transactions are used to ensure data consistency. The data is stored in partitions according to time, segment, and grid dimensions, providing real-time and high-quality data input for upper-layer AI analysis.

[0028] (ii) Data Knowledge Layer The data knowledge layer is the core knowledge carrier of the system, responsible for transforming unstructured documents and structured construction data into reusable, searchable, and reasonable digital knowledge assets. It is the core foundation for achieving accurate AI question answering and intelligent decision-making. This layer includes four core knowledge bases, a vector database, a real-time data lake, and supporting knowledge base construction and management modules.

[0029] (II.1). Four Core Knowledge Bases 1. Standards and Specifications Library: Stores national, industry, and local safety standards, standard drawings, and policy documents for fields such as building construction, port engineering, and municipal engineering, for compliance checks, standard lookup, and construction guidance; 2. Project Policy Library: Stores the current project's safety management policies, emergency plans, safety production responsibility system, and grid management scheme, for the dissemination, implementation, and verification of project management requirements; 3. Construction Plan Library: Stores project construction organization designs, special construction plans for each sub-item of the project, and safety technical briefing documents, for use in plan retrieval, briefing generation, and on-site operation guidance; 4. Risk Case Library: Stores historical hazard data, safety accident cases, and rectification and disposal plans for this project and the industry, for use in risk identification, warning education, and guidance on hazard disposal.

[0030] (II.2) Core Modules for Building the Knowledge Base 1. Intelligent Document Parsing Unit: Supports parsing multiple document formats such as Word, PDF, Excel, images, and videos. It extracts text content from images and scanned documents through an OCR engine, removes invalid headers and footers, watermarks, and garbled characters through a format cleaning algorithm, restores the document's chapter structure and hierarchical relationship, and realizes the structured processing of unstructured documents. 2. Text Segmentation and Vectorization Unit: A sliding window segmentation strategy is used to segment the structured text into blocks. The window size is set to 512 tokens, and the stride is set to 128 tokens to avoid semantic breaks and information redundancy. A pre-trained vector encoder, bge-large-zh-v1.5, is used to transform the segmented text into 1024-dimensional dense vectors. The vector generation formula is as follows: in, For the first The 1024-dimensional feature vector of each text block. For the first The text content of each text block These are the pre-trained parameters for the vector encoder. It is a bidirectional text encoder based on the Transformer architecture.

[0031] 3. Knowledge Association Modeling Unit: Employing a bottom-up knowledge graph construction method, entities in the construction safety domain are extracted through a Named Entity Recognition (NER) model, including sub-projects, construction procedures, risk sources, hazard types, control measures, regulatory clauses, and responsible positions. Relationships between entities are constructed through a relation extraction model, including inclusion, cause, control, application, and responsibility. Finally, the knowledge graph is stored in the Neo4j graph database, providing structured knowledge support for risk tracing and compliance checks. 4. Incremental Training and Update Unit: A daily incremental update mechanism is set up to automatically collect newly added construction plans, hazard rectification records, performance data, and management documents for the project. This completes incremental parsing, vectorization processing, and knowledge graph updates of the documents, enabling continuous learning and iterative optimization of the knowledge base and ensuring that the knowledge base is synchronized with the actual progress of the project.

[0032] (II. III. Vector Databases and Real-Time Data Lakes) The vector database uses the Milvus distributed vector database to store 1024-dimensional vectors, original text, and metadata (document type, project, effective time, and tags) corresponding to text blocks. It supports millisecond-level high-performance vector similarity retrieval, providing core support for RAG enhanced retrieval. The real-time data lake adopts the Delta Lake architecture to store structured dynamic data generated by the project in real time. It supports batch and stream data processing, providing real-time data support for AI time-series analysis and performance quantification calculations.

[0033] (III) AI Capability Layer The AI ​​capability layer is the intelligent core of the system, integrating AI models and algorithm engines adapted to construction safety management scenarios. It provides standardized capability outputs for upper-layer multi-agent systems, including seven core functional units: (III.1) Large Language Model Reasoning Unit The inference unit utilizes open-source, commercially viable large language models such as DeepSeek-R1 and Tongyi Qianwen 4, supporting localized and private deployment. It is responsible for semantic understanding, logical reasoning, text generation, question-and-answer interaction, report generation, and decision-making suggestion output. The prompt word template has been optimized for construction safety management scenarios to address the "large model illusion" problem and ensure the professionalism, compliance, and feasibility of the output content. The core prompt word template is as follows (example): Plain Text <System Prompt> You are a dedicated AI expert specializing in construction safety management. When handling related issues, you must strictly adhere to the provided project knowledge base context and real-time construction data in your responses. Fabricating information is strictly prohibited, and you must not answer questions outside the scope of the knowledge base. Your answers must comply with construction industry safety regulations, employ a structured output format, ensure the content is directly implementable, and use a language style appropriate for the management environment of a construction site.

[0034] <knowledge base context>{context} <Project Real-time Data>{realtime_data} <User Question>{question}. (III.2) RAG Enhanced Search and Reordering Unit This unit is the core unit for improving the accuracy of question answering and solving the illusion of large models. It adopts a two-level retrieval mechanism of "vector retrieval + keyword retrieval" and combines a re-ranking model to achieve accurate filtering of highly relevant contexts.

[0035] like Figure 3 As shown, the core operating mechanism of this unit is as follows: 1. First-level vector retrieval: The user's question is transformed into a query vector. Using cosine similarity calculation, the top 50 semantically relevant text blocks are retrieved from the vector database. The cosine similarity calculation formula is: in, User question vector With knowledge base text block vectors The cosine similarity, with a value range of [-1, 1], indicates that the higher the value, the stronger the semantic relevance; The dot product of two vectors. , These are the L2 norms of the two vectors, respectively.

[0036] 2. Secondary Keyword Retrieval: The BM25 algorithm is used to perform keyword matching retrieval of user questions, recalling the top 30 text blocks in the knowledge base. The BM25 algorithm score calculation formula is as follows:

[0037] in, For users to query text, For the first in the query One keyword, For the document to be matched; Inverse document frequency, Keywords In the document Frequency of occurrence in For document Length, The average length of all documents in the corpus; , To enable adjustable hyperparameters, this system is configured with... , .

[0038] 3. Re-ranking and filtering: The results of the two searches are merged and deduplicated. The relevance of the merged text blocks is scored by the bge-reranker-large re-ranking model. The top 10 highly relevant text blocks are selected as the context input for the large language model, which greatly improves the accuracy of the answer.

[0039] The specific mechanism by which the bge-reranker-large reordering model scores the relevance of merged text blocks is as follows: The model is based on the bidirectional encoding mechanism of the Transformer architecture and uses a cross-attention mechanism to achieve deep semantic interaction between queries and text blocks. First, the user query and candidate text blocks are transformed into 768-dimensional dense vectors through a pre-trained language model. Then, contextual features are fused through a multi-layer Transformer encoder to capture the semantic associations of key entities in areas such as "construction safety specifications", "risk and hazard types", and "disposal measures".

[0040] The specific implementation process includes: 1) The input layer performs length normalization processing on the merged text block (truncated / padded to 512 tokens); 2) The feature extraction layer generates context-aware vectors through 12 Transformer modules, with the output of the 8th layer serving as intermediate features; 3) The stratified approach uses a bilinear attention mechanism to calculate the query-text similarity score, with the formula Score(Q,T)=Q T ·W·T, where: Q represents the user query vector (768-dimensional dense vector), generated by encoding through a pre-trained language model; T represents the candidate text block vector (768-dimensional dense vector), output from the merged text block through a feature extraction layer; W is a 32×768-dimensional learnable weight matrix used to capture the cross-semantic association between the query and the text block. A bilinear attention mechanism is used to calculate the semantic similarity score between the query and the text block; a higher value indicates a stronger relevance between the text block and the query.

[0041] 4) The sorting layer arranges text blocks in descending order of score, retaining the Top 10 results as the knowledge source for RAG enhancement.

[0042] The process achieves an average processing time of 87ms per query on an NVIDIA A100 GPU, which is at least 4-50% higher than the relevance accuracy of the traditional BM25 algorithm.

[0043] (III.3) Construction Risk Timing Prediction Unit An Attention-LSTM network is employed to perform time-series modeling of project hazards, risks, and performance data, predicting risk evolution trends and enabling early warning. The model input consists of daily time-series features from the past 30 days, including the number of newly identified hazards, the number of rectified hazards, the number of overdue hazards, the personnel performance completion rate, the number of high-risk operations, and the number of equipment anomalies. The output is the probability of risk occurrence and hazard trends for the next 7 days.

[0044] The basic structure of an LSTM network model includes a forget gate, an input gate, an output gate, and cell states, which will not be elaborated here.

[0045] Based on the probability of risk occurrence output by the model, the risk is divided into three levels: high risk (probability ≥ 80%), medium risk (50% ≤ probability < 80%), and low risk (probability < 50%), corresponding to red, yellow, and green warnings.

[0046] (III. IV) Quantitative Analysis Unit for Safety Performance The core calculation formulas for quantifying the performance of responsible personnel and grids, assigning security codes, and identifying weak links include: 1. Formula for calculating the completion rate of grid-based duties: in, This represents the completion rate of the grid's duties within the statistical period. This represents the number of tasks actually completed within the period. This refers to the number of tasks that must be completed within a specified period.

[0047] 2. Formula for calculating the overall score of personnel performance: in, The overall score for personnel performance is 100 points. Weighting for individual job performance completion rate ; Weighting the completion rate of hazard rectification ; For violations that result in point deductions, the weight is... ; Weighting for safety training completion rate .

[0048] 3. Personnel security code assignment rules: Green code: The score is 5, and there are no overdue hidden dangers or serious violations; Yellow code: The score is 1 point, or there is 1 general hidden danger that has not been rectified within the time limit; Red code: The score is low, or there are serious violations, or there are two or more unrectified potential hazards.

[0049] (III. V) Other functional units Knowledge graph construction engine: Responsible for the automated construction, updating and reasoning of knowledge graphs in the field of construction safety, supporting risk tracing, root cause analysis of hidden dangers, and compliance path query; Speech recognition and semantic understanding unit: It adopts a fine-tuned Whisper-large-v3 Chinese speech recognition model, optimized for the noise environment of construction sites, supports mobile voice input, and has a speech-to-text accuracy of ≥98%; at the same time, through the intent recognition model, it identifies the core intent of the user's voice and automatically matches the corresponding question-and-answer template and search scope; Intelligent recommendation engine: Based on collaborative filtering and content matching algorithms, it pushes matching safety regulations, training content, and control measures to users, while also pushing targeted management suggestions based on project risk trends.

[0050] (iv) Multi-agent cooperative scheduling layer The multi-agent collaborative scheduling layer is the central scheduling core of the system. Based on the multi-agent architecture, it designs four dedicated agents and realizes information interaction and task collaboration between agents through a standardized message bus, completing the closed-loop management of the entire process from knowledge accumulation, interactive question answering, analysis and decision-making to message push.

[0051] Each intelligent agent possesses independent task objectives, perception capabilities, decision-making capabilities, and execution capabilities, while simultaneously accepting unified coordination from the intelligent agent collaborative scheduling center.

[0052] (IV.1) Four Core Intelligent Agents 1. Knowledge Base Management Intelligent Agent Core objective: To automate the construction, updating, maintenance, and optimization of a project-specific knowledge base, thereby accumulating and reusing enterprise security knowledge assets.

[0053] It automatically receives uploaded documents and performs intelligent parsing, segmentation, and vectorization, writing them into the vector database and updating the knowledge graph; it automatically completes incremental data collection and knowledge base updates daily; it receives user feedback data from the question-and-answer interactive agent to optimize knowledge base content and retrieval accuracy.

[0054] The agent's inputs and outputs are as follows: Inputs include: user-uploaded project documents in various formats, industry standards, solutions, hazard data, and accident cases; and real-time project data collected by the system. Outputs: a structured project-specific knowledge base, a vector database index, and a construction safety knowledge graph.

[0055] 2. Security Question-Answering Interactive Intelligent Agent Core objective: To provide on-site personnel with accurate and professional intelligent Q&A services for construction safety management 24 / 7, and to enable the widespread reuse of expert capabilities.

[0056] After receiving user questions, the system performs speech recognition and intent recognition, calls the RAG enhanced retrieval unit to obtain highly relevant knowledge base context, and calls the large language model reasoning unit to generate structured professional answers. For questions beyond the scope, the system calls the network search unit to obtain authoritative information, and simultaneously records interaction data and feeds it back to the knowledge base management agent.

[0057] The operating mechanism is as follows: Input: User's text / voice questions, user identity and permission information, and real-time project data; Output: Professional answers matching user needs, on-site safety guidance, standardized construction suggestions, and data query results.

[0058] 3. Performance Analysis and Decision-Making Intelligent Agent Core objective: To conduct in-depth analysis of the security management situation based on multi-dimensional data, generate actionable intelligent decision-making suggestions, and upgrade from "experience-based decision-making" to "data-driven decision-making".

[0059] The system automatically calculates the daily performance completion rate and comprehensive score of grid members and personnel, and assigns safety codes; it calls the time-series prediction unit to generate risk warnings for the next 7 days; it identifies potential hazards and management shortcomings through clustering algorithms; and it generates targeted management suggestions and standardized safety management reports by combining knowledge bases.

[0060] The operating mechanism is as follows: Inputs: Real-time data on personnel, equipment, potential hazards, performance of duties, and risks from the data lake; and standards and regulations from the knowledge base. Outputs: Personnel performance heatmap, equipment health index, risk trend warning, hazard clustering analysis results, AI decision suggestions, and comprehensive project safety management report.

[0061] 4. Message Push Intelligent Agent Core objective: To achieve automated and precise delivery of safety management information, ensuring that management instructions and early warning information are delivered without any omissions.

[0062] Daily safety briefings are generated according to preset templates and schedules and pushed to the management team via WeChat bot; high-risk warnings and overdue potential hazards are pushed to the corresponding responsible persons in real time; hierarchical and precise push notifications are implemented based on user roles and permissions; message reading status is tracked, and unprocessed information is given secondary reminders, forming a closed-loop management system. The operating mechanism is as follows: Inputs: Early warning information, problem summaries, and management suggestions generated by the performance analysis and decision-making intelligent agent; daily system operation data; and user-defined push rules and templates. Outputs include daily security briefings, real-time risk warnings, task rectification notices, and management updates. Push channels include WeChat bots, apps, mini-programs, and SMS.

[0063] (IV.2) Intelligent Agent Cooperative Scheduling Center The scheduling center is responsible for task scheduling, resource allocation, message communication, and status monitoring of the four intelligent agents. It adopts an event-driven scheduling mechanism, automatically scheduling the corresponding intelligent agent to complete the collaborative task when a specific business event is triggered. For example, when a user raises a question, the scheduling center first schedules the knowledge base management intelligent agent to complete the knowledge base retrieval, and then schedules the security question-and-answer interaction intelligent agent to complete the answer generation; at a fixed time each day, the performance analysis and decision-making intelligent agent is first scheduled to complete the data analysis and report generation, and then the message push intelligent agent is scheduled to complete the briefing push.

[0064] (v) Application Interaction Layer The application interaction layer serves as the entry point for interaction between the system and users. It provides multi-terminal adapted interfaces for different user roles and scenarios, including mobile apps, WeChat mini-programs, PC management terminals, WeChat robots, and on-site smart terminals. Customized functional modules are developed for different roles such as frontline safety officers, grid workers, project managers, and company management. For frontline workers: the focus is on voice Q&A, safety training, and hazard reporting functions; Safety Officer / Grid Member: Focuses on fulfilling duties, identifying and rectifying potential hazards, and providing on-site safety guidance; Project Leader: Focuses on data visualization dashboards, risk warnings, AI decision-making suggestions, and management reporting functions; Company management focuses on functions such as summarizing and analyzing multiple projects, reusing knowledge across projects, and evaluating management performance.

[0065] The core functional interfaces include: AI intelligent question and answer interface, grid-based duty performance interface, hidden danger investigation and rectification interface, dual prevention mechanism interface, data visualization dashboard, personnel / equipment / work point code management interface, knowledge base management interface, and system configuration interface.

[0066] The above is a detailed introduction to the system architecture.

[0067] like Figure 4 As shown, the system adopts a three-layer communication architecture of "cloud-edge-device" collaboration to ensure the real-time performance, security, and reliability of data transmission. 1. Edge side: Includes mobile devices, field IoT devices, sensors, and video monitoring terminals, responsible for field data collection and user interaction. It communicates with the cloud system via 4G / 5G / WiFi and uses HTTPS / TLS 1.3 encryption protocol to ensure transmission security. For remote construction sites without network access, edge computing nodes are deployed to handle local data collection, preprocessing, and caching. Once the network is restored, the data is automatically synchronized with the cloud to ensure the system's offline availability. 2. Edge computing nodes deploy lightweight AI models to be responsible for real-time hazard identification of on-site videos and edge preprocessing of IoT data, reducing cloud computing pressure and data transmission bandwidth usage; 3. Cloud: The core services of the system are deployed in the cloud. The modules at each level communicate with the intelligent agents at high speed using the gRPC protocol. The data lake and the vector database communicate using JDBC / RESTful interfaces. The large language model service uses OpenAI-compatible API interfaces for calls, supporting horizontal scaling to adapt to concurrent access from multiple projects.

[0068] The system employs four event-driven closed-loop control logics to realize the PDCA cycle throughout the entire construction safety management process: 1. Knowledge Accumulation Closed Loop: Document Upload → Intelligent Parsing → Vectorized Storage → Retrieval Feedback → Incremental Optimization, realizing the continuous accumulation and iteration of project security knowledge; 2. Question-and-answer interaction closed loop: User asks a question → Intent recognition → RAG retrieval → Large model answers → User feedback → Model optimization, achieving continuous improvement in question-and-answer accuracy; 3. Safety Management Closed Loop: Data collection → AI analysis → risk warning → decision-making suggestions → implementation of rectification → effect verification → data update, realizing closed-loop control of the entire process of construction safety management; 4. Closed-loop message delivery: Event triggering → Information generation → Targeted push → Reading confirmation → Secondary reminder → Closed-loop archiving, achieving zero omissions in the delivery of management information. The specific implementation process will be described in detail below.

[0069] II. Construction AI-Based Risk Analysis and Decision-Making Methods This method is implemented based on the above system, such as Figure 5 As shown, the method includes 8 core steps, and the specific implementation process is as follows: Step 1: Initialization and Construction of Project-Specific Knowledge Base 1. After the system is deployed, the project administrator creates a project space on the PC management terminal and configures the basic project information (project type, section division, grid division, personnel organizational structure, and permission allocation). 2. Upload project-related documents, including bidding documents, contracts, construction organization design, special construction plans, safety management systems, emergency plans, historical hazard data, and accident cases. At the same time, connect to the industry standard library and import the applicable national, industry, and local safety standards for the project. 3. The knowledge base management agent automatically initiates the document processing workflow: First, it extracts the text content of documents in various formats through OCR and document parsing engines, completing format cleaning, deduplication, and filtering of invalid content; then, it uses a sliding window strategy to divide the text into blocks, and uses a vector encoder to convert the text blocks into 1024-dimensional dense vectors; finally, it writes the vectors, original text, and metadata into the Milvus vector database, and at the same time builds the project knowledge graph to complete the initialization of the knowledge base. 4. Configure the knowledge base update rules, set up automatic incremental updates every morning at midnight, and synchronize newly added documents, hazard data, and performance records for the project to achieve continuous learning of the knowledge base. This transforms project safety knowledge scattered in paper and electronic documents into reusable and searchable digital assets, solving the pain point of knowledge loss after project completion and providing dedicated knowledge support for subsequent AI question answering and analysis decision-making.

[0070] Step 2: Initialization of Multi-Agent Cooperative System and System Parameters 1. Based on the project's organizational structure and grid division, assign corresponding system permissions and intelligent agent usage scope to users with different roles. For example, front-line safety officers are given access to Q&A interaction and job performance functions, while project leaders are given access to data analysis and decision-making suggestion functions. 2. Configure the operating parameters of each agent, including: the incremental update cycle of the knowledge base management agent, the recall quantity and similarity threshold of RAG retrieval, the performance calculation cycle of the performance analysis and decision-making agent, the security code assignment rules, the time window for risk prediction, and the push time, template, and recipient of the message push agent; 3. Integrate with the project's existing personnel management system, video surveillance system, IoT devices, and construction management platform; complete the debugging of data interfaces; and achieve real-time synchronization of multi-source data. 4. Conduct system trial runs to test the collaborative operation of each intelligent agent, optimize the large model prompt word template and algorithm hyperparameters, and ensure that the system output meets project management requirements. Initialization can complete the deep adaptation of the system with the actual project management scenario, realize the collaborative linkage of each intelligent agent, lay the foundation for subsequent full-process operation, and solve the problem of the disconnect between traditional systems and project management.

[0071] Step 3: Real-time acquisition and preprocessing of multi-source construction safety data 1. Each data collection unit in the basic data layer synchronizes project personnel attendance data, job performance completion data, equipment operation status data, hidden danger investigation and rectification data, high-risk operation approval data, and IoT sensor data in real time through standardized interfaces; 2. Preprocess the collected raw data: remove outliers using the 3σ principle, fill missing values ​​using linear interpolation, and map data of different dimensions to the [0,1] interval using min-max normalization to complete the data standardization process; 3. The preprocessed structured data is written to Delta Lake in real time and stored in partitions according to time, segment, and grid dimensions to provide standardized real-time data support for subsequent AI analysis; 4. For unstructured on-site inspection videos and images, a target detection model is used to identify on-site violations and safety hazards, and structured information is extracted and written into the data lake. Preprocessing enables real-time collection and standardized processing of safety management data across all dimensions of the project, solving the pain points of traditional data being scattered, dormant, and unusable, and providing a high-quality data foundation for AI deep analysis.

[0072] Step 4: RAG-enhanced intelligent question-and-answer interaction for security management 1. Frontline users at the construction site can use the voice input function of the mobile APP / mini-program to ask questions, such as "What are the key points of safety management for working at heights?", "What are the responsibilities of grid 3 today?", and "What are the rectification measures for the hidden danger of not setting up edge protection?". 2. The speech recognition unit converts speech into text, and the intent recognition model identifies the user's core intent and search scope, transforming the user's question into a query vector through a vector encoder; 3. RAG Enhanced Retrieval and Re-ranking Unit initiates a two-level retrieval: First, the top 50 highly semantically relevant text blocks are retrieved from the vector database through cosine similarity calculation. Then, the top 30 keyword-matching text blocks are retrieved through the BM25 algorithm. After merging and deduplication, the top 10 most relevant texts are selected through the re-ranking model. 4. The large language model reasoning unit fills the prompt word template with the user's question, the retrieved knowledge base context, and the project's real-time data, generates a structured answer that conforms to construction safety specifications, and at the same time matches the corresponding images, videos, and specification clauses, and pushes them to the user; 5. For users' questions that exceed the scope of the question, the system utilizes online search to retrieve the latest industry standards, policy documents, and incident reports. After verification, these are added to the answer, and the new, valid content is simultaneously synchronized to the knowledge base. This step provides on-site personnel with 24 / 7 AI expert support, enabling millisecond-level responses to on-site issues. It addresses the pain points of traditional methods of searching for standards and experts inefficiently, democratizing expert capabilities and significantly improving the responsiveness and standardization of on-site safety management.

[0073] Step 5: AI-based quantitative analysis of security performance based on multi-dimensional data 1. Every day at dawn, the performance analysis and decision-making intelligent agent automatically extracts the full amount of data from the previous day from the real-time data lake, calls the safety performance quantitative analysis unit, and calculates the performance completion rate of personnel in each level-three grid and each position within the period. ; 2. Calculate the overall performance score of personnel by combining the completion rate of hazard rectification, records of violations, and completion rate of safety training. Complete the assignment of red, yellow, and green security codes for personnel and grids according to the coding rules; 3. Conduct multi-dimensional analysis of performance data, generate personnel performance heatmaps, identify weak grids with a performance completion rate of less than 80% and personnel with high frequency of overdue performance, and analyze obstacles to performance. 4. Analyze equipment data, calculate equipment health index, identify equipment that is due for maintenance or is malfunctioning, and generate early warning information; The analysis results are stored in a data lake and simultaneously pushed to a message push intelligent agent to issue early warnings to personnel with red codes and those in charge of vulnerable areas. Through automated and quantitative analysis of project safety performance, the work of traditional manual statistical reporting has been replaced, increasing efficiency by more than 10 times. At the same time, it accurately identifies management shortcomings and provides data support for subsequent management optimization.

[0074] Step 6: Dynamic Prediction and Intelligent Decision Generation of Construction Risks 1. The performance analysis and decision-making intelligent agent calls the construction risk time series prediction unit to extract the daily time series data of the past 30 days, including 8 core features such as the number of hidden dangers, the performance completion rate, the number of high-risk operations, and the number of equipment anomalies, and inputs them into the trained Attention-LSTM model; 2. The model outputs the probability of hazard occurrence and risk evolution trend for the next 7 days, and provides graded early warnings according to high, medium and low risks, and identifies high-risk construction sites and operation types; 3. Use the DBSCAN clustering algorithm to cluster historical hazard data of the project, identify high-frequency hazard types, locations of occurrence, time patterns, and responsible work teams, and uncover the underlying causes of hazard occurrences; 4. The large language model reasoning unit combines risk prediction results, hidden danger clustering analysis results, and control standards in the knowledge base to generate actionable AI decision-making suggestions, including personnel interviews, adjustment of inspection frequency, special training, and optimization of resource allocation; 5. Automatically generates comprehensive project safety management reports, including full-dimensional data on personnel performance, equipment management, risk control, and hazard rectification. Users can adjust analysis dimensions and report templates through custom formulas. By shifting from "post-event remediation" to "pre-event prevention," AI predicts and identifies risks in advance, and data drives precise management decisions, it solves the pain points of traditional management relying on experience and guesswork, significantly reducing the probability of safety accidents.

[0075] Step 7: Automated Push and Closed-Loop Control of Management Information 1. The message push intelligent agent automatically summarizes the project safety management data of the previous day according to the preset daily push time, generates a standardized daily safety briefing, and pushes it to the project management team through WeChat robot; 2. For urgent information such as high-risk warnings, overdue rectified hazards, and individuals with red health codes, push notifications will be sent to the relevant responsible persons in real time via APP, mini-program, and SMS, clearly specifying rectification requirements and deadlines; 3. Based on users' roles and permissions, information can be pushed to them in a precise and hierarchical manner. For example, grid workers can only receive task notifications within their own grids, while project leaders can receive overall project security status and warnings of major risks. 4. Track the reading and execution status of messages, provide secondary reminders for warning messages that have not been processed for more than 24 hours, and push them to superiors simultaneously, forming a closed-loop management system of "warning - push - execution - feedback - archiving"; 5. Automatically record push and execution data, store it in a data lake, and use it as a basis for personnel performance evaluation. By ensuring zero omissions in the delivery of security management information, it solves the pain points of traditional management relying on human reporting and information delays, forms a standardized management rhythm, and improves the efficiency of management instruction execution.

[0076] Step 8: Continuous Iteration of the Knowledge Base and Incremental Optimization of the Model 1. The knowledge base management intelligent agent automatically collects newly added construction plans, hidden danger rectification records, performance data, and accident cases of the project every day, completes incremental parsing and vectorization of documents, updates the vector database and knowledge graph, and realizes the continuous accumulation of project knowledge. 2. Collect user question and answer feedback and answer correction data, optimize high-frequency low-match knowledge base content, and incrementally fine-tune the vector model and re-ranking model based on project-specific data to improve the accuracy of retrieval and question answering; 3. Based on the newly added hidden dangers and risks data of the project, the time series prediction model is incrementally trained to optimize the model parameters and improve the accuracy of risk prediction; 4. Upon project completion, a complete safety management knowledge base is established, becoming a digital knowledge asset for the enterprise. This knowledge can be reused in subsequent similar projects, achieving "knowledge remains even after the project ends." Through the continuous evolution of the knowledge base and AI model, the knowledge becomes increasingly sophisticated and accurate with each use, addressing the pain point of traditional project knowledge loss due to personnel turnover and enabling the continuous accumulation and reuse of the enterprise's safety knowledge assets.

[0077] Example This embodiment uses the xx port construction project as an application scenario to explain in detail the specific implementation process and application effects of this system: 1. Project Overview: The XX Port project comprises multiple unit projects including container yards, wharf hydraulic engineering, land reclamation, and building construction. It is divided into 1 primary grid, 5 secondary grids, and 10 tertiary grids, with over 1200 construction workers and 124 pieces of special equipment on site. The project's traditional safety management model suffers from problems such as inadequate performance of duties, untimely rectification of hidden dangers, difficulty in verifying on-site personnel's compliance, and insufficient coverage of expert resources, resulting in significant safety management pressure.

[0078] 2. System Deployment and Initialization: (1) The system is deployed locally on the project's local server to meet the project's data security requirements. The system completes the creation of the project space, grid division, personnel organization structure and permission configuration, and connects to the project personnel management system, special equipment management system and on-site video monitoring system. (2) Upload the project construction organization design, 23 special construction plans, project safety management system, emergency plan, national / industry standards and norms related to port engineering, historical hidden dangers and accident cases, totaling 126 documents. (3) The knowledge base management agent completes the intelligent parsing and block division of the documents, generating a total of 3200+ text blocks, which are converted into 1024-dimensional vectors and stored in the Milvus vector database. At the same time, a port construction safety knowledge graph containing 8 types of entities and 12 types of relationships is constructed, and the initialization of the project's exclusive knowledge base is completed. (4) Configure system parameters: the duty calculation cycle is daily, the safety code assignment rule is executed according to the formula of this system, the risk prediction window is the next 7 days, the daily safety briefing is pushed at 8 am every day, and the RAG retrieval similarity threshold is set to 0.6.

[0079] 3. Core process of system operation: (1) Intelligent question and answer application: Before the dynamic compaction construction of No. 20 yard, the on-site safety officer asked "What are the key points of safety management and control for dynamic compaction construction?" through the APP voice. After the system voice recognition is converted into text, the relevant content of dynamic compaction construction in the project special construction plan and industry standard is retrieved through RAG. The large model generates a structured answer, including 12 key points of management and control in three major modules: pre-construction inspection, operation process control, and emergency response. At the same time, the corresponding standard clauses and teaching videos are pushed. The entire response time is ≤2 seconds, which solves the problem of no expert guidance for on-site personnel. (2) Quantitative analysis of performance: The system automatically calculates the performance completion rate of each grid every day. Example calculation: No. 20 berth third-level grid should complete 17 performance tasks on the day, and actually completed 6 tasks, with a performance completion rate of The system automatically marks the grid as a red code grid and pushes early warning information to the grid leader and project manager. At the same time, the AI ​​analyzes the reasons for the weak performance of duties and suggests adjusting the grid personnel configuration and strengthening on-site supervision. (3) Risk prediction and decision-making: The system extracts the time series data of the past 30 days and inputs it into the Attention-LSTM model to predict that the probability of high-altitude operation hazards in the bridge substructure construction area in the next 7 days is 86%, which is a high risk. The AI ​​automatically generates decision-making suggestions: ① Talk to the grid leader of the grid to understand the obstacles to the performance of duties; ② Increase the number of high-altitude operation special inspections in the area by 2 times a day in the next 3 days; ③ Push high-altitude operation special training to all personnel in the grid; ④ Adjust the inspection frequency of safety protection facilities in the area. (4) Message push: At 8 am every day, the WeChat robot automatically pushes the daily safety briefing of the project, including the number of duties completed in the previous day (124), 7 hidden dangers found, 4 rectified, 3 overdue hidden dangers, and 1 high-risk warning. At the same time, it pushes the outstanding issues and AI management suggestions. The project management team can grasp the overall safety dynamics of the project without logging into the system.

[0080] 4. Application Results: After the system was online for 3 months, the project personnel's job performance completion rate increased from 38% to 92%, the timely rectification rate of hidden dangers increased from 65% to 100%, the response time for on-site personnel to query safety regulations was shortened from an average of 30 minutes to 2 seconds, the labor cost of safety training and management was reduced by 70%, and no safety accidents occurred in the project, realizing the digital and intelligent upgrade of safety management.

[0081] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A construction AI risk analysis and decision-making system based on multi-agent collaboration, characterized in that, It includes a basic data layer, a data knowledge layer, an AI capability layer, a multi-agent collaborative scheduling layer, and an application interaction layer, among which: The basic data layer is used for the collection, cleaning and standardized storage of multi-source construction safety-related data, providing raw data support for the upper layer. It has a built-in data preprocessing module to complete the removal of outliers, filling of missing values ​​and standardization of raw data. The data knowledge layer is used to transform unstructured documents and structured data into reusable digital knowledge assets. It includes four core libraries: a standard library, a project system library, a construction plan library, and a risk case library, as well as a vector database and a real-time data lake. It also includes a document intelligent parsing unit, a text segmentation and vectorization unit, a knowledge association modeling unit, and an incremental training and update unit. The AI ​​capability layer provides intelligent algorithms and model support for the system, including a large language model reasoning unit, a RAG enhanced retrieval and re-ranking unit, a construction risk time-series prediction unit, a safety performance quantitative analysis unit, a knowledge graph construction engine, a speech recognition and semantic understanding unit, and an intelligent recommendation engine. The multi-agent collaborative scheduling layer is the central scheduling core of the system, including a knowledge base management agent, a security question-and-answer interaction agent, a performance analysis and decision-making agent, a message push agent, and an agent collaborative scheduling center. The agent collaborative scheduling center is used to realize the task scheduling, message communication and collaborative operation of the four agents, and complete the closed-loop management of the entire process from knowledge accumulation, interactive question and answer, analysis and decision-making to message push. The application interaction layer serves as the entry point for interaction between the system and users, including mobile apps, WeChat mini-programs, PC management terminals, WeChat robots, and on-site smart terminals.

2. The system according to claim 1, characterized in that, The text segmentation and vectorization unit adopts a sliding window segmentation strategy with a window size of 512 tokens and a stride of 128 tokens. The segmented text is converted into 1024-dimensional dense vectors by a vector encoder. The vector generation formula is as follows: in, For the first A 1024-dimensional feature vector of a text block, For the first The text content of each text block These are the pre-trained parameters for the vector encoder. It is a bidirectional encoder based on the Transformer architecture.

3. The system according to claim 1, characterized in that, The RAG enhanced retrieval and reordering unit adopts a two-level retrieval mechanism. The first level is vector retrieval based on cosine similarity, and the second level is keyword matching retrieval based on the BM25 algorithm. After merging the retrieval results, highly relevant text blocks are selected through a reordering model. The formula for calculating cosine similarity is: in, User question vector With knowledge base text block vectors cosine similarity, The dot product of two vectors. , Let L2 norm be the L2 norm of the two vectors respectively; the BM25 algorithm scoring formula is as follows: ； in, For users to query text, For the first in the query One keyword, For the document to be matched, Inverse document frequency, Keywords In the document Frequency of occurrence in For document Length, Let be the average length of all documents in the corpus, and k1 and b are adjustable hyperparameters.

4. The system according to claim 1, characterized in that, The construction risk time series prediction unit adopts an LSTM long short-term memory network based on the attention mechanism. The model input is the daily time series data of the past 30 days, and the output is the probability of risk occurrence in the next 7 days. The daily time-series characteristics include the number of newly added hidden dangers, the number of rectified hidden dangers, the number of overdue hidden dangers, the personnel performance completion rate, the number of high-risk operations, and the number of equipment malfunctions; The probability of risk occurrence is used to classify risks into three levels: high risk, medium risk, and low risk, corresponding to red, yellow, and green warning colors.

5. The system according to claim 1, characterized in that, The quantitative analysis unit for safe performance of duties includes a built-in formula for calculating the grid performance completion rate, a formula for calculating the comprehensive score of personnel performance, and a security code assignment rule. The formula for calculating the grid performance completion rate is as follows: in, The percentage of tasks completed within the grid cycle. This represents the number of tasks actually completed within the period. The number of tasks to be completed within the specified period; The formula for calculating the overall performance score of personnel is as follows: in, As a comprehensive score for personnel performance, Weighting for individual job performance completion rate ; Weighting the completion rate of hazard rectification ; For violations that result in point deductions, the weight is... ; Weighting for safety training completion rate .

6. The system according to claim 1, characterized in that, The knowledge base management intelligent agent is used to realize the automated construction, incremental update and optimization of the project-specific knowledge base. Its input is multi-format project documents uploaded by users and project dynamic data collected by the system in real time, and its output is a structured project-specific knowledge base, vector database index and knowledge graph. The aforementioned safety question-and-answer interactive intelligent agent is used to provide users with intelligent question-and-answer services for construction safety management. Its inputs are the user's text / voice questions, user identity and permission information, and real-time project data, and its outputs are professional answers that match the needs and on-site safety guidance. The aforementioned performance analysis and decision-making intelligent agent is used to complete in-depth analysis of the safety management situation and generate intelligent decision suggestions. Its input is multi-dimensional construction data from the real-time data lake and knowledge base control requirements, and its output is risk warning, data analysis results, AI decision suggestions, and standardized safety management reports. The message push intelligent agent is used to realize the automated and accurate push of security management information. Its input is the output data of the performance analysis and decision-making intelligent agent and the user-defined push rules, and its output is daily security briefing, real-time risk warning and task rectification notice.

7. A construction AI-based risk analysis and decision-making method based on multi-agent collaboration, characterized in that, The system implementation based on any one of claims 1-6 includes the following steps: Initialization and construction of project-specific knowledge base: Create project space, upload project-related documents and industry standards, complete intelligent document parsing, text segmentation and vectorization processing through knowledge base management intelligent agent, build project-specific knowledge base and knowledge graph, and configure knowledge base incremental update rules; Multi-agent collaborative system and system parameter initialization: Configure user permissions, operating parameters of each agent, connect with the project's existing management system and IoT devices, and complete the adaptation of the system to the project management scenario; Real-time acquisition and preprocessing of multi-source construction safety data: Multi-source construction data is acquired in real time through the basic data layer, and preprocessing such as outlier removal, missing value filling and data normalization is completed. The standardized data is then written into the real-time data lake. RAG-enhanced intelligent question-and-answer interaction for security management: It receives text / voice questions from users, completes speech recognition, intent recognition, RAG-enhanced retrieval and reordering, and generates professional answers that match the knowledge base and real-time project data through a large language model; AI-based quantitative analysis of safe performance based on multi-dimensional data: The intelligent decision-making agent automatically calculates the performance completion rate and comprehensive score of the grid and personnel, completes the assignment of security codes, and identifies weak links in performance. Dynamic prediction of construction risks and intelligent decision generation: The construction risk time series prediction unit predicts the risk evolution trend in the next 7 days, performs cluster analysis on the hidden danger data, and generates practical AI decision suggestions and safety management reports by combining the knowledge base; Automated push and closed-loop control of management information: The daily safety briefing is pushed on a scheduled basis and real-time warning of emergency risks is given through a message push intelligent agent, and the execution status is tracked to form a closed-loop control. Continuous iteration of the knowledge base and incremental optimization of the model: Daily incremental updates to the knowledge base, and incremental fine-tuning of the model based on user feedback and new project data, to achieve continuous evolution of system capabilities.

8. The method according to claim 7, characterized in that, The specific process of RAG enhanced retrieval and re-ranking is as follows: The user's question is transformed into a query vector using a vector encoder; By calculating cosine similarity, the top 50 highly semantically relevant text blocks are retrieved from the vector database. Keyword matching and retrieval were performed using the BM25 algorithm to retrieve the top 30 text blocks; The two search results are merged and deduplicated. The relevance of the merged text blocks is scored using a re-ranking model, and the top 10 highly relevant text blocks are selected as the context input for the large language model.

9. The method according to claim 7, characterized in that, The dynamic prediction of construction risks adopts an LSTM model based on an attention mechanism, and the specific process is as follows: Extract daily time-series data from the past 30 days, including the number of newly added hazards, the number of rectifications completed, the performance completion rate, the number of high-risk operations, and the number of equipment anomalies, and construct an input feature vector; The input feature vector is fed into the trained Attention-LSTM model, the LSTM unit extracts temporal features, and the attention mechanism assigns weights to key features. The model outputs the probability of risk occurrence in the next 7 days, and issues warnings in three levels: high risk ≥80%, medium risk 50% ≤ probability < 80%, and low risk < 50%.

10. The method according to claim 7, characterized in that, The message push intelligent agent adopts a hierarchical and precise push mechanism, which pushes information within the corresponding scope based on the user's position and permissions; for urgent information such as high-risk warnings and overdue unrectified hidden dangers, it pushes information to the corresponding responsible persons in real time, and provides secondary reminders for information that has not been handled in a timely manner, thus forming a closed-loop management.

Images