CN-121979851-A - Construction log generation method based on agent and large model

CN121979851ACN 121979851 ACN121979851 ACN 121979851ACN-121979851-A

Abstract

The invention provides a construction log generation method based on an intelligent agent and a large model, which comprises the steps of firstly constructing a multi-source heterogeneous data acquisition layer, automatically acquiring needed original data from a built-in interface, an external application programming interface, internet of things equipment and a manual input channel through the intelligent agent, secondly, establishing a data preprocessing and structuring layer, processing unstructured data by utilizing an AI technology, extracting key information and packaging the key information into a standardized data unit, thirdly, constructing a field model of proficiency construction terms and log wind, fourthly, generating a log primary draft by the large model, fifthly, providing a man-machine cooperation interface for an engineering staff to carry out auditing, correction and confirmation on the generated log primary draft, and sixthly, establishing an optimization closed loop based on manual feedback, and carrying out continuous iterative optimization on a data acquisition, preprocessing rule, a prompt word template or the large model. The method realizes full-automatic acquisition, and various data sources are automatically docked through the intelligent agent, so that manpower is thoroughly liberated.

Inventors

GUO YUANHAO
YU FANGQIANG
FANG TINGCHEN
CHEN YUANHONG

Assignees

上海建工集团股份有限公司

Dates

Publication Date: 20260505
Application Date: 20251222

Claims (5)

1. The construction log generation method based on the agent and the large model is characterized by comprising the following steps: step S1, constructing a multi-source heterogeneous data acquisition layer, wherein required original data are automatically acquired from a built-in interface, an external application programming interface, internet of things equipment and a manual input channel through an intelligent agent; Step S2, a data preprocessing and structuring layer is established, namely unstructured data are processed by utilizing an AI technology, key information is extracted, and the key information is packaged into a standardized data unit; Step S3, constructing a field model of proficiency construction terms and log wind, namely building or fine-tuning a professional large language model of a construction field to proficiency engineering expertise of the field; s4, the large model generates a journal manuscript, namely, a standardized data unit is injected into a preset structured prompt word and submitted to a professional large language model, and the construction journal full-text manuscript which accords with a specified format and content is automatically generated by the professional large language model; Step S5, man-machine collaborative auditing and optimization, namely providing a man-machine collaborative interface for an engineering person to audit, revise and confirm the generated log primary draft; and S6, establishing an optimization closed loop based on manual feedback, and performing continuous iterative optimization on the data acquisition, the preprocessing rule, the prompt word template or the large model.
2. The method according to claim 1, wherein in said step S1, developing or configuring a series of data acquisition agents comprises: Step 1.1, constructing a weather Agent, namely requesting yesterday weather data of a place where a project is located by using a Chinese weather bureau or a weather API, and returning structured JSON data; Step 1.2, BIM Agent construction, namely reading model component information corresponding to the present planning progress in a project BIM management platform through an IFC standard or a Revit API interface, automatically extracting space positioning information of the model component information, and generating construction _positioning; Step 1.3, constructing an attendance Agent, namely connecting a database of a building site access control system through an ODBC interface, executing SQL query, and summarizing query results into attendance data in a JSON format; Step 1.4, meeting Agent, which is to automatically access the audio and video stream of the meeting, record the engineering example meeting in the afternoon of the day and store the audio file into the queue to be processed; Step 1.5, constructing a document Agent, namely simulating a login project management system through a requests library of Python or an RPA technology, capturing a document newly added today, and storing the document to a local place; and 1.6, manually inputting, and providing a Web or mobile terminal form for filling out sudden faults and special measures.
3. The method according to claim 1, wherein in said step S2, the data preprocessing and generating standardized data units comprises: step 2.1, converting conference audio into text through ASR service, abstracting through NLP model, extracting key resolution and instruction, and generating standardized data unit; step 2.2, identifying the field picture through a CV model to generate a standardized data unit; Step 2.3, the structured data from the API and database are directly encapsulated into corresponding standardized data units; And 2.4, marking all generated standardized data units with a timestamp and a source label, and storing the standardized data units in corresponding partitions of the central knowledge base in time sequence.
4. A method according to claim 3, characterized in that in said step S4, the large model generation log-up comprises the steps of: Step 4.1, loading a professional large model, namely loading the trimmed professional large model into a GPU memory by a system; Step 4.2, constructing a prompt word, namely, the system retrieves all generated standardized data units from a knowledge base and fills the standardized data units into the following preset prompt word templates according to modules, and at the moment, the system automatically calls related preset data; And 4.3, calling a model to generate, namely inputting the assembled ultralong Prompt into a large model, and enabling the large model to understand and digest all data to generate a construction log full text which is 500 words or more and has a strict structure.
5. The method according to claim 4, wherein in the step S5, the human-computer collaborative audit and optimization includes the steps of: Step 5.1, constructing an audit interface, wherein a generated journal draft is presented on the right side of a Web audit interface, and all generated standardized data units for generating journals are listed on the left side as a fact basis; Step 5.2, performing manual auditing, and performing expression fine adjustment on the abnormal part; And 5.3, the system performs feedback learning, records the correction behavior, desensitizes the final log, stores the final log in a high-quality training sample library, and performs incremental fine adjustment on the large professional model by using new samples regularly so as to continuously improve the performance of the large professional model.

Description

Construction log generation method based on agent and large model Technical Field The invention belongs to the technical field of intelligent construction sites and construction engineering informatization, and particularly relates to a construction log generation method based on an intelligent body and a large model. Background The construction log is one of the most important original records in the building construction process, and comprehensively records daily construction activities, resource investment, quality safety conditions and major events. At present, the filling of the construction log is generally finished manually by constructors or operators, and has the problems of complicated and low-efficiency work, time and labor consumption for recall and summarizing various information all day, recordation of quality variation, easy omission of key matters, difficult information integration, difficult manual collection of information required by the log in different systems (such as monitoring, attendance checking, meeting, supervision notification and the like), difficult utilization of unstructured data, difficult conversion of valuable information such as meeting record, field picture, group chat record and the like into texts and extraction of points. At present, although individual software adopts a form filling form to simplify the record, the record is still manually recorded in nature, and automatic sensing, fusion and intelligent generation of data cannot be realized. Therefore, how to provide a construction log generating method based on an agent and a large model, to automatically generate a high-quality and structured construction log is a technical problem to be solved by those skilled in the art. Disclosure of Invention The invention provides a construction log generation method based on an agent and a large model, which solves the problems of low efficiency, low standardization degree, easy forgetting of data and high error rate of manual filling of the construction log, and realizes automation, intellectualization and standardization of construction log writing. The technical scheme of the construction log generation method based on the intelligent agent and the large model is as follows: A construction log generation method based on an agent and a large model comprises the following steps: step S1, constructing a multi-source heterogeneous data acquisition layer, wherein required original data are automatically acquired from a built-in interface, an external application programming interface, internet of things equipment and a manual input channel through an intelligent agent; Step S2, a data preprocessing and structuring layer is established, namely unstructured data are processed by utilizing an AI technology, key information is extracted, and the key information is packaged into a standardized data unit; Step S3, constructing a field model of proficiency construction terms and log wind, namely building or fine-tuning a professional large language model of a construction field to proficiency engineering expertise of the field; s4, the large model generates a journal manuscript, namely, a standardized data unit is injected into a preset structured prompt word and submitted to a professional large language model, and the construction journal full-text manuscript which accords with a specified format and content is automatically generated by the professional large language model; Step S5, man-machine collaborative auditing and optimization, namely providing a man-machine collaborative interface for an engineering person to audit, revise and confirm the generated log primary draft; and S6, establishing an optimization closed loop based on manual feedback, and performing continuous iterative optimization on the data acquisition, the preprocessing rule, the prompt word template or the large model. Further, in the step S1, developing or configuring a series of data acquisition agents includes: Step 1.1, constructing a weather Agent, namely requesting yesterday weather data of a place where a project is located by using a Chinese weather bureau or a weather API, and returning structured JSON data; Step 1.2, BIM Agent construction, namely reading model component information corresponding to the present planning progress in a project BIM management platform through an IFC standard or a Revit API interface, automatically extracting space positioning information of the model component information, and generating construction _positioning; Step 1.3, constructing an attendance Agent, namely connecting a database of a building site access control system through an ODBC interface, executing SQL query, and summarizing query results into attendance data in a JSON format; Step 1.4, meeting Agent, which is to automatically access the audio and video stream of the meeting, record the engineering example meeting in the afternoon of the day and store the audio file into the queue to be processed; Step 1.