Dynamic supervision method for construction quality of building engineering

Abstract

The invention discloses a building engineering construction quality dynamic supervision method, relates to the technical field of building engineering supervision, and aims to solve the problems that a traditional method depends on manual inspection, quality defects are often formed when problems are found, and the rectification cost is high. According to the technical scheme, a dynamic supervision system is adopted, and the method comprises the following steps that S1, physical states, images and check record data of a construction site are collected through an Internet of Things sensor, a mobile terminal and an unmanned aerial vehicle, the physical states, the images and the check record data are cleaned and aligned with design standard data in a building information model, and a standardized quality state data stream is generated; according to the method, fusion analysis is carried out on sensor data, images and texts through the multi-modal machine learning model, the probability of risks such as concrete cracking and supporting system instability can be predicted in advance, and early warning is given out before problems occur. Therefore, the management center of gravity can be moved from post-event remedy to in-event control and pre-event prevention, and the occurrence probability of quality accidents and rework loss are remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of building construction supervision technology, and more specifically, to a dynamic supervision method for the construction quality of building projects. Background Technology

[0002] Construction quality supervision is a crucial link in ensuring the structural safety and functionality of building projects. Traditional supervision models mainly rely on manual inspections, visual observations, and paper records by supervisors. Their quality judgment largely depends on personal experience. This human-centered supervision approach has significant limitations when dealing with large, complex, and time-sensitive modern engineering projects.

[0003] Currently, with the development of information technology, some digital assistance methods have been attempted to be applied in engineering supervision. These include using sensors to monitor specific parameters or employing BIM models for visualization. However, these technological applications are often isolated, lacking effective integration and linkage between data. Various types of data (such as sensor readings, on-site images, and text records) are usually in a separate state, unable to be comprehensively analyzed and deeply mined, leading to delayed and one-sided identification of quality risks.

[0004] Therefore, a new solution is needed to address this problem. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a dynamic supervision method for construction quality of building engineering, which has the advantage of being able to monitor construction risks in advance.

[0006] The above-mentioned technical objective of this invention is achieved through the following technical solution: a dynamic supervision method for construction quality of building engineering, which adopts a dynamic supervision system and includes the following steps: S1. Collect physical status, images and inspection record data of the construction site through IoT sensors, mobile terminals and drones, and clean and align them with the design standard data in the building information model to generate a standardized quality status data stream. S2. Utilize the standardized data stream of S1 to drive and update the building information model in real time, construct a digital twin that is synchronized with the physical project and carries real-time quality information, and perform visualization rendering. S3. Based on the digital twin, perform the following operations and analysis: S3.1 Automatically compare real-time monitoring data with the design specifications embedded in the model to complete immediate compliance judgment; S3.2 Input sensor data, image data and text records into a multimodal machine learning model. The multimodal machine learning model uses time series analysis for trend prediction, uses convolutional neural networks to identify defect features in images, and uses natural language processing technology to parse text records. Finally, it outputs a quality risk probability assessment for a specific part. S3.3 When the risk probability exceeds the threshold, the potential root cause is automatically traced based on the pre-built quality association knowledge graph, and a graded early warning information is generated accordingly. S4. Based on the early warning information, rectification tasks are automatically generated and dispatched. The task information is automatically pushed to the responsible person. The rectification process is tracked and the results are verified before the task status is updated to complete, thus realizing closed-loop management. The verification data after the rectification is completed is used as a new training sample to perform incremental learning on the multimodal machine learning model, so as to achieve continuous self-optimization of model performance.

[0007] The present invention is further configured such that: the Internet of Things sensor in S1 includes one or more of tilt sensors, stress sensors, strain sensors, GPS displacement sensors and temperature and humidity sensors, and image data is collected by drone in S1, specifically including: using drones equipped with high-definition cameras and thermal imaging cameras to periodically and automatically cruise and collect multispectral image data of the construction site, and automatically identify material stacking, support structure deformation and concrete surface defects through image recognition algorithms.

[0008] The present invention is further configured such that: the digital twin constructed in S2 uses different color codes to distinguish the quality status in its visualization rendering, where green represents compliance with specifications, yellow represents slight deviations that need to be observed, and red represents serious deviations that need to be dealt with immediately.

[0009] The present invention is further configured such that the time series analysis in S3.2 specifically involves using an LSTM neural network model to analyze the continuous data collected by the sensor in order to predict the trend of key quality indicators at future time points.

[0010] The present invention is further configured such that: the convolutional neural network in S3.2 is a pre-trained neural network model, which is fine-tuned using a dataset of images of defects in building engineering through transfer learning to identify apparent defects such as cracks, honeycombing, and pitting.

[0011] The present invention is further configured to: use natural language processing technology to parse text records in S3.2, specifically including: extracting key quality entities and their status descriptions from the electronic inspection forms filled out by the supervisors and the text converted from speech using named entity recognition technology.

[0012] The present invention is further configured such that the quality association knowledge graph in S3.3 is a computable network constructed with construction procedures, engineering entities, personnel, machinery, and materials as nodes, and their logical relationships, technological relationships, and causal relationships as edges.

[0013] The present invention is further configured such that the closed-loop management in S4 is specifically implemented as follows: the rectification task order generated by the dynamic supervision system is automatically pushed to the mobile terminal APP of the person in charge. The task order is marked with a geofence. The person in charge needs to upload a photo after rectification. The dynamic supervision system updates the task status to complete only after the image recognition verification is passed.

[0014] The present invention is further configured to include an initialization step S0: in advance, in the building information model, using its API interface or extended attribute set function, add custom quality attribute fields for key components and processes, write the quality acceptance standard threshold and parameter rules into the field, and form a standard rule library that can be called by the dynamic supervision system.

[0015] In summary, the present invention has the following beneficial effects: First, by fusing and analyzing sensor data, images, and text through multimodal machine learning models, the probability of risks such as concrete cracking and support system instability can be predicted in advance, issuing early warnings before problems occur. This allows the management focus to shift from post-event remediation to in-process control and pre-event prevention, significantly reducing the probability of quality accidents and rework losses. Second, by automatically tracing the root causes of problems through a quality-related knowledge graph, and automatically generating clearly targeted rectification tasks, combined with a closed-loop management process on mobile terminals, it achieves accurate task assignment, process tracking, and automatic result verification, significantly reducing personnel communication costs and subjective judgment errors, and making the quality control process more standardized, refined, and efficient. Third, by using the verification data after each rectification as new samples, the multimodal machine learning model is incrementally learned. This allows the dynamic supervision system to continuously learn new defect patterns and more complex causal relationships from practice, thereby continuously optimizing the accuracy and adaptability of its predictive model. Ultimately, the system can continuously evolve along with the project, becoming increasingly intelligent with use, providing a powerful technological impetus for the continuous improvement of construction project quality management. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention; Figure 2 This is a flowchart of S0 in this invention; Figure 3 This is a flowchart of S1 in this invention; Figure 4 This is a flowchart of S3 in this invention; Figure 5 This is a flowchart of S4 in this invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0018] In the description of this invention, it should be noted that the terms "upper", "lower", "inner", "outer", "top / bottom", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0019] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "sleeved / connected," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] A dynamic supervision method for construction quality of building projects, such as Figures 1-2 As shown, a dynamic monitoring system was adopted.

[0021] Initialization step S0: In the building information model, use its API interface or extended attribute set function to add custom quality attribute fields for key components and processes in advance, write the quality acceptance standard threshold and parameter rules into the field, and form a standard rule library that can be called by the dynamic supervision system.

[0022] In one embodiment, initialization step S0 is a prerequisite for building the intelligent decision-making foundation of the entire dynamic supervision system. Its core lies in transforming abstract, textual quality acceptance specifications into structured, machine-invoked digital rules. Specifically, this step uses programming to call the API interface provided by Building Information Modeling (BIM) software, such as Revit, or utilizes its built-in extended attribute set function to deeply empower key components in the model, such as beams, slabs, columns, and walls, and construction procedures such as concrete pouring, formwork installation, and rebar tying. Operators need to pre-define a complete set of quality attributes in the dynamic supervision system, that is, create a series of custom dynamic parameter fields for each type of component or construction procedure. These fields are not the original geometric or physical attributes of the model, but rather control dimensions specifically designed for quality supervision, such as allowable stress limits, current curing time requirements, flatness deviation thresholds, and design strength grades. Subsequently, the specific numerical thresholds and evaluation rules stipulated in the national, industry, and project-specific quality acceptance specifications are written into their corresponding custom attribute fields.

[0023] After embedding the aforementioned attributes, the dynamic supervision system constructs a centrally managed and uniformly invoked standard rule base in memory or a database. This standard rule base is not a static data table, but a computable knowledge system closely linked to BIM model component IDs and possessing a clear logical hierarchy. This standard rule base provides real-time services to other core modules of the dynamic supervision system through standardized web service interfaces or internal functions. In the subsequent S3.1 real-time compliance judgment step, the analysis engine does not need to manually consult paper specifications; instead, it directly calls this standard rule base via API to obtain the specific threshold parameters corresponding to the currently monitored component. This allows for millisecond-level automatic comparison of the real-time monitoring data with the standard values ​​in the rule base, instantly determining whether a violation has occurred. This embedded mechanism ensures the uniqueness, accuracy, and efficiency of the quality assessment standards, providing an indispensable standard basis for the automated and intelligent analysis and decision-making of the entire dynamic supervision system.

[0024] like Figure 1 and Figure 3 As shown, S1 collects physical status, images, and inspection record data of the construction site through IoT sensors, mobile terminals, and drones, and cleans and aligns them with the design standard data in the building information model to generate a standardized quality status data stream. Among them, the IoT sensors include one or more of tilt sensors, stress sensors, strain sensors, GPS displacement sensors, and temperature and humidity sensors. In S1, image data is collected by drones, specifically including: using drones equipped with high-definition cameras and thermal imaging cameras to periodically and automatically cruise and collect multispectral image data of the construction site, and automatically identify material stacking, support structure deformation, and concrete surface defects through image recognition algorithms.

[0025] In one embodiment, step S1 constitutes the perceptual neural network of the entire dynamic supervision system. Its core task is to achieve efficient collection and preliminary integration of massive amounts of raw data from multiple dimensions and sources at the construction site. At the level of physical state data collection, various Internet of Things sensors are deployed on key nodes and important components of the project to form a dense monitoring network.

[0026] Among them, tilt sensors monitor the attitude changes and stability of the support system and structure in real time; stress and strain sensors are embedded inside the concrete or on the surface of the steel structure to directly sense the internal force state and deformation of the components; GPS displacement sensors accurately capture the settlement and horizontal displacement of the entire building or a part thereof; and temperature and humidity sensors are deployed in the concrete curing area to track environmental parameters that affect strength development throughout the process. These sensors are networked wirelessly via LoRa or via wired connections and continuously upload monitoring data at a frequency of minutes.

[0027] Meanwhile, drones equipped with high-definition visible light cameras and thermal imaging cameras regularly and automatically cruise along preset routes, collecting not only high-resolution orthophotos and oblique photography data, but also capturing invisible thermal information such as the distribution of concrete hydration heat and defects in the insulation layer through thermal imaging modules, forming a multispectral image dataset. The onboard computing unit or the data transmitted back to the cloud server then runs an image recognition algorithm based on computer vision, automatically identifying quality and safety hazards such as haphazardly piled materials, abnormal deformation of the support structure, honeycomb pitting and cracks on the concrete surface, achieving intelligent inspection.

[0028] The collected raw data exhibits significant heterogeneity. For example, sensor data consists of timestamped numerical streams, drone images are raster data containing spatial information, and data uploaded by mobile terminals includes structured checklists as well as unstructured on-site photos and audio recordings. To achieve the fusion and analysis of this multi-source data, cleaning and alignment processes were performed. Data cleaning first involves outlier removal, noise filtering, and missing value imputation of the raw data based on a pre-defined standard rule base to ensure data reliability. The crucial alignment process follows: time alignment uses a high-precision network time protocol to provide a unified clock reference for all data sources, and interpolates and synchronizes asynchronously acquired data streams based on timestamps to ensure strict correspondence of all data on the timeline. Spatial alignment utilizes laser point cloud scanning and SLAM technology to precisely register the image frames acquired by the drone, the physical installation location of each sensor, and the global coordinate system of the BIM model. This ensures that every on-site photo and every sensor reading can find its accurate spatial location and corresponding component in the 3D BIM model. Ultimately, the multi-source data, after being cleaned and spatiotemporally aligned, was transformed into a standardized quality status data stream with unified spatiotemporal labels and a standardized format, providing a high-quality data foundation for subsequent construction of digital twins and deep intelligent analysis.

[0029] like Figure 1 As shown, S2 utilizes the standardized data stream from S1 to drive and update the Building Information Model in real time, constructing a digital twin that is synchronized with the physical project and carries real-time quality information, and then performing visualization rendering. The visualization rendering of the digital twin uses different color codes to distinguish quality status, where green represents compliance with specifications, yellow represents minor deviations requiring observation, and red represents serious deviations requiring immediate action.

[0030] In one embodiment, step S2 is the core of achieving deep integration between the physical entity and the virtual model. It maps the standardized data stream generated in S1 to the Building Information Model (BIM) in real time, driving its dynamic updates and thus constructing a digital twin that is synchronized with the progress of the physical project and carries real-time quality information. The dynamic supervision system, through predefined data interfaces and mapping rules, precisely binds sensor readings, image recognition results, etc., to specific component objects in the BIM model, transforming the static design model into a comprehensive information carrier integrating real-time status, historical data, and predictive information.

[0031] To achieve intuitive quality status awareness, the dynamic monitoring system uses real-time status-based visualization to render the twin: the system automatically compares monitored values ​​with embedded design specification thresholds and uses color coding to dynamically render components, where green represents compliance with specifications, yellow represents minor deviations requiring observation, and red represents serious deviations requiring immediate action. This mechanism allows managers to clearly understand the quality health status of the entire construction site in three-dimensional space, quickly locate problems, and achieve a fundamental shift from reading data reports to understanding the quality status, significantly improving supervision efficiency and decision-making intuitiveness.

[0032] like Figure 1 and Figure 4 As shown, S3, based on the digital twin, perform the following operations and analyze them: S3.1 Automatically compare real-time monitoring data with the design specifications embedded in the model to complete immediate compliance judgment.

[0033] This dynamic supervision system first relies on a standard rule library embedded in the BIM model components during the initialization phase. This standard rule library, through API interfaces or extended attribute set functions, adds custom quality attribute fields to each key component and process, containing specific numerical thresholds and logical judgment rules from design drawings and construction acceptance specifications. When the digital twin built by S2 receives the real-time monitoring data stream from S1, the analysis engine immediately and automatically locates the corresponding component object in the BIM model based on the unique identifier in the data packet, and accurately retrieves all relevant acceptance standards and specification parameters from its custom attribute fields. This process is fully automated, avoiding the delays and errors caused by manually reviewing paper documents in traditional methods, laying a solid foundation for real-time judgment.

[0034] S3.2. Sensor data, image data, and text records are input into a multimodal machine learning model. This model uses time series analysis for trend prediction, employs a convolutional neural network to identify defect features in the images, and uses natural language processing (NLP) technology to parse the text records, ultimately outputting a quality risk probability assessment for specific locations. The time series analysis specifically involves using an LSTM neural network model to analyze continuous data collected by sensors to predict trends in key quality indicators at future points in time. The convolutional neural network is a pre-trained neural network model that is fine-tuned using a construction engineering defect image dataset through transfer learning to identify surface defects such as cracks, honeycombing, and pitting. The NLP technology used to parse the text records specifically includes extracting key quality entities and their state descriptions from electronic inspection forms filled out by supervisors and speech-to-text conversions using named entity recognition (NAME) technology.

[0035] In one embodiment, S3.2 achieves deep intelligent diagnosis of engineering quality risks through a multimodal machine learning model. Its core innovation lies in the fusion of data from three different modalities for comprehensive judgment. This model employs an advanced late-stage fusion architecture to process heterogeneous data in parallel: for continuous monitoring data generated by sensors, the dynamic monitoring system uses a Long Short-Term Memory (LSTM) neural network model for time series analysis. The LSTM neural network model, through its unique gating mechanism, can effectively capture the long-term dependencies and dynamic changes of parameters such as stress, displacement, temperature, and humidity, thereby enabling prediction of future trends of key quality indicators and outputting a risk trend vector for future time points. Simultaneously, for high-definition and thermal imaging image data collected by drones, the dynamic monitoring system uses a convolutional neural network optimized through transfer learning for defect feature extraction. This convolutional neural network is first pre-trained on a large general-purpose image dataset to obtain basic feature extraction capabilities, and then fine-tuned using a professionally constructed architectural engineering defect image dataset to specialize in the accurate identification of apparent defects in engineering scenarios, outputting a high-dimensional visual feature vector.

[0036] In terms of text data processing, the dynamic supervision system uses natural language processing (NLP) technology to parse inspection records from mobile terminals. Named entity recognition (NER) technology automatically extracts key quality entities and their status descriptions from electronic forms filled out by supervisors and converted speech text, transforming them into structured text feature vectors. Finally, the model concatenates and reduces the dimensionality of the temporal feature vectors from the LSTM neural network, the visual feature vectors from the convolutional neural network, and the text feature vectors from entity recognition, then inputs them into a fully connected layer classifier for joint inference. This multi-source information complementarity mechanism significantly improves the accuracy of the assessment, ultimately outputting a comprehensive risk probability value for quality defects in a specific engineering location. This provides a quantitative basis for subsequent early warning and decision-making, achieving a transformation from single-point detection to multi-source fusion prediction.

[0037] S3.3 When the risk probability exceeds the threshold, the potential root cause is automatically traced based on the pre-built quality association knowledge graph, and a graded early warning information is generated accordingly. The quality association knowledge graph is a computable network constructed with construction procedures, engineering entities, personnel, machinery, and materials as nodes, and their logical relationships, technological relationships, and causal relationships as edges.

[0038] In one embodiment, the core of step S3.3 lies in constructing a quality-related knowledge graph capable of simulating expert thinking and applying it to intelligent tracing of quality problems. The construction of this quality-related knowledge graph is a dynamic and systematic process of supervision: First, a digital identity system for all project elements is established, using construction procedures, project entities, personnel, machinery, and materials as basic nodes. Then, by analyzing data such as construction organization designs, process specifications, and databases of common quality defects, natural language processing technology is used to automatically or semi-automatically extract various semantic relationships between nodes as edges, including logical relationships, process relationships, and causal relationships. These nodes and edges together constitute a structured, computable, and scalable network, integrating fragmented project information into an organic knowledge system.

[0039] When the risk probability assessed in S3.2 exceeds a preset threshold, the dynamic monitoring system immediately activates an automatic tracing mechanism based on the quality-related knowledge graph. The tracing process begins by setting the risk location as the starting point for the query. The dynamic monitoring system utilizes the efficient traversal capabilities of the graph database, performing multi-hop queries and reasoning along causal relationships, process relationships, and other connections. For example, the dynamic monitoring system automatically explores all nodes related to the risk location: tracing back to the pouring process, the responsible construction team, the batch of building materials used, and the applied curing methods. By calculating the correlation strength between these related nodes and the current quality problem, the dynamic monitoring system can sort potential root causes and output quantitative analysis results according to the weight of the potential causes. Finally, the dynamic monitoring system integrates the risk probability value, the type of root cause traced, and its weight to automatically generate precise early warning information including risk level (e.g., yellow, orange, red), problem location, root cause analysis, and handling suggestions. This provides clear action guidance for subsequent closed-loop management, achieving an intelligent upgrade from problem discovery to root cause location.

[0040] like Figure 1 and Figure 5 As shown, S4 automatically generates and assigns rectification tasks based on early warning information, tracks the rectification process and verifies the results via mobile terminals to achieve closed-loop management, and feeds the result data back to the dynamic supervision system for model optimization. Specifically, the closed-loop management is implemented as follows: the rectification task sheet generated by the dynamic supervision system is automatically pushed to the responsible person's mobile terminal APP. The task sheet includes a geofence marker, and the responsible person needs to upload a photo after rectification. Only after the dynamic supervision system verifies the image through image recognition will the task status be updated to "completed".

[0041] The results data are fed back to the dynamic monitoring system for model optimization. Specifically, the verification data after rectification is used as new training samples to incrementally learn the multimodal machine learning model, so as to achieve continuous self-optimization of model performance.

[0042] In one embodiment, the core of step S4 is to achieve a digital closed-loop management system from risk warning to problem resolution. When the dynamic monitoring system generates tiered warning information, it immediately triggers the automatic creation and dispatch process of rectification task orders. The task order not only includes basic information such as problem description, location of related components, and compliance standards, but also integrates key geofence markers. This task order is sent in real-time to the responsible person's mobile terminal APP via push notification service. The responsible person arrives at the problem area according to the APP's navigation guidance. The mobile terminal automatically senses and confirms that it has entered the preset geofence range via GPS or Bluetooth beacons, thereby activating the task status. After rectification is completed, the responsible person must upload multi-angle photos or short videos that clearly show the post-rectification status through the APP. Upon receiving the media files, the dynamic monitoring system does not rely on manual review but instead activates its built-in image recognition verification module. This module calls a trained convolutional neural network model to compare and analyze the uploaded image data with the initial problem state image and samples of the standard completed state, intelligently determining whether the rectification is truly in place and whether the defects have been eliminated. Only after passing this automatic verification will the dynamic monitoring system officially update the task status to "completed." This entire process is digitally documented, ensuring the authenticity, traceability, and efficiency of the rectification work.

[0043] The endpoint of closed-loop management is not merely the elimination of problems, but more importantly, the transformation of practical results into nutrients for the intelligent evolution of the dynamic supervision system. Verification data after rectification constitutes a feedback loop. For each successful rectification case, the complete data chain—from initial warning information, multimodal risk indicators, and source analysis results to the final rectification measures and verified video evidence—is automatically collected, anonymized, and labeled by the dynamic supervision system, forming a new, high-quality training sample. These training samples are continuously injected into the incremental learning framework of the multimodal machine learning model described in S3.2. This framework avoids costly retraining of the entire model, instead employing strategies such as online learning or small-batch updates to fine-tune the model parameters using new sample data, enabling it to continuously learn new defect patterns, more complex causal relationships, and more effective risk characteristics. This mechanism allows the model to continuously adapt to changes brought about by new materials and processes, continuously optimizing its predictive accuracy and generalization ability, ultimately achieving a qualitative leap in the entire dynamic supervision system from static rule execution to dynamic self-optimization, forming a virtuous cycle of becoming increasingly intelligent with use.

[0044] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic supervision of construction quality in a construction project, characterized in that, A dynamic supervision system is adopted, including the following steps: S1, collecting physical state, image and inspection record data of the construction site through Internet of Things sensors, mobile terminals and unmanned aerial vehicles, and cleaning and aligning with design standard data in the building information model to generate standardized quality state data flow; S2, using the standardized data flow of S1 to drive and update the building information model in real time, constructing a digital twin that is synchronized with the physical engineering and carrying real-time quality information, and performing visual rendering; S3, based on the digital twin, the following operations are performed and analyzed: S3.1, automatically compare real-time monitoring data with design specifications embedded in the model to complete immediate compliance judgment; S3.2, input sensor data, image data and text records into a multi-modal machine learning model, which uses time series analysis for trend prediction, uses a convolutional neural network to identify defect features in images, and uses natural language processing technology to analyze text records, and finally outputs a quality risk probability assessment for a specific part; S3.3, when the risk probability exceeds the threshold, automatically trace the potential root cause based on the pre-constructed quality association knowledge graph, and generate a hierarchical warning information accordingly; S4, automatically generate and distribute rectification tasks according to the warning information, automatically push task information to the responsible person, track the rectification process and verify the results, and update the task status to completed to realize closed-loop management, and the verification data after the rectification is completed as new training samples, the multi-modal machine learning model is incrementally learned to realize continuous self-optimization of the model performance.

2. The construction engineering construction quality dynamic supervision method according to claim 1, characterized in that: The Internet of Things sensors in S1 include one or more of an inclination sensor, a stress sensor, a strain sensor, a GPS displacement sensor, and a temperature and humidity sensor, and the image data in S1 is collected by an unmanned aerial vehicle, specifically including: using an unmanned aerial vehicle equipped with a high-definition camera and a thermal imaging camera to automatically patrol and collect multispectral image data of the construction site regularly, and automatically identifying material stacking, support structure deformation and concrete surface defects through image recognition algorithms.

3. The construction engineering construction quality dynamic supervision method according to claim 1, characterized in that: The digital twin constructed in S2 uses different color coding to distinguish quality states, with green representing compliance with specifications, yellow representing slight deviation for observation, and red representing serious deviation for immediate treatment.

4. The construction engineering construction quality dynamic supervision method according to claim 1, characterized in that: The time series analysis in S3.2 is specifically: using an LSTM neural network model to analyze continuous data collected by sensors to predict the trend of key quality indicators at future time points.

5. The construction engineering construction quality dynamic supervision method according to claim 1, characterized in that: The convolutional neural network in S3.2 is a pre-trained neural network model that is fine-tuned using a building engineering defect image dataset through transfer learning to identify surface defects such as cracks, honeycombs and pitted surfaces.

6. A construction engineering construction quality dynamic supervision method according to claim 5, characterized in that: In S3.2, natural language processing technology is used to analyze text records, specifically including: extracting key quality entities and their state descriptions from electronic inspection forms filled out by supervisors and voice-converted text through named entity recognition technology.

7. The construction engineering construction quality dynamic supervision method according to claim 1, characterized in that: The quality association knowledge graph in S3.3 is a computable network with construction procedures, engineering entities, personnel, machinery and materials as nodes and logical relationships, technological relationships and causal relationships therebetween as edges.

8. The construction engineering construction quality dynamic supervision method according to claim 1, characterized in that: The closed-loop management in S4 is specifically implemented as follows: the rectification task list generated by the dynamic supervision system is automatically pushed to the mobile terminal APP of the person in charge, the task list is attached with a geographic fence identifier, the person in charge needs to upload a photo after rectification, and the dynamic supervision system updates the task state to completed only after image recognition verification.

9. The construction engineering construction quality dynamic supervision method according to claim 1, characterized in that: The method further comprises an initialization step S0: in advance, in the building information model, a self-defined quality attribute field is added to key components and procedures by using an API interface or an extended attribute set function, and quality acceptance specification thresholds and parameter rules are written into the field to form a standard rule library that can be called by the dynamic supervision system.

Images