An external wall thermal insulation construction quality and safety intelligent identification method

Abstract

The application provides a kind of external wall thermal insulation construction quality and safety intelligent identification method, it is related to external wall thermal insulation construction technical field, based on construction site video stream, personnel safety state and key process quality state in external wall thermal insulation construction process are jointly identified, and hierarchical early warning and abnormal trace management are realized, including: collecting construction site video stream;Video data is processed to multi-task identification, including personnel safety identification, plastering area quantitative identification, anchor bolt and anchor nail form state identification and anchor bolt and anchor nail assembly timing identification;Based on the identification result, hierarchical decision and early warning linkage are carried out;And evidence is stored and reviewed for the early warning event management.The beneficial effects of the present application are: unified identification and collaborative determination of multiple abnormalities such as safety helmet wearing, safety clothing wearing, number of workers, plastering area and fixed part construction state, thereby improving the automation degree of abnormal identification on construction site and reducing the dependence on manual inspection.

Description

Technical Field

[0001] This invention relates to the field of external wall insulation construction technology, and in particular to an intelligent identification method for the quality and safety of external wall insulation construction. Background Technology

[0002] Exterior wall insulation construction is a typical high-altitude operation scenario, characterized by an open working environment, close-knit process connections, highly concealed quality, and numerous on-site interference factors. The construction process involves both worker safety management and quality control of key procedures. Specifically, the proper wearing of safety helmets and clothing, and whether the number of people on the suspended platform or work platform is overloaded, directly affect the safety of working at heights. Whether the plastering area of ​​the insulation board meets the standards, whether there are protrusions or insufficient depth after the anchor bolts and nails are installed, and whether the order of installation of the anchor bolts and nails is correct, directly affect the bonding quality and anchoring effect between the insulation board and the base wall. If these abnormalities are not detected in time, they may lead to rework, project delays, and increased costs, or even cause insulation layer detachment, fall injuries, or other construction safety accidents.

[0003] Currently, the management of exterior wall insulation construction sites still mainly relies on manual inspections, spot checks, and post-construction verification. This management method is highly dependent on the experience of on-site management personnel and is easily affected by factors such as observation angle, working distance, construction pace, and environmental interference, making it difficult to achieve continuous monitoring of the entire construction process. Furthermore, it is difficult to make continuous, objective, and quantitative judgments on issues such as the order of anchor bolt and nail installation, whether the plastering area reaches the specified threshold, and local protrusions or insufficient depth of anchor bolts and nails based solely on short-term manual observation. Especially in construction environments with strong light, shadows, obstructions, swaying work platforms, and cluttered backgrounds, traditional manual management methods struggle to detect various anomalies in a timely manner and fail to generate standardized and complete anomaly documentation.

[0004] To address the identification of anomalies in external wall insulation construction, existing technologies have begun to utilize video image analysis to monitor localized anomalies at the construction site. These technologies can monitor plastering anomalies by identifying incomplete plastering and panel movement. They can also identify anomalies in anchor bolts and nails by detecting protrusions or incorrect installation sequences to provide early warnings. Other technologies attempt to integrate plastering anomalies, anchor bolt and nail anomalies, and worker safety status into a single monitoring framework. These technologies provide a foundation for intelligent supervision of the external wall insulation construction process. However, existing technologies generally still focus on single anomalies or single anomaly types. Even with some solutions involving comprehensive monitoring frameworks, shortcomings remain in areas such as precise identification of the local morphological state of anchor bolts and nails, identification of the assembly sequence of anchor bolts and nails, quantitative determination of plastering area, and graded response and record-keeping verification of early warning events. These limitations make it difficult to meet the actual needs of real-time identification, collaborative handling, and full-process traceability management of multiple types of safety and quality anomalies at the construction site.

[0005] Currently, while existing technologies can identify some abnormalities during exterior wall insulation construction, such as abnormal plastering, abnormal anchor bolts and nails, and abnormal personnel safety status, they are still largely focused on single objects or single types of abnormalities. A comprehensive quality and safety identification and early warning solution tailored to the actual management needs of construction sites has not yet been developed. In particular, existing technologies still have shortcomings in areas such as unified identification of personnel safety status and process quality status, precise judgment of anchor bolt and nail abnormalities, quantitative analysis of plastering area, and graded handling and record-keeping verification of early warning events. These shortcomings make it difficult to fully meet the application requirements for automatic identification, graded response, and full-process management of multiple types of abnormalities in exterior wall insulation construction scenarios. Specifically, these shortcomings are reflected in the following aspects: 1. Existing technologies typically detect personnel safety status and construction quality status separately, lacking a technical framework for unified identification, unified judgment, and collaborative handling of multiple types of abnormal events, making it difficult to meet the needs of construction sites for differentiated responses and intelligent management of different types of abnormalities. 2. Anchor bolts and nails used in external wall insulation construction are small in size, and their abnormal conditions usually manifest as local bulges, insufficient depth, or subtle morphological changes. Existing methods mostly use conventional target detection or overall classification for identification, which has relatively limited ability to extract local detailed features of anchor bolts and nails, and is prone to false detection, missed detection, or unstable identification results, making it difficult to achieve stable and accurate identification of the morphological state of anchor bolts and nails; 3. Anomalies in the installation of anchor bolts and anchor nails are not only reflected in the final result, but also in whether the assembly sequence during construction meets the requirements. Existing technologies for identifying such anomalies often remain at the level of judging single-frame image results or simple state comparison, lacking effective analytical capabilities for state changes during continuous construction. Therefore, they are still insufficient in terms of the completeness and accuracy of identifying assembly sequence anomalies. 4. Existing methods for monitoring abnormal plastering are mostly focused on identifying states such as "plastering is full" or "plastering is not full", or judging typical abnormal situations by combining the movement of the slab. However, the quantitative analysis of whether the plastering area has reached the preset ratio threshold is still insufficient. This makes the evaluation of plastering quality still rely on manual experience to a certain extent, which is not conducive to direct correspondence with construction quality standards. 5. Existing technologies mainly focus on anomaly identification and alerts, but do not adequately consider the preservation of image evidence and subsequent review and processing when warning events occur. This makes it difficult to form a complete closed loop for anomaly event tracking and management, thus limiting its effectiveness in tracing responsibility and managing rectification at construction sites.

[0006] How to solve the above-mentioned technical problems is the challenge facing this invention. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an intelligent identification and early warning method for external wall insulation construction scenarios. This method addresses the problem that existing technologies struggle to achieve unified identification, collaborative judgment, and tiered handling of various types of anomalies during external wall insulation construction. The method enables automatic identification, tiered early warning, and traceable management of abnormal events at the construction site.

[0008] The technical solution adopted by this invention to solve its technical problem is as follows: This invention provides an intelligent identification method for the construction quality and safety of external wall insulation, including the following steps: S1. Acquire video streams from the construction site; S2. Perform multi-task recognition processing on the collected video data to obtain personnel safety status information and construction process quality status information; S3. Generate abnormal event information based on the identification results; S4. Classify and make decisions on abnormal event information and generate early warning events; S5. Conduct evidence retention and review management for early warning events to achieve traceable management of abnormal events at the construction site.

[0009] In step S1, video is acquired through a video acquisition module. The video acquisition module is used to deploy camera equipment at the construction site to acquire video streams of the construction site and transmit the video streams to an edge computing terminal or server. The video acquisition module includes several camera devices deployed at the construction site, preferably including: Monitoring cameras are used to record the working conditions of personnel on suspended platforms or work platforms, as well as to monitor the areas where insulation boards are plastered and the areas where anchor bolts and anchor nails are installed.

[0010] In step S2, personnel safety status information is obtained through the personnel safety identification module, and construction process quality status information is obtained through the plastering area quantification identification module, anchor bolt and anchor nail morphology status identification module, and anchor bolt and anchor nail assembly sequence identification module.

[0011] The personnel safety identification module is used to detect targets such as workers, safety helmets and safety clothing in video frames, and to determine the wearing status based on the spatial correspondence between the personnel and the safety helmets and safety clothing. At the same time, it counts the number of personnel in the work area. When any of the following situations occur, the system determines it as a work stoppage-level abnormal event: not wearing a safety helmet, not wearing a safety clothing, or the number of people on the work platform exceeds a preset threshold. Specifically, the personnel safety identification module uses an object detection model to identify the input video frames. The detected objects include at least personnel, safety helmets, and safety clothing, and obtains the bounding box position and category information of each object. In one embodiment, the object detection model uses the YOLOv5 deep learning-based object detection network to detect the video frames and outputs the bounding box coordinates, category label, and confidence information of each object.

[0012] During the identification of protective equipment wearing status, the system establishes a corresponding human body determination area for each person target. Preferably, the head and torso areas can be determined based on the positioning results of key human body parts; or, if key body part positioning is not used, the head and torso areas can be divided according to a preset ratio based on the person detection frame.

[0013] After obtaining the detection results, the personnel and their protective equipment are matched through spatial correspondence. Specifically, the safety helmet target is searched in the head area or its adjacent area, and the safety clothing target is searched in the torso area or its adjacent area.

[0014] The personnel count is based on the target detection results and target tracking methods. Specifically, personnel targets are detected in each frame of the image, and the same person is continuously tracked by a cross-frame association algorithm to avoid double counting.

[0015] The plastering area quantification and recognition module is used to extract the area of ​​the insulation board to be constructed. The system extracts the area of ​​the insulation board to be constructed from the video frame, segments the plastering coverage area, determines whether the plastering area of ​​the insulation board reaches a preset ratio, calculates the ratio of the plastering area to the total area of ​​the insulation board to obtain the plastering coverage rate, and generates rectification-level abnormal event information when the plastering coverage rate is lower than a preset threshold. Let the plastering area be... The total area of ​​the insulation board is Then the coverage rate of the plastering It can be represented as: ,when When the area is less than the preset threshold, it is determined that the plastering area is substandard, and a rectification-level early warning candidate event is generated. In this way, the quality of plastering can be quantitatively judged.

[0016] Specifically, the input video stream is first parsed frame by frame to obtain image frames arranged in chronological order, and the image frames are preprocessed to reduce the impact of environmental interference on the recognition results. The preprocessing operations include, but are not limited to, image size normalization, brightness normalization, noise suppression, and region of interest cropping. After obtaining the preprocessed image frame, the insulation board area and the mortar-covered area in the image are segmented and extracted. The insulation board area is used to define the spatial range of the object to be analyzed, and the mortar-covered area is used to characterize the area of ​​the insulation board surface that has been covered by mortar. The segmentation process can be implemented based on an image segmentation model. By performing pixel-level classification on the image, the image is divided into the insulation board area, the mortar-covered area, and the background area. After extracting the insulation board area and the grout-covered area, an ungrouted area is further constructed. The ungrouted area is the part of the insulation board area that is not covered by grout, which is used to characterize the grout leakage on the board surface.

[0017] The unplastered area can be obtained by performing a difference operation between the insulation board area and the plastered area.

[0018] Based on this, the overall grout coverage rate is calculated according to the area relationship between the grout coverage area and the insulation board area, and the quality is judged in combination with the distribution of ungrouted areas in key areas, thereby realizing the effective extraction of grout areas and subsequent quality analysis.

[0019] The anchor bolt and anchor nail morphology recognition module is used to detect the position of anchor bolts and anchor nails to obtain candidate areas, and to perform local analysis on the candidate areas to make fine-grained judgments on the morphology of anchor bolts and anchor nails. When anchor bolts and anchor nails have protrusions or insufficient depth, rectification-level abnormal event information is generated. Specifically, firstly, the anchor bolts and anchor nails in the image are located using a target detection model to obtain the corresponding target areas; then, local feature analysis is performed on the target areas to determine their installation status.

[0020] By analyzing the contour morphology of the target area of ​​the anchor bolt or anchor nail and its relative positional relationship with the surface of the insulation board, it can be determined whether there is a protrusion. Specifically, when the outer contour of the anchor bolt or anchor nail is detected to be significantly protruding from the surface area of ​​the insulation board, and its boundary shows obvious protrusion features in the image, it can be determined that there is a protrusion anomaly.

[0021] By analyzing the grayscale distribution or illumination change characteristics of the anchor bolt or anchor nail area, the installation depth can be determined. Since protruding or unflattened anchor bolts form obvious brightness changes or shadow features in the image, by analyzing the brightness gradient or grayscale difference of the target area, it can be determined whether there is insufficient depth or unflattened condition.

[0022] By using the analysis method based on local image features, the installation status of anchor bolts and anchor nails can be identified, thereby determining whether there are any abnormalities such as protrusions, insufficient depth, or improper installation.

[0023] The anchor bolt and anchor nail assembly sequence recognition module is used to analyze and judge the assembly sequence of anchor bolts and anchor nails based on the relationship between the state changes of anchor bolts and anchor nails before and after construction in continuous video frames. When the assembly sequence does not meet the preset construction requirements, rectification-level abnormal event information is generated. The assembly sequence recognition of anchor bolts and anchor nails is achieved based on the target detection results and time series analysis in the video sequence. Specifically, the construction video is first analyzed frame by frame to obtain the image frame sequence arranged in time order, and target detection is performed on each frame to identify anchor bolt and anchor nail targets and obtain the corresponding detection result sequence.

[0024] The above method enables the identification of the assembly sequence of anchor bolts and anchor nails based on video sequences, and can automatically determine the situation where the assembly sequence does not meet the preset construction requirements, thereby providing a basis for subsequent graded early warning.

[0025] In step S4, a hierarchical decision-making and early warning linkage module is used to make hierarchical decisions and generate early warning events. The hierarchical decision-making and early warning linkage module is used to fuse and determine abnormal event information. Specifically, the hierarchical decision-making and early warning linkage module receives abnormal event information output by the personnel safety identification module, the plastering area quantification identification module, the fastener morphology and status identification module, and the fastener assembly sequence identification module, and classifies and processes them according to the abnormality type.

[0026] When the abnormal event is a personnel safety abnormality, a work stoppage level warning is output and a work stoppage prompt and management warning push are executed. When the abnormal event is a construction process quality abnormality, such as the plastering area not meeting the standard, the anchor bolt and anchor nail shape abnormality, or the anchor bolt and anchor nail assembly sequence abnormality, a rectification level warning is output and on-site audible and visual alarms and rectification prompts are executed. The hierarchical decision-making and early warning linkage module sends the early warning results to the evidence retention and review module for further processing.

[0027] In step S5, the evidence retention and verification module manages the evidence retention and verification of the early warning event. The evidence retention and verification module is used to automatically capture and save the corresponding video footage of the early warning event. The video footage contains information about the work area when the abnormality occurs, so as to identify the suspended platform or construction location corresponding to the abnormal event.

[0028] Specifically, when the hierarchical decision-making and early warning linkage module generates an early warning event, the system automatically captures the corresponding video frame as an alarm screenshot based on the trigger time of the early warning event; It includes a model training and scene fine-tuning module, which is used to train and optimize the recognition model corresponding to the multi-task recognition processing based on sample data collected at the construction site, and to iteratively adjust the data under different construction environments through transfer learning and data augmentation to improve the recognition accuracy and stability in complex lighting, occlusion and dynamic construction scenarios. The model training is based on video data collected at the construction site to build a training dataset. Specifically, image frames are extracted from the construction video, and the personnel, safety helmets, safety clothes, insulation boards, plastering areas, anchor bolts and anchor nails in the images are labeled to form training sample data for target detection or image segmentation. The labeled data includes target category information and its corresponding spatial location information. For target detection tasks, the labels are in the form of bounding boxes; for image segmentation tasks, the labels are in the form of pixel-level region masks. The above labeling methods provide supervision information for subsequent model training. During model training, a deep learning-based object detection model and an image segmentation model are used for training. The model parameters are iteratively optimized to make the model output results gradually approach the labeled results. The difference between the model prediction results and the labeled results is measured by the loss function, and the model parameters are updated based on the gradient descent method to complete the model training process.

[0029] Transfer learning is based on existing pre-trained models, which means using the model parameters trained on a general dataset as initial parameters and then retraining them with construction site data to improve the model's recognition performance in specific scenarios. After the model is deployed, in order to adapt to the environmental differences between different construction sites, the model is continuously optimized or fine-tuned based on the new data collected on site. Specifically, sample data in the actual operation process is collected regularly, and the identification results are manually reviewed and re-labeled. The updated data is used for model retraining or parameter adjustment, thereby realizing the continuous optimization of the model in different environments.

[0030] Through the above training and optimization mechanisms, each recognition model can adapt to changes in lighting, occlusion, and dynamic operating conditions in different construction environments, thereby improving the accuracy and stability of the overall recognition results of the system.

[0031] The beneficial effects of this invention are as follows: 1. The method of the present invention performs multi-task recognition processing on the video stream of the construction site, combining personnel safety recognition with construction quality recognition, and realizes unified recognition and collaborative judgment of multiple anomalies such as wearing safety helmets, wearing safety clothing, number of workers, plastering area, and construction status of fasteners, thereby improving the automation level of anomaly recognition at the construction site and reducing the reliance on manual inspection. 2. This invention constructs a hierarchical decision-making and early warning linkage mechanism, classifying abnormal events into work stoppage-level early warnings and rectification-level early warnings, and triggering different processing methods such as work stoppage prompts and rectification prompts respectively, so that personnel safety abnormalities and construction quality abnormalities can be responded to in a classified manner, thereby improving the pertinence and response efficiency of risk handling at the construction site. 3. This invention detects the fastener target and combines it with local area feature analysis to identify morphological anomalies such as protrusions and insufficient depth of the fastener, thereby enhancing the ability to identify fine-grained quality anomalies during construction and improving the completeness of anomaly detection. 4. This invention achieves quantitative judgment of plastering construction quality by dividing the plastering area of ​​the insulation board and calculating the plastering coverage ratio, thus transforming the construction quality judgment from experience-based judgment to a data-based evaluation method, thereby improving the consistency and comparability of quality assessment. 5. This invention automatically captures images or video clips of early warning events and establishes event records, associating and storing abnormal information with corresponding work areas and target objects to achieve traceability management of abnormal events, thereby facilitating subsequent verification and review and improving the traceability of construction management. 6. This invention trains and fine-tunes the recognition model to adapt to changes in lighting, occlusion, and dynamic interference in different construction environments, thereby improving the stability and applicability of the system in actual construction scenarios. Attached Figure Description

[0032] Figure 1 This is the overall flowchart of the intelligent identification and early warning method for external wall insulation construction of the present invention.

[0033] Figure 2 This is a flowchart illustrating the hierarchical decision-making, early warning linkage, and evidence retention process of this invention. Detailed Implementation

[0034] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0035] Example 1 See Figures 1 to 2 As shown, this embodiment takes the exterior wall insulation construction site of a high-rise residential project as an example. This embodiment is a method for intelligent identification of exterior wall insulation construction quality and safety, including the following steps: S1. Acquire video stream from the construction site: Video is acquired via a video acquisition module, which includes several cameras deployed at the construction site. Two video acquisition devices and one edge computing server are deployed at the site. The first camera, installed on the front side of the work platform, captures video of personnel working on the suspended platform or work platform, as well as video of the anchor bolt and anchor nail installation area after the insulation board is installed on the wall. This camera captures video of the personnel working on the platform and the anchor bolt and anchor nail installation process. The second camera, installed above the grouting work platform, captures video of the insulation board plastering work area. The edge computing server is equipped with a graphics processor and uses a deep learning framework to parse and infer the video stream. The cameras are Dahua 4MP high-definition P40A2-WT [3.6mm] cameras, installed at a height of 2.5m and tilted at a 45° angle.

[0036] S2. Perform multi-task recognition processing on the collected video data to obtain personnel safety status information and construction process quality status information: The personnel safety status information is obtained through the personnel safety identification module, and the construction process quality status information is obtained through the plastering area quantitative identification module, the anchor bolt and anchor nail morphology status identification module, and the anchor bolt and anchor nail assembly sequence identification module. S3. Generate abnormal event information based on the identification results: The personnel safety identification module is used to detect targets such as workers, safety helmets, and safety clothing in video frames, namely whether workers are wearing safety helmets and safety clothing in accordance with regulations, and whether the number of people on the work platform exceeds the limit. It also determines the wearing status based on the spatial correspondence between personnel and safety helmets and safety clothing, and counts the number of people in the work area. When any of the following situations occur, the system determines it as a work stoppage-level abnormal event: not wearing a safety helmet, not wearing a safety clothing, or the number of people on the work platform exceeds a preset threshold. Specifically, the personnel safety identification module uses an object detection model to identify the input video frames. The detected objects include personnel, safety helmets, and safety clothing, and obtain the bounding box position and category information of each object. The object detection model uses the YOLOv5 deep learning-based object detection network to detect the video frames and outputs the bounding box coordinates, category label, and confidence information of each object.

[0037] During the process of identifying the wearing status of protective equipment, the system establishes a corresponding human body determination area for each person target. Preferably, the head area and torso area are determined based on the positioning results of key human body parts; or, if key body parts positioning is not used, the head area and torso area are divided according to a preset ratio based on the personnel detection frame.

[0038] After obtaining the detection results, the personnel and their protective equipment are matched through spatial correspondence. Specifically, the safety helmet target is searched in the head area or its adjacent area, and the safety clothing target is searched in the torso area or its adjacent area.

[0039] In this embodiment, matching is determined based on the following rules: 1. Positional constraints: The safety helmet target is located within the head area and overlaps with the personnel detection frame in the horizontal direction; 2. Distance constraint: The distance between the target center point of the safety helmet and the center position of the head area is less than a preset threshold; 3. Overlap Constraint: The intersection-to-union ratio or overlap ratio between the helmet target and the head area is greater than a preset threshold; When at least one or more of the above constraints are met, the safety helmet is determined to match the corresponding person, thus confirming that the person is wearing a safety helmet; otherwise, it is determined that the person is not wearing a safety helmet. The same method is used to identify safety clothing: by determining whether the safety clothing target is located within the torso area or has a spatial overlap or containment relationship with it, it is determined whether the person is wearing safety clothing.

[0040] The personnel count is based on the target detection results and target tracking methods. Specifically, personnel targets are detected in each frame of the image, and the same person is continuously tracked by a cross-frame association algorithm to avoid double counting.

[0041] In this embodiment, a unique identifier ID is assigned to each person using a target tracking algorithm, and cross-frame matching is performed based on the target's location, size, and appearance features.

[0042] During the personnel counting process on the work platform, the system pre-determines the monitoring area of ​​the suspended platform or work platform, and determines whether the person is located within the work platform area based on the center point, bottom position, or positional relationship between the detection frame and the work platform area.

[0043] Personnel who enter the work platform area and continuously reach a preset number of frames or a preset time threshold are recorded as valid personnel present. For the same personnel trajectory identifier, it is counted only once during the period when it stays in the work platform area, thereby avoiding duplicate counting caused by repeated detection of adjacent video frames, short-term shaking of personnel, or restoration of occlusion.

[0044] When personnel leave the work platform area and the preset disappearance time is exceeded, the system cancels the presence status corresponding to the trajectory identifier; when they re-enter, the count is restarted.

[0045] By combining the above-mentioned target detection and target tracking methods, the number of personnel at the construction site can be accurately counted, and the problem of double counting can be effectively avoided. The plastering area quantification and identification module is used to extract the area of ​​the insulation board to be constructed. The system extracts the area of ​​the insulation board to be constructed from the video frame, segments the plastering coverage area, and determines whether the plastering area of ​​the insulation board reaches a preset ratio (90% in this embodiment). The ratio of the plastering area to the total area of ​​the insulation board is calculated to obtain the plastering coverage rate. When the plastering coverage rate is lower than a preset threshold, a rectification-level abnormal event information is generated. Let the plastering area be... The total area of ​​the insulation board is Then the coverage rate of the plastering It can be represented as: ,when When the area is less than the preset threshold, it is determined that the plastering area is substandard, and a rectification-level early warning candidate event is generated. In this way, the quality of plastering can be quantitatively judged.

[0046] Specifically, the input video stream is first parsed frame by frame to obtain image frames arranged in chronological order, and the image frames are preprocessed to reduce the impact of environmental interference on the recognition results. The preprocessing operations include, but are not limited to, image size normalization, brightness normalization, noise suppression, and region of interest cropping. After obtaining the preprocessed image frame, the insulation board area and the mortar-covered area in the image are segmented and extracted. The insulation board area is used to define the spatial range of the object to be analyzed, and the mortar-covered area is used to characterize the area of ​​the insulation board surface that has been covered by mortar. The segmentation process can be implemented based on an image segmentation model. By performing pixel-level classification on the image, the image is divided into the insulation board area, the mortar-covered area, and the background area. After extracting the insulation board area and the mortar-covered area, an un-mortared area is further constructed. The un-mortared area is the part of the insulation board area that is not covered by mortar, and is used to characterize the mortar leakage on the board surface. In this embodiment, the unplastered area can be obtained by performing a difference operation between the insulation board area and the plastered area: In this embodiment, an unplastered area is constructed based on the extracted insulation board area and the grout-covered area. Specifically, the insulation board area is first represented as a first binary mask image, where the pixel value of the insulation board is set to 1 and the other areas are set to 0; the grout-covered area is represented as a second binary mask image, where the pixel value of the grout-covered area is set to 1 and the other areas are set to 0.

[0047] Subsequently, a set difference operation is performed on the first binary mask image and the second binary mask image to obtain the mask for the unpainted area, i.e.:

[0048] in, Indicates the masking of the insulation board area. This indicates a mask covering the area to be plastered. This represents the mask for the unplastered area. In image processing, the difference operation can be implemented through pixel-by-pixel logical operations. That is, the pixel positions within the insulation board area that are not covered by plaster are assigned a value of 1, and the remaining positions are assigned a value of 0, thus forming a binary representation of the unplastered area, which is used for subsequent area statistics and local defect analysis.

[0049] Based on this, the overall grout coverage rate is calculated according to the area relationship between the grout coverage area and the insulation board area, and the quality is judged in combination with the distribution of ungrouted areas in key areas, thereby realizing the effective extraction of grout areas and subsequent quality analysis.

[0050] The anchor bolt and anchor nail morphology recognition module is used to detect the position of anchor bolts and anchor nails to obtain candidate areas, and to perform local analysis on the candidate areas to make fine-grained judgments on the morphology of anchor bolts and anchor nails. When anchor bolts and anchor nails have protrusions or insufficient depth, rectification-level abnormal event information is generated. Specifically, firstly, the anchor bolts and anchor nails in the image are located using a target detection model to obtain the corresponding target areas; then, local feature analysis is performed on the target areas to determine their installation status.

[0051] By analyzing the contour morphology of the target area of ​​the anchor bolt or anchor nail and its relative positional relationship with the surface of the insulation board, it can be determined whether there is a protrusion. Specifically, when the outer contour of the anchor bolt or anchor nail is detected to be significantly protruding from the surface area of ​​the insulation board, and its boundary shows obvious protrusion features in the image, it can be determined that there is a protrusion anomaly.

[0052] By analyzing the grayscale distribution or illumination change characteristics of the anchor bolt or anchor nail area, the installation depth can be determined. Since protruding or unflattened anchor bolts form obvious brightness changes or shadow features in the image, by analyzing the brightness gradient or grayscale difference of the target area, it can be determined whether there is insufficient depth or unflattened condition.

[0053] In this embodiment, the analysis is based on the boundary continuity between the anchor bolt or anchor nail and the surrounding insulation board area. When there is a clear discontinuity or boundary discontinuity between the anchor bolt or anchor nail area and the insulation board surface, it can be determined as an installation abnormality.

[0054] By using the analysis method based on local image features, the installation status of anchor bolts and anchor nails can be identified, thereby determining whether there are any abnormalities such as protrusions, insufficient depth, or improper installation.

[0055] The anchor bolt and anchor nail assembly sequence recognition module is used to analyze and judge the assembly sequence of anchor bolts and anchor nails based on the relationship between the state changes of anchor bolts and anchor nails before and after construction in continuous video frames. When the assembly sequence does not meet the preset construction requirements, rectification-level abnormal event information is generated. The assembly sequence recognition of anchor bolts and anchor nails is achieved based on the target detection results and time series analysis in the video sequence. Specifically, the construction video is first analyzed frame by frame to obtain the image frame sequence arranged in time order, and target detection is performed on each frame to identify anchor bolt and anchor nail targets and obtain the corresponding detection result sequence.

[0056] In this embodiment, for each frame of image, the target detection results include the category information and spatial location information of anchor bolts and anchor nails, thereby forming a detection result sequence arranged in time. Based on the detection result sequence, a corresponding event sequence is constructed to characterize the occurrence of each key component in the time dimension. Furthermore, time series analysis is performed on the event sequence to convert the anchor bolt detection event and the anchor nail detection event into corresponding time-series signals. When an anchor bolt is detected in a certain frame or several consecutive frames, it is determined that an anchor bolt event has occurred; when an anchor nail is detected, it is determined that an anchor nail event has occurred. In this way, the spatial recognition result is converted into an event sequence representation in the time dimension. Based on this, the system performs timing determination on the event sequence according to the preset assembly process logic. The assembly process logic includes at least the following: anchor bolt installation should be completed before anchor nail installation, that is, anchor nail operation can only be carried out after the anchor bolt is in the installed position. The system judges whether the assembly timing meets the preset requirements by analyzing the order of event occurrence and the relationship between their state changes.

[0057] The above method enables the identification of the assembly sequence of anchor bolts and anchor nails based on video sequences, and can automatically determine the situation where the assembly sequence does not meet the preset construction requirements, thereby providing a basis for subsequent graded early warning.

[0058] S4. Perform hierarchical decision-making on abnormal event information and generate early warning events: The hierarchical decision-making and early warning linkage module performs hierarchical decision-making and generates early warning events. The hierarchical decision-making and early warning linkage module is used to fuse and determine abnormal event information. Specifically, the hierarchical decision-making and early warning linkage module receives abnormal event information output by the personnel safety identification module, the plastering area quantification identification module, the fastener morphology and status identification module, and the fastener assembly sequence identification module, and classifies and processes them according to the abnormality type.

[0059] When the abnormal event is a personnel safety abnormality, a work stoppage level warning is output and a work stoppage prompt and management warning push are executed. When the abnormal event is a construction process quality abnormality, such as the plastering area not meeting the standard, the anchor bolt and anchor nail shape abnormality, or the anchor bolt and anchor nail assembly sequence abnormality, a rectification level warning is output and on-site audible and visual alarms and rectification prompts are executed. In this embodiment, when there is only a single abnormal event at the construction site, the system directly generates a corresponding early warning event based on the level of the abnormal event. When multiple abnormal events occur at the construction site, the hierarchical decision-making and early warning linkage module performs fusion processing based on the target object corresponding to the abnormal event, the work area to which it belongs, and the time of occurrence. Specifically: (1) When multiple abnormal events belong to the same warning level and correspond to the same operator, the same construction object or the same work area, the multiple abnormal events shall be merged into the same warning event, and the multiple abnormal triggering reasons shall be recorded in the warning event. (2) When multiple abnormal events belong to the same warning level, but correspond to different workers, different construction objects or different work areas, separate warning events are generated respectively; The same worker is determined by the unique identifier ID of the worker target, and the same construction object or work area is determined by the target detection area or the preset area range.

[0060] When both a work stoppage warning and a rectification warning exist, the system will prioritize outputting the work stoppage warning and proceed to the work stoppage handling stage, while simultaneously recording the corresponding rectification-level anomaly. If the rectification-level anomaly persists after the work stoppage anomaly is resolved, then the system will output the rectification warning and proceed to the rectification handling stage. For the same abnormal event, the system determines whether it persists based on the detection results of consecutive frames or a time window. That is, when the abnormality persists for several consecutive frames or within a preset time, the same warning event is not generated repeatedly; when the abnormality disappears after a preset time threshold, the warning event is terminated.

[0061] The hierarchical decision-making and early warning linkage module sends the early warning results to the evidence retention and review module for further processing. S5. Implement evidence retention and review management for early warning events: The evidence retention and review module manages the retention and review of evidence for early warning events. The evidence retention and review module is used to automatically capture and save the corresponding video footage of the early warning event. The video footage includes information about the work area when the anomaly occurred.

[0062] Specifically, when the hierarchical decision-making and early warning linkage module generates an early warning event, the system automatically captures the corresponding video frame as an alarm screenshot based on the trigger time of the early warning event; In this embodiment, a preset time window is extended forward and backward according to the triggering time to extract the corresponding video clips or key frame sequences as supplementary evidence data.

[0063] The system generates a unique event identifier for each warning event and associates and stores the screenshots, video clips and related information with the event identifier, thereby achieving a one-to-one correspondence between warning events and evidence data.

[0064] For each warning event, the system establishes a corresponding event record, which includes the warning level, anomaly type, occurrence time, work area information, and corresponding screenshot or video information; it also includes the target object identifier, triggering reason, and event status information.

[0065] When the same early warning event contains multiple triggering reasons, the system creates the same event record for the early warning event and saves the multiple triggering reasons in the record; when multiple early warning events are independent of each other and correspond to different operators, different construction objects or different work areas, the system creates independent event records for each.

[0066] During the review process, managers retrieve screenshots, video clips, and event records of the corresponding warning events through the management platform to verify whether the anomaly actually occurred, whether the anomaly category was accurate, and whether on-site rectification had been completed.

[0067] The system updates the event status based on the review results, including unprocessed, processing, and processed status; when it is confirmed that the anomaly has been eliminated or rectified, the corresponding event status is updated to processed, thus forming a closed-loop management of anomaly events.

[0068] By employing the above methods, automatic recording, data association, storage, and verification of early warning events can be achieved, thereby improving the traceability and closed-loop management capabilities of abnormal events at construction sites.

[0069] This embodiment also includes a model training and scene fine-tuning module. The model training and scene fine-tuning module is used to train and optimize the recognition model corresponding to the multi-task recognition processing based on sample data collected at the construction site, and to iteratively adjust the data under different construction environments through transfer learning and data augmentation to improve the recognition accuracy and stability in complex lighting, occlusion and dynamic construction scenarios. The model training is based on video data collected at the construction site to build a training dataset. Specifically, image frames are extracted from the construction video, and the personnel, safety helmets, safety clothes, insulation boards, plastering areas, anchor bolts and anchor nails in the images are labeled to form training sample data for target detection or image segmentation. The labeled data includes target category information and its corresponding spatial location information. For target detection tasks, the labels are in the form of bounding boxes; for image segmentation tasks, the labels are in the form of pixel-level region masks. The above labeling methods provide supervision information for subsequent model training. During model training, a deep learning-based object detection model and an image segmentation model are used for training. The model parameters are iteratively optimized to make the model output results gradually approach the labeled results. The difference between the model prediction results and the labeled results is measured by the loss function, and the model parameters are updated based on the gradient descent method to complete the model training process.

[0070] In this embodiment, the training data is augmented, including but not limited to one or more of brightness adjustment, contrast adjustment, image flipping, scale change and noise perturbation, to enhance the robustness of the model to changes in illumination, occlusion and shooting angle.

[0071] Transfer learning is based on existing pre-trained models, which means using the model parameters trained on a general dataset as initial parameters and then retraining them with construction site data to improve the model's recognition performance in specific scenarios. After the model is deployed, in order to adapt to the environmental differences between different construction sites, the model is continuously optimized or fine-tuned based on the new data collected on site. Specifically, sample data in the actual operation process is collected regularly, and the identification results are manually reviewed and re-labeled. The updated data is used for model retraining or parameter adjustment, thereby realizing the continuous optimization of the model in different environments.

[0072] Through the above training and optimization mechanisms, each recognition model can adapt to changes in lighting, occlusion, and dynamic operating conditions in different construction environments, thereby improving the accuracy and stability of the overall recognition results of the system.

[0073] Example 2 Based on Example 1, in order to improve the stability of the identification of the installation status of anchor bolts and anchor nails, a continuous frame analysis method is combined. That is, when the abnormal features mentioned in Example 1 continue to exist in several consecutive frames, it is determined to be a morphological abnormality.

[0074] Example 3 Based on Example 1, in identifying the installation sequence of anchor bolts and anchor nails, a state transition model is constructed to model the assembly process. The state transition model is used to describe the state change relationship at each stage of construction and drives state updates based on event sequences. When an anchor bolt event is detected, the state transitions from the initial state to the anchor bolt installation state; when an anchor nail event is detected in the anchor bolt installation state, the state transitions to the anchor nail installation state; when an anchor nail event is detected but the anchor bolt installation completion state is not detected, it is determined to be an assembly sequence abnormality.

[0075] Example 4 Based on the implementation examples, in order to reduce the impact of false detections in a single frame or short-term occlusion on the timing determination results, a continuous frame confirmation mechanism is adopted. That is, the corresponding event is determined to be valid only when the anchor bolt event or anchor nail event is continuously established in several consecutive frames, thereby improving the stability of timing recognition.

[0076] The technical features of this invention not described can be implemented by or using existing technology, and will not be repeated here. Of course, the above description is not a limitation of this invention, and this invention is not limited to the examples above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention should also be within the protection scope of this invention.

Claims

1. A method for intelligent identification of construction quality and safety of external wall insulation, characterized in that, Includes the following steps: S1. Acquire video streams from the construction site; S2. Perform multi-task recognition processing on the collected video data to obtain personnel safety status information and construction process quality status information; S3. Generate abnormal event information based on the identification results; S4. Classify and make decisions on abnormal event information and generate early warning events; S5. Conduct evidence retention and review management for early warning events.

2. The intelligent identification method for construction quality and safety of external wall insulation according to claim 1, characterized in that, In step S1, video is acquired through a video acquisition module, which includes several camera devices deployed at the construction site.

3. The intelligent identification method for construction quality and safety of external wall insulation according to claim 2, characterized in that, In step S2, personnel safety status information is obtained through the personnel safety identification module, and construction process quality status information is obtained through the plastering area quantification identification module, anchor bolt and anchor nail morphology status identification module, and anchor bolt and anchor nail assembly sequence identification module.

4. The intelligent identification method for construction quality and safety of external wall insulation according to claim 3, characterized in that, The personnel safety identification module is used to detect targets such as workers, safety helmets, and safety clothing in video frames. When any of the following situations occur, the system determines it as a work stoppage-level abnormal event: not wearing a safety helmet, not wearing safety clothing, or the number of people on the work platform exceeds a preset threshold.

5. The intelligent identification method for construction quality and safety of external wall insulation according to claim 4, characterized in that, The plastering area quantification and identification module is used to extract the area of ​​the insulation board to be constructed. The system extracts the area of ​​the insulation board to be constructed from the video frame, segments the plastering coverage area, determines whether the plastering area of ​​the insulation board reaches the preset ratio, calculates the ratio of the plastering area to the total area of ​​the insulation board to obtain the plastering coverage rate, and generates rectification-level abnormal event information when the plastering coverage rate is lower than the preset threshold.

6. The intelligent identification method for construction quality and safety of external wall insulation according to claim 5, characterized in that, The anchor bolt and anchor nail morphology recognition module is used to detect the position of anchor bolts and anchor nails to obtain candidate areas, and to perform local analysis on the candidate areas to determine the morphology of anchor bolts and anchor nails. When anchor bolts and anchor nails have protrusions or insufficient depth, rectification-level abnormal event information is generated.

7. The intelligent identification method for construction quality and safety of external wall insulation according to claim 6, characterized in that, The anchor bolt and anchor nail assembly sequence recognition module is used to analyze and judge the assembly sequence of anchor bolts and anchor nails based on the relationship between the state changes of anchor bolts and anchor nails before and after construction in continuous video frames. When the assembly sequence does not meet the preset construction requirements, rectification-level abnormal event information is generated.

8. The intelligent identification method for construction quality and safety of external wall insulation according to claim 7, characterized in that, In step S4, a hierarchical decision-making and early warning linkage module is used to make hierarchical decisions and generate early warning events. The hierarchical decision-making and early warning linkage module is used to fuse and determine abnormal event information. When the abnormal event is a personnel safety-related abnormality, a work stoppage-level warning will be issued and a work stoppage notification will be sent to the management terminal. When the abnormal event is a construction process quality abnormality, a rectification-level warning will be issued and an on-site audible and visual alarm will be sent and a rectification notification will be sent. The hierarchical decision-making and early warning linkage module sends the early warning results to the evidence retention and review module for further processing.

9. The intelligent identification method for construction quality and safety of external wall insulation according to claim 8, characterized in that, In step S5, the evidence retention and review module manages the evidence retention and review of the early warning event. The evidence retention and review module is used to automatically capture and save the corresponding video footage of the early warning event. The video footage contains information about the work area when the abnormality occurs.

10. The intelligent identification method for construction quality and safety of external wall insulation according to claim 1, characterized in that, It includes a model training and scene fine-tuning module, which is used to train and optimize the recognition model corresponding to the multi-task recognition processing based on sample data collected at the construction site, and to iteratively adjust the data under different construction environments through transfer learning and data augmentation.

Images