Risk monitoring prediction analysis method for intelligent construction site and image recognition

Abstract

The invention provides a risk monitoring prediction analysis method for an intelligent construction site and image recognition, and belongs to the technical field of data processing, and the method specifically comprises the steps: employing a target construction site, based on the recognition matching data of different construction sites in a professional scene of construction quality, and the recognition association data in other professional scenes; the method comprises the following steps: determining a construction site focusing on construction quality monitoring, taking the construction site as the construction quality monitoring site, and determining the construction site needing to adjust the monitoring scheme of the professional scene according to the identification association condition of the construction quality monitoring site in other professional scenes; and the reliability of verification processing of the identification reliability degree of the quality risk is improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a risk monitoring and prediction analysis method for smart construction sites and image recognition. Background Technology

[0002] When conducting risk monitoring and safety early warning at construction sites, it is necessary to combine monitoring data from monitoring devices to achieve risk prediction and processing for multiple professional scenarios. Existing technical solutions often rely on the analysis of monitoring images and the use of predictive maintenance and accident prevention technologies to avoid safety accidents. However, the following technical problems exist: For construction management units, the quality risk factors at different construction sites are not isolated, but rather interconnected. Therefore, how to achieve collaborative management of monitoring perspectives at different construction sites and determine whether the reliability of deep learning models for monitoring images in identifying different risk factors meets the requirements have become urgent technical problems to be solved.

[0003] Therefore, there is an urgent need for a risk monitoring and prediction analysis method for smart construction sites and image recognition. Summary of the Invention

[0004] To achieve the objectives of this invention, the following technical solution is adopted: Specifically, this application provides a risk monitoring and prediction analysis method for smart construction sites and image recognition, which includes: S1 utilizes the degree of overlap in monitoring targets of construction quality professional scenarios between construction sites to determine the need for assessment and analysis of the reliability of deep learning model identification. Based on risk identification data of construction sites in different professional scenarios, it identifies construction sites that cannot be monitored and analyzed solely in the professional scenario of construction quality and uses them as target construction sites. S2 uses the target construction site to identify and match the construction quality in different professional scenarios and identify and associate the construction sites in other professional scenarios, and determines the construction site to be focused on for construction quality monitoring and designates it as the construction quality monitoring site. Based on the identification and correlation of the construction quality monitoring site in other professional scenarios, S3 determines the construction sites that need to have their monitoring schemes adjusted for professional scenarios.

[0005] Furthermore, the monitoring targets for the professional scenarios of construction quality are determined based on the construction materials and construction projects that affect the construction quality. Specifically, these include concrete compressive strength, steel bar specifications, weather resistance of waterproof membrane, steel structure installation deviation, waterproof coating, waterproof construction time, steel bar quality, and door and window installation deviation.

[0006] Furthermore, the correlation is determined based on the construction projects at different construction sites. For specific construction projects of the same type, the monitoring targets for the professional scenarios of construction quality are consistent.

[0007] Furthermore, it is necessary to determine the reliability assessment and analysis required for the deep learning model's recognition capabilities, specifically including: Based on the aforementioned correlation, the overlap of construction projects at different construction sites is determined; Based on the aforementioned overlap, different construction sites are divided into different groups; Based on the construction site data in different combinations, determine whether deep learning model verification is required.

[0008] Specifically, such as Figure 3 As shown, the method for determining the construction quality monitoring site is as follows: The construction sites in the combination with the largest number of construction sites are selected as candidate targets. Based on the target construction site data in the candidate targets, the target construction sites in the candidate targets are determined. Based on the identification and matching status of the monitoring devices in the professional construction quality scenario of the candidate targets, determine the angle from which the monitoring devices of the candidate targets can monitor the largest number of monitoring targets in the professional construction quality scenario, and use this angle as the monitoring angle. Based on the monitoring target data in the professional scenario of construction quality under the monitoring angle and the monitoring target data in other professional scenarios under the monitoring angle, the ranking result of the candidate targets is determined. Based on the ranking result and the target construction sites among the candidate targets, the construction quality monitoring sites are determined.

[0009] The beneficial effects of this invention are as follows: Based on the identification and matching of monitoring devices in the construction quality-specific scenario and the identification and correlation in other professional scenarios, the construction site to be focused on for construction quality monitoring is identified and designated as the construction quality monitoring site. This approach considers both the monitoring reliability of the monitoring devices in the construction quality-specific scenario and the identification reliability of risk targets in other professional scenarios, provided that the monitoring reliability of the monitoring devices meets the requirements. This approach effectively identifies the construction quality monitoring site while ensuring reliable monitoring of construction quality and other professional scenarios.

[0010] Based on the identification and correlation of construction quality monitoring sites in other professional scenarios, construction sites that require adjustments to their professional scenario monitoring schemes are identified. This fully considers the deviations of monitoring targets in other professional scenarios that can be monitored from different monitoring angles of the monitoring devices at the construction quality monitoring site. Therefore, the construction sites requiring adjustments to their professional scenario monitoring schemes are determined from the perspective of the deviations of the monitoring targets. This allows for adjustments to the monitoring angles of the monitoring devices at the construction quality monitoring site based on the verification results of the monitoring targets in the deep learning model, ensuring the reliability of monitoring targets with high levels of deviation while also meeting the monitoring needs of construction quality.

[0011] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0012] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0013] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0014] Figure 1 This is a flowchart of a risk monitoring and predictive analysis method for smart construction sites and image recognition. Figure 2 It is a flowchart for determining whether an assessment and analysis of the reliability of deep learning model identification is needed; Figure 3 This is a flowchart illustrating the method for determining construction quality monitoring sites; Figure 4 This is a flowchart illustrating the method for identifying construction sites that require adjustments to the monitoring solutions for specific scenarios. Detailed Implementation

[0015] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0016] Example 1 like Figure 1As shown, this application provides a risk monitoring and prediction analysis method for smart construction sites and image recognition, specifically including: S1 utilizes the degree of overlap in monitoring targets of construction quality professional scenarios between construction sites to determine the need for assessment and analysis of the reliability of deep learning model identification. Based on the risk identification data of construction sites in various professional scenarios, it identifies construction sites that cannot be monitored and analyzed solely in the professional scenarios of construction quality and uses them as target construction sites. Furthermore, the monitoring targets for the professional scenarios of construction quality are determined based on the construction materials and construction projects that affect the construction quality. Specifically, these include concrete compressive strength, steel bar specifications, weather resistance of waterproof membrane, steel structure installation deviation, waterproof coating, waterproof construction time, steel bar quality, and door and window installation deviation.

[0017] Furthermore, the correlation is determined based on the construction projects at different construction sites. For specific construction projects of the same type, the monitoring targets for the professional scenarios of construction quality are consistent.

[0018] Specifically, determining the need for an assessment and analysis of the reliability of deep learning models in identification involves the following steps: The core objective of this solution is not to directly detect construction problems, but rather to evaluate the reliability of the deep learning model used for monitoring and to ensure the accuracy and effectiveness of large-scale, multi-site monitoring through dynamic adjustments. It addresses the fundamental question of "how to trust the eyes of AI."

[0019] Core logic: Monitor target → Link to project → Group construction sites by project → Assess model verification needs → Trigger verification to ensure identification reliability.

[0020] S11 determines the overlap of construction projects at different construction sites based on the aforementioned correlation. This is the first step in conducting group analysis. The system needs to identify commonalities among all managed construction sites from a macro perspective. By analyzing "project overlap," it can be discovered which sites are facing the same quality control points and safety risks, thus laying the foundation for addressing the reliability issues of subsequent batch processing models.

[0021] Overlap: This refers to the degree of overlap between two or more construction projects currently underway at different sites. If two sites are both carrying out "main structure construction," then they overlap in this "main structure construction" project.

[0022] S12 uses the overlapping situation to divide different construction sites into different combinations; This is the foundation of the entire system. All subsequent identification, analysis, and decision-making are based on these clearly defined objectives. Without clear definitions, deep learning models wouldn't know what to learn or identify, and the entire system would lose its direction. This is equivalent to establishing an "exam syllabus" for the entire quality monitoring system.

[0023] Definition of Construction Quality in Professional Contexts: This refers to the key aspects or objects requiring focused attention and control during construction, and those adhering to professional technical standards. It emphasizes whether the physical structure of the project itself meets specifications, encompassing safety equipment testing, hazardous area intrusion detection, equipment status monitoring (tower crane tilt warning, elevator equipment anomaly detection), and construction quality monitoring.

[0024] Monitoring targets: Technical indicators or risk points that need to be specifically examined and quantified in specific professional scenarios. These are the specific objects or states that deep learning models need to identify from images or data.

[0025] Examples of monitoring targets given: Rebar specifications: checking whether the diameter, grade, spacing, etc. of the rebar meet the design requirements; Steel structure installation deviation: measuring the error between the actual position of the steel structure component after installation and the design position; Safety risk events: unsafe conditions or behaviors that may cause personal injury or property damage during construction, such as "workers not wearing safety belts" or "foundation pit collapse".

[0026] Grouping (or clustering) is a core tool for data analysis and efficient management. By grouping construction sites with the same monitoring objectives into the same set, the system can evaluate and manage deep learning models on a "group" basis. This makes resource allocation and decision-making more efficient—focusing only on the "set" rather than hundreds of individual "construction sites."

[0027] A portfolio is a collection of all construction sites currently undertaking the same project. For example, "all construction sites currently installing doors and windows" constitute a portfolio.

[0028] S13 determines whether it is necessary to perform an assessment and analysis of the reliability of the deep learning model based on the construction site data in different combinations.

[0029] It is understandable that construction sites with similar projects are grouped into the same category.

[0030] It should be noted that when there is no need to evaluate and analyze the reliability of the deep learning model's recognition, there is no need to adjust the monitoring angles of different construction sites.

[0031] Furthermore, such as Figure 2As shown, based on the construction site data in different combinations, it is determined whether an evaluation and analysis of the reliability of the deep learning model's recognition is necessary. Specifically, this includes: Based on the construction site data in the combination, determine the verification requirement value of the combination; Based on the verification requirements of different combinations, determine whether it is necessary to conduct an evaluation and analysis of the reliability of deep learning models.

[0032] It should be noted that the verification requirement value of the combination is determined according to the proportion of the number of construction sites in the combination to all construction sites. When there is a combination whose verification requirement value is greater than the preset requirement threshold, since there are many construction sites with the same monitoring target that need to be observed, in order to ensure the reliability of the deep learning model in identifying quality risks, it is determined that the reliability of the deep learning model in identifying quality risks needs to be evaluated and analyzed.

[0033] This is the core of the entire process decision-making. The system's goal is to determine whether deep learning models used in specialized construction quality scenarios require evaluation and analysis. It is based on a risk management principle: the broader the potential impact of a problem, the higher its priority.

[0034] Deep learning models are artificial intelligence algorithms capable of automatically learning features from large amounts of data and performing complex tasks such as recognition and classification. In this system, it is responsible for identifying specific monitoring targets from surveillance images (determining whether construction quality defects exist).

[0035] It should be noted that when using deep learning models for risk identification and processing in multiple scenarios, the specific steps include: The system deploys multiple source sensors: Video source: 4K high-definition wide-angle cameras deployed on tower cranes and floor edges, covering the work area.

[0036] Sensors: tilt and pressure sensors installed on hoisting equipment; anemometers deployed on rooftops.

[0037] Data preprocessing: Video processing: Extract keyframes (1 frame / second) and perform perspective correction on each frame to eliminate wide-angle distortion.

[0038] Labeling work: Boundary box labeling: Label "Personnel", "Safety Rope", "Lifting Suction Cup", "Curtain Wall Unit", etc.

[0039] Key point annotation: Mark the "waist attachment point" of the person (used to determine the position of the seat belt).

[0040] Segmentation annotation: Pixel-level annotation for "installed curtain wall joints".

[0041] Labeling: Add risk labels to each frame of image or each video clip, such as [“Seatbelt violation”, “Seam too wide”].

[0042] Model selection, training and fusion strategies Phase 1: Real-time Risk Behavior and State Recognition (Computer Vision Model) Model: A combination of YOLOv8-Pose (keypoint detection version) and DeepLabV3+ (semantic segmentation) is used.

[0043] YOLOv8-Pose Training: Input: A single frame of a construction site image.

[0044] Output: Bounding boxes of personnel, categories (workers, commanders), and 17 body key points (especially waist coordinates).

[0045] Loss functions: Bounding box loss (CIoU), class loss (cross-entropy), keypoint loss (MSE).

[0046] Training method: Using weights pre-trained on the COCO dataset, transfer learning was performed using the labeled data from this project, with fine-tuning for approximately 100 epochs.

[0047] DeepLabV3+ Training: Input: A single frame image.

[0048] Output: Pixel-level mask of the "curtain wall seam" area.

[0049] Backbone network: ResNet-101.

[0050] Training method: The pre-trained model is also used for fine-tuning to optimize segmentation accuracy (Dice Loss + cross-entropy).

[0051] Reasoning and Rule Engine: Seatbelt violation detection: After YOLO detects a person, it calculates the pixel distance between the person's waist key point and the nearest "safety rope anchor point" (detected by another YOLO). If the distance exceeds a threshold and the person is in the edge area of ​​the floor slab, a "seatbelt not fastened" alarm is triggered.

[0052] Seam quality assessment: After DeepLabV3+ segments the seam area, it calculates the average pixel width. If the width exceeds the design allowable range (e.g., 20mm ± 2mm), the seam is marked as "width unacceptable".

[0053] Phase 2: Time-Series Risk Early Warning (Time-Series Prediction Model) Risk: Predict the overall risk level of the hoisting operation in the next 5 minutes.

[0054] Model: Multivariate time series Transformer.

[0055] Training: Input feature sequence (sliding window, length 30 minutes, step size 1 minute): Visual features: a temporal sequence of counts of “dangerous behaviors” (such as the number of people not wearing seat belts) in each frame, extracted by the YOLO model in stage 1.

[0056] Sensor sequence: wind speed (m / s), hoisting equipment tilt angle (degrees), suction cup pressure value (MPa).

[0057] Environmental sequence: time period (day / night), working face height.

[0058] Output label: Risk level (0-Normal, 1-Low risk, 2-Medium risk, 3-High risk) at the 5th minute in the future. The label is generated retrospectively based on historical accident records and expert rules.

[0059] Model architecture: The encoder uses a Transformer layer to capture long-term dependencies among multiple variables; the decoder uses a fully connected layer for regression.

[0060] Training method: Supervised learning is adopted to minimize the cross-entropy loss between the predicted risk level and the actual risk level.

[0061] Phase 3: Multimodal decision fusion (information aggregation), model / policy: late-stage fusion based on attention mechanism.

[0062] Each modality is scored independently: Visual risk score: Output by the Phase 1 rule engine, normalized to 0-1.

[0063] Time series prediction risk score: output from the Phase 2 Transformer model and converted into probability.

[0064] Text quality inspection score: The pass rate of the inspection items uploaded by the quality inspector (e.g., "Sealant inspection passed" is 1, otherwise it is 0).

[0065] Integrated Decision Making: Design a lightweight synthetic neural network (several fully connected layers) that takes the three scores mentioned above and the original features (such as the current wind speed) as input. The network learns to automatically assign weights through attention layers. For example, when the wind speed is extremely high, the model will give more attention to the "time-series predicted score" and the "original wind speed value".

[0066] The network ultimately outputs a global risk score and the main risk sources (e.g., "80% due to strong winds, 20% due to personnel violations").

[0067] System Deployment and Workflow Edge computing: In Phase 1, the YOLO and DeepLabV3+ models are deployed on the edge server at the construction site to perform real-time analysis of the video stream and achieve local alarms (such as on-site audible and visual alarms) within 200ms.

[0068] Cloud-based collaboration: The models for phases 2 and 3 run in the cloud. Edge servers package and upload the extracted time-series features and key events to the cloud every 30 seconds.

[0069] Real-time cockpit: The cloud-integrated model updates the global risk map every minute and pushes it to the mobile apps of project managers and safety management personnel. When the risk level is "high," the operation is automatically suspended, and the tower crane operator is notified.

[0070] Model Iteration: The system collects all false alarms and missed alarms, which are then confirmed or corrected by security personnel on the app. This corrected data automatically generates new training samples weekly, initiating a round of online incremental learning to continuously optimize model performance.

[0071] Verification processing: This involves evaluating and validating the performance of deployed deep learning models, including checking their recognition accuracy, false negative rate, and false positive rate. If necessary, retraining with new data can improve reliability. A quantitative and objective metric is needed to compare the priority of different combinations, avoiding subjective assumptions. The proportional algorithm simply and directly reflects the influence range of a particular model.

[0072] Verification Requirement Value: A quantitative metric used to measure the urgency and importance of verifying the deep learning models used in a portfolio. It is calculated by dividing the number of projects within the portfolio by the total number of projects.

[0073] Preset demand threshold: A manually set threshold value. When the verification demand value exceeds this threshold, the system automatically triggers the verification process. This threshold is a key parameter for balancing "monitoring reliability" and "verification cost".

[0074] Explanation of decision-making logic: "When there are combinations of verification requirement values ​​greater than the preset requirement threshold, since there are many construction sites that need to observe the same monitoring target, it is determined that deep learning models need to be verified in order to ensure the reliability of quality risk identification." This follows the "risk concentration" principle. If there are combinations of verification requirement values ​​greater than the preset threshold, such as 0.5 or higher, it is determined that the reliability of deep learning models needs to be evaluated and analyzed.

[0075] Specifically, the risk identification data of the construction site in various professional scenarios is determined based on the safety risk events identified in the monitoring images of the construction site, such as construction workers without safety protection measures, abnormal shaking of construction elevators, tilting of construction towers, foundation pit collapse, and leakage of electricity.

[0076] Specifically, the method for determining the target construction site is as follows: S21 Based on the risk identification data of the construction site in various professional scenarios, determine the safety risk events of the construction site; Specifically, identifying safety risk events at construction sites is the data foundation for safety assessments. The system cannot judge whether a site is dangerous based on intuition; it must rely on objective factual evidence identified and recorded by AI. This step involves extracting a structured list of risk events from raw monitoring data, providing "ammunition" for subsequent decision-making.

[0077] Risk identification data refers to the raw results generated by the AI ​​model after analyzing surveillance images. It may be a series of data records containing time, location, event type, and confidence level. For example: Timestamp: 2023-10-27 10:15:00, Camera ID: Cam_05, Identified event: not wearing a safety helmet, Coordinates: (x1,y1,x2,y2), Confidence level: 0.96.

[0078] Safety risk events: Specific, confirmed safety violations or dangerous situations extracted from "risk identification data." These are specific items such as "construction workers without safety protection measures" (i.e., not wearing safety helmets / safety belts) and "abnormal shaking of construction elevators" mentioned above. "Risk identification data" is the raw material, while "safety risk events" are the processed and confirmed finished product.

[0079] S22 determines whether the construction site is the target construction site based on the safety risk events at the construction site.

[0080] Specifically, identifying target construction sites based on safety risk events is a data-driven decision-making process. The system needs a clear set of rules to accurately pinpoint "key sites" with weak safety management and frequent risks. Focusing solely on monitoring construction quality might lead to unreliable monitoring of safety risk events, making it impossible to concentrate on construction quality monitoring at the aforementioned sites.

[0081] Target construction site: refers to a construction site that, according to preset rules, is determined by the system to have a high safety risk and needs to be closely monitored and managed.

[0082] It should be noted that when the safety risk events at the construction site do not meet the requirements, due to the large number of safety risk events occurring there, if the focus is on monitoring, analyzing and processing in the professional scenario of construction quality, it will inevitably lead to poor reliability in the identification of other professional scenarios, making it difficult to carry out timely and effective early warning and processing of safety risk events. Therefore, the construction site is identified as the target construction site.

[0083] Requirements not met: A general conclusion is that the safety risk event record at the construction site triggered the system's preset alarm conditions. The specific triggering conditions are described in detail in the examples below.

[0084] In one possible embodiment, determining that the safety risk event at the construction site does not meet the requirements specifically includes: If a safety risk event is detected on different dates, then the construction site is identified as the target construction site.

[0085] Example 1: Assessment of Persistent Risk Specific method: If a safety risk event is detected on different dates, the construction site is identified as the target construction site.

[0086] This rule aims to identify construction sites with persistent, systemic safety issues. While an occasional problem might be accidental, problems occurring on multiple separate workdays indicate that the safety issues are not isolated incidents but rather stem from deeper causes such as lax management and habitual violations. Such sites represent a higher level of risk.

[0087] Definition: Different Dates: Refers to two or more independent calendar days. This emphasizes the continuity and recurrence of risk over time.

[0088] Example: Safety incident records at construction site X: 2023-11-01: 2 incidents of "not wearing safety belts" were found. 2023-11-02: 1 incident of "electrical leakage risk" was found. 2023-11-03: 1 incident of "construction elevator malfunction" was found. Analysis: Although the number of incidents per day is not large, the fact that safety incidents occur every day indicates that safety problems repeatedly occur at this construction site and have not been effectively rectified.

[0089] Judgment: According to the rules of Example 1, the system determines that construction site X is the target construction site.

[0090] In another possible embodiment, determining that the safety risk event at the construction site does not meet the requirements specifically includes: When there are more than a preset number of dates on which multiple safety risk events have been detected, for example, when there are more than three dates on which multiple safety risk events have been detected within the most recent month, the construction site is determined to be the target construction site.

[0091] Example 2: High-Frequency Risk Assessment Specific method: When there are more than a predetermined number of days on which multiple safety risk events have been detected, the construction site is designated as a target construction site. This rule aims to identify construction sites with concentrated, explosive safety problems. It focuses not on the persistence of the risk, but on the frequency of severe outbreaks. Even if the problems are concentrated on only a few days, if there are too many such "high-risk days," it indicates a significant loophole in the site's safety management.

[0092] Definition: Preset quantity: A threshold set in advance by the system administrator, such as "more than 3".

[0093] Dates of multiple safety risk events: This refers to the detection of more than one safety risk event on the same day. This indicates that on that day, safety management at the construction site was completely out of control.

[0094] Example: Preset rule: If there are more than 3 dates within the past month where multiple safety risk events are detected, the site is identified as a target construction site. Construction site Y's safety event records for the past month (e.g., October) are as follows: 2023-10-10: 5 safety events detected (e.g., not wearing a safety helmet, smoking in violation of regulations, etc.); 2023-10-18: 3 safety events detected; 2023-10-25: 4 safety events detected; 2023-10-28: 1 safety event detected.

[0095] Analysis: In October, construction site Y experienced "multiple safety risk events" (i.e., the number of events > 1) on three dates (10th, 18th, and 25th). This exactly reached the preset threshold of "more than 3". Determination: According to the rules of Example 2, the system determines that construction site Y is the target construction site.

[0096] S2 uses the target construction site to determine the construction site to be focused on for construction quality monitoring based on the identification and matching of monitoring devices in the professional scenario of construction quality and the identification and association in other professional scenarios, and designates it as the construction quality monitoring site. Overall, this process builds upon the results of the first two processes: from process one, we obtain a "combination" that needs to be verified (i.e., the combination with the most identical construction projects and the largest number of construction sites); from process two, we obtain which "target construction sites" (i.e. problem sites) in this combination have frequent safety issues.

[0097] Now, the task of this process is to determine which construction sites to select from this "group" of sites that need to be verified to perform the verification task. Core logic: Identify candidate sites → Find the best monitoring angle → Rank construction sites → Combine safety conditions to finally determine the list of construction sites to be used for verification of deep learning models.

[0098] Specifically, such as Figure 3 As shown, the method for determining the construction quality monitoring site is as follows: S31 selects the construction sites in the combination with the largest number of construction sites as candidate targets, and determines the target construction sites in the candidate targets based on the target construction site data in the candidate targets. Specifically, the process involves identifying candidate targets and the problematic construction sites within them. The "combination with the largest number of construction sites" is a conclusion from the previous process, meaning that the AI ​​model for verifying this combination has the highest cost-effectiveness. Using these sets of construction sites as "candidate targets" means that the verification data will be generated from them.

[0099] At the same time, it is necessary to identify the "target construction sites" (problem sites) because these sites will be handled specially in subsequent decision-making.

[0100] Candidate sites: This refers to the set of all construction sites selected from the "combinations" that need to be verified. They are the final pool of "construction quality monitoring sites".

[0101] Target construction site data: refers to the information of those construction sites that are determined to be "target construction sites" according to the rules in process two within the "candidate target" set.

[0102] S32 determines the angle from which the monitoring device of the candidate target can detect the largest number of monitoring targets in the professional scenario of construction quality based on the identification and matching of the candidate target in the professional scenario of construction quality, and uses it as the monitoring angle. Specifically, determining the optimal monitoring angle is crucial to ensuring high-quality verification data. The AI ​​model's verification needs to capture the monitoring targets (such as steel bars and concrete) in the professional construction quality scenario as clearly and comprehensively as possible. Therefore, the system needs to simulate the perspective of a "best photographer" to find an angle that allows the camera to see the most quality monitoring targets. This angle will be uniformly applied to all candidate construction sites to ensure the standardization of data collection.

[0103] Monitoring angle: refers to a specific perspective formed by the setting of parameters such as the orientation and tilt angle of a surveillance camera in space.

[0104] Matching identification: refers to the ability and effectiveness of an AI model to successfully identify the monitored target in an image from a certain monitoring angle.

[0105] "The angle from which the monitoring device can detect the most monitoring targets": This is an optimization goal, which is to find a globally optimal perspective by analyzing monitoring images from different angles, so that the number of quality monitoring targets (such as steel bars and formwork) that can be seen from this perspective is maximized.

[0106] S33 determines the ranking result of the candidate targets based on the monitoring target data in the professional scenario of construction quality under the monitoring angle and the monitoring target data in other professional scenarios under the monitoring angle, and determines the construction quality monitoring site based on the ranking result and the target construction site among the candidate targets.

[0107] In the above steps, sorting the construction sites and finalizing the list is the most refined screening step. After finding the best "monitoring angle," the system needs to further evaluate which construction site can provide the "richest" verification data from this angle. An ideal data collection site should not only clearly show the quality objectives, but also ideally "bundle" some objectives from other scenarios, thus achieving the effect of multiple models benefiting from a single verification.

[0108] Ranking Results: The ranking of all "candidate target" construction sites is based on a quantitative indicator: the number of monitoring targets in the professional scenario of construction quality under the monitoring perspective + the number of other professional scenarios with monitoring targets under the monitoring perspective.

[0109] The first part (quantity of quality targets): This is the core indicator, directly determining the amount of data used to verify the quality model. The second part (quantity of other targets): This is an added value indicator. If targets from other scenarios such as security and equipment can be seen simultaneously, it means that the collected data can also be used to verify other AI models, improving the overall efficiency of data utilization.

[0110] Scenario 1: When the proportion of the target construction site among the candidate targets is greater than the preset proportion threshold, since there are a large number of target construction sites, in order to ensure the reliability of the verification processing of the deep learning model for professional scenarios of construction quality, all candidate targets other than the target construction site are determined to be construction quality monitoring sites.

[0111] The system adopted two different selection strategies based on the proportion of "problematic construction sites" in the "candidate targets," which reflects the flexibility of management.

[0112] Scenario 1: When the proportion of problematic construction sites is high (>60%) – “Safety First” strategy; Rule: When the proportion of the target construction site is greater than 0.6, all candidate targets other than the target construction site shall be used as construction quality monitoring sites.

[0113] This is a conservative and safe option. If more than 60% of the sites in a portfolio are "problem sites" with frequent safety issues, it indicates that the overall safety management level of the portfolio is very low. Conducting detailed quality data collection on such sites may pose safety risks, or the chaotic site environment may lead to poor data quality (e.g., disorderly stacking of materials obstructing visibility).

[0114] Therefore, the system adopts a "one-size-fits-all" strategy, directly excluding all problematic construction sites and selecting only from the remaining less than 40% of "relatively safe" sites. Even if the data collection conditions at these sites are not optimal, they can still basically meet the verification requirements, while ensuring the reliability of the process and data.

[0115] It should also be noted that when the proportion of the target construction site among the candidate targets is greater than the preset proportion threshold, the candidate target ranked before the preset position will be used as the construction quality monitoring site.

[0116] It should be noted that the sorting results are sorted from largest to smallest based on the sum of the number of monitoring targets in the professional scenario of construction quality under the monitoring angle and the number of monitoring targets in other professional scenarios under the monitoring angle.

[0117] Scenario 2: When the proportion of problematic construction sites is not high (≤60%) – “Selection based on merit” strategy Rule: Select the candidate targets that rank before the preset position (e.g., the top 40%) as construction sites for construction quality monitoring.

[0118] This is an efficiency-driven approach. When most construction sites are in acceptable condition, the system can confidently select the best among the best. Based on the previously calculated "ranking results" (i.e., data richness ranking), it selects only the top-ranked sites (e.g., the top 40%), using the fewest site samples to obtain the largest amount and most diverse verification data, greatly improving the efficiency and data value of the verification work.

[0119] In one possible specific embodiment, S31: The combination of "constructing the main structure" is the candidate target, which includes 100 construction sites. After analysis, 20 of them were identified as "target construction sites" (problem sites), accounting for 20% (<60%).

[0120] S32: System analysis revealed that setting a monitoring angle that "looks down at the work surface from the southeast corner" provides the clearest view of the most rebar and formwork.

[0121] S33: Ranking: The system uses this "optimal angle" to simulate and analyze 100 candidate construction sites, calculates the (number of visible steel bars + number of other visible targets such as construction workers and tower cranes) for each site, and ranks them.

[0122] Ultimately, it was determined that since problematic construction sites accounted for 20% (<60%), the system would employ a "selection based on merit" strategy. Assuming the top 40% are pre-selected, the system will choose the top 40 construction sites from the ranking as the "construction quality monitoring sites" for this verification of the "rebar tying" quality identification model.

[0123] This process demonstrates a highly intelligent decision-making process that comprehensively considers the urgency of model verification (combined selection), on-site safety risks (target site screening), data collection efficiency (optimal perspective), and maximizing data value (comprehensive ranking), ultimately achieving the optimal allocation of limited resources in a complex environment.

[0124] Example 2 Based on the identification and correlation of construction quality monitoring sites in other professional scenarios, determine the construction sites that require adjustments to the monitoring schemes for those professional scenarios.

[0125] It should be noted that the construction sites that require adjustments to the monitoring scheme for specific scenarios are those excluding construction quality monitoring sites and target construction sites.

[0126] The core logic of this application is to determine which direction is most effective when adjusting the camera angle at a "construction quality monitoring site," and to prioritize the monitoring quality of targets other than those specifically related to construction quality. Solution: Analyze monitoring data from other ordinary construction sites to identify targets that are prevalent but have poor monitoring reliability. These targets represent the "weak links" in the entire monitoring system and should be prioritized for coverage as "high-value targets" after adjusting the perspective of "construction quality monitoring sites." Based on the analysis results, dynamically adjust the monitoring schemes for ordinary construction sites with a large number of "high-value targets" to determine the monitoring reliability of high-value targets in ordinary construction sites. This reliability will then guide the adjustment of the monitoring perspective for construction quality monitoring sites.

[0127] Specifically, such as Figure 4 As shown, the method for determining the construction site requiring adjustment of the monitoring scheme for the specific scenario is as follows: S41 determines the monitoring targets in other professional scenarios under the current monitoring angle based on the monitoring data of the construction quality monitoring site, and uses them as matching monitoring targets. In the above steps, identifying the "matching monitoring targets" refers to the monitoring objects (such as safety, equipment, and environment) that can be identified from the current monitoring perspective of the construction quality monitoring site, in addition to its core construction quality scenario. This step aims to establish an evaluation benchmark. By taking stock of the additional monitoring responsibilities that the core monitoring points are currently "incidentally" undertaking, a baseline is provided for subsequently evaluating the impact and benefits of adjusting their perspective.

[0128] Scenario: In Project A (Commercial and Financial Center), the construction quality monitoring camera Q_A is currently focused on "core tube steel structure welding". Process: The system analyzes the real-time footage of Q_A and identifies that while monitoring the welding quality, it can also see: the safety helmet wearing status of the tower crane operator in the background (personnel safety scenario), and the standard status of the precast component stacking area below (material stacking scenario). Output: Matched monitoring target = {safety helmet, precast component stacking}; S42 takes the angle of the construction quality monitoring site as the professional scene of construction quality under other monitoring angles as the available angle, and determines the available angle where the number of monitoring targets is greater than the number of matching monitoring targets based on the number of monitoring targets in other professional scenes under the available angle. In the above steps, find "the available angles where the number of monitoring targets is greater than the number of matching monitoring targets". Available angles: other physical angles that the cameras at the construction quality monitoring site can be adjusted to without seriously affecting their core construction quality monitoring tasks.

[0129] Available angles where the number of monitored targets exceeds the number of matched monitored targets: Among these available angles, those that can cover a number of other professional scene targets exceeding the number determined in S41.

[0130] This step aims to identify optimization opportunities. Through simulation analysis, it identifies potential angles that can maximize the monitoring coverage and improve overall utilization efficiency of core monitoring points while fulfilling their primary responsibilities.

[0131] Scenario: Continuing from the previous example, the system simulates the angle of the Q_A camera in project A. Process: It finds an available angle α that can still clearly monitor the key process of core tube welding while covering all matching monitoring targets in S41: {safety helmet, prefabricated component stacking}. New targets are added: {tower crane hook (equipment status), foundation pit dewatering well water level (environmental safety)}. Judgment: The number of targets under the new angle (4) > the number of targets under the current angle (2). Output: The available angle α is a better angle, and its monitoring target set = {safety helmet, prefabricated component stacking, tower crane hook, foundation pit dewatering well water level}.

[0132] Based on the available angles where the number of monitoring targets at different construction quality monitoring sites exceeds the number of matching monitoring targets, S43 determines the construction sites that require adjustments to the monitoring schemes for their respective professional scenarios, using monitoring target data from other professional scenarios.

[0133] Understandably, based on the premise that the number of monitoring targets at different construction quality monitoring sites exceeds the available number of matching monitoring targets, and considering the monitoring target data in other professional scenarios, the construction sites that require adjustments to their professional scenario monitoring plans are identified. These specifically include: Based on the available angle where the number of monitoring targets in different construction quality monitoring sites is greater than the number of matching monitoring targets, the overlap of monitoring targets in other professional scenarios is used to determine the number of construction quality monitoring sites corresponding to different monitoring targets. The "Switching Monitoring Requirement Targets" are identified. These targets, within the "Better Angle Target Set" discovered in S42, are those whose corresponding number of construction quality monitoring sites (i.e., the number of construction quality monitoring sites with switching requirement targets) exceeds a preset threshold. These are high-value targets that are universally needed across projects. From the local optimization opportunities in Q_A, monitoring targets with global value are selected. Subsequent decisions will revolve around these high-value targets.

[0134] Scenario 1: When there are no construction quality monitoring sites with a number of monitoring targets greater than the preset threshold, there is no need to adjust the angle of the construction quality monitoring sites based on the identification deviation of the monitoring targets in other construction sites. Therefore, it is determined that no professional scene monitoring scheme needs to be adjusted for other construction sites. In the above steps, there is no universal target and no need for adjustment. This situation is triggered when there is no "switching monitoring requirement target". If the target targeted by the Q_A angle adjustment is not universal, it means that this is only a local optimization and will not have a significant impact on the global monitoring network. Therefore, there is no need to carry out cross-project adjustments.

[0135] Scenario: Suppose that in S42, the newly added target for the available angle α of Q_A is "special curtain wall installation machine". This target is only of concern to project A, while construction quality monitoring sites E and F are not concerned. That is, the above monitoring target exists. The system determines that there is no "switching monitoring requirement target". Therefore, it is determined that the monitoring scheme of construction sites B, C and D does not need to be adjusted.

[0136] Scenario 2: When there are monitoring targets with a number of corresponding construction quality monitoring sites that exceed the preset threshold, the monitoring targets with a number of corresponding construction quality monitoring sites that exceed the preset threshold are used as switching monitoring demand targets. When the number of monitoring targets for different switching monitoring demands in other construction sites all exceeds the preset threshold, the deep learning model verification of the switching monitoring demand targets is deemed to have high reliability. Therefore, it is determined that no adjustment of the monitoring scheme for the professional scenario is required for other construction sites. In the above steps, the number of monitored construction sites refers to the number of ordinary construction sites in other projects (B, C, D) that can effectively monitor a particular "switching monitoring target." Even if a target is important and widespread, if it is already reliably covered by existing monitoring points in other projects, it indicates that the monitoring network is robust. Adjusting it at this point would be a waste of resources.

[0137] Example: Scenario: Following the main process, if the monitoring target exists in A / E / F, that is, when the monitoring requirement target is "tower crane hook".

[0138] System checks revealed that in projects B, C, and D, there were three ordinary cameras that could clearly monitor the "tower crane hooks" in their respective areas. The number of monitored construction sites was three, which is greater than the preset threshold (2). Therefore, the monitoring reliability was high, and it was determined that no adjustment to the monitoring plan was needed for the construction sites in projects B, C, and D.

[0139] Scenario 3: When there are other construction sites with a number of monitored construction sites that do not exceed the preset threshold for switching monitoring requirements, the number of other construction sites with a number of monitored construction sites that do not exceed the preset threshold for switching monitoring requirements is taken as the switching monitoring requirement value. When the switching monitoring requirement value is not greater than the preset monitoring requirement threshold, the number of switching monitoring requirements with poor monitoring reliability that the construction quality monitoring site needs to consider when adjusting the monitoring angle is small. Therefore, it is determined that other construction sites do not need to adjust the monitoring scheme for professional scenarios. In other words, there is no need to consider switching monitoring requirements with poor monitoring reliability when switching angles. In the above steps, the number of unreliable targets is small, the impact is controllable, and no adjustment is required. Switch monitoring requirement value: refers to the total number of targets identified as "targets requiring switching monitoring", but which are unreliable to monitor in ordinary construction sites of other projects (i.e., the number of construction sites under monitoring is ≤ the preset number threshold).

[0140] This is a cost-benefit analysis. When there are few unreliable high-value objectives, adjusting the monitoring plans for multiple projects to achieve a single objective will result in management costs and operational risks that outweigh the benefits.

[0141] Example: Scenario: The target of the monitoring switch is "tower crane hook". Reliability check: It was found that only one camera in Project C could vaguely see "tower crane hook". The number of construction sites under monitoring = 1, which is not greater than the threshold (2). Calculate the demand value: The only target of the monitoring switch due to unreliable monitoring is "tower crane hook". The monitoring switch demand value = 1.

[0142] Judgment: The demand value (1) is not greater than the preset monitoring demand threshold (1). Therefore, it is determined that the construction sites of projects B, C, and D do not need to adjust the monitoring plan.

[0143] Scenario 4: When the switching monitoring demand value is greater than the preset monitoring demand threshold, if the switching monitoring demand value is within the preset demand threshold range, that is, when the switching monitoring demand value is relatively large, then the number of construction sites monitored that are not greater than the preset number threshold for switching monitoring demand targets, and the number of construction sites within the preset demand target number range, are all considered as construction sites that require adjustment of the monitoring scheme for professional scenarios. Through dynamic adjustment of the monitoring scheme, the monitoring reliability of the switching monitoring demand targets is determined, thereby determining whether the monitoring angle of the construction quality monitoring sites needs to be adjusted.

[0144] In the above embodiments, the number of unreliable targets is moderate, allowing for precise adjustments. When the switching monitoring demand value is within a preset demand threshold range, the system only adjusts those construction sites directly related to specific unreliable targets. This is a "precision radiotherapy" strategy. When the problem reaches a certain scale but has not yet become widespread, targeted adjustments are made, resulting in controllable costs and significant effects.

[0145] Example: Assume that after analysis of S42 / S43, it is found that "tower crane hook" and "construction elevator entrance / exit" are both targets requiring switching monitoring, and their coverage in other projects is unreliable (the number of construction sites monitored is ≤2).

[0146] Calculate the demand value: There are 2 unreliable switching monitoring demand targets. Switching monitoring demand value = 2, judgment: Demand value (2) is within the preset demand threshold range [2,3]. The system accurately locates the construction site of project C where the monitoring of "tower crane hook" is unreliable and the construction site of project B where the monitoring of "construction elevator entrance and exit" is unreliable. It requires these two specific construction sites to adjust the monitoring scheme. By adjusting the monitoring angle according to the preset cycle, such as 1 hour, the monitoring reliability in the above specific construction sites is determined.

[0147] Scenario 5: It is also understood that if the switching monitoring demand value is not within the preset demand threshold range, then all construction sites with a number of monitored construction sites not exceeding the preset number threshold for switching monitoring demand targets will be considered as construction sites requiring adjustment of the monitoring scheme for professional scenarios. Through dynamic adjustment of the monitoring scheme, the monitoring reliability of the switching monitoring demand target will be determined, thereby determining whether the monitoring angle of the construction quality monitoring site needs to be adjusted.

[0148] In the above embodiment: With a large number of unreliable targets, extensive adjustments are made. When the switching monitoring demand value is within the preset demand threshold range, the system adopts a stricter strategy, treating all construction sites containing any unreliable targets as adjustment targets. When the system discovers that multiple high-value targets generally have monitoring risks, it indicates that the problem is systemic. Extensive adjustments are then made to quickly and comprehensively address the shortcomings of the global monitoring network.

[0149] Scenario: Assuming that after analysis of S42 / S43, it is found that "tower crane hook", "construction elevator entrance / exit", "hazardous materials warehouse", and "water level in foundation pit dewatering well" are all switching monitoring requirements, and their coverage in other projects is unreliable, the required value is calculated as follows: There are 3 unreliable switching monitoring requirements. Switching monitoring requirement value = 3.

[0150] Judgment: The demand value (3) is not within the preset demand threshold range [2,3]. Decision (case 5): The system generates a list. All construction sites in projects B, C, and D that can monitor any of the three targets are required to adjust their monitoring plans. By adjusting the monitoring plans, the monitoring reliability of the switching monitoring demand targets is identified. When the monitoring reliability of the switching monitoring demand targets is poor, the monitoring angle of the construction quality monitoring site is automatically adjusted to realize the monitoring processing of the switching monitoring demand targets with poor monitoring reliability.

[0151] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0152] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0153] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A risk monitoring and prediction analysis method for smart construction sites and image recognition, characterized in that, Specifically, it includes: By utilizing the degree of overlap in the monitoring targets of construction quality professional scenarios between construction sites, when determining the need for assessment and analysis of the reliability of deep learning model identification, risk identification data of construction sites in different professional scenarios are used to identify construction sites that cannot be monitored and analyzed solely in the professional scenarios of construction quality, and these sites are designated as target construction sites. Using the target construction site, based on the identification and matching data of different construction sites in professional construction quality scenarios and the identification and association data in other professional scenarios, the construction site to be focused on for construction quality monitoring is identified and designated as the construction quality monitoring site. Based on the identification and correlation of construction quality monitoring sites in other professional scenarios, determine the construction sites that require adjustments to the monitoring schemes for those professional scenarios.

2. The risk monitoring and prediction analysis method for smart construction sites and image recognition as described in claim 1, characterized in that, The process involves determining the reliability assessment of the deep learning model's recognition capabilities, specifically including: Based on the aforementioned correlation, the overlap of construction projects at different construction sites is determined; Based on the aforementioned overlap, different construction sites are divided into different groups; Based on the construction site data in different combinations, determine whether deep learning model verification is required.

3. The risk monitoring and prediction analysis method for smart construction sites and image recognition as described in claim 1, characterized in that, The risk identification data of the construction site in various professional scenarios is determined based on the safety risk events identified in the monitoring images of the construction site, such as construction workers without safety protection measures, abnormal shaking of construction elevators, tilting of construction towers, foundation pit collapse, and leakage of electricity.

4. The risk monitoring and prediction analysis method for smart construction sites and image recognition as described in claim 1, characterized in that, The method for determining the construction site for construction quality monitoring is as follows: The construction sites in the combination with the largest number of construction sites are selected as candidate targets. Based on the target construction site data in the candidate targets, the target construction sites in the candidate targets are determined. Based on the identification and matching status of the monitoring devices in the professional construction quality scenario of the candidate targets, determine the angle from which the monitoring devices of the candidate targets can monitor the largest number of monitoring targets in the professional construction quality scenario, and use this angle as the monitoring angle. Based on the monitoring target data in the professional scenario of construction quality under the monitoring angle and the monitoring target data in other professional scenarios under the monitoring angle, the ranking result of the candidate targets is determined. Based on the ranking result and the target construction sites among the candidate targets, the construction sites to be monitored for construction quality are determined.

5. The risk monitoring and prediction analysis method for smart construction sites and image recognition as described in claim 4, characterized in that, When the proportion of the target construction site among the candidate targets is greater than a preset proportion threshold, all candidate targets other than the target construction site are determined as construction quality monitoring sites.

6. The risk monitoring and prediction analysis method for smart construction sites and image recognition as described in claim 4, characterized in that, When the proportion of the target construction site among the candidate targets is not greater than a preset proportion threshold, the candidate target ranked before the preset position is selected as the construction site for construction quality monitoring.

7. The risk monitoring and prediction analysis method for smart construction sites and image recognition as described in claim 6, characterized in that, The sorting results are sorted from largest to smallest based on the sum of the number of monitoring targets in the professional scenario of construction quality under the monitoring angle and the number of monitoring targets in other professional scenarios under the monitoring angle.

8. The risk monitoring and prediction analysis method for smart construction sites and image recognition as described in claim 1, characterized in that, The method for determining the construction sites that require adjustments to the monitoring scheme for specific scenarios is as follows: Using the construction quality monitoring site as the construction quality monitoring site, based on the monitoring data of the construction quality monitoring site under the current monitoring angle, determine the monitoring targets in other professional scenarios under the current monitoring angle, and use them as matching monitoring targets; The angles from which construction quality monitoring sites exhibit professional scenarios of construction quality under other monitoring angles are taken as available angles. Based on the number of monitoring targets in other professional scenarios under the available angles, the available angles with a number of monitoring targets greater than the number of matching monitoring targets are determined. Based on the available angles where the number of monitoring targets at different construction quality monitoring sites exceeds the number of matching monitoring targets, and considering the monitoring target data in other professional scenarios, we can determine the construction sites that require adjustments to their professional scenario monitoring plans.

9. The risk monitoring and prediction analysis method for smart construction sites and image recognition as described in claim 8, characterized in that, Based on the available angle where the number of monitoring targets at different construction quality monitoring sites exceeds the number of matching monitoring targets, and considering the monitoring target data in other professional scenarios, we identify construction sites that require adjustments to their professional scenario monitoring plans. These specifically include: Based on the available angle where the number of monitoring targets in different construction quality monitoring sites is greater than the number of matching monitoring targets, the overlap of monitoring targets in other professional scenarios is used to determine the number of construction quality monitoring sites corresponding to different monitoring targets. If the number of construction sites for which there is no corresponding construction quality monitoring exceeds the preset threshold for the number of construction sites, then it is determined that no adjustments to the monitoring scheme for the other construction sites are required.

Images