Dynamic weight distribution multi-modal construction site safety early warning method and system

Abstract

The present application relates to the technical field of construction site safety early warning, in particular to a multi-modal construction site safety early warning method and system with dynamic weight distribution, comprising the following steps: deploying a multi-modal data acquisition module, collecting and transmitting construction site data to a central processing unit in real time; constructing a correlation model according to a historical accident database; pre-processing real-time data and detecting abnormalities, and dynamically adjusting the multi-modal weight according to the results; inputting the adjusted weight into the model, calculating the comprehensive early warning score and grading the early warning; correcting the model according to the early warning and accident matching degree. The present application collects three types of data, namely sensors, vision and text, by deploying a multi-modal data acquisition module, and combines a dynamic weight distribution mechanism, which can adjust the weight of each mode in real time according to the abnormal detection results, significantly improves the accuracy and comprehensiveness of safety hazard identification, and effectively avoids the one-sidedness and missed reporting problems of single-mode monitoring.

Description

Technical Field

[0001] This invention relates to the field of construction site safety early warning technology, and more specifically, to a multimodal construction site safety early warning method and system with dynamic weight allocation. Background Technology

[0002] With the development of construction site informatization, information technology has penetrated into all aspects of construction site production. However, existing construction site safety early warning technologies have many problems that urgently need to be solved, resulting in insufficient accuracy, real-time performance, and adaptability of early warnings.

[0003] First, traditional methods often rely on a single data source, such as sensor modalities, which only collect environmental or equipment data and cannot link it to key risk factors such as personnel behavior and management documents. This makes it easy to miss accidents caused by "unsafe human behavior" and "management deficiencies." Visual modalities, on the other hand, rely solely on camera monitoring and lack linkage analysis with equipment operating parameters and environmental data, resulting in a high false alarm rate. At the same time, textual data such as safety inspection records and construction plans are not used in a structured way and cannot be mapped to dynamic construction scenarios in real time, forming "data silos."

[0004] Secondly, existing multimodal fusion technologies use fixed weights and do not consider the impact of real-time anomalies and the differences in construction stages. For example, when sensor values ​​change abruptly or high-risk behaviors are detected by vision, the corresponding modal weights are not increased in a targeted manner, resulting in the weakening of key risk signals. The risk modal priorities are different in different construction stages, and static weights cannot be dynamically adapted.

[0005] In view of this, a multimodal construction site safety early warning method and system with dynamic weight allocation is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a multimodal construction site safety early warning method and system with dynamic weight allocation, so as to solve the problems of insufficient accuracy, real-time performance and adaptability in existing early warning technologies.

[0007] To address the aforementioned technical problems, this invention provides a multimodal construction site safety early warning method with dynamic weight allocation, comprising the following steps:

[0008] S1. By deploying a multimodal data acquisition module, multimodal data from the construction site is collected in real time and transmitted to the central processing unit;

[0009] S2. Based on the historical accident database, construct a correlation model between multimodal data and accident types;

[0010] S3. Perform preprocessing and anomaly detection on the real-time acquired multimodal data, and dynamically adjust the multimodal weights based on the anomaly detection results;

[0011] S4. Input the adjusted multimodal weights into the correlation model, calculate the comprehensive early warning score, and trigger the graded early warning.

[0012] S5. Based on the matching degree between the early warning results and the actual accident, revise the correlation model.

[0013] As a further improvement to this technical solution, in step S1, the multimodal data includes sensor modal data, visual modal data, and text modal data, wherein:

[0014] Sensor modal data includes dust, noise, and gas concentration data collected by environmental sensors, mechanical vibration and load data collected by equipment sensors, and position data of personnel and equipment collected by positioning sensors.

[0015] Visual modal data includes real-time video from cameras and human behavior features extracted through target detection and pose estimation, the wearing status of safety protective equipment, and intrusion information into dangerous areas;

[0016] Text modal data includes construction plans, safety inspection records, hazard rectification reports, and structured safety indicators extracted through natural language processing.

[0017] As a further improvement to this technical solution, in step S2, the correlation model establishes a mapping relationship table between multimodal features and accident types by statistically analyzing the sensor features, visual features, and text features corresponding to each accident type; and calculates the causal relationship strength between sensor feature nodes, visual feature nodes, and text feature nodes to construct a correlation network including multimodal nodes.

[0018] The initial weights are calculated based on the triggering frequency of each modality in historical accidents and data quality indicators, which are quantified by sensor signal strength, video image clarity, and text field missing rate.

[0019] As a further improvement to this technical solution, in step S3, the preprocessing includes:

[0020] The sensor modal data is denoised, and the short-term change of the detected value after denoising is marked as an abnormal sensor feature node when it exceeds a preset threshold dynamically adjusted according to the construction stage. The preset threshold is obtained by matching the current construction stage with the equipment operating parameter standard library, which generates a dynamic threshold range based on the construction stage type, equipment model, and environmental parameters.

[0021] The visual modal data is processed for illumination and occlusion detection. When unsafety equipment is not worn, intrusion into dangerous areas, or violation of operating procedures is detected, it is marked as an abnormal visual feature node. Unsafety equipment includes not wearing a safety helmet, protective gloves, or high-altitude work safety belt.

[0022] Keyword extraction and timeliness assessment are performed on text modal data. When a keyword matching of a security risk is detected and the timeliness is lower than the preset standard, it is marked as an abnormal text feature node.

[0023] As a further improvement to this technical solution, step S3, dynamically adjusting the multimodal weights, includes:

[0024] For the marked abnormal sensor feature nodes, abnormal visual feature nodes, and abnormal text feature nodes, the weight increase is positively correlated with the magnitude of numerical change, the target danger level, and the keyword matching degree, respectively. The weight increment is then transmitted to the associated sensor feature nodes, visual feature nodes, or text feature nodes according to the strength of the causal relationship through the correlation network.

[0025] For unassociated feature nodes, the weights are reduced by a preset decay rate, which is set according to the mode type.

[0026] As a further improvement to this technical solution, in step S3, a real-time priority queue for weight adjustment tasks is established, and the priority is calculated based on the anomaly type, the percentage of numerical change, and the level of associated accidents; weight adjustment and propagation operations are executed in priority order, and tasks with a priority higher than a preset threshold trigger real-time priority scheduling of computing resources.

[0027] As a further improvement to this technical solution, the priority calculation method includes:

[0028] For tasks involving sensor modal anomalies, the priority is determined by a weighted sum of the proportion of the numerical change to the sensor's measurement range and the severity level of the associated incident.

[0029] For visual modal anomaly tasks, priority is determined by the proximity of the target to the danger zone and the data acquisition time decay coefficient;

[0030] For text modal anomaly tasks, priority is determined by the matching degree between keywords and incident types and the timeliness of the text.

[0031] As a further improvement to this technical solution, in step S4, the correlation network is divided into logical subgraphs according to sensor modality, visual modality, and text modality. The weight propagation within each subgraph is processed in parallel through distributed computing nodes. For cross-modal correlation nodes, an asynchronous message passing mechanism is used to update the weights.

[0032] Sensor modal data, visual modal data, and text modal data correspond to sensor feature nodes, visual feature nodes, and text feature nodes, respectively, and output warning probabilities. A comprehensive warning score is generated by weighted averaging. The corresponding level of warning is triggered by comparing the comprehensive warning score with the preset multi-level warning thresholds.

[0033] As a further improvement to this technical solution, in step S5, matching data between early warning results and actual accidents are collected, false alarms, missed alarms and causal modes are labeled, a correction sample set is formed, and the causal relationship strength of correct association nodes is enhanced and the strength of incorrect association nodes is weakened based on the correction sample set.

[0034] A multimodal construction site safety early warning system with dynamic weight allocation, wherein the multimodal construction site safety early warning system with dynamic weight allocation is used to implement the above-mentioned multimodal construction site safety early warning method with dynamic weight allocation, including:

[0035] The multimodal data acquisition module includes a sensor array, a camera group, and a text data interface;

[0036] The central processing unit, connected to the multimodal data acquisition module, includes a data preprocessing module, a correlation network construction module, a dynamic weight allocation module, a correlation early warning module, and a closed-loop correction module.

[0037] The early warning output module, connected to the central processing unit, includes an audible and visual alarm, a mobile terminal push interface, and a large visual screen.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. The multimodal construction site safety early warning method and system with dynamic weight allocation utilizes a correlation network to construct a mapping relationship between multimodal features and accident types, and processes each modal subgraph in parallel through distributed computing nodes. At the same time, a real-time priority queue for weight adjustment tasks is established, which greatly reduces the computation time and breaks through the bottleneck of slow response and low efficiency of traditional fixed weight early warning methods.

[0040] 2. In this multimodal construction site safety early warning method and system with dynamic weight allocation, a multimodal data acquisition module is deployed to collect three types of data: sensor, visual, and text data. Combined with the dynamic weight allocation mechanism, the weights of each modality can be adjusted in real time according to the anomaly detection results, which significantly improves the accuracy and comprehensiveness of safety hazard identification and effectively avoids the one-sidedness and missed reporting problems of single-modal monitoring.

[0041] 3. In this multimodal construction site safety early warning method and system with dynamic weight allocation, a correction sample set is formed by collecting matching data between early warning results and actual accidents. The correlation model is then closed-loop corrected to continuously enhance the causal relationship strength of correct correlation nodes and weaken erroneous correlations. This ensures that the early warning system can continuously adapt to changes in the construction site environment, making the early warning results highly consistent with actual safety risks and improving the long-term reliability of the system. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0043] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all secondary embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Currently, with the development of construction site informatization, information technology has penetrated into all aspects of construction site production. However, existing construction site safety early warning technologies have many problems that urgently need to be solved, resulting in insufficient accuracy, real-time performance, and adaptability of early warnings.

[0046] In view of this, please refer to Figure 1 As shown, one of the objectives of this invention is to provide a multimodal construction site safety early warning method with dynamic weight allocation. This method includes the following steps:

[0047] S1. By deploying a multimodal data acquisition module, multimodal data from the construction site is collected in real time and transmitted to the central processing unit;

[0048] S2. Based on the historical accident database, construct a correlation model between multimodal data and accident types;

[0049] S3. Perform preprocessing and anomaly detection on the real-time acquired multimodal data, and dynamically adjust the multimodal weights based on the anomaly detection results;

[0050] S4. Input the adjusted multimodal weights into the correlation model, calculate the comprehensive early warning score, and trigger the graded early warning.

[0051] S5. Based on the matching degree between the early warning results and the actual accident, revise the correlation model.

[0052] This dynamic weighted multimodal construction site safety early warning method utilizes a correlation network to construct a mapping relationship between multimodal features and accident types. Distributed computing nodes process each modal subgraph in parallel, while a real-time priority queue for weight adjustment tasks is established, significantly reducing computation time and overcoming the bottlenecks of slow response and low efficiency in traditional fixed-weight early warning methods. Furthermore, by deploying a multimodal data acquisition module to collect sensor, visual, and textual data, combined with a dynamic weighted allocation mechanism, the weights of each modality can be adjusted in real time based on anomaly detection results, significantly improving the accuracy and comprehensiveness of safety hazard identification and effectively avoiding the one-sidedness and missed reporting problems of single-modal monitoring. Simultaneously, by collecting matching data between early warning results and actual accidents to form a correction sample set, the correlation model is closed-loop corrected, continuously strengthening the causal relationship strength of correctly correlated nodes and weakening erroneous correlations. This ensures that the early warning system can continuously adapt to changes in the construction site environment, making the early warning results highly consistent with actual safety risks and improving the long-term reliability of the system.

[0053] Considering that traditional construction site safety early warning technologies rely solely on single-modal data, such as monitoring environmental parameters only through sensors or monitoring personnel behavior only through cameras, there are problems such as one-sided monitoring dimensions, isolated information, and incomplete hazard identification. Specifically, sensors cannot capture behavioral risks such as "personnel not wearing safety equipment", cameras cannot correlate with mechanical operating states such as "equipment overload" in real time, and safety hazards in text records cannot be dynamically matched with real-time scenarios, resulting in high false negative rates and incomplete risk assessment.

[0054] Therefore, step S1 employs multimodal data fusion technology to incorporate sensor modality, visual modality, and text modality into the data acquisition system, achieving comprehensive risk perception through the following technologies:

[0055] For sensor modalities, environmental sensors (dust, noise, and gas concentration sensors) are deployed to collect construction environment parameters in real time, equipment sensors (mechanical vibration and load sensors) monitor the operating status of machinery, and positioning sensors (UWB or GPS positioning) track the location of personnel and equipment. This enables real-time quantitative monitoring of "environmental risks (such as excessive dust and leakage of harmful gases)," "equipment risks (such as abnormal vibration and overload)," and "location risks (such as personnel entering high-risk areas and equipment collision risks)." This provides basic data on "unsafe conditions of objects" for the early warning system and solves the problem that traditional single sensors cannot cover multi-dimensional equipment and environmental risks.

[0056] For visual modalities, real-time video streams are captured via cameras, and combined with computer vision technologies such as object detection (YOLO algorithm) and pose estimation (OpenPose), behavioral features such as "personnel violations (e.g., throwing objects from heights)," "lack of safety equipment (not wearing a helmet / safety belt)," and "intrusion into dangerous areas (entering the foundation pit without permission)" are extracted. This technology fills the gap in traditional sensors' inability to monitor "unsafe human behavior." For example, when the sensor detects that the tower crane load is approaching a threshold, the vision module simultaneously verifies whether the operator is operating according to regulations, avoiding accidents caused by the superposition of risks from "equipment malfunction + personnel violation." By identifying the safety belt wearing status of personnel working at heights through pose estimation, it directly correlates with the risk of "fall from height," upgrading the early warning system from "object monitoring" to "human-object linkage monitoring."

[0057] For text-based modalities, unstructured texts such as construction plans, safety inspection records, and hazard rectification reports are processed using Natural Language Processing (NLP) technology to extract structured indicators such as "hazard type," "rectification timeliness," and "responsible entity" (e.g., locating hazard locations through named entity recognition and extracting rectification deadlines through time expressions). This transforms static management data into dynamic risk indicators. For example, if the text modal detects "a hazard of loose scaffolding bolts in a certain area not rectified within 48 hours," combined with real-time visual data showing "frequent personnel entering and exiting the area," a higher-level warning can be triggered. By assessing the timeliness of text data (e.g., hazards not rectified within the time limit), a correlation analysis of "historical hazards - current status" is formed with real-time sensor and visual data, solving the problem of "disconnect between recording and execution" in traditional management data.

[0058] By integrating three types of modal data, we have achieved data complementarity, full-scenario coverage, and precise early warning, breaking through the bottleneck of traditional early warning systems that are "one-sided data and fragmented scenarios". Through the integration of multiple technologies, we have achieved "full-element perception and full-risk correlation", providing a solid data foundation for subsequent dynamic weight allocation and intelligent early warning.

[0059] Considering that traditional construction site safety early warning models are mostly based on independent analysis of single-modal data, lacking modeling of the correlation between multimodal features, and failing to consider the differences in data quality and accident triggering frequency (e.g., when sensor signals have high noise, video images are blurry, or text records are missing), fixed weights are still used, leading to inaccurate mapping between features and accident types, and significant fluctuations in model early warning accuracy with data quality. Therefore, step S2 employs multimodal correlation modeling technology, using statistical feature mapping, causal relationship analysis, and data quality weighting to construct a highly interpretable and robust correlation model. The specific technical path and effects are as follows:

[0060] Traditional models rely solely on the direct correlation between a single modality's "feature-accident," neglecting the combined effect of multimodal features. For example, "abnormal equipment vibration + personnel not evacuated from the danger zone + overdue rectification of hidden dangers" all point to high-risk accidents. Therefore, based on a historical accident database, we statistically analyze the sensor features (such as vibration amplitude, gas concentration), visual features (such as not wearing a safety belt, intrusion into a danger zone), and text features (such as keywords like "scaffolding hidden dangers" and overdue rectification records) corresponding to each accident type (such as falls from heights, being struck by objects, and mechanical injuries), forming a "feature-accident" mapping matrix. Through this matrix, we achieve cross-indexing of multimodal features. For instance, when "tower crane load sensor values ​​exceed limits" (sensor feature) and "operators not wearing safety helmets enter under the crane boom" (visual feature) are triggered simultaneously, the mapping table quickly locates the "mechanical injury" accident type, avoiding misjudgment based on a single feature and enabling the model to initially possess multi-dimensional risk correlation capabilities.

[0061] Traditional models assume that each modal feature is independent, neglecting potential causal relationships between features. For example, a "missed maintenance plan" in management text may lead to abnormal equipment sensor data. This deficiency results in a lack of risk transmission path analysis, failing to capture the chain risk of "management deficiency → equipment failure → personnel injury". Therefore, Granger causality tests or structural causality models (SCM) are used to calculate the strength of causal relationships between sensor, vision, and text feature nodes. (Value range [0,1]), for example, the causal influence of the "dust concentration exceeding the standard" node on the "workers not wearing dust masks" node, or the transmission probability of the "missing hazard rectification report" node on the "equipment failure" node, and construct a directed acyclic graph (DAG) containing cross-modal associations, thereby clarifying the causal transmission path of multimodal features. For example, when the system identifies "the safety inspection record does not mention edge protection" (text feature), the correlation network will automatically enhance the association weight of "personnel fall risk area" (visual feature), so that the model can simulate the causal chain of "management loophole → behavioral risk → accident occurrence", and improve the foresight and logical interpretability of the early warning;

[0062] Furthermore, traditional models do not differentiate between data quality and accident triggering frequency, treating all features equally. Therefore, in step S2, initial weights are calculated based on the triggering frequency of each modality in historical accidents and data quality indicators. These data quality indicators are quantified using sensor signal strength, video image clarity, and text field missing rate. Features that frequently appear in historical accidents are assigned higher initial weights to ensure that core risk signals are not ignored; features with low data quality are automatically downweighted, such as sensor data with signal strength below a threshold, visual features corresponding to blurry video images, and text features with missing fields, to avoid model misjudgments due to low-quality data and effectively improve the reliability and robustness of the model's initial input.

[0063] Considering the potential for spurious causal relationships among multimodal feature nodes—for example, changes in ambient temperature simultaneously affecting sensor signals and human behavior, leading to incorrect associations of nodes with no direct causal relationship—and the tendency for traditional correlation networks to experience computational delays due to an explosion in the number of nodes in complex scenarios, resulting in inaccurate warnings or delayed responses, this paper employs causal relationship denoising techniques and a hierarchical response optimization architecture when constructing the correlation network. This improves network performance in terms of both association accuracy and computational efficiency.

[0064] From the perspective of causal relationship denoising technology, firstly, the strength of the initially calculated causal relationship (such as the causal strength of node A → node B) is verified through counterfactual reasoning: assuming node A does not experience an anomaly, does the anomaly probability of node B decrease significantly? If the decrease is less than 10%, it is judged as a false association and removed. Simultaneously, domain expert knowledge is introduced to construct a "causal prior rule base," for example, clarifying that "textual hidden danger rectification overdue" necessarily precedes "equipment failure," and there is no reverse causal relationship. This is used to logically verify the causal relationships automatically calculated by the algorithm, filtering out associations with contradictory time sequences. In this way, false causal edges such as the accidental co-occurrence association between low sensor battery power and personnel violations are successfully removed, making the causal relationships in the network closer to actual physical logic. For example, the unidirectional causal chain of "not wearing a safety helmet" → "object strike injury" is strengthened, effectively avoiding weight mispropagation caused by incorrect associations. For instance, a construction site once mistakenly associated "canteen meal time" with "equipment shutdown," and such erroneous warnings were reduced after denoising. Secondly, the minimum lifespan principle in graph theory is applied... The MST (Multi-Tree Tree) algorithm sparsifies dense local subgraphs (such as fully connected subgraphs with more than 10 nodes), retaining core causal paths like "overload → abnormal vibration → loose bolts" while removing irrelevant edges such as "overload → operator's mood". Simultaneously, it sets a dynamic threshold for the strength of causal relationships (initially 0.4, gradually adjusted during model training), retaining only edges with strength greater than the threshold. For cross-modal correlation edges (such as sensor → visual node), the threshold is increased by 0.1 because cross-modal physical correlations are more complex. This series of operations significantly reduces the number of network edges, lowering computational complexity while preserving genuine causal relationships. It avoids confusion in warning logic caused by "proliferation of low-probability correlations." For example, a construction site once triggered simultaneous alarms from 30+ feature nodes due to excessive correlations; after pruning, this problem was completely resolved.

[0065] From the perspective of hierarchical response optimization architecture, firstly, the correlation network is divided into intra-modal sub-networks and inter-modal bridge nodes. Intra-modal sub-networks encompass node associations within sensors, vision systems, and text processing. The sensor sub-network is grouped by equipment type (tower crane / elevator), the vision sub-network by risk area (high-altitude / foundation pit), and the text sub-network by hazard type (equipment / management). Each sub-network independently performs weight propagation on distributed computing nodes, achieving parallel processing. Inter-modal propagation is implemented through an asynchronous message queue. Cross-modal weight propagation is only activated when an intra-modal anomaly triggers a bridge node (e.g., a tower crane load sensor anomaly triggers a risk at the "tower crane operation area" vision node), avoiding meaningless cross-modal calculations and effectively improving computational efficiency. Secondly, accident level weights are labeled for causal relationship edges, such as adding 30% weight to edges pointing to "mass casualties" accidents and subtracting 10% weight to edges pointing to "minor injuries," thus constructing priority-based propagation paths. Dijkstra's algorithm is used. The algorithm variant automatically plans "high-accident-level priority transmission paths" for early warning tasks. For example, when "combustible gas concentration exceeds the standard" (sensor node) is detected, it is prioritized to strongly correlated high-risk nodes such as "personnel in hot work area not wearing protective masks" (visual node) and "fire equipment inspection overdue" (text node), rather than prioritizing low-priority nodes such as "irrelevant equipment vibration". In a simulated gas leak scenario at a chemical construction site, the system prioritizes activating the causal chain of "gas concentration → personnel protection → fire management", quickly completes multimodal weight aggregation and triggers a level-one early warning. Compared with traditional non-priority networks, it responds earlier, gaining crucial and valuable time for emergency response and significantly improving the weight transmission efficiency in high-risk scenarios.

[0066] Considering that traditional models, when processing multimodal data, fail to distinguish between "high-frequency key features" and "low-frequency interference features," and lack effective filtering mechanisms for noisy data (such as sensor signal drift and disordered text formatting), they are often misled by low-quality data. For example, they may mistakenly identify noise in equipment vibration signals as faults, leading to inaccurate warning results. Therefore, this approach assigns initial weights to each modal feature by statistically analyzing its frequency of occurrence in historical accidents. For instance, "not wearing a seatbelt" has an 80% occurrence rate in falls from heights, so this feature is given a higher initial weight to highlight its importance. Simultaneously, a data quality index system is constructed based on the characteristics of different modal data: for sensor modalities, signal strength is used for evaluation; for example, the weight of location data is reduced when GPS positioning signals are weak. For visual modalities, image clarity is used for judgment; the system dynamically reduces the weight of video recognition accuracy in low-light environments. For text modalities, the credibility of text data is reduced based on the field missing rate, particularly for cases where key fields in hazard reports are missing. These methods effectively improve the reliability of the model's initial input. For example, frequently occurring visual features such as "working at heights without a safety belt" are given high weight to ensure that core risk signals are not ignored; while sensor data with signal strength below a threshold are automatically downweighted to avoid false alarms caused by equipment failure or signal noise. Ultimately, this significantly improves the model's robustness in complex scenarios with unstable data quality, reduces false warnings caused by data quality issues, and enhances the accuracy and stability of the early warning system.

[0067] Considering the noise interference, environmental changes, and timeliness differences in multimodal data from construction sites—for example, sensor signals are susceptible to electromagnetic interference causing abnormal fluctuations, cameras may miss violations in low-light or obstructed environments, and hazard information in text records may lose its reference value due to overdue processing—and the inability of traditional fixed-threshold preprocessing methods to adapt to dynamic changes during the construction phase (such as significant differences in equipment operating parameters between pile foundation construction and high-altitude operations), resulting in high false alarm rates and significant missed detection risks, step S3 employs a modal adaptive preprocessing technique. Through dynamic threshold matching, environmental robustness processing, and timeliness analysis, accurate anomaly labeling of multimodal data is achieved. The specific technical path and effects are as follows:

[0068] For sensor modal data, wavelet transform or Kalman filtering is used to reduce noise in the original sensor signal, filtering out high-frequency noise (such as electromagnetic interference) and low-frequency drift (such as sensor temperature drift); the short-term variation amplitude of the values ​​is calculated based on a sliding window. =|Current Value - Average Value of the Previous 5 Minutes|), suppressing interference from random fluctuations; simultaneously, a three-dimensional standard library of "Construction Stage - Equipment Model - Environmental Parameters" is constructed. For example, during the pile foundation construction stage, the vibration threshold of the rotary drilling rig is dynamically adjusted based on soil hardness (obtained through environmental sensors); during high-altitude operations, the load threshold of the suspended platform is combined with wind speed (environmental parameter) and suspended platform model (equipment parameter) to generate a dynamic range. By matching the current construction stage (such as process information obtained from the BIM system) with equipment operating parameters in real time, the corresponding threshold range is called; thereby improving the accuracy of anomaly marking and adapting to complex working conditions;

[0069] For visual modal data, an environment-adaptive preprocessing technique is employed. To mitigate environmental interference, histogram equalization and CycleGAN generative adversarial networks are used to enhance the contrast of low-light images, ensuring high differentiation between the safety helmet (red / yellow) and the background. For occlusion issues, YOLOv8+Deformable DETR is used to capture the outline of partially occluded safety equipment (e.g., only 1 / 3 of the safety helmet edge is visible). For behavioral complexity issues, OpenPose is used to extract keypoint coordinates, and LSTM neural networks are combined to analyze pose sequences of more than 5 consecutive frames (e.g., determining whether "high-altitude workers have been continuously not wearing safety belts") for pose estimation. Semantic segmentation is used to delineate ROIs (regions of interest) such as foundation pits and tower crane operating areas, and optical flow is used to track personnel movement trajectories, distinguishing between "brief passage" and "illegal stay," and detecting whether intrusion into dangerous areas has occurred.

[0070] For text modal data, a BERT-wwm pre-trained model is used to identify specific hazard terms in areas such as "fall risk" and "missing edge," and TF-IDF is used to calculate keyword weights (e.g., the higher the frequency of "not accepted" in the text, the greater the risk weight) to achieve keyword extraction. A knowledge graph is used to link "tower crane model QTZ80" to the corresponding equipment in the sensor modality, achieving cross-modal data association between "text hazard" and "real-time equipment status." Simultaneously, spaCy is used to identify time entities in the text, such as "rectification deadline 2025-04-20" and "inspection date 2025-04-15," and timeliness coefficients are defined. ,in, For the current time, The rectification time should be set for potential hazards. (where S is the decay factor), and when S < 0.5, it is marked as "potential for exceeding the expiration date";

[0071] By employing modal adaptive preprocessing technology, we have achieved "spurious data removal, dynamic adaptation, and multi-source alignment" for multimodal data. Compared with traditional fixed preprocessing methods, this comprehensively improves the accuracy of anomaly detection and lays a solid data foundation for subsequent dynamic weight allocation and comprehensive early warning, ensuring the reliability and real-time performance of the entire early warning system from the source.

[0072] Considering that traditional multimodal early warning models use fixed weight allocation, they cannot respond in real time to the dynamic risk differences of anomalies in different modalities (such as the urgency of sudden changes in sensor values ​​differing from the risk level of expired textual hazards), and that unrelated redundant feature nodes introduce noise interference and consume computational resources, leading to delayed early warnings or a lack of prioritization. Therefore, step S3 employs a dynamic weight adaptive adjustment technique, using a three-pronged mechanism of "anomaly-driven weight enhancement - irrelevant node attenuation - causal network propagation" to achieve precise focusing and coordinated response of multimodal risk signals. The specific technical logic and effects are as follows:

[0073] The weight increase of the marked abnormal sensor feature nodes, abnormal visual feature nodes, and abnormal text feature nodes is positively correlated with the magnitude of numerical change, the target hazard level, and the keyword matching degree, respectively.

[0074] For sensor modes, define a weight boosting function. ,in The change in value. For sensor range, Sensitivity coefficients during the construction phase (such as those of equipment load sensors during high-altitude operations). Set to 1.5, and set to 1 during pile foundation construction. For example: if the tower crane load sensor value suddenly increases by 80% (approaching the fracture threshold), the weight increase reaches 80% × 1.5 = 120%, far exceeding the weight increase of normal anomalies, ensuring that high-risk signals are processed first;

[0075] For visual modalities, a "target hazard level matrix" is established, classifying violations into different risk levels (fall from height risk level 5, not wearing a helmet level 3, not wearing gloves level 1), with varying weight increases. = Grade value × Time decay coefficient. For example, if the violation of "not wearing a helmet" occurs frequently recently, the time decay coefficient will increase, further increasing the weight of the behavior.

[0076] For text-based modalities, keyword matching scores (ranging from 0 to 1) are calculated using the BERT model, combined with timeliness scores (with an additional 30% weighting for potential delays), and the weighting increase is [not specified]. = Matching degree × (1 + Timeliness correction coefficient). For example, when the keyword "loose scaffolding bolts" is detected and the rectification period has expired, the weight increase is significantly increased.

[0077] The above strategy enables the precise amplification of risk signals. For example, if the vibration sensor value of a construction site elevator suddenly increases (high risk), the weight increases sharply, and the vision module is triggered to detect whether "personnel are in the elevator car". Compared with the fixed weight model, the response speed of high-risk anomalies is improved, and key signals are avoided from being submerged due to the average weight distribution.

[0078] Traditional models only independently weight single-modal anomalous nodes, neglecting "cross-modal causal relationships" (such as "overloading of equipment" potentially leading to "personnel misjudging the safe distance and entering a danger zone"), failing to create a "ripple effect" of risk transmission, and resulting in the underestimation of multimodal superposition risks. Therefore, a correlation network (causal relationship strength) is used to address this issue. ∈[0,1]), the weight increments are based on the strength of causality. The proportion is transmitted to related nodes; for example, the causal strength between "not wearing a safety belt at height" (visual node, weight increased by 50%) and the "fall from height accident" node is 0.8, while the causal strength between "safety net not set up" (text node) and "safety net not set up" is 0.6. Then the latter receives a weight increase of 50% × 0.6 = 30%.

[0079] Considering that traditional models do not perform weight decay on irrelevant feature nodes, leading to computational redundancy, and that random fluctuations of irrelevant features (such as irrelevant sensor data far from the work area) may mislead early warning results, the weights of unrelated feature nodes are reduced according to a preset decay rate, which is set according to the modality type. For sensor modalities, unrelated nodes are decayed at 5% / minute (because equipment status data is highly real-time, and outdated signals have rapidly reduced value). For visual modalities, the decay rate for unrelated routine behavior nodes (such as "normal walking") is 3% / minute, and the decay rate for dangerous behavior nodes is 1% / minute (to retain short-term risk memory). For text modalities, historical hazard nodes are decayed at a rate of 1%. attenuation( (The number of days remaining until the rectification deadline) Nodes that are more than 30 days overdue have weights that approach 0. At the same time, nodes with a weight of less than 0.1 after attenuation are automatically masked and do not participate in subsequent comprehensive early warning calculations, thereby improving calculation efficiency and reducing noise interference.

[0080] Considering the significant differences in risk levels and urgency of handling various multimodal anomalies at construction sites (e.g., sudden spikes in sensor values ​​may trigger immediate accidents, while delayed textual risks represent progressive risks), traditional indiscriminate sequential processing methods easily lead to "high-risk tasks being blocked by low-priority tasks," resulting in delays or even missed detections of critical warnings. Furthermore, computing resources cannot be allocated as needed, leading to inefficiency and waste. Therefore, step S3 employs real-time priority scheduling technology. By constructing a dynamic priority queue and resource triggering mechanism, it achieves "risk level-driven task scheduling and on-demand computing resource allocation." The specific technical path and effects are as follows:

[0081] Establish a real-time priority queue for weight adjustment tasks. The priority is calculated based on the anomaly type, the percentage of numerical change, and the level of associated incidents. Weight adjustment and propagation operations are executed in priority order. Tasks with a priority higher than a preset threshold trigger real-time priority scheduling of computing resources.

[0082] For sensor modes, their priority is calculated by a weighted sum of the proportion of the numerical change amplitude to the sensor's measurement range and the severity level of the associated accident. The formula is: Priority Where β+γ=1, The change in value. The range is defined by Sr, which represents the severity level of the associated accident, from 1 to 5. For example, a sudden 80% increase in tower crane load (approaching the fracture threshold) =0.8) and associated with a "mass casualty" accident level 5, =0.6×0.8+0.4×5=2.48 (highest priority);

[0083] For visual modalities, priority is determined by the proximity of the target to the danger zone and the data acquisition time attenuation coefficient, calculated using the following formula: Priority ,in, The distance of the target from the danger zone. Safety threshold distance The data collection time interval For example, when personnel are only 1 meter from the edge of the pit (d / D=0.2), and the latest frame of the real-time video stream is detected (…). =0.1), =0.8 × 0.9 = 0.72 (medium to high priority);

[0084] For text modalities, priority is determined by the matching degree between keywords and accident types, as well as the timeliness of the text. The calculation formula is: Priority =Keyword matching degree × (1 + timeliness correction coefficient), the correction coefficient increases by 0.1 for each day of delay, up to a maximum of +0.5;

[0085] This priority quantification model enables precise quantification of the urgency of abnormal task risks, ensuring that high-risk tasks are handled first and avoiding delays in critical early warnings.

[0086] Based on the above priority quantization model, a real-time priority queue is established: the priority queue is implemented using efficient data structures such as binary heaps or Fibonacci heaps, ensuring that the insertion / retrieval time complexity of the highest priority task is O(logn), and supporting the concurrent processing of tens of thousands of tasks; the system performs weight adjustment and propagation operations according to priority order, ensuring that high-risk abnormal tasks are processed first, effectively avoiding the situation where "high-risk tasks are blocked by low-priority tasks".

[0087] Traditional fixed resource allocation models suffer from two major problems: First, high-priority tasks must wait for current computation tasks to complete, leading to delays in critical early warnings. For example, an urgent task involving sensor values ​​exceeding their range by 50% may fail to trigger an early warning in time due to waiting for text data preprocessing. Second, distributed computing nodes experience uneven load distribution, with some nodes idle and others overloaded. To address these issues, this technology employs threshold-triggered resource scheduling, which sets priority thresholds... =1.5 (can be preset), when the task When triggered, the following actions are taken: local resource preemption, pausing low-priority processes and releasing CPU / GPU resources to prioritize urgent tasks; distributed scaling, dynamically requesting additional computing nodes through the Kubernetes cluster to achieve parallel weight propagation across nodes, for example, synchronously computing sensor subgraphs and visual subgraphs on different nodes; and resource usage isolation technology, allocating dedicated cache space for high-priority tasks, such as partitioning GPU memory, to avoid bandwidth contention caused by sharing resources with regular tasks and to ensure the computing resource requirements of high-priority tasks.

[0088] Considering that traditional construction site safety early warning systems often suffer from false alarms and missed alarms due to independent judgment of a single modality (such as relying solely on sensor readings without considering personnel behavioral risks), and lack fine-grained differentiation of risk levels (such as uniformly triggering the same level of warning leading to misallocation of emergency resources), they cannot meet the actual needs of "precise early warning and tiered response." Therefore, step S4 adopts multimodal probability fusion and tiered early warning technology. Through weighted aggregation of early warning probabilities and intelligent matching of multi-level thresholds, it achieves a leap from "single indicator judgment" to "comprehensive risk assessment." The specific technical logic and effects are as follows:

[0089] Traditional early warning systems directly use raw data (such as sensor values ​​and visual detection results) for judgment, which has significant drawbacks: First, the data is not comparable; continuous sensor values, discrete labels from visual detection, and matching scores from text analysis cannot be directly integrated. Second, there are limitations due to single-modality limitations; a single modality is prone to misjudgment due to data noise or environmental interference, such as temporary sensor malfunctions causing abnormal values ​​or obstructed cameras leading to missed violations. Third, there is a lack of confidence; the reliability of the detection results is not quantitatively assessed. For example, the reliability of visual detection of safety helmets is low in low-light conditions, but it is still included in the judgment with equal weight. Therefore, it is necessary to first perform probabilistic transformation on the data from each modality.

[0090] For sensor modes, based on the Gaussian mixture model (GMM), the probability Ps of the current sensor value deviating from the normal range is calculated. For example, by analyzing the distribution of historical data, the possibility that the current tower crane load value exceeds the safe range can be determined.

[0091] For visual modalities, the confidence score Pv output by the object detection algorithm (such as YOLOv8) is directly used as the probability of the corresponding behavior (such as not wearing a seat belt) occurring, with a value ranging from 0 to 1;

[0092] For the text modality, the matching probability Pt between hidden danger keywords and accident types in the text is calculated using the BERT classifier, and the final probability is adjusted by combining the timeliness score.

[0093] This step transforms heterogeneous multimodal data into probability values ​​in the 0-1 range, laying the foundation for subsequent fusion calculations.

[0094] Considering that traditional simple average weighting (such as equal weighting of each modality) ignores real-time risk differences, the risk contribution of each modality varies in different construction stages and scenarios. For example, in high-altitude operations, the visual modality (safety belt wearing status) is more critical than the sensor modality. Furthermore, the risk level of a minor anomaly in a single modality differs from that of simultaneous anomalies in multiple modalities, but fixed weights cannot reflect this difference. Therefore, the real-time adjusted weights output in step S3 are directly adopted. , , This includes enhancement of abnormal nodes, incremental causal transmission, and attenuation of unrelated nodes, ensuring that the weights reflect the actual importance of each mode under the current operating conditions; the comprehensive early warning score formula is: This formula, through normalization, ensures that changes in the total weight do not affect the comparability of scores, and achieves accurate aggregation of risk signals through dynamic weights. For example, when a sensor detects abnormal equipment vibration ( =0.7), while visually detected personnel had not evacuated the danger zone ( =0.8), the text indicates that the relevant hidden dangers have not been rectified ( =0.6), dynamic weights will enhance the contribution of these high-risk modes, significantly increasing the overall score S and accurately reflecting the superimposed risks.

[0095] Traditional early warning systems use a single threshold to trigger alarms, which has significant drawbacks. First, it lacks risk grading; minor anomalies (such as dust concentration exceeding the standard by 10%) and severe anomalies (gas concentration exceeding the standard by 100%) trigger the same level of alarm, leading to a workload for safety personnel. Second, the response measures are simplistic, failing to link different handling plans according to risk levels; for example, high-risk scenarios require immediate shutdown, while low-risk scenarios only require recording and alerting. Therefore, this embodiment constructs a three-level threshold system:

[0096] Level 1 Warning (S≥0.8): Red warning, triggering audible and visual alarms, automatic equipment shutdown, and SMS notification to the project manager (response time <1 second).

[0097] Level 2 warning (0.6≤S<0.8): Yellow warning, push notification to APP pop-up and voice broadcast, and arrange for the safety supervisor to conduct a special verification (response time <5 seconds).

[0098] Level 3 warning (0.4≤S<0.6): Blue warning, the system records and archives the data, and generates an inspection work order (response time <30 seconds).

[0099] Meanwhile, the system automatically optimizes threshold boundaries based on historical accident data and updates them quarterly to adapt to long-term changes such as the skill level of construction site personnel and the aging of equipment.

[0100] Considering that traditional construction site safety early warning models rely on initial training data and lack a feedback mechanism for actual accidents, they are prone to problems such as "accumulated false alarms and missed alarms" and "failure to adapt to new scenarios" in long-term use (e.g., risk features introduced by new processes are not recognized by the model, and old related nodes still dominate the early warning logic), making it impossible to achieve closed-loop optimization of "data-model-effect". Therefore, step S5 adopts a model self-evolution technology based on a corrected sample set. Through a closed-loop process of "abnormal data labeling - causal relationship recalibration - model parameter iteration", an intelligent early warning system that can learn autonomously and continuously optimize is constructed. The specific technical path and effects are as follows:

[0101] Traditional models have significant shortcomings when processing real-world accident data. Firstly, they fail to handle false alarms effectively. False alarms caused by sensor noise, for example, are not properly labeled, leading to persistent incorrect associations between "noise characteristics" and "accident types." For instance, electromagnetic interference noise from sensors might be incorrectly associated with equipment malfunctions. Secondly, the causes of missed alarms are unclear. For missed alarms, the causal modality is not traced. For example, when multimodal risks are not identified, factors such as "missed management text entries" are not included in the association model, resulting in repeated missed detections of similar risks. To address these issues, a revised sample set is constructed and structured annotations are applied. These structured annotations include:

[0102] The warning result type is clearly labeled as correct warning, false alarm (no accident triggered), and missed alarm (an accident occurred but no warning was issued), clearly distinguishing the warning effect of the model;

[0103] Cause modality labeling accurately determines whether it is a single modality misjudgment (such as visual occlusion causing the seat belt to be missed) or a lack of multimodal linkage (such as sensor malfunction not being transmitted to related text nodes), thus locating the root cause of the problem;

[0104] Feature node contribution labeling analyzes the impact of each feature node on the early warning results using SHAP values, accurately identifying nodes that are "overly correlated" (contribution > reasonable threshold) or "lacking correlation" (contribution < reasonable threshold). For example, if a feature node has an excessively high SHAP value, it indicates over-correlation in the early warning process, potentially leading to false alarms; conversely, if the SHAP value is too low, it may indicate missing correlation, resulting in missed alarms.

[0105] The above annotations enable the corrected sample set to comprehensively and accurately reflect the false positives and false negatives of the model and their causes, providing strong data support for subsequent model optimization.

[0106] In traditional models, causal relationships are fixed and cannot adapt to new scenarios and actual accident feedback. Therefore, the strength of causal relationships is dynamically adjusted based on the modified sample set.

[0107] For "correctly associated nodes" (such as "not wearing a seatbelt - fall from height"), follow the formula: This strengthens the causal relationship (A is the sample contribution, obtained through SHAP value analysis, and η is the learning rate). For example, if the association between "not wearing a seatbelt" and "falling from a height" is correct and has a high contribution in the corrected samples, this formula can further strengthen the causal relationship between the two, making the model pay more attention to such associations in subsequent warnings;

[0108] For "erroneous association nodes" (such as "irrelevant equipment vibration - object impact"), according to the formula This weakens the causal relationship (B is the number of misclassifications), reducing it to a minimum of 0.1. For example, if "irrelevant equipment vibration" and "object impact" are repeatedly misclassified as a correlation in the corrected samples, this formula can gradually weaken the relationship between the two, preventing the model from being misled by erroneous associations.

[0109] To more effectively process the corrected sample set, the corrected samples are input into a graph neural network (GNN). Specifically, the edge weights of the entire correlation network are updated through graph convolution operations. For "missing correlation edges corresponding to missed incidents," additional weights are added to compensate for their absence, ensuring they can be captured by the model in subsequent warnings. For "redundant edges with high-frequency false alarms," ​​suppression is implemented to reduce their impact on warning results. Simultaneously, the application of GNNs enables the model to better handle the complex relationships between multimodal data. Through optimization of the correlation network, the model can more quickly adapt to changes in risk characteristics brought about by new scenarios and processes, reducing the occurrence of false alarms and missed alarms.

[0110] Please see Figure 2As shown, a second objective of this invention is to provide a multimodal construction site safety early warning system with dynamic weight allocation. This system is used to implement the aforementioned multimodal construction site safety early warning method with dynamic weight allocation, and includes:

[0111] The multimodal data acquisition module includes a sensor array, a camera group, and a text data interface;

[0112] The central processing unit, connected to the multimodal data acquisition module, includes a data preprocessing module, a correlation network construction module, a dynamic weight allocation module, a correlation early warning module, and a closed-loop correction module.

[0113] The early warning output module, connected to the central processing unit, includes an audible and visual alarm, a mobile terminal push interface, and a large visual screen.

[0114] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multimodal construction site safety early warning method with dynamic weight allocation, characterized in that, Includes the following steps: S1. By deploying a multimodal data acquisition module, multimodal data from the construction site is collected in real time and transmitted to the central processing unit; S2. Based on the historical accident database, construct a correlation model between multimodal data and accident types; in step S2, the correlation model is established by statistically analyzing the sensor features, visual features, and text features corresponding to each accident type to create a mapping table between multimodal features and accident types; calculate the causal relationship strength between sensor feature nodes, visual feature nodes, and text feature nodes to construct a correlation network including multimodal nodes; The initial weights are calculated based on the triggering frequency of each modality in historical accidents and data quality indicators, which are quantified by sensor signal strength, video image clarity, and text field missing rate. S3. Perform preprocessing and anomaly detection on the real-time acquired multimodal data, and dynamically adjust the multimodal weights based on the anomaly detection results; In step S3, the preprocessing includes: The sensor modal data is denoised, and the short-term change of the detected value after denoising is marked as an abnormal sensor feature node when it exceeds a preset threshold dynamically adjusted according to the construction stage. The preset threshold is obtained by matching the current construction stage with the equipment operating parameter standard library, which generates a dynamic threshold range based on the construction stage type, equipment model, and environmental parameters. The visual modal data is processed for illumination and occlusion detection. When unsafety equipment is not worn, intrusion into dangerous areas, or violation of operating procedures is detected, it is marked as an abnormal visual feature node. Unsafety equipment includes not wearing a safety helmet, protective gloves, or high-altitude work safety belt. Keyword extraction and timeliness assessment are performed on text modal data. When a keyword match for a security risk is detected and the timeliness is lower than the preset standard, it is marked as an abnormal text feature node. In step S3, dynamically adjusting the multimodal weights includes: For the marked abnormal sensor feature nodes, abnormal visual feature nodes, and abnormal text feature nodes, the weight increase is positively correlated with the magnitude of numerical change, the target danger level, and the keyword matching degree, respectively. The weight increment is then transmitted to the associated sensor feature nodes, visual feature nodes, or text feature nodes according to the strength of the causal relationship through the correlation network. For unassociated feature nodes, the weights are reduced by a preset decay rate, which is set according to the mode type; In step S3, a real-time priority queue for weight adjustment tasks is established. The priority is calculated based on the anomaly type, the percentage of numerical change, and the level of associated accidents. Weight adjustment and propagation operations are executed in priority order. Tasks with a priority higher than a preset threshold trigger real-time priority scheduling of computing resources. S4. Input the adjusted multimodal weights into the correlation model, calculate the comprehensive early warning score, and trigger the graded early warning. S5. Based on the matching degree between the early warning results and the actual accident, revise the correlation model.

2. The multimodal construction site safety early warning method with dynamic weight allocation according to claim 1, characterized in that, In step S1, the multimodal data includes sensor modal data, visual modal data, and text modal data, wherein: Sensor modal data includes dust, noise, and gas concentration data collected by environmental sensors, mechanical vibration and load data collected by equipment sensors, and position data of personnel and equipment collected by positioning sensors. Visual modal data includes real-time video from cameras and human behavior features extracted through target detection and pose estimation, the wearing status of safety protective equipment, and intrusion information into dangerous areas; Text modal data includes construction plans, safety inspection records, hazard rectification reports, and structured safety indicators extracted through natural language processing.

3. The multimodal construction site safety early warning method with dynamic weight allocation according to claim 1, characterized in that: The priority calculation method includes: For tasks involving sensor modal anomalies, the priority is determined by a weighted sum of the proportion of the numerical change to the sensor's measurement range and the severity level of the associated incident. For visual modal anomaly tasks, priority is determined by the proximity of the target to the danger zone and the data acquisition time decay coefficient; For text modal anomaly tasks, priority is determined by the matching degree between keywords and incident types and the timeliness of the text.

4. The multimodal construction site safety early warning method with dynamic weight allocation according to claim 1, characterized in that: In step S4, the correlation network is divided into logical subgraphs according to sensor modality, visual modality, and text modality. The weight propagation within each subgraph is processed in parallel through distributed computing nodes. For cross-modal correlation nodes, the weights are updated using an asynchronous message passing mechanism. Sensor modal data, visual modal data, and text modal data correspond to sensor feature nodes, visual feature nodes, and text feature nodes, respectively, and output warning probabilities. A comprehensive warning score is generated by weighted averaging. The corresponding level of warning is triggered by comparing the comprehensive warning score with the preset multi-level warning thresholds.

5. The multimodal construction site safety early warning method with dynamic weight allocation according to claim 1, characterized in that: In step S5, matching data between early warning results and actual accidents are collected, false alarms, missed alarms, and causal modes are labeled, and a correction sample set is formed. Based on the correction sample set, the causal relationship strength of correct correlation nodes is enhanced, and the strength of incorrect correlation nodes is weakened.

6. A multimodal construction site safety early warning system with dynamic weight allocation, wherein the multimodal construction site safety early warning system with dynamic weight allocation is used to implement the multimodal construction site safety early warning method with dynamic weight allocation according to any one of claims 1-5, characterized in that, include: The multimodal data acquisition module includes a sensor array, a camera group, and a text data interface; The central processing unit, connected to the multimodal data acquisition module, includes a data preprocessing module, a correlation network construction module, a dynamic weight allocation module, a correlation early warning module, and a closed-loop correction module. The early warning output module, connected to the central processing unit, includes an audible and visual alarm, a mobile terminal push interface, and a large visual screen.

Images