Intelligent construction site danger source identification and early warning system based on deep learning

Abstract

The invention relates to the technical field of building construction safety management, in particular to an intelligent building construction site danger source identification and early warning system based on deep learning. Comprising a multi-dimensional dynamic target sensing module which is used for collecting real-time video stream data and resolving a state vector of a dynamic target; the spatio-temporal trajectory prediction module is used for predicting a future state sequence of the dynamic target in a preset time window; the dynamic danger coupling quantitative evaluation module is used for determining predicted coupling duration and judging whether the predicted coupling duration is greater than a preset early warning time threshold value or not so as to generate an early warning decision signal; and the multi-channel intelligent early warning module is used for responding to the early warning decision signal and feeding back the early-warned environment change as new real-time video stream data to the multi-dimensional dynamic target sensing module. According to the system, traditional afterward response type safety management is converted into beforehand pre-judgment type active intervention, intervention can be carried out before a dangerous coupling state is evolved into an actual accident, and the safety of a construction site is improved.

Description

Technical Field

[0001] This invention relates to the field of construction safety management technology, specifically to a deep learning-based intelligent construction site hazard identification and early warning system. Background Technology

[0002] With the deepening development of a new round of technological revolution and industrial transformation, new-generation information technologies such as artificial intelligence, big data, and the Internet of Things are accelerating their integration with traditional industries, driving all sectors toward intelligent and digital transformation. Against this backdrop, intelligent construction has become the core direction for the transformation and upgrading of the construction industry, aiming to deeply integrate advanced information technology with engineering construction technology to achieve informatization and intelligent management covering the entire life cycle of a project, thereby improving the quality, efficiency, and safety of engineering projects.

[0003] As a dynamic, open, and complex working environment, construction sites involve frequent interactions and cross-operations among various production elements such as people, machines, and materials, which constitute the main source of accident risks. Therefore, how to use intelligent means to improve the inherent safety level of construction sites and transform traditional passive, post-event response-based safety management into proactive, pre-event prediction-based risk intervention is a key issue that urgently needs to be addressed in the field of intelligent construction.

[0004] However, current safety management methods at construction sites still have significant limitations. Traditional methods rely heavily on manual inspections by safety officers and routine video surveillance. Manual inspections suffer from limited monitoring range, susceptibility to subjective factors, and untimely responses. Traditional video surveillance systems are limited in function, mostly used only for post-event tracing and evidence collection, lacking the ability to analyze and predict dynamic risks in real time. These existing technologies struggle to comprehensively model and assess key risk factors such as the movement trajectories and working postures of multiple dynamic targets on the construction site, including workers, construction machinery, and transport vehicles, as well as blind spots caused by equipment structures. This results in hazard identification remaining at a static and isolated level, unable to effectively address the sudden and complex hazards arising from the dynamic coupling of multiple targets.

[0005] Therefore, there is an urgent need in this field for an innovative technical solution that can be deeply integrated with the concept of intelligent construction to address the pain points of the existing safety management model. This solution should be able to achieve comprehensive and continuous perception of dynamic targets at the construction site, accurate prediction of movement trends, interactive risk quantification assessment, and multi-channel intelligent early warning, and build a closed-loop management system from risk prediction to proactive intervention, truly moving the safety checkpoint forward, thereby improving the safety assurance capabilities of intelligent construction sites without significantly increasing labor costs. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses a deep learning-based intelligent construction site hazard identification and early warning system. Specifically, the technical solution of this invention is as follows:

[0007] A deep learning-based intelligent construction site hazard identification and early warning system includes:

[0008] The multi-dimensional dynamic target perception module is used to collect real-time video stream data and calculate the state vector of the dynamic target based on a preset hybrid deep learning model.

[0009] The spatiotemporal trajectory prediction module is used to predict the future state sequence of a dynamic target within a preset time window based on the historical state sequence of the target and using a preset trajectory prediction model.

[0010] The dynamic hazard coupling quantification assessment module is used to determine the predicted coupling duration based on the future state sequence and the preset hazard triggering threshold, and to determine whether the predicted coupling duration is greater than the preset warning time threshold, so as to generate a warning decision signal.

[0011] The multi-channel intelligent early warning module is used to respond to early warning decision signals, execute preset early warning strategies based on the category and location of dynamic targets, and feed back the environmental changes after the early warning as new real-time video stream data to the multi-dimensional dynamic target perception module.

[0012] Preferably, the state vector represents the centroid position coordinates, instantaneous velocity vector, instantaneous acceleration vector, attitude and orientation vector, and category label of the dynamic target;

[0013] Preferably, the spatiotemporal trajectory prediction module adopts a trajectory prediction model based on a long short-term memory network, which takes the historical state sequence of the dynamic target as input and outputs the future position and future velocity within a preset time window;

[0014] Preferably, the dynamic hazard coupling quantitative assessment module is also used to generate a dynamic interaction risk sequence based on the future state sequence; the predicted coupling duration is determined based on the dynamic interaction risk sequence and the hazard triggering threshold;

[0015] Preferably, the dynamic interactive risk sequence is generated based on predicted relative speed, predicted distance, weight of hazardous operation status, and visual blind spot influence factor;

[0016] Preferably, the weight of hazardous operation status is determined based on the predicted posture and category label of dynamic target, combined with a pre-set industry safety operation procedure knowledge base;

[0017] Preferably, the steps for determining the visual blind spot impact factor include:

[0018] Based on the predicted attitude and category label of the dynamic target, the preset mechanical three-dimensional blind zone geometric model is invoked;

[0019] Determine whether the predicted position of the target falls within the mechanical three-dimensional blind zone geometric model;

[0020] If it falls into the range, the visual blind spot impact factor will be set as the preset penalty value;

[0021] If it does not fall within the range, the visual blind spot impact factor will be set as the baseline value;

[0022] Preferably, the step of determining the predicted coupling duration includes: determining the maximum time span during which the dynamic interaction risk sequence continuously exceeds the danger triggering threshold within a preset time window as the predicted coupling duration;

[0023] Preferred, preset early warning strategies include:

[0024] Send targeted alerts to the smart safety helmets of specific workers;

[0025] Activate the audible and visual alarm in the cab of the construction machinery;

[0026] And audible and visual warning posts that activate the boundaries of dangerous areas.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. This system uses a deep learning model to predict the trajectory of dynamic targets at the construction site and transforms traditional reactive safety management into proactive intervention based on pre-emptive prediction. It can intervene before a dangerous coupling state evolves into an actual accident, thereby improving the safety of the construction site.

[0029] 2. This system can perform intelligent monitoring of the site at all times and without blind spots, and can detect and track various dynamic targets such as workers and construction machinery in real time and accurately. It overcomes the shortcomings of traditional manual inspection, such as insufficient coverage and untimely response, as well as the fact that traditional video surveillance can only trace back after the fact.

[0030] 3. The system establishes a multi-dimensional dynamic hazard quantification assessment model, which comprehensively considers multiple factors such as predicted relative speed, predicted distance, hazardous operation status and visual blind spots, and transforms abstract hazard relationships into precise numerical values ​​for assessment, thereby accurately identifying and predicting complex hazards that emerge from the interaction of multiple dynamic elements.

[0031] 4. The system constructs a closed-loop control system from perception, prediction, assessment to early warning and feedback. It can send directional, multi-channel, and precise early warnings to devices such as smart safety helmets, mechanical cabs, and sound and light stakes at the boundaries of dangerous areas based on the target category and location involved in the hazard. It achieves efficient and automated proactive safety intervention without increasing manpower costs. Attached Figure Description

[0032] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0033] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0035] Example 1:

[0036] Please see Figure 1 A deep learning-based intelligent construction site hazard identification and early warning system includes:

[0037] The multi-dimensional dynamic target perception module is used to collect real-time video stream data and calculate the state vector of the dynamic target based on a preset hybrid deep learning model.

[0038] The spatiotemporal trajectory prediction module is used to predict the future state sequence of a dynamic target within a preset time window based on the historical state sequence of the target and using a preset trajectory prediction model.

[0039] The dynamic hazard coupling quantification assessment module is used to determine the predicted coupling duration based on the future state sequence and the preset hazard triggering threshold, and to determine whether the predicted coupling duration is greater than the preset warning time threshold, so as to generate a warning decision signal.

[0040] The multi-channel intelligent early warning module is used to respond to early warning decision signals, execute preset early warning strategies based on the category and location of dynamic targets, and feed back the environmental changes after the early warning as new real-time video stream data to the multi-dimensional dynamic target perception module.

[0041] This invention provides a deep learning-based intelligent construction site hazard identification and early warning system. The system is designed as a closed-loop, predictive intelligent safety monitoring system. Its technical purpose is to transform traditional post-event response-based safety management into pre-event predictive proactive intervention by continuously sensing, predicting, quantifying risks, and providing intelligent early warnings of dynamic targets at the construction site. Without significantly increasing the cost of human supervision, it effectively suppresses sudden dangers caused by high-frequency and disorderly interactions between dynamic elements and proactively controls the actual duration of hazard coupling below the safety threshold.

[0042] In this embodiment, the system includes four core modules that work together to form a complete closed loop of perception-prediction-evaluation-early warning-feedback technology;

[0043] The multi-dimensional dynamic target perception module is designed to detect and continuously track various dynamic targets within the construction site in real time and accurately from uninterrupted video information. In this embodiment, the module collects real-time video stream data through multiple high-frame-rate, wide-angle cameras deployed on-site. This data is fed into a pre-trained hybrid deep learning model. This hybrid deep learning model refers to a composite model consisting of a target detection network and a multi-target tracking network cascaded together. The target detection network can be a YOLO series model, and the multi-target tracking network can be the DeepSORT algorithm.

[0044] The model is trained through supervised learning using a dataset of images and videos containing a large number of construction site scenes, enabling it to efficiently identify and distinguish targets such as workers, various types of construction machinery, transport vehicles, and large materials. The output of this module is a state vector of the dynamic target generated for each identified and continuously tracked dynamic target, which is continuously transmitted to the spatiotemporal trajectory prediction module.

[0045] The spatiotemporal trajectory prediction module is designed to predict the trajectory of a dynamic target within a short future time window based on its past motion state. In this embodiment, the module receives a sequence of historical states of the dynamic target from the multi-dimensional dynamic target perception module. The core of this module is a pre-trained trajectory prediction model, which learns and masters the typical motion patterns of different types of dynamic targets under various working conditions, thereby achieving high-precision trajectory prediction. The module outputs a sequence of future states of the dynamic target within a preset time window, which is a key input for subsequent hazard quantification assessment.

[0046] The dynamic hazard coupling quantification assessment module is designed to quantify and determine the probability and severity of hazardous interactions between any two or more dynamic targets within a future time window. In this embodiment, the module receives a future state sequence output by the spatiotemporal trajectory prediction module. Based on this future state sequence and a hazard triggering threshold set according to statistical laws, it calculates a predicted coupling duration. The predicted coupling duration refers to the length of time a high-risk state is expected to last continuously. The module compares this predicted coupling duration with a warning time threshold set to filter out transient interference. The warning time threshold is determined by analyzing historical data, statistically analyzing the duration distribution of false hazardous coupling events caused by sensor noise or instantaneous model prediction bias, and selecting the 99th percentile of this distribution as the threshold to ensure that only stable hazardous events with a duration exceeding the threshold trigger a warning. If the former is greater than the latter, it means that a stable hazardous event requiring intervention is about to occur, at which point the module generates a warning decision signal.

[0047] The multi-channel intelligent early warning module selects and executes the most effective early warning intervention measures after confirming that a hazard is about to occur. In this embodiment, after responding to the early warning decision signal generated by the dynamic hazard coupling quantitative assessment module, the module automatically executes a set of predefined early warning strategies based on the category and real-time location of the dynamic target involved in the hazard. After the early warning strategy is executed, the behavior of on-site personnel or machinery will change. This change will be captured by the multi-dimensional dynamic target perception module as new environmental information, forming new real-time video stream data and feeding it back to the system entry point, thus forming a complete closed-loop control. The current model assumes that the early warning can effectively guide avoidance behavior and form an ideal negative feedback. Future research can combine behavioral science models to model the complex reactions that personnel may have after the early warning, such as delay, panic, or incorrect operation, in order to further improve the effectiveness and robustness of the system in real-world scenarios.

[0048] This embodiment constructs a complete closed-loop intelligent monitoring system through the collaborative work of the above modules, from pre-event prediction to in-event intervention and post-event feedback. It overcomes the shortcomings of traditional manual inspections, such as insufficient coverage and untimely response, as well as the limitations of traditional video surveillance, which can only trace back after the fact. This system can proactively predict complex dangers arising from the interaction of multiple dynamic elements and intervene in potential dangerous coupling states before they evolve into actual accidents. Thus, without increasing manpower costs, it can proactively suppress the duration of dangerous coupling within a safe threshold, greatly improving the inherent safety level of construction sites.

[0049] Example 2:

[0050] The state vector represents the centroid position coordinates, instantaneous velocity vector, instantaneous acceleration vector, attitude and orientation vectors, and category label of a dynamic target;

[0051] Based on Example 1, this embodiment provides a more specific definition of the state vector. The state vector refers to a data structure used to comprehensively and accurately describe the multi-dimensional physical state of a dynamic target at a specific moment. Its function is to provide standardized and information-rich input for subsequent trajectory prediction and risk assessment.

[0052] In this embodiment, the state vector specifically represents the centroid position coordinates, instantaneous velocity vector, instantaneous acceleration vector, attitude and orientation vector, and category label of the dynamic target;

[0053] Centroid position coordinates: refers to the geometric center position of the target in the three-dimensional world coordinate system. It is obtained by the multi-dimensional dynamic target perception module through visual algorithms to reconstruct the three-dimensional position information in the camera coordinate system from the two-dimensional position information in the camera coordinate system.

[0054] Instantaneous velocity vector and instantaneous acceleration vector: These refer to the kinematic parameters of the target in three-dimensional space. Together, they constitute the basic description of the target's motion trend. They are obtained by performing first-order and second-order difference calculations on the changes in the centroid position coordinates between consecutive frames.

[0055] Attitude and orientation vectors: These are parameters that describe the target's rotational state and direction of movement. For mechanical equipment, these parameters are further used to describe the relative angles and angular velocities of its key operating components. They are derived from video images through attitude estimation algorithms.

[0056] Category label: refers to the classification information of the target, which comes from the direct output of the target detection network in the hybrid deep learning model.

[0057] Example 3:

[0058] The spatiotemporal trajectory prediction module adopts a trajectory prediction model based on long short-term memory network. It takes the historical state sequence of the dynamic target as input and outputs the future position and future velocity within a preset time window.

[0059] Based on Example 1, this embodiment provides a detailed description of the trajectory prediction model used in the spatiotemporal trajectory prediction module.

[0060] In this embodiment, the trajectory prediction model employs a Long Short-Term Memory (LSTM) network-based model. This LSTM-based trajectory prediction model is a recurrent neural network particularly suitable for processing and predicting time-series data. Its underlying logic lies in its unique gated recurrent unit structure, which enables it to effectively learn and memorize long-term dependencies in time-series data. This is crucial for accurately predicting the future trajectory of a target based on its motion state over the past few seconds. The model is pre-trained using a dataset containing a large number of real construction site dynamic target trajectories, and the training process aims to minimize the error between the model's predicted trajectory and the actual trajectory.

[0061] During runtime, the model takes the historical state sequence of the dynamic target over a period of time, i.e., the time sequence of the state vector defined in Example 2, as its input; through forward propagation calculation of the network, the model can output the future state sequence of the target within a preset time window, which specifically includes the future position and future velocity of the target at each future moment.

[0062] Example 4:

[0063] The dynamic hazard coupling quantification assessment module is also used to generate dynamic interaction risk sequences based on future state sequences; the predicted coupling duration is determined based on the dynamic interaction risk sequence and the hazard triggering threshold.

[0064] The steps to determine the predicted coupling duration include: determining the maximum time span during which the dynamic interaction risk sequence continuously exceeds the danger triggering threshold within a preset time window as the predicted coupling duration;

[0065] Based on Example 1, this embodiment further elaborates on the working method of the dynamic hazard coupling quantitative assessment module;

[0066] In this embodiment, after receiving the future state sequence, the dynamic hazard coupling quantification assessment module generates a dynamic interaction risk sequence based on the sequence. The dynamic interaction risk sequence refers to a time series data used to describe the change of the interaction risk of two or more dynamic targets over time within a future time window. Its function is to transform abstract hazard relationships into quantifiable values.

[0067] The steps for determining the predicted coupling duration are as follows: The system compares each risk value in the dynamic interactive risk sequence with a hazard trigger threshold. This hazard trigger threshold is a critical risk value used to define the dangerous and safe states. It is determined based on the statistical distribution analysis of risk indices in a large amount of historical safety and near-miss event data, selecting a quantile value that can achieve a preset balance between detection sensitivity and false alarm rate, such as the 95th percentile. The system identifies all time intervals in the dynamic interactive risk sequence that are continuously higher than the hazard trigger threshold within a preset time window and calculates the duration of each interval. Finally, the maximum time span among these continuous time intervals is determined as the predicted coupling duration.

[0068] Example 5:

[0069] The dynamic interactive risk sequence is generated based on predicted relative speed, predicted distance, weight of hazardous operation status, and visual blind spot influence factor;

[0070] The weights of hazardous operation states are determined based on the predicted posture and category labels of dynamic targets, combined with a pre-set industry safety operation procedure knowledge base;

[0071] Based on Example 1, this embodiment provides a detailed explanation of the basis for generating the dynamic interactive risk sequence. For a typical human-computer interaction scenario involving a worker and a piece of construction machinery, at a certain point in the future... The risks of dynamic interaction are quantified and defined as follows:

[0072]

[0073] in:

[0074] The dynamic interaction risk between personnel p and machinery m at a future time τ, with the dimension of time. The reciprocal of, that is In physical terms, it can be understood as the frequency of dangerous events or the rate of risk growth.

[0075] A non-negative small constant is set to prevent the denominator from being zero, which is used to enhance the numerical stability of the model. Its dimensions are consistent with those of the distance.

[0076] The static risk term, which is only related to distance, is used to characterize the potential danger of close contact even when the relative velocity is zero; its dimensions are consistent with those of the dynamic risk term. For example, this function can be specifically defined as a piecewise function:

[0077]

[0078] in It is a safe distance threshold preset according to industry standards. It is a static risk constant determined through historical data statistics or expert evaluation;

[0079] The system risk calibration coefficient is a dimensionless global adjustment parameter used to calibrate the overall risk sensitivity of the system. To clarify the technical significance of this parameter, its calibration process is as follows: A calibration dataset is constructed consisting of historical near-miss events and safety events. Each event sample contains a set of independent observation variables, such as the relative speed and distance between humans and machines, the machine's operating status, and blind zone relationships during the event. Based on this dataset, regression analysis methods such as least squares are used to fit the optimal coefficient. The value ensures that the risk value calculated by the model has the highest correlation with the actual risk level of the event;

[0080] : respectively personnel With machinery In the future The predicted velocity vector;

[0081] :personnel With machinery Center of mass in the future moment The predicted Euclidean distance;

[0082] :mechanical The weights of hazardous operation states are dimensionless parameters.

[0083] :personnel Compared to machinery The visual blind spot influence factor is a dimensionless parameter.

[0084] In the above formula, the independent variable parameter and The source is calculated by the spatiotemporal trajectory prediction module, and the parameters are... It is calculated based on the future position output by the spatiotemporal trajectory prediction module;

[0085] As shown in the formula above, the dynamic interactive risk sequence is generated based on predicted relative speed, predicted distance, weights of hazardous operation states, and visual blind spot influence factors. The design of this formula ensures that the model's behavior conforms to physical principles and system safety requirements even in extreme cases such as when the predicted distance dpm(τ) approaches 0 or the relative speed is 0. In the risk quantification model, risk is directly proportional to relative speed and inversely proportional to distance, as shown in the formula. The model uses weighting factors... and Multiplicative adjustment aligns with the fundamental safety logic that higher speeds, closer distances, hazardous operating conditions, and blind spots all carry higher risks. In other embodiments, more complex nonlinear functions can be employed to improve physical fidelity; for example, the distance term 1 / (dpm(τ)+ Replace α / dpm(τ) with exp(α / dpm(τ)) or a piecewise function to more accurately characterize the characteristic that the risk increases sharply as it approaches the critical safety distance;

[0086] Hazardous operation status weights The purpose is to integrate industry safety standards into the risk model, so that the risk assessment can reflect the additional dangers brought about by the specific operation of machinery. The weight is determined based on the predicted posture and category label of the dynamic target, combined with a pre-built industry safety operation procedure knowledge base. The industry safety operation procedure knowledge base is a database or lookup table that stores the risk coefficients of different models of construction machinery under different operating conditions. For example, when the system determines that the excavator is in a swing state based on the predicted posture, it will query the corresponding higher weight value from the knowledge base; when it is in a stationary state, it is assigned a baseline value of 1. For other types of dynamic target interactions, such as interactions between construction machinery (machine-to-machine), a similar risk quantification model can be used, but the weighting factors will be different.

[0087] For complex scenarios involving three or more objectives, the system can calculate the interaction risk of all objective pairs and define the global dynamic interaction risk within a preset time window as the maximum value among all risk pairs. Where i and j are any different targets within the scene; the general risk function is used here. Depending on the type of interaction objective, such as human-machine or machine-machine interaction, the corresponding risk quantification model is adopted; for example, two pieces of construction machinery. and The interaction risk between them can be quantified as follows:

[0088]

[0089] The weighting factors are determined independently based on the respective hazardous operation status;

[0090] It should be noted that using the maximum value among all risk pairs as the global risk is a simplified approach that is effective in most scenarios. In subsequent implementations, more complex risk fusion models can be studied to assess the synergistic risk effects that may arise when multiple dynamic objectives interact simultaneously.

[0091] To further approximate complex real-world physical scenarios, those skilled in the art will understand that the aforementioned risk quantification model can be further expanded in subsequent implementations to incorporate risk factors from more dimensions. For example, the directionality of relative motion can be further considered in risk assessment by projecting the relative velocity vector onto the vector connecting the centers of mass of the two targets to obtain the radial relative velocity, assigning a higher risk weight when the targets are close to each other. For engineering machinery with large working components, the predicted distance... This can be optimized to the distance between the center of mass of the personnel and the nearest point on the outline of the mechanical 3D model, in order to address the risk misjudgment caused by the irregular geometry of the target. In addition, by introducing environmental sensors, such as light sensors and humidity sensors, data such as light influence factors and ground slippage coefficients can be added to the risk model, thereby constructing a more comprehensive and realistic dynamic risk assessment system, further improving the system's practicality.

[0092] To ensure the robustness of the system, in addition to setting small constants... In addition to handling computational stability and instantaneous noise through warning time thresholds, the system should also include abnormal input processing logic in actual deployment. For example, setting reasonable thresholds for the velocity and acceleration output by the spatiotemporal trajectory prediction module can smooth out or eliminate abrupt data caused by problems such as re-identification after target tracking loss. When the perception module briefly loses the target, a state estimation method should be used to maintain the target's virtual trajectory for a short period of time to ensure the continuity of risk assessment. These measures together constitute a stress test and robustness guarantee for the model.

[0093] Example 6:

[0094] The steps to determine the impact factors of visual blind spots include:

[0095] Based on the predicted attitude and category label of the dynamic target, the preset mechanical three-dimensional blind zone geometric model is invoked;

[0096] Determine whether the predicted position of the target falls within the mechanical three-dimensional blind zone geometric model;

[0097] If it falls into the range, the visual blind spot impact factor will be set as the preset penalty value;

[0098] If it does not fall within the range, the visual blind spot impact factor will be set as the baseline value;

[0099] Based on Example 1, this embodiment focuses on the influence factors of visual blind spots. The determination steps are explained in detail; the purpose of the blind spot impact factor is to quantify the additional risks caused by obstructed vision of the mechanical driver;

[0100] The steps to determine this factor include:

[0101] Based on the predicted attitude and category label of the dynamic target, the system calls a pre-built mechanical three-dimensional blind zone geometric model. This mechanical three-dimensional blind zone geometric model is a three-dimensional spatial geometric model that is pre-built according to the factory specifications of a specific model of machinery and can change in real time with the attitude of the machinery.

[0102] The system determines whether the predicted position of the target falls within the spatial range covered by the aforementioned mechanical three-dimensional blind zone geometric model.

[0103] If the judgment result is "falling into," it means that the person will enter the driver's blind spot, and the risk is significantly amplified. The system will determine the visual blind spot impact factor as a preset penalty value. The preset penalty value is a constant greater than the baseline value of 1, for example, a value of 2. Its specific value can be calibrated by expert scoring or combined with historical accident data. The principle is that this value should be able to significantly increase the risk level of interactive events that occur in the blind spot. If the judgment result is "not falling into," the visual blind spot impact factor is determined to be the baseline value of 1.

[0104] Example 7:

[0105] The preset early warning strategies include:

[0106] Send targeted alerts to the smart safety helmets of specific workers;

[0107] Activate the audible and visual alarm in the cab of the construction machinery; and

[0108] Activate the audible and visual warning posts at the boundary of the danger zone;

[0109] Based on Example 1, this embodiment provides a detailed description of the early warning strategy executed by the multi-channel intelligent early warning module. This early warning strategy is a set of rules that automatically selects the optimal early warning channel and method based on the specific participants, location, and severity of a dangerous event. Its purpose is to achieve accurate, effective, and low-interference early warning.

[0110] In this embodiment, the early warning strategy includes, but is not limited to, a combination or individual execution of the following methods:

[0111] One strategy is to send directional alarms to the smart helmets of specific workers; when the system identifies that a hazard mainly involves a specific person, it will trigger the smart helmet worn by that person to emit a directional vibration or voice alarm via the wireless communication module.

[0112] Another strategy is to activate the audible and visual alarm in the cab of the construction machinery; when the main party responsible for the danger or the party avoiding it is the machine operator, the system will activate the buzzer or warning light in their cab and can highlight the location and type of the hazard on the cab screen.

[0113] Another strategy is to activate the audible and visual warning posts at the boundaries of dangerous areas. For wide-area risks that may affect multiple targets within an area, or when danger occurs at critical passages, the system will activate the audible and visual warning posts deployed at the boundaries of the area to warn all personnel in the vicinity.

[0114] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0115] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based intelligent construction site hazard identification and early warning system, characterized in that, include: The multi-dimensional dynamic target perception module is used to collect real-time video stream data and calculate the state vector of the dynamic target based on a preset hybrid deep learning model. The spatiotemporal trajectory prediction module is used to predict the future state sequence of a dynamic target within a preset time window based on the historical state sequence of the target and using a preset trajectory prediction model. The dynamic hazard coupling quantification assessment module is used to determine the predicted coupling duration based on the future state sequence and the preset hazard triggering threshold, and to determine whether the predicted coupling duration is greater than the preset warning time threshold, so as to generate a warning decision signal. The multi-channel intelligent early warning module is used to respond to early warning decision signals, execute preset early warning strategies based on the category and location of dynamic targets, and feed back the environmental changes after the early warning as new real-time video stream data to the multi-dimensional dynamic target perception module.

2. The intelligent construction site hazard identification and early warning system based on deep learning according to claim 1, characterized in that, The state vector represents the centroid position coordinates, instantaneous velocity vector, instantaneous acceleration vector, attitude and orientation vectors, and category label of a dynamic target.

3. The intelligent construction site hazard identification and early warning system based on deep learning according to claim 1, characterized in that, The spatiotemporal trajectory prediction module adopts a trajectory prediction model based on long short-term memory networks. It takes the historical state sequence of the dynamic target as input and outputs the future position and future velocity within a preset time window.

4. The intelligent construction site hazard identification and early warning system based on deep learning according to claim 1, characterized in that, The dynamic hazard coupling quantification assessment module is also used to generate dynamic interaction risk sequences based on future state sequences; the predicted coupling duration is determined based on the dynamic interaction risk sequence and the hazard triggering threshold.

5. The intelligent construction site hazard identification and early warning system based on deep learning according to claim 4, characterized in that, The dynamic interactive risk sequence is generated based on predicted relative speed, predicted distance, weight of hazardous operation status, and visual blind spot influence factor.

6. The intelligent construction site hazard identification and early warning system based on deep learning according to claim 5, characterized in that, The weights of hazardous operation states are determined based on the predicted posture and category labels of dynamic targets, combined with a pre-set industry safety operation procedure knowledge base.

7. The intelligent construction site hazard identification and early warning system based on deep learning according to claim 5, characterized in that, The steps to determine the impact factors of visual blind spots include: Based on the predicted attitude and category label of the dynamic target, the preset mechanical three-dimensional blind zone geometric model is invoked; Determine whether the predicted position of the target falls within the mechanical three-dimensional blind zone geometric model; If it falls into the range, the visual blind spot impact factor will be set as the preset penalty value; If it does not fall within the range, the visual blind spot impact factor will be set as the baseline value.

8. The intelligent construction site hazard identification and early warning system based on deep learning according to claim 4, characterized in that, The steps for determining the predicted coupling duration include: determining the maximum time span during which the dynamic interaction risk sequence continuously exceeds the danger trigger threshold within a preset time window as the predicted coupling duration.

9. The intelligent construction site hazard identification and early warning system based on deep learning according to claim 1, characterized in that, The preset early warning strategies include: Send targeted alerts to the smart safety helmets of specific workers; Activate the audible and visual alarm in the cab of the construction machinery; And audible and visual warning posts that activate the boundaries of dangerous areas.

Images