Building construction risk assessment method based on deep learning

Abstract

The invention discloses a building construction risk assessment method based on deep learning. The method comprises the following steps: collecting multi-source data of a building construction site and preprocessing the multi-source data; segmenting the preprocessed multi-source data, and recombining the segmented multi-source data into a space-time sample in combination with a construction area identifier; performing feature extraction and fusion to form a fusion feature vector; constructing a training data set and a verification data set; inputting the training data set into the improved ASTGCN model for training, and verifying through the verification data set to obtain a trained improved ASTGCN model; outputting a risk assessment result; obtaining a risk prediction value through a risk trend prediction network; the risk classification early warning information is generated, multi-source data driving characterization, dynamic quantitative evaluation and prospective trend early warning of construction site risks are achieved, and the accuracy of risk identification and the timeliness of early warning in a complex construction scene are improved.

Description

Technical Field

[0001] This invention relates to the field of building construction safety and intelligent risk assessment, and in particular to a building construction risk assessment method based on deep learning. Background Technology

[0002] Construction processes are characterized by openness, dynamism, and high risk. Construction sites often involve multiple overlapping work processes, dense movement of personnel and equipment, continuously changing environmental conditions, and high-risk operations such as those near edges, at heights, in deep foundation pits, and involving lifting and hoisting. To prevent accidents, construction companies typically need to continuously assess on-site risks and implement corresponding early warning and control measures accordingly. Existing construction risk assessment technologies mainly include manual inspection and experience-based judgment, rule-based assessment methods based on specifications and checklists, and risk prediction methods based on statistics or traditional machine learning.

[0003] Manual inspections and experience-based judgments rely on safety officers' on-site observations and subjective judgments, typically combining construction safety specifications, hazard identification checklists, and risk classification guidelines for risk identification and classification. While these methods are widely used in engineering practice, their assessment results are significantly influenced by personnel experience and a sense of responsibility, making it difficult to establish stable and consistent quantitative standards. Furthermore, manual inspections are limited by time and space, often exhibiting random and intermittent nature, making it difficult to respond in real-time to rapid changes in risk conditions, resulting in a significant lag. Especially in large-scale, complex projects or high-intensity, overlapping work scenarios, manual inspections cannot cover all areas and time periods, increasing the probability of missed or misjudged risks. Summary of the Invention

[0004] One objective of this invention is to propose a construction risk assessment method based on deep learning. This invention fully utilizes multi-source data fusion from the construction site, an improved ASTGCN model, and risk trend prediction technology to achieve automatic identification, quantitative assessment, and forward-looking prediction of risks at the construction site. It has the advantages of strong risk characteristic representation capabilities and high real-time assessment results.

[0005] A construction risk assessment method based on deep learning according to an embodiment of the present invention includes the following steps:

[0006] Collect and preprocess multi-source data from construction sites;

[0007] The preprocessed multi-source data is segmented according to the time window and recombined into spatiotemporal samples by combining the construction area identifiers;

[0008] Feature extraction and fusion of spatiotemporal samples are performed to form a fused feature vector;

[0009] Training and validation datasets were constructed based on fused feature vectors and corresponding risk annotations.

[0010] The improved ASTGCN model is trained by inputting the training dataset and then validated using the validation dataset to obtain the trained improved ASTGCN model.

[0011] Real-time collection of multi-source data from the construction site and input into the trained improved ASTGCN model; output risk assessment results.

[0012] The risk assessment results are combined in chronological order to form a risk assessment sequence, which is then input into a risk trend prediction network to obtain risk prediction values.

[0013] Based on the predicted risk values, risk classification early warning information is generated through preset risk classification rules.

[0014] Optionally, the multi-source data includes construction site video image data, environmental and structural monitoring sensor data, construction equipment operating condition data, personnel operation behavior data, construction progress data, and BIM model data. The preprocessing includes data denoising, missing data completion, and abnormal data removal.

[0015] Optionally, the formation of the spatiotemporal samples specifically includes:

[0016] The preprocessed multi-source data are aligned according to a unified time base to form a continuous multi-source data sequence;

[0017] Set the time window length and the time window sliding step size, extract continuous multi-source data sequences according to the time window length to obtain the first time window, and successively move the corresponding continuous multi-source sequences according to the time window sliding step size to form time window multi-source data;

[0018] Window samples are formed by recombining and aggregating multi-source data within a time window with construction area identifiers.

[0019] The window samples are associated with the corresponding spatial location information to form spatiotemporal samples that include both temporal continuity and spatial distribution.

[0020] Optionally, the formation of the fused feature vector specifically includes:

[0021] The spatiotemporal samples are decomposed according to the data source to obtain subsets of video images, sensor and equipment operating conditions, personnel work behavior, construction progress, and BIM model.

[0022] Visual information is extracted from a subset of video images to generate a visual feature set. The extraction process involves detecting and locating targets such as personnel, equipment, components, and hazards in the subset of video images within each time window, obtaining target categories and spatial locations, continuously tracking detected targets to obtain target motion trajectories, performing time-series identification of personnel and equipment work behaviors, obtaining hazard behavior labels, and forming a visual feature set.

[0023] Numerical information is extracted from a subset of sensor and equipment operating conditions to generate a numerical feature set. The extraction process involves performing unified dimension and normalization processing on the sampling sequences of sensors and equipment within each time window to obtain the statistical characteristics and rate of change characteristics of each sampling sequence, and generating threshold exceedance markers and equipment anomaly markers to form a numerical feature set.

[0024] Behavioral information is extracted from a subset of personnel work behaviors to generate a set of process and component status features. The extraction process involves processing the process and component status information of a subset of construction progress and a subset of BIM models. Within each time window, the personnel work location sequence, action sequence, work duration, and violation work markers are obtained to form a set of behavioral features. The component type, spatial relationship, component status, and location of dangerous parts corresponding to the process are extracted from the BIM model to form a set of process and component status features.

[0025] The visual feature set, numerical feature set, behavioral feature set, and process and component status feature set are unified in dimension and aligned in feature, and then combined according to the preset fusion rules to obtain the fused feature vector.

[0026] Optionally, the construction of the training dataset and the validation dataset specifically includes:

[0027] Number each sample of the fusion feature vector, and record the corresponding time identifier and construction area identifier for each fusion feature vector to obtain the fusion feature sample set.

[0028] The risk labeling records corresponding to the fusion feature sample set are organized. Based on the time identifier and construction area identifier of each sample in the fusion feature sample set, the risk level label is obtained from the risk labeling records. The risk level labels are then summarized by sample number to form a risk label set.

[0029] The risk labeling set is divided into a training dataset and a validation dataset according to a preset division ratio.

[0030] Optionally, obtaining the improved ASTGCN model after training specifically includes:

[0031] The training dataset is input into the improved ASTGCN model for training. The improved ASTGCN model includes a dynamic graph construction module, an adaptive spatial attention module, and a multi-scale temporal attention and temporal convolution module. The construction dynamic graph construction module refers to establishing and updating an initial graph structure based on spatial adjacency, process dependence, and interaction strength, using construction area identifiers as nodes, to obtain a dynamic graph structure. The adaptive spatial attention module refers to introducing an evidence-based attention calibration mechanism under the dynamic graph structure, calculating the spatial association strength between nodes, and obtaining spatially enhanced node features. The multi-scale temporal attention and temporal convolution module refers to performing multi-temporal-scale attention modeling on the spatially enhanced node features and obtaining a spatiotemporal joint representation of the nodes through temporal convolution.

[0032] In the construction dynamic graph construction module, the construction area identifiers in the training dataset are used as graph nodes, and the spatial adjacency, process dependency, and interaction intensity relationships between construction areas within the time window are used as node connection relationships to form an initial graph structure. During the training iteration process, the node connection relationships are updated according to the fusion feature vector of the current time window to obtain the construction dynamic graph structure.

[0033] In the adaptive spatial attention module, risk propagation direction constraint relationships between nodes are constructed based on the construction dynamic map structure and the fused feature vectors within the corresponding time window. The construction process involves obtaining adjacent nodes based on the construction dynamic map structure and combining the spatial positional relationship, process sequence relationship, and interaction intensity change trend in the fused feature vectors to form risk propagation direction constraint relationships between nodes. By introducing an evidential attention calibration mechanism, the spatial association strength between nodes is calculated, and spatially enhanced node features are generated. The calculation process involves using the evidential attention calibration mechanism to calculate the feature correlation score of adjacent node pairs that satisfy the risk propagation direction constraints, thereby obtaining the spatial coupling degree of the node pairs and forming the spatial association strength.

[0034] In the multi-scale temporal attention and temporal convolution module, the spatially enhanced node features are arranged into a temporal feature sequence according to the order of the time window. The temporal attention weights are calculated on the temporal feature sequence according to a preset time scale to obtain a multi-scale temporal attention representation. The multi-scale temporal attention representation is then subjected to temporal convolution to obtain a spatiotemporal joint representation of the node.

[0035] Nonlinear feature compression is performed on the spatiotemporal joint representation of nodes to generate risk output values ​​corresponding to each training sample. The sample error is calculated based on the deviation between the risk level label of the training sample and the risk output value. The average value of all sample errors is used as the training loss.

[0036] The improved ASTGCN model is iteratively trained based on the training loss. The validation error is calculated using the validation dataset. The iterative training stops when the validation error meets the preset convergence condition, resulting in the trained improved ASTGCN model.

[0037] Optionally, the output of the risk assessment results specifically includes:

[0038] Real-time acquisition of multi-source data from the construction site and feature extraction and fusion are performed to obtain a real-time fused feature vector corresponding to the current time window;

[0039] The real-time fused feature vector is input into the trained improved ASTGCN model. The construction dynamic map structure is updated by the construction dynamic map construction module, the spatially enhanced node features are extracted by the adaptive spatial attention module, and the spatiotemporal joint representation of the nodes is obtained by the multi-scale temporal attention and temporal convolution module, generating real-time risk output values.

[0040] The real-time risk output value is associated with the corresponding time marker and construction area marker to form the risk assessment result.

[0041] Optionally, obtaining the risk prediction value specifically includes:

[0042] The risk assessment results are cached and arranged in chronological order to form a risk assessment sequence;

[0043] The risk assessment sequence is input into the risk trend prediction network, which performs time-series inference based on the risk assessment sequence and outputs the risk prediction value for each prediction time within the corresponding prediction time window.

[0044] Optionally, the generation of the risk classification and early warning information specifically includes:

[0045] The risk prediction value is identified. When the risk prediction value is in the form of a level, each level is converted into a corresponding numerical measure according to the preset risk classification rules. When the risk prediction value is in the form of a probability, the numerical value is used as the basic measure, and the basic measure is subjected to a unified scale conversion process to form a risk measure value.

[0046] The risk measurement value is compared with the risk classification threshold in the preset risk classification rules to obtain the warning level. The comparison process is to determine the risk measurement value and the risk classification threshold from low to high interval. When the risk measurement value is within the interval of the risk classification threshold, the classification label corresponding to the interval is used as the warning level.

[0047] By combining the warning level, time marker, construction area marker, and risk assessment results, a risk-level warning information is generated.

[0048] The beneficial effects of this invention are:

[0049] This invention preprocesses, segments time windows, and reassembles spatiotemporal samples from multiple sources, including construction site video images, sensor and equipment operating conditions, personnel work behavior, construction progress, and BIM data, under a unified time reference and spatial identifier. It further completes cross-modal feature extraction, alignment, and fusion, solving the problems of difficulty in uniformly expressing multi-source information and difficulty in covering complex scenarios with manual indicators in existing technologies from a data perspective. This makes the input representation of construction risks more comprehensive, objective, and spatiotemporally continuous.

[0050] At the model level, this invention employs an improved ASTGCN model for risk assessment training. A dynamic graph construction module builds a time-updated dynamic graph structure of the construction area and its spatial adjacency, process dependencies, and interaction strength relationships. Furthermore, an adaptive spatial attention module introduces risk propagation direction constraints and an evidence-based attention calibration mechanism to screen and reconstruct effective spatial interaction features, significantly enhancing the model's ability to characterize the propagation and coupling triggering mechanisms of risk across different regions. Simultaneously, multi-scale temporal attention and temporal convolution modules jointly model the short-term abrupt changes and long-term cumulative effects of risk, improving the accuracy of identifying and fitting the dynamic risk evolution process. Therefore, this invention achieves automated, nonlinear, and spatiotemporally correlated risk assessment of complex construction scenarios without relying on fixed rules and manual weights, reducing missed and false positives.

[0051] In terms of application effectiveness, this invention not only outputs the risk assessment results corresponding to the current time and construction area, but also inputs the continuous risk assessment results into a sequence into a risk trend prediction network to obtain the risk prediction values ​​within the future prediction time window. Based on preset risk grading rules, it generates risk grading warning information including warning levels, corresponding times, and construction area identifiers, achieving a forward-looking closed loop of "current assessment—trend prediction—grading warning." Compared to existing static or post-event assessment methods, this invention significantly improves the real-time nature, accuracy, and predictability of risk assessment, providing earlier and more actionable warning windows for construction sites, thereby enhancing the proactive prevention and control capabilities and overall safety level of construction safety management. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is an overall flowchart of a deep learning-based construction risk assessment method proposed in this invention;

[0054] Figure 2This is a schematic diagram illustrating the construction of a fusion feature vector for a deep learning-based construction risk assessment method proposed in this invention.

[0055] Figure 3 This is a schematic diagram of the improved ASTGCN model for a deep learning-based construction risk assessment method proposed in this invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figures 1-3 A deep learning-based method for assessing construction risks includes the following steps:

[0058] Collect and preprocess multi-source data from construction sites;

[0059] The preprocessed multi-source data is segmented according to the time window and recombined into spatiotemporal samples by combining the construction area identifiers;

[0060] Feature extraction and fusion of spatiotemporal samples are performed to form a fused feature vector;

[0061] Training and validation datasets were constructed based on fused feature vectors and corresponding risk annotations.

[0062] The improved ASTGCN model is trained by inputting the training dataset and then validated using the validation dataset to obtain the trained improved ASTGCN model.

[0063] Real-time collection of multi-source data from the construction site and input into the trained improved ASTGCN model; output risk assessment results.

[0064] The risk assessment results are combined in chronological order to form a risk assessment sequence, which is then input into a risk trend prediction network to obtain risk prediction values.

[0065] Based on the predicted risk values, risk classification early warning information is generated through preset risk classification rules.

[0066] In this embodiment, the multi-source data includes construction site video image data, environmental and structural monitoring sensor data, construction equipment operating condition data, personnel operation behavior data, construction progress data, and BIM model data. The preprocessing includes data denoising, missing data completion, and abnormal data removal.

[0067] In this embodiment, the formation of the spatiotemporal sample specifically includes:

[0068] The preprocessed multi-source data are aligned according to a unified time base to form a continuous multi-source data sequence;

[0069] Set the time window length and the time window sliding step size, extract continuous multi-source data sequences according to the time window length to obtain the first time window, and successively move the corresponding continuous multi-source sequences according to the time window sliding step size to form time window multi-source data;

[0070] Window samples are formed by recombining and aggregating multi-source data within a time window with construction area identifiers.

[0071] The window samples are associated with the corresponding spatial location information to form spatiotemporal samples that include both temporal continuity and spatial distribution.

[0072] In this embodiment, the formation of the fused feature vector specifically includes:

[0073] The spatiotemporal samples are decomposed according to the data source to obtain subsets of video images, sensor and equipment operating conditions, personnel work behavior, construction progress, and BIM model.

[0074] Visual information is extracted from a subset of video images to generate a visual feature set. The extraction process involves detecting and locating targets such as personnel, equipment, components, and hazards in the subset of video images within each time window, obtaining target categories and spatial locations, continuously tracking detected targets to obtain target motion trajectories, performing time-series identification of personnel and equipment work behaviors, obtaining hazard behavior labels, and forming a visual feature set.

[0075] Numerical information is extracted from a subset of sensor and equipment operating conditions to generate a numerical feature set. The extraction process involves performing unified dimension and normalization processing on the sampling sequences of sensors and equipment within each time window to obtain the statistical characteristics and rate of change characteristics of each sampling sequence, and generating threshold exceedance markers and equipment anomaly markers to form a numerical feature set.

[0076] Behavioral information is extracted from a subset of personnel work behaviors to generate a set of process and component status features. The extraction process involves processing the process and component status information of a subset of construction progress and a subset of BIM models. Within each time window, the personnel work location sequence, action sequence, work duration, and violation work markers are obtained to form a set of behavioral features. The component type, spatial relationship, component status, and location of dangerous parts corresponding to the process are extracted from the BIM model to form a set of process and component status features.

[0077] The visual feature set, numerical feature set, behavioral feature set, and process and component status feature set are unified in dimension and aligned in feature, and then combined according to the preset fusion rules to obtain the fused feature vector.

[0078] In this embodiment, the construction of the training dataset and the validation dataset specifically includes:

[0079] The fusion feature vectors are numbered one by one according to the samples, and the corresponding time identifier and construction area identifier are recorded for each fusion feature vector to obtain the fusion feature sample set.

[0080] The risk labeling records corresponding to the fusion feature sample set are organized. Based on the time identifier and construction area identifier of each sample in the fusion feature sample set, the risk level label is obtained from the risk labeling records. The risk level labels are then summarized by sample number to form a risk label set.

[0081] The risk labeling set is divided into a training dataset and a validation dataset according to a preset division ratio.

[0082] In this embodiment, obtaining the improved ASTGCN model after training specifically includes:

[0083] The training dataset is input into the improved ASTGCN model for training. The improved ASTGCN model includes a dynamic graph construction module, an adaptive spatial attention module, and a multi-scale temporal attention and temporal convolution module. The construction dynamic graph construction module refers to establishing and updating an initial graph structure based on spatial adjacency, process dependence, and interaction strength, using construction area identifiers as nodes, to obtain a dynamic graph structure. The adaptive spatial attention module refers to introducing an evidence-based attention calibration mechanism under the dynamic graph structure, calculating the spatial association strength between nodes, and obtaining spatially enhanced node features. The multi-scale temporal attention and temporal convolution module refers to performing multi-temporal-scale attention modeling on the spatially enhanced node features and obtaining a spatiotemporal joint representation of the nodes through temporal convolution.

[0084] In the construction dynamic graph construction module, the construction area identifiers in the training dataset are used as graph nodes, and the spatial adjacency, process dependency, and interaction intensity relationships between construction areas within the time window are used as node connection relationships to form an initial graph structure. During the training iteration process, the node connection relationships are updated according to the fusion feature vector of the current time window to obtain the construction dynamic graph structure.

[0085] In the adaptive spatial attention module, risk propagation direction constraint relationships between nodes are constructed based on the construction dynamic map structure and the fused feature vectors within the corresponding time window. The construction process involves obtaining adjacent nodes based on the construction dynamic map structure and combining the spatial positional relationship, process sequence relationship, and interaction intensity change trend in the fused feature vectors to form risk propagation direction constraint relationships between nodes. By introducing an evidential attention calibration mechanism, the spatial association strength between nodes is calculated, and spatially enhanced node features are generated. The calculation process involves using the evidential attention calibration mechanism to calculate the feature correlation score of adjacent node pairs that satisfy the risk propagation direction constraints, thereby obtaining the spatial coupling degree of the node pairs and forming the spatial association strength.

[0086] In the multi-scale temporal attention and temporal convolution module, the spatially enhanced node features are arranged into a temporal feature sequence according to the order of the time window. The temporal attention weights are calculated on the temporal feature sequence according to a preset time scale to obtain a multi-scale temporal attention representation. The multi-scale temporal attention representation is then subjected to temporal convolution to obtain a spatiotemporal joint representation of the node.

[0087] Nonlinear feature compression is performed on the spatiotemporal joint representation of nodes to generate risk output values ​​corresponding to each training sample. The sample error is calculated based on the deviation between the risk level label of the training sample and the risk output value. The average value of all sample errors is used as the training loss.

[0088] The improved ASTGCN model is iteratively trained based on the training loss. The validation error is calculated using the validation dataset. The iterative training stops when the validation error meets the preset convergence condition, resulting in the trained improved ASTGCN model.

[0089] In this embodiment, the output of the risk assessment results specifically includes:

[0090] Real-time acquisition of multi-source data from the construction site and feature extraction and fusion are performed to obtain a real-time fused feature vector corresponding to the current time window;

[0091] The real-time fused feature vector is input into the trained improved ASTGCN model. The construction dynamic map structure is updated by the construction dynamic map construction module, the spatially enhanced node features are extracted by the adaptive spatial attention module, and the spatiotemporal joint representation of the nodes is obtained by the multi-scale temporal attention and temporal convolution module, generating real-time risk output values.

[0092] The real-time risk output value is associated with the corresponding time marker and construction area marker to form the risk assessment result.

[0093] In this embodiment, obtaining the risk prediction value specifically includes:

[0094] The risk assessment results are cached and arranged in chronological order to form a risk assessment sequence;

[0095] The risk assessment sequence is input into the risk trend prediction network, which performs time-series inference based on the risk assessment sequence and outputs the risk prediction value for each prediction time within the corresponding prediction time window.

[0096] In this embodiment, the generation of the risk classification and early warning information specifically includes:

[0097] The risk prediction value is identified. When the risk prediction value is in the form of a level, each level is converted into a corresponding numerical measure according to the preset risk classification rules. When the risk prediction value is in the form of a probability, the numerical value is used as the basic measure, and the basic measure is subjected to a unified scale conversion process to form a risk measure value.

[0098] The risk measurement value is compared with the risk classification threshold in the preset risk classification rules to obtain the warning level. The comparison process is to determine the risk measurement value and the risk classification threshold from low to high interval. When the risk measurement value is within the interval of the risk classification threshold, the classification label corresponding to the interval is used as the warning level.

[0099] By combining the warning level, time marker, construction area marker, and risk assessment results, a risk classification warning information is generated.

[0100] Example 1:

[0101] This embodiment uses a super high-rise commercial complex project in a coastal city as an application scenario. The project has 52 floors above ground and 3 floors underground, and the construction phase is in a high-risk period where the main structure and electromechanical installation are intersecting. The site is characterized by high-frequency operation of tower cranes and construction hoists, dense rebar and formwork operations, many high-altitude edge operations, mixed traffic of vehicles and personnel around the deep foundation pit, and fatigue of personnel due to night-time rush work. In the past, the project safety management team mainly relied on safety officer inspections, video spot checks, and weekly risk list scoring to assess risks. However, when there are dense overlapping operations and rapidly changing risks, there are often situations where "it was safe yesterday, but suddenly something goes wrong today." In particular, the identification of hidden risks formed by the superposition of "personnel violations + equipment malfunctions + environmental disturbances" is not timely, resulting in many false alarms and missed alarms, short warning windows, and delayed response. This is a typical manifestation of the two minor technical problems that this invention aims to solve: first, insufficient risk characteristic representation in complex dynamic scenarios; and second, delayed assessment and lack of trend prediction.

[0102] In this scenario, the invention is deployed on the project's existing digital safety platform. The construction site is equipped with 48 high-definition cameras covering the tower crane's slewing radius, edge protection zones, material storage areas, and unloading platforms. A total of 126 displacement, tilt, stress, and support axial force sensors are deployed in the deep foundation pit and high-formwork areas. Key equipment such as tower cranes and elevators upload load, speed, start / stop status, and alarm signals via an industrial gateway. The personnel management system provides the work location and duration of work teams through location tags and attendance records. The construction progress and BIM system provide the daily process stages, key component status, and hazardous location locations. The system uses a 5-minute time window, aligning, cleaning, denoising, and filling in missing data from all sources within that 5-minute window, and then reconstructing it into a spatiotemporal sample. Visual features, numerical features, behavioral features, and process and component status features are extracted from each spatiotemporal sample and then aligned and fused within a unified feature space to form a fused feature vector.

[0103] In terms of model application, the system was first trained offline using historical data from the past four months since the project commenced. This historical data contained approximately 180,000 spatiotemporal samples, with risk labels derived from past hazard records and event logs, categorized into four risk levels: low, medium, high, and critical. During training, the improved ASTGCN model constructed a dynamic graph structure with construction areas as nodes. The connections between nodes were jointly determined by spatial adjacency, process dependence, and interaction strength, and dynamically updated according to changes in fusion features within the time window. The adaptive spatial attention module introduced risk propagation direction constraints on top of the dynamic graph, allowing risks to only move from "strong interaction / upstream process / adjacent hazardous area" to "affected area." An evidence-based attention calibration mechanism was used to filter out node pairs with higher spatial correlation strength, thus more accurately learning the risk transmission patterns between areas. Multi-scale temporal attention and temporal convolution modules simultaneously captured minute-level sudden violations and hour-level cumulative risks. After training, the model performed real-time inference on-site, outputting the risk output value for the current time window every 5 minutes, and binding it with time markers and construction area markers to form a risk assessment result. The continuous risk assessment results are then combined into a risk assessment sequence and input into the risk trend prediction network to predict risk changes within the next 30 minutes, and generate graded early warning information to be pushed to safety officers and team leaders.

[0104] To verify the beneficial effects, this embodiment underwent a 6-week comparative trial run on the same project. The comparison object was the project's original method of "rule list + manual scoring + video spot checks". Evaluation indicators included the accuracy of high-risk event identification, false alarm rate, warning lead time, and safety management response effectiveness. During the trial run, a total of 214 medium-to-high-risk events were recorded on-site after verification, including 36 high-risk events. The original method mainly focused on safety officer inspections and post-event reviews for identifying high-risk events, with an average warning lead time of only about 7 minutes, and significant false alarms for superimposed and coupled risks. The method of this invention significantly improves the accuracy of high-risk event identification and early warning capability without increasing the frequency of manual inspections, especially for the identification of combined risks of "personnel behavior + equipment operating conditions + environmental trends". The trial run results show that the average warning lead time for high-risk events of this invention is increased to more than 23 minutes, allowing more time for on-site handling; at the same time, the false alarm rate is reduced, and safety officers report "fewer but more accurate warnings", reducing fatigue response. More importantly, slow-changing risks such as the gradual accumulation of foundation pit displacement and the increase in nighttime fatigue violation rate can be continuously given "high-risk upward trend" warnings before reaching the critical point by the trend prediction network. This allows the project management to adjust the support and night shift rotation strategies in advance, and no major hidden dangers requiring work stoppage and rectification occurred within 6 weeks.

[0105] Table 1 Comparison of Risk Assessment and Early Warning Effects During Operation Phase

[0106] index Existing methods Method of the present invention Increase Accuracy of identifying medium- and high-risk events 81.3% 92.5% +11.2% High-risk event identification accuracy (36 cases) 77.8%（28 / 36） 94.4%（34 / 36） +16.6% Missed rate of coupled high-risk events (18 events) 33.3%（6 / 18） 5.6%（1 / 18） −27.7% Average false alarm rate (based on daily warnings) 14.6% 7.9% −6.7% Average warning lead time 7.2 minutes 23.6 minutes +16.4 minutes Accuracy of identifying violations at heights and near edges 79.1% 93.0% +13.9% Accuracy of identifying intrusions into restricted areas during hoisting operations 83.5% 95.2% +11.7% Accuracy of identifying risks associated with equipment malfunctions 74.6% 90.8% +16.2% Accuracy of early warning for foundation pit displacement trends (12 times) 58.3%（7 / 12） 91.7%（11 / 12） +33.4% Average response time of safety officers 19.4 minutes 11.1 minutes -8.3 minutes Number of major hidden dangers requiring work stoppage and rectification 2 0 Significant decline

[0107] Based on the identification results in Table 1, the method of this invention achieves an overall identification accuracy of 92.5% for medium- and high-risk events, an improvement of 11.2% compared to the original method. This demonstrates that multi-source feature fusion and improved spatiotemporal modeling of ASTGCN effectively enhance the sufficiency of risk representation. Particularly at the high-risk event level, this invention correctly identified 34 out of 36 high-risk events, achieving an accuracy of 94.4%, while the original method only identified 28, reflecting the model's higher sensitivity to highly dangerous situations.

[0108] The data that best reflects the "solution to technical problems" is the underreporting rate of coupled high-risk events. In this embodiment, there were 18 coupled high-risk events triggered by the superposition of "personnel violations + equipment malfunctions / environmental trends". The original method missed 6 of these events, with an underreporting rate as high as 33.3%; the present invention missed only 1 event, reducing the underreporting rate to 5.6%. This directly demonstrates that the present invention can automatically extract and fuse implicit coupling features in complex dynamic scenarios, overcoming the shortcomings of traditional rules / manual indicators in capturing multi-factor linkage risks.

[0109] In terms of early warning capabilities, the average early warning lead time of this invention is 23.6 minutes, more than 16 minutes longer than the original method. Furthermore, the accuracy rate of early warning for foundation pit displacement trends has increased from 58.3% to 91.7%, demonstrating that the trend prediction network expands risk assessment from "current judgment" to "future window prediction," effectively solving the problems of assessment lag and short early warning windows. The increased lead time also leads to improved management effectiveness; the average response time for safety officers has decreased from 19.4 minutes to 11.1 minutes, and on-site feedback and early warnings are more concentrated and actionable, reducing ineffective emergency responses.

[0110] Finally, regarding the results of major hidden dangers, the original method resulted in two work stoppages for rectification during the trial operation, while the method of this invention resulted in zero. Although the number of hidden dangers is affected by multiple factors such as operating conditions and management, the aforementioned data on underreporting and advance warning reasonably demonstrate that this invention, through more accurate and forward-looking risk assessment and graded early warning, improves the on-site proactive prevention and control capabilities, achieving the technical improvements and beneficial effects required in the background technology.

Claims

1. A construction risk assessment method based on deep learning, characterized in that, Includes the following steps: Collect and preprocess multi-source data from construction sites; The preprocessed multi-source data is segmented according to the time window and recombined into spatiotemporal samples by combining the construction area identifiers; Feature extraction and fusion of spatiotemporal samples are performed to form a fused feature vector; Training and validation datasets were constructed based on fused feature vectors and corresponding risk annotations. The improved ASTGCN model is trained by inputting the training dataset and then validated using the validation dataset to obtain the trained improved ASTGCN model. Real-time collection of multi-source data from the construction site and input into the trained improved ASTGCN model; output risk assessment results. The risk assessment results are combined in chronological order to form a risk assessment sequence, which is then input into a risk trend prediction network to obtain risk prediction values. Based on the predicted risk values, risk classification early warning information is generated through preset risk classification rules.

2. The construction risk assessment method based on deep learning according to claim 1, characterized in that, The multi-source data includes construction site video image data, environmental and structural monitoring sensor data, construction equipment operating condition data, personnel operation behavior data, construction progress data, and BIM model data. The preprocessing includes data denoising, missing data completion, and abnormal data removal.

3. The construction risk assessment method based on deep learning according to claim 1, characterized in that, The formation of the spatiotemporal sample specifically includes: The preprocessed multi-source data are aligned according to a unified time base to form a continuous multi-source data sequence; Set the time window length and the time window sliding step size, extract continuous multi-source data sequences according to the time window length to obtain the first time window, and successively move the corresponding continuous multi-source sequences according to the time window sliding step size to form time window multi-source data; Window samples are formed by recombining and aggregating multi-source data within a time window with construction area identifiers. The window samples are associated with the corresponding spatial location information to form spatiotemporal samples that include both temporal continuity and spatial distribution.

4. The construction risk assessment method based on deep learning according to claim 1, characterized in that, The formation of the fused feature vector specifically includes: The spatiotemporal samples are decomposed according to the data source to obtain subsets of video images, sensor and equipment operating conditions, personnel work behavior, construction progress, and BIM model. Visual information is extracted from a subset of video images to generate a visual feature set. The extraction process involves detecting and locating targets such as personnel, equipment, components, and hazards in the subset of video images within each time window, obtaining target categories and spatial locations, continuously tracking detected targets to obtain target motion trajectories, performing time-series identification of personnel and equipment work behaviors, obtaining hazard behavior labels, and forming a visual feature set. Numerical information is extracted from a subset of sensor and equipment operating conditions to generate a numerical feature set. The extraction process involves performing unified dimension and normalization processing on the sampling sequences of sensors and equipment within each time window to obtain the statistical characteristics and rate of change characteristics of each sampling sequence, and generating threshold exceedance markers and equipment anomaly markers to form a numerical feature set. Behavioral information is extracted from a subset of personnel work behaviors to generate a set of process and component status features. The extraction process involves processing the process and component status information of a subset of construction progress and a subset of BIM models. Within each time window, the personnel work location sequence, action sequence, work duration, and violation work markers are obtained to form a set of behavioral features. The component type, spatial relationship, component status, and location of dangerous parts corresponding to the process are extracted from the BIM model to form a set of process and component status features. The visual feature set, numerical feature set, behavioral feature set, and process and component status feature set are unified in dimension and aligned in feature, and then combined according to the preset fusion rules to obtain the fused feature vector.

5. The construction risk assessment method based on deep learning according to claim 1, characterized in that, The construction of the training and validation datasets specifically includes: The fusion feature vectors are numbered one by one according to the samples, and the corresponding time identifier and construction area identifier are recorded for each fusion feature vector to obtain the fusion feature sample set. The risk labeling records corresponding to the fusion feature sample set are organized. Based on the time identifier and construction area identifier of each sample in the fusion feature sample set, the risk level label is obtained from the risk labeling records. The risk level labels are then summarized by sample number to form a risk label set. The risk labeling set is divided into a training dataset and a validation dataset according to a preset division ratio.

6. The construction risk assessment method based on deep learning according to claim 1, characterized in that, The improved ASTGCN model obtained after training specifically includes: The training dataset is input into the improved ASTGCN model for training. The improved ASTGCN model includes a dynamic graph construction module, an adaptive spatial attention module, and a multi-scale temporal attention and temporal convolution module. The construction dynamic graph construction module refers to establishing and updating an initial graph structure based on spatial adjacency, process dependence, and interaction strength, using construction area identifiers as nodes, to obtain a dynamic graph structure. The adaptive spatial attention module refers to introducing an evidence-based attention calibration mechanism under the dynamic graph structure, calculating the spatial association strength between nodes, and obtaining spatially enhanced node features. The multi-scale temporal attention and temporal convolution module refers to performing multi-temporal-scale attention modeling on the spatially enhanced node features and obtaining a spatiotemporal joint representation of the nodes through temporal convolution. In the construction dynamic graph construction module, the construction area identifiers in the training dataset are used as graph nodes, and the spatial adjacency, process dependency, and interaction intensity relationships between construction areas within the time window are used as node connection relationships to form an initial graph structure. During the training iteration process, the node connection relationships are updated according to the fusion feature vector of the current time window to obtain the construction dynamic graph structure. In the adaptive spatial attention module, risk propagation direction constraint relationships between nodes are constructed based on the construction dynamic map structure and the fused feature vectors within the corresponding time window. The construction process involves obtaining adjacent nodes based on the construction dynamic map structure and combining the spatial positional relationship, process sequence relationship, and interaction intensity change trend in the fused feature vectors to form risk propagation direction constraint relationships between nodes. By introducing an evidential attention calibration mechanism, the spatial association strength between nodes is calculated, and spatially enhanced node features are generated. The calculation process involves using the evidential attention calibration mechanism to calculate the feature correlation score of adjacent node pairs that satisfy the risk propagation direction constraints, thereby obtaining the spatial coupling degree of the node pairs and forming the spatial association strength. In the multi-scale temporal attention and temporal convolution module, the spatially enhanced node features are arranged into a temporal feature sequence according to the order of the time window. The temporal attention weights are calculated on the temporal feature sequence according to a preset time scale to obtain a multi-scale temporal attention representation. The multi-scale temporal attention representation is then subjected to temporal convolution to obtain a spatiotemporal joint representation of the node. Nonlinear feature compression is performed on the spatiotemporal joint representation of nodes to generate risk output values ​​corresponding to each training sample. The sample error is calculated based on the deviation between the risk level label of the training sample and the risk output value. The average value of all sample errors is used as the training loss. The improved ASTGCN model is iteratively trained based on the training loss. The validation error is calculated using the validation dataset. The iterative training stops when the validation error meets the preset convergence condition, resulting in the trained improved ASTGCN model.

7. The construction risk assessment method based on deep learning according to claim 1, characterized in that, The output of the risk assessment results specifically includes: Real-time acquisition of multi-source data from the construction site and feature extraction and fusion are performed to obtain a real-time fused feature vector corresponding to the current time window; The real-time fused feature vector is input into the trained improved ASTGCN model. The construction dynamic map structure is updated by the construction dynamic map construction module, the spatially enhanced node features are extracted by the adaptive spatial attention module, and the spatiotemporal joint representation of the nodes is obtained by the multi-scale temporal attention and temporal convolution module, generating real-time risk output values. The real-time risk output value is associated with the corresponding time marker and construction area marker to form the risk assessment result.

8. The construction risk assessment method based on deep learning according to claim 1, characterized in that, The risk prediction value is obtained specifically through: The risk assessment results are cached and arranged in chronological order to form a risk assessment sequence; The risk assessment sequence is input into the risk trend prediction network, which performs time-series inference based on the risk assessment sequence and outputs the risk prediction value for each prediction time within the corresponding prediction time window.

9. The construction risk assessment method based on deep learning according to claim 1, characterized in that, The generation of the risk classification and early warning information specifically includes: The risk prediction value is identified. When the risk prediction value is in the form of a level, each level is converted into a corresponding numerical measure according to the preset risk classification rules. When the risk prediction value is in the form of a probability, the numerical value is used as the basic measure, and the basic measure is subjected to a unified scale conversion process to form a risk measure value. The risk measurement value is compared with the risk classification threshold in the preset risk classification rules to obtain the warning level. The comparison process is to determine the risk measurement value and the risk classification threshold from low to high interval. When the risk measurement value is within the interval of the risk classification threshold, the classification label corresponding to the interval is used as the warning level. By combining the warning level, time marker, construction area marker, and risk assessment results, a risk classification warning information is generated.

Images