A construction safety risk real-time early warning method and system

Abstract

The application discloses a kind of building construction safety risk real-time early warning method and system, wherein method includes the video monitoring data and construction log data of target building construction obtained respectively are analyzed and handled, obtain standardized video data and standard log data;Based on double-channel attention mechanism, the feature fusion of standardized video data and standard log data is carried out, and fusion data is obtained;Risk feature identification is carried out by inputting fusion data into pre-trained risk identification model, and risk feature vector is obtained;Real-time process data and risk feature vector obtained are associated and analyzed, and associated analysis result is obtained;Risk grade determination is carried out based on associated analysis result, and real-time early warning of target building construction is carried out based on risk grade determination result.The method of the application significantly improves the accuracy of building construction safety warning.

Description

Technical Field

[0001] This invention relates to the field of construction management technology, and in particular to a method and system for real-time early warning of safety risks in building construction. Background Technology

[0002] The construction industry is characterized by numerous high-altitude operations, a high density of specialized equipment, and complex overlapping processes, making safety risk management extremely difficult. The complexity and high risk of its working environment remain the core challenges for safety management.

[0003] Currently, construction sites rely solely on video data for early warning during construction. However, the video data collected by cameras at the construction site cannot fully cover the construction area and is easily affected by environmental factors such as lighting and obstructions, resulting in poor early warning effectiveness and failing to guarantee the safety of life and property during construction. Summary of the Invention

[0004] This invention provides a real-time early warning method and system for construction safety risks, in order to solve the technical problem of how to predict construction safety risks and improve the accuracy of construction safety early warning.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a method for real-time early warning of construction safety risks, comprising:

[0006] In response to risk video recognition signals, standardized video data of the target building's construction process is acquired;

[0007] In response to the risk log identification signal, the content of the construction log data corresponding to the construction process of the target building is identified to obtain standard log data;

[0008] The standardized video data is processed using spatiotemporal enhancement technology to obtain spatiotemporal enhanced video features;

[0009] The standard log data is subjected to time-series feature extraction to obtain time-series enhanced process features, and process relationship modeling is performed on the time-series enhanced process features to obtain structured process features;

[0010] The spatiotemporal enhanced video features and the structured process features are fused to obtain fused data;

[0011] The fused data is input into a pre-trained risk identification model to identify risk features and obtain a risk feature vector.

[0012] In response to the risk warning signal, the system performs correlation analysis on the acquired real-time process data and the risk feature vector to obtain the correlation analysis results. Based on the correlation analysis results, the system determines the risk level and provides real-time warnings for the construction of the target building according to the risk level determination results.

[0013] As one preferred embodiment, acquiring standardized video data of the target building construction process includes:

[0014] The acquired video monitoring data is dynamically denoised using an adaptive Gaussian filtering algorithm to obtain denoised video frames.

[0015] The denoised video frames are labeled with target bounding boxes to obtain target video segments, and the target video segments are standardized to obtain standardized video data; wherein the target bounding boxes include target bounding boxes for construction personnel and target bounding boxes for special equipment.

[0016] As one preferred embodiment, the step of performing content recognition on the construction log data corresponding to the construction process of the target building to obtain standard log data includes:

[0017] Based on the BERT model, target information is extracted from construction log data to obtain the core log dataset.

[0018] Entity relationship mapping is performed on the core log dataset based on the pre-built construction database to obtain structured log entries;

[0019] The structured log entries are then time-aligned to obtain standard log data.

[0020] As one preferred embodiment, the process of processing the standardized video data based on spatiotemporal enhancement technology to obtain spatiotemporally enhanced video features includes:

[0021] The standardized video data is subjected to global information aggregation processing to generate channel descriptor vectors, and the channel descriptor vectors are subjected to nonlinear transformation processing to generate channel attention weight matrices.

[0022] The channel attention weight matrix is ​​multiplied element-wise with the original feature map of the standardized video data to obtain a spatial feature enhancement map;

[0023] Spatiotemporal joint feature extraction is performed on the spatial feature enhancement map to generate a spatiotemporal feature cube, and frame-level importance weighting is performed on the spatiotemporal feature cube to generate temporal attention weights;

[0024] The spatiotemporal feature cube is fused with the temporal attention weight to obtain a spatiotemporally enhanced feature video.

[0025] As one preferred embodiment, the step of extracting time-series features from the standard log data to obtain time-series enhanced process features, and then modeling process relationships based on these time-series enhanced process features to obtain structured process features, including:

[0026] The standard log data is semantically vectorized to obtain process feature vectors, and the process feature vectors are then modeled with temporal dependencies to capture the relationships between processes before and after execution and generate hidden state sequences.

[0027] The hidden state sequence is subjected to key process identification processing to obtain time-enhanced process features, and the time-enhanced process features are mapped to graph node features to construct an initial process relationship graph.

[0028] The initial process relationship graph is calculated based on the graph attention mechanism, and the node weights are updated according to the dependency strength between processes to obtain the weighted process relationship graph.

[0029] The process relationship graph is subjected to feature aggregation processing to obtain structured process features, wherein the feature aggregation processing is designed to integrate the association information of neighboring nodes into the features of each node in the process relationship graph.

[0030] As one preferred embodiment, the step of performing feature fusion processing on the spatiotemporal enhanced video features and the structured process features to obtain fused data includes:

[0031] The spatiotemporal enhanced video features and the structured process features are subjected to dimensional unification processing to generate feature pairs of the same dimension;

[0032] The same-dimensional feature pairs are subjected to correlation strength quantification based on mutual information method to obtain cross-modal correlation matrix, and the same-dimensional feature pairs are interactively processed with the cross-modal correlation matrix to obtain cross-dimensional interactive features;

[0033] Based on predefined risk labels, the correlation between each cross-dimensional interaction feature and construction risk is calculated, and a dynamic weight vector corresponding one-to-one with the dimensions of the cross-dimensional interaction feature is generated.

[0034] The cross-dimensional interactive features are multiplied element-wise with the dynamic weight vector to obtain adaptive weighted features.

[0035] The adaptive weighted features are subjected to nonlinear transformation to obtain deep fusion features, and the deep fusion features are then subjected to residual connection processing with the spatiotemporal enhanced video features and the structured process features to obtain enhanced fusion features.

[0036] The enhanced fusion features are normalized to obtain fused data.

[0037] As one preferred embodiment, the step of inputting the fused data into the pre-trained risk identification model for risk feature identification to obtain a risk feature vector includes:

[0038] To construct a training dataset, historical fusion data and corresponding risk event records in the construction field are obtained.

[0039] An initial risk identification model based on the TimeSformer model is constructed, and the initial risk identification model is transferred to the training dataset to obtain a trained risk identification model.

[0040] The fused data is segmented according to time series and input into the risk identification model to output multi-scale risk features.

[0041] The multi-scale risk features are sequentially processed by feature concatenation and feature dimensionality reduction to obtain the risk feature vector.

[0042] As one preferred embodiment, the correlation analysis processing of the acquired real-time process data and the risk feature vector to obtain the correlation analysis results includes:

[0043] Entity extraction is performed on real-time process data to construct a process feature entity set;

[0044] The correlation strength between the process feature entity set and the risk feature vector is calculated based on the cosine similarity algorithm to obtain the correlation matrix between the process and the risk.

[0045] Cluster analysis is performed on the real-time process data based on the correlation matrix to obtain high-correlation process groups and low-correlation process groups;

[0046] Based on the historical risk occurrence probability distribution, the threshold interval of the first risk feature vector corresponding to each highly correlated process group is calculated to obtain the dynamic risk threshold for process matching.

[0047] The deviation between the second risk feature vector and the feature entity set corresponding to each low-association process group is fitted and calculated using the least squares method to obtain the feature correction coefficient used to correct the risk feature weight.

[0048] The dynamic risk threshold and the feature correction coefficient are associated and stored to generate an association analysis result containing the process ID, dynamic risk threshold, and feature correction coefficient.

[0049] As one preferred embodiment, the step of determining the risk level based on the correlation analysis results and providing real-time early warning for the construction of the target building based on the risk level determination results includes:

[0050] Extract the dynamic risk threshold and feature correction coefficient matching the current process from the correlation analysis results, and perform a weighted operation on the risk feature vector and feature correction coefficient to obtain the corrected risk value;

[0051] The corrected risk value is compared with the dynamic risk threshold to obtain the risk level;

[0052] Real-time early warning of target building construction based on risk level.

[0053] Another embodiment of the present invention provides a real-time early warning system for construction safety risks, comprising:

[0054] The first acquisition module is used to acquire standardized video data of the target building construction process in response to the risk video recognition signal;

[0055] The second acquisition module is used to respond to the risk log identification signal, perform content identification on the construction log data corresponding to the construction process of the target building, and obtain standard log data.

[0056] The first extraction module is used to process the standardized video data based on spatiotemporal enhancement technology to obtain spatiotemporal enhanced video features;

[0057] The second extraction module is used to extract time-series features from the standard log data to obtain time-series enhanced process features, and to model the process relationships of the time-series enhanced process features to obtain structured process features.

[0058] The fusion module is used to perform feature fusion processing on the spatiotemporal enhanced video features and the structured process features to obtain fused data;

[0059] The identification module is used to input the fused data into a pre-trained risk identification model to identify risk features and obtain a risk feature vector.

[0060] The early warning module is used to respond to risk warning signals by performing correlation analysis on the acquired real-time process data and the risk feature vector to obtain correlation analysis results, and to determine the risk level based on the correlation analysis results, and to provide real-time early warning for the construction of the target building based on the risk level determination results.

[0061] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0062] 1) This invention employs a dual-channel attention mechanism to fuse video monitoring data and construction log data at the feature level, which not only retains key visual features (such as personnel violations) but also integrates process logic features (such as high-altitude operation specifications), thus solving the one-sidedness problem of traditional single-source data early warning.

[0063] 2) This invention uses correlation analysis to analyze real-time process and risk characteristics, customizes dynamic risk thresholds for different construction processes (such as the threshold for not wearing a safety belt in high-altitude operations is much lower than that for ground operations), and introduces feature correction coefficients to adjust risk weights, so that the early warning standard can adapt to changes in the construction scenario and significantly reduce the false alarm rate.

[0064] 3) This invention captures the spatiotemporal dynamic features of video based on the TimeSformer model (such as personnel not wearing safety equipment for 3 consecutive minutes), and combines it with the temporal information of construction logs (such as the order of process execution) to achieve accurate identification of complex risk scenarios, such as multiple risk superposition scenarios of "violation of regulations + equipment abnormality + high-risk process". Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating a method for real-time early warning of construction safety risks in one embodiment of the present invention.

[0066] Figure 2 This is a schematic diagram of the high-altitude hoisting process in one embodiment of the present invention;

[0067] Figure 3 This is a flowchart of a method for real-time early warning of construction safety risks in one embodiment of the present invention;

[0068] Figure 4 This is a schematic diagram of a real-time early warning system for construction safety risks in one embodiment of the present invention;

[0069] Figure label:

[0070] Among them, 11, first acquisition module; 12, second acquisition module; 13, first extraction module; 14, second extraction module; 15, fusion module; 16, identification module; 17, early warning module. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0072] In the description of this invention, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0073] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0074] One embodiment of the present invention provides a method for real-time early warning of safety risks in building construction. For details, please refer to [link / reference]. Figure 1 , Figure 1 The diagram shown is a flowchart illustrating a real-time early warning method for construction safety risks according to one embodiment of the present invention, which includes steps S1-S7:

[0075] S1: In response to risk video recognition signals, acquire standardized video data of the target building construction process;

[0076] S2: In response to the risk log identification signal, perform content identification on the construction log data corresponding to the construction process of the target building to obtain standard log data;

[0077] S3: Process the standardized video data based on spatiotemporal enhancement technology to obtain spatiotemporal enhanced video features;

[0078] S4: Extract time-series features from standard log data to obtain time-series enhanced process features, and model the process relationships of the time-series enhanced process features to obtain structured process features;

[0079] S5: Perform feature fusion processing on the spatiotemporal enhanced video features and structured process features to obtain fused data;

[0080] S6: Input the fused data into the pre-trained risk identification model to identify risk features and obtain risk feature vectors;

[0081] S7: In response to the risk warning signal, perform correlation analysis on the acquired real-time process data and the risk feature vector to obtain the correlation analysis result, and determine the risk level based on the correlation analysis result, and issue a real-time warning for the construction of the target building based on the risk level determination result.

[0082] Currently, construction sites rely solely on video data for early warning during construction. However, the video data collected by on-site cameras cannot fully cover the construction area and is easily affected by environmental factors such as lighting and obstructions, resulting in poor early warning effectiveness and failing to guarantee the safety of life and property during construction. More importantly, safety risks in construction are often not isolated but closely related to the rationality and standardization of construction procedures. For example, the early or late execution of a certain procedure, or the omission of connections between procedures, can trigger a chain reaction of safety hazards.

[0083] Work log data serves as a timeline recorder of the construction process, clearly presenting the planned arrangements, actual progress, personnel allocation, and key milestones for each procedure. This information constructs the framework of the construction process, providing a benchmark for understanding the rationality of on-site actions. For example, the detailed record in the log of the sequence of procedures—"foundation treatment → rebar tying → formwork installation"—can help determine whether on-site operations conform to standard procedures and avoid safety risks caused by reversing the order of procedures.

[0084] However, while work log data can supplement video data in terms of information gaps regarding process logic and standardization, it also has its own shortcomings: First, logs are manually recorded by construction personnel, which may lead to information omissions, recording delays, or data distortion due to human negligence (such as concealing potential hazards to avoid responsibility); Second, log data is mostly structured / semi-structured data in text or tabular form, and its timeliness depends on the efficiency of manual input, making it difficult to reflect the dynamics on site in real time; Third, single log data lacks an intuitive connection with the actual on-site operation scenario, and it is difficult to accurately locate the specific links and manifestations of risks through textual descriptions alone, resulting in a high degree of subjectivity in risk assessment.

[0085] Since video data and log data are complementary in terms of "visual intuitiveness and logical structure" in representing risk information, and each has its own shortcomings that are difficult to overcome alone, this embodiment of the invention deeply integrates the two types of data. This allows for the capture of real-time visual features of on-site operations using video data, and the mining of structured logic of process flows using log data. This enables comprehensive and high-precision identification and early warning of construction safety risks, thus overcoming the limitations of a single data source in risk management.

[0086] Upon receiving a risk identification signal, the computer system will initiate the parsing and processing of video monitoring data and construction log data of the target building construction. The purpose is to transform the raw and messy data into standardized, orderly, and easily analyzed and utilized standardized video data and standard log data, laying a solid foundation for the subsequent accurate identification of construction safety risks.

[0087] Preferably, in one embodiment of the present invention, acquiring standardized video data of the target building construction process includes:

[0088] The video monitoring data is dynamically denoised based on the adaptive Gaussian filtering algorithm to obtain denoised video frames.

[0089] The denoised video frames are labeled with target bounding boxes to obtain target video segments, and the target video segments are then standardized to obtain standardized video data; the target bounding boxes include target bounding boxes for construction personnel and target bounding boxes for special equipment.

[0090] Adaptive Gaussian filtering is an image processing technique that uses a Gaussian function to smooth images, reduce noise and details, and highlight key features. Unlike traditional Gaussian filtering, it dynamically adjusts filter parameters (such as standard deviation) based on local image characteristics. If the variance in a local area is small, it indicates that the area is relatively homogeneous, and a larger variance can be used. The values ​​should be smoothed; if the variance is large, it indicates that the region contains edge information and should be reduced. The value is used to avoid edge blurring, thus preserving edge information while smoothing out noise.

[0091] In this embodiment, a series of detailed and critical parsing processes are performed on the video monitoring data and construction log data obtained from the construction of the target building, in order to obtain standardized video data and standard log data.

[0092] Specifically, in terms of video monitoring data processing, the first step is to use an adaptive Gaussian filtering algorithm to dynamically denoise the video monitoring data. Construction site environments are complex, and videos are easily affected by various factors, such as electromagnetic interference from construction equipment and variable lighting conditions. The adaptive Gaussian filtering algorithm can dynamically adjust the filtering parameters based on the local features of the video image.

[0093] Taking the tower crane operation area at the construction site as an example, the image in this area changes frequently and has a large local variance. At this time, the algorithm will automatically reduce the standard deviation of the filter, so as to remove noise while preserving the key details of the tower crane operation to the greatest extent, such as the rotation trajectory of the tower crane boom and the position change of the hook. In the relatively stable material stacking area, the local variance is small, so the algorithm increases the standard deviation to make the video image in this area smoother, effectively remove noise, and finally output clear noise-reduced video frames.

[0094] After obtaining the denoised video frames, target bounding boxes are then annotated. Considering the unique characteristics of construction scenarios, target bounding boxes are primarily divided into those for construction workers and those for special equipment. When annotating the construction worker bounding boxes, established annotation standards are strictly followed to ensure the boxes accurately fit the outlines of the workers. Even if some limbs are obscured, reasonable inferences are made based on human anatomy to ensure the bounding boxes cover key areas such as the head and limb joints. This allows for accurate identification of worker actions, such as whether safety helmets are worn correctly or whether unauthorized climbing is occurring. For special equipment bounding boxes, taking tower cranes as an example, key components such as the tower body, boom, counterweight boom, and hook are fully annotated, precisely defining their position and extent within the video frame to obtain target video clips containing crucial target information.

[0095] Furthermore, the target video segments undergo standardization processing, unifying the resolution, frame rate, and other aspects of videos captured by different cameras. For example, videos with varying resolutions are uniformly adjusted to 1920×1080, and videos with different frame rates are standardized to 25 frames per second, ensuring a consistent video data format. This facilitates efficient and unified analysis and processing by subsequent computer systems and related algorithms, thereby obtaining standardized video data.

[0096] Preferably, in one embodiment of the present invention, content recognition is performed on the construction log data corresponding to the construction process of the target building to obtain standard log data, including:

[0097] Based on the BERT model, target information is extracted from construction log data to obtain the core log dataset.

[0098] Entity relationship mapping is performed on the core log dataset based on the pre-built construction database to obtain structured log entries;

[0099] The structured log entries are time-series aligned to obtain standard log data.

[0100] Among them, the BERT model is a bidirectional language representation model based on the Transformer architecture. It learns rich contextual representations by pre-training on massive amounts of text data, and then fine-tunes them for downstream tasks to achieve high performance. Its core advantage lies in its bidirectionality, which allows it to learn contextual information from both the left and right sides of the text simultaneously, thus better understanding the semantics of each word in a sentence; and it adopts a pre-training and fine-tuning mode, which allows for rapid fine-tuning on various downstream tasks through pre-training tasks.

[0101] In terms of construction log data processing, this embodiment preferably uses the BERT model to extract target information from the construction log data. Of course, other large language models can also be selected for processing; no specific limitation is made in this embodiment. Construction logs are typically recorded in natural language and contain complex information, covering various aspects such as construction time, location, procedures, personnel arrangements, and equipment usage. The BERT model, with its powerful bidirectional encoder structure in the field of natural language processing, can deeply understand the semantics of log text. Specifically, the BERT model can extract construction information from the construction log data, including but not limited to process-related information, personnel information, equipment and material information, safety and quality records, and information on anomalies and emergencies.

[0102] For example, for a log entry like "On July 10, 2022, at the construction site of Building 3, Engineer Wang led 5 workers to carry out the rebar binding operation for the first-floor beams and slabs, using 2 rebar bending machines," the BERT model can accurately extract key information such as the construction time "July 10, 2022," the construction location "construction site of Building 3," the construction procedure "rebar binding operation for the first-floor beams and slabs," the person in charge "Engineer Wang," the number of participants "5," and the equipment used "2 rebar bending machines," forming the core log dataset.

[0103] Furthermore, entity relationship mapping is performed on the core log dataset based on a pre-built construction database. The construction database encompasses rich entity information and their interrelationships within the construction field, such as construction personnel, equipment, materials, and procedures, as well as the usage relationships between procedures and equipment, and the consumption relationships between materials and procedures. Key information extracted from the logs is matched and associated with entities in the database. For example, the "rebar tying operation" procedure is linked to the corresponding procedure entity in the database, and its usage relationship with the "rebar bending machine" equipment entity is clarified. This transforms loose log information into structured log entries, such as "Rebar tying operation - Use - Rebar bending machine (equipment)" and "Rebar tying operation - Performers - Engineer Wang and 5 workers."

[0104] Structured log entries are aligned according to their time sequence. Construction processes follow strict sequential order and logical temporal relationships. By analyzing the timestamp information in the log entries and the logical connections between processes, the structured log entries are sorted and integrated. For example, foundation excavation must precede foundation pouring. When processing log data, based on these logical relationships and the order of recording, relevant log entries are arranged in the correct sequence to ensure the continuity and accuracy of construction log data in the temporal dimension. This results in standardized log data that accurately reflects the actual progress and timeline of the construction process, providing a reliable basis for subsequent comprehensive and accurate safety risk analysis combined with video data.

[0105] Furthermore, in one embodiment of the present invention, feature fusion is performed on standardized video data and standard log data based on a dual-channel attention mechanism to obtain fused data; wherein the dual-channel attention mechanism is designed to assign differentiated weights to visual key features and structured key features.

[0106] This embodiment uses a dual-channel attention mechanism to perform feature-level fusion of standardized video data (visual information) and standard log data (structured process information). The purpose is to enable the computer system to be like a human safety officer, able to understand both the on-site footage and the construction process, and to give higher attention to high-risk associated features (such as "violations + high-risk processes"), ultimately outputting fused data that comprehensively reflects the essence of construction risks.

[0107] The dual-channel attention mechanism includes a spatial attention module based on the SENet network and a temporal attention module based on the LSTM network. Specifically, in this embodiment, the dual-channel attention mechanism includes two parallel channels, each adapted to the feature extraction needs of different types of data. The spatial attention channel focuses on the "visual key features" of standardized video data, such as the illegal actions of construction workers and the dangerous state of special equipment, highlighting risk-related areas through spatial dimension weight allocation. The temporal attention channel focuses on the "structured key features" of standard log data, such as process type and operation parameters, highlighting risk-related process nodes through temporal dimension weight allocation.

[0108] Preferably, in one embodiment of the present invention, the standardized video data is processed based on spatiotemporal enhancement technology to obtain spatiotemporally enhanced video features, including:

[0109] The standardized video data is subjected to global information aggregation processing to generate channel descriptor vectors, and the channel descriptor vectors are subjected to nonlinear transformation processing to generate channel attention weight matrices.

[0110] The spatial feature enhancement map is obtained by multiplying the channel attention weight matrix element-wise with the original feature map of the standardized video data.

[0111] Spatiotemporal joint feature extraction is performed on the spatial feature enhancement map to generate a spatiotemporal feature cube, and frame-level importance weighting is applied to the spatiotemporal feature cube to generate temporal attention weights;

[0112] By fusing the spatiotemporal feature cube with the temporal attention weight, a spatiotemporally enhanced feature video is obtained.

[0113] Specifically, the standardized video data undergoes "global average pooling," compressing the spatial information of each feature channel into a scalar value, ultimately generating a channel descriptor vector (with dimension [C], where C is the number of feature channels). For example, the feature channels (color channels) related to "worker's safety helmet" will aggregate to a higher scalar value, while the channels related to "background sky" will have a lower scalar value.

[0114] The channel descriptor vectors are input into an activation network consisting of fully connected layers and activation functions, and a channel attention weight matrix (dimension [C], numerical range [0,1]) is generated through nonlinear mapping. For example, the weight of the "edge of high-altitude work area" feature channel is increased to 0.8, while the weight of the "ground non-work area" channel is suppressed to 0.1.

[0115] Next, the channel attention weight matrix, which reflects the importance of the channel, is applied to the original video features to enhance the feature representation of high-risk areas and suppress low-risk areas.

[0116] Specifically, the original feature map of the standardized video data has dimensions [H, W, C] (H is height, W is width, and C is the number of channels), and the channel attention weight matrix has dimensions [C]. The channel attention weight matrix is ​​multiplied element-wise with the original feature map of the standardized video data. During element-wise multiplication, the feature value of each spatial location (x, y) is multiplied by the weight of the corresponding channel, resulting in a spatial feature enhancement map. For example, the spatial location (x1, y1) where "the worker is not wearing a safety belt" has a "edge feature channel" weight of 0.9; after multiplication, the edge feature value at this location is magnified by 9 times. Meanwhile, the "texture channel" weight of the "background wall" location (x2, y2) is 0.1, and the feature value is reduced by 10 times.

[0117] The spatial feature enhancement map is processed using 3D convolution to capture both spatial and temporal features, generating a spatiotemporal feature cube (dimensions [H, W, T, C], where T is the number of time frames). For example, it can capture continuous spatiotemporal features of "worker moving from the safety zone (frame t1) → entering the crane blind zone (frame t2) → staying (frame t3)".

[0118] Furthermore, "temporal dimension attention calculation" is performed on the spatiotemporal feature cube, and temporal attention weights (dimension [T], numerical range [0,1]) are generated through the Softmax function. For example, the weight of "worker entering the blind zone t2 frame" is 0.8, the weight of "stationary t1 frame" is 0.1, and the weight of "stationary t3 frame" is 0.6, highlighting the key frames where dynamic risks occur.

[0119] Finally, the "spatial enhancement features" are combined with "temporal keyframe weights" to generate the final features that simultaneously contain key spatial regions and key temporal dynamics. Specifically, the temporal attention weights are "weighted summation along the temporal dimension" with the spatiotemporal feature cube. That is, the feature values ​​of each spatial location (x, y) in the temporal dimension are accumulated according to the weights of the corresponding frames to obtain the spatiotemporally enhanced feature video. For example, the "worker in blind spot" feature in frame t2 (weight 0.8) and the "worker lingering" feature in frame t3 (weight 0.6) are emphasized and preserved, while the feature influence of frame t1 (weight 0.1) is weakened.

[0120] The spatiotemporal augmented video features highlight high-risk spatial areas, such as blind spots of tower cranes and high-altitude work platforms, while also condensing key temporal dynamics, such as the occurrence and continuation of violations.

[0121] Taking the video processing of "tower crane hoisting operations" as an example, the specific effect of this step is as follows:

[0122] 1) Generation of channel attention weight matrix: The weights of "tower crane boom edge" (edge ​​feature channel) and "worker's safety helmet color" (color channel) in the video are increased to 0.9, while the weight of "background trees" (texture channel) is reduced to 0.1;

[0123] 2) Spatial feature enhancement image generation: The features of the crane boom edge and worker outline are magnified, while the features of the background trees are weakened, and the outlines of high-risk targets (crane boom, worker) in the image are clearer;

[0124] 3) Spatiotemporal feature cube and temporal weight generation: capture consecutive frames of "crane starts to rotate (t1) → worker suddenly enters crane range (t2) → crane does not decelerate (t3)", with temporal weights of 0.8 and 0.7 for frames t2 and t3 respectively;

[0125] 4) Spatiotemporal Enhanced Video Feature Output: In the final features, the spatiotemporal linkage features of "worker intrusion in frame t2" and "crane boom not decelerating in frame t3" are significantly enhanced, providing core basis for subsequent identification of "tower crane collision risk".

[0126] In summary, this embodiment transforms raw video data into "risk-oriented spatiotemporal features" by highlighting risk-related spatial features and capturing risk-related temporal dynamics, enabling subsequent feature fusion and risk identification to focus on core risks and ignore invalid information.

[0127] Preferably, in one embodiment of the present invention, time-series features are extracted from standard log data to obtain time-series enhanced process features, and process relationship modeling is performed on the time-series enhanced process features to obtain structured process features, including:

[0128] The standard log data is semantically vectorized to obtain process feature vectors, and the process feature vectors are then modeled with temporal dependencies to capture the relationships between processes before and after execution and generate hidden state sequences.

[0129] The hidden state sequence is processed to identify key processes, and the time-series enhanced process features are obtained. These time-series enhanced process features are then mapped to graph node features to construct an initial process relationship graph.

[0130] The importance of nodes in the initial process relationship graph is calculated based on the graph attention mechanism, and the node weights are updated according to the dependence strength between processes to obtain the weighted process relationship graph.

[0131] The process relationship graph is subjected to feature aggregation processing to obtain structured process features. The feature aggregation processing is designed to integrate the features of each node in the process relationship graph with the association information of its neighboring nodes.

[0132] In this embodiment, through the progressive operations of "semantic vectorization, temporal dependency modeling, graph structure relationship modeling, and feature aggregation", deep features of "key processes + dependencies" are extracted from the structured logs to achieve accurate representation of the risk logic of construction processes.

[0133] Specifically, the unstructured log text is first transformed into computer-understandable vectors, capturing the "sequential logical relationships" between processes. Specifically, standard log data is encoded using a BERT model, transforming it into fixed-dimensional process feature vectors. Each vector contains semantic and attribute information about the process. For example, the "spreader inspection" vector has a high cosine similarity to the "lifting operation" vector (due to their strong correlation), but a low similarity to the "ground cleaning" vector.

[0134] The process feature vectors are input into an LSTM (Long Short-Term Memory) network in chronological order. The model learns the temporal relationships between processes through a gating mechanism. For example, the LSTM remembers that "spreading tool inspection" is a prerequisite process for "lifting operation." If the log shows an error in the order of "lifting without inspection," it will be reflected as an abnormal fluctuation in the hidden state. Finally, a sequence of hidden states is generated, which contains the "contextual information" of each process in the temporal chain.

[0135] Furthermore, "high-risk associated processes" (such as "high-altitude hoisting" and "deep foundation pit excavation") are selected from the time-series data and transformed into nodes in a graph structure, thus initially establishing a framework for the association between processes. Specifically, a "temporal attention mechanism" is applied to the hidden state sequence to calculate the attention weight of each process. The higher the weight, the stronger the association between the process and risk. The temporal enhancement process features are obtained through weighted summation.

[0136] Using time-series enhanced process features as "graph nodes", an initial process relationship graph is constructed based on the process connection relationships clearly defined in the logs. The "edges" between nodes represent "preceding-following" dependencies, such as the edge from node 1 to node 2 indicating that "lifting equipment inspection is a preceding process of lifting".

[0137] Furthermore, the "dependency strength" between processes is quantified, and key dependencies are strengthened through weight updates. Specifically, for each node in the initial process relationship graph, the model dynamically adjusts its own weight based on the "importance of neighboring nodes" and the "dependency strength of edges." For example, the "lifting operation" node will focus on the neighboring nodes of "lifting equipment inspection" (strong dependency, edge weight 0.9) and "safety net erection" (strong dependency, edge weight 0.8), while weakening the impact of "material handling" (weak dependency, edge weight 0.1). Figure 2 As shown, Figure 2 This is a schematic diagram of a high-altitude hoisting process in an embodiment of the present invention.

[0138] Each node is assigned an "importance score" through a self-attention mechanism. Nodes with higher dependency strength contribute more to the weight of their neighbors. For example, if the importance score of the "spreading equipment inspection" node is 0.9, the weight of the "lifting operation" node will increase to 0.8 due to the influence of its neighbor (initial weight 0.6 + 0.9 × 0.2). The resulting weighted process relationship graph shows that node weights reflect the "core position of the process in the risk chain," while edge weights reflect the "strength of risk transmission between processes."

[0139] Ultimately, the features of each process node are integrated with information from its "strongly dependent neighbors" to form a composite feature containing both "its own risk" and "related process risks." Specifically, a "mean aggregation" or "maximum aggregation" strategy is used to merge the features of each node with the features of its neighboring nodes (weighted according to edge weights). For example, the features of the "steel structure hoisting" node will be integrated with the "whether the inspection result is qualified" feature of "lifting equipment inspection" (edge ​​weight 0.9) and the "whether the protection range meets the standard" feature of "safety net erection" (edge ​​weight 0.8), ultimately generating a structured process feature containing multi-dimensional related information.

[0140] Structured process characteristics include both the individual attributes of a process and the related information of its dependent processes, fully reflecting the "contextual logic" of process risks.

[0141] Taking the log processing of "High-altitude steel structure hoisting of Building 3" as an example, the specific effect of this step is as follows:

[0142] 1) Semantic vectorization and temporal dependency modeling: The text "lifting equipment inspection → lifting operation → bolt tightening" in the log is converted into a vector sequence. LSTM captures the temporal dependency that "lifting operation must be after lifting equipment inspection". If the log shows an abnormal order of "lifting before inspection", the hidden state sequence will fluctuate significantly.

[0143] 2) Identification of key processes and construction of initial diagram: "Lifting operation" is identified as a key process (attention weight 0.9), and forms an initial relationship diagram with "lifting equipment inspection" and "bolt tightening" (edges represent "preceding-following" relationships);

[0144] 3) Graph attention weight update: The "Lifting Operation" node focuses on the neighbors of "Lifting Gear Inspection" (edge ​​weight 0.9) and "Safety Net Erection" (edge ​​weight 0.8), and the node weight is increased from 0.6 to 0.85;

[0145] 4) Feature aggregation: The structured process feature of "lifting operation" integrates neighboring information such as "lifting equipment inspection pass rate (90%)" and "safety net erection range (covering the lifting area)". The final feature can directly reflect that "the risk of lifting operation depends not only on itself, but also on the pre-inspection and protection measures".

[0146] In summary, this step transforms discrete log information into "structured features with risk logic" through "temporal modeling + graph structure analysis," enabling the system not only to "know what process is currently in progress" but also to "understand which related processes the risk of this process comes from," laying a core foundation for subsequent fusion with video features and accurate identification of compound risks.

[0147] Preferably, in one embodiment of the present invention, the spatiotemporally enhanced video features and structured process features are subjected to feature fusion processing to obtain fused data, including:

[0148] The spatiotemporal enhanced video features and structured process features are processed to unify their dimensions, generating feature pairs of the same dimension;

[0149] Based on the mutual information method, the correlation strength of feature pairs of the same dimension is quantified to obtain the cross-modal correlation matrix. The feature pairs of the same dimension are then interactively processed with the cross-modal correlation matrix to obtain cross-dimensional interactive features.

[0150] The correlation between each cross-dimensional interaction feature and construction risk is calculated based on predefined risk labels, and a dynamic weight vector corresponding one-to-one with the dimensions of the cross-dimensional interaction features is generated.

[0151] The cross-dimensional interactive features are multiplied element-wise with the dynamic weight vector to obtain adaptive weighted features.

[0152] The adaptive weighted features are subjected to nonlinear transformation to obtain deep fusion features. The deep fusion features are then combined with spatiotemporal enhanced video features and structured process features through residual connection to obtain enhanced fusion features.

[0153] The enhanced fusion features are normalized to obtain fused data.

[0154] This embodiment transforms spatiotemporally enhanced video features and structured process features into fused data that accurately reflects the risks associated with the linkage between "scene and process" through progressive operations of "dimensional adaptation, correlation quantification, dynamic weighting, and deep fusion".

[0155] Specifically, the spatiotemporal enhanced video features and structured process features are first processed to unify their dimensions, generating feature pairs of the same dimension. Specifically, linear projection (such as a fully connected layer) is used to map the dimensions of the two types of features: the spatiotemporal enhanced video features are reduced in dimension by convolution or flattened, and then projected to D dimensions through a fully connected layer; the structured process features are also projected to D dimensions through a fully connected layer, where D is a preset dimension.

[0156] Furthermore, the dependency relationship between video features and process features is quantified (e.g., the correlation between "workers not wearing safety belts" and "working at heights" is much higher than that with "working on the ground"), and the collaborative expression of highly correlated features is strengthened through interactive operations. Mutual information is used to measure the "shared information entropy" between two features; a higher value indicates a stronger correlation.

[0157] Specifically, mutual information is calculated for feature pairs (V, P) of the same dimension to generate a cross-modal correlation matrix M, where M[i,j] represents the correlation degree between the i-th dimension of V and the j-th dimension of P. Matrix multiplication (e.g., V×M×P^T) is then performed on the feature pairs (V, P) of the same dimension and the correlation matrix M to generate cross-dimensional interactive features. This operation amplifies the influence of highly correlated features through the correlation matrix.

[0158] This embodiment achieves risk-oriented feature selection by fusing features to "focus on high-risk associations" and suppress "low-risk associations." The predefined risk labels cover typical risks in construction, such as "falls from heights," "tower crane collisions," and "scaffolding collapses."

[0159] Specifically, the correlation between each cross-dimensional interactive feature and the risk label (range [0,1]) is calculated based on cosine similarity or multilayer perceptron (MLP). For example, the feature "worker under crane boom + hoisting process" has a correlation of 0.95 with the label "tower crane collision" and 0.05 with the label "scaffolding collapse". The maximum correlation between each interactive feature and all risk labels is taken to form a dynamic weight vector.

[0160] By amplifying the impact of high-risk interaction features with weights while suppressing low-risk features, adaptive enhancement is achieved, where "the higher the risk, the more prominent the feature." Specifically, element-wise multiplication is performed between cross-dimensional interaction features and dynamic weight vectors to obtain adaptive weighted features. In these adaptive weighted features, the feature values ​​of high-risk combinations are significantly amplified, while the feature values ​​of low-risk combinations are compressed, effectively filtering redundant information.

[0161] Finally, the "non-linear risk relationships" between features are extracted (e.g., the combined risks of "overload + strong wind + hoisting" are not simply linear superpositions), while residual connections are used to avoid the loss of original feature information. Specifically, adaptive weighted features are input into a network composed of a multilayer perceptron and activation functions to learn deep non-linear associations (e.g., the interactive risks of "tower crane tilt angle × wind speed × hoisting weight"), generating deep fusion features. These deep fusion features are then residually connected with the original spatiotemporally enhanced video features and structured process features.

[0162] The numerical values ​​of the enhanced fusion features are scaled to a uniform range to avoid the impact of differences in feature value magnitudes on subsequent model training and inference. Specifically, L2 normalization or batch normalization is used. L2 normalization scales the feature vectors to the unit norm; batch normalization stabilizes the feature distribution by adjusting the mean and variance.

[0163] Preferably, in one embodiment of the present invention, the fused data is input into a pre-trained risk identification model to identify risk features, thereby obtaining a risk feature vector, including:

[0164] To construct a training dataset, historical fusion data and corresponding risk event records in the construction field are obtained.

[0165] An initial risk identification model based on the TimeSformer model is constructed, and the initial risk identification model is transferred to the training dataset to obtain a trained risk identification model.

[0166] The fused data is segmented according to time series and input into the risk identification model to output multi-scale risk features.

[0167] The risk features at multiple scales are sequentially processed by feature concatenation and feature dimensionality reduction to obtain the risk feature vector.

[0168] This embodiment constructs a risk identification model based on TimeSformer, accurately extracts risk features linked by "time-space-process" from fused data, and finally outputs a risk feature vector that can be directly used for risk assessment.

[0169] In this embodiment, firstly, the model is provided with learning samples consisting of "fused data and risk labels," allowing it to learn to identify risk characteristics through historical cases. Specifically, fused data processed using the above steps is collected from construction projects over the past 3-5 years, covering different processes and risk scenarios. Each piece of historical fused data is labeled with a corresponding risk label, including:

[0170] Risk type (e.g., "fall from height risk" or "tower crane collision risk");

[0171] Time of risk occurrence (e.g., "Tower crane collision warning occurred at 14:30 on 2023-05-10");

[0172] Risk level (e.g., 1-5, with 5 being the highest).

[0173] The dataset is divided into training, validation, and test sets in a 7:2:1 ratio, and the model's generalization ability is improved through data augmentation (such as shuffling the time sequence and adding noise to features).

[0174] Next, leveraging TimeSformer's strong ability to capture "spatiotemporal correlation features," and combining it with data from the construction industry, we conducted customized training to obtain a model that can accurately identify construction risks.

[0175] Among them, the risk identification model is based on the Transformer architecture and supports multiple attention mechanisms, such as spatial attention and temporal attention. It can capture the spatial and temporal features of the video at the same time, making it particularly suitable for handling dynamic scenes with "time-space" linkage.

[0176] Based on the TimeSformer pre-trained on a general video dataset (such as Kinetics-400), retaining its bottom-level feature extraction layer, and replacing the top-level classification head with a "construction risk classification layer," an initial risk identification model is constructed. The initial model is then fine-tuned using the constructed training dataset, including:

[0177] Freeze 30% of the parameters at the bottom layer and train only the top-level classifier head and intermediate layer parameters; adopt the "learning rate decay" strategy, that is, the initial learning rate is 1e-4, and it decays by 50% every 10 rounds. Optimize the model through the cross-entropy loss function until the accuracy of the validation set is stable.

[0178] Transfer training allows the model to quickly adapt to the specificities of building construction scenarios while retaining its "spatiotemporal feature extraction capabilities," significantly reducing the amount of training data and improving the accuracy of risk identification.

[0179] Furthermore, it captures risk characteristics across different time windows, such as instantaneous risk, short-term cumulative risk, and long-term trend risk, to avoid missed risks due to a single time scale. Specifically, the fused data (a feature sequence arranged chronologically) is divided into segments of three time scales:

[0180] Fine-grained segments: 1 segment per second (capturing instantaneous risks, such as "a worker suddenly entering the crane's range");

[0181] Medium-grained segments: 1 segment every 10 seconds (capturing short-term cumulative risks, such as "workers staying in a danger zone for more than 10 seconds");

[0182] Coarse-grained segments: 1 segment every 60 seconds (capturing long-term trend risks, such as "the tilt angle of the tower crane boom continues to increase").

[0183] Segments of different scales are fed into the trained TimeSformer model, and the model outputs risk features at the corresponding scales through different levels of attention mechanisms:

[0184] Fine-grained features: focusing on "spatial location + instantaneous action" (e.g., "worker coordinates × boom position");

[0185] Medium-grained characteristics: Focus on "action sequence + process status" (e.g., "worker fails to wear safety belt 3 times in a row + high-altitude operation continues");

[0186] Coarse-grained characteristics: Focus on "trend changes + environmental impacts" (such as "the boom tilt angle increases by 2 degrees every 5 minutes + wind speed continues to rise").

[0187] Finally, complementary information from multi-scale features is integrated to generate a compact and efficient risk feature vector. Specifically, fine-grained, medium-grained, and coarse-grained risk features are concatenated dimensionally to form a high-dimensional feature matrix, which is then dimensionality-reduced using principal component analysis or linear discriminant analysis. The concatenated features contain both instantaneous risk signals and cumulative and trend information, fully reflecting the temporal evolution of risk.

[0188] In summary, this embodiment transforms the risk information contained in the fused data into a risk feature vector that can be understood and calculated by machines through "data-driven model training + multi-scale feature extraction + quantized vector output". It is a key bridge connecting feature fusion and risk judgment, and directly determines the accuracy of subsequent early warning.

[0189] Preferably, in one embodiment of the present invention, the acquired real-time process data and risk feature vector are subjected to correlation analysis processing to obtain correlation analysis results, including:

[0190] Entity extraction is performed on real-time process data to construct a process feature entity set;

[0191] The correlation strength between the process feature entity set and the risk feature vector is calculated based on the cosine similarity algorithm to obtain the correlation matrix between the process and the risk.

[0192] Cluster analysis was performed on the real-time process data based on the correlation matrix to obtain high-correlation process groups and low-correlation process groups;

[0193] Based on the historical risk occurrence probability distribution, the threshold interval of the first risk feature vector corresponding to each highly correlated process group is calculated to obtain the dynamic risk threshold for process matching.

[0194] The deviation between the second risk feature vector and the feature entity set corresponding to each low-association process group is fitted and calculated using the least squares method to obtain the feature correction coefficient used to correct the risk feature weight.

[0195] The dynamic risk threshold and feature correction coefficient are linked and stored to generate a correlation analysis result containing process ID, dynamic risk threshold and feature correction coefficient.

[0196] This embodiment dynamically correlates real-time process data with risk feature vectors to generate risk assessment criteria (dynamic risk thresholds) and feature optimization parameters (feature correction coefficients) that are highly matched to the current process scenario. Specifically, different construction processes (such as "foundation pouring" and "high-altitude hoisting") have different risk characteristics, requiring customized risk thresholds; at the same time, risk features that are weakly correlated with the current process are weighted and corrected to avoid interference from irrelevant risk signals.

[0197] In this embodiment, firstly, key entities are extracted from unstructured real-time process text to form machine-processable structured features. Specifically, a BERT-based Named Entity Recognition (NER) model is used to identify the following entity types from the real-time process data: time entities; location entities; process type entities; equipment entities; and parameter entities. Relationships between entities are established through dependency parsing to form a process feature entity set.

[0198] It should be noted that although the real-time process data processed in this step and the standard log data in step S2 are both related to construction processes, they differ significantly in terms of data source, timeliness, content granularity, and core function.

[0199] 1) The standard log data in S2 mainly comes from structured / semi-structured logs manually recorded before or during construction. It is subject to lag, as it may be pre-construction planning records or post-construction supplementary information, and is not real-time synchronized field data. The real-time process data in this step mainly comes from dynamic data collected in real-time at the construction site. It has strong real-time performance, and the data is synchronized with the on-site construction status, with a delay of no more than 1 minute.

[0200] 2) The standard log data in S2 mainly consists of macro-level process plans and static information, with a relatively coarse granularity. Although it is structured after processing, it focuses on the logical relationships between processes and does not include real-time dynamic parameters. The real-time process data in this step mainly consists of micro-level real-time status and dynamic parameters, with an extremely fine granularity. After entity extraction, it has a higher degree of structure and focuses on the real-time status of the current process, directly related to on-site risk parameters.

[0201] 3) The core function of the standard log data in S2 is to fuse with video data, providing a static logical background for the process. This data serves as a "feature-level input," assisting the model in understanding risk scenarios within the video. The core function of the real-time process data in this step is to correlate with risk feature vectors, dynamically adapting to risk assessment criteria. This data serves as a "decision-level input," dynamically adjusting the thresholds and weights for risk assessment, directly determining the warning results.

[0202] Next, the correlation between the "current process" and "various risk characteristics" is quantified to provide a basis for subsequent clustering. Specifically, the process feature entity set is transformed into a vector representation to obtain the process feature vector. The cosine similarity is calculated between each dimension in P and each risk feature dimension in the risk feature vector to generate a correlation matrix.

[0203] Using the correlation matrix as input, the K-means algorithm is used to divide the process features into two categories, and the clustering quality is evaluated by the silhouette coefficient to ensure that the discriminant between the two categories is >0.6.

[0204] Highly correlated process groups: those with an average correlation degree greater than or equal to a threshold (e.g., 0.7) with the risk feature vector, and containing process features that directly affect the current risk;

[0205] Low-association process groups: those with an average correlation to the risk feature vector less than the threshold, containing indirect or irrelevant features.

[0206] Furthermore, risk assessment criteria are customized for the current process. Specifically, historical risk data is grouped by process type, a probability distribution curve is fitted to the risk feature vector of each group, and quantile thresholds corresponding to different risk levels are calculated, including:

[0207] Low-risk threshold: corresponds to the 90th percentile, meaning that 90% of normal samples in historical data are below this value;

[0208] Medium risk threshold: corresponding to the 95th percentile;

[0209] High-risk threshold: corresponding to the 99th percentile.

[0210] Based on the risk level of the current process, select the corresponding threshold as the dynamic risk threshold.

[0211] Furthermore, the accuracy of early warnings is improved by adjusting the weights of features that are "weakly correlated with the current process but may affect risk assessment." Specifically, for each feature in a low-correlation process group, its deviation from the risk feature vector is calculated. The relationship between the deviation and the historical probability of risk occurrence is then fitted using the least squares method to obtain a correction coefficient. For example:

[0212] When "actual wind speed > standard wind speed + 2 m / s", the weight of the "risk of falling from height" dimension should be multiplied by a correction factor of 1.2;

[0213] When the actual temperature is <5℃, the weight of the "equipment failure risk" dimension should be multiplied by a correction factor of 1.3.

[0214] The correction coefficients of all low-correlation features are combined into a matrix for subsequent risk feature weight adjustment.

[0215] Ultimately, customized risk assessment criteria and feature correction parameters are linked to specific processes to form a traceable and reusable risk analysis knowledge base. Specifically, dynamic risk thresholds and feature correction coefficients are associated with process IDs and stored in JSON format. Newly generated association analysis results are stored in the risk knowledge base, and the thresholds and coefficients are updated periodically based on new data to ensure system adaptability.

[0216] Preferably, in one embodiment of the present invention, risk level determination is performed based on the correlation analysis results, and real-time early warning of target building construction is provided based on the risk level determination results, including:

[0217] Extract the dynamic risk threshold and feature correction coefficient matching the current process from the correlation analysis results, and perform a weighted operation on the risk feature vector and feature correction coefficient to obtain the corrected risk value;

[0218] The risk level is obtained by comparing the corrected risk value with the dynamic risk threshold.

[0219] Real-time early warning of target building construction based on risk level.

[0220] This embodiment combines risk feature vectors with correlation analysis results to complete a closed loop from "risk feature quantification" to "level determination" and then to "real-time early warning".

[0221] In this embodiment, the original value of the risk feature vector is adjusted to better reflect the actual risk level on site, taking into account the specific circumstances of the current process. Specifically, dynamic risk thresholds and feature correction coefficients matching the current process are extracted from the correlation analysis results. Each dimension in the risk feature vector is weighted according to the corresponding correction coefficient. For example, if the original value of the "fall from height risk" dimension in the risk feature vector is 0.7 and the corresponding correction coefficient is 1.2, then the corrected R[fall from height] = 0.7 × 1.2 = 0.84.

[0222] Furthermore, the risk level is determined based on whether the "corrected risk value exceeds the dynamic threshold" and the degree to which it exceeds the threshold, providing a basis for subsequent early warning measures. Specifically, the dynamic risk threshold is divided into four levels according to risk level, including:

[0223] Low-risk threshold T1: corresponds to "no intervention required" (e.g., if the corrected value is ≤ T1, the risk level is 1).

[0224] Medium risk threshold T2: corresponds to "warning for attention" (T1 < revised value ≤ T2, risk level = 2);

[0225] High-risk threshold T3: corresponds to "suspension of work" (T2 < revised value ≤ T3, risk level = 3);

[0226] Extremely high risk threshold T4: corresponds to "emergency evacuation" (corrected value > T3, risk level = 4).

[0227] Each corrected risk value is compared with the threshold of the corresponding dimension, and the highest risk level is taken as the final result. For example, the corrected value of "crane tilt risk" is 0.85 > T3 (0.8) → Level 3; the corrected value of "fall from height risk" is 0.75 ≤ T2 (0.8) → Level 2; then the final risk level is Level 3.

[0228] Differentiated early warning measures are triggered based on the risk level to minimize disruption to construction while ensuring safety. This embodiment pre-sets four early warning levels, each corresponding to a risk level, as shown in Table 1. Table 1 is the early warning plan level table.

[0229] Table 1

[0230]

[0231] In this embodiment, after an early warning is issued, the system continues to monitor the on-site status until the risk is eliminated. For example, in a Level 3 early warning, after the equipment is suspended and the workers have completed the rectification, the system will automatically resume operations upon confirmation.

[0232] For details, please see Figure 3 , Figure 3The diagram shows the flowchart of a real-time early warning method for construction safety risks in one embodiment of the present invention.

[0233] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0234] 1) This invention employs a dual-channel attention mechanism to fuse video monitoring data and construction log data at the feature level, which not only retains key visual features (such as personnel violations) but also integrates process logic features (such as high-altitude operation specifications), thus solving the one-sidedness problem of traditional single-source data early warning.

[0235] 2) This invention uses correlation analysis to analyze real-time process and risk characteristics, customizes dynamic risk thresholds for different construction processes (such as the threshold for not wearing a safety belt in high-altitude operations is much lower than that for ground operations), and introduces feature correction coefficients to adjust risk weights, so that the early warning standard can adapt to changes in the construction scenario and significantly reduce the false alarm rate.

[0236] 3) This invention captures the spatiotemporal dynamic features of video based on the TimeSformer model (such as personnel not wearing safety equipment for 3 consecutive minutes), and combines it with the temporal information of construction logs (such as the order of process execution) to achieve accurate identification of complex risk scenarios, such as multiple risk superposition scenarios of "violation of regulations + equipment abnormality + high-risk process".

[0237] Another embodiment of the present invention provides a real-time early warning system for construction safety risks. For details, please refer to [link / reference]. Figure 4 , Figure 4 The diagram shown is a schematic diagram of a real-time early warning system for construction safety risks according to one embodiment of the present invention, wherein:

[0238] The first acquisition module is used to acquire standardized video data of the target building construction process in response to the risk video recognition signal;

[0239] The second acquisition module is used to respond to the risk log identification signal, perform content identification on the construction log data corresponding to the construction process of the target building, and obtain standard log data.

[0240] The first extraction module is used to process the standardized video data based on spatiotemporal enhancement technology to obtain spatiotemporal enhanced video features;

[0241] The second extraction module is used to extract time-series features from the standard log data to obtain time-series enhanced process features, and to model the process relationships of the time-series enhanced process features to obtain structured process features.

[0242] The fusion module is used to perform feature fusion processing on the spatiotemporal enhanced video features and the structured process features to obtain fused data;

[0243] The identification module is used to input the fused data into a pre-trained risk identification model to identify risk features and obtain a risk feature vector.

[0244] The early warning module is used to respond to risk warning signals by performing correlation analysis on the acquired real-time process data and the risk feature vector to obtain correlation analysis results, and to determine the risk level based on the correlation analysis results, and to provide real-time early warning for the construction of the target building based on the risk level determination results.

[0245] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A real-time pre-warning method for construction safety risks, characterized in that, Comprising: in response to a risk video identification signal, obtaining standardized video data of a target building construction process; in response to a risk log identification signal, performing content identification on construction log data corresponding to the target building construction process to obtain standard log data; processing the standardized video data based on a spatio-temporal enhancement technique to obtain spatio-temporal enhanced video features; extracting time sequence features from the standard log data to obtain time sequence enhanced process features, and modeling process relationships of the time sequence enhanced process features to obtain structured process features; performing feature fusion processing on the spatio-temporal enhanced video features and the structured process features to obtain fusion data; inputting the fusion data into a pre-trained risk identification model to perform risk feature identification and obtain a risk feature vector; in response to a risk warning signal, performing correlation analysis processing on the obtained real-time process data and the risk feature vector to obtain a correlation analysis result, and based on the correlation analysis result, performing risk level determination, and according to the risk level determination result, performing real-time warning of the target building construction; the processing of the standardized video data based on the spatio-temporal enhancement technique to obtain the spatio-temporal enhanced video features, comprising: performing global information aggregation processing on the standardized video data to generate a channel descriptor vector, and performing nonlinear transformation processing on the channel descriptor vector to generate a channel attention weight matrix; element-wise multiplying the channel attention weight matrix and the original feature map of the standardized video data to obtain a spatial feature enhancement map; performing spatio-temporal joint feature extraction on the spatial feature enhancement map to generate a spatio-temporal feature cube, and performing frame-level importance weighting processing on the spatio-temporal feature cube to generate a time sequence attention weight; fusing the spatio-temporal feature cube and the time sequence attention weight to obtain a spatio-temporal enhanced feature video; the time sequence feature extraction from the standard log data to obtain the time sequence enhanced process features, and the modeling of the process relationships of the time sequence enhanced process features to obtain the structured process features, comprising: performing semantic vectorization processing on the standard log data to obtain a process feature vector, and performing time sequence dependence modeling on the process feature vector to capture the before-and-after correlation of process execution, and generating a hidden state sequence; performing key process identification processing on the hidden state sequence to obtain time sequence enhanced process features, and mapping the time sequence enhanced process features to graph node features to construct an initial process relationship graph; performing node importance calculation on the initial process relationship graph based on a graph attention mechanism, updating node weights according to the dependence strength between processes, and obtaining a weighted process relationship graph; performing feature aggregation processing on the process relationship graph to obtain structured process features, wherein the feature aggregation processing is designed to fuse the correlation information of neighbor nodes for each node feature in the process relationship graph; the correlation analysis processing on the real-time process data and the risk feature vector to obtain the correlation analysis result, comprising: performing entity extraction on the real-time process data to construct a process feature entity set; Based on the cosine similarity algorithm, the process feature entity set and the risk feature vector are associated to calculate the correlation strength, and a process and risk correlation matrix is obtained; According to the correlation matrix, the real-time process data is clustered and analyzed to obtain a high correlation process group and a low correlation process group; Based on the historical risk occurrence probability distribution, the first risk feature vector corresponding to each high correlation process group is calculated to obtain a dynamic risk threshold for process matching; According to the least square method, the second risk feature vector corresponding to each low correlation process group and the deviation value of the process feature entity set are fitted and calculated to obtain a feature correction coefficient for correcting the risk feature weight; The dynamic risk threshold and the feature correction coefficient are associated and stored to generate an association analysis result including process ID, dynamic risk threshold and feature correction coefficient.

2. The real-time pre-warning method for construction safety risk according to claim 1, characterized in that, The standardization video data of the target building construction process is obtained, including: Based on the adaptive Gaussian filter algorithm, the obtained video monitoring data is dynamically denoised to obtain a denoised video frame; The target frame is labeled and processed to obtain a target video segment, and the target video segment is standardized to obtain standardized video data; wherein the target frame includes a construction personnel target frame and a special equipment target frame.

3. The real-time pre-warning method for construction safety risk according to claim 1, characterized in that, The content of the construction log data corresponding to the target building construction process is identified to obtain standard log data, including: Target information extraction is performed on the construction log data to obtain a log core data set; According to the pre-constructed building construction database, the log core data set is mapped to obtain a structured log item; The structured log item is time-aligned to obtain standard log data.

4. The real-time pre-warning method for construction safety risk according to claim 1, characterized in that, The spatio-temporal enhanced video feature and the structured process feature are fused to obtain fusion data, including: The spatio-temporal enhanced video feature and the structured process feature are dimensionally unified to generate a same-dimension feature pair; Based on the mutual information method, the same-dimension feature pair is quantitatively processed to obtain a cross-modal correlation matrix, and the same-dimension feature pair and the cross-modal correlation matrix are interactively operated to obtain a cross-dimension interaction feature; Based on the pre-defined risk label, the correlation degree of each cross-dimension interaction feature and the construction risk is calculated to generate a dynamic weight vector corresponding to the cross-dimension interaction feature dimension by one; The cross-dimension interaction feature and the dynamic weight vector are element-level multiplied to obtain an adaptive weighted feature; The adaptive weighted feature is nonlinearly transformed to obtain a deep fusion feature, and the deep fusion feature, the spatio-temporal enhanced video feature and the structured process feature are residual connected to obtain an enhanced fusion feature; The enhanced fusion feature is normalized to obtain fusion data.

5. The real-time pre-warning method for construction safety risk according to claim 1, characterized in that, The fusion data is input into the pre-trained risk identification model to identify the risk feature, and a risk feature vector is obtained, including: The historical fusion data and the corresponding risk event records in the field of building construction are obtained to construct a training data set; An initial risk identification model based on a TimeSformer model is constructed, and the initial risk identification model is migrated and trained based on the training data set to obtain a trained risk identification model; The fusion data is input into the risk identification model in a time sequence to output multi-scale risk features; The multi-scale risk features are sequentially subjected to feature splicing and feature dimension reduction processing to obtain a risk feature vector.

6. The real-time pre-warning method of construction safety risk according to claim 1, characterized in that, The risk level is determined based on the correlation analysis result, and real-time warning of the target building construction is performed according to the risk level determination result, including: The dynamic risk threshold and the feature correction coefficient matched with the current process are extracted from the correlation analysis result, the risk feature vector is weighted with the feature correction coefficient to obtain a corrected risk value; The corrected risk value is compared with the dynamic risk threshold to obtain a risk level; Real-time warning of the target building construction is performed based on the risk level.

7. A real-time pre-warning system for construction safety risks, characterized in that, It includes: The first acquisition module is configured to acquire standardized video data of a target building construction process in response to a risk video identification signal; The second acquisition module is configured to perform content identification on construction log data corresponding to the target building construction process in response to a risk log identification signal to obtain standard log data; The first extraction module is configured to process the standardized video data based on a spatiotemporal enhancement technique to obtain spatiotemporal enhanced video features; The second extraction module is configured to extract time sequence features from the standard log data to obtain time sequence enhanced process features, and model process relationships of the time sequence enhanced process features to obtain structured process features; The fusion module is configured to perform feature fusion processing on the spatiotemporal enhanced video features and the structured process features to obtain fusion data; The identification module is configured to input the fusion data into a pre-trained risk identification model to identify risk features and obtain a risk feature vector; The warning module is configured to perform correlation analysis on real-time process data and the risk feature vector obtained in response to a risk warning signal to obtain a correlation analysis result, determine a risk level based on the correlation analysis result, and perform real-time warning of the target building construction according to the risk level determination result; The standardized video data is processed based on the spatiotemporal enhancement technique to obtain spatiotemporal enhanced video features, including: The standardized video data is subjected to global information aggregation processing to generate a channel descriptor vector, and the channel descriptor vector is subjected to nonlinear transformation processing to generate a channel attention weight matrix; The channel attention weight matrix is multiplied element by element with an original feature map of the standardized video data to obtain a spatial feature enhancement map; The spatial feature enhancement map is subjected to spatiotemporal joint feature extraction to generate a spatiotemporal feature cube, and the spatiotemporal feature cube is subjected to frame-level importance weighting processing to generate a time sequence attention weight; The spatiotemporal feature cube and the time sequence attention weight are fused to obtain a spatiotemporal enhanced feature video; The standard log data is subjected to time sequence feature extraction to obtain time sequence enhanced process features, and process relationship modeling is performed on the time sequence enhanced process features to obtain structured process features, including: The standard log data is subjected to semantic vectorization processing to obtain process feature vectors, and time sequence dependence modeling is performed on the process feature vectors to capture the before-and-after correlation of process execution, thereby generating a hidden state sequence; The hidden state sequence is subjected to key process identification processing to obtain time sequence enhanced process features, and the time sequence enhanced process features are mapped to graph node features to construct an initial process relationship graph; Based on the graph attention mechanism, node importance calculation is performed on the initial process relationship graph, and the node weights are updated according to the dependence strength between processes to obtain a weighted process relationship graph; The process relationship graph is subjected to feature aggregation processing to obtain structured process features, wherein the feature aggregation processing is designed to fuse the correlation information of neighbor nodes for each node feature in the process relationship graph; The obtained real-time process data and the risk feature vector are subjected to correlation analysis processing to obtain a correlation analysis result, including: Entity extraction is performed on the real-time process data to construct a process feature entity set; Based on the cosine similarity algorithm, correlation strength calculation is performed on the process feature entity set and the risk feature vector to obtain a correlation degree matrix of processes and risks; According to the correlation degree matrix, clustering analysis is performed on the real-time process data to obtain high-correlation process groups and low-correlation process groups; Based on the historical risk occurrence probability distribution, threshold interval calculation processing is performed on the first risk feature vector corresponding to each high-correlation process group to obtain a dynamic risk threshold for process matching; According to the least squares method, fitting calculation is performed on the deviation values of the second risk feature vector corresponding to each low-correlation process group and the process feature entity set to obtain a feature correction coefficient for correcting the risk feature weight; The dynamic risk threshold and the feature correction coefficient are stored in association to generate a correlation analysis result containing process ID, dynamic risk threshold, and feature correction coefficient.

Images