Steel structure construction safety risk real-time identification and early warning method

Abstract

The invention discloses a steel structure construction safety risk real-time identification and early warning method, which comprises the following steps of: constructing a space-time group behavior model by collecting multi-source context data such as a construction stage, an environmental parameter, a task type and personnel allocation and combining space trajectories and behavior characteristics of multi-role personnel through standardization and feature coding; context-enhanced behavior anomaly detection is realized through deep learning and a graph neural network, and abnormal behaviors are discriminated in real time and early-warned and pushed based on an early-warning information generation mechanism adaptive to a dynamic threshold value and a context; on-site feedback is utilized to realize online incremental learning and adaptive optimization of the model, the sensitivity of behavior recognition to on-site state and task change is enhanced, the accuracy and response timeliness of anomaly detection are improved, and intelligent upgrading of construction safety management is assisted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel structure construction safety monitoring, and particularly relates to a steel structure construction safety risk real-time identification and early warning method. BACKGROUND

[0002] At present, the field of steel structure construction safety monitoring and group behavior analysis is gradually developing towards intelligence, precision and automation. With the wide application of large and complex steel structure projects, multi-role collaboration, real-time safety warning and behavior anomaly identification on the construction site have become a highly concerned problem in the industry. Traditional steel structure construction safety control relies on manual inspection, on-site video monitoring and post-accident investigation. Although some high-end projects have introduced personnel positioning, environmental sensing and automatic alarm information technology, the intelligent level of group behavior anomaly analysis still has a lot of room for improvement. In recent years, personnel positioning systems represented by UWB, RFID and other technologies, as well as environmental parameter monitoring systems based on various sensor networks have emerged in the industry. These systems can realize the basic collection of the position, motion trajectory and related environmental information of the construction site workers, providing data support for subsequent data analysis and safety warning. At the same time, with the help of emerging technologies such as deep learning, graph neural networks and spatio-temporal behavior recognition, some academic research and enterprise products have initially possessed the ability to automatically identify and detect anomalies in individual behavior and local group activities. For example, the disclosed behavior anomaly detection method often uses trajectory clustering, spatio-temporal graph modeling, LSTM or Transformer network to model personnel motion patterns, realizing automatic discrimination of abnormal walking, residence, gathering or work conflict behaviors. However, the existing group behavior analysis methods have obvious defects and are difficult to adapt to the actual needs of complex and variable steel structure construction sites. The core pain points are: (1) Lack of context awareness, the mainstream behavior anomaly detection scheme mostly relies on personnel trajectory, speed, acceleration and other motion features, ignoring key context factors such as construction phase, current environmental state (temperature, humidity, wind speed, light, etc.), specific task type and work role allocation. This leads to misjudgment or false reporting of the model under different construction scenarios or external influences (such as bad weather, special task nodes), and limited ability to identify complex collaborative behaviors; (2) Single feature fusion, imperfect decision mechanism, existing technologies mostly rely on simple feature splicing or statistical indicators in multi-source data fusion, lacking effective spatio-temporal alignment, multi-modal feature weighting and graph structure collaborative relationship modeling. Most schemes are difficult to accurately capture the dynamic interaction patterns among multiple roles and their deep coupling relationship with the scene context, affecting the comprehensive discrimination ability of collaborative anomalies and safety risks; (3) The early warning mechanism lacks adaptability and pertinence. At present, the system generally uses fixed thresholds or static response templates in the disposal of abnormal events and the push of early warning, ignoring the influence of factors such as dynamic changes in construction progress, task complexity, and environmental mutations on safety decision-making, and there are problems such as early warning generalization and low precision. (4) The model is difficult to continuously adapt to dynamic changes on site. The site environment, operation process, and personnel team structure have high time variability. Most existing models use offline training parameters and lack online self-learning ability combined with real-time feedback and manual review, making it difficult to respond to the influence of sudden events and long-term environmental changes on group behavior patterns, and easily leading to model "aging" and reducing safety protection effect. SUMMARY

[0003] The present application provides a steel structure construction safety risk real-time identification and early warning method to solve the above technical problems.

[0004] The technical scheme of the present application is as follows: a steel structure construction safety risk real-time identification and early warning method, comprising: S1: Real-time acquisition of multi-source context data of steel structure construction site, the context data including construction phase information, environmental parameters, task type and personnel allocation information; S2: Preprocessing the collected context data, including missing value filling, outlier removal and standardization conversion, to generate a unified format context feature vector; S3: Obtain the positioning data and behavior trajectory information of the multi-role operating personnel on the construction site, and construct a group behavior space-time model based on the space-time coordinates; S4: Feature fusion of the context feature vector and the group behavior space-time model to form a context-enhanced group behavior representation dataset; S5: Based on the deep learning algorithm, the context-enhanced group behavior representation dataset is used to train a context-aware behavior anomaly detection model; S6: Input the real-time collected context feature vector and the current group behavior trajectory into the trained behavior anomaly detection model to generate a behavior anomaly score result; S7: Determine whether the behavior anomaly score exceeds the preset threshold, if it exceeds, trigger an abnormal behavior recognition signal; S8: Based on the triggered abnormal behavior recognition signal, generate targeted early warning information and response suggestions combined with the current context features, and push them to the terminal of the site manager; S9: Periodically collect model prediction results and manual review feedback information, and perform online incremental training and parameter optimization on the behavior anomaly detection model.

[0005] The steel structure construction safety risk real-time identification and early warning method provided by the application has the following beneficial effects: (1) Significantly improve the accuracy and robustness of abnormality identification. Traditional group behavior analysis methods often ignore key context factors such as construction stage, environmental changes, and task allocation, which can easily lead to misjudgment of normal behavior patterns in different construction stages. The present application effectively represents the complex and variable spatio-temporal relationship of the work scene by standardizing the fusion of construction stage one-hot encoding, environmental parameters such as temperature, humidity, and wind speed, task type, and personnel allocation information into high-dimensional context features, greatly reducing false positives caused by isolated behavior analysis.

[0006] (2) Strengthen the context adaptive ability of the model. The present application designs a multi-modal feature fusion and graph neural network structure, which can adaptively perceive and focus on the key safety context of the current construction scene by dynamically adjusting the weights of context information and behavior features during model reasoning. For example, automatically increase the proportion of environmental factors in high-risk stages (such as high-altitude hoisting, severe weather, etc.), to ensure that the model can accurately identify context-related abnormal group behavior, avoid misjudgment caused by single feature changes, and significantly improve the practicality and generalization ability of the system in complex construction site environments such as different process flows, seasonal changes, and climate mutations.

[0007] (3) Automation and high-quality support for data collection and preprocessing. The present application innovatively integrates multi-source sensor real-time data, intelligent missing value repair, automatic outlier removal, and dynamic normalization, etc., to ensure the continuity and accuracy of input data, and significantly overcome common technical bottlenecks such as packet loss, asynchronous sampling of heterogeneous devices, etc. This mechanism makes the subsequent model's learning of the spatio-temporal structure of behavior more stable, improves the accuracy of trajectory feature extraction, and ensures the reliability of the abnormality detection results.

[0008] (4) Intelligence and pertinence of abnormal response mechanism. The present application not only can output abnormal scores and adaptively determine thresholds in real time, but also deeply combines the abnormality identification results with the current stage and task type of construction, and automatically matches the most suitable warning information and disposal suggestions for the current scene through a rule engine. In practical applications, it can significantly improve the speed and accuracy of management personnel's response to sudden safety incidents, and reduce safety hazards caused by response delays and high false positive rates.

[0009] (5) Reduce the cost of manual monitoring and improve the efficiency of early warning. Through automated data fusion, abnormality detection, and response suggestions, the whole process realizes high-frequency real-time discrimination and early warning push, greatly reducing the need for manual patrols and relying solely on video inspections. In large, widely distributed, or dangerous construction sites, it can achieve real-time safety control in the whole process and in the whole area, improve productivity, and reduce the probability of safety accidents. BRIEF DESCRIPTION OF DRAWINGS

[0010] Fig. 1 A flow chart of a steel structure construction safety risk real-time identification and early warning method of the present application is shown in the figure; Fig. 2 A sub-flow chart of a steel structure construction safety risk real-time identification and early warning method of the present application is shown in the figure; Fig. 3 Another sub-flow chart of a steel structure construction safety risk real-time identification and early warning method of the present application is shown in the figure. DETAILED DESCRIPTION

[0011] In order to make the purpose and advantages of the present application more clear and obvious, the present application will be further described below in conjunction with examples; it should be understood that the specific examples described herein are only used to explain the present application, and do not limit the present application.

[0012] The preferred implementation methods of the present application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these implementation methods are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0013] As used herein, the singular forms "a", "an" and "the" can also include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the terms "comprise / contain" or "have" and the like specify the presence of stated features, integers, steps, operations, components, parts or combinations thereof, but do not exclude the presence or addition of one or more other features, integers, steps, operations, components, parts or combinations thereof. At the same time, the term "and / or" used in the specification includes any and all combinations of the related listed items.

[0014] Please refer to Figs. 1-3 A steel structure construction safety risk real-time identification and early warning method is shown in the figure, which comprises: S1: Real-time acquisition of multi-source context data of steel structure construction site, the context data comprising construction phase information, environmental parameters, task type and personnel allocation information; S2: Preprocessing of the acquired context data, including missing value filling, outlier removal and standardization conversion, to generate a context feature vector in a unified format; S3: Obtaining positioning data and behavior trajectory information of multi-role workers at the construction site, and constructing a group behavior spatio-temporal model based on spatio-temporal coordinates; S4: Feature fusion of the context feature vector and the group behavior spatio-temporal model to form a context-enhanced group behavior representation dataset; S5: Training a context-aware behavior anomaly detection model based on deep learning algorithm using the context-enhanced group behavior representation dataset; S6: input the real-time collected context feature vector and the current group behavior trajectory into the trained behavior anomaly detection model to generate a behavior anomaly score result; S7: determine whether the behavior anomaly score exceeds a preset threshold, and if so, trigger an abnormal behavior recognition signal; S8: based on the triggered abnormal behavior recognition signal, generate targeted warning information and response suggestions in combination with the current context features, and push them to the terminal of the on-site management personnel; S9: periodically collect model prediction results and manual review feedback information, and perform online incremental training and parameter optimization on the behavior anomaly detection model.

[0015] The step S1: real-time collection of multi-source context data of the steel structure construction site, the context data including construction phase information, environmental parameters, task type and personnel allocation information. Specifically, it includes: S1.1: obtain the current construction phase information based on the construction progress management system interface, the construction phase information including but not limited to steel structure hoisting, welding, correction and other operation phase identification data, and the construction phase identification data is processed by encoding conversion to generate a phase state encoding vector that can be used for model input; Based on the construction progress management system interface to access the construction phase information data stream, using API data grabbing method (parameters: interface authentication key, data refresh frequency 1 Hz, data buffer size 1024 KB) to realize real-time acquisition of steel structure construction site operation phase identification data; Further, through the phase identification data analysis algorithm (parameters: regular expression matching rule set, phase dictionary library version V2.3), automatic analysis and mapping of the original phase identification string is realized, and a structured phase code list is obtained; Further, the phase state quantization encoding method (parameters: total number of phase categories , one-to-one correspondence between one-hot encoding dimension and category) is used to realize the conversion of phase code list to standardized numerical code, and generate a phase binary encoding matrix; Further, the phase sequence timestamp synchronization algorithm (parameters: system master clock accuracy ms, allowed time difference threshold s) is used to realize the alignment of phase encoding matrix and construction global time base, and obtain time-synchronized phase state sequence data; Through the phase state vector construction formula, the quantification of model input features is realized: Wherein, is the one-hot encoding vector of the th phase, is the total number of phases, is the average stage state vector; Through the above-mentioned encoding conversion and time synchronization processing, the construction stage text identifier is converted into a numerical, time-aligned and structured stage state encoding vector, which realizes the provision of high-precision stage input features for the subsequent context-aware group behavior analysis module; For example, in a high-rise steel structure hoisting operation site, the construction progress management system refreshes the stage identifier signal every second, and the stage types include steel beam hoisting, temporary fixing, weld welding, weld detection, component correction, bolt fastening, painting, and acceptance, a total of 8 types. The "weld welding" stage is represented by one-hot encoding as , and a time series matrix is formed by sampling at 1 Hz. The system synchronization module aligns all stage vectors with the global clock, so that the maximum time difference does not exceed seconds. The average stage state vector calculated by the above formula can improve the stage recognition accuracy by about % in the model training set, and reduce the probability of group behavior misjudgment caused by time misalignment by about %, effectively supporting the context feature fusion analysis of subsequent abnormal behavior detection; S1.2: Collect environmental parameter data through the environmental sensor network deployed on the construction site, including temperature, humidity, wind speed, and light intensity. Perform sliding window mean filtering on the collected environmental parameter data to generate denoised environmental state time series data; The input condition is that multiple types of environmental sensor node networks have been deployed on the construction site, including temperature and humidity sensors, three-cup or ultrasonic anemometers, and light intensity sensors, and communicate with the data collection gateway through a wireless Mesh network; A multi-channel synchronous acquisition method (parameters: sampling frequency 1-10 Hz, time synchronization error ≤ s, buffer capacity 2048 samples) is used to realize parallel acquisition of temperature, humidity, wind speed, and light intensity raw signals; Further, through a signal calibration algorithm (parameters: temperature calibration coefficient K = , humidity linear compensation threshold %RH, wind speed zero point offset correction value m / s, light intensity gain coefficient ), the consistency of different types of sensors in output characteristics is adjusted to obtain environmental parameter sequences (t) with unified physical units; Further, through a sliding window mean filtering algorithm (parameters: window length seconds, window sliding step The window weighting coefficient adopts a triangular weighting mode), and the denoising and smoothing processing is performed on the calibrated environmental parameter sequence; The sliding mean calculation formula adopted is: wherein, is the filter output value at time , is the weighting coefficient of the th sample in the window, is the original parameter value of the corresponding delay sample; Further, through the timestamp interpolation alignment algorithm (parameters: interpolation method Lagrange third-order interpolation, maximum interpolation interval seconds), the data frame alignment of the multi-source environmental parameters on the same time axis is realized, and the environmental state sequence matrix with a unified time base is obtained; Through the sliding mean filtering and time alignment processing described above, the original environmental parameter signal is converted into noise-reduced and synchronized environmental state time sequence data, which realizes stable and reliable environmental input features for group behavior spatiotemporal modeling; For example, in a bridge steel structure welding operation monitoring, the temperature sensor sampling frequency is set to 1 Hz, and the original temperature signal contains ± ℃ high-frequency jitter in a certain period of time. After the sliding mean filtering processing of the triangular weighting window with a length of 5 seconds, the jitter amplitude is reduced to ± ℃. The humidity sensor near the air outlet of the machine room produces ± %RH random fluctuation due to air flow fluctuation. After linear compensation and sliding mean, the fluctuation is reduced to ± %RH. Two consecutive missing points (collection interval 2 seconds) appear in the wind speed signal due to temporary shielding, and the missing values are supplemented by third-order Lagrange interpolation, ensuring the integrity under the unified time axis. The environmental state time sequence matrix obtained after processing is used as the key input of the subsequent S1.4 step feature fusion, which reduces the false alarm rate of abnormal behavior detection by about % in the model training test. S1.3: Obtain the task type and personnel allocation information based on the construction task scheduling platform, perform multi-classification label coding processing on the task type data to generate a task type feature vector, and perform structured analysis on the personnel allocation information to generate a mapping relationship matrix of each operation role and task; S1.4: Perform time sequence alignment and data fusion processing on the construction phase state coding vector, environmental state time sequence data, task type feature vector, and personnel allocation relationship matrix to generate a context feature data frame marked with a unified timestamp; ​S1.5: Perform normalization transformation processing on the fused context feature data frame based on the preset data standardization strategy to generate a context feature vector that meets the model input requirements as the input basis for subsequent group behavior spatiotemporal model feature fusion.

[0016] The step S2: pre-processing the collected context data, including missing value filling, outlier removal and standardization conversion, to generate a unified format context feature vector. Specifically, it includes: S2.1: Detect missing values in multi-source context data such as construction phase information, environmental parameters, task types and personnel allocation information, and fill in the missing data based on time series interpolation algorithm to obtain complete and continuous context data stream; For the collected multi-source context raw data such as construction phase information, environmental parameters, task types and personnel allocation information, use the missing value detection algorithm (parameters: time stamp consistency threshold seconds, allowed null value proportion threshold %)to identify the location and number of missing samples in the continuous data stream; Further, through the multi-source data synchronization window division method (parameters: window length seconds, sliding step seconds), the missing position mapping of different data sources under the unified time reference system is realized, and the missing value index matrix is generated; Further, the time series interpolation algorithm (interpolation method: cubic spline interpolation, boundary processing strategy mirror extension, maximum interpolation interval seconds) is used to realize the numerical reconstruction of the missing value interval and obtain the continuous and smooth time series estimated value; Further, the interpolation residual error evaluation method is used to calculate the deviation between the interpolation result and the historical adjacent sample, and the interpolation residual error calculation formula is adopted: Where, is the residual value at time , is the true known observation value, is the interpolation estimated value; Further, the residual threshold detection (parameters: residual tolerance threshold is times the historical standard deviation of this feature) is used to remove unreliable interpolation and use the historical sliding window mean instead to ensure the stability and accuracy of the data stream; Through the above interpolation and correction processing method, the original multi-source context data is converted into time-continuous and non-missing context data stream, which provides complete and consistent data input for subsequent outlier detection and standardization processing; Exemplary, in the real-time monitoring of a certain steel beam hoisting construction site, the environmental parameter sampling frequency is Hz, the wind speed sensor has 4 seconds of continuous missing data due to temporary network packet loss, and the corresponding missing proportion is less than % threshold, the third-order spline interpolation is used to generate an interpolation value sequence, and 2 high deviation samples are removed through residual detection and replaced with the mean value of a sliding window length of 10 seconds, finally forming a complete and uninterrupted environmental parameter time sequence; there is a 1 second delay value when the construction stage identification signal is converted from hoisting to welding, which is filled by the mirror extension method to ensure the integrity and continuity of the stage feature sequence, and it is verified that the accuracy of the subsequent S2.2 step abnormal value detection is improved by about %, and the time sequence misjudgment rate caused by missing data is reduced by about % ; S2.2: Perform Z-score-based abnormal value identification algorithm on the filled context data, identify abnormal data points that exceed the normal fluctuation range, and use the sliding window median replacement method to correct the abnormal values to obtain a context data set with enhanced stability; S2.3: Perform standardization conversion processing on the corrected context data, and map each dimension of data to the [0, 1] interval based on the min-max normalization method to generate a standardized context feature matrix with uniform dimensions; S2.4: Perform feature encoding processing on the standardized context feature matrix, convert category type variables (such as construction stage, task type) into numerical feature vectors based on the one-hot encoding method, to obtain a context feature representation compatible with the input format of the deep learning model; S2.5: Perform feature concatenation and dimension alignment operations on the encoded context feature vector, fuse continuous and discrete features into a unified dimension context feature vector based on the feature concatenation algorithm, to generate a standardized context feature output suitable for group behavior spatio-temporal model fusion analysis.

[0017] The step S3: Obtain the positioning data and behavior trajectory information of the multi-role workers at the construction site, and construct a group behavior spatio-temporal model based on the spatio-temporal coordinates. As shown in Fig. 2 , specifically comprising: S3.1: Obtain the three-dimensional space coordinate sequence of each worker at the construction site through the UWB ultra-wideband positioning system, to obtain the continuous movement trajectory data of the personnel in the construction area; S3.2: Perform sliding window segmentation processing on the obtained three-dimensional space coordinate sequence to generate a behavior segment sequence based on the time window, which is used for subsequent trajectory feature extraction; The three-dimensional space coordinate sequence of the multi-role workers obtained by the UWB ultra-wideband positioning system is segmented by a sliding window based on the time domain (parameters: window length second, sliding step second), time slicing of continuous trajectory data is achieved, and preliminary time-blocked trajectory data sub-sequences are formed; Further, through the window-in-time index mapping algorithm (parameters: timestamp precision second), the sampling points in each time window are mapped to a set of ordered coordinate vector sets , wherein is a three-dimensional coordinate vector; Further, the time window smoothing interpolation resampling algorithm (interpolation method: linear interpolation, resampling frequency Hz) is used to achieve uniform sampling of trajectory points in the window and enhance trajectory continuity, obtaining a resampled behavior trajectory sub-sequence matrix ; Further, for the window resampling result, the time span and the number of samples of each window trajectory sub-sequence are calculated, and based on the condition ( is a preset minimum sampling frequency threshold) is used for effectiveness screening to eliminate window segments with insufficient sampling frequency and ensure trajectory data quality; Through the above sliding window segmentation and resampling processing method, the original continuous three-dimensional space trajectory data is converted into a time sequence consistent, uniformly sampled and quality controllable behavior segment sequence based on time window, providing a high confidence time sequence data basis for subsequent trajectory feature extraction; For example, in a certain bridge steel structure girder hoisting construction scene, the UWB system obtains the three-dimensional position coordinates of 10 workers at a sampling frequency of 20 Hz, and the total trajectory duration is seconds. Using a segmentation strategy with a window length of seconds and a sliding step of seconds, each window contains original sampling points. After applying timestamp mapping and linear interpolation resampling to 10 Hz, the number of trajectory points per window is fixed at , forming windowed behavior segments. After effectiveness screening, 12 segments with sampling rate reduced to 8 Hz due to temporary signal obstruction are removed, and finally high-quality behavior segment sequences are obtained, providing complete and uniformly sampled spatiotemporal input data for the S3.3 Kalman filter and velocity vector calculation stage; S3.3: Based on the segmented behavior segment sequence, a Kalman filter algorithm is used to smooth the trajectory data to eliminate positioning noise interference and improve trajectory continuity and accuracy; ​S3.4: Calculate the velocity vector and acceleration rate of change of the smoothed trajectory data to extract key dynamic behavior characteristics such as personnel movement direction, speed change and residence behavior, etc. Based on the Kalman filter smoothed trajectory data matrix , the three-dimensional difference method (parameters: time interval takes trajectory sampling period ) is used to calculate the displacement vector of adjacent sampling points, and the instantaneous velocity vector sequence is derived; Specifically, for any two adjacent position vectors and , the velocity calculation formula is used: Where, is the velocity vector of the k-th sampling point; Further, the same sequence three-point central difference algorithm (parameters: difference interval sampling points) is applied to the velocity vector sequence to calculate the acceleration rate vector sequence, and the smoothed acceleration vector is obtained, which describes the second-order change characteristics of the individual motion state; Further, the arctangent function is used to calculate the direction angle of the velocity vector in the horizontal and vertical planes, and the direction angle formula is: Where, is the horizontal azimuth angle, is the vertical inclination angle; Further, the low-speed threshold (parameters: threshold value takes the empirical walking speed m / s) is set for the velocity module length sequence , and the continuous low threshold time period detection method is used to identify the residence behavior, and the residence behavior time interval set is output; Further, the mean and variance of the acceleration rate sequence are calculated to form a dynamic behavior stability index, which is used to describe the rhythm change characteristics of the individual in the task execution process; Through the chain calculation of the above velocity, acceleration, direction and residence characteristics, the low-noise trajectory data is converted into a multi-dimensional dynamic behavior feature vector, and the movement pattern of the multi-role personnel in the construction site is fully quantized; For example, in a certain bridge steel structure main beam butt joint construction scene, the UWB system sampling frequency is Hz, and the smoothed trajectory point time interval = seconds. For the 5-second window trajectory sequence processed by Kalman filter, the velocity vector is calculated according to the formula, and the mean velocity module length is about m / s, the direction angle change continuously distributed in ° to °, the average of acceleration m / s², the variance m 2 / s 4 . In the interval of 3 to 3.8 seconds, the speed is lower than m / s and the duration is more than 0.5 seconds, which is identified as the residence behavior. The dynamic feature extraction result is used for subsequent construction of group behavior spatiotemporal atlas, which can obviously distinguish normal cooperative walking from local stagnation mode caused by obstacle avoidance, and provide high-resolution input features for abnormal behavior discrimination. The final verification shows that the moving direction recognition accuracy of the feature extraction method in this scene is %, the residence behavior detection recall rate is more than %, which effectively supports the dynamic analysis of group behavior with context awareness; S3.5: Based on the extracted dynamic behavior features and corresponding timestamp information, construct a group behavior spatiotemporal atlas to form a group behavior modeling basis with spatial relationship and time evolution; S3.6: Model the constructed group behavior spatiotemporal atlas with graph neural network to generate group behavior representation vector containing role attributes and interaction relationship, which is used for subsequent fusion analysis with context features.

[0018] The step S4: fuse the context feature vector and the group behavior spatiotemporal model to form a context-enhanced group behavior representation dataset. As shown in Fig. 3 , specifically includes: S4.1: Perform time alignment operation on the context feature vector, align the context feature and group behavior trajectory data on the unified time axis based on the timestamp sequence, to eliminate the time sequence deviation of multi-source data collection, and obtain the time-synchronized context-behavior correlation data; The linear interpolation alignment algorithm based on global time axis (parameters: interpolation method is set to piecewise linear interpolation, maximum allowed time deviation seconds) is adopted to realize the accurate alignment of different data source sampling times for the context feature vector set obtained by S3 step and the node timestamp sequence of group behavior spatiotemporal atlas; Further, the time index mapping matrix construction method (matrix element represents the time index of the th trajectory sampling point corresponding to the th context vector) is adopted to realize the position mapping relationship calculation of the context vector on the unified time axis; Further, the least mean square error time offset correction algorithm (iteration step seconds, the iteration termination condition is that the time alignment residual is less than seconds), the residual time drift of multi-source data is corrected to generate a high-precision synchronous timestamp sequence; Further, the index sequence under the unified sampling period (taking the least common multiple of the context sampling period and the trajectory sampling period) is established by the timestamp registration function, so that the data from different sources have corresponding matching context features and behavior trajectory features at each unified time point; Through the above interpolation, index mapping, offset correction and registration processing, the original time-synchronous context feature vector and trajectory data are converted into time-synchronous context-behavior correlation data, realizing the accurate matching of time consistency and feature correspondence in the subsequent S4.2 feature fusion link; For example, in a certain high-rise steel structure hoisting scene, the context feature sampling frequency is Hz, the group behavior trajectory sampling frequency is Hz, and the maximum initial collection time deviation is about seconds. Using piecewise linear interpolation and index mapping method, each 0.5 second sampling context feature vector is mapped to the corresponding 5 trajectory sampling point index position. Through the least mean square error time offset correction, the global time residual is reduced from the initial seconds to seconds. Under the registration time axis of the unified sampling period seconds, the corresponding context features and group behavior features are obtained at each time point to form a complete time-synchronous context-behavior correlation matrix, providing accurate alignment input data for subsequent multi-modal feature fusion; S4.2: Based on the time-synchronous context-behavior correlation data, the multi-modal feature splicing method is used to fuse the context feature vector and the spatial coordinate feature of the group behavior trajectory to generate a preliminary fused context-enhanced behavior feature matrix; Based on the context-behavior correlation data time-aligned by S4.1 step, the multi-modal feature splicing method (parameters: the feature splicing dimension order is set as context feature first and behavior trajectory spatial feature second) is used to realize the feature-level fusion processing of heterogeneous data; Further, through the feature alignment index mapping table, the context feature vector at each time-synchronous point is matched with the corresponding group behavior trajectory spatial coordinate feature , wherein is an m-dimensional normalized context feature vector, is an n-dimensional smooth spatial and dynamic feature combination vector; Further, the vector splicing operation formula is used: wherein, represents the concatenation operation of feature vectors in dimension, generating a multimodal fusion feature vector with length ; Further, the concatenated is subjected to feature rescaling normalization processing (parameters: normalization method is Z-score standardization, mean and variance are calculated based on global time series), to eliminate the modeling bias caused by the numerical range and distribution difference of different modalities, to obtain a normalized fusion vector with mean 0 and standard deviation 1; Further, a weighted fusion enhancement method (parameters: context feature weight =0.4, behavior trajectory feature weight =0.6, constraint condition ) is adopted, and the modal contribution degree adjustment based on weight proportion is realized through the formula: to generate a weighted fusion context-enhanced behavior feature matrix ; Through the chain processing mode of the above multi-modal splicing, normalization and weighted fusion, the time-synchronized context features and group behavior spatial coordinates and dynamic features are jointly encoded to transform into a preliminary fusion context-enhanced behavior feature matrix that can be directly input into a graph neural network, realizing the expected technical effects of modal complementation and feature expression richness; For example, in a certain bridge steel structure main beam installation construction scene, the context feature vector dimension m of each time-synchronized point is 12, including 6-dimensional construction phase one-hot encoding features, 4-dimensional task type one-hot encoding features and 2-dimensional environmental parameters; the group behavior trajectory feature vector dimension n is 15, including 3-dimensional position coordinates, 3-dimensional velocity, 3-dimensional acceleration, 2-dimensional direction angle, 1-dimensional residence state and 3-dimensional dynamic stability indicators. Based on the index mapping relationship, the and of each 0.1 second sampling moment are spliced to obtain a 27-dimensional fusion vector . After normalization, weighted fusion is applied with weight =0.4 and =0.6 to generate a fusion matrix Each row vector of the fusion matrix reflects the comprehensive features of the context state and real-time behavior. In this embodiment, the fusion matrix is used as the input of the graph neural network of S4.3, which makes the model improve the group abnormal behavior recognition accuracy by 7.8% in the test set environment compared with using only behavior features, verifying the effectiveness of multi-modal fusion in improving context awareness and abnormal detection performance; S4.3: Apply a graph neural network modeling method to the preliminarily fused context-enhanced behavior feature matrix to construct an interaction relationship graph structure between group members, wherein the nodes represent individual behavior trajectories, and the edge weights represent context-driven collaboration correlation strength, to obtain a context-aware group interaction graph representation; Based on the weighted fused context-enhanced behavior feature matrix obtained in step S4.2 , a graph neural network modeling method (parameters: adjacency matrix normalization strategy is symmetric normalization, node feature input dimension is ) is used to realize the structured expression of the interaction relationship between group members; Further, by constructing a context-driven collaboration relationship matrix between nodes , a correlation weight calculation method based on feature similarity (parameters: similarity measure is cosine similarity, minimum threshold ) is used to obtain an unweighted original adjacency matrix , wherein the matrix element represents the collaboration correlation between the th node and the th node; Further, a context weighting coefficient modulation method (parameters: construction phase weight , task type weight , and environment state weight ) is used to correct the weight values of each element in to generate a context-driven weighted adjacency matrix , realizing the fusion of collaboration strength and scene semantics; Further, symmetric normalization processing is performed on to obtain a normalized adjacency matrix ; Further, based on the normalized adjacency matrix and the feature matrix , a graph convolution network (parameters: number of convolution layers is 2, number of hidden units is 64, and activation function is ReLU) is used to perform node feature aggregation operation; Further, the node representation is aggregated into a graph-level representation vector by a global graph embedding generation module (parameters: pooling method is mean pooling), to obtain a context-aware group interaction graph global feature vector , which is used for subsequent spatio-temporal attention mechanism calculation; Through the above graph neural network construction and weighting processing, the preliminarily fused feature matrix is mapped to a context-aware group interaction graph representation that combines individual behavior details and cooperative mode structure, realizing unified modeling of collaboration relationship and scene semantics; Exemplarily, in the high wind speed construction scene of hoisting the main girder of a certain bridge steel structure, the number of nodes of the preliminary fusion feature matrix is 12 (corresponding to 12 workers), and the feature dimension of each node is 27 dimensions. The cosine similarity is used to calculate the original cooperation relationship matrix , and the connections with similarity less than 0.2 are removed to obtain an adjacency matrix with a sparsity of about 46%. The context weighting is performed on the basis of the construction stage, task type and environment state weight to generate , and the average weight is improved by about 15%. The degree matrix is calculated and symmetrically normalized, and after inputting into a two-layer GCN structure, the node embedding dimension is mapped from 27 dimensions to 64 dimensions, and the global graph representation is obtained by mean pooling . In subsequent abnormal behavior identification, the representation vector significantly improves the detection rate of group cooperation imbalance behavior caused by sudden change of wind speed, and improves the abnormal identification accuracy by about 9.3% compared with the model without introducing context weighting; S4.4: Based on the context-aware group interaction graph representation, the spatio-temporal attention mechanism is calculated to weight and aggregate the group behavior features in the time dimension and the space dimension respectively, so as to highlight the representation strength of the key behavior mode under the current context condition, and generate a spatio-temporally enhanced context fusion behavior representation vector; S4.5: The spatio-temporally enhanced context fusion behavior representation vector is normalized and packaged into a context-enhanced group behavior representation dataset according to a preset data format, so as to serve as the training and reasoning input of the context-aware behavior anomaly detection model.

[0019] The step S5: based on the deep learning algorithm, the context-aware behavior anomaly detection model is trained by using the context-enhanced group behavior representation dataset. Specifically, it comprises: S5.1: The context-enhanced group behavior representation dataset is divided into data, and the sliding window method is used to divide the continuous behavior sequence into time-aligned training sample segments to construct a training dataset suitable for time series modeling; S5.2: Based on the Transformer architecture, the context-aware behavior anomaly detection model structure is constructed, and the multi-head self-attention mechanism calculation is performed on the input behavior feature sequence to extract the long-time dependence relationship in the group behavior; Based on the input of the context-enhanced group behavior representation dataset, the Transformer encoding structure (parameters: the number of encoding layers is 6, the number of multi-head attention mechanism heads in each layer is 8, the hidden layer dimension is 512, the feedforward network dimension is 2048, and the residual connection and LayerNorm standardization are both enabled) is adopted to realize the deep time series modeling capability of the group behavior feature sequence; Furthermore, the position information at each time step is calculated by the position encoding generation module (parameter: the position encoding type is sine-cosine function encoding, and the encoding dimension is consistent with the input feature dimension), and added element by element to the input feature sequence to explicitly preserve the temporal order information; Furthermore, a weighted correlation calculation is performed on the input behavioral feature sequence based on a multi-head self-attention mechanism to calculate the attention weight matrix; Furthermore, the input features are transformed by trainable linear transformation matrices. , , Generate multiple groups , , Attention outputs are computed independently within each attention head, and concatenation is performed along the head count dimension, followed by a linear mapping matrix. Map the spliced ​​result back to the original dimension; Furthermore, a feedforward fully connected sub-network (parameters: two-layer structure, activation function is ReLU, dropout ratio is 0.1) is used to perform nonlinear transformation on the attention output step by step to enhance the expressive power of features. Residual connections and LayerNorm are introduced to normalize the input and output of each sub-layer element by element to ensure the numerical stability of the training process and the effectiveness of gradient propagation. Through the above Transformer-based multi-head self-attention modeling, the original feature sequence is encoded into a deep temporal embedding vector that can capture long-term dependencies and global interaction patterns, providing a highly discriminative structured representation for subsequent context feature fusion and anomaly detection; For example, in a high-altitude collaborative operation scenario for hoisting a steel bridge cable saddle, the input is a context-enhanced group behavior representation dataset encapsulated using S4.5, with a feature sequence length of 100 time steps and a feature vector dimension of 256 for each step. An 8-head attention mechanism is employed, with each head having a feature dimension of 32. After positional encoding and stacking, a 100×256 input matrix is ​​formed. A linear mapping is used to generate Q, K, and V matrices, each with a dimension of 100×32. The attention weight matrix, with a size of 100×100, is calculated, normalized using softmax, and multiplied by the V matrix to obtain the time-weighted behavioral feature output. The multi-head results are concatenated to a 100×256 matrix, and then processed... Mapped back to 256 dimensions, the result is added to the original input residual and normalized using LayerNorm. After ReLU activation via a feedforward network and two linear transformations, the output dimension remains 100×256, which is then input into S5.3 to perform context-aware interactive fusion modeling. In test set evaluation, this Transformer encoding layer improved the long-term dependency feature capture rate of group anomaly patterns by approximately 12.4%, and the accuracy of anomaly behavior recognition was improved by 8.1% compared to the traditional bidirectional LSTM model. S5.3: Embedding the context feature vector as auxiliary input into the middle layer of the Transformer model to model the interaction between context information and behavior trajectory, so as to enhance the model's perception ability of construction phase, environment state and task type; Based on the group behavior deep time sequence embedding vector sequence obtained through the S5.2 step Transformer encoding, a context embedding fusion method (parameters: context embedding dimension is 64, fusion method is cross-modal attention mechanism) is used to realize the middle layer fusion of context feature vector and group behavior embedding; Further, the input context feature vector is mapped to the feature dimension consistent with the behavior sequence embedding through the context feature linear transformation module (parameters: weight matrix size is , and the activation function is ReLU), so as to ensure the operation compatibility between multi-modal features, and obtain the context embedding matrix ; Further, the cross-modal attention mechanism is used to calculate the interaction weight between the context feature and the behavior embedding; Further, the cross-modal attention output result and the original behavior embedding are element-wise weighted and added, and a residual connection and a LayerNorm normalization operation (parameters: normalization epsilon value is ) are introduced to realize stable training of the fusion feature; Further, the fusion feature is nonlinearly transformed and refined through the fusion feedforward network module (parameters: two-layer fully connected network, hidden dimension is 512, activation function is GELU, and Dropout ratio is 0.1) to generate the context-enhanced behavior representation sequence ; Through the above context middle layer fusion processing method, the Transformer behavior embedding is closely combined with the context information such as environment, task and phase, and is converted into a high representation feature sequence with context perception ability, so as to realize the improvement of scene adaptability and recognition accuracy of the model in abnormal behavior detection; For example, in a bridge steel structure segment hoisting cooperation construction scene, the input context feature vector has a dimension of 20, including 6-dimensional construction phase one-hot encoding, 4-dimensional task type one-hot encoding and 10-dimensional environment parameters. The feature vector is mapped to 256 dimensions through a linear transformation matrix (size ) and the behavior embedding sequence (time step 100, feature dimension 256) obtained by Transformer encoding is calculated by cross-modal attention at each step. The behavior embedding is taken as the query , the context embedding is taken as the key and value , , and the feature dimension is is 32, the attention weight matrix size is 100x1, and after softmax normalization, it is multiplied by to obtain a 100x256 context attention feature, which is input into a two-layer GELU activated feedforward network after residual connection and LayerNorm, and the output sequence size remains 100x256. Experimental results show that under this context intermediate layer fusion strategy, the model's accuracy in identifying abnormal cooperative behavior in two high-risk context environments, wind speed > 12 m / s and insufficient night lighting, is improved by about 11.6%, and the false positive rate caused by context changes is effectively reduced, realizing the adaptive optimization of the abnormal behavior detection model for complex construction scenes; S5.4: Use cross-entropy loss function and contrast learning strategy to optimize model parameters, perform negative sampling contrast learning calculation on the behavior embedding features output by the model, and improve the model's ability to distinguish between normal and abnormal behavior patterns; S5.5: Adjust the model output threshold based on the behavior anomaly score distribution of the validation set, and perform Sigmoid normalization processing on the behavior anomaly score output by the model to generate a standardized abnormal probability output.

[0020] The step S6: input the real-time collected context feature vector and the current group behavior trajectory into the trained behavior anomaly detection model to generate a behavior anomaly score result. Specifically, it includes: S6.1: Perform format checking and dimension alignment processing on the real-time collected context feature vector to adapt to the input interface requirements of the behavior anomaly detection model; S6.2: Perform spatio-temporal alignment and coordinate normalization processing on the current group behavior trajectory of the multi-role workers in the construction site to generate standardized group behavior sequence data; For the current group behavior trajectory data of the multi-role workers in the construction site, a unified clock synchronization mechanism (parameters: NTP network time protocol synchronization period of seconds, maximum time difference of milliseconds) is used to achieve consistency of time references of different collection nodes and obtain trajectory original sequences with global time consistency; Further, linear interpolation resampling algorithm (parameters: time step of milliseconds, interpolation method of cubic spline) is used to perform resampling processing on the time axis of different personnel trajectory data, realize complete alignment of trajectory data at each time step, and generate an aligned trajectory matrix with uniform sampling frequency; Further, for different coordinate system data in the trajectory space coordinates, a rigid body transformation registration algorithm (parameters: rotation matrix and translation vector estimated based on least squares method, registration error threshold of cm), to realize the unified conversion of multi-coordinate system data in the global reference coordinate system, and obtain trajectory point cloud data with consistent spatial positions; Further, based on the normalization processing formula: wherein, is the original value of the spatial coordinate, and are the minimum and maximum values of the spatial coordinate, respectively, and the three-dimensional coordinate is normalized in the interval [0, 1] to eliminate the differences of different ranging ranges and units, and obtain the dimensionless standard spatial features; Further, a sliding window smoothing algorithm (parameters: window size is time steps, weight distribution is Gaussian type, standard deviation ) is used to process the normalized trajectory sequence dimension by dimension, to reduce the interference of instantaneous measurement jitter on subsequent feature extraction, and to improve the trajectory continuity and stability; Through the above synchronization, resampling, registration, normalization and smoothing processing, the multi-role trajectory data of different collection sources are converted into standardized group behavior sequence data with consistent structure, unified time, consistent space and dimensionless, to realize the high-robustness input data basis for subsequent local pattern feature extraction and multi-modal fusion; For example, in the welding cooperation scene of a bridge steel structure, three UWB positioning base stations are deployed on the construction site, with a positioning accuracy of cm, a sampling frequency of Hz, and all base stations are time-synchronized to an error of less than ms through the NTP protocol. There are workers, and the positioning data format is three-dimensional Cartesian coordinates, and some nodes use local coordinate systems. First, perform three times of spline interpolation to unify all trajectories to a sampling rate of 10 Hz. After alignment, the trajectory matrix dimension is (time steps) x (personnel) x (coordinate dimensions). Through the rigid transformation registration algorithm, the local coordinate system data is converted to the global construction reference system, and the registration error is controlled within cm. During normalization processing, = meters, = meters, and other parameters are substituted into the formula to obtain the dimensionless value of each coordinate dimension. Finally, through the Gaussian weighted sliding window smoothing, the error variance caused by trajectory noise is reduced by about %, effectively improving the sensitivity and accuracy of subsequent abnormal behavior detection; S6.3: Based on the attention mechanism, the standardized group behavior sequence data is locally pattern feature extracted to obtain a group behavior local feature vector; S6.4: The context feature vector and the group behavior local feature vector are multi-modal feature spliced to form a context-enhanced group behavior fusion feature tensor; S6.5: The context-enhanced group behavior fusion feature tensor is input into the trained behavior anomaly detection neural network model, and the group behavior anomaly score result is generated through the full connection layer and the Softmax activation function.

[0021] The step S7: judging whether the behavior anomaly score exceeds the preset threshold value, if exceeding, triggering an abnormal behavior recognition signal. Specifically, it includes: S7.1: The behavior anomaly score result is normalized to eliminate the deviation caused by uneven distribution of scores on threshold judgment; S7.2: Based on the historical behavior score data and the context feature distribution, a dynamic threshold calculation algorithm is used to generate an adaptive anomaly threshold value matched with the current construction stage and task type.

[0022] S7.3: The normalized current behavior anomaly score is compared with the adaptive anomaly threshold value to determine whether there is a group behavior anomaly; Based on the processing target of S7.3, the input normalized group behavior anomaly score is compared with the dynamically calculated adaptive anomaly threshold value to determine whether there is a group behavior anomaly state in the current construction scene; The numerical comparison method (parameters: comparison accuracy is set to ) is used to realize the one-by-one comparison function of the normalized anomaly score value and the adaptive threshold value; Further, through the difference calculation method (parameters: the decimal place accuracy is set to bit), the difference between the normalized anomaly score and the adaptive threshold value is calculated, and the formula is: Among them, is the normalized current group behavior anomaly score, is the adaptive anomaly threshold value obtained according to S7.2; Further, the sign function judgment algorithm is used to determine the sign of , if the function output is positive, it is marked as an abnormal candidate; Further, combined with the logic threshold filtering mechanism (parameters: the hysteresis interval width is ), a buffer zone is introduced in the abnormal and normal state edge value region to reduce the misjudgment probability caused by value fluctuation; Through the above comparison and logical judgment method, the model output score result of the previous step is converted into binary judgment data of whether there is a group behavior anomaly, realizing the judgment basis of subsequent abnormal signal triggering; For example, in the high-altitude steel structure welding cooperation scene, the normalized group behavior anomaly score obtained by real-time detection is , the adaptive abnormal threshold value corresponding to the construction stage is calculated as , the comparison precision is set as . Through the difference calculation formula = , the sign function output is positive. Since the difference is higher than the hysteresis interval width , the logical threshold filtering judgment is an obvious abnormal state, thereby providing a clear triggering condition for S7.4 to trigger the abnormal behavior recognition signal. In the continuous 200 times of measurement in this scene, the response false alarm rate of the edge value fluctuation of this method is reduced to % or less, significantly improving the stability and field applicability of abnormal identification; S7.4: If the current behavior anomaly score exceeds the adaptive abnormal threshold value, generate a behavior anomaly recognition signal, and mark the signal as a pre-alarm state; S7.5: Based on the triggering state of the behavior anomaly recognition signal, combine the construction stage information and task type information in the context feature vector to generate an abnormal event record with a context label, which is called by the subsequent pre-alarm information generation module.

[0023] The step S8: based on the triggered abnormal behavior recognition signal, generate targeted pre-alarm information and response suggestions combined with the current context features, and push to the terminal of the site manager. Specifically, it includes: S8.1: classify and analyze the triggered abnormal behavior recognition signal, extract the abnormal type label, occurrence timestamp, involved operation personnel number and spatial coordinates, to generate structured abnormal event description information; S8.2: Based on the construction stage information, environmental parameters and task type in the current context feature vector, perform multi-dimensional context matching analysis to identify the environmental constraints and task execution background associated with the abnormal event; S8.3: Use the rule engine and historical abnormal response knowledge base to generate preliminary pre-alarm level division and response suggestion templates combined with the structured abnormal event description information and multi-dimensional context matching results; The input is the structured abnormal event description information and multi-dimensional context matching analysis results generated after S8.1 and S8.2 processing, including abnormal type label, occurrence timestamp, personnel number, spatial coordinates, construction phase code, environmental parameters, and task type feature vector; The rule engine-based reasoning mechanism (parameter: rule matching priority strategy is a conflict resolution method based on weight) is used to realize the rapid matching of abnormal event types and historical response modes; Further, the rule analysis function is used to perform pattern matching on the input abnormal type label and the pre-defined rule conditions in the knowledge base (parameter: semantic matching tolerance threshold is ), to realize accurate classification of abnormal categories; Further, the historical abnormal response knowledge base retrieval function (parameter: similarity calculation method is a multi-dimensional cosine similarity based on the fusion of environmental parameters, task type, and construction phase information) is called to calculate the similarity score between the current event and each historical record; The multi-dimensional similarity is calculated using the following formula: wherein, is the weight of the th feature, is the cosine similarity value of the feature; Further, the top most similar historical event templates are selected based on the similarity score, and the preliminary warning level is generated according to the historical disposition level distribution and the current context constraint conditions ; The warning level score is calculated using the following weighted summation formula: wherein, is the weight of the th similar historical event, is the corresponding warning level quantization value; Further, according to the calculated value and the preset interval threshold, the event warning level is divided into three levels: high, medium, and low, and the corresponding response suggestion template is generated. The content of the response suggestion template includes the disposition responsible person, the execution time limit, the required safety protection measures, and the backup scheme options; Through the rule engine and historical knowledge base fusion processing method, the current abnormality is associated with historical experience to generate a response suggestion template with a level label, realizing the construction of a fast and targeted warning scheme; For example, in the high-altitude steel structure welding scenario, the input abnormal event type is "multiple people synchronous displacement anomaly", the construction phase code is 3 (corresponding to the steel structure hoisting stage), and the environmental parameter is temperature ℃, wind speed m / s, the task type is cross-beam collaborative welding. Based on the rule engine detection, the event type is accurately matched with the historical event template "multi-person collaborative high-altitude displacement conflict", and the semantic matching degree is . The multi-dimensional similarity score of the computing environment and the task = , the historical events with a similarity score of are screened out, and the weight and the grade value are weighted to obtain = , which corresponds to the medium level warning level after interval mapping. The response suggestion template includes: the person in charge is designated as the site safety officer and the welding group leader, the disposal time limit is minutes, the execution measures include immediately stopping the welding operation, adjusting the personnel station, enabling the anti-falling safety net, and replacing the unstable post by the standby personnel to ensure the construction safety closed loop disposal. The target output is a structured response template containing the warning level "medium", four suggestion contents, and two constraint conditions; S8.4: Perform semantic enhancement processing on the generated preliminary warning level and response suggestion template, and embed specific parameter values and personnel role information in the current context features to generate structured warning information with context association; S8.5: Based on the construction site communication protocol, encapsulate the structured warning information into a warning message package conforming to the receiving format of the management personnel terminal, and push it to the corresponding site management personnel terminal device through the wireless communication network.

[0024] The step S9: periodically collect model prediction results and artificial review feedback information, and perform online incremental training and parameter optimization on the behavior anomaly detection model. Specifically, it includes: S9.1: Based on the abnormal score results output by the behavior anomaly detection model and the feedback information of the site management personnel, construct an incremental training sample set to obtain the deviation data between the model prediction error and the actual judgment; S9.2: Perform feature consistency calibration processing on the context feature vectors and group behavior trajectory data in the incremental training sample set to eliminate the data distribution deviation problem caused by sensor drift or environmental changes; S9.3: Use online incremental learning algorithm to fine-tune the feature extraction layer of the behavior anomaly detection model, update the spatio-temporal perception ability of the model based on the calibrated incremental samples, and improve its adaptability to dynamic changes of group behavior; S9.4: Perform context-sensitive weight adjustment mechanism on the classification decision layer of the model, combine the current construction stage and task type information, and optimize the abnormal recognition sensitivity and false alarm control ability of the model in different scenarios; S9.5: Based on the optimized model performance indicators, generate a model update version and deploy it to the edge computing nodes at the construction site to realize the continuous improvement of the model's closed-loop iterative optimization and real-time early warning capabilities.

[0025] The technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after these changes or replacements will fall within the protection scope of the present application.

[0026] The above description is only for the preferred embodiments of the present application and is not intended to limit the present application; for those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present application shall be included within the protection scope of the present application.

Claims

1. A method for real-time identification and early warning of safety risks in steel structure construction, characterized in that, Includes the following steps: S1: Real-time acquisition of multi-source context data at the steel structure construction site, including construction stage information, environmental parameters, task type, and personnel allocation information; S2: Preprocess the collected context data to generate context feature vectors in a unified format; S3: Obtain the location data and behavioral trajectory information of multi-role workers at the construction site, and construct a spatiotemporal model of group behavior based on spatiotemporal coordinates; S4: The context feature vectors are fused with the spatiotemporal model of group behavior to form a context-enhanced group behavior representation dataset; S5: Based on deep learning algorithms, train a context-aware behavior anomaly detection model using the aforementioned group behavior representation dataset; S6: Input the real-time collected context feature vectors and the current group behavior trajectory into the trained behavior anomaly detection model to generate behavior anomaly scoring results; S7: Determine whether the abnormal behavior score exceeds a preset threshold. If it does, trigger an abnormal behavior recognition signal. S8: Based on the triggered abnormal behavior recognition signal, combined with the current context features, generate targeted early warning information and response suggestions, and push them to the on-site management personnel terminal; S9: Periodically collect model prediction results and manual review feedback information to perform online incremental training and parameter optimization of the behavior anomaly detection model.

2. The method for real-time identification and early warning of safety risks in steel structure construction according to claim 1, characterized in that, Step S1 specifically includes: The current construction stage information is obtained based on the interface of the construction progress management system. The construction stage information includes work stage identification data. The construction stage identification data is encoded and converted to generate a construction stage status encoding vector. Environmental parameter data is collected by an environmental sensor network deployed at the construction site. The collected environmental parameter data is then subjected to sliding window mean filtering to generate noise-reduced environmental state time series data. Based on the construction task scheduling platform, task type and personnel allocation information are obtained. Multi-class label encoding is performed on the task type data to generate task type feature vectors. The personnel allocation information is then structured and parsed to generate a mapping matrix between each work role and task. The construction phase state coding vector, the environmental state time series data, the task type feature vector, and the personnel allocation relationship matrix are subjected to time-series alignment and data fusion processing to generate a context feature data frame with a unified timestamp. The context feature data frame is normalized based on a preset data standardization strategy to generate a context feature vector.

3. The method for real-time identification and early warning of safety risks in steel structure construction according to claim 2, characterized in that, In step S1, environmental parameters are collected, including temperature, humidity, wind speed and light intensity. The sampling frequency is 1-10Hz. After multi-channel synchronous acquisition and physical quantity calibration, the multi-source parameter data is aligned and normalized by mean filtering with a 5-second sliding window and Lagrange interpolation for noise reduction.

4. The method for real-time identification and early warning of safety risks in steel structure construction according to claim 1, characterized in that, Step S2 specifically includes: Missing values ​​are detected in multi-source context data, and missing data are filled in based on time series interpolation algorithms to obtain a complete and continuous context data stream; An outlier identification algorithm is executed on the context data stream to identify outlier data points that exceed the normal fluctuation range, and the outlier is corrected by the sliding window midpoint substitution method to obtain a context dataset with enhanced stability. The stability-enhanced context dataset is subjected to a normalization transformation to map the data of each dimension to a unified interval, thereby generating a normalized context feature matrix; The standardized context feature matrix is ​​subjected to feature encoding processing. Based on the one-hot encoding method, the categorical variables are transformed into numerical feature vectors to obtain the context feature representation. The encoded context feature vector is subjected to feature concatenation and dimension alignment operations. Based on the feature concatenation algorithm, continuous and discrete features are fused into a context feature vector of uniform dimension, generating a standardized context feature output.

5. The method for real-time identification and early warning of safety risks in steel structure construction according to claim 1, characterized in that, Step S3 specifically includes: The three-dimensional spatial coordinate sequence of each worker at the construction site is obtained through the UWB ultra-wideband positioning system, and the continuous movement trajectory data of the workers in the construction area is obtained. The three-dimensional spatial coordinate sequence is subjected to sliding window segmentation to generate a sequence of behavioral segments based on a time window; Based on the aforementioned behavior segment sequence, the trajectory data is smoothed using the Kalman filter algorithm; Calculate the velocity vector and rate of change of acceleration of the smoothed trajectory data, and extract key dynamic behavior features; Based on the key dynamic behavioral features and corresponding timestamp information, a spatiotemporal map of group behavior is constructed, forming a foundation for group behavior modeling with spatial relationships and temporal evolution. A graph neural network is used to model the spatiotemporal map of the group behavior to generate a group behavior representation vector.

6. The method for real-time identification and early warning of safety risks in steel structure construction according to claim 5, characterized in that, In step S3, trajectory processing based on 5-second window segmentation and Kalman filtering is adopted, and the velocity, acceleration, movement direction, dwell state and dynamic stability index are calculated through three-dimensional difference. Based on the extracted features and timestamps, a spatiotemporal map of group behavior with role attributes and interaction relationships is constructed.

7. The method for real-time identification and early warning of safety risks in steel structure construction according to claim 1, characterized in that, Step S4 specifically includes: Perform time alignment on the context feature vector, align the context features with the group behavior trajectory data on a unified time axis based on the timestamp sequence, and obtain context-behavior association data; Based on the aforementioned context-behavior association data, a multimodal feature concatenation method is used to fuse the context feature vector with the spatial coordinate features of the group's behavioral trajectory to generate a context-enhanced behavioral feature matrix. Applying graph neural network modeling methods to the context-enhanced behavioral feature matrix, an interaction graph structure among group members is constructed to obtain a group interaction graph representation; Based on the group interaction graph representation, a spatiotemporal attention mechanism is executed to calculate and aggregate group behavior features in the time and space dimensions respectively, generating a context-fused behavior representation vector. The context-fused behavior representation vectors are normalized and packaged into a context-enhanced group behavior representation dataset according to a preset data format.

8. The method for real-time identification and early warning of safety risks in steel structure construction according to claim 7, characterized in that, In step S4, multimodal feature splicing and context / behavioral feature weighted fusion are used. The feature splicing dimension order is set as context features first, followed by behavioral trajectory space features. The fused features are modeled by a graph neural network. The node input is context-enhanced behavioral features. The edge weight set reflects the context-driven collaboration strength of construction stage, task type, and environmental state. After normalization, a global group interaction feature vector is generated.

9. The method for real-time identification and early warning of safety risks in steel structure construction according to claim 1, characterized in that, Step S5 specifically includes: The context-enhanced group behavior representation dataset is partitioned by using a sliding window method to divide continuous behavior sequences into time-aligned training sample segments to construct a training dataset. A context-aware behavior anomaly detection model is built based on the Transformer architecture. A multi-head self-attention mechanism is performed on the input behavior feature sequence to extract long-term dependencies in group behavior. The context feature vector is embedded as an auxiliary input into the intermediate layer of the Transformer model to fuse and model the interaction between context information and behavioral trajectory. The model parameters are jointly optimized by the cross-entropy loss function and the contrastive learning strategy, and negative sampling contrastive learning computation is performed on the behavioral embedding features output by the model. The model output threshold is adjusted based on the distribution of behavioral anomaly scores on the validation set, and the behavioral anomaly scores output by the model are normalized using Sigmoid to generate standardized anomaly probability output.

10. A method for real-time identification and early warning of safety risks in steel structure construction according to claim 9, characterized in that, In step S5, the Transformer-encoded multi-head attention mechanism has 6 encoding layers, 8 heads per layer, a hidden layer dimension of 512, a feedforward network dimension of 2048, and residual connections and LayerNorm normalization are both enabled.

Images