Online management method and system for engineering project

Abstract

The invention provides an on-line management method and system for an engineering project, and the method comprises the steps: obtaining a continuous video stream of a monitoring region, and collecting environment data through an environment sensing network; performing timestamp alignment and space coordinate binding through a space-time synchronization unit to obtain a space-time synchronization data stream; inputting the time-space synchronization data stream into equipment for edge calculation to obtain an initial contour data set, and generating a dynamic deformation feature vector through continuous inter-frame displacement vector calculation and deformation rate analysis; inputting a pre-trained machine learning classifier, and analyzing to obtain a risk cause classification result; and performing grading response matching on the risk cause classification result, triggering a grading alarm mechanism to execute a corresponding instruction, starting environment linkage control equipment when the risk is caused by environmental factors, and generating a security event report containing a behavior mode label when the risk is caused by human factors, and pushing the security event report to the terminal. According to the invention, real-time monitoring of material deformation of construction safety risks, intelligent identification of risk causes and triggering of graded response are realized.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an online management method and system for an engineering project. Background Art

[0002] Currently, the safe management of stacked materials such as steel, cement, sand and gravel at construction sites faces many technical challenges. The traditional management model mainly relies on regular manual inspections and post-review of fixed surveillance videos. This passive management method has obvious technical defects.

[0003] In terms of material deformation monitoring, the low frequency of manual inspections cannot capture instantaneous pile deformation. Sudden situations such as sand pile collapse caused by strong winds are often difficult to detect in time. In terms of environmental parameter monitoring, the existing system fails to achieve spatiotemporal synchronization of video surveillance data and environmental data, resulting in the inability to accurately quantify the specific impact of environmental factors such as wind speed and humidity on material pile deformation. In terms of risk identification, existing technologies lack the ability to intelligently classify risk causes and cannot distinguish between safety hazards caused by environmental factors and human factors. In terms of response mechanisms, simple threshold alarm systems cannot achieve hierarchical disposal, which is prone to false alarms and difficult to trace the responsible parties.

[0004] These problems seriously restrict the efficiency and accuracy of on-site safety management of engineering projects, and there is an urgent need for a comprehensive solution that can achieve multi-source data fusion analysis, intelligent risk identification and graded response. Summary of the Invention

[0005] In order to solve the problems existing in the prior art, the present invention provides an online management method and system for engineering projects, which can realize real-time monitoring of material deformation of construction safety risks, intelligently identify risk causes, and trigger graded responses.

[0006] In a first aspect, the present invention provides an online management method for an engineering project, comprising: The video acquisition device deployed at the construction site acquires a continuous video stream of the monitored area, and simultaneously collects environmental data, including wind speed, humidity, temperature, and rainfall, through an environmental sensor network; The continuous video stream and the environmental data are timestamp aligned and spatial coordinate bound by a spatiotemporal synchronization unit to obtain a spatiotemporal synchronization data stream; Inputting the spatiotemporal synchronized data stream into a device for edge computing, extracting contour features of the resource stacking area through an image processing unit to obtain an initial contour dataset, and subjecting the initial contour dataset to continuous inter-frame displacement vector calculation and deformation rate analysis to generate a dynamic deformation feature vector; Inputting the dynamic deformation feature vector into a pre-trained machine learning classifier, and analyzing the association rules of coupling the deformation feature with the environmental data to obtain a risk cause classification result, wherein the risk cause classification result includes a human factor category and an environmental factor category; The risk cause classification results are matched with hierarchical responses to trigger a hierarchical alarm mechanism and execute corresponding instructions. The corresponding instructions include environmental factors and human factors. When the classification result is an environmental factor risk, the environmental linkage control device is started. When the classification result is a human factor risk, a security incident report containing a behavior pattern label is generated and pushed to the terminal.

[0007] In a second aspect, the present invention further provides an online management system for engineering projects, which is applied to the online management method for engineering projects as described in the first aspect; the online management system for engineering projects comprises: The front-end acquisition module is used to obtain a continuous video stream of the monitored area through the video acquisition device deployed at the construction site, and collect environmental data such as wind speed, humidity, temperature and rainfall through the environmental sensor network; An edge computing module is used to align the timestamps and bind the spatial coordinates of the continuous video stream and the environmental data through a spatiotemporal synchronization unit to obtain a spatiotemporal synchronized data stream; an intelligent analysis module for inputting the spatiotemporal synchronized data stream into a device for edge computing, extracting contour features of the resource stacking area through an image processing unit to obtain an initial contour dataset, and subjecting the initial contour dataset to continuous inter-frame displacement vector calculation and deformation rate analysis to generate a dynamic deformation feature vector; a risk decision module, configured to input the dynamic deformation feature vector into a pre-trained machine learning classifier, and obtain a risk cause classification result by coupling the deformation feature with the association rule analysis of the environmental data, wherein the risk cause classification result includes a human factor category and an environmental factor category; It is used to match the risk cause classification results through hierarchical responses, trigger a hierarchical alarm mechanism, and execute corresponding instructions. The corresponding instructions include environmental factors and human factors. When the classification result is an environmental factor risk, the environmental linkage control device is started. When the classification result is a human factor risk, a security incident report containing a behavior pattern label is generated and pushed to the terminal.

[0008] The online management method for engineering projects provided by the embodiments of the present invention significantly improves the efficiency of engineering safety management by constructing an intelligent analysis closed loop of multi-source data fusion, utilizes a spatiotemporal synchronization mechanism to accurately associate video streams with environmental sensor data, and realizes the coupled analysis of deformation dynamic characteristics and environmental parameters of material stacking areas, breaking through the limitations of data fragmentation in traditional monitoring systems. Based on a machine learning classifier, it performs two-dimensional tracing of risk causes, accurately distinguishes between human factors and environmental factors, and reduces the misjudgment rate. It relies on edge computing to optimize bandwidth resources, and only uploads key video clips with deformation characteristics exceeding the threshold. Combined with a hierarchical response mechanism, it achieves precise disposal, reducing the time efficiency of the entire process from risk identification to response, and greatly improving efficiency compared to manual inspections, forming an active defense system of "intelligent monitoring-precise attribution-rapid stop-loss", which comprehensively solves the three major problems of response delay, difficulty in tracing responsibility, and waste of resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flow chart of an online management method for an engineering project provided by an embodiment of the present invention; Figure 2 A schematic structural diagram of an online management system for an engineering project provided by an embodiment of the present invention.

[0010] Figure 3 This is a flow chart of performing edge computing on the spatiotemporal synchronous data stream input device to generate a dynamic deformation feature vector in an online management method for an engineering project provided by an embodiment of the present invention; Figure 4 This is a flow chart of analyzing the dynamic deformation feature vector input into a pre-trained machine learning classifier to obtain a risk cause classification result in an online management method for an engineering project provided by an embodiment of the present invention; Figure 5 It is a flow chart of triggering a hierarchical alarm mechanism to execute corresponding instructions based on hierarchical response matching of the risk cause classification results in an online management method for an engineering project provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0012] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Therefore, features specified as "first" or "second" may explicitly or implicitly include one or more features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0013] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or illustration". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0014] Reference Figure 1 , Figure 1 The figure is a flow chart of an online management method for an engineering project provided by the present invention, wherein the online management method for an engineering project comprises: Step 10: A video acquisition device deployed at the construction site acquires a continuous video stream of the monitored area, and simultaneously collects environmental data including wind speed, humidity, temperature, and rainfall through an environmental sensor network. Step 20, aligning timestamps and binding spatial coordinates of the continuous video stream and the environmental data by a spatiotemporal synchronization unit to obtain a spatiotemporal synchronized data stream; Step 30: Input the spatiotemporal synchronized data stream into the device for edge computing. The image processing unit extracts the contour features of the resource stacking area to obtain an initial contour dataset. The initial contour dataset is subjected to continuous inter-frame displacement vector calculation and deformation rate analysis to generate a dynamic deformation feature vector. Step 40: Input the dynamic deformation feature vector into a pre-trained machine learning classifier, and obtain a risk cause classification result by coupling the deformation feature with the association rule analysis of the environmental data. The risk cause classification result includes a human factor category and an environmental factor category. Step 50, the risk cause classification result is matched with the hierarchical response, the hierarchical alarm mechanism is triggered, and the corresponding instructions are executed. The corresponding instructions include environmental factors and human factors. When the classification result is an environmental factor risk, the environmental linkage control device is started. When the classification result is a human factor risk, a security event report containing a behavior pattern tag is generated and pushed to the terminal.

[0015] Specifically, in step 10, multi-source data acquisition equipment is deployed in the steel yard, sand and gravel area, and cement storage area of ​​the construction site through a distributed hardware system. The video acquisition device uses an explosion-proof pan-tilt camera with an IP67 protection grade to record H.264 encoded video streams. It is equipped with an automatic zoom lens and a wide dynamic range function to ensure that the outline of materials can be clearly captured in strong backlight environments. The environmental sensing network includes an ultrasonic anemometer installed on the top bracket of the pile, a capacitive humidity sensor embedded in the cement pile, a PT100 temperature sensor arranged on the surface of the material and the shady area, and a piezoelectric rain gauge fixed on the anti-vibration platform in the open air area. The industrial-grade transmission equipment uploads sensor data to the edge computing gateway every 500ms via the RS485 bus using the Modbus protocol. The video stream is independently transmitted to the local storage server via gigabit optical fiber to form the original data acquisition layer.

[0016] In a specific embodiment, independent collection operations are implemented using steel yard monitoring as an example. The video collection end uses a preset cruise control pan-tilt camera to scan the yard in an S-shaped path. When the infrared thermal imager detects that the steel temperature is greater than 60°C, it automatically switches to thermal imaging mode. The original video stream is saved as an MP4 file at 15 minutes per segment. The environmental sensor end installs an anemometer array 5 meters from the ground at the four corners of the yard to record the three-dimensional wind speed components in real time. At the same time, a temperature sensor chain is deployed in the gaps between steels to monitor the thermal conduction gradient inside the pile. In the local preprocessing stage, the edge computing device performs background modeling, separates dynamic targets from static piles through a Gaussian mixture model, and generates a binary image sequence containing only the outline of the steel. The environmental data is then subjected to median filtering to eliminate pulse interference, thereby achieving physical isolation between collection and primary processing.

[0017] During implementation, not only is data collection required, but differentiated data collection is also performed for multiple material scenarios. For example, in the sand and gravel area, fisheye cameras cover the material piles, combined with LiDAR to generate point cloud data. When a vibration sensor detects the entry of a loaded vehicle, the camera switches to HD recording mode. In the cement storage area, fog-proof infrared cameras are deployed to penetrate moisture interference. A three-layer humidity sensor chain (surface, middle, and bottom) is simultaneously activated to monitor humidity gradients. During rainfall, the tarpaulin gap detection function is activated, identifying gaps through pixel-level comparison of the real-time image with a dry cover template. For bandwidth optimization, keyframe extraction is performed on the video stream, and incremental encoding is used for environmental data. Ultimately, the data volume of a single monitoring point is compressed to ensure efficient operation of the collection system.

[0018] Specifically, in step 20, the continuous video stream and environmental sensor data obtained in step 10 are integrated in a spatiotemporal dimension by a spatiotemporal synchronization unit. The spatiotemporal dimension integration includes: Adopting the NTP-PTP hybrid clock protocol, a unified microsecond-level timestamp is applied to the video frame sequence based on the H.264 encoding stream and the environmental data packet based on the Modbus protocol. Establish a spatial coordinate mapping model to bind the 3D physical coordinates of equipment such as anemometers and humidity sensors to the corresponding pixel areas of the video image after measurement and calibration by a total station. For example, the coordinates of the steel yard (X12, Y07) are mapped to a 128×128 pixel block in the upper right quadrant of the image. Through the data injection engine of the edge computing gateway, environmental parameters are embedded in the non-image data area of ​​the video stream in SEI (supplementary enhancement information) format, generating synchronized data packets with spatial anchors and realizing millisecond-level spatiotemporal alignment of original data streams.

[0019] In a specific embodiment, the time-space binding operation is performed using a steel yard as an example: The 3D coordinates of the anemometer array (e.g., anemometer #1: X=12.3m, Y=7.5m, Z=5.2m) are converted to the video image coordinate system (pixel coordinates u=780-908, v=120-248) using a pre-configured spatial mapping table. The PTP precision clock protocol is used to synchronize the device clock, and the camera generates a timestamp for each frame.<frame_epoch> #<frame number> (e.g. 1688793655.125083#F12093), the environment data packet is marked with the same timestamp<sensor_epoch> @<device ID>; A data fusion engine runs within the edge gateway to analyze the GPS positioning information of the video frames (the camera's built-in RTK module has an accuracy of ±2cm), dynamically correct the coordinate offset caused by pile displacement, and output a synchronized data stream with a temporal and spatial error of less than 3cm / 5ms.

[0020] During the specific implementation process, dynamic calibration is implemented to meet the differentiated synchronization requirements of multiple regions, including: For mobile piles in sand and gravel areas, the point cloud of the pile is scanned in real time by LiDAR (updated at 10Hz), and the point cloud coordinate system is aligned with the camera coordinate system using the ICP algorithm to establish a dynamic spatial mapping relationship. For fixed monitoring points in the cement storage area, the preset affine transformation matrix is ​​used to constantly map the humidity sensor coordinates (e.g., surface sensor S01: X=5.2m, Y=3.1m, Z=1.0m) to the image coordinates; For time synchronization, a hardware-level clock synchronization module (based on the 1588v2 protocol on an FPGA) is deployed to eliminate software protocol stack delays and ensure that the deviation between the video frame exposure time and the environmental data collection time is ≤1ms. Finally, the time-space locked data packets are output through a customized encapsulation format (video stream H.264+SEI metadata area embedded with JSON format environment parameters).

[0021] Specifically, in step 30, the spatiotemporal synchronized data stream is input into the device for edge computing, and the contour features of the resource stacking area are extracted by the image processing unit to obtain an initial contour data set. The initial contour data set is subjected to continuous inter-frame displacement vector calculation and deformation rate analysis to generate a dynamic deformation feature vector.

[0022] In one embodiment, Figure 3 As shown, the dynamic deformation vector generation is based on the edge calculation of the spatiotemporal synchronous data stream input, and the contour features of the resource stacking area are extracted by the image processing unit to obtain an initial contour data set. The initial contour data set is subjected to continuous inter-frame displacement vector calculation and deformation rate analysis, including: step 301, data reception and preprocessing; step 302, contour feature extraction; step 303, displacement field calculation; step 304, deformation rate parameter generation; step 305, feature vector construction; step 306, threshold determination and data upload.

[0023] Specifically, in step 301, an encrypted spatiotemporal synchronized data stream is received through a dedicated data channel on the edge computing device. The spatiotemporal synchronized data stream integrates the multi-channel high-definition video from the video acquisition unit and the real-time parameters of the environmental sensing unit. The original data packet is processed by a multi-level parsing engine, separating the protocol encapsulation of the video transmission layer and the environmental data layer. The video layer is disassembled into independent frame units and the timestamp continuity is verified. The environmental layer extracts sensor data such as wind speed, humidity, and temperature and performs time calibration. The video frames are sent to the hardware decoding cluster and converted into a processable image matrix through parallel decoding technology. At the same time, the environmental data stream is filtered through a sliding window filtering algorithm to eliminate high-frequency noise interference. The decoded image is intelligently downsampled, and a scene stability assessment model is established based on background modeling. The sampling rate is dynamically adjusted in combination with a motion target detection algorithm. When the static area ratio exceeds a set threshold, only the key frame sequence is retained. Finally, a time-aligned lightweight dataset is output, with the original data volume compressed by 96% and the time synchronization error controlled to the millisecond level. This step breaks the bottleneck of massive data transmission and provides a high signal-to-noise ratio input source for real-time analysis.

[0024] In a specific embodiment, the implementation process utilizes an industrial-grade edge computing platform. The video decoding module integrates a hardware acceleration architecture, supporting the parallel processing of multiple 4K video streams. The environmental data processing unit deploys an adaptive Kalman filter to effectively suppress sensor signal jitter. Key frame screening uses a hybrid strategy combining structural similarity analysis and motion vector detection, dynamically discarding static frames with redundancy exceeding 95%. The system compresses and reduces the average daily data transmission volume while ensuring feature integrity. This solution significantly reduces network bandwidth pressure, stabilizes preprocessing latency within 30 milliseconds, and establishes an efficient data foundation for subsequent analysis.

[0025] Specifically, in step 302, the preprocessed image sequence is input into the multimodal contour analysis system to perform illumination adaptive enhancement, which dynamically adjusts the gamma curve according to the image grayscale distribution histogram to optimize the detail visibility of the backlight and strong reflection areas. The enhanced image enters the cascade edge detection process. In the first stage, the adaptive dual-threshold Canny operator is used to generate the initial contour map, and in the second stage, the directional gradient histogram algorithm is applied to refine the edge geometric features.

[0026] The contour binary image is subjected to morphological optimization, wherein the morphological optimization fills internal holes through a closing operation, and then filters out isolated noise points through a connected domain analysis.

[0027] A high-precision Hough transform analysis is performed on the regular pile. The Hough transform analysis sets a 0.01° resolution sampling in the inclination sensitive range (0°-30°) and determines the main axis orientation through accumulator peak detection.

[0028] Convex hull hierarchical modeling is used for irregular piles. The convex hull hierarchical modeling first constructs the minimum convex boundary, then calculates the surface curvature distribution characteristics based on the triangulation algorithm, and finally outputs a geometric parameter set with sub-pixel accuracy. The contour positioning error is better than 0.3 pixels.

[0029] This step enables accurate capture of deformation features in complex scenarios and provides reliable spatial parameters for risk assessment.

[0030] In a specific embodiment, for dust interference environment, the system activates a multi-spectral fusion mechanism, which includes providing texture details through visible light images and assisting boundary positioning with infrared thermal images; the convex hull calculation adopts an incremental Graham scanning algorithm to support dynamic tracking of the pile deformation process.

[0031] Curvature analysis incorporates a machine learning correction module. Using a pre-trained model to identify specific material pile characteristics, the system automatically optimizes the curvature calculation weight coefficient. This system can detect subtle surface deformations (such as a 5-centimeter drop in a 10-meter diameter sand pile) with high efficiency. This solution maintains contour recognition accuracy even in harsh environments, significantly improving the robustness of feature extraction.

[0032] Specifically, in step 303, sequential keyframes are fed into a hierarchical optical flow computation framework to construct a three-level image pyramid. A global displacement is initially estimated at the top, low-resolution layer to generate a coarse-grained displacement vector field. This initial estimate serves as the initial condition for optimization of the lower layers, and a variational optical flow algorithm is used to iteratively solve for sub-pixel displacement in a local window. The generated full-resolution displacement field is then subjected to motion consistency verification. A random sampling consistency algorithm is used to separate device jitter components from the true target displacement. Feature points are then optimized for the verified displacement data, and texture richness is assessed based on a corner response function. A displacement statistical model is constructed by selecting 200 spatially uniformly distributed feature points. Finally, a complete feature set containing the average displacement modulus, maximum displacement, and displacement acceleration is output, with a displacement detection sensitivity of 0.05 pixels per frame. This step achieves precise quantification of motion vectors, providing a dynamic basis for deformation trend analysis.

[0033] In one specific embodiment, the implementation integrates optically assisted positioning technology, constructing a displacement correction model using a laser reference matrix. The calculations employ a partitioning strategy, dividing the monitored area into 1m×1m grid cells. The displacement characteristics of each cell are independently analyzed, and abnormal areas are marked. This system can reliably detect minute hourly tilt changes even in heavy rain, with a low false detection rate. The processing engine utilizes GPU parallel acceleration, minimizing the time required for displacement analysis of 1080P images. This solution overcomes the challenge of displacement detection in vibration-prone environments and provides precise kinematic parameters for engineering safety.

[0034] Specifically, in step 304, the contour feature point set is input into the intelligent area calculation module, the scan line filling algorithm is used for regular polygonal piles, and the Monte Carlo regional integration method is applied to irregular piles; the current area value is compared with the reference state, and the physical deformation rate is calculated in combination with the displacement characteristics. The material elastic modulus parameter is introduced to realize the conversion of pixel displacement to actual deformation variable, and the deformation rate index is derived in combination with the time interval.

[0035] The deformation rate sequence over consecutive time periods is input into the time-frequency analysis engine. Wavelet transforms are used to separate the sudden deformation component from the gradual deformation component, extracting acceleration characteristic parameters. The environmental coupling unit calculates the wind speed-pile height critical curve and the humidity-bearing capacity attenuation model in real time, outputting a deformation risk indicator compensated for environmental factors. This step maps image features to physical risks, providing a quantitative basis for early warning decisions.

[0036] In one specific embodiment, the system integrates a material mechanics database containing pre-set physical parameters of engineering materials such as steel and concrete. Temperature field compensation is incorporated into deformation rate calculations, using an infrared thermal imager to monitor temperature gradients and automatically correct the thermal expansion coefficient when local temperature differences exceed a certain limit. During high-temperature testing, the system can successfully warn of potential micro-deformation in the rail pile and trigger a response in advance. This solution improves deformation detection sensitivity and reduces false alarm rates.

[0037] Specifically, in step 305, a multi-source feature fusion channel is constructed. The geometric feature channel receives inclination and curvature data, the motion feature channel inputs displacement statistics, the deformation channel accesses rate and acceleration parameters, and the environmental channel integrates wind speed and humidity information. Weights are dynamically assigned through the feature importance assessment module, and a random forest algorithm is used to analyze historical accident data, assigning a 40% decision weight to the displacement and acceleration features. The weighted feature set is fed into an adaptive normalization processor, which dynamically updates the extreme value boundaries of each feature dimension based on a sliding time window. Finally, a principal component analysis engine performs dimensionality reduction, retaining principal components with a cumulative contribution rate greater than 95%, and outputting 3-5 dimensional lightweight feature vectors to reduce data volume. This step effectively reduces the dimensionality of high-dimensional features, providing efficient data representation for real-time decision-making at the edge.

[0038] In one specific embodiment, the system configures adaptive material templates. These templates include a three-dimensional vector of [inclination rate, displacement acceleration, and wind speed] for steel piles and a three-dimensional vector of [curvature rate, average displacement, and humidity] for sand and gravel piles. Normalized parameters are stored in dual redundancy, with local devices caching real-time parameters and cloud servers storing historical optimal values. Automatically switching to offline mode in the event of a network interruption ensures continuous system operation. This solution overcomes the computing power limitations of edge devices and enables rapid completion of complex feature analysis.

[0039] Specifically, in step 306, the feature vector is input into the risk quantification model, and the Mahalanobis distance in the feature space is calculated as a comprehensive risk indicator. The comprehensive risk indicator integrates feature correlation and distribution characteristics. The dynamic threshold adjustment module integrates a meteorological warning interface, receiving real-time rainfall intensity and wind level data, and automatically lowering the judgment threshold in adverse conditions. When the risk indicator exceeds the limit, the intelligent event capture engine is triggered. Based on PTP precision clock positioning, it accurately captures the original video from 10 minutes before the risk to 5 minutes after the risk occurs. Video compression uses perceptual coding technology, allocating a low bitrate to static background areas and enhancing compression for moving target areas. The upload strategy implements a four-level priority control: real-time transmission for level 1 risks (collapse probability >80%), off-peak transmission for level 2 risks (30%-80%), and local storage for level 3 man-made events. This step establishes a hierarchical response closed loop to maximize the utilization of network resources.

[0040] In one specific embodiment, risk assessment utilizes a Bayesian optimization mechanism, with an initial threshold set at 0.75 and parameters dynamically updated based on on-site response feedback. Event video packets are generated using inter-frame predictive coding, with the first frame transmitted intact and subsequent frames transmitting only the differenced regions. The system supports SM4 encryption, ensuring data security. This solution achieves a 97% improvement in bandwidth utilization, establishing an efficient and reliable risk response system.

[0041] Specifically, in step 40, the dynamic deformation feature vector is input into a pre-trained machine learning classifier, and the risk cause classification result is obtained by coupling the deformation feature with the association rule analysis of the environmental data. The risk cause classification result includes a human factor category and an environmental factor category.

[0042] In one embodiment, Figure 4 As shown, the risk cause classification result is based on inputting the dynamic deformation feature vector into a pre-trained machine learning classifier, and is obtained by coupling the deformation feature with the association rule analysis of the environmental data, including: step 401, feature input and preprocessing; step 402, machine learning classifier loading; step 403, human factor segmentation processing; step 404, environmental coupling rule analysis; step 405, risk cause classification decision; step 406, result confidence verification.

[0043] Specifically, in step 401, the standardized dynamic deformation feature vector output by the edge computing device enters the feature selection engine, which automatically selects core indicators based on a pre-trained feature importance scoring model. The full-dimensional feature vector is input into the gain rate evaluation module, which calculates the mutual information between each feature and historical risk events and retains the feature subset whose information gain value exceeds the set threshold. For the risk cause classification task, four strongly correlated features, namely displacement acceleration, deformation rate, wind speed, and humidity, are prioritized, and redundant dimensions are eliminated. The selected features are calibrated in real time using sliding window normalization technology to eliminate device measurement bias. Finally, a lightweight core feature subset is generated, reducing the data volume. This step significantly improves feature quality and lays the data foundation for accurate classification.

[0044] In one specific embodiment, the implementation process uses a dynamic feature screening strategy. The system automatically updates the feature importance scoring model every 24 hours and recalculates the gain rate based on the data of 1,000 recent risk events. For example, the feature selection threshold is set to 0.35, and only feature dimensions above this value are retained. The standardization process introduces an environmental reference value calibration mechanism. When the temperature and humidity sensors detect extreme environments, the normalization parameter boundaries are automatically adjusted. The feature dimensions are compressed from 12 to 4, reducing processing latency by 60% while maintaining 98% of the effective information volume. This solution solves the problem of high-dimensional feature processing efficiency and improves the real-time performance of edge devices.

[0045] Specifically, in step 402, a pre-trained random forest classifier set containing 500 deeply optimized decision trees is retrieved from a cloud-based model repository. The original model is processed by an edge adaptation engine, complex subtrees exceeding 10 layers in depth are pruned, and 32-bit floating-point parameters are quantized to 8-bit fixed-point numbers. Key decision rules are extracted using knowledge distillation technology and reconstructed into a lightweight decision forest. The model loading process uses dynamic memory mapping technology to activate only decision paths related to the current features. Finally, an optimized model is generated that can run on low-computing power devices, improving inference speed. This step overcomes the limitations of edge computing resources and enables efficient deployment of complex models.

[0046] In one specific embodiment, the implementation process utilizes a hierarchical loading mechanism, with the base decision tree resident in memory and subtrees for specific scenarios loaded on demand. A structured pruning algorithm is applied during the model optimization phase to remove subtree branches with a contribution rate below 5%. A non-uniform quantization strategy is used for parameter quantization, retaining higher precision for key decision nodes. This shortens model loading time and reduces memory usage. This solution enables low-end devices to possess advanced classification capabilities, expanding the system's applicable scenarios.

[0047] Specifically, in step 403, the video analysis unit extracts real-time data on personnel activity trajectories and feeds it into a three-level behavior classification pipeline. The first-level processor detects the number and distribution density of personnel, distinguishing between collaborative work and individual activities through group motion pattern recognition. The second-level analysis unit calculates the rate of change of personnel movement speed and the consistency of movement direction, combining the interaction status of the equipment to identify unintentional collision characteristics. The third-level verification module compares operation time with safety regulations and detects abnormal contact behavior through a transport vehicle identification algorithm. The final output is a fine-grained label set containing active handling, unintentional collision, and abnormal contact. This step enables precise tracing of human risk and supports hierarchical management and control decisions.

[0048] In one specific implementation, the system employs spatiotemporal behavior modeling, constructing a three-dimensional relationship matrix between personnel, equipment, and the pile. A trajectory clustering algorithm identifies typical behavior patterns. Unintentional collision detection uses dual thresholds, triggering an alarm when, for example, the speed mutation rate exceeds 1.5 m / s² and the distance to the pile is less than 2 meters. Abnormal contact detection incorporates a nighttime work permit database, automatically marking unauthorized operations as high-risk. Even in complex construction sites, the system can successfully distinguish most types of human-induced incidents with a low false positive rate. This solution addresses the inability of traditional monitoring systems to trace behavior to its source.

[0049] Specifically, in step 404, the core feature subset is input into the physical rule engine for dual analysis using a material mechanics model and an environmental effects model. The material mechanics pathway calculates the critical stack height curve for the current wind speed and compares it to the actual stack height in real time to generate a stability index. The environmental effects pathway analyzes the attenuation of material strength due to humidity changes and outputs a bearing capacity reduction factor. The results of these two pathways are then combined in a weighted fusion module, dynamically adjusting the weight ratio based on the type of material on site, ultimately generating a quantitative environmental risk indicator. This step establishes a quantitative correlation between deformation characteristics and environmental impacts, revealing the underlying physical mechanisms.

[0050] In one specific implementation, the implementation process integrates an engineering materials database, pre-setting physical parameters for 12 material categories, including steel, cement, and sand and gravel. A wind speed response model leverages fluid dynamics simulation data to establish a nonlinear mapping between pile height and wind speed. Humidity impact analysis incorporates a time-varying attenuation factor, automatically increasing the risk level with sustained exposure to high humidity. This enables the system to accurately warn of weather-related sand pile collapses. This solution will improve the accuracy of environmental risk identification and provide a basis for subsequent response.

[0051] Specifically, in step 405, core features, human labels, and environmental indicators are input into a multi-source decision engine, where causal traceability is achieved through association rule mining. First, a feature correlation matrix is ​​constructed, and the correlation coefficient between human factors and environmental indicators is calculated. When the correlation coefficient is below a threshold and the human label is valid, the risk is determined to be human-dominated. When the environmental indicator exceeds the threshold and the deformation characteristics conform to physical laws, the risk is determined to be environmental-dominated. For conflicting cases, a deep traversal of the decision tree is initiated, parsing complex scenarios through multiple key judgment nodes, and ultimately outputting a binary classification result and the basis for determining the dominant factor. This step achieves dual verification of machine reasoning and physical rules, ensuring the interpretability of the conclusions.

[0052] In one specific embodiment, the implementation process utilizes a confidence-driven strategy. Initial classification results are generated by a random forest, and when the confidence level is less than 85%, a rule engine review is triggered. The correlation analysis module integrates a library of typical accident cases to compare the similarity between the event under test and historical patterns. For example, in analyzing a steel hoisting accident, the system accurately identified the coupling between human error (an unlocked hook) and a sudden gust of wind (force 8), resulting in a precise attribution of 65% human error and 35% environmental error. This approach improves classification accuracy and provides more robust evidence than traditional methods.

[0053] Specifically, in step 406, the initial classification results are input into a multi-layered verification system. The first layer calculates a baseline confidence level using the probability output of a random forest. The second layer compares the distributional similarity of feature values ​​with historical cases. The third layer detects the conflict index between environmental and human factors. These three results are weighted and integrated to generate a final confidence level between 0 and 1. For example, if the confidence level is ≥ 0.85, the result is directly output; if it is between 0.75 and 0.85, the video review process is triggered; if it is less than 0.75, the expert consultation mechanism is initiated. Verified conclusions are automatically stored in the case library optimization model. This step establishes a continuously evolving quality assurance system to ensure the reliability of risk attribution.

[0054] In one specific embodiment, a dynamic weighting strategy is implemented during the implementation process, such as a 60% weighting for base confidence, a 25% weighting for distribution similarity, and a 15% weighting for the contradiction index. Video review utilizes intelligent capture of key segments and automatically annotates controversial feature points for manual confirmation. This reduces the rate of rechecks triggered by insufficient confidence and improves the efficiency of model self-optimization. This solution establishes a closed-loop quality improvement mechanism by reducing high-risk misjudgments.

[0055] Specifically, in step 50, the risk cause classification result is matched with the hierarchical response, the hierarchical alarm mechanism is triggered, and the corresponding instructions are executed. The corresponding instructions include environmental factors and human factors. When the classification result is an environmental factor risk, the environmental linkage control device is started. When the classification result is a human factor risk, a security incident report containing a behavior pattern label is generated and pushed to the terminal.

[0056] In one embodiment, Figure 5 As shown, the hierarchical alarm mechanism is based on matching the risk cause classification results with hierarchical responses, triggering the hierarchical alarm mechanism and executing corresponding instructions. The corresponding instructions include environmental factors and human factors. When the classification result is an environmental factor risk, the environmental linkage control device is started. When the classification result is a human factor risk, a security incident report containing a behavior pattern label is generated and pushed to the terminal, including: step 501, risk classification result input; step 502, environmental risk linkage response; step 503, human risk report generation; step 504, multi-level response execution.

[0057] Specifically, in step 501, the risk cause classification results output by the machine learning classifier enter the hierarchical strategy matching engine. The risk cause classification results include a dominant risk type label and a confidence score. The classification results are matched with the dynamic response rule library, the risk type label is parsed, and the dominant environmental or human factor is determined. The initial response level is then determined based on the confidence score. For example, when the confidence level is higher than 0.85, the classification result is directly adopted. When the confidence level is between 0.75 and 0.85, auxiliary rule verification is triggered, comparing the real-time data from environmental sensors with the intermediate results of behavioral analysis. If the confidence level is lower than 0.75, a manual review process is initiated. The matching process uses a multi-dimensional decision matrix, comprehensively considering factors such as risk type, location, and time sensitivity, to output an initial response level code, which includes level 1, level 2, and level 3. This step achieves a precise mapping of risk conclusions to response strategies, laying the foundation for decision-making in hierarchical disposal.

[0058] In a specific embodiment, the implementation process adopts an adaptive matching mechanism, and the system automatically loads the response rule template according to the type of construction site. For example, bridge projects focus on wind speed response, and underground projects focus on humidity control.

[0059] When a red alert for heavy rain is detected, the confidence level of environmental risks is automatically increased; the matching engine has a built-in case reasoning module to compare the similarity of the event to be handled with 500 historical accidents, and optimize the response level determination; this solution improves the accuracy of level determination and controls the decision delay within seconds, solving the problem of rigid response levels in traditional systems and achieving dynamic and accurate matching.

[0060] Specifically, in step 502, when environmental factors are determined to be the dominant risk, the environmental coupling analysis module obtains quantitative risk values ​​for parameters such as wind speed and humidity. These risk values ​​are input into a threshold classifier. Exceeding a critical threshold triggers a primary response, initiating a hard shutdown protocol for the equipment. Moderate overshooting triggers a secondary response, generating an early warning report. The primary response signal is transmitted via the industrial control bus, activating the audible and visual alarm array to emit a high-frequency warning tone and simultaneously sending a safety shutdown command to construction equipment. Tower cranes activate mechanical locking devices, while mixers initiate an emergency shutdown and power outage. The secondary response signal is processed by the early warning engine, automatically generating a structured report containing risk coordinates, environmental parameters, and recommended measures. This report is then pushed to the manager's mobile terminal via the message-based middleware. This step establishes an automated closed-loop environmental risk management system to minimize disaster losses.

[0061] In a specific implementation, for example, a coordinated response is implemented for rainstorm scenarios, triggering a primary response when humidity exceeds 90% and rainfall exceeds 50 mm / h. Equipment control utilizes a three-tiered protection mechanism: PLCs directly power off core equipment, the Internet of Things shuts down secondary equipment, and mechanical locking devices activate physical protection. Early warning reports are automatically linked to BIM models, highlighting high-risk areas in 3D visualization. For example, at one construction site, the system automatically activated a rain shelter 12 minutes before a rainstorm, preventing the loss of 200 tons of cement from hardening. This solution shortens the time required to address environmental disasters and reduces economic losses.

[0062] Specifically, in step 503, if the risk is determined to be dominated by human factors, the behavior analysis unit obtains fine-grained behavior tags, which are categorized as active handling, unintentional collision, and abnormal contact. The behavior tags are input into the trajectory reconstruction engine, which generates the person's motion path based on the video tracking data. Kalman filtering is used to eliminate jitter noise and annotate key behavior points, such as the moment of contact with the pile. The optimized trajectory and behavior tags are input into the report generator, which automatically generates a structured event report using a natural language processing template, including elements such as the behavior type, occurrence time, geographic coordinates, and risk level. The report is digitally signed and timestamped by the security audit module and encrypted and pushed to the responsible terminal via the enterprise service bus. This step enables accurate tracing of human responsibility and supports safety management decisions.

[0063] In a specific implementation, for example, when handling unusual contact incidents at night, the system automatically captures the first five minutes of video footage when it detects an unauthorized person entering the storage yard. Trajectory reconstruction uses heatmap overlay technology to visually display hotspots of activity. A report generator uses a pre-set template: "[Behavior type] detected at [time] [location], risk level [level], recommended action [measure]." Once the report is pushed, it automatically triggers a work order system and assigns it to the on-duty security personnel. This system improves the efficiency of tracing violations and enhances the relevance of safety education, resolving the industry challenge of unclear demarcation of human responsibility.

[0064] Specifically, in step 504, the response level code triggers a differentiated execution chain. The first-level response initiates a hard linkage at the device level, directly connecting to the device controller via the industrial bus, forcing physical protection such as shutdown, power outage, and lockout, and simultaneously activating the emergency broadcast system to play directional evacuation instructions. The second-level response triggers a soft warning at the management level, parsing the risk report into a visual alarm interface, calibrating the risk radius in the BIM model, and pushing disposal recommendations to the responsible person. The third-level response initiates the security disposal flow, automatically linking video evidence to generate a disposal work order, assigning it to the nearest security personnel's mobile terminal, and simultaneously initiating continuous tracking video recording. All response processes are recorded in real time on the blockchain evidence storage platform to ensure traceability of operations. This step achieves precise and hierarchical risk disposal and forms a complete security closed loop.

[0065] In a specific implementation, such as a steel hoisting incident, the system detected a sudden change in wind speed, triggering a hoisting risk. A Level 1 response was triggered within one second: the crane's rotating mechanism was locked, an audible and visual alarm was activated within three seconds, and a warning report with 3D positioning was sent to the engineer within five seconds. Simultaneously, a Level 3 response was triggered, tracking and recording the operator's illegal overloading behavior and generating an accountability report. The entire process was completed within seconds, preventing a potential overturning accident. This system shortens incident response time, improves accountability, restructures engineering safety response standards, and establishes a new paradigm for tiered response.

[0066] The embodiments of the present invention revolutionize engineering safety management through an innovative multi-source data fusion and intelligent analysis architecture. At the risk monitoring level, they overcome the time constraints of traditional manual inspections, enabling continuous monitoring of material deformation and environmental risks in seconds, accurately capturing early-stage hazards such as micron-level displacement and millimeter-level settlement. At the causal analysis level, a dual-dimensional traceability mechanism for environmental and human factors, combined with a physical rule engine and machine learning model, accurately analyzes complex coupling relationships such as wind speed-pillar instability and humidity-material degradation, while simultaneously achieving fine-grained identification of human misconduct. At the response and disposal level, a hierarchical, interconnected closed-loop system is established, with environmental risks automatically triggering equipment hard shutdown protection and human risk intelligently generating behavior tracing reports, significantly improving the timeliness and accuracy of response. Furthermore, the edge computing architecture optimizes resource consumption, while intelligent video stream compression and model lightweighting technologies significantly reduce bandwidth and storage requirements. This enables the low-cost deployment of a full-coverage monitoring network for large-scale projects, ultimately forming a self-evolving, full-process "risk perception-intelligent diagnosis-precise disposal" system. This system transforms reactive accident response into preventative safety management, comprehensively enhancing the inherent safety of engineering construction.

[0067] Furthermore, the online management system of the engineering project provided by the present invention is described below. The online management system of the engineering project described below and the online management method of the engineering project described above can be referred to each other. Figure 2 , Figure 2 This is a schematic diagram of the structure of the online management system for engineering projects provided by the present invention. The online management system for engineering projects includes: The front-end acquisition module 210 is used to obtain a continuous video stream of the monitored area through the video acquisition device deployed at the construction site, and collect environmental data such as wind speed, humidity, temperature and rainfall through the environmental sensor network; The edge computing module 220 is configured to align the timestamps of the continuous video stream and the environmental data and bind the spatial coordinates thereof to obtain a spatiotemporal synchronized data stream through a spatiotemporal synchronization unit; An intelligent analysis module 230 is configured to input the spatiotemporal synchronized data stream into a device for edge computing, extract contour features of the resource storage area through an image processing unit to obtain an initial contour dataset, and perform inter-frame displacement vector calculation and deformation rate analysis on the initial contour dataset to generate a dynamic deformation feature vector; A risk decision module 240 is configured to input the dynamic deformation feature vector into a pre-trained machine learning classifier, and obtain a risk cause classification result by coupling the deformation feature with the association rule analysis of the environmental data, wherein the risk cause classification result includes a human factor category and an environmental factor category; The response execution module 250 is used to match the risk cause classification results with hierarchical responses, trigger a hierarchical alarm mechanism, and execute corresponding instructions. The corresponding instructions include environmental factors and human factors. When the classification result is an environmental factor risk, the environmental linkage control device is started. When the classification result is a human factor risk, a security incident report containing a behavior pattern label is generated and pushed to the terminal.

[0068] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An online management method for an engineering project, characterized in that: include: The video acquisition device deployed at the construction site acquires a continuous video stream of the monitored area, and simultaneously collects environmental data, including wind speed, humidity, temperature, and rainfall, through an environmental sensor network; The continuous video stream and the environmental data are timestamp aligned and spatial coordinate bound by a spatiotemporal synchronization unit to obtain a spatiotemporal synchronization data stream; Inputting the spatiotemporal synchronized data stream into a device for edge computing, extracting contour features of the resource stacking area through an image processing unit to obtain an initial contour dataset, and subjecting the initial contour dataset to continuous inter-frame displacement vector calculation and deformation rate analysis to generate a dynamic deformation feature vector; Inputting the dynamic deformation feature vector into a pre-trained machine learning classifier, and analyzing the association rules of coupling the deformation feature with the environmental data to obtain a risk cause classification result, wherein the risk cause classification result includes a human factor category and an environmental factor category; The risk cause classification results are matched with hierarchical responses to trigger a hierarchical alarm mechanism and execute corresponding instructions. The corresponding instructions include environmental factors and human factors. When the classification result is an environmental factor risk, the environmental linkage control device is started. When the classification result is a human factor risk, a security incident report containing a behavior pattern label is generated and pushed to the terminal.

2. The method according to claim 1, characterized in that Inputting the spatiotemporal synchronized data stream into a device for edge computing includes: By real-time monitoring of the dynamic deformation feature vector, a feature value change curve is obtained, and the feature value change curve is compared with a predetermined threshold. When the feature value exceeds the threshold, a video clip upload instruction is triggered, and the video clip associated with the instruction is intercepted with continuous frames from a set time period before the risk occurs to a set time period after the risk occurs, to obtain an event video package, and the event video package is uploaded to the central server.

3. The method according to claim 1, characterized in that The step of calculating the displacement vectors and analyzing the deformation rate of the initial contour data set between consecutive frames to generate a dynamic deformation feature vector includes: The displacement acceleration data of the stacked materials is obtained by calculating the displacement acceleration using the optical flow method; Processing the displacement acceleration data through a contour matching algorithm to identify the position offsets of the vertices of the stack to obtain a vertex offset set; The vertex offset set is combined with the wind speed gradient, humidity change rate and material type code in the environmental sensor data to construct a spatiotemporal feature matrix; The spatiotemporal feature matrix is ​​normalized to generate a standardized dynamic deformation feature vector.

4. The method according to claim 1, wherein The step of inputting the dynamic deformation feature vector into a pre-trained machine learning classifier includes: The original accident dataset is obtained by collecting surveillance video clips of historical safety accidents and corresponding environmental sensor data. The original accident dataset is annotated with two dimensions of human factors and environmental factors to generate an annotated dataset. The annotated dataset is extracted from the pile deformation feature sequence and environmental feature vector of the video clips to obtain a training sample set. The training sample set is trained with a random forest algorithm to learn the mapping rules between deformation features and environmental data, and a classifier model containing multiple decision trees is generated.

5. The method according to claim 1, wherein The human factors categories include: Analyzing the human activities in the video stream using a behavior recognition model to obtain an initial behavior feature set, and processing the initial behavior feature set through a multi-level classifier; Identify active transport behaviors through the characteristics of personnel number changes and equipment interaction status; Identify unintentional collision behaviors through the characteristics of movement speed mutation and pile contact distance; Identify abnormal contact behaviors through operation time compliance analysis and transportation tool frequency; Finally, the refined human factor subcategory labels are output.

6. The method according to claim 1, characterized in that The risk cause classification results are matched with hierarchical responses, a hierarchical alarm mechanism is triggered, and corresponding instructions are executed, including: The risk decision module receives the risk cause classification result to obtain a risk level determination signal, and matches the risk level determination signal with the response strategy rule library. When it is determined to be a level one environmental risk, the sound and light alarm device is triggered and an equipment shutdown instruction is output. When it is determined to be a level two environmental risk, the risk location coordinates and environmental parameter abnormality report are generated and pushed to the management terminal. When it is determined to be a level three man-made risk, video tracking is started to record the activity trajectory of the personnel to obtain behavior trajectory data, and the behavior trajectory data is processed through behavior pattern recognition to generate a security event log.

7. An online management system for a construction project, applied to the online management method for a construction project according to any one of claims 1 to 6; the online management system for a construction project comprises: The front-end acquisition module is used to obtain a continuous video stream of the monitored area through the video acquisition device deployed at the construction site, and collect environmental data such as wind speed, humidity, temperature and rainfall through the environmental sensor network; An edge computing module is used to align the timestamps and bind the spatial coordinates of the continuous video stream and the environmental data through a spatiotemporal synchronization unit to obtain a spatiotemporal synchronized data stream; an intelligent analysis module for inputting the spatiotemporal synchronized data stream into a device for edge computing, extracting contour features of the resource stacking area through an image processing unit to obtain an initial contour dataset, and subjecting the initial contour dataset to continuous inter-frame displacement vector calculation and deformation rate analysis to generate a dynamic deformation feature vector; a risk decision module, configured to input the dynamic deformation feature vector into a pre-trained machine learning classifier, and obtain a risk cause classification result by coupling the deformation feature with the association rule analysis of the environmental data, wherein the risk cause classification result includes a human factor category and an environmental factor category; The response execution module is used to match the risk cause classification results through hierarchical response, trigger the hierarchical alarm mechanism, and execute corresponding instructions. The corresponding instructions include environmental factors and human factors. When the classification result is an environmental factor risk, the environmental linkage control device is started. When the classification result is a human factor risk, a security incident report containing a behavior pattern label is generated and pushed to the terminal.

Images