Building construction quality safety risk management system

Abstract

The invention relates to an intelligent management system, in particular to a building construction quality safety risk management system which comprises a data acquisition module, a convolution analysis module, a risk assessment module, an early warning decision module and a feedback optimization module. The data acquisition module generates a sensing signal. And the convolution analysis module receives and adopts a multi-scale feature pyramid and a cross-modal attention mechanism to process the sensing signal to form a four-dimensional fusion signal. And the risk assessment module carries out comparative analysis on the four-dimensional fusion signal under a digital twinborn framework and forms a five-dimensional risk early warning signal. The early warning decision module receives the risk early warning signal and forms a multi-level risk early warning level signal. And the feedback optimization module receives the risk early warning level signal and optimizes the building construction quality safety risk management system through an incremental learning technology. The building construction quality safety risk management system can solve the problems that the dynamic characteristic adaptability is low in the construction process, and misinformation or missing information is likely to be generated.

Description

Technical Field

[0001] The present invention relates to an intelligent management system, specifically to a construction quality and safety risk management system, and more specifically to a comprehensive management system integrating multi-modal perception, intelligent analysis, and closed-loop optimization. Background Art

[0002] In the field of construction, quality and safety management has always been the core link of project management. Traditional construction safety management mainly relies on manual inspections, regular checks, and static regulatory constraints. This approach has obvious limitations: Firstly, the coverage of manual inspections is limited, making it difficult to achieve round-the-clock and all-round monitoring; Secondly, the inspection results are greatly affected by personnel experience, resulting in subjective judgment deviations; Moreover, there is a time lag in risk identification and response to handling, making it difficult to detect and respond to emergencies in a timely manner. With the continuous expansion of the scale of construction projects and the increasing complexity of construction technologies, traditional management methods are no longer able to meet the needs of modern engineering construction.

[0003] In recent years, new technologies such as the Internet of Things and big data have been initially applied in the field of construction. Some projects have tried to deploy video surveillance systems and sensor networks to achieve digital monitoring of some construction scenarios. However, the application of these technologies is often limited to the collection and analysis of single data sources, lacking the deep integration of multi-modal data and being difficult to comprehensively reflect the complex state of the construction site. For example, relying solely on video surveillance cannot accurately evaluate the internal stress changes of structures, and simply relying on sensor data is difficult to identify behavioral risks such as personnel's illegal operations. In addition, existing systems mostly use fixed thresholds to judge risks, with low adaptability to the dynamic characteristics of the construction process, and are prone to false alarms or missed alarms. Summary of the Invention

[0004] In view of the above disadvantages of the prior art, the purpose of the present invention is to provide a construction quality and safety risk management system to solve the problems of low adaptability to the dynamic characteristics of the construction process and being prone to false alarms or missed alarms. The present invention collects multi-dimensional data of the construction process in real time through a sensor network deployed at the construction site, extracts risk features using deep learning methods, combines with a knowledge graph in the construction field for risk assessment, conducts risk early warning based on the risk assessment, and establishes a dynamic optimization mechanism to achieve the continuous improvement of the system.

[0005] The present invention provides a construction quality and safety risk management system, including: A data acquisition module, which includes a vision unit, a radar unit, a mechanics unit, and an environmental perception unit, and generates perception signals therefrom; A convolutional analysis module, which receives and processes the perception signals using a multi-scale feature pyramid and a cross-modal attention mechanism to form a four-dimensional fusion signal; A risk assessment module, which includes a structure assessment unit, a material assessment unit, a process detection unit, an environment assessment unit, and a personnel assessment unit. The risk assessment module conducts comparative analysis on the four-dimensional fusion signal under the digital twin framework and forms a five-dimensional risk warning signal; A warning decision-making module, which receives the risk warning signal and forms a multi-level risk warning level signal; A feedback optimization module, which receives the risk warning level signal and optimizes the building construction quality and safety risk management system through incremental learning technology; Among them, the convolutional analysis module includes four steps: multi-modal data preprocessing and alignment, multi-scale spatio-temporal feature extraction, cross-modal attention fusion, and construction prediction.

[0006] In an embodiment of the present invention, the perception signals include visual perception signals, radar perception signals, mechanical perception signals, and environmental perception signals. The visual unit includes a collaborative acquisition architecture of a high-resolution panoramic camera array and a mobile drone platform. Among them, the panoramic camera array is deployed at the high point of the construction area in a hexagonal topology structure, and each camera node is equipped with an automatic cleaning pan-tilt head and a dynamic exposure adjustment module. The drone platform is equipped with a three-axis stabilized pod and a laser rangefinder. The visual unit forms visual perception signals, and the acquisition frequency of the visual perception signals is dynamically adjusted according to the construction stage; the radar unit includes a lidar and a millimeter-wave radar. The lidar outputs a three-dimensional density map, and the millimeter-wave radar outputs a wide-angle coverage map. The three-dimensional density map and the wide-angle coverage map are fused by the Kalman filtering algorithm for the motion target trajectory to form a radar perception signal; the mechanical unit includes a distributed fiber optic sensor network and a piezoelectric vibration sensor. The distributed fiber optic sensor is laid on the surface of the load-bearing structure in a snake-shaped path, and the piezoelectric vibration sensor adopts a resonance frequency adaptive adjustment technology. The mechanical unit forms a mechanical perception signal; the environmental perception unit includes a micro-meteorological station and a multi-gas detection device. The micro-meteorological station includes an ultrasonic wind speed and direction sensor and a laser disdrometer. The multi-gas detection device detects the concentration of hazardous gases through an NDIR (non-dispersive infrared gas analysis technology) infrared sensor array. The environmental perception unit forms an environmental perception signal.

[0007] In an embodiment of the present invention, when performing multi-modal data preprocessing and alignment, spatio-temporal calibration and data formatting processing are included. When performing spatio-temporal calibration, the convolutional analysis module receives perception signals, aligns the spatial coordinate systems of visual perception signals and radar perception signals through a checkerboard calibration board, synchronizes the timestamps of each unit under the data acquisition module to microsecond-level accuracy using the NTP (Network Time) protocol. Visual data is cut into a sequence of spatio-temporal cubes of 256×256 pixels, and every 15 frames form an analysis unit. Point cloud data is processed by voxelization to generate a three-dimensional density grid with an accuracy of 0.1 m³, and visual feature signals and radar feature signals are formed. When performing data formatting processing, mechanical perception signals are subjected to Z-score (standard score or zero-mean unit variance) normalization to eliminate dimension differences and form mechanical feature signals, and environmental perception signals are converted into a structured tensor format and form environmental feature signals.

[0008] In an embodiment of the present invention, when performing multi-scale spatio-temporal feature extraction, a parallel dual-path feature pyramid architecture is used to extract visual feature signals, radar feature signals, mechanical feature signals, and environmental feature signals and form four-dimensional feature signals. The spatial path of the multi-scale spatio-temporal features deploys an improved ResNet-101 (a deep residual network) network. The bottom layer of the dual-path feature pyramid uses a 3×3 standard convolution kernel to extract fine-grained features such as steel bar spacing. The middle layer of the dual-path feature pyramid applies dilated convolution with a dilation rate of 2 to detect the overall deformation of the formwork support system. The top layer of the dual-path feature pyramid integrates deformable convolution layers to capture the features of special-shaped structures. The temporal path of the multi-scale spatio-temporal features constructs a 3DConvNeXt (an efficient convolutional neural network model) network, sets a sliding window in the temporal dimension to analyze the process timing deviation of the perception signals, and the size of the sliding window is dynamically configured according to the construction stage, and synchronously fuses the acceleration features of the construction machinery movement trajectory generated by optical flow estimation. The multi-scale spatio-temporal features achieve semantic aggregation through a cross-layer fusion formula as follows: , where, is the multi-scale feature tensor after fusion at the l-th layer, is the learnable weight coefficient of the i-th spatial feature layer at the l-th layer, is a 1×1 convolution operation, is the feature map of the i-th layer in the spatial path, is the fusion weight of the temporal path feature at the l-th layer, is a three-dimensional convolution operation, is the temporal feature of the l-th layer in the temporal path.

[0009] In one embodiment of the present invention, the convolution analysis module includes a historical knowledge base, and a two-way feature interaction mechanism is implemented when performing cross-modal attention fusion to fuse the four-dimensional feature signals and form a four-dimensional fusion signal. The high-dimensional semantic information of the visual feature signal in the four-dimensional fusion signal is output through the visual feature extraction network, and the temporal dynamic features of the radar feature signal and the mechanical feature signal in the four-dimensional feature signal are generated through the radar and mechanical sensor processing network. At the same time, a cross-attention association model of visual, radar and mechanical data is established, and then the dynamic allocation of feature importance is realized through a learnable attention weight matrix. In the fusion process, the environmental feature signal in the four-dimensional feature signal is introduced as a bias item to form a four-dimensional fusion signal containing the joint feature representation of visual, radar, sensor and environmental elements. The four-dimensional fusion signal is stored in the historical knowledge base.

[0010] In one embodiment of the present invention, a convolutional recurrent network with time memory function is used to capture the dynamic evolution law of the four-dimensional fusion signal and predict the construction process when performing construction prediction. When the deviation value between the prediction result and the actual detection result exceeds a preset deviation value threshold, a feature calibration process based on the historical knowledge base is started.

[0011] In one embodiment of the present invention, the structure assessment unit receives the four-dimensional fusion signal, and calculates the stress-displacement deviation rate of the beam component and the column component and forms a structural risk signal by comparing the real-time monitoring data of the finite element simulation model with the distributed optical fiber sensor. The material assessment unit periodically scans the components on the construction site through a handheld spectrometer and forms a material risk signal. The process detection unit receives the four-dimensional fusion signal, and detects the process parameters such as the steel bar spacing and the template overlap in real time by deploying a deep learning-based visual recognition algorithm, and forms a process risk signal. The environmental assessment unit constructs a multi-physical field coupling model by integrating the meteorological station and the micro-environment sensor network, and forms an environmental risk signal through the multi-physical field coupling model. The personnel assessment unit uses multi-perspective video analysis technology to identify violations such as not wearing safety equipment and entering dangerous areas, and forms a personnel risk signal in combination with UWB (wireless carrier communication) positioning data.

[0012] In one embodiment of the present invention, the risk assessment module weightedly fuses the structural risk signal, the material risk signal, the process risk signal, the environmental risk signal and the personnel risk signal to form a five-dimensional risk warning signal.

[0013] In an embodiment of the present invention, the early warning decision-making module receives the five-dimensional risk early warning signal from the risk assessment module, calculates the comprehensive risk index using the fuzzy comprehensive algorithm, and defines a three-level early warning mechanism. The three-level early warning mechanism includes first-level early warning, second-level early warning, and third-level early warning. The first-level early warning triggers the audible and visual alarm device at the construction site and sends a text message notification to the safety officer. The second-level early warning additionally issues an electronic rectification work order on the basis of the first-level early warning. The third-level early warning activates the emergency plan and broadcasts the comprehensive risk index through the radio.

[0014] In an embodiment of the present invention, the feedback optimization module receives the five-dimensional risk early warning signal and the perception signal collected by the data acquisition module after the risk assessment module completes risk processing. By comparing the change trends of the perception signal before and after risk processing and forming a risk optimization index, when the risk optimization index is lower than the preset risk optimization threshold, the feedback optimization module forms an optimization parameter signal through incremental learning technology and inputs the optimization parameter signal to the data acquisition module, the convolutional analysis module, the risk assessment module, and the early warning decision-making module.

[0015] A construction quality and safety risk management system provided by the present invention constructs a comprehensive management system integrating multi-modal perception, intelligent analysis, and closed-loop optimization, and focuses on solving the problems of low adaptability to the dynamic characteristics of the construction process and easy false alarms or missed alarms. Specifically, multi-dimensional data of the construction process are collected in real time through a sensor network deployed at the construction site, risk features are extracted using deep learning methods, risk assessment is combined with a knowledge graph in the construction field, risk early warning is carried out according to the risk assessment, and a dynamic optimization mechanism is established to achieve continuous improvement of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings without creative efforts based on these drawings.

[0017] Figure 1 It is a system architecture diagram of the construction quality and safety risk management system; Figure 2 It is a schematic diagram showing the working principle of the convolutional analysis module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0019] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Therefore, only the components related to the present invention are shown in the diagrams, rather than being drawn according to the number, shape, and size of the components in actual implementation. The types, quantities, and proportions of the components in actual implementation can be arbitrarily changed, and the layout type of the components may also be more complex.

[0020] In the following description, a large number of details are explored to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0021] The present invention relates to a construction quality and safety risk management system, which is mainly used in large-scale construction management scenarios. In the prior art, quality and safety management has always been the core link of engineering management. Traditional construction safety management mainly relies on manual inspections, regular inspections and static specification constraints. This method has obvious limitations: first, the coverage of manual inspections is limited, and it is difficult to achieve all-weather and all-round monitoring; second, the inspection results are greatly affected by personnel experience, and there is a subjective judgment bias; third, there is a time lag in risk identification and disposal response, and it is difficult to timely discover and respond to emergencies. With the continuous expansion of the scale of construction projects and the increasing complexity of construction processes, traditional management methods have been difficult to meet the needs of modern engineering construction. In terms of risk analysis, the prior art mainly adopts a rule-based method to judge the risk level by pre-set conditions. Although this method is simple to implement, it is difficult to handle complex correlations in construction scenarios. For example, in deep foundation pit construction, factors such as rainfall, soil moisture content, and support structure deformation affect each other, and simple threshold judgment cannot accurately assess the overall risk status. At the same time, the traditional method lacks effective use of historical data, and cannot optimize the risk identification model through continuous learning, resulting in system performance being difficult to improve over time. The early warning decision link also faces challenges. Most existing systems use a single alarm method and lack a hierarchical response mechanism, which can easily lead to overreaction or underreaction. For example, directly triggering the highest level alarm when a minor risk is detected may cause unnecessary downtime losses; while slow response to major risks may lead to serious accidents. In addition, the matching of emergency plans mostly relies on manual selection, lacks an intelligent recommendation mechanism, and it is difficult to quickly formulate the optimal disposal plan in an emergency. In terms of system optimization, existing technologies generally lack a closed-loop feedback mechanism. System parameters mostly rely on manual settings and cannot be automatically adjusted according to actual operating results. For example, key parameters such as sensor sampling frequency and risk judgment threshold often remain unchanged throughout the construction cycle, making it difficult to adapt to changes in risk characteristics at different construction stages. At the same time, the system's adaptability to new processes and new materials is insufficient, requiring frequent manual intervention and parameter adjustment, which increases the cost of use. In summary, the existing construction quality and safety management system has obvious deficiencies in data collection, risk analysis, early warning decision-making and system optimization, and is difficult to meet the needs of modern engineering construction. Developing a comprehensive management system that can achieve deep fusion of multimodal data, intelligent risk analysis, hierarchical early warning response and continuous optimization and improvement has become a technical problem that needs to be solved in the field of construction. It is in this context that this patent proposes an innovative construction quality and safety risk management system, aiming to overcome the shortcomings of existing technologies and provide more reliable safety guarantees for engineering construction.

[0022] See also Figure 1, shown is the construction quality and safety risk management system of the present invention. The construction quality and safety risk management system of the present invention includes a data acquisition module, a convolution analysis module, a risk assessment module, a warning decision module, and a feedback optimization module. The data acquisition module includes a vision unit, a radar unit, a mechanics unit, and an environmental perception unit, and generates perception signals therefrom. The convolution analysis module receives and processes the perception signals by using a multi-scale feature pyramid and a cross-modal attention mechanism to form a four-dimensional fusion signal. The risk assessment module includes a structure assessment unit, a material assessment unit, a process detection unit, an environmental assessment unit, and a personnel assessment unit. The risk assessment module conducts a comparative analysis on the four-dimensional fusion signal under the digital twin framework and forms a five-dimensional risk warning signal. The warning decision module receives the risk warning signal and forms a multi-level risk warning level signal. The feedback optimization module receives the risk warning level signal and optimizes the construction quality and safety risk management system through incremental learning technology. Among them, the convolution analysis module includes four steps: multi-modal data preprocessing and alignment, multi-scale spatio-temporal feature extraction, cross-modal attention fusion, and construction prediction.

[0023] As Figure 1As shown, the data acquisition module is the core of the perception layer of the system of the present invention, responsible for obtaining real-time status information of the construction site in all directions and multi-dimensions. This module consists of four subsystems: a vision unit, a radar unit, a mechanics unit, and an environmental perception unit. Each unit works in coordination to form a three-dimensional data acquisition network. In an embodiment of the present invention, the vision unit uses a high-resolution panoramic camera array, which is deployed at key positions of the construction site in a hexagonal topology. Each camera node is equipped with an automatic cleaning pan-tilt head and a dynamic exposure adjustment module to ensure the imaging quality under various lighting conditions. The camera array continuously acquires video data at a frequency of 5Hz and automatically switches to the event trigger mode when a moving target is detected, improving the capture accuracy of key frames. To expand the coverage range, the system is also equipped with a drone platform with a three-axis stabilized pod, which can flexibly inspect the high-altitude operation area. The vision data is transmitted to the edge computing node after being encoded and compressed by H.265, and the bit rate of a single video stream is controlled within 4Mbps to ensure the real-time nature of network transmission. The radar unit consists of a heterogeneous detection system composed of a 32-line lidar and a 77GHz millimeter-wave radar. The lidar scans the construction area at a frequency of 10Hz to generate three-dimensional point cloud data with centimeter-level accuracy, and the point cloud density reaches 200 points per square meter. The millimeter-wave radar uses a MIMO antenna array to achieve a coverage of a 120-degree horizontal field of view and a 20-degree vertical field of view, and can accurately track the movement trajectory of construction machinery. The data of the two radars are fused through the Kalman filtering algorithm to improve the robustness of moving target detection. In the tower crane operation area, a rotating millimeter-wave radar is additionally deployed to monitor the swing amplitude of the hook and the load status. The mechanics unit includes a composite measurement system of a distributed fiber Bragg grating sensor network and a piezoelectric vibration sensor. The fiber sensors are laid on the surface of the load-bearing structure in a serpentine path, with a spatial resolution of 2cm, and can monitor the stress and strain distribution inside the concrete in real time. The vibration sensor uses a resonance frequency adaptive adjustment technology, with a measurement range covering 0.1Hz to 1kHz, and is used to capture the structural vibration caused by the operation of construction machinery. The mechanics data is sampled at a frequency of 10Hz and transmitted to the gateway node through the LoRaWAN protocol, and the power consumption of a single sensor is controlled within 5mW to ensure long-term stable operation. The environmental perception unit integrates a micro weather station and a multi-gas detection module. The weather station is equipped with an ultrasonic anemometer and a laser disdrometer, which can accurately measure the wind speed (accuracy ±0.1m / s), wind direction (accuracy ±1°), and rainfall intensity. The multi-gas detection module uses an NDIR infrared sensor array to simultaneously monitor the concentrations of 8 kinds of hazardous gases such as CO and CH4, and the detection limit reaches the ppm level. The environmental data is uploaded once every 30 seconds, and the sampling frequency is automatically increased to 1 time per second when an abnormal value is detected.

[0024] Specifically, the data of each sensing unit is aligned through a unified spatio-temporal calibration system. The spatial coordinate system uses the BIM model as a reference benchmark, and the registration of visual and radar data is achieved through the checkerboard calibration method, with the registration error controlled within ±2 mm. The time synchronization is achieved through the high-precision network time protocol, and the clock deviation of each node does not exceed 1 ms. The preprocessed multi-modal data is encapsulated into standardized sensing signals, including metadata such as timestamps, spatial coordinates, data types, and measurement values, and is transmitted to the convolutional analysis module through an encrypted channel. To adapt to different construction scenarios, the data acquisition module supports dynamic configuration functions. During the earthwork construction stage, ground settlement and slope stability are key monitored, and the deployment density of mechanical sensors is increased; during the main structure construction period, the visual and radar coverage of the high-altitude operation area is increased; during the decoration and finishing stage, the gas detection ability of the environmental sensing unit is enhanced. The system is also equipped with a self-diagnosis function, which can monitor the operating status of each sensor in real time. When a device failure or data anomaly is detected, it automatically activates the backup node or adjusts the sampling strategy to ensure the continuity of data acquisition. Through the above design, the data acquisition module realizes the all-round perception of the construction site, providing a high-quality data basis for subsequent risk analysis. The deep fusion and intelligent processing of multi-modal data enable the system to more accurately identify and evaluate various risks during the construction process, providing reliable technical support for engineering safety management.

[0025] As Figure 2 shown, the convolutional analysis module is the core processing unit of the present invention, responsible for converting the original sensing data into a high-level feature representation that can be used for risk assessment. This module realizes the intelligent analysis of the construction scene through four steps: multi-modal data preprocessing and alignment, multi-scale spatio-temporal feature extraction, cross-modal attention fusion, and construction prediction. In an embodiment of the present invention, in the multi-modal data preprocessing and alignment stage, the system first performs standardization processing on the original data. The visual data is cut into spatio-temporal cubes of 256×256 pixels, and each cube contains 15 consecutive frames of images with a time span of 3 seconds (sampling rate of 5 Hz). The point cloud data is converted into a three-dimensional density grid with an accuracy of 0.1 m through voxelization, and each voxel stores the point cloud density and reflection intensity information. The mechanical sensing signals are standardized by Z-score to eliminate the dimension difference. The environmental data is converted into a multi-dimensional tensor format, including the temporal variations of parameters such as temperature, humidity, and wind speed. All data is aligned through a unified spatio-temporal calibration system: the spatial coordinate system uses the BIM model as a reference benchmark, and the registration of visual and radar data is achieved through the checkerboard calibration method, with the registration error controlled within ±2 mm; the time synchronization is achieved through the high-precision network time protocol, and the clock deviation of each node does not exceed 1 ms.

[0026] Specifically, the preprocessed data is encapsulated into standardized perception signals, including metadata such as timestamps, spatial coordinates, data types, and measurement values. In the multi-modal data preprocessing and alignment stage, the system first standardizes the original data. Visual data is cut into spatio-temporal cubes of 256×256 pixels, each cube containing 15 consecutive frames of images with a time span of 3 seconds (sampling rate of 5Hz). Point cloud data is transformed into a three-dimensional density grid with an accuracy of 0.1m through voxelization, and each voxel stores point cloud density and reflection intensity information. Mechanical sensing signals are standardized by Z-score to eliminate dimensional differences. Environmental data is converted into a multi-dimensional tensor format, including the temporal variations of parameters such as temperature, humidity, and wind speed. All data is aligned through a unified spatio-temporal calibration system: the spatial coordinate system uses the BIM model as a reference benchmark, and the registration of visual and radar data is achieved through the checkerboard calibration method, with the registration error controlled within ±2mm; time synchronization is achieved through the high-precision network time protocol, and the clock deviation of each node does not exceed 1ms. The preprocessed data is encapsulated into standardized perception signals, including metadata such as timestamps, spatial coordinates, data types, and measurement values. The multi-scale spatio-temporal feature extraction adopts a parallel dual-channel architecture. The spatial channel is based on the improved ResNet-101 network, which includes three feature extraction levels: the bottom layer uses 3×3 standard convolutional kernels to extract fine-grained features such as steel bar spacing and formwork joints; the middle layer introduces dual-path dilated convolution branches with dilation rates of 2 and 4 to expand the receptive field to capture the overall deformation of the formwork support system; the upper layer deploys deformable convolutional layers to adapt to the feature extraction of curved surface structures through learnable offsets. The time channel uses the 3DConvNeXt network, and a sliding window is set in the time dimension to analyze the process timing deviation: a 10-frame window (2-second span) is used to detect the compliance of the interval time during the concrete pouring stage, and it is adjusted to a 30-frame window (6-second span) during the decoration stage to monitor the construction rhythm. The multi-level features achieve semantic aggregation through the cross-layer fusion formula: , where, is the multi-scale feature tensor after fusion at the l-th layer, is the learnable weight coefficient of the i-th spatial feature layer at the l-th layer, is a 1×1 convolution operation, is the feature map of the i-th layer of the spatial channel, is the fusion weight of the time channel feature at the l-th layer, is a three-dimensional convolution operation, is the temporal feature of the l-th layer of the time channel. This formula uses backpropagation optimization to adjust the contribution degrees of features at each level.

[0027] The cross-modal attention fusion mechanism realizes the deep interaction of visual, physical sensing and environmental data. The visual feature stream extracts 1024-dimensional semantic information using the ResNeXt-50 architecture, and the physical sensing stream processes the time-series signals sampled at 10 Hz through one-dimensional depthwise separable convolution. A bidirectional attention model is established: the attention weight from vision to sensing is calculated as: , where is the visual feature query vector, and are the physical sensing key-value pair matrices. The reverse attention path establishes the correlation model from sensing to vision. The final fusion weight is generated by a two-layer fully connected network: , Normalized to the 0-1 interval through the Sigmoid (a sigmoid function) function to achieve dynamic weighted recombination of feature channels. Environmental data is injected into the fusion process as a bias term. For example, when the temperature sensor detects a high temperature above 35 °C, the weight coefficient of the features related to the degradation of material properties is enhanced. Construction prediction uses the ConvLSTM network to capture the dynamic evolution law of the construction process. Set a 30-second time step to match the typical process cycle, the hidden state dimension is 256, and the cell state uses the tanh activation function to control the information flow. Introduce the physical law constraint term: , where λ = 0.001 is the balance coefficient, σ is the predicted stress field, is the measured external force. The spatio-temporal consistency score is calculated as: , When < 0.7, start the feature recalibration process and call the historical normal mode data in the knowledge graph for interpolation compensation. Through the above processing flow, the convolutional analysis module realizes the intelligent perception and feature extraction of the construction scene, providing high-quality feature representations for subsequent risk assessment. The deep fusion of multi-modal data and the dynamic modeling of spatio-temporal features enable the system to more accurately predict potential risks during the construction process, providing reliable technical support for engineering safety management.

[0028] Such as Figure 1As shown, the risk assessment module is the core analysis unit of the present invention, responsible for quantitatively assessing various risks during the construction process. In an embodiment of the present invention, this module includes five subsystems: a structure assessment unit, a material assessment unit, a process detection unit, an environmental assessment unit, and a personnel assessment unit. Each unit conducts specialized analysis on specific types of risks and realizes multi-dimensional risk association through a knowledge graph. The structure assessment unit compares the finite element simulation results with the real-time monitoring data through digital twin technology. It calls the ANSYS solver every 15 minutes for static analysis, divides the BIM model into finite element meshes with a 10-cm accuracy, and calculates the stress and displacement distributions of each node. The distributed fiber optic sensor network monitors the actual structural response with a 2-cm spatial resolution and generates a three-dimensional strain field cloud map. The difference between the simulation and the measured data is calculated through a deviation matrix, and a secondary warning is triggered when the deviation of the beam member exceeds 15% or the deviation of the column member exceeds 10%. The unit also integrates a failure mode recognition function, which uses a graph attention network to analyze the structural node correlation and identify abnormal fluctuations in the key force transmission paths. The material assessment unit implements a dual verification mechanism of spectral feature matching and performance degradation prediction. A handheld X-ray fluorescence spectrometer is configured to scan the composition of the incoming steel bars every 2 hours to detect the content of harmful elements such as sulfur and phosphorus. After the spectral data is pre-filtered by Savitzky-Golay (a high-precision conformal filter), it is input into a one-dimensional convolutional neural network for classification and recognition, and the recognition accuracy reaches 95%. At the same time, the Arrhenius acceleration aging model is combined to predict the material life, and a material defect warning is generated when the measured strength value is lower than the lower limit of the predicted curve. The process detection unit deploys a collaborative verification framework based on a deep learning-based visual recognition algorithm and an RFID process tracking system. The visual algorithm uses an improved Mask R-CNN (a neural network) architecture to identify 23 process parameters such as the steel bar binding spacing and the formwork lap width, and the detection accuracy reaches 92%. The RFID system monitors the chronological order of construction nodes, and a process violation is determined when the detected process time deviation exceeds 30% of the standard or the visual parameter error rate is greater than 20%.

[0029] Specifically, the unit is also configured with a function for monitoring the concrete pouring process. By analyzing the data of vibration sensors, the uniformity of pouring is evaluated, and quality defects such as cold joints and honeycombs are identified. The environmental assessment unit constructs a multi-physical-field coupling model to evaluate the impact of external conditions on construction safety in real time. An integrated weather station and a micro-environment sensor network are used to collect parameters such as wind speed, rainfall, and temperature every 30 seconds. A computational fluid dynamics model is established to predict the wind load distribution, and when the wind speed exceeds level 6, a wind prevention emergency plan is activated. The rainfall impact assessment uses an improved Green-Ampt infiltration model to predict the impact of changes in soil moisture content on the stability of the foundation pit. The unit is also configured with a noise monitoring function, and when the equivalent continuous A-weighted sound level exceeds 85 dB for 8 consecutive hours, an occupational health warning is triggered. The personnel assessment unit implements a dual monitoring mechanism for behavior recognition and positioning tracking. An improved SlowFast network is used to analyze the video stream to identify 17 types of dangerous actions such as not wearing a safety helmet and not fastening the safety belt during high-altitude operations, and the recognition accuracy reaches 90%. The UWB positioning system tracks the personnel position with an accuracy of 10 cm, calculates the personnel density index in the intensive operation area, and generates an evacuation prompt when it exceeds 2 people per square meter. The unit also integrates a physiological monitoring function. By collecting heart rate variability (HRV) data through an intelligent bracelet, a fatigue warning is triggered when the HRV value is continuously 30% lower than the baseline for 2 hours. Each assessment unit realizes risk correlation analysis through a knowledge graph. For example, when the environmental unit detects heavy rainfall, it automatically correlates the foundation pit monitoring data of the structural unit to predict the impact of precipitation on the stability of the support system; when the material unit discovers steel bar corrosion, it correlates the binding quality data of the process unit to evaluate the impact on the overall performance of the structure. The knowledge graph uses graph neural networks for representation learning to construct a semantic relationship network containing "risk inducement - disposal measures - effect evaluation" to support the reasoning and analysis of complex risk scenarios. The output of the risk assessment module is a five-dimensional risk vector R=(r1, r2, r3, r4, r5), corresponding to the risk values of structure, material, process, environment, and personnel respectively. The value range of each dimension is [0,1]. The comprehensive risk index is calculated through a fuzzy comprehensive evaluation algorithm. The comprehensive risk index is used to trigger hierarchical warnings and provide a quantitative basis for subsequent decision-making support. Through the above design, the risk assessment module realizes a comprehensive assessment of construction quality and safety risks. The multi-dimensional and multi-scale analysis framework, combined with the semantic association ability of the knowledge graph, enables the system to accurately identify potential risks in complex construction scenarios and provides scientific and reliable decision-making support for project safety management.

[0030] Such as Figure 1As shown, the early warning decision module is the core response unit of the present invention, which is responsible for converting the risk assessment results into executable early warning instructions and disposal plans. This module realizes intelligent response to construction safety risks through four functional components: hierarchical early warning mechanism, risk visualization, emergency plan matching and rectification measures recommendation. In one embodiment of the present invention, the hierarchical early warning mechanism adopts a three-level response strategy, and triggers the corresponding level of early warning signals according to the comprehensive risk index CI. The first-level early warning (0.4≤CI<0.6) activates the on-site sound and light alarm device, including an 85dB rotating siren and a red strobe warning light, and pushes a text message notification to the project safety officer, which contains the risk type, location information and preliminary disposal suggestions. The second-level early warning (0.6≤CI<0.8) retains the first-level response measures, adds a tower crane cab HUD projection warning, displays the three-dimensional positioning information and safety operation instructions of the risk area, and automatically generates an electronic rectification work order, which is distributed to the relevant responsible person through a mobile terminal. Level 3 warning (CI ≥ 0.8) initiates the highest level of response: the linkage equipment control system performs an emergency shutdown, activates the drone cluster to conduct aerial monitoring of the risk area, and starts the emergency plan broadcast system to play evacuation instructions and safety precautions. The lifting of the warning signal requires double confirmation: the system automatically verifies that the risk indicator has returned to the normal range, and it can be revoked only after manual confirmation by the safety supervisor. The risk visualization interface uses augmented reality technology to overlay the BIM model with the real-time risk heat map. Users can adjust the viewing angle through gestures or voice commands to view the three-dimensional structural details of the risk area. The system also provides a historical risk trajectory playback function to support visual analysis of the risk evolution process in the past 24 hours. The emergency plan matching engine uses the BERT pre-trained model to calculate semantic similarity and retrieve the disposal plan that best matches the current risk scenario from the structured plan library. The plan library contains more than 500 standard emergency plans, each of which is annotated with metadata such as applicable scenarios, implementation steps, and required resources. The matching process comprehensively considers the similarity of risk characteristics (weight 0.6), historical disposal effects (weight 0.3) and resource availability (weight 0.1), outputs the top three plans with the highest matching degree and sorts them by comprehensive score. For example, when the risk of scaffold instability is detected, the system may recommend the following plans: (1) immediately evacuate the workers and set up a warning area (matching degree 0.92); (2) temporarily reinforce the support system (matching degree 0.85); (3) fully stop work for inspection (matching degree 0.78). The corrective measures recommendation algorithm is based on the reinforcement learning framework and optimizes policy selection through Q-learning. Define the reward function: , in, Risk is the risk reduction rate, and They are time cost and resource consumption respectively. The algorithm selects the plan with the largest reward value from the set of feasible measures and preferentially recommends intervention strategies with a comprehensive utility value higher than 7.5. For example, when it is detected that the early strength of concrete is insufficient, the system may recommend the following measures: (1) Extend the curing time and cover with heat preservation materials (R = 8.2); (2) Add early strength agent (R = 7.8); (3) Demolish and rebuild (R = 5.1). The recommended results are accompanied by detailed implementation plans, required resources and evaluation of expected effects. The early warning decision-making module is also equipped with a blockchain evidence storage function, and uses the Hyperledger Fabric framework to record the key decision-making process. Each early warning event generates a data block containing the timestamp, geographical location, disposal plan, and responsible person. The block header contains the Merkle tree root hash and the previous block hash value to ensure the immutability of the data. The smart contract sets automatic trigger conditions: when two consecutive blocks record similar risk events, automatically increase the relevant process detection frequency to 200%; when the same responsible person triggers more than 3 early warnings within 24 hours, start a special safety training process. Through the above design, the early warning decision-making module realizes the intelligent and hierarchical response to construction safety risks. The multi-level early warning mechanism, intuitive risk visualization, intelligent plan matching and optimized rectification recommendations significantly improve the efficiency and effectiveness of construction safety management. The actual application of the system shows that compared with the traditional management method, the risk response time is shortened by 67%, and the false alarm rate is reduced by 82%, providing a more reliable safety guarantee for engineering construction.

[0031] As Figure 1 shown, the feedback optimization module constructs a closed-loop learning system. The disposal effect evaluation unit compares the changes in monitoring data before and after risk disposal, and calculates indicators such as the improvement degree of structural deviation and the material repair rate. The incremental learning framework uses a dynamic replay buffer to store historical features, and the elastic weight consolidation technology constrains the update amplitude of key weights through a parameter importance matrix. The knowledge graph evolution subsystem scans new accident reports daily, extracts the "risk cause - disposal measure - effect evaluation" triple, and updates the node association strength through a graph attention mechanism. The optimization signal distribution controller dynamically adjusts parameters according to the performance indicators of each module. For example, when the material evaluation error exceeds 5%, the spectral acquisition frequency is increased to 3 times per hour. The blockchain evidence storage unit records each optimization process and uses a threshold signature mechanism to ensure the immutability of the data.

[0032] In an embodiment of the present invention, the system significantly improves the efficiency of construction safety management. In a super high-rise building project, the system detected abnormal displacement rates of the core tube shear wall formwork during the concrete pouring stage: the lidar monitored that the east formwork shifted 3.2 mm per hour, and the fiber optic sensor showed that the stress value in the corresponding area exceeded the limit by 15%. The spatio-temporal convolutional analysis engine identified the accelerating trend of displacement, and the knowledge graph associated with historical cases to judge the risk of instability of the support system, triggering a secondary warning. On-site inspection found that the torque of the scaffolding fasteners was insufficient. After emergency reinforcement, the displacement rate returned to normal. The entire process from risk detection to disposal took only 47 minutes, which was 6 times more efficient than the traditional manual inspection method. In another embodiment of the present invention, the construction scenario occurred during the steel structure hoisting operation. The millimeter-wave radar detected that the gust wind speed reached 12 m / s instantaneously, and the environmental assessment sub-module predicted that the swing amplitude of the tower crane jib would exceed the safety threshold. The system immediately initiated a level-three warning: stop the hoisting operation, retract the hook, and activate the windproof anchoring device. At the same time, the personnel assessment unit found through video analysis that two workers did not correctly fasten their safety belts and sent a special rectification notice. In this incident, the system predicted the risk 9 minutes in advance and avoided a possible major accident.

[0033] A building construction quality and safety risk management system of the present invention collects multi-dimensional data of the construction process in real time through a sensor network deployed at the construction site, uses deep learning methods to extract risk characteristics, combines with a knowledge graph in the construction field for risk assessment, conducts risk early warning based on the risk assessment, and establishes a dynamic optimization mechanism to achieve continuous improvement of the system.

[0034] Therefore, through the building construction quality and safety risk management system of the present invention, the problems of low adaptability to the dynamic characteristics of the construction process and easy generation of false alarms or missed alarms can be solved.

[0035] The above embodiments are only illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A construction quality safety risk management system, characterized in that: include: A data acquisition module, wherein the data acquisition module includes a visual unit, a radar unit, a mechanical unit and an environmental perception unit, and generates a perception signal therefrom; A convolution analysis module receives and processes the perception signal using a multi-scale feature pyramid and a cross-modal attention mechanism to form a four-dimensional fusion signal; A risk assessment module, wherein the risk assessment module includes a structure assessment unit, a material assessment unit, a process detection unit, an environment assessment unit, and a personnel assessment unit. The risk assessment module compares and analyzes the four-dimensional fusion signal under the digital twin framework and forms a five-dimensional risk warning signal; An early warning decision module receives the risk early warning signal and generates a multi-level risk early warning level signal; A feedback optimization module, which receives the risk warning level signal and optimizes the construction quality and safety risk management system through incremental learning technology; Among them, the convolution analysis module includes four steps: multimodal data preprocessing and alignment, multi-scale spatiotemporal feature extraction, cross-modal attention fusion, and construction prediction.

2. The construction quality safety risk management system according to claim 1 is characterized in that: The perception signals include visual perception signals, radar perception signals, mechanical perception signals and environmental perception signals. The visual unit includes a collaborative acquisition architecture of a high-resolution panoramic camera array and a mobile UAV platform, wherein the panoramic camera array is deployed at a commanding height in the construction area in a hexagonal topology structure, and each camera node is equipped with an automatic cleaning gimbal and a dynamic exposure adjustment module. The UAV platform is equipped with a three-axis stabilization pod and a laser rangefinder. The visual unit forms the visual perception signal, and the acquisition frequency of the visual perception signal is dynamically adjusted according to the construction stage; the radar unit includes a laser radar and a millimeter-wave radar, the laser radar outputs a three-dimensional density map, the millimeter-wave radar outputs a wide-angle coverage map, and the three-dimensional density The figure and the wide-angle coverage figure describe the Kalman filtering algorithm for fusion of moving target trajectories and formation of radar perception signals; the mechanical unit includes a distributed fiber optic sensor network and a piezoelectric vibration sensor, the distributed fiber optic sensor is arranged on the surface of the load-bearing structure in a serpentine path, the pressure point vibration sensor adopts a resonant frequency adaptive adjustment technology, and the mechanical unit forms the mechanical perception signal; the environmental perception unit includes a micro-meteorological station and a multi-gas detection device, wherein the micro-meteorological station includes an ultrasonic anemometer and a laser raindrop spectrometer, the multi-gas detection device detects the concentration of hazardous gases through an NDIR (non-dispersive infrared gas analysis technology) infrared sensor array, and the environmental perception unit forms the environmental perception signal.

3. The construction quality safety risk management system according to claim 2 is characterized in that: The multimodal data preprocessing and alignment include spatiotemporal calibration and data formatting. When performing the spatiotemporal calibration, the convolution analysis module receives the perception signal, and realizes the spatial coordinate system alignment of the visual perception signal and the radar perception signal through a checkerboard calibration plate. The timestamps of each unit under the data acquisition module are synchronized to microsecond accuracy using the NTP (network time) protocol. The visual data is cut into a 256×256 pixel spatiotemporal cube sequence, and every 15 frames constitute an analysis unit. The point cloud data is voxelized to generate a three-dimensional density grid with an accuracy of 0.1m³, and form a visual feature signal and a radar feature signal. When performing the data formatting, the mechanical perception signal performs Z-score (standard score or zero mean unit variance) standardization to eliminate dimensional differences and form a mechanical feature signal. The environmental perception signal is converted into a structured tensor format to form an environmental feature signal.

4. The construction quality safety risk management system according to claim 3 is characterized in that: When extracting the multi-scale spatiotemporal features, a parallel dual-path feature pyramid architecture is used to extract the visual feature signal, the radar feature signal, the mechanical feature signal and the environmental feature signal, and form a four-dimensional feature signal. The spatial path of the multi-scale spatiotemporal features deploys an improved ResNet-101 (a deep residual network) network. The bottom layer of the dual-path feature pyramid uses a 3×3 standard convolution kernel to extract fine-grained features such as the spacing between steel bars. The middle layer of the dual-path feature pyramid uses a dilated convolution with a dilated rate of 2 to detect the overall deformation of the template support system. The high layer of the dual-path feature pyramid integrates a deformable convolution layer to capture the features of the special-shaped structure. The time path of the multi-scale spatiotemporal features constructs a 3D ConvNeXt (an efficient convolutional neural network model) network, and a sliding window is set in the time dimension to analyze the process timing deviation of the perception signal. The size of the sliding window is dynamically configured according to the construction stage, and the acceleration features of the motion trajectory of the construction machinery generated by the optical flow estimation are synchronously fused. The multi-scale spatiotemporal features are semantically aggregated through a cross-layer fusion formula, as follows: ， in, is the multi-scale feature tensor after fusion of the lth layer, is the learnable weight coefficient of the i-th spatial feature layer at the l-th layer, is a 1×1 convolution operation, is the i-th layer feature map of the spatial pathway, is the fusion weight of the temporal path feature at layer l, is a three-dimensional convolution operation, is the lth layer timing feature of the time path.

5. The construction quality safety risk management system according to claim 4 is characterized in that: The convolution analysis module includes a historical knowledge base, and a two-way feature interaction mechanism is implemented when performing the cross-modal attention fusion. The four-dimensional feature signals are fused to form a four-dimensional fusion signal. The high-dimensional semantic information of the visual feature signal in the four-dimensional fusion signal is output through a visual feature extraction network. The temporal dynamic features of the radar feature signal and the mechanical feature signal in the four-dimensional feature signal are generated through a radar and mechanical sensor processing network. At the same time, a cross-attention association model of visual, radar and mechanical data is established, and then the dynamic allocation of feature importance is realized through a learnable attention weight matrix. In the fusion process, the environmental feature signal in the four-dimensional feature signal is introduced as a bias item to form the four-dimensional fusion signal containing the joint feature representation of visual, radar, sensor and environmental elements. The four-dimensional fusion signal is stored in the historical knowledge base.

6. The construction quality safety risk management system according to claim 5 is characterized in that: When performing the construction prediction, a convolutional recurrent network with time memory function is used to capture the dynamic evolution law of the four-dimensional fusion signal and predict the construction process. When the deviation value between the prediction result and the actual detection result exceeds the preset deviation value threshold, the feature calibration process based on the historical knowledge base is started.

7. The construction quality safety risk management system according to claim 1 is characterized in that: The structural assessment unit receives the four-dimensional fusion signal, and calculates the stress-displacement deviation rate of the beam component and the column component and forms a structural risk signal by comparing the finite element simulation model with the real-time monitoring data of the distributed optical fiber sensor. The material assessment unit periodically scans the components on the construction site through a handheld spectrometer and forms a material risk signal. The process detection unit receives the four-dimensional fusion signal, and detects process parameters such as steel bar spacing and template overlap in real time by deploying a deep learning-based visual recognition algorithm, and forms a process risk signal. The environmental assessment unit constructs a multi-physical field coupling model by integrating a meteorological station and a micro-environment sensor network, and forms an environmental risk signal through the multi-physical field coupling model. The personnel assessment unit uses multi-perspective video analysis technology to identify violations such as not wearing safety equipment and entering dangerous areas, and forms a personnel risk signal in combination with UWB (wireless carrier communication) positioning data.

8. The construction quality safety risk management system according to claim 7 is characterized in that: The risk assessment module weightedly fuses the structural risk signal, the material risk signal, the process risk signal, the environmental risk signal and the personnel risk signal to form the five-dimensional risk warning signal.

9. The construction quality safety risk management system according to claim 1, characterized in that: The early warning decision module receives the five-dimensional risk early warning signal of the risk assessment module, and adopts a fuzzy comprehensive algorithm to calculate a comprehensive risk index, and defines a three-level early warning mechanism, which includes a first-level early warning, a second-level early warning and a third-level early warning. The first-level early warning triggers the sound and light alarm device at the construction site and pushes a text message notification to the safety officer. The second-level early warning adds an electronic rectification work order on the basis of the first-level early warning. The third-level early warning activates the emergency plan and broadcasts the comprehensive risk index.

10. The construction quality safety risk management system according to claim 1, characterized in that: The feedback optimization module receives the five-dimensional risk warning signal and receives the perception signal collected by the data acquisition module after the warning decision module completes the risk processing, and forms a risk optimization index by comparing the change trend of the perception signal before and after the risk processing. When the risk optimization index is lower than the preset risk optimization threshold, the feedback optimization module forms an optimization parameter signal through the incremental learning technology and inputs the optimization parameter signal to the data acquisition module, the convolution analysis module, the risk assessment module and the warning decision module.

Images