A highway engineering construction safety data acquisition system and method

Abstract

The application discloses a kind of highway engineering construction safety data acquisition system and method, belong to highway engineering construction safety monitoring technical field.System uses "end-edge-cloud" three-layer collaborative architecture: perception terminal layer is through multi-source sensor and collects environment, vibration, positioning, video and physiological data;Edge intelligent layer is deployed in edge computing gateway, and original data is excluded, feature extraction and light compression coding are carried out, and millisecond level local early warning is realized by running lightweight risk preliminary judgment model;Cloud analysis layer receives edge upload data, and runs depth analysis model to carry out global risk research and model iteration update.Collecting terminal is realized sub-microsecond level time synchronization by PTP protocol, and unified space coordinate association is established, and the multi-modal data fusion of "same time, same place" is realized.The application has the advantages of high data compression rate, strong real-time, unified space-time benchmark, etc., and can effectively improve the real-time and reliability of construction safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of highway construction safety monitoring technology, specifically to a highway construction safety data acquisition system and method, and more particularly to a highway construction safety data acquisition system and method based on edge intelligent preprocessing, spatiotemporal synchronization and fusion acquisition of multi-source heterogeneous data, and lightweight data transmission protocol. Background Technology

[0002] Highway engineering construction involves complex environments, wide work areas, and numerous personnel and machinery, resulting in numerous and widespread safety risks. This places extremely high demands on the collection and real-time monitoring of safety data during the construction process. Currently, highway engineering construction safety data collection mainly relies on deploying various sensors (environmental sensors, vibration sensors, positioning terminals, video monitoring equipment, etc.) at the construction site. The collected data is then uploaded to a cloud platform via wireless network for analysis and processing.

[0003] However, the existing technology still has the following shortcomings: (i) The passivity and redundancy of data acquisition. Most existing systems rely on passive data acquisition at fixed frequencies, with sensors continuously transmitting data. This results in a large amount of invalid or low-value data consuming transmission bandwidth and storage resources, leading to low effective information density. Furthermore, the data from various subsystems (environmental monitoring, personnel positioning, and video surveillance) are isolated, lacking a unified data acquisition protocol and collaborative mechanism.

[0004] (ii) Weak edge intelligent processing capabilities. The vast majority of existing systems adopt an "edge-to-cloud" direct transmission mode, where all raw data collected by sensors is uploaded to the cloud for processing. This mode results in limited real-time performance (due to delays in round-trip transmission between the cloud and the cloud) and consumes a large amount of network bandwidth. Although some literature mentions the application of edge computing in construction monitoring, it mainly focuses on data forwarding or simple preprocessing, and has not yet formed a systematic edge intelligent preprocessing architecture.

[0005] (iii) Lack of unified spatiotemporal reference for multi-source heterogeneous data. Various sensors deployed at the construction site come from different manufacturers and use different communication protocols, resulting in different time references and spatial coordinate systems for data acquisition. Existing technologies lack a unified time synchronization mechanism and spatial coordinate association model, making it difficult to effectively correlate multi-source data at the same time and location, which seriously restricts the accuracy of subsequent multimodal data fusion analysis and risk assessment.

[0006] (iv) The data transmission protocol lacks a tiered protection mechanism. The existing system adopts a uniform transmission strategy for all types of data, failing to differentiate data based on its security relevance and urgency. Emergency warning data and routine monitoring data are transmitted in the same channel, which cannot guarantee the real-time performance and reliability of critical data.

[0007] (v) Insufficient data quality assurance mechanisms. Sensors operate in harsh outdoor environments for extended periods, making them susceptible to data drift or malfunctions due to factors such as vibration, temperature and humidity changes, and dust. Existing systems lack online assessment and credibility labeling mechanisms for collected data, potentially leading to the misuse of low-quality data for security decisions. Summary of the Invention

[0008] (a) Purpose of the invention: The purpose of this invention is to provide a highway engineering construction safety data acquisition system and method. It aims to comprehensively improve the real-time performance, reliability, data quality, and transmission efficiency of highway engineering construction safety data acquisition by constructing a three-layer collaborative architecture of "edge-cloud". This architecture achieves anomaly removal, feature extraction, compression encoding, and lightweight initial risk assessment at the edge; differentiated transmission guarantees based on data importance classification at the transmission layer; and accurate spatiotemporal synchronization and unified spatial coordinate association of multi-source heterogeneous data at the acquisition layer. It also addresses the technical problems existing in current highway engineering construction safety data acquisition systems, such as weak intelligent processing capabilities at the edge, inconsistent spatiotemporal benchmarks for multi-source heterogeneous data, lack of hierarchical data transmission guarantees, and the absence of data quality assessment mechanisms.

[0009] (II) Technical Solution: A highway engineering construction safety data acquisition system, adopting a three-layer collaborative architecture of end-edge-cloud, includes: The sensing terminal layer is deployed along the entire highway construction line. The sensing terminal layer includes a data acquisition terminal, which is used to collect multi-source heterogeneous raw data at a preset sampling frequency. The data acquisition terminal has a built-in high-precision real-time clock module. The edge intelligence layer, deployed at the construction site's edge computing gateway, includes: an edge data preprocessing module for outlier removal, format standardization, timestamp marking, and keyframe extraction and target detection of the video stream; an edge feature extraction and compression encoding module for time-domain / frequency-domain feature extraction and lightweight encoding compression of the preprocessed data; an edge lightweight risk assessment module for running a lightweight risk assessment model to perform real-time risk level assessment and millisecond-level local early warning of the preprocessed data; and an edge communication protocol adaptation module for enabling data protocol access and bidirectional communication with the cloud platform. The cloud analytics layer, deployed on a central server or public cloud platform, is used to receive data uploaded by edge nodes, run deep analytics models for global situational awareness and cross-segment collaborative risk assessment, complete iterative training of the model, and distribute the updated model to edge nodes.

[0010] Preferably, the sensing terminal layer includes: environmental sensor nodes fixedly deployed along the construction section for collecting data on temperature and humidity, wind speed and direction, dust concentration, and concentrations of harmful gases including carbon monoxide and nitrogen oxides; structural vibration sensors fixedly installed on key structures for collecting data on structural vibration frequency and amplitude; personnel positioning terminals and smart wearable terminals worn by construction workers for collecting high-precision location information and physiological parameters including heart rate, body temperature, and body posture; the positioning accuracy of the high-precision location information is no greater than 30cm; the body posture physiological parameters include triaxial acceleration and angular velocity data; a video acquisition terminal for collecting real-time video streams from the construction site; and mechanical equipment status sensors fixedly installed on the machinery for collecting mechanical operating parameters; the machinery includes road rollers, pavers, and tower cranes; the operating parameters include the vibration frequency and compaction speed of the road rollers, the paving thickness and paving temperature of the pavers, and the lifting weight and tilt angle of the tower cranes.

[0011] Preferably, the edge computing gateway has a built-in GPS / BeiDou time synchronization module to obtain a high-precision UTC time reference, and uses it as a PTP master clock to distribute synchronization time to the acquisition terminals within the coverage area via Ethernet. The acquisition terminals have built-in PTP hardware clocks to achieve sub-microsecond time synchronization. The acquisition terminals obtain precise spatial coordinates through RTK positioning, and the video acquisition terminals establish a mapping relationship between image pixel coordinates and three-dimensional spatial coordinates through camera calibration. All acquired data is appended with a dual-association index of timestamp + spatial coordinates.

[0012] Preferably, the edge computing gateway covers a construction area of ​​500 meters to 1 kilometer, and the acquisition terminal communicates with the edge computing gateway via wired Ethernet or 5G wireless network; the edge communication protocol adaptation module supports the conversion of Modbus, CAN, RS485 industrial protocols to MQTT standardized protocols, and adopts a publish / subscribe model to decouple the device from the cloud platform.

[0013] A method for collecting safety data during highway construction, implemented using a highway construction safety data collection system, includes the following steps: S1. Collect multi-source heterogeneous data through the sensing terminal layer at a preset sampling frequency, and add a local hardware timestamp when the data is generated. S2. After receiving the raw data, the edge computing gateway sequentially performs outlier removal, feature extraction, and lightweight encoding compression to obtain compressed feature data. S3. Under a unified spatiotemporal reference, sensor data are aligned to the same time axis according to a unified timestamp and mapped to a unified three-dimensional spatial coordinate system to form a structured fusion data package at the same time and location. S4. Edge nodes classify the fused data according to security relevance and urgency, and use the MQTT protocol to transmit the data to the cloud platform with differentiated QoS levels based on the data level. After receiving data, the S5 cloud platform runs a deep analysis model to perform global situational awareness and cross-segment collaborative risk assessment, and regularly retrains and updates the edge model.

[0014] Preferably, in S2, outlier removal specifically involves: using a statistical threshold-based filtering method for numerical data, marking data exceeding a preset normal range as outliers and removing them; using the isolated forest algorithm for outlier detection for time-series data; and using a lightweight CNN model for target detection for video data, identifying and discarding invalid or blurry frames.

[0015] Preferably, in S2, feature extraction includes: extracting time-domain features such as mean, standard deviation, peak value, and root mean square value; extracting frequency-domain features of the main frequency component and spectral energy distribution through fast Fourier transform; extracting statistical features such as skewness, kurtosis, and coefficient of variation; and extracting safety semantic features such as helmet wearing status, personnel fall behavior, and fireworks events from the video stream.

[0016] Preferably, in S2, lightweight coding compression includes: using differential coding to reduce temporal data redundancy, using fixed-point quantization to compress data precision for floating-point numbers, and using optimized coding formats to compress video keyframes.

[0017] Preferably, in S4, the data is categorized according to safety relevance and urgency: Level I Emergency Data includes personnel falls, electronic fence intrusion, structural vibration exceeding thresholds, and hazardous gas exceeding standards; structural vibration exceeding thresholds refers to vibration amplitude exceeding the limit specified in the standard, or exceeding 80% of the design value; hazardous gas exceeding standards refers to carbon monoxide concentration exceeding 24 ppm or nitrogen dioxide concentration exceeding 3 ppm; Level II Important Data includes equipment malfunctions, sudden environmental changes, and personnel physiological abnormalities; equipment malfunctions include equipment temperature exceeding 85℃ or vibration frequency deviating from the normal operating range by ±15%; sudden environmental changes refer to wind speed... The data must be classified as follows: Level 1 data changes exceeding Level 4 within 10 minutes or temperature changes exceeding 10°C within 30 minutes; Level 2 data is defined as a heart rate consistently exceeding 120 beats / min or below 50 beats / min for 1 minute, or a body temperature exceeding 38.5°C; Level 3 data is classified as routine data, including routine environmental monitoring data, normal vibration data, and routine video streams; Level 3 data is transmitted using the MQTT protocol with differentiated QoS levels, with Level 1 data transmitted exactly once using QoS2, Level 2 data transmitted at least once using QoS1, and Level 3 data transmitted at most once using QoS0, and then compressed and uploaded in batches according to a preset period. Edge nodes only upload Level 1 and Level 2 data in real time.

[0018] Preferably, in S5, the cloud platform performs real-time verification and confirmation of Level I data and triggers emergency plans; it extracts the spatiotemporal and semantic features of multi-source data from the runtime Transformer fusion model for Level II and Level III data, and generates dynamic risk scores; based on the accumulated full data, the cloud regularly retrains and optimizes the lightweight risk assessment model at the edge, and distributes the updated model parameters to the edge computing nodes, forming a closed loop of data collection → edge processing → cloud analysis → model optimization → edge update.

[0019] (III) Beneficial Effects: Therefore, compared with the prior art, the present invention has the following beneficial effects: 1. Significantly reduces data transmission volume and bandwidth consumption. By performing outlier removal, feature extraction, and lightweight encoding compression at the edge, the original high-dimensional data is compressed into low-dimensional feature vectors, achieving a data compression rate of 75%, which greatly reduces data transmission bandwidth consumption and cloud storage pressure.

[0020] 2. Achieve millisecond-level local real-time early warning. A lightweight risk assessment model runs at the edge, performing real-time risk level evaluation on preprocessed data. Millisecond-level local early warning can be achieved without waiting for cloud transmission, significantly improving the response speed of security incidents.

[0021] 3. Ensuring the spatiotemporal consistency and fusion of multi-source heterogeneous data. Sub-microsecond time synchronization accuracy is achieved through GPS / PTP precise time synchronization protocol, and centimeter-level spatial coordinate unification is achieved through RTK positioning and camera calibration. A dual-link index of "timestamp + spatial coordinates" is constructed, realizing the association and collection of multimodal data "at the same time and in the same location", providing a high-quality data foundation for subsequent review and risk analysis.

[0022] 4. Ensure reliable transmission of critical data. Through data importance classification and differentiated QoS guarantees using the MQTT protocol, Level I urgent data is transmitted precisely once using QoS2 to ensure no loss or duplication; Level II important data is transmitted at least once using QoS1 to guarantee integrity; and Level III routine data is transmitted at most once using QoS0 to reduce network overhead, thus achieving on-demand allocation of transmission resources.

[0023] 5. Establishing a closed-loop learning feedback mechanism through cloud-based collaboration. The cloud-based system periodically retrains the edge model based on full data and distributes updated parameters, enabling continuous iterative evolution of the edge-based risk assessment model. This allows the system to adapt to changes in safety risk characteristics under different construction stages and environmental conditions.

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall architecture of a highway engineering construction safety data acquisition system and method according to the present invention; Figure 2 This is a schematic diagram of the internal module composition of the edge computing gateway of the highway engineering construction safety data acquisition system and method of the present invention; Figure 3 This is an overall flowchart of a highway engineering construction safety data acquisition system and method according to the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the following description is provided in conjunction with the appendix. Figures 1-3 The embodiments of the present invention will be further described in detail below.

[0027] Example 1: This embodiment provides a highway engineering construction safety data acquisition system.

[0028] like Figure 1 As shown, this embodiment adopts a three-layer collaborative architecture of "device-edge-cloud", including a sensing terminal layer, an edge intelligence layer, and a cloud analysis layer.

[0029] (a) Deployment of the sensing terminal layer: A sensing terminal layer was deployed on a 5-kilometer section of a highway reconstruction and expansion project, from K10+000 to K15+000. The specific deployment is as follows: An environmental sensor node was deployed every 200 meters along the construction section, totaling 25 nodes. Each node integrated a temperature and humidity sensor, a wind speed and direction sensor, a PM2.5 / PM10 dust sensor, and a CO gas sensor. Twenty structural vibration sensors, using piezoelectric accelerometers, were installed at key structural locations on the three bridges and two tunnels within the section, with a sampling frequency of 1kHz. Each of the 120 construction workers was equipped with a personnel positioning terminal (UWB positioning tag, positioning accuracy 30cm) and a smart wearable terminal (integrating heart rate monitoring, body temperature monitoring, and posture sensors). A video acquisition terminal (high-definition network camera) was installed every 100 meters along the construction section, totaling 50, and two inspection drones were also deployed for high-altitude panoramic inspections. Mechanical equipment status sensors were installed on three tower cranes, five road rollers, and four pavers to collect parameters such as load, vibration frequency, and temperature. All acquisition terminals have built-in high-precision real-time clock modules and support the PTP protocol.

[0030] (II) Deployment of the edge intelligence layer: Ten edge computing gateways (using an industrial-grade ARM architecture edge AI computing platform, equipped with a quad-core processor, 8GB of memory, and an NPU acceleration unit) were deployed every 500 meters along the construction section. Each edge computing gateway covered a 500-meter construction area and connected to the various data acquisition terminals in the sensing terminal layer via wired Ethernet or a 5G wireless network.

[0031] Each edge computing gateway integrates a GPS / BeiDou time synchronization module, acquiring a high-precision UTC time reference via satellite signals. The GPS PPS pulse signal accuracy reaches ±50ns. The edge computing gateway acts as the PTP master clock, distributing synchronization time to all acquisition terminals within its coverage area via Ethernet. Each acquisition terminal has a built-in PTP hardware clock, achieving sub-microsecond time synchronization accuracy. For terminals that do not support PTP (such as some older sensors), the NTP protocol is used for millisecond-level time synchronization.

[0032] Each acquisition terminal obtains precise spatial coordinates via RTK positioning during deployment, achieving centimeter-level accuracy. Video surveillance equipment uses calibration methods to calibrate cameras, establishing a mapping relationship between image pixel coordinates and three-dimensional spatial coordinates.

[0033] The edge computing gateway integrates the following modules (such as...) Figure 2 (as shown) Edge data preprocessing module: Receives raw data uploaded from each terminal, and processes numerical data using 3D printing technology. Outlier removal is performed according to the principle. The isolated forest algorithm is used to detect outliers in time series data. The lightweight YOLOv5s model is used to detect objects in video streams (identifying helmet wearing, people falling, fireworks events, etc.). Invalid or blurry frames are identified and discarded. Edge feature extraction and compression coding module: Extracts time-domain features (mean, standard deviation, peak value, root mean square value), frequency-domain features (main frequency component and spectral energy distribution extracted by FFT), and statistical features (skewness, kurtosis, coefficient of variation) from the cleaned effective data. Compressed coding is performed using differential coding and fixed-point quantization, achieving a data compression rate of 75%. Lightweight edge risk assessment module: Runs a lightweight risk assessment model based on decision tree to assess the risk level (normal / concern / early warning / alarm) of preprocessed data in real time, and achieves millisecond-level local early warning - when the vibration amplitude is detected to exceed the preset threshold or personnel enter the electronic fence restricted area, the edge computing gateway issues a local audible and visual alarm within 50ms; Edge communication protocol adaptation module: realizes the conversion of industrial protocols such as Modbus, CAN, and RS485 to the standardized MQTT protocol.

[0034] (III) Deployment of the cloud analytics layer: The cloud-based analytics layer is deployed on a central cloud server and configured with GPU-accelerated computing resources. The cloud platform receives compressed feature data and anomaly event fragments uploaded from various edge computing gateways. It uses a runtime Transformer fusion model to extract spatiotemporal and semantic features from the multi-source data to generate dynamic risk scores. It integrates data from across the entire network to conduct cross-section collaborative risk assessment and trend prediction. It also updates the digital twin model to achieve a three-dimensional visualization of construction safety.

[0035] The cloud platform retrains and optimizes the lightweight risk assessment model at the edge every week, and distributes the updated model parameters to each edge computing gateway through a secure channel.

[0036] Example 2: This embodiment provides a method for collecting safety data during highway construction.

[0037] This embodiment, based on Embodiment 1 above, provides a method for collecting safety data during highway construction. For example... Figure 3 As shown, the method includes the following steps: Step S1: Multi-source heterogeneous data acquisition: The sensing terminal layer collects various types of data at a preset sampling frequency. Environmental sensor nodes collect temperature, humidity, wind speed and direction, PM concentration, and harmful gas concentration data every second. Structural vibration sensors continuously collect vibration acceleration data at a sampling frequency of 1 kHz. Personnel positioning terminals report UWB location coordinates every 100 ms. Video acquisition terminals collect real-time video streams of the construction site at a frame rate of 25 fps. Smart wearable terminals collect heart rate, body temperature, and body posture data every second. Mechanical equipment status sensors collect load, vibration frequency, and temperature data every 500 ms. Each acquisition terminal adds a local hardware timestamp when the data is generated.

[0038] Step S2: Edge data preprocessing: After receiving the raw data uploaded by each terminal within its coverage area, each edge computing gateway performs the following sub-steps: S21: Outlier removal. For numerical data (environmental parameters, vibration data, etc.), a 3x3 outlier removal method is used. Outlier detection is performed in principle, and data exceeding the mean ± 3 standard deviations are marked as outliers and removed. For time-series data, the Isolation Forest algorithm is used for outlier detection, with an outlier ratio set at 5%, identifying anomalous jumps caused by sensor malfunctions or environmental interference (such as lightning strikes or strong electromagnetic pulses). For video data, a lightweight YOLOv5s model is used for target detection, identifying and discarding invalid frames caused by lens occlusion, overexposure due to strong light, or motion blur.

[0039] S22: Feature Extraction. Feature extraction is performed on the cleaned, valid data. For vibration data, time-domain features such as mean, standard deviation, peak value, peak-to-peak value, and root mean square value are extracted within a 1-second time window. The dominant frequency component (range 0-500Hz) and spectral energy distribution are extracted using FFT. For environmental data, statistical features including skewness, kurtosis, and coefficient of variation are extracted. For video streams, safety semantic features are extracted, including helmet wearing status (wearing / not wearing), personnel fall behavior (normal / fall), and smoke / fire events (no / present). Detection results are output every 1 second.

[0040] S23: Lightweight Encoding and Compression. The extracted feature data is compressed and encoded. Differential encoding is used for temporal feature data, storing only the difference between the current value and the previous value to reduce redundant information. Floating-point features are quantized using 8-bit fixed-point quantization, compressing the data precision from 32-bit floating-point numbers to 8-bit integers. Video keyframes are compressed using H.265 encoding. After the above processing, the original data compression rate reaches 75%.

[0041] Step S3: Spatiotemporal synchronization and fusion of multi-source heterogeneous data: S31: Establishment of a unified time base. The edge computing gateway obtains the UTC time base through the GPS / BeiDou time synchronization module. The accuracy of the GPS PPS pulse signal is ±50ns. The edge computing gateway acts as the PTP master clock, distributing synchronization time to all acquisition terminals within its coverage area via Ethernet. The PTP hardware clock of each acquisition terminal achieves sub-microsecond time synchronization accuracy. For terminals that do not support PTP, the NTP protocol is used to achieve millisecond-level time synchronization.

[0042] S32: Unified Spatial Coordinate Association. A global three-dimensional spatial reference system (CGCS2000 National Geodetic Coordinate System plus elevation) is established based on the design coordinates of the construction section. Each acquisition terminal obtains precise spatial coordinates via RTK positioning during deployment, with centimeter-level accuracy. Video surveillance equipment undergoes camera calibration using a calibration method, establishing a mapping relationship between image pixel coordinates and three-dimensional spatial coordinates. The UWB coordinates reported in real-time by personnel positioning terminals are directly mapped to the unified coordinate system. Each data entry is appended with a dual association index of "timestamp + spatial coordinates".

[0043] S33: Multimodal Data Spatiotemporal Fusion. Under a unified spatiotemporal reference, data from various sensors are aligned to the same timeline using a unified timestamp, forming a multimodal data snapshot at the same moment. The spatial coordinates of targets (personnel, equipment) detected by the video are spatiotemporally associated and bound with the sensor data. For example, when an environmental sensor at a certain location detects that the concentration of harmful gases exceeds the standard, the system automatically associates the location data and physiological data of all personnel within a 10-meter radius of that location, forming a structured fusion data package of "same time, same location".

[0044] Step S4: Lightweight transmission based on data importance classification: S41: Data Importance Classification. Edge nodes classify the merged data according to security relevance and urgency: Level I (Emergency) data includes incidents of people falling, intrusion into electronic fences (people entering restricted areas), structural vibration exceeding the threshold (vibration amplitude exceeding 80% of the design value), and excessive levels of harmful gases (CO concentration exceeding 50 ppm). This type of data is reported in real time. Level II (Important) data includes abnormal equipment operation (temperature exceeding limits, load exceeding limits), sudden environmental changes (wind speed exceeding level 6, temperature drop exceeding 10℃ / h), and abnormal physiological conditions of personnel (heart rate exceeding 120 bpm or below 50 bpm, body temperature exceeding 38.5℃). This type of data adopts a real-time reporting strategy. Level III (Routine) Data: This includes normal environmental temperature and humidity values, normal vibration monitoring data, and routine video streams. This type of data is reported in batches every 5 minutes.

[0045] S42: Lightweight Protocol-Based Tiered Transmission. It uses the MQTT protocol as the core data transmission protocol. The MQTT header is only 2 bytes, supporting binary data transmission. A publish / subscribe model is used to decouple devices from the cloud platform. Differentiated QoS levels are applied based on data importance: Level I data uses QoS2 (Exact Once), ensuring no data loss or duplication through a four-way handshake between the sender and receiver; Level II data uses QoS1 (At Least Once), ensuring the integrity of critical data through sender storage and retransmission mechanisms; Level III data uses QoS0 (At Most Once), requiring no acknowledgment or retransmission, suitable for batches of non-critical data to reduce network overhead. The edge computing gateway only uploads Level I and Level II data in real time, while Level III data is batch compressed and uploaded at 5-minute intervals.

[0046] Step S5: Cloud-based deep analysis and model iteration: After receiving data uploaded from each edge computing gateway, the cloud analytics layer performs immediate verification and confirmation of Level I data. Upon confirmation, it automatically triggers emergency plans (including pushing alarm information to on-site management personnel, initiating broadcast evacuation instructions, and notifying nearby emergency personnel). It conducts in-depth analysis of Level II and Level III data, using the runtime Transformer fusion model to extract the spatiotemporal and semantic features of multi-source data and generate dynamic risk scores (0-100 points). It integrates data from the entire line to conduct cross-section collaborative risk assessment and trend prediction. It updates the digital twin model to achieve a three-dimensional visualization of construction safety.

[0047] The cloud analytics layer retrains and optimizes the lightweight risk assessment model at the edge every week based on the accumulated full data. The updated model parameters are then distributed to each edge computing gateway through a secure channel, enabling continuous iterative evolution of the edge model.

[0048] Example 3: This embodiment provides a specific application scenario for edge data preprocessing and local early warning.

[0049] During the construction of a highway tunnel, edge computing gateways were deployed at the tunnel entrance and every 500 meters inside the tunnel. Hazardous gas sensors (CO, NO2, CH4) deployed inside the tunnel collect gas concentration data in real time at a sampling frequency of 1Hz.

[0050] During a construction operation, the CO concentration inside the tunnel rose from the normal value of 8 ppm to 65 ppm within 3 seconds (exceeding the Level I threshold of 50 ppm). Upon receiving this data, the edge data preprocessing module of the edge computing gateway first determined that the data was valid (not due to sensor failure) by outlier removal. Subsequently, the edge lightweight risk assessment module ran a decision tree model and assessed the risk level as "alarm" (Level I emergency) within 35 ms.

[0051] The edge communication protocol adaptation module immediately reports the warning information to the cloud analysis layer in real time via the MQTT protocol at QoS2 level. Simultaneously, local alarms are triggered at the tunnel entrance and inside the tunnel by audible and visual alarms (without relying on cloud feedback). On-site management personnel initiate the tunnel evacuation procedure within 10 seconds of the alarm. Upon receiving the warning information, the cloud analysis layer automatically correlates the location data of all personnel within a 50-meter radius of the location with the physiological data of smart wearable devices. If any person is found to have an abnormal heart rate or cessation of movement, the precise location information is immediately pushed to the nearest emergency medical personnel.

[0052] At the same time, Level III routine data (such as temperature and humidity inside the tunnel, and normal vibration data) are uploaded in batches every 5 minutes, without occupying the bandwidth resources of the emergency warning channel.

[0053] Example 4: This embodiment provides a specific application scenario for spatiotemporal synchronization and fusion of multi-source heterogeneous data.

[0054] During the construction of a high slope on a highway, six environmental sensor nodes, four structural vibration sensors, 15 personnel positioning terminals, three video acquisition terminals, and 15 smart wearable terminals were deployed simultaneously within a 500-meter section covered by the edge computing gateway.

[0055] Without a unified time reference, each sensor may have a time deviation of tens to hundreds of milliseconds due to independent clock drift. Taking the slope micro-seismic signal (frequency 100-300Hz) collected by the structural vibration sensor as an example, a time deviation of 100ms is equivalent to the offset of 10-30 sampling points, which will make it impossible to accurately determine the precise time of occurrence of the vibration event and its causal relationship with personnel location and environmental parameters.

[0056] In this embodiment, the edge computing gateway obtains the UTC time reference through a GPS / BeiDou time synchronization module, with a PPS pulse signal accuracy of ±50ns. The edge computing gateway acts as the PTP master clock, distributing synchronization time to all acquisition terminals within its coverage area via Ethernet. Each acquisition terminal has a built-in PTP hardware clock, achieving sub-microsecond time synchronization accuracy. All data is stamped with a unified timestamp accurate to the microsecond level upon generation.

[0057] Regarding spatial coordinate unification, each acquisition terminal obtains precise spatial coordinates through RTK positioning, with an accuracy of centimeters. The video acquisition terminal performs camera calibration using a calibration method, establishing a mapping relationship between image pixel coordinates and three-dimensional spatial coordinates. When the video acquisition terminal detects a person falling at a certain location (spatial coordinates X=3542167.32, Y=495832.15, Z=125.67), the system automatically associates data from all sensors within a 5-meter radius of that coordinate—including vibration data at that location (to determine if it was caused by a falling rock impact), environmental data (to determine if it was caused by harmful gases leading to unconsciousness), and data from the person's smart wearable device (to determine if heart rate and body temperature are abnormal)—forming a structured fusion data packet of "same time, same location" and uploading it to the cloud analysis layer, providing a complete multimodal evidence chain for accident cause analysis.

[0058] Example 5: This embodiment provides a specific application scenario for cloud-based model iterative updates.

[0059] During the 18-month construction period of a highway reconstruction and expansion project, the construction phases successively included four stages: roadbed excavation (months 1-4), roadbed filling (months 5-8), pavement paving (months 9-12), and traffic safety facility installation (months 13-18). The safety risk characteristics of each construction stage differed significantly: the main risks during the roadbed excavation stage were slope collapse and falls from heights; the main risks during the pavement paving stage were high-temperature burns and mechanical injuries; and the main risk during the traffic safety facility installation stage was working at heights.

[0060] The lightweight risk assessment model deployed on each edge computing gateway (the initial version was a decision tree model trained based on general construction safety data) achieved a risk identification accuracy of 92% during the initial stage of roadbed excavation. However, when construction entered the pavement paving stage, the risk identification accuracy of the initial model dropped to 78% due to significant changes in risk characteristics (increased risk of heatstroke in high-temperature asphalt paving environment and increased risk of mechanical injury in the roller operation area).

[0061] The cloud analytics layer retrains the edge model based on the full dataset accumulated each week: At the end of week 10, the cloud uses the training set accumulated over the previous nine weeks, including high-temperature environment data and machinery operation data, to incrementally learn the decision tree model, adjusting the risk feature weights and decision thresholds; on Monday morning of week 11, the cloud distributes the updated model parameters (version v2.0) to all edge computing gateways via a secure channel. After the update, the model's accuracy in identifying risks during the road paving stage rebounded to 89%.

[0062] Subsequently, the cloud performs model retraining and distribution every weekend, forming a closed-loop learning system of "data collection → edge processing → cloud analysis → model optimization → edge update". By the end of construction, the edge model had been iterated and updated 26 times, and could adapt to changes in safety risk characteristics under different construction stages and environmental conditions. The risk identification accuracy at each stage remained stable at over 90%.

[0063] In summary, this invention discloses a highway engineering construction safety data acquisition system and method, employing a three-layer collaborative architecture of "edge-cloud". The edge layer performs anomaly removal, feature extraction, and lightweight compression (75% compression rate) and runs a lightweight risk assessment model to achieve millisecond-level local early warning. The cloud layer performs global risk assessment and model iteration. This system offers the following advantages: significantly reduced bandwidth usage and storage costs; millisecond-level local early warning, eliminating reliance on cloud-based data transmission and improving emergency response speed; sub-microsecond time synchronization via the PTP protocol, combined with RTK positioning to establish unified spatial coordinates, enabling precise spatiotemporal fusion of multimodal data; a three-level data hierarchical transmission mechanism, with emergency data transmitted precisely once using the MQTT protocol with QoS2, ensuring zero data loss in early warnings; and regular retraining and updating of the edge model in the cloud, forming a closed-loop learning system that adapts to changes in risk characteristics at different construction stages. This effectively solves the problems of weak edge processing capabilities, inconsistent spatiotemporal benchmarks, and insufficient critical data protection in existing technologies.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A highway engineering construction safety data acquisition system, characterized in that, The architecture adopts a three-layer collaborative structure of endpoint-edge-cloud, including: The sensing terminal layer is deployed along the entire highway construction line. The sensing terminal layer includes a data acquisition terminal, which is used to collect multi-source heterogeneous raw data at a preset sampling frequency. The data acquisition terminal has a built-in high-precision real-time clock module. The edge intelligence layer, deployed at the construction site's edge computing gateway, includes: an edge data preprocessing module for outlier removal, format standardization, timestamp marking, and keyframe extraction and target detection of the video stream; an edge feature extraction and compression encoding module for time-domain / frequency-domain feature extraction and lightweight encoding compression of the preprocessed data; an edge lightweight risk assessment module for running a lightweight risk assessment model to perform real-time risk level assessment and millisecond-level local early warning of the preprocessed data; and an edge communication protocol adaptation module for enabling data protocol access and bidirectional communication with the cloud platform. The cloud analytics layer, deployed on a central server or public cloud platform, is used to receive data uploaded by edge nodes, run deep analytics models for global situational awareness and cross-segment collaborative risk assessment, complete iterative training of the model, and distribute the updated model to edge nodes.

2. The highway engineering construction safety data acquisition system according to claim 1, characterized in that, The sensing terminal layer includes: environmental sensor nodes fixedly deployed along the construction section for collecting data on temperature and humidity, wind speed and direction, dust concentration, and concentrations of harmful gases including carbon monoxide and nitrogen oxides; structural vibration sensors fixedly installed on key structures for collecting structural vibration frequency and amplitude data; personnel positioning terminals and smart wearable terminals worn by construction workers for collecting high-precision location information and physiological parameters including heart rate, body temperature, and body posture; the positioning accuracy of the high-precision location information is no greater than 30cm; the body posture physiological parameters include triaxial acceleration and angular velocity data; a video acquisition terminal for collecting real-time video streams from the construction site; and mechanical equipment status sensors fixedly installed on machinery for collecting mechanical operating parameters; the machinery includes road rollers, pavers, and tower cranes; the operating parameters include the vibration frequency and compaction speed of the road rollers, the paving thickness and paving temperature of the pavers, and the lifting weight and tilt angle of the tower cranes.

3. The highway engineering construction safety data acquisition system according to claim 1, characterized in that, The edge computing gateway has a built-in GPS / BeiDou time synchronization module to obtain a high-precision UTC time reference, and distributes synchronization time to the acquisition terminals within the coverage area via Ethernet as a PTP master clock. The acquisition terminals have built-in PTP hardware clocks to achieve sub-microsecond time synchronization. The acquisition terminals obtain precise spatial coordinates through RTK positioning, and the video acquisition terminals establish a mapping relationship between image pixel coordinates and three-dimensional spatial coordinates through camera calibration. All acquired data is appended with a dual-link index of timestamp and spatial coordinates.

4. The highway engineering construction safety data acquisition system according to claim 1, characterized in that, The edge computing gateway covers a construction area of ​​500 meters to 1 kilometer. The acquisition terminal communicates with the edge computing gateway via wired Ethernet or 5G wireless network. The edge communication protocol adaptation module supports the conversion of Modbus, CAN, RS485 industrial protocols to the MQTT standardized protocol and adopts a publish / subscribe model to decouple the device from the cloud platform.

5. A method for collecting safety data during highway construction, characterized in that, The implementation of the highway engineering construction safety data acquisition system according to any one of claims 1 to 4 includes the following steps: S1. Collect multi-source heterogeneous data through the sensing terminal layer at a preset sampling frequency, and add a local hardware timestamp when the data is generated. S2. After receiving the raw data, the edge computing gateway sequentially performs outlier removal, feature extraction, and lightweight encoding compression to obtain compressed feature data. S3. Under a unified spatiotemporal reference, sensor data are aligned to the same time axis according to a unified timestamp and mapped to a unified three-dimensional spatial coordinate system to form a structured fusion data package at the same time and location. S4. Edge nodes classify the fused data according to security relevance and urgency, and use the MQTT protocol to transmit the data to the cloud platform with differentiated QoS levels based on the data level. After receiving data, the S5 cloud platform runs a deep analysis model to perform global situational awareness and cross-segment collaborative risk assessment, and regularly retrains and updates the edge model.

6. The method for collecting safety data during highway construction according to claim 5, characterized in that, In S2, outlier removal specifically involves: using a statistical threshold-based filtering method for numerical data, marking data that exceeds the preset normal range as outliers and removing them; using the isolated forest algorithm for outlier detection for time-series data; and using a lightweight CNN model for target detection for video data, identifying and discarding invalid or blurry frames.

7. The method for collecting safety data during highway construction according to claim 5, characterized in that, In S2, feature extraction includes: extracting time-domain features such as mean, standard deviation, peak value, and root mean square value; extracting frequency-domain features of the main frequency component and spectral energy distribution through fast Fourier transform; extracting statistical features such as skewness, kurtosis, and coefficient of variation; and extracting safety semantic features such as helmet wearing status, personnel fall behavior, and fireworks events from the video stream.

8. A method for collecting safety data during highway construction according to claim 5, characterized in that, In S2, lightweight coding compression includes: using differential coding to reduce temporal data redundancy, using fixed-point quantization to compress data precision for floating-point numbers, and using optimized coding formats to compress video keyframes.

9. A method for collecting safety data during highway construction according to claim 5, characterized in that, In S4, data is categorized by safety relevance and urgency: Level I Emergency Data includes personnel falls, electronic fence intrusion, structural vibration exceeding thresholds, and hazardous gas exceeding standards; structural vibration exceeding thresholds refers to vibration amplitude exceeding the limit specified in the standard, or exceeding 80% of the design value; hazardous gas exceeding standards refers to carbon monoxide concentration exceeding 24 ppm or nitrogen dioxide concentration exceeding 3 ppm; Level II Important Data includes equipment malfunction, sudden environmental changes, and personnel physiological abnormalities; equipment malfunction includes equipment temperature exceeding 85℃ or vibration frequency deviating from the normal operating range by ±15%; sudden environmental changes refer to wind speed changing by more than level 4 within 10 minutes or temperature changing by more than 10℃ within 30 minutes; personnel physiological abnormalities refer to heart rate continuously exceeding 120 beats / min or below 50 beats / min for 1 minute, or body temperature exceeding 38.5℃; Level III Routine Data includes routine environmental monitoring data, normal vibration data, and routine video streams. The MQTT protocol is used for transmission with differentiated QoS levels. Level I data is transmitted exactly once using QoS2, Level II data is transmitted at least once using QoS1, and Level III data is transmitted at most once using QoS0 and uploaded in batches with compression according to a preset period. Edge nodes only upload Level I and Level II data in real time.

10. A method for collecting safety data during highway construction according to claim 5, characterized in that, In S5, the cloud platform performs real-time review and confirmation of Level I data and triggers emergency plans; for Level II and Level III data, the runtime space-time Transformer fusion model extracts the spatiotemporal and semantic features of multi-source data and generates dynamic risk scores. Based on the accumulated full data, the cloud periodically retrains and optimizes the lightweight risk assessment model at the edge, and distributes the updated model parameters to the edge computing nodes, forming a closed loop of data collection → edge processing → cloud analysis → model optimization → edge update.

Images