A construction safety management and control method and system for expressways

Abstract

The application discloses a construction safety management and control method and system for expressways, and relates to the field of intelligent traffic safety; the method collects high-precision positioning fusion sensing data of roadside radars, vehicle network connection and construction entities in real time, and maps the data to a digital twin voxel space; hidden features are extracted by using a space-time variation autoencoder model to predict the uncertainty probability distribution of target trajectories; a dynamic space-time collision time matrix of the intersection area is calculated based on the probability distribution, and corresponding hierarchical collaborative risk avoidance strategies are generated and issued according to the extreme values of the matrix; meanwhile, the deceleration compliance rate of vehicles after the execution of the strategies is extracted as a feedback signal to adaptively fine-tune the model. The application effectively quantifies the random divergent boundary of complex traffic flow, overcomes the defect of low fault tolerance of traditional prediction models, realizes deep closed-loop coupling of safety management and control strategies and real driving feedback, and significantly improves the active defense strength and system generalization ability of the construction area.

Description

Technical Field

[0001] This invention belongs to the field of highway construction safety, and in particular relates to a method and system for construction safety management and control of highways. Background Technology

[0002] With the deep evolution of intelligent transportation systems and vehicle-road cooperative technologies, safety management in highway construction areas is gradually shifting from traditional manual inspections to digital monitoring. Because highway reconstruction and expansion operations often take place under complex traffic flow conditions of "construction while traffic is flowing," the site not only faces a large number of unstructured construction machinery and personnel scheduling issues, but also severe spatiotemporal conflicts caused by varying speeds and limited visibility of other vehicles. Currently, the mainstream construction safety management solutions in the industry mainly rely on deploying roadside radar and visual perception equipment at the front end of the construction area. These systems use conventional kinematic models, such as Kalman filtering, to perform deterministic trajectory extrapolation of passing vehicles and delineate static virtual electronic fences in the cloud. Once a vehicle's trajectory is detected to geometrically intersect the fence, a one-way audible and visual alarm is triggered, or a hazard avoidance command is issued via V2X communication.

[0003] Existing highway construction management and control schemes suffer from serious technical defects, including low trajectory prediction fault tolerance and open-loop management links. Specifically, traditional deterministic trajectory extrapolation models cannot effectively characterize the highly nonlinear and random micro-driving behaviors of human drivers when facing complex scenarios such as speed limits and lane changes in construction zones. They are also unable to accurately quantify the divergent uncertainty range of target trajectory in digital space, leading to frequent false alarms and missed reports when vehicles make abnormal incursions or need to make emergency avoidances. Furthermore, existing systems generally lack continuous observation and quantitative evaluation of the "deceleration compliance rate" of physical vehicles after issuing avoidance commands, failing to feed back the real avoidance feedback from the physical space into the prediction algorithm. This results in the prediction model's generalization ability stagnating under different traffic flow densities and weather conditions, making it difficult to build a truly robust and adaptive proactive defense system for construction safety. Summary of the Invention

[0004] The purpose of this invention is to provide a construction safety management method and system for highways, aiming to solve the problems mentioned in the background art.

[0005] This invention is implemented as follows: On one hand, a construction safety management method for highways, the method comprising:

[0006] Real-time collection of traffic flow and construction operation data in highway construction areas;

[0007] The traffic flow and construction operation fusion perception data includes at least roadside lidar point cloud data, vehicle network trajectory data, and high-precision positioning data of construction entities.

[0008] The traffic flow and construction operation fusion perception data are mapped to a pre-constructed digital twin voxel space, and the hidden spatial features within continuous time steps are extracted using a spatiotemporal variational autoencoder model to predict and output the trajectory uncertainty probability distribution of past vehicles and construction entities.

[0009] Based on the trajectory uncertainty probability distribution, the dynamic spatiotemporal collision time matrix of the intersection area is calculated, and a matching hierarchical collaborative risk avoidance strategy is generated according to the extreme value of the dynamic spatiotemporal collision time matrix.

[0010] The hierarchical collaborative risk avoidance strategy is issued and executed, while the deceleration compliance rate of passing vehicles after the strategy is executed is monitored. The deceleration compliance rate is used as a feedback signal to adaptively fine-tune the spatiotemporal variational autoencoder model.

[0011] As a further aspect of the present invention, the real-time acquisition of traffic flow and construction operation fusion sensing data in the highway construction area specifically includes:

[0012] The dynamic three-dimensional bounding boxes of passing vehicles are obtained by roadside perception computing units deployed upstream and to the side of the construction area, which serve as the point cloud data of the roadside lidar.

[0013] The vehicle receives basic safety messages broadcast by surrounding intelligent connected vehicles through the vehicle-road cooperative communication protocol, and extracts vehicle speed, heading angle and brake pedal status as vehicle connected trajectory data.

[0014] Differential positioning coordinates are obtained by using positioning wristbands worn by construction workers and mobile stations mounted on machinery, which serve as high-precision positioning data for the construction entity.

[0015] As a further aspect of the present invention, the predicted output of the trajectory uncertainty probability distribution of past vehicles and construction entities specifically includes:

[0016] After spatiotemporal alignment of the roadside lidar point cloud data, vehicle network trajectory data, and high-precision positioning data of construction entities, they are transformed into a rasterized tensor sequence in the digital twin voxel space.

[0017] The rasterized tensor sequence is input into the encoder of the spatiotemporal variational autoencoder model to generate latent variables that follow a multivariate Gaussian distribution.

[0018] The latent variables are reconstructed and extrapolated to future time steps by the decoder of the spatiotemporal variational autoencoder model, and the output is the trajectory uncertainty probability distribution containing the expected value of the predicted trajectory and the covariance matrix, so as to characterize the spatial probability boundary that the target may appear in the future.

[0019] As a further aspect of the present invention, the generation of a matching hierarchical collaborative risk avoidance strategy specifically includes:

[0020] In the digital twin voxel space, the minimum collision time of the overlapping region of each probability boundary is calculated, and the dynamic spatiotemporal collision time matrix is ​​constructed.

[0021] When the extreme value of the dynamic spatiotemporal collision time matrix is ​​greater than the first threshold, a first-level avoidance strategy is generated, which includes sending a lane change warning command to the connected terminal of the passing vehicle.

[0022] When the extreme value of the dynamic spatiotemporal collision time matrix is ​​between the first threshold and the second threshold, a second-level risk avoidance strategy is generated, which includes dynamically expanding the virtual electronic fence range of the construction area and triggering the roadside directional millimeter-wave audible and visual alarm device.

[0023] When the extreme value of the dynamic spatiotemporal collision time matrix is ​​less than the second threshold, a third-level avoidance strategy is generated, which includes issuing the highest priority emergency vibration and voice evacuation commands to the wearable and vehicle-mounted equipment of the construction entity.

[0024] As a further aspect of the present invention, the process of adaptively fine-tuning the spatiotemporal variational autoencoder model by using the deceleration compliance rate as a feedback signal specifically includes:

[0025] Within the observation window after the graded collaborative risk avoidance strategy is issued, the ratio of the number of vehicles that actually decelerate to the total number of vehicles warned is used to obtain the deceleration compliance rate.

[0026] Calculate the deviation loss between the deceleration compliance rate and the target expected compliance rate;

[0027] The bias loss is introduced into the loss function of the spatiotemporal variational autoencoder model, and the model network weights are updated through the backpropagation algorithm, so that the model can automatically expand the covariance range of trajectory prediction in low compliance scenarios.

[0028] As a further aspect of the present invention, another option is a construction safety management and control system for highways, the system comprising:

[0029] The perception data fusion and aggregation module is used to collect real-time fusion perception data of traffic flow and construction operations in the highway construction area;

[0030] The traffic flow and construction operation fusion perception data includes at least roadside lidar point cloud data, vehicle network trajectory data, and high-precision positioning data of construction entities.

[0031] The twin mapping and probability prediction module is used to map the traffic flow and construction operation fusion perception data to a pre-constructed digital twin voxel space, and use a spatiotemporal variational autoencoder model to extract hidden spatial features within continuous time steps, and predict and output the trajectory uncertainty probability distribution of past vehicles and construction entities.

[0032] The collision matrix and strategy generation module is used to calculate the dynamic spatiotemporal collision time matrix of the intersection area based on the trajectory uncertainty probability distribution, and generate a matching hierarchical collaborative risk avoidance strategy according to the extreme value of the dynamic spatiotemporal collision time matrix.

[0033] The execution feedback and adaptive fine-tuning module is used to issue and execute the hierarchical collaborative risk avoidance strategy, while monitoring the deceleration compliance rate of passing vehicles after the strategy is executed, and using the deceleration compliance rate as a feedback signal to adaptively fine-tune the spatiotemporal variational autoencoder model.

[0034] As a further aspect of the present invention, the sensing data fusion and aggregation module specifically includes:

[0035] Roadside laser scanning unit: used to acquire dynamic 3D bounding boxes of passing vehicles, its purpose is to provide robust non-connected vehicle contour detection under complex lighting conditions;

[0036] Vehicle-road cooperative analysis unit: used to receive basic safety messages broadcast by intelligent vehicles and extract the internal operating status of the vehicles. Its purpose is to obtain the advance movement intentions of vehicles beyond line of sight and those that are obscured.

[0037] Physical high-precision positioning unit: used to obtain differential coordinates of construction personnel and machinery, and its purpose is to provide high-frequency centimeter-level position references for moving targets within the construction area.

[0038] As a further aspect of the present invention, the twin mapping and probability prediction module specifically includes:

[0039] Twin space tensor transformation unit: used to transform aligned data into a rasterized tensor sequence in digital twin voxel space;

[0040] Spatiotemporal variational encoding / decoding unit: used to generate latent variables and output prediction results containing expected values ​​and covariance matrices.

[0041] As a further aspect of the present invention, the collision matrix and strategy generation module specifically includes:

[0042] The dynamic collision time calculation unit is used to calculate the minimum collision time of the overlapping regions of each probability boundary in the digital twin voxel space and construct the matrix.

[0043] The tiered strategy delivery unit is used to trigger network-connected early warning, electronic fence linkage, or emergency evacuation commands based on the extreme values ​​of the collision time.

[0044] This invention provides a construction safety management method and system for highways. By introducing a spatiotemporal variational autoencoder model into the digital twin space, it overcomes the technical bottleneck of traditional mechanical extrapolation algorithms being unable to predict sudden driving behaviors and effectively quantifies the random divergence boundary of complex traffic flow trajectories. At the same time, it uses the vehicle deceleration compliance rate in the real physical space as a negative feedback signal to feed back the model weights, realizing the adaptive closed-loop coupling of safety management strategies with dynamic road conditions, weather conditions, and driver psychology. Attached Figure Description

[0045] Figure 1 This is the main flowchart of a construction safety management method for highways.

[0046] Figure 2 This is a main structure diagram of a construction safety management and control system for highways.

[0047] Figure 3 This is a schematic diagram illustrating the trajectory uncertainty prediction principle of an embodiment of the present invention.

[0048] Figure 4 This is a schematic diagram of scene linkage and closed loop in an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0050] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0051] This invention provides a construction safety management method and system for highways, which solves the technical problems in the background art.

[0052] like Figure 1 The diagram shown is a main flowchart of a construction safety management method for highways, provided in one embodiment of the present invention. The method includes:

[0053] Step S100: Real-time collection of traffic flow and construction operation fusion sensing data in the highway construction area;

[0054] The traffic flow and construction operation fusion perception data includes at least roadside lidar point cloud data, vehicle network trajectory data, and high-precision positioning data of construction entities.

[0055] Step S200: Map the traffic flow and construction operation fusion perception data to a pre-constructed digital twin voxel space, and use a spatiotemporal variational autoencoder model to extract hidden spatial features within continuous time steps, and predict and output the trajectory uncertainty probability distribution of past vehicles and construction entities.

[0056] Step S300: Based on the trajectory uncertainty probability distribution, calculate the dynamic spatiotemporal collision time matrix of the intersection area, and generate a matching hierarchical collaborative risk avoidance strategy according to the extreme value of the dynamic spatiotemporal collision time matrix;

[0057] Step S400: Issue and execute the hierarchical collaborative risk avoidance strategy, while monitoring the deceleration compliance rate of passing vehicles after the strategy is executed, and use the deceleration compliance rate as a feedback signal to adaptively fine-tune the spatiotemporal variational autoencoder model.

[0058] In this embodiment, the first step is to collect real-time fused perception data of traffic flow and construction operations within the construction area and at least 1.5 kilometers upstream at a high frequency of no less than 20Hz at the perception layer. This data not only covers traditional macroscopic traffic density but also delves into the three-dimensional contours of microscopic vehicles and the centimeter-level poses of individual construction entities. Subsequently, the system's edge computing nodes uniformly transform the aforementioned fused perception data into a digital twin voxel space based on the WGS84 ellipsoidal datum and projected onto the local ENU coordinate system, ensuring a precise 1:1 three-dimensional scale mapping between the physical and digital worlds. Within the digital twin space, instead of using traditional Kalman filtering for linear extrapolation, a spatiotemporal variational autoencoder model pre-deployed on a high-performance GPU is invoked to perform nonlinear dimensionality reduction and latent variable feature extraction on historical trajectory features across multiple consecutive frames. This predicts the probability distribution of trajectory uncertainty for past vehicles and construction entities within the next 3 to 5 seconds, representing the spatial envelope of possible abnormal lane changes or emergency avoidance by vehicles in the form of a multidimensional covariance matrix. Then, multi-target three-dimensional envelope surface cross-collision tests are conducted in digital space to calculate the dynamic spatiotemporal collision time matrix of the intersection area. Based on this matrix, a preset rule engine is triggered within millisecond delays to generate a graded collaborative risk avoidance strategy that matches the current risk extreme value. Finally, control commands are sent to the execution terminals at all levels on site via a 5G private network or directly via fiber optic cable. Within a preset time window after the strategy is executed, the actual deceleration acceleration and trajectory deviation of passing vehicles are continuously tracked. The deceleration compliance rate is extracted as a reward feedback signal for the reinforcement learning mechanism, and the hyperparameters of the spatiotemporal variational autoencoder model are dynamically fine-tuned. This forms an active safety closed-loop control system with continuous self-evolution capabilities that can automatically calibrate the prediction boundary according to driving habits in different areas.

[0059] like Figure 3As shown, in a preferred embodiment of the present invention, the real-time acquisition of traffic flow and construction operation fusion sensing data in the highway construction area specifically includes:

[0060] Step S101: Obtain the dynamic three-dimensional bounding boxes of passing vehicles by roadside perception computing units deployed upstream and to the side of the construction area, as the point cloud data of the roadside lidar;

[0061] Step S102: Receive basic safety messages broadcast by surrounding intelligent connected vehicles through the vehicle-road cooperative communication protocol, and parse and extract vehicle speed, heading angle and brake pedal status as vehicle network trajectory data;

[0062] Step S103: Differential positioning coordinates are obtained through the positioning wristbands worn by construction workers and the mobile station mounted on the machinery vehicle, which serve as high-precision positioning data for the construction entity.

[0063] In this embodiment, during application, 128-line equivalent high-resolution solid-state LiDAR and edge-side AI processing units are deployed at equal intervals of 150 meters along the wave-shaped guardrail in the upstream transition zone and periphery of the core operation area of ​​the construction zone. The LiDAR point cloud data is first filtered by VoxelGrid to remove redundant noise, and the ground point cloud is removed using the RANSAC algorithm. Then, the dynamic three-dimensional bounding boxes (including length, width, height, center point coordinates, and heading angle) of passing vehicles are extracted using Euclidean clustering and deep learning 3D target detection algorithms. At the same time, the roadside units deployed on the roadside continuously receive basic safety messages or cooperative perception messages broadcast by surrounding intelligent connected vehicles at a frequency of 10Hz based on the C-V2X PC5 direct communication interface. The parsing scripts accurately extract the underlying control parameters of the vehicles from the ASN.1 encoded code stream, especially the brake pedal depth, steering wheel angle, and ESP system intervention status, thereby obtaining the advanced intentions of targets beyond the line of sight of non-connected vehicles. Furthermore, for high-frequency dynamic entities at the construction site, each worker's smart safety helmet integrates an RTK high-precision positioning module supporting BDS / GPS dual-mode, while large road rollers and pavers on site have mobile vehicle-mounted reference stations installed on their roof antennas. By receiving RTCM3.2 format differential correction data broadcast from a temporary differential base station set up in the construction section, the system can calculate the centimeter-level three-dimensional coordinates of the construction personnel and machinery with fixed solutions in real time. All three data streams are stamped with nanosecond-level high-precision timestamps based on the IEEE 1588 PTP protocol, ensuring tight alignment of spatiotemporal data.

[0064] like Figure 3 As shown, in a preferred embodiment of the present invention, the predicted output of the trajectory uncertainty probability distribution of past vehicles and construction entities specifically includes:

[0065] Step S201: After spatiotemporally aligning the roadside lidar point cloud data, vehicle network trajectory data, and high-precision positioning data of construction entities, convert them into a rasterized tensor sequence in the digital twin voxel space.

[0066] Step S202: Input the rasterized tensor sequence into the encoder of the spatiotemporal variational autoencoder model to generate latent variables that follow a multivariate Gaussian distribution;

[0067] Step S203: The latent variables are reconstructed and extrapolated to future time steps by the decoder of the spatiotemporal variational autoencoder model, and the trajectory uncertainty probability distribution containing the expected value of the predicted trajectory and the covariance matrix is ​​output to characterize the spatial probability boundary that the target may appear in the future.

[0068] In this embodiment, the multi-source aligned perception data is first discretized and mapped into a rasterized tensor sequence in a three-dimensional digital twin voxel space. For example, the construction area is divided into extremely small voxel units of 0.2m × 0.2m × 0.2m, and each cell is assigned a feature tensor containing occupancy probability, velocity vector, and acceleration. This tensor sequence is then fed into the encoder network of a spatiotemporal variational autoencoder model. The encoder adopts a cascaded architecture of multi-layer graph convolutional networks and long short-term memory networks to capture the spatiotemporal dependencies of the game between vehicles, and maps the high-dimensional input to a low-dimensional latent space through fully connected layers. To ensure the model's generative capability, the latent space does not output deterministic feature vectors, but instead outputs the mean and variance vectors following a multivariate Gaussian distribution. During training and inference, random sampling is performed from this Gaussian distribution using reparameterization techniques to obtain latent variables carrying uncertain features. The decoder network receives these latent variables and, combined with the current road network topology constraints, reconstructs the target trajectory for future multiple frames through deconvolution and temporal unrolling operations. The final output prediction result not only includes a center line of the expected trajectory with the highest probability of occurrence, but more importantly, it outputs a covariance matrix that continuously expands with the prediction time step. This matrix is ​​visualized in three-dimensional space as a dynamically expanding "probability ellipsoid," which quantifies the spatial boundary of extreme divergent behaviors that human drivers may exhibit when facing traffic cone restrictions in construction zones, such as tentative lane changes and sudden braking with rear-end skidding.

[0069] like Figure 4 As shown, in a preferred embodiment of the present invention, the generation of a matching hierarchical collaborative risk avoidance strategy specifically includes:

[0070] Step S301: In the digital twin voxel space, calculate the minimum collision time of the overlapping regions of each probability boundary, and construct the dynamic spatiotemporal collision time matrix;

[0071] Step S302: When the extreme value of the dynamic spatiotemporal collision time matrix is ​​greater than the first threshold, a first-level avoidance strategy is generated, which includes sending a lane change warning command to the connected terminal of the passing vehicle.

[0072] Step S303: When the extreme value of the dynamic spatiotemporal collision time matrix is ​​between the first threshold and the second threshold, a second-level risk avoidance strategy is generated, which includes dynamically expanding the virtual electronic fence range of the construction area and triggering the roadside directional millimeter wave sound and light alarm device.

[0073] Step S304: When the extreme value of the dynamic spatiotemporal collision time matrix is ​​less than the second threshold, a third-level avoidance strategy is generated, which includes issuing the highest priority emergency vibration and voice evacuation commands to the wearable and vehicle-mounted equipment of the construction entity.

[0074] It should be understood that after predicting the probability boundary of future trajectories, the twin engine performs continuous Boolean intersection operations on the predicted probability ellipsoids of all passing vehicles and the virtual electronic fence of the construction area. When spatial overlap is detected, the future time point corresponding to the overlap is extracted and defined as the dynamic spatiotemporal collision time. The system has a built-in three-level defense threshold system: when the extreme value of the calculated dynamic spatiotemporal collision time matrix is ​​between 3 and 5 seconds (i.e., the first threshold range), the system determines it as a potential risk at a distance. At this time, the system sends a low-latency early warning message to the OBU of the connected vehicle involved through the edge-side RSU, triggering a visual and gentle voice prompt of "Construction ahead, lane change risk" on the vehicle's in-vehicle infotainment screen or HUD, realizing soft guidance of macro traffic flow; when the extreme value of the dynamic spatiotemporal collision time is shortened to between 1.5 and 3 seconds (i.e., the second threshold range), the system determines it as a moderate approach risk. At this time, the control platform extends the electronic fence by 0.5 meters in the digital twin space for redundant defense, and at the same time directly triggers the industrial Internet of Things protocol. In the real physical world, a directional millimeter-wave alarm deployed at the edge of the lane involved in the accident precisely projects a sharp sound wave of up to 120 decibels and a high-frequency strobe laser into the driver's cab of the vehicle, awakening a distracted or fatigued driver through strong physical stimulation. When the extreme value of the dynamic spatiotemporal collision time is less than 1.5 seconds (i.e., the third threshold, defined as an unavoidable critical collision state), the system instantly bypasses vehicle intervention and directly issues the highest priority survival command to the construction area. Through the LoRaWAN protocol, it triggers the violent vibration module of the construction worker's exoskeleton or safety helmet to force the personnel to lie down and avoid the collision. It also uses a hard connection via the CAN bus to instantly cut off the hydraulic main valve of the excavator on site, forcibly locking its boom to prevent mechanical rollover or secondary crush injuries after the vehicle collision.

[0075] In a preferred embodiment of the present invention, the process of adaptively fine-tuning the spatiotemporal variational autoencoder model by using the deceleration compliance rate as a feedback signal specifically includes:

[0076] Step S401: Within the observation window after the hierarchical collaborative risk avoidance strategy is issued, the ratio of the number of vehicles that actually decelerate to the total number of vehicles warned is calculated to obtain the deceleration compliance rate.

[0077] Step S402: Calculate the deviation loss between the deceleration compliance rate and the target expected compliance rate;

[0078] Step S403: Introduce the bias loss into the loss function of the spatiotemporal variational autoencoder model, and update the model network weights through the backpropagation algorithm so that the model can automatically expand the covariance range of trajectory prediction in low compliance scenarios.

[0079] In this embodiment, after each issuance of the tiered collaborative risk avoidance strategy, an 8-second observation window is initiated to continuously track the longitudinal acceleration and lateral yaw rate of the warned vehicles using lidar and V2X data. The system is configured such that if a vehicle exhibits a continuous deceleration exceeding 2.5 m / s² within 2 seconds of receiving the instruction, or safely moves laterally away from the conflict lane, it is recorded as a valid compliance sample. Within each statistical period (e.g., every 10 minutes), the ratio of the actual number of vehicles decelerating to the total number of warned vehicles is calculated to obtain the true deceleration compliance rate. Subsequently, the system subtracts this actual compliance rate from the platform's preset expected compliance rate (e.g., set to 95%) to calculate the deviation loss. In the subsequent online fine-tuning of the spatiotemporal variational autoencoder model, not only is the conventional variational lower bound used as the loss function, but the aforementioned deviation loss is also introduced as a penalty term into the backpropagation process. When the system detects a significant drop in deceleration compliance during rainy or snowy weather or at night, the bias loss surges. Gradient updates force the spatiotemporal variational autoencoder model to actively increase variance weights when generating latent variables. Macroscopically, this manifests as the system recognizing the current sluggishness of the driver group and automatically amplifying the probability envelope (covariance range) of all vehicle predictions. This allows the vehicle to spatially overlap with the virtual fence earlier, effectively advancing the warning timing and adaptively extending the protection depth, greatly improving the model's robustness under long-tailed adverse conditions.

[0080] like Figure 2 As shown, in another preferred embodiment of the present invention, a construction safety management and control system for highways is provided, the system comprising:

[0081] The perception data fusion and aggregation module 100 is used to collect real-time fusion perception data of traffic flow and construction operations in the highway construction area.

[0082] The traffic flow and construction operation fusion perception data includes at least roadside lidar point cloud data, vehicle network trajectory data, and high-precision positioning data of construction entities.

[0083] The twin mapping and probability prediction module 200 is used to map the traffic flow and construction operation fusion perception data to a pre-constructed digital twin voxel space, and use a spatiotemporal variational autoencoder model to extract hidden spatial features within continuous time steps, and predict and output the trajectory uncertainty probability distribution of past vehicles and construction entities.

[0084] The collision matrix and strategy generation module 300 is used to calculate the dynamic spatiotemporal collision time matrix of the intersection area based on the trajectory uncertainty probability distribution, and generate a matching hierarchical collaborative risk avoidance strategy according to the extreme value of the dynamic spatiotemporal collision time matrix.

[0085] The execution feedback and adaptive fine-tuning module 400 is used to issue and execute the hierarchical collaborative risk avoidance strategy, while monitoring the deceleration compliance rate of passing vehicles after the strategy is executed, and using the deceleration compliance rate as a feedback signal to adaptively fine-tune the spatiotemporal variational autoencoder model.

[0086] In this embodiment, the front-end data multidimensional perception and aggregation module 100 consists of hardware gateways and drivers that access various protocols. It is responsible for deduplicating, cleaning, and timestamping massive concurrent data, and distributing the cleaned formatted data through an internal high-speed message bus. The twin mapping and probability prediction module 200, at the core of edge computing, is deployed as a Docker container on an MEC server located close to a highway gantry. This module embeds a high-performance TensorRT inference engine, specifically responsible for maintaining the memory state of the digital twin voxel space and efficiently performing matrix multiplication and probability density calculations of the spatiotemporal variational autoencoder model, ensuring that the end-to-end latency from data input to trajectory distribution output is controlled within 50 milliseconds. The collision matrix and policy generation module 300 serves as the rule determination center, using a high-concurrency in-memory database to access the dynamic TTC matrix in real time, and generating the optimal solution under non-deterministic conditions based on a complex expert rule base and threshold conditions. Ultimately, the execution feedback and adaptive fine-tuning module 400 acts as a scheduling gateway for equipment back-control, sending instructions at all levels to the corresponding PLC controllers or RSU antennas through multi-threaded asynchronous concurrency. On the other hand, it collects trajectory feedback streams in the background, accumulates a certain batch of positive and negative samples in the cloud GPU cluster, and periodically recalculates the network gradient and sends out updated model weight files, thereby realizing the organic operation and continuous evolution of the entire complex system.

[0087] In another preferred embodiment of the present invention, the sensing data fusion and aggregation module 100 specifically includes:

[0088] The roadside laser scanning unit 101 is used to acquire the dynamic three-dimensional bounding boxes of passing vehicles, and its purpose is to provide highly robust non-connected vehicle contour detection under complex lighting conditions.

[0089] The vehicle-road cooperative analysis unit 102 is used to receive basic safety messages broadcast by intelligent vehicles and extract the internal operating status of the vehicles. Its purpose is to obtain the forward movement intentions of vehicles beyond line of sight and those that are obscured.

[0090] The physical high-precision positioning unit 103 is used to acquire the differential coordinates of construction personnel and machinery, and its purpose is to provide a high-frequency centimeter-level position reference for moving targets within the construction area.

[0091] In this embodiment, the roadside laser scanning unit 101 employs solid-state or hybrid solid-state LiDAR hardware, which effectively filters artifact noise caused by water mist kicked up by vehicles ahead and construction dust. Its main purpose is to overcome the blinding failure problem of pure vision cameras in nighttime, backlight, and tunnel entrance strong light gradient environments, providing a highly robust non-connected vehicle contour detection base with absolute physical depth. The vehicle-road cooperative parsing unit 102 is a protocol stack service that includes channel listening and ASN.1 decoding. Its practical significance lies in solving the line-of-sight occlusion problem. For example, when a heavy truck obstructs a car behind, the LiDAR cannot directly scan the car, but the parsing unit can directly penetrate the physical obstacle through the V2X radio waves emitted by the car to obtain its position, speed, and even the trigger signal of the anti-lock braking system (ABS), extending the system's advanced perception capability. The physical high-precision positioning unit 103 relies on the RTK differential base station network and uses carrier phase differential technology to eliminate ionospheric and tropospheric delay errors. Its core purpose is to provide an absolutely stable anchor point for the digital twin space. Because the movement of construction workers is often disordered and frequent, only by relying on centimeter-level high-frequency differential coordinates can we accurately determine whether a worker has already half-exposed the protective boundary of the safety cone when turning around, thereby completely eliminating false alarms and missed alarms caused by positioning drift.

[0092] In another preferred embodiment of the present invention, the twin mapping and probability prediction module 200 specifically includes:

[0093] The twin space tensor transformation unit 201 is used to transform aligned data into a rasterized tensor sequence in the digital twin voxel space;

[0094] The spatiotemporal variational encoding and decoding unit 202 is used to generate latent variables and output prediction results containing expected values ​​and covariance matrices.

[0095] In this embodiment, the twin space tensor transformation unit 201 is a high-intensity memory computing module. Its working principle is to discretize the continuous three-dimensional physical world coordinates according to a fixed resolution, constructing a three-dimensional grid matrix (rasterized tensor). Each grid not only stores the Boolean value indicating space occupancy but also incorporates the velocity component and acceleration partial derivative of the vehicle currently occupying that grid. Its core purpose is to forcibly normalize scattered data such as radar point clouds and GPS latitude and longitude into a high-dimensional matrix format that convolutional neural networks can directly process, greatly reducing the alignment overhead of subsequent deep learning models. The spatiotemporal variational encoding and decoding unit 202 abandons the traditional deterministic thinking of "given an initial state, output a unique future trajectory." Its variational encoding part maps the preceding tensor sequence to a mean and variance, thereby constructing a latent variable probability space; the decoding part reconstructs multiple reasonable trajectories by sampling multiple times from this probability space. Its core purpose is to deeply understand that driving behavior is inherently unpredictable. By outputting the probability distribution boundary containing the expected value and covariance matrix, it can intuitively see the probability of a car turning the steering wheel to the left and making an emergency evasive maneuver to the right when it encounters an obstacle in front of it. Thus, it uses this quantified uncertainty to accommodate extremely complex and sudden road conditions, and establishes the safety redundancy of the construction zone on a rigorous probability theory basis.

[0096] In another preferred embodiment of the present invention, the collision matrix and strategy generation module 300 specifically includes:

[0097] The dynamic collision time calculation unit 301 is used to calculate the minimum collision time of the overlapping region of each probability boundary in the digital twin voxel space and construct a matrix;

[0098] The tiered strategy distribution unit 302 is used to trigger network early warning, electronic fence linkage, or emergency evacuation commands based on the extreme values ​​of the collision time.

[0099] In this embodiment, the operation of the dynamic collision time calculation unit 301 relies on an efficient three-dimensional spatial intersection algorithm. In this unit, the probability prediction boundary of the vehicle is represented as a convex polyhedron or ellipsoid that increases with time, while the virtual electronic fence of the construction area is also a static three-dimensional geometry. This unit uses the GJK algorithm or the SAT separating axis theorem to perform iterative calculations on multiple future time slices to find the timestamp of the first spatial overlap between the two geometries, thereby constructing a dynamic spatiotemporal collision time matrix reflecting the urgency of risk across the entire road segment. Its core purpose is to intuitively transform the abstract trajectory coordinates into a unique, decision-making urgency indicator: how many seconds remain before a physical impact. The tiered strategy distribution unit 302 incorporates a high-response MQTT message queue and a Hardware Abstraction Layer (HAL). Upon receiving an extreme value alarm from the collision matrix, it accurately routes the execution to the corresponding branch based on the preset threshold range into which the extreme value falls. For minor risks, the system routes commands to the cloud-based vehicle networking platform via network interface; for serious risks, it routes commands directly to the edge gateway on-site, controlling flashing lights and loudspeakers via industrial relays; and for extremely dangerous risks, it delivers commands directly to the worker's device via a low-latency wireless radio frequency channel. This ensures that in any emergency situation where communication is restricted or extremely dangerous, the system can find the most effective and fundamental means to protect the lives of construction workers.

[0100] The above embodiments of the present invention provide a construction safety management method and a construction safety management system for highways. By introducing a spatiotemporal variational autoencoder model into the digital twin space, the technical bottleneck of traditional mechanical extrapolation algorithms being unable to predict sudden driving behaviors is overcome, and the random divergence boundary of complex traffic flow trajectories is effectively quantified. At the same time, the vehicle deceleration compliance rate in the real physical space is used as a negative feedback signal to feed back the model weights, realizing the adaptive closed-loop coupling of safety management strategies with dynamic road conditions, weather conditions and driver psychology.

[0101] In order for the above methods and systems to operate smoothly, the system may include more or fewer components than those described above, or combine certain components, or different components, in addition to the various modules mentioned above. For example, it may include input / output devices, network access devices, buses, processors, and memory.

[0102] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (OPGs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the system, connecting various parts via various interfaces and lines.

[0103] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0104] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A construction safety management method for expressways, characterized by, The method includes: Real-time collection of traffic flow and construction operation data in highway construction areas; The traffic flow and construction operation fusion perception data includes at least roadside lidar point cloud data, vehicle network trajectory data, and high-precision positioning data of construction entities. The traffic flow and construction operation fusion perception data are mapped to a pre-constructed digital twin voxel space, and the hidden spatial features within continuous time steps are extracted using a spatiotemporal variational autoencoder model to predict and output the trajectory uncertainty probability distribution of past vehicles and construction entities. Based on the trajectory uncertainty probability distribution, the dynamic spatiotemporal collision time matrix of the intersection area is calculated, and a matching hierarchical collaborative risk avoidance strategy is generated according to the extreme value of the dynamic spatiotemporal collision time matrix. The hierarchical collaborative risk avoidance strategy is issued and executed, while the deceleration compliance rate of passing vehicles after the strategy is executed is monitored. The deceleration compliance rate is used as a feedback signal to adaptively fine-tune the spatiotemporal variational autoencoder model. The real-time collection of traffic flow and construction operation fusion sensing data in the highway construction area specifically includes: The dynamic three-dimensional bounding boxes of passing vehicles are obtained by roadside perception computing units deployed upstream and to the side of the construction area, which serve as the point cloud data of the roadside lidar. The vehicle receives basic safety messages broadcast by surrounding intelligent connected vehicles through the vehicle-road cooperative communication protocol, and extracts vehicle speed, heading angle and brake pedal status as vehicle connected trajectory data. Differential positioning coordinates are obtained by using positioning wristbands worn by construction workers and mobile stations mounted on machinery, which serve as high-precision positioning data for the construction entity. The predicted output of the trajectory uncertainty probability distribution of past vehicles and construction entities specifically includes: After spatiotemporal alignment of the roadside lidar point cloud data, vehicle network trajectory data, and high-precision positioning data of construction entities, they are transformed into a rasterized tensor sequence in the digital twin voxel space. The rasterized tensor sequence is input into the encoder of the spatiotemporal variational autoencoder model to generate latent variables that follow a multivariate Gaussian distribution. The latent variables are reconstructed and extrapolated to future time steps by the decoder of the spatiotemporal variational autoencoder model, and the trajectory uncertainty probability distribution containing the expected value of the predicted trajectory and the covariance matrix is ​​output to characterize the spatial probability boundary that the target may appear in the future. The generation of matching hierarchical collaborative risk avoidance strategies specifically includes: In the digital twin voxel space, the minimum collision time of the overlapping region of each probability boundary is calculated, and the dynamic spatiotemporal collision time matrix is ​​constructed. When the extreme value of the dynamic spatiotemporal collision time matrix is ​​greater than the first threshold, a first-level avoidance strategy is generated, which includes sending a lane change warning command to the connected terminal of the passing vehicle. When the extreme value of the dynamic spatiotemporal collision time matrix is ​​between the first threshold and the second threshold, a second-level risk avoidance strategy is generated, which includes dynamically expanding the virtual electronic fence range of the construction area and triggering the roadside directional millimeter-wave audible and visual alarm device. When the extreme value of the dynamic spatiotemporal collision time matrix is ​​less than the second threshold, a third-level risk avoidance strategy is generated, which includes issuing the highest priority emergency vibration and voice evacuation commands to the wearable and vehicle-mounted equipment of the construction entity. The process of adaptively fine-tuning the spatiotemporal variational autoencoder model by using the deceleration compliance rate as a feedback signal specifically includes: Within the observation window after the graded collaborative risk avoidance strategy is issued, the ratio of the number of vehicles that actually decelerate to the total number of vehicles warned is used to obtain the deceleration compliance rate. Calculate the deviation loss between the deceleration compliance rate and the target expected compliance rate; The bias loss is introduced into the loss function of the spatiotemporal variational autoencoder model, and the model network weights are updated through the backpropagation algorithm, so that the model can automatically expand the covariance range of trajectory prediction in low compliance scenarios.

2. A construction safety management and control system for expressways, characterized by, The system, employing the construction safety management method for highways as described in claim 1, comprises: The perception data fusion and aggregation module is used to collect real-time fusion perception data of traffic flow and construction operations in the highway construction area; The traffic flow and construction operation fusion perception data includes at least roadside lidar point cloud data, vehicle network trajectory data, and high-precision positioning data of construction entities. The twin mapping and probability prediction module is used to map the traffic flow and construction operation fusion perception data to a pre-constructed digital twin voxel space, and use a spatiotemporal variational autoencoder model to extract hidden spatial features within continuous time steps, and predict and output the trajectory uncertainty probability distribution of past vehicles and construction entities. The collision matrix and strategy generation module is used to calculate the dynamic spatiotemporal collision time matrix of the intersection area based on the trajectory uncertainty probability distribution, and generate a matching hierarchical collaborative risk avoidance strategy according to the extreme value of the dynamic spatiotemporal collision time matrix. The execution feedback and adaptive fine-tuning module is used to issue and execute the hierarchical collaborative risk avoidance strategy, while monitoring the deceleration compliance rate of passing vehicles after the strategy is executed, and using the deceleration compliance rate as a feedback signal to adaptively fine-tune the spatiotemporal variational autoencoder model. The sensing data fusion and aggregation module specifically includes: The roadside laser scanning unit is used to acquire dynamic 3D bounding boxes of passing vehicles, and its purpose is to provide robust non-connected vehicle contour detection under complex lighting conditions. The vehicle-road cooperative analysis unit is used to receive basic safety messages broadcast by intelligent vehicles and extract the internal operating status of the vehicles. Its purpose is to obtain the advance movement intentions of vehicles beyond line of sight and those that are obscured. The physical high-precision positioning unit is used to obtain the differential coordinates of construction personnel and machinery. Its purpose is to provide a high-frequency centimeter-level position reference for moving targets within the construction area. The twin mapping and probability prediction module specifically includes: The twin space tensor transformation unit is used to transform aligned data into a rasterized tensor sequence in the digital twin voxel space. The spatiotemporal variational coding and decoding unit is used to generate latent variables and output prediction results containing expected values ​​and covariance matrices; The collision matrix and strategy generation module specifically includes: A dynamic collision time calculation unit is configured to calculate the minimum collision time of each probability boundary overlap region in the digital twin voxel space and construct a matrix. A echelon strategy issuing unit is configured to trigger the networked early warning, the electronic fence linkage, or the emergency evacuation instruction according to the extreme value of the collision time.

Images