Road paving system based on multi-head self-attention mechanism

Abstract

The invention belongs to the technical field of road paving, and particularly relates to a road paving system based on a multi-head self-attention mechanism, which comprises a compaction robot, a paving robot, a multi-modal data sensing layer, an AI dynamic decision core and intelligent execution equipment, the AI dynamic decision core fuses construction time sequence data and a spatial topological relation through a space-time diagram neural network to realize construction progress prediction and multi-device cooperative control, and the intelligent execution device comprises a self-adaptive screed control unit, a dynamic material distribution system, a compaction robot cluster and a non-uniform material compensation control device. According to the method, autonomous optimization and dynamic correction of the whole process of paving, compacting and detecting can be achieved, manual intervention is reduced, the intelligent level is high, the flatness control precision can be improved, the construction efficiency can be improved, material waste can be reduced, and the engineering cost can be reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of road paving, and particularly relates to a road paving system based on a multi-head self-attention mechanism. Background Art

[0002] Traditional road paving technology relies on manual operation and mechanical control, and has disadvantages such as insufficient precision, low efficiency, and lagging quality control.

[0003] In the prior art, although auxiliary means such as GPS and laser positioning have been introduced, there is still a lack of efficient fusion ability for multi-source heterogeneous data, and the control strategies are mostly static feedback, making it difficult to cope with dynamic construction scenarios. Therefore, the present invention proposes an intelligent 3D printing system based on a multi-head self-attention mechanism, which realizes closed-loop optimization of the whole construction process through multi-modal perception, dynamic decision-making, and physical constraint reinforcement learning. Summary of the Invention

[0004] The purpose of the present invention is to provide a road paving system based on a multi-head self-attention mechanism, which can realize autonomous optimization, dynamic correction of the whole process of paving, compaction, and detection, reduce manual intervention, has a high level of intelligence, can improve the flatness control accuracy, improve the construction efficiency, reduce material waste, and reduce the project cost.

[0005] The technical solution adopted by the present invention is specifically as follows:

[0006] A road paving system based on a multi-head self-attention mechanism includes a compaction robot, a paving robot, a multi-modal data perception layer, an AI dynamic decision-making core, and intelligent execution devices;

[0007] The multi-modal data perception layer is used to collect construction environment, material, and equipment status data;

[0008] The AI dynamic decision-making core fuses construction time-series data and spatial topological relationships through a spatio-temporal graph neural network to realize construction progress prediction and multi-device collaborative control;

[0009] The intelligent execution devices include an adaptive screed control unit, a dynamic batching system, a compaction robot cluster, and a non-uniform material compensation control device.

[0010] Further, the multi-modal data perception layer includes a lidar + 3D vision scanning module, an embedded sensor matrix, an on-line material gradation detection module, and an environment perception unit;

[0011] The lidar + 3D vision scanning module is used to generate a three-dimensional point cloud model in real time, realize real-time monitoring of paving quality, and monitor paving thickness and material distribution uniformity;

[0012] The embedded sensor matrix is used to collect the pressure distribution and soil compactness in real time during the compaction process;

[0013] The on-line material gradation detection module is used to monitor the aggregate gradation and asphalt content in real time;

[0014] The environmental perception unit is used to monitor the temperature, humidity of the construction environment and the state of the underground structural layer in real time.

[0015] Furthermore, the embedded sensor matrix includes an embedded pressure sensor module, an acceleration sensor, a temperature sensor, a high-precision positioning and orientation receiver, and a VCV compactness sensor;

[0016] The embedded pressure sensor module is installed on the roller of the road roller to monitor the pressure distribution in real time during the compaction process;

[0017] The acceleration sensor monitors the vibration of the steel wheel of the road roller, and judges the density and compactness of the soil according to the changes in amplitude and frequency;

[0018] The pressure sensor measures the pressure of the tire on the ground to judge the compactness of the soil;

[0019] The temperature sensor monitors the temperature during the operation of the road roller;

[0020] The high-precision positioning and orientation receiver and the VCV compactness sensor are installed on the road roller.

[0021] Furthermore, the on-line material gradation detection module includes a high-speed camera, an edge computing image processing algorithm, and a near-infrared spectrometer;

[0022] The high-speed camera is installed at the aggregate inlet, and is used to capture the image data of the aggregate in real time and process it, and extract the characteristics of the aggregate including particle size and shape;

[0023] The edge computing image processing algorithm is used to calculate the particle size distribution of the aggregate according to the image data of the aggregate captured by the high-speed camera, evaluate the aggregate gradation, and realize the on-line real-time monitoring of the aggregate quality;

[0024] The near-infrared spectrometer is installed at the aggregate outlet, and determines the asphalt content through spectral data analysis.

[0025] Furthermore, the environmental perception unit includes a temperature and humidity sensor and a ground radar;

[0026] The temperature and humidity sensor monitors the temperature and humidity parameters of the construction environment in real time;

[0027] The ground radar collects information about the underground structure.

[0028] Furthermore, the AI dynamic decision-making core includes physical constraint enhanced learning, distributed model predictive control, and spatio-temporal graph neural network;

[0029] The physical constraint enhanced learning embeds the material mechanics equation as a constraint condition into the deep reinforcement learning to optimize the paving thickness control;

[0030] Objective function:

[0031]

[0032] where represents the sensitivity of stress to paving thickness, f(ρ, E) is a function of material density and elastic modulus, σ is stress, h is paving thickness, ρ is density, E is elastic modulus, and Q(s, a) is the action value function, which is used to evaluate the expected cumulative return that the agent can obtain in the future after executing action a in state s;

[0033] Among them, the distributed model predictive control is used to coordinate multiple devices to optimize the global flatness and balance the control input and construction continuity;

[0034] Minimize the error between the device output and the target value:

[0035]

[0036] where y(t) is the actual output of the device, y ref is the reference output, u i (t) is the control input of the device, and λ is the regularization factor, which is used to balance the magnitude of the control input while ensuring the smoothness and continuity of the device control input;

[0037] The spatio-temporal graph neural network is used to fuse the construction time-series data and the spatial topological relationship to predict the construction progress, material requirements, and coordinate the actions of multiple devices.

[0038] Furthermore, the construction and application of the spatio-temporal graph neural network include the following steps:

[0039] Step 1: Capture the spatial topological relationship in the construction scene through the graph convolutional network, aggregate the feature information of neighboring nodes through the graph convolutional operation, extract the spatial correlation between devices, and provide support for dynamic decision-making in the spatial dimension. The specific elements include:

[0040] Device location: The paving robot and the compaction robot are used as graph nodes, and the node features include the real-time coordinates and status parameters of the devices;

[0041] Flow direction: Using the material transportation path as the edge, define the interaction relationship between devices;

[0042] Adjacency matrix: An adjacency matrix is constructed based on the device location and the material flow direction to quantify the connection strength between nodes;

[0043] Step 2: Time-dependence modeling. Process the time-series data during the construction process through a temporal convolutional network to capture the temporal evolution law of construction parameters, providing a basis in the time dimension for prediction and optimization. The key elements include:

[0044] Time-series data: Collect the temporal changes in device status, environmental parameters, and material properties;

[0045] Dynamic feature extraction: Use one-dimensional convolution or recurrent neural networks to extract temporal features;

[0046] Step 3: Construction progress prediction and optimization. Based on the spatio-temporal feature fusion ST-GNN, achieve the following prediction and control objectives:

[0047] Construction progress prediction: Combine historical data and real-time features to predict the construction completion degree in the future time period;

[0048] Material distribution prediction: Dynamically calculate the material requirements according to the grading test results and device status, and optimize the resource scheduling;

[0049] Coordinated control: Through distributed model predictive control, coordinate the actions of multiple devices to ensure construction continuity and reduce conflicts;

[0050] Step 4: Architecture interaction process input layer. The multi-modal perception layer uploads spatial data, time-series data, and environmental data in real time;

[0051] Spatio-temporal feature fusion: Aggregate spatial features including device location and material flow direction; Extract the temporal dependence of construction parameters, and weight and fuse spatio-temporal features to generate a global construction state representation;

[0052] Decision output: Generate path planning instructions for the compaction robot, issue height adjustment parameters for the screed, and feedback the grading compensation signal to the dynamic batching system.

[0053] Furthermore, the adaptive screed control unit includes:

[0054] A hydraulic servo system, installed on the paver, drives the lifting and attitude adjustment of the paver screed through a pressure closed-loop control mechanism, dynamically adjusting the paving thickness and slope;

[0055] A high-precision laser positioning module, used to obtain the three-dimensional spatial coordinates of the screed in real time;

[0056] A feedforward control module, based on the material settlement prediction model Calculate the target adjustment amount Δh of the screed in advance pre, and perform pre-compensation through a hydraulic servo system to offset the expected settlement error caused by the self-weight of the material and temperature changes, where ρ is the material density and k is the settlement coefficient;

[0057] The real-time feedback PID control module continuously monitors the actual height h of the paving layer through a high-precision laser positioning module real , calculate the deviation value e(t) = h ref from the target height h ref - h real , and output a correction amount based on the PID algorithm:

[0058]

[0059] where K p , K i , K d are the proportional, integral, and differential gain coefficients respectively, e(t) represents the real-time error signal, dτ represents the time variable τ, represents the instantaneous change rate of the error e(t) with respect to time, and the correction amount u(t) is fed back to the hydraulic servo system in real time to achieve dynamic error elimination, ensuring that the paving flatness ≤ ±2 mm / 3 m.

[0060] Furthermore, the intelligent compaction robot cluster includes multiple robots. Based on the compaction energy dynamic allocation algorithm of the density cloud map, it coordinates each other's paths and actions through the algorithm to achieve multi-agent collaborative path planning.

[0061] Furthermore, when the non-uniform material compensation control device detects abnormal gradation in a local area, it realizes micro-compensation by adjusting the parameters of the feeding system:

[0062]

[0063] where d 10 and d 60 are the 10% and 60% sieve particle sizes of the material particles respectively, represents the continuity of the particle size distribution of the material.

[0064] The technical effects achieved by the present invention are:

[0065] A road paving system based on the multi-head self-attention mechanism of the present invention realizes the full-process autonomous optimization, dynamic correction of paving, compaction, and detection through the "perception - decision - execution" closed-loop process, reduces manual intervention, has a high level of intelligence, can improve the flatness control accuracy, improve the construction efficiency, reduce material waste, and reduce the project cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is the system architecture diagram of the present invention;

[0067] Figure 2 It is the system architecture diagram of the ST-GNN of the present invention. Specific implementation manners

[0068] In order to make the objectives and advantages of the present invention clearer and more understandable, the present invention will be specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific implementation manners of the present invention, and does not strictly limit the scope of protection specifically claimed by the present invention.

[0069] As Figure 1-2 shown, a road paving system based on the multi-head self-attention mechanism includes a compaction robot, a paving robot, and three major modules: a multi-modal data perception layer, an AI dynamic decision-making core, and intelligent execution devices. Taking the multi-head self-attention mechanism as the technical main line, it realizes the integrated closed-loop control of "perception - decision - execution".

[0070] Among them, the multi-modal data perception layer is used to collect construction environment, material, and equipment status data;

[0071] Specifically, the multi-modal data perception layer includes a lidar + 3D vision scanning module, an embedded sensor matrix, an on-line material gradation detection module, and an environment perception unit;

[0072] In this system, the lidar + 3D vision scanning module is the "visual center" of the intelligent construction system. Through high-precision three-dimensional data acquisition and analysis, it is used to generate a high-precision three-dimensional point cloud model in real time, realize real-time monitoring of paving quality, and monitor paving thickness and material distribution uniformity. By comprehensively scanning the construction site, three-dimensional data of the road surface is obtained, and the thickness and uniformity of the paving layer are accurately evaluated to ensure construction quality.

[0073] Lidar (LiDAR) is a technology that uses laser beams to measure the distance of objects. By emitting laser pulses and receiving their reflected signals, the exact position of the object is calculated. Compared with traditional two-dimensional imaging, lidar three-dimensional imaging has advantages such as rich information, good initiative, and strong anti-interference ability, and has been widely used in multiple fields such as remote sensing reconnaissance, unmanned driving, and aerospace. 3D vision scanning technology captures the three-dimensional shape of an object through a high-speed camera and generates a high-precision three-dimensional model in combination with computer vision algorithms. This technology can provide rich visual information and is suitable for object recognition and modeling in complex environments.

[0074] Among them, the embedded sensor matrix is the "tactile nerve" of the intelligent construction system, which is used to collect the pressure distribution and soil compaction degree during the compaction process in real time, provide key data for construction quality control, dynamic decision-making, and equipment coordination, and finally realize efficient, accurate, and safe road paving and compaction operations;

[0075] The embedded sensor matrix includes an embedded pressure sensor module, a pressure sensor, an acceleration sensor, a temperature sensor, a high-precision positioning and orientation receiver, and a VCV compaction degree sensor;

[0076] The embedded pressure sensor module is installed on the roller of the road roller and can monitor the pressure distribution during the compaction process in real time. These sensors judge the compaction degree by sensing the physical properties of the soil. The acceleration sensor is used to monitor the vibration of the steel wheel of the road roller, and judge the density and compaction degree of the soil according to the changes in amplitude and frequency. The pressure sensor directly measures the pressure of the tire on the ground to judge the compactness of the soil. The temperature sensor monitors the temperature during the operation of the road roller to prevent equipment damage or affect the construction quality caused by overheating. By collecting sensor data during the compaction process in real time, a compaction energy distribution map can be drawn to visually display the compaction degree of different areas. By installing a high-precision positioning and orientation receiver and a VCV compaction degree sensor on the roadbed road roller, indicators such as the compaction pile number position, compaction speed, compaction passes, and VCV compaction degree are recorded in real time, and the relevant data is uploaded to the system platform to realize real-time monitoring of the construction quality.

[0077] Among them, the online material gradation detection module is the "quality goalkeeper" of the intelligent construction system. By monitoring the aggregate gradation and asphalt content in real time, processing data with edge computing, dynamically adjusting construction parameters, and realizing the closed-loop control of material quality, the paving accuracy, construction efficiency, and material utilization rate are ultimately improved.

[0078] The online material gradation detection module includes a high-speed camera, an edge computing image processing algorithm, and a near-infrared spectrometer;

[0079] The high-speed camera is installed at the aggregate inlet to capture and process the image data of the aggregate in real time, and extract the characteristics such as the particle size and shape of the aggregate;

[0080] The edge computing image processing algorithm is used to calculate the particle size distribution of the aggregate according to the image data of the aggregate captured by the high-speed camera, evaluate the aggregate gradation situation, and realize the online real-time monitoring of the aggregate quality;

[0081] The near-infrared spectrometer is installed at the aggregate outlet. By analyzing the spectral data, the asphalt content is determined to ensure the quality of the mixture discharged;

[0082] By processing and fusing data from different sensors through a road paving system based on the multi-head self-attention mechanism, the model's ability to capture complex spatio-temporal features is improved. For example, a spatio-temporal graph neural network model based on the multi-head self-attention mechanism has been used for traffic flow prediction and can effectively extract the spatio-temporal features of the road network.

[0083] The environmental perception unit is the "environmental sentinel" of the intelligent construction system. By monitoring the temperature, humidity and the state of the underground structure layer in real time, it provides key data support for construction process optimization, safety guarantee and dynamic decision-making, and ultimately improves the project quality, efficiency and environmental adaptability.

[0084] The environmental perception unit includes a temperature and humidity sensor and a ground radar;

[0085] Temperature and humidity have an important impact on the setting and hardening process of concrete. By monitoring the temperature and humidity parameters of the construction environment in real time through the temperature and humidity sensor, the construction process can be optimized to ensure the quality of concrete;

[0086] Regarding the state of the underground structure layer, including cracks, cavities, etc., the ground radar can be used to collect detailed information of the underground structure, which helps to evaluate the bearing capacity and stability of the foundation.

[0087] In the pavement paving project, the underground layer structure usually refers to the load-bearing and functional layer between the surface layer and the subgrade in the pavement structure, mainly including the base course, cushion course and possibly the sub-base course. These structure layers jointly bear the traffic load and improve the mechanical and environmental performance of the subgrade, and are the key part of the overall stability of the pavement.

[0088] Among them, the AI dynamic decision-making core fuses construction time-series data and spatial topological relationships through a spatio-temporal graph neural network. The multi-head self-attention mechanism processes spatio-temporal features in different sub-spaces in parallel through multiple groups of independent attention heads, dynamically assigns weights to strengthen key associations (such as the collaborative relationship between the equipment position and the material flow direction), and realizes construction progress prediction and coordinated control of multiple devices (compaction robots, paving robots and intelligent execution devices);

[0089] The AI dynamic decision-making core includes physics-informed reinforcement learning, distributed model predictive control and spatio-temporal graph neural network;

[0090] Physics-Informed Reinforcement Learning (PIRL) is a method that embeds physical laws into deep reinforcement learning. By adding physical constraints to the loss function, PIRL can guide the agent to follow known physical laws during the learning process, thereby improving the robustness and interpretability of the model.

[0091] Embed the material mechanics equation (compaction work transfer model) as a constraint condition into deep reinforcement learning to optimize the paving thickness control.

[0092] Objective function:

[0093]

[0094] Among them, Indicates the sensitivity of stress to paving thickness, and f(ρ, E) is a function of material density and elastic modulus.

[0095] Among them, σ is stress, h is paving thickness, ρ is density, E is elastic modulus, and Q(s, a) is the Action-Value Function, which is used to evaluate the expected cumulative return (long-term benefit) that the agent can obtain in the future after performing action a in state s.

[0096] Among them, Distributed Model Predictive Control (DMPC) is used to coordinate multiple devices to optimize the global flatness and balance the control input and construction continuity.

[0097] Minimize the error between the device output and the target value:

[0098]

[0099] As mentioned above, it involves the kinematic and dynamic models of the device, and these models define the behavior limitations of the device. y(t) is the actual output of the device, and y ref is the reference output, that is, the target or expected construction effect (such as flatness, thickness, etc.). u i (t) is the control input of the device (such as rotational speed, position, etc.), and λ is the regularization factor, which is used to balance the magnitude of the control input while ensuring the smoothness and continuity of the device control input. Ensuring these constraints helps prevent the device from being damaged due to excessive operation and ensures the stable operation of the device.

[0100] The role of the Spatio-Temporal Graph Neural Network (ST-GNN) is to fuse construction time-series data and spatial topological relationships to predict construction progress, material requirements, and coordinate the actions of multiple devices.

[0101] ST-GNN is an architecture that combines Graph Neural Network (GNN) and sequence models, aiming to process graph data with spatio-temporal dependencies. It captures spatial relationships through graph convolution operations and captures time dependencies through temporal convolution or Recurrent Neural Network (RNN). This system uses ST-GNN to jointly model construction time-series data and spatial topological relationships (device positions, material flow directions) ( Figure 2 )

[0102] Specifically, the construction and application of ST-GNN include the following steps:

[0103] Step 1: This module captures the spatial topological relationships in the construction scene through the Graph Convolution Network (GraphConv), aggregates the feature information of neighboring nodes through graph convolution operations, extracts the spatial correlation between devices, and provides spatial dimension support for dynamic decision-making. Specifically, it includes the following elements:

[0104] Equipment Location: Consider equipment such as paving robots and compaction robots as graph nodes. The node features include the real-time coordinates of the equipment and status parameters (such as paving speed, compaction energy).

[0105] Flow Direction: Define the interaction relationships between equipment with the material transportation path as the edge (such as the paving robot transferring settlement correction parameters to the compaction robot).

[0106] Adjacency Matrix (adj_matrix): Construct an adjacency matrix based on equipment location and material flow direction to quantify the connection strength between nodes. For example, the interaction weights of adjacent equipment are higher, while those of distant equipment are lower.

[0107] Step 2: Temporal Dependency Modeling. This module processes the time-series data during the construction process through a Temporal Convolutional Network (TemporalConv) to capture the temporal evolution laws of construction parameters and provide a time dimension basis for prediction and optimization. The key elements include:

[0108] Time-Series Data: Collect the temporal variations of equipment status (such as screed height Δh, feeder speed RPM), environmental parameters (temperature and humidity), and material properties (grading, density).

[0109] Dynamic Feature Extraction: Use one-dimensional convolution or a Recurrent Neural Network (RNN) to extract temporal features, such as the fluctuation trend of paving thickness and the cumulative effect of compaction energy.

[0110] Step 3: Construction Progress Prediction and Optimization. Based on spatio-temporal feature fusion ST-GNN, achieve the following prediction and control objectives:

[0111] Construction Progress Prediction: Combine historical data and real-time features to predict the construction completion in future time periods (such as paving area, compaction passes).

[0112] Material Allocation Prediction: Dynamically calculate material requirements (such as asphalt consumption, aggregate ratio) based on grading test results and equipment status to optimize resource scheduling.

[0113] Collaborative Control: Through Distributed Model Predictive Control (DMPC), coordinate the actions of multiple devices to ensure construction continuity and reduce conflicts. If a conflict occurs, ST-GNN will trigger a dynamic material distribution adjustment system to adjust the speed (ΔRPM) and baffle angle of the spiral feeder to achieve local compensation.

[0114] Step 4: Input Layer of the Architecture Interaction Process. The multi-modal perception layer uploads spatial data (3D point cloud), temporal data (equipment status), and environmental data (temperature and humidity) in real time.

[0115] Spatio-temporal feature fusion (GraphConv): Aggregate spatial features such as device location and material flow direction; extract the temporal dependencies of construction parameters, and weightedly fuse spatio-temporal features to generate a global construction state representation;

[0116] Decision output, generate compaction robot path planning instructions, issue screed height adjustment parameters (±Δh), and feedback gradation compensation signals to the dynamic batching system (screw feeder).

[0117] Among them, the intelligent execution devices include an adaptive screed control unit, a dynamic batching system, a compaction robot cluster, and a non-uniform material compensation control device;

[0118] The adaptive screed control unit adopts a collaborative architecture of a hydraulic servo system and high-precision laser positioning technology, and combines a composite strategy of feedforward control based on predicted settlement and real-time error feedback PID control to achieve precise adjustment of paving flatness. Its technical composition and functions include:

[0119] ① Hydraulic servo system:

[0120] The hydraulic servo system is installed on the paving robot. Through a pressure closed-loop control mechanism, this system drives the lifting and attitude adjustment of the paving robot's screed, outputs a high-precision linear force (accuracy up to ±0.1 kN), and ensures millimeter-level control of the device position and movement trajectory (positioning error ≤ ±0.5 mm). During the paving operation, the hydraulic servo system directly acts on the screed, dynamically adjusting the paving thickness and slope according to the decision-making instructions to adapt to the rheological characteristics of the material and environmental disturbances.

[0121] ② High-precision laser positioning module:

[0122] Configure a lidar and a multi-axis laser scanner to obtain the three-dimensional spatial coordinates of the screed in real time (resolution 0.1 mm), and construct a three-dimensional point cloud model of the paving layer. The laser positioning data is compared with the construction design (BIM / CAD) in real time to generate a position deviation signal, providing an input reference for the control algorithm.

[0123] ③ Feedforward control module:

[0124] Based on the material settlement prediction model Calculate the target adjustment amount Δh of the screed in advance pre , and perform pre-compensation through the hydraulic servo system to offset the expected settlement error caused by material self-weight and temperature changes.

[0125] ρ(τ)dτ is the differential-integral expression of the material density with respect to time, where ρ is the material density, k is the sedimentation coefficient, ρ(τ) represents the material density, and dτ is the time, that is, the density changes with time (for example, during the paving process, factors such as the compaction degree, paving speed, and temperature of the material will cause the density to change in real time). Its core significance lies in quantifying the dynamic cumulative effect of the material density during the construction process and providing a mathematical basis for the predictive control of the intelligent paving system.

[0126] ④ Real-time feedback PID control module:

[0127] Continuously monitor the actual height h of the paving layer through the high-precision laser positioning module real , calculate the deviation value e(t) = h ref - h ref with the target height h real , and output the correction amount based on the PID algorithm:

[0128]

[0129] where K p , K i , K d are the proportional, integral, and differential gain coefficients respectively.

[0130] u(t) represents the output correction amount of the PID controller (unit: such as voltage, hydraulic pressure, or displacement), and the correction amount u(t) is fed back to the hydraulic servo system in real time to achieve dynamic error elimination and ensure that the paving flatness ≤ ±2mm / 3m.

[0131] e(t) represents the real-time error signal, which characterizes the deviation between the actual thickness of the current paving layer and the design value. e(t) = href(t) - hreal(t), where:

[0132] href(t): Target paving thickness (based on the design model);

[0133] hreal(t): Paving thickness measured by the laser positioning system.

[0134] dτ is the differential of the time variable τ, representing a very short time segment.

[0135] represents the instantaneous change rate of the error e(t) with respect to time, that is:

[0136]

[0137] By calculating the change rate of the error, the system can predict the increasing or decreasing direction of future deviations.

[0138] The dynamic fabric automatic adjustment system includes a variable-frequency drive for the spiral feeder + an angle-adjustable baffle, which dynamically adjusts the rotational speed (RPM) and the material throwing angle according to the gradation detection result. The spiral feeder is a common material conveying device, often used in civil engineering, construction, and other engineering fields. It conveys materials from the feeding end to the output end through spiral conveyor blades. Variable-frequency drive (VFD) technology controls the conveying speed of the spiral feeder by adjusting the rotational speed of the motor. The angle-adjustable baffle is a baffle set at the discharge port of the spiral feeder, which can be rotated by a motor drive to adjust the angle of the baffle and dynamically adjust the material throwing angle.

[0139] The intelligent compaction robot cluster includes multiple robots. Based on the compaction energy dynamic distribution algorithm of the density cloud map, it coordinates each other's paths and actions through the algorithm to achieve multi-agent collaborative path planning, so as to ensure the overall efficiency and task optimization. Each robot acts as an agent and can exchange information during the collaboration process to avoid conflicts and maximize the task completion rate.

[0140] When the non-uniform material compensation control device detects abnormal gradation in a local area, it realizes microscopic compensation by adjusting the fabric system parameters:

[0141]

[0142] where d 10 and d 60 are the 10% and 60% sieve sizes of the material particles respectively, represents the continuity of the particle size distribution of the material and reflects the uniformity of the material. By adjusting the rotational speed (RPM), the system can compensate for these local non-uniformities and fine-tune the operating parameters of the feeder in real time to ensure the uniformity of the laid material.

[0143] In summary, the application of this system can achieve a flatness control accuracy of ±2 mm / 3 m (±5 mm / 3 m for traditional methods), with an expected construction efficiency increase of 40% and a 30% reduction in material waste, reducing project costs. At the same time, it has a high level of intelligence and realizes autonomous optimization, dynamic correction, and reduced manual intervention in the whole process of paving, compaction, and detection through the "perception - decision - execution" closed-loop process.

[0144] This system can be applied to road projects with high-precision requirements such as highways and airport runways; it can be extended to complex scenarios such as bridge paving and tunnel lining, and is compatible with various materials such as asphalt and concrete. The intelligent compaction technology reduces carbon emissions (with an expected 22% reduction in energy consumption), meeting the concept of green construction.

[0145] To further verify the effect of this technical solution, experiments were conducted on the above technical solution, and the experimental results are as follows:

[0146]

[0147]

[0148] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. The structures, devices, and operation methods not specifically described and explained in the present invention are implemented by conventional means in the art without special description and limitation.

Claims

1. A road paving system based on a multi-head self-attention mechanism, characterized by: It includes a compaction robot, a paving robot, a multi-modal data perception layer, an AI dynamic decision-making core, and intelligent execution devices; The multi-modal data perception layer is used to collect construction environment, material, and equipment status data; The AI dynamic decision-making core fuses construction time-series data and spatial topological relationships through a spatio-temporal graph neural network to achieve construction progress prediction and multi-device collaborative control; The intelligent execution devices include an adaptive screed control unit, a dynamic material distribution system, a compaction robot cluster, and a non-uniform material compensation control device.

2. The road paving system based on the multi-head self-attention mechanism according to claim 1, characterized in that: The multi-modal data perception layer includes a lidar + 3D vision scanning module, an embedded sensor matrix, an on-line material gradation detection module, and an environment perception unit; The lidar + 3D vision scanning module is used to generate a three-dimensional point cloud model in real time, realize real-time monitoring of paving quality, and monitor paving thickness and material distribution uniformity; The embedded sensor matrix is used to collect the pressure distribution and soil compaction degree during the compaction process in real time; The on-line material gradation detection module is used to monitor the aggregate gradation and asphalt content in real time; The environment perception unit is used to monitor the temperature, humidity of the construction environment and the state of the underground structural layer in real time.

3. The road paving system based on the multi-head self-attention mechanism according to claim 2, wherein: The embedded sensor matrix includes an embedded pressure sensor module, an acceleration sensor, a temperature sensor, a high-precision positioning and orientation receiver, and a VCV compaction degree sensor; The embedded pressure sensor module is installed on the roller of the road roller to monitor the pressure distribution during the compaction process in real time; The acceleration sensor monitors the vibration of the steel wheel of the road roller, and judges the density and compaction degree of the soil according to the changes in amplitude and frequency; The pressure sensor measures the pressure of the tire on the ground to judge the compaction degree of the soil; The temperature sensor monitors the temperature during the operation of the road roller; The high-precision positioning and orientation receiver and the VCV compaction degree sensor are installed on the road roller.

4. The road paving system based on a multi-head self-attention mechanism according to claim 3, characterized in that: The on-line material gradation detection module includes a high-speed camera, an edge computing image processing algorithm, and a near-infrared spectrometer; The high-speed camera is installed at the aggregate inlet to capture and process the image data of the aggregate in real time, and extract the characteristics of the aggregate including particle size and shape; The edge computing image processing algorithm is used to calculate the particle size distribution of the aggregate according to the image data of the aggregate captured by the high-speed camera, evaluate the aggregate gradation situation, and realize on-line real-time monitoring of the aggregate quality; The near-infrared spectrometer is installed at the aggregate outlet to determine the asphalt content through spectral data analysis.

5. The road paving system based on a multi-head self-attention mechanism according to claim 4, characterized in that: The environment perception unit includes a temperature and humidity sensor and a ground radar; The temperature and humidity sensor monitors the temperature and humidity parameters of the construction environment in real time; The ground radar collects information on the underground structure.

6. The road paving system based on the multi-head self-attention mechanism according to claim 1, characterized in that: The AI dynamic decision-making core includes physical constraint enhanced learning, distributed model predictive control, and spatio-temporal graph neural network; The physical constraint enhanced learning embeds the material mechanics equation as a constraint condition into the deep reinforcement learning to optimize the paving thickness control; Objective function: Among them, represents the sensitivity of stress to paving thickness, f(ρ, E) is a function of material density and elastic modulus, σ is stress, h is paving thickness, ρ is density, E is elastic modulus, and Q(s, a) is the action value function, which is used to evaluate the expected cumulative return that the agent can obtain in the future after executing action a in state s; Among them, the distributed model predictive control is used to collaborate with multiple devices to optimize the global flatness and balance the control input and construction continuity; Minimize the error between the device output and the target value: where y(t) is the actual output of the device, y ref is the reference output, u i (t) is the control input of the device, and λ is the regularization factor used to balance the magnitude of the control input while ensuring the smoothness and continuity of the device control input; The spatiotemporal graph neural network is used to fuse construction time series data with spatial topological relationships, predict construction progress, material requirements, and coordinate the actions of multiple devices.

7. The road paving system based on a multi-head self-attention mechanism according to claim 6, characterized in that: The construction and application of the spatiotemporal graph neural network includes the following steps: Step 1: Capture the spatial topological relationships in the construction scene through a graph convolutional network. Aggregate the feature information of neighboring nodes through graph convolution operations, extract the spatial correlation between devices, and provide spatial dimension support for dynamic decision-making. Specifically, it includes the following elements: Equipment location: The paving robot and compacting robot are used as graph nodes, and the node features include the real-time coordinates and status parameters of the equipment; Flow direction: Define the interaction between devices using the material conveying path as the edge; Adjacency matrix: Construct an adjacency matrix based on equipment location and material flow direction to quantify the connection strength between nodes; Step 2: Time-dependent modeling. This involves processing time-series data from the construction process using a temporal convolutional network to capture the temporal evolution of construction parameters and provide a temporal basis for prediction and optimization. Key elements include: Time series data: collects time series changes in equipment status, environmental parameters, and material properties; Dynamic feature extraction: Extracting temporal features using one-dimensional convolutional or recurrent neural networks; Step 3: Construction progress prediction and optimization, based on spatiotemporal feature fusion ST-GNN to achieve the following prediction and control objectives: Construction progress forecasting, combining historical data with real-time features to predict construction completion in future time periods; Material allocation prediction: dynamically calculate material demand and optimize resource scheduling based on grading test results and equipment status; Collaborative control, through distributed model predictive control, coordinates the actions of multiple devices to ensure construction continuity and reduce conflicts; Step 4: Build an interactive process input layer and a multimodal perception layer to upload spatial data, time series data, and environmental data in real time; Spatiotemporal feature fusion: Aggregate spatial features including equipment location and material flow direction; extract the temporal dependency of construction parameters, weightedly fuse spatiotemporal features, and generate a global construction status representation; The decision output generates the compaction robot path planning instructions, issues the screed height adjustment parameters, and feeds back the grading compensation signal to the dynamic material distribution system.

8. The road paving system based on a multi-head self-attention mechanism according to claim 6, characterized in that: The adaptive screed control unit comprises: The hydraulic servo system, installed on the paving robot, drives the lifting and posture adjustment of the paving robot's screed through a pressure closed-loop control mechanism, dynamically adjusting the paving thickness and slope; High-precision laser positioning module, used to obtain the three-dimensional coordinates of the screed in real time; The feedforward control module, based on the material settlement prediction model Calculates in advance the target adjustment amount Δh of the screed pre , and performs pre-compensation through the hydraulic servo system to offset the expected settlement error caused by the self-weight of the material and temperature changes, where ρ is the material density and k is the settlement coefficient; Real-time feedback PID control module, continuous monitoring of the actual height of the paving layer through the high-precision laser positioning module real , calculate the target height h ref The deviation value e(t)=h ref -h real , and output the correction value based on the PID algorithm: Where K p , K i , K d are proportional, integral, and differential gain coefficients respectively, e(t) represents the real-time error signal, dτ represents the time variable τ, It represents the instantaneous rate of change of the error e(t) over time, and the correction value u(t) is fed back to the hydraulic servo system in real time to achieve dynamic error elimination.

9. The road paving system based on the multi-head self-attention mechanism according to claim 8, wherein: The intelligent compaction robot cluster includes multiple robots, and a dynamic compaction energy distribution algorithm based on a density cloud map. The algorithm coordinates each other's paths and actions to achieve multi-agent collaborative path planning.

10. A road paving system based on a multi-head self-attention mechanism according to claim 8, characterized in that: When the non-uniform material compensation control device detects abnormal gradation in a local area, it implements micro-compensation by adjusting the material distribution system parameters: where d 10 and d 60 are the 10% and 60% sieve sizes of the material particles respectively, indicating the continuity of the particle size distribution of the material.

Images