A method and system for adjusting parameters of an intelligent road roller based on neural networks

Abstract

This invention relates to the field of artificial intelligence technology and discloses a method and system for adjusting the parameters of an intelligent road roller based on a neural network. The method includes: acquiring image data, temperature distribution data, vibration distribution data, and moisture content distribution data of the road surface to be constructed; identifying the degree of segregation based on the image data and temperature distribution data using a pre-constructed cross-modal spatiotemporal cross-attention model to obtain a segregation distribution heatmap; performing temperature compensation on the compaction distribution data calculated from the vibration distribution data based on the temperature distribution data to obtain compensated compaction distribution data; and inputting the segregation distribution heatmap, temperature distribution data, moisture content distribution data, and compensated compaction distribution data into a pre-constructed dynamic coupling model to obtain optimal road roller operation parameters. The method of this application improves the accuracy of segregation detection and compaction measurement, and realizes real-time, adaptive, and refined adjustment of road roller operation parameters.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for adjusting the parameters of an intelligent road roller based on neural networks. Background Technology

[0002] Road surface compaction is a core process in road construction to ensure its long-term durability and performance. Traditional compaction quality control mainly relies on post-construction sampling and operator experience, which has obvious lag and subjectivity, and cannot achieve real-time and continuous quality control across the entire work surface.

[0003] With the development of sensor technology, real-time monitoring and dynamic control of operating parameters have become the core direction for improving compaction accuracy. However, existing technologies mainly rely on data collected by a single sensor for monitoring and control, resulting in the following systemic defects: Segregation detection methods are limited; existing technologies typically rely solely on visible light image analysis to identify the degree of segregation. When lighting changes, the surface is wet, or stains are present, the reliability of visual features drops sharply, leading to insufficient segregation detection accuracy. Compaction degree measurement is severely affected by operating conditions, raising doubts about its accuracy. Existing technologies acquire vibration data through acceleration sensors installed on the roller and calculate compaction degree based on soil mechanics models. However, the true compaction degree is affected by the temperature and moisture content of the mixture, causing distortion in compaction degree measurements under varying operating conditions. Data is isolated, decision-making is one-sided, and systematic coupling optimization is lacking. The compaction process is a dynamic process with strong coupling of multiple physical fields; isolated data analysis cannot reveal its inherent correlation, resulting in a mismatch between the obtained roller operation parameters and the real-time state of the pavement material.

[0004] Therefore, improving the accuracy of segregation detection and compaction measurement during road rolling operations, as well as the matching degree between road rolling operation parameters and real-time road surface conditions, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method and system for adjusting the parameters of an intelligent road roller based on a neural network, in order to solve the technical problems of how to improve the accuracy of segregation detection and compaction measurement in road rolling operations, as well as the matching degree between road rolling operation parameters and real-time road surface conditions, thereby achieving the effect of improving the accuracy of segregation detection and compaction measurement in road rolling operations, and real-time, adaptive, and refined control of road roller operation parameters.

[0006] In a first aspect, the present invention provides a method for adjusting the parameters of an intelligent road roller based on a neural network, the method comprising:

[0007] Acquire multi-source data of the road surface to be constructed during roller operation, including image data, temperature distribution data, vibration distribution data, and moisture content distribution data;

[0008] Based on the image data and the temperature distribution data, the degree of segregation is identified through a pre-constructed cross-modal spatiotemporal cross-attention model to obtain a segregation distribution heatmap. The cross-modal spatiotemporal cross-attention model is configured to perform cross-attention fusion of spatiotemporal visual features and spatiotemporal thermal features to generate the segregation distribution heatmap. The spatiotemporal visual features are obtained by extracting dual-channel features from the image data using a two-dimensional convolutional neural network and an optical flow network. The spatiotemporal thermal features are obtained by extracting and fusing features from the temperature distribution data sequentially using a convolutional long short-term memory network based on an attention mechanism.

[0009] Based on the vibration distribution data, compaction degree distribution data is obtained, and temperature compensation is performed on the compaction degree distribution data based on the temperature distribution data to obtain compensated compaction degree distribution data.

[0010] The segregation distribution heat map, the temperature distribution data, the moisture content distribution data, and the compensated compaction degree distribution data are used as real-time feedback data input into a pre-constructed dynamic coupling model to obtain the optimal rolling operation parameters.

[0011] Preferably, the acquisition of multi-source data on the road surface to be constructed during roller operation includes:

[0012] Receive the second pulse signal and time message output by the GNSS timing module, and use the second pulse signal and time message as the absolute time reference to synchronize the master clock;

[0013] Based on the master clock, continuous sampling is performed through the positioning module and the piezoelectric accelerometer to obtain position timing data and vibration distribution data with the same timestamp.

[0014] Based on the location timing data, corresponding trigger signals are generated according to different preset spatial intervals, and each trigger signal is sent to the image acquisition module, temperature acquisition module, and near-infrared spectrometer to control the image acquisition module to perform triggered acquisition to obtain image data, control the temperature acquisition module to perform triggered acquisition to obtain temperature distribution data, and control the near-infrared spectrometer to perform triggered acquisition to obtain water content distribution data. All data obtained by triggered acquisition are stamped with a unified timestamp based on the master clock.

[0015] Preferably, the step of identifying the degree of segregation based on the image data and the temperature distribution data using a pre-constructed cross-modal spatiotemporal cross-attention model to obtain a segregation distribution heatmap includes:

[0016] The cross-modal spatiotemporal cross-attention model also performs the following operations:

[0017] Using the spatiotemporal visual features as the query vector, and using the spatiotemporal hot features as both the key vector and the value vector, attention weights are calculated to obtain a first attention weight. Based on the first attention weight, the spatiotemporal visual features are enhanced to obtain enhanced visual features.

[0018] Using the spatiotemporal thermal features as the query, and using the spatiotemporal visual features as both the key vector and the value vector, attention weights are calculated to obtain a second attention weight. Based on the second attention weight, the spatiotemporal thermal features are enhanced to obtain enhanced thermal features.

[0019] The enhanced visual features and the enhanced thermal features are fused to obtain multimodal fusion features. A hierarchical spatiotemporal transform encoder is used to extract contextual spatiotemporal features from the multimodal fusion features to obtain long-term dependent spatiotemporal features. Based on the long-term dependent spatiotemporal features, a dissociated distribution heatmap is obtained.

[0020] Preferably, the step of using a two-dimensional convolutional neural network and an optical flow network to perform dual-channel feature extraction on the image data includes:

[0021] The image data is repaired by using a pre-constructed image inpainting generative adversarial network to obtain repaired image data;

[0022] A two-dimensional convolutional neural network is used to extract the spatial features of each frame of the repaired image data, and the spatial features are stacked along the time dimension to extract temporal features, thus obtaining spatiotemporal features.

[0023] An optical flow network is used to extract dense motion fields between consecutive frames of the repaired image data. The dense motion fields are stacked and then feature extraction is performed to obtain motion evolution features.

[0024] The spatiotemporal features and the motion evolution features are fused in the channel dimension to obtain spatiotemporal visual features.

[0025] Preferably, the use of an attention-based convolutional long short-term memory network to sequentially extract and fuse features from the temperature distribution data includes:

[0026] A lightweight two-dimensional convolutional neural network is used to extract the temperature change features of the temperature distribution data, and a long short-term memory network is used to extract the temperature distribution features of the temperature distribution data.

[0027] Based on the temperature change characteristics, a temporal attention weight is generated; based on the temperature distribution characteristics, a spatial attention weight is generated; and based on the spatial attention weight and the temporal attention weight, the temperature change characteristics and the temperature distribution characteristics are fused to obtain spatiotemporal thermal characteristics.

[0028] Preferably, the step of performing temperature compensation on the compaction degree distribution data based on the temperature distribution data to obtain compensated compaction degree distribution data includes:

[0029] The pre-built compaction compensation model was calibrated using a small batch of training datasets collected on the road surface to be constructed, resulting in a calibrated compaction compensation model.

[0030] Based on the temperature distribution data and the compaction distribution data, the calibrated compaction compensation model is used to obtain the compensated compaction distribution data.

[0031] Preferably, the step of inputting the segregation distribution heat map, the temperature distribution data, the moisture content distribution data, and the compensated compaction degree distribution data as real-time feedback data into a pre-constructed dynamic coupling model to obtain optimal rolling operation parameters includes:

[0032] Based on the moisture content distribution data and the preset optimal moisture content, a first set of correction coefficients for the roller operating parameters is generated. Based on the temperature distribution data and the preset optimal operating temperature, a second set of correction coefficients for the roller operating parameters is generated. Based on the segregation distribution heat map and the preset segregation degree threshold, the coupling relationship between segregation degree and energy is simulated to generate a third set of correction coefficients for the roller operating parameters. Based on the compensated compaction degree distribution data and the preset target compaction degree value, the coupling relationship between compaction degree and energy is simulated to generate a fourth set of correction coefficients for the roller operating parameters.

[0033] The roller operating parameters are adjusted according to the first combination of correction coefficients, the second combination of correction coefficients, the third combination of correction coefficients, and the fourth combination of correction coefficients to obtain the optimal roller operating parameters.

[0034] Preferably, the step of generating the first set of correction coefficients for the roller operating parameters based on the moisture content distribution data and the preset optimal moisture content includes:

[0035] Based on the moisture content distribution data, the current moisture content of each pre-divided grid cell is obtained. Based on the pre-calibrated humidity correction coefficient gradient model, the gradient corresponding to the humidity difference between the current moisture content and the preset optimal moisture content is determined. The humidity correction coefficient gradient model is set to map the humidity difference to a combination of humidity correction coefficients with different gradients.

[0036] Based on the attribution gradient, the first set of correction coefficients for the roller operation parameters is obtained.

[0037] Preferably, the step of generating the third set of correction coefficients for the roller operating parameters based on the segregation distribution heat map and a preset segregation degree threshold, simulating the coupling relationship between segregation degree and energy, includes:

[0038] Based on the segregation distribution heatmap, the current segregation index of each pre-divided grid cell is obtained, and based on the current segregation index and the preset segregation threshold, the segregation improvement requirement of each grid cell is obtained.

[0039] Based on the required segregation degree improvement, the first total required energy is obtained using a pre-constructed energy segregation degree improvement mapping model;

[0040] Based on the remaining number of compaction passes and the preset energy allocation principle, the first total energy demand is allocated to obtain the first compaction energy demand for this compaction, and the first compaction energy demand is mapped to the third correction coefficient combination of the roller operation parameters.

[0041] Secondly, the present invention also provides a system for adjusting the parameters of an intelligent road roller based on a neural network, used to implement the above-mentioned method for adjusting the parameters of an intelligent road roller based on a neural network, the system comprising:

[0042] Multi-source data acquisition module, segregation degree identification module, compaction degree compensation module, and dynamic coupling module;

[0043] The multi-source data acquisition module is used to acquire multi-source data of the road surface to be constructed during the operation of the road roller. The multi-source data includes image data, temperature distribution data, vibration distribution data and moisture content distribution data.

[0044] The segregation degree recognition module is used to identify the segregation degree based on the image data and the temperature distribution data through a pre-constructed cross-modal spatiotemporal cross-attention model to obtain a segregation distribution heatmap. The cross-modal spatiotemporal cross-attention model is set to perform cross-attention fusion of spatiotemporal visual features and spatiotemporal thermal features to generate the segregation distribution heatmap. The spatiotemporal visual features are obtained by extracting dual-channel features from the image data using a two-dimensional convolutional neural network and an optical flow network. The spatiotemporal thermal features are obtained by extracting and fusing features from the temperature distribution data sequentially using a convolutional long short-term memory network based on an attention mechanism.

[0045] The compaction compensation module is used to obtain compaction distribution data based on the vibration distribution data, and to perform temperature compensation on the compaction distribution data based on the temperature distribution data to obtain compensated compaction distribution data.

[0046] The dynamic coupling module is used to input the segregation distribution heat map, the temperature distribution data, the moisture content distribution data, and the compensated compaction degree distribution data as real-time feedback data into the pre-constructed dynamic coupling model to obtain the optimal rolling operation parameters.

[0047] This application provides a method and system for adjusting the parameters of an intelligent road roller based on a neural network. Compared with the prior art, the beneficial effects of the embodiments of this application are as follows:

[0048] This application discloses a neural network-based intelligent road roller parameter adjustment method, which constructs a multimodal pavement state perception system. Through multi-source information complementarity and collaborative verification, it achieves a comprehensive and in-depth understanding of the microscopic mechanism and evolution trend of the compaction process, possessing strong causal reasoning ability and anti-interference capability. It achieves deep, bidirectional, and adaptive fusion of visual modal data and temperature modal data in the spatiotemporal dimension through a cross-modal spatiotemporal cross-attention model. Utilizing the cross-attention mechanism, visual features and thermal features guide and reinforce each other in the spatiotemporal dimension. A hierarchical spatiotemporal transformation encoder determines whether an anomaly is isolated noise or part of a continuous segregation zone, thus assigning... The cross-modal spatiotemporal cross-attention model, similar to the global judgment capability of experts, greatly improves accuracy. Temperature distribution data compensates for the compaction distribution data calculated from vibration distribution data to eliminate the interference of temperature on pavement compaction, providing a real and consistent decision-making basis for real-time assessment and closed-loop control of pavement compaction quality. Through multi-physics coupling analysis of the influence of moisture content, temperature, segregation degree, and real-time pavement compaction on roller operating parameters, real-time, adaptive, and refined control of roller operating parameters is achieved. The entire compaction process is digitally recorded, and the compaction quality at any point is traceable, transforming post-event quality control into in-process assurance. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the steps of a neural network-based intelligent road roller parameter adjustment method provided in a preferred embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of a neural network-based intelligent road roller parameter adjustment system provided in a preferred embodiment of the present invention;

[0051] Figure label:

[0052] 1-Multi-source data acquisition module, 2-Segregation degree identification module, 3-Compaction degree compensation module, 4-Dynamic coupling module. Detailed Implementation

[0053] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. The embodiments are provided for illustrative purposes only and should not be construed as limiting the scope of the invention. The accompanying drawings are for reference and illustration only and do not constitute a limitation on the scope of protection of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of this invention.

[0054] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0055] Please see Figure 1 The diagram illustrates the steps of a neural network-based intelligent road roller parameter adjustment method. In an embodiment of the present invention, a neural network-based intelligent road roller parameter adjustment method is provided, the method comprising:

[0056] S1. Acquire multi-source data of the road surface to be constructed during roller operation. This multi-source data includes image data, temperature distribution data, vibration distribution data, and moisture content distribution data. In road construction, pavers and rollers typically work together. The roller's speed and vibration parameters directly determine the road surface's smoothness, density, and final quality. The coordination between the paver and roller is an orderly and continuous process, following the basic principles of "close following, slow compaction, high frequency, and low amplitude." The roller generally follows the paver closely, performing initial compaction, secondary compaction, and final compaction. Initial compaction is performed 3-5 meters behind the paver, usually when the mixture temperature is above 150℃, to initially stabilize the mixture, prevent displacement, and establish initial density. Secondary compaction follows immediately after initial compaction, usually when the mixture temperature is between 120℃ and 140℃, achieving the preset density requirement, typically above 96% of the maximum density. Final compaction is performed after secondary compaction, when the mixture temperature drops to 90℃-110℃, to eliminate all wheel tracks and form a final smooth and even surface. For different road surfaces, the number of compaction passes for initial, secondary, and final compaction is fixed, and controlling road compaction quality by changing the number of passes is costly. However, dynamically adjusting the vibration parameters and travel speed of the roller can effectively improve compaction, reduce road segregation, and lower the road defect rate. In a preferred embodiment of this application, multiple data acquisition devices are deployed, breaking through the traditional single vibration feedback to construct a comprehensive real-time perception of materials, machinery, and the environment. In a preferred embodiment of this application, the data acquisition devices include an image acquisition device, a temperature acquisition device, a near-infrared spectrometer, a positioning module, and a piezoelectric accelerometer. The near-infrared spectrometer is used to collect moisture content distribution data, and the piezoelectric accelerometer is used to collect vibration distribution data. To ensure that the collected multi-source data share the same spatiotemporal reference system, during the compaction operation, the time synchronization module receives a 1PPS pulse signal and a time message containing the year, month, day, hour, minute, and second from the global navigation satellite system's timing module. This signal serves as the absolute time source, synchronizing and calibrating the internal master clock to ensure all data acquisition is conducted within a unified, high-precision time coordinate system. During the operation, based on this master clock, the positioning module (such as an RTK-GNSS receiver) continuously samples at a preset frequency, generating position time-series data with a unified timestamp, including latitude, longitude, elevation, velocity, and heading. Piezoelectric accelerometers synchronously and continuously sample at the same or a specific high frequency, generating vibration time-series data with a unified timestamp, such as triaxial or uniaxial acceleration. The system processes the position time-series data provided by the positioning module in real time. When the cumulative displacement reaches different preset spatial intervals (e.g., 0.2 meters or 0.5 meters per step), it automatically generates corresponding trigger signals. For different data acquisition devices, the same spatial interval or different spatial intervals can be used. It is only necessary to generate corresponding trigger signals and send each trigger signal to the image acquisition module, temperature acquisition module and near-infrared spectrometer.Upon receiving a trigger signal, each acquisition module immediately performs an acquisition action and automatically adds the current timestamp from the master clock to the acquisition results. Specifically, the image acquisition module acquires road surface images in front of and behind the current compaction wheel, generating image data; the temperature acquisition module acquires road surface temperature at positions before and after compaction, generating temperature distribution data before and after compaction; the near-infrared spectrometer acquires the reflectance spectrum of the current road surface points, and generates moisture content data at each point before compaction through inversion using a built-in model, and constructs moisture content distribution data through a series of data points.

[0057] This application uses GNSS (Global Navigation Satellite System) as the absolute time reference to construct a unified spatiotemporal coordinate system. It adopts a "position-triggered, synchronous acquisition" data acquisition mode, which fundamentally solves the persistent problem of spatiotemporal mismatch of multi-source data in existing intelligent compaction systems. It achieves pixel-level precise alignment of all sensing data in the time and space dimensions, providing a reliable data foundation for subsequent accurate state diagnosis and intelligent decision control.

[0058] In a preferred embodiment of this application, image data, temperature distribution data, vibration distribution data, and moisture content distribution data are divided into grids, mapping the continuous physical working surface into discrete data grids with spatiotemporal attributes to facilitate algorithm processing and analysis. Specifically, at the starting point of the construction section, a global coordinate system is established using GNSS. The positioning modules on the roller and paver provide real-time centimeter-level position information. Using the working surface as the XY plane, regular square or rectangular grids are defined as unit grids. The size of the grid is usually the smallest unit for compaction quality detection. For image data and temperature distribution data, the ground coordinates corresponding to each pixel are calculated using sensor calibration parameters and positioning data, and assigned to the corresponding grid. The average or weighted values ​​of multiple pixels within the same grid are used to obtain the representative value of that grid. For vibration distribution data, the vibration data carries the position coordinates of the roller at the time of acquisition. When a trajectory point passes through a grid, the vibration data of that point represents the vibration distribution data of that grid. For multiple trajectory point data within the same grid, they can be fused according to time or space. For the water content distribution data, Kriging interpolation is used to convert the discrete point data into a continuous field covering the entire working surface, and then the values ​​are distributed to each grid.

[0059] The image acquisition device includes an industrial camera and a multispectral structured light projection module. The device comprises two sets. One set is mounted on a fixed rigid gantry spanning the entire width of the lane, with the gantry positioned at both ends of the road surface to be compacted. This gantry acquires high-precision image data of key nodes along the entire road surface to provide baseline data. The other set of image acquisition devices is mounted on a drone. The drone flies along the construction route, collecting continuous image data during the construction process. However, the continuous image data acquired by the drone is subject to vibration and positioning errors. Therefore, the ultra-high-precision data acquired by the rigid gantry is used to calibrate and correct the continuous image data acquired by the drone, ultimately obtaining continuous image data of the road surface to be constructed with high-precision baselines.

[0060] Calibration and correction of continuous data are crucial for improving the accuracy of image data. In a preferred embodiment of this application, high-precision image data acquired by a rigid gantry is converted into a high-precision ground truth point cloud, with the coordinate system being the geographic coordinate system. Continuous image data acquired by a UAV is converted into a point cloud to be corrected, with each point carrying an initial geographic coordinate system and attitude angle. Based on the GNSS timestamp and location, all data frames within a certain time frame before and after the UAV flies over the entire fixed rigid gantry are identified. The point cloud to be corrected corresponding to these data frames is projected onto the high-precision ground truth point cloud. Stable 3D feature points are extracted within the overlapping area of ​​the high-precision ground truth point cloud and the point cloud to be corrected. Through nearest neighbor search and consistency checks, mismatched feature points are removed, resulting in a series of high-confidence 3D matching point pairs. A 3D affine transformation model is selected, and an overdetermined system of equations is constructed based on the 3D matching point pairs. The optimal parameters of the 3D affine transformation model are solved using the least squares method. Based on the optimal parameters, the 3D affine transformation model is used to correct the entire point cloud to be corrected, ultimately obtaining the image data of the road surface to be constructed during the entire roller operation process.

[0061] In a preferred embodiment of this application, the image acquisition device mounted on a rigid gantry can accommodate multiple cameras spaced along the crossbeam, with adjacent cameras overlapping their acquisition areas by 10%-15% to avoid missed shots. The camera lenses are fitted with anti-fouling protective covers, the surface of which is coated with a hydrophobic and oil-resistant coating to reduce the adhesion of asphalt mixture splashes. A small compressed air-driven blowing device is also provided, automatically blowing the lenses every 5 minutes to ensure cleanliness.

[0062] Temperature distribution data acquisition during roller operation is crucial for controlling the compaction window, determining the compactability of hot-mixed materials, preventing temperature segregation, and setting roller vibration parameters. The selection and arrangement of temperature acquisition devices must meet the requirements of high dynamics, resistance to compaction, and real-time reliability. In the preferred embodiment of this application, an infrared temperature array sensor is selected. The infrared temperature array sensor is a linear or small-area array of infrared detectors, which is cost-effective and has good vibration resistance. A single infrared temperature array sensor cannot provide comprehensive coverage. This application deploys an infrared temperature array sensor above the front of the roller wheel, behind the roller wheel, and on the side of the roller body to capture the complete temperature dynamic process. Specifically, the infrared temperature array sensor above the front of the roller wheel is installed on a crossbeam about 0.5m-1.0m in front of the roller wheel, collecting the lateral temperature distribution of the road surface before compaction, and obtaining pre-compaction temperature distribution data. This pre-compaction temperature distribution data is used for roller parameter adjustment. The infrared temperature measurement array sensor behind the steel wheel is installed approximately 0.2m-0.5m behind the wheel in a transverse array. It collects the transverse temperature distribution of the road surface after compaction, obtaining post-compaction temperature distribution data, which is used to verify the control effect. The side infrared temperature measurement array sensor is specifically installed at a higher position on the side of the roller body, viewed from the side and covering an area of ​​several meters from the steel wheel to the rear, forming a longitudinal temperature band. Through the infrared temperature measurement array sensor group deployed above, the temperature distribution data of the road surface to be constructed can be obtained throughout the entire operation of the roller.

[0063] The acquisition of vibration distribution data during road roller operation is crucial for compaction control and dynamic adjustment of operating parameters. In a preferred embodiment of this application, a triaxial integrated electronic piezoelectric accelerometer is selected as the piezoelectric accelerometer. Its range is adapted to the vibration parameters of the road roller, enabling it to capture the complete acceleration waveform of the steel wheel vibration, including the positive peak, negative peak, and vibration period. The sampling frequency is set to 1000Hz to ensure that no vibration details are missed. The triaxial integrated electronic piezoelectric accelerometer is installed on the bearing housing or vibration shaft housing of the vibrating steel wheel, which is the closest position to the vibration source, directly measuring the signal of acceleration in the X / Y / Z directions of the vibrating steel wheel as a function of time.

[0064] The acquisition of moisture content distribution data during road roller operation is crucial for road quality control and dynamic adjustment of rolling operation parameters. In the preferred embodiment of this application, a near-infrared spectrometer is used to collect moisture content distribution data. Near-infrared spectroscopy is an analytical method based on the selective absorption of near-infrared light (wavelength typically 780-2500 nm) by molecules. The OH bonds in water molecules have very significant characteristic absorption peaks in the near-infrared spectral region (such as around 1450 nm and 1940 nm), and the absorption intensity is highly correlated with the moisture content. The near-infrared spectrometer emits one or more beams of near-infrared light of specific wavelengths towards the road surface. Some of the light is absorbed by water molecules, while some is scattered or reflected back to the near-infrared spectrometer. The near-infrared spectrometer analyzes the reflection and absorption spectra, and by analyzing the intensity of the OH bond characteristic peaks or the shape changes of the entire spectrum, the moisture content of the road material can be calculated in real time using a pre-established calibration model. The near-infrared spectrometer is installed on the front crossbeam of the road roller, approximately 30 cm-100 cm above the ground, which avoids contact with uncompacted materials, reducing interference with the road surface and sensor wear. A single near-infrared spectrometer provides the road surface moisture content at a point directly below. To generate moisture content distribution data, multiple near-infrared spectrometers need to be installed side-by-side on the crossbeam, covering the entire wheel width at once, and simultaneously collecting road surface moisture content at different transverse locations. Combined with a GNSS positioning system, the precise coordinates of each measurement point are recorded. The time-series measurement data is then bound to the location information to generate moisture content point cloud data with geographic coordinates. Continuous three-dimensional moisture content distribution data is generated using the Kriging interpolation algorithm.

[0065] The data acquisition method of this application yields a set of multi-source data that is strictly synchronized in time and aligned in spatial location. Each set of data collected at a specific spatial point can be linked to the precise location of the road roller and the road surface condition at that moment through its timestamp, thus constructing a multimodal road surface condition perception system. Through the complementarity and collaborative verification of multi-source information, a comprehensive and in-depth understanding of the micro-mechanism and evolution trend of the compaction process is achieved, which has strong causal reasoning ability and anti-interference ability.

[0066] S2. Based on the image data and the temperature distribution data, the degree of segregation is identified through a pre-constructed cross-modal spatiotemporal cross-attention model to obtain a segregation distribution heatmap. The cross-modal spatiotemporal cross-attention model is set to perform cross-attention fusion of spatiotemporal visual features and spatiotemporal thermal features to generate the segregation distribution heatmap. The spatiotemporal visual features are obtained by extracting dual-channel features from the image data using a two-dimensional convolutional neural network and an optical flow network. The spatiotemporal thermal features are obtained by extracting and fusing features from the temperature distribution data sequentially using a convolutional long short-term memory network based on an attention mechanism. In a preferred embodiment of this application, the collected image data may have quality problems such as local missing parts, blurring, and noise interference due to complex working conditions such as dust, sudden changes in light, mechanical vibration, water droplet splashing, or instantaneous occlusion. These defective images will directly affect the accuracy of subsequent vision-based analysis of the degree of road segregation. To address the aforementioned issues, in a preferred embodiment of this application, a pre-constructed image inpainting generative adversarial network (GAN) is used. This network can automatically identify defective regions in an image without human intervention and generate a visually coherent and content-appropriate high-quality inpainted image based on the overall contextual semantics of the image and information from temporally adjacent frames. The final output is temporal data of the inpainted image. The image inpainting GAN comprises a generator network and a discriminator network. The generator network employs an encoder-decoder structure with an attention mechanism. The encoder performs multi-level downsampling on the input defective image, extracting multi-level feature representations. At the feature level, the attention mechanism references features from preceding and following frames in temporal sequence, as well as spatial context information, to predict the features of the defective region for inpainting the feature map. The decoder upsamples the inpainted feature map to reconstruct a complete RGB image, ensuring that the inpainted area maintains consistency with the overall texture and lighting. The discriminator network employs a PatchGAN classifier (Local Receptive Field Discriminator) structure. The PatchGAN classifier does not distinguish between real and fake images as a whole, but rather between real and fake images of each local "block," thus forcing the generator network to produce realistic outputs even at finer-grained local textures. The image inpainting generative adversarial network is trained using a large publicly available image dataset and a self-built dataset of damaged images from engineering scenes, enabling the network to learn general image inpainting capabilities. Image data collected on-site is input frame-by-frame into the trained image inpainting generative adversarial network, which automatically identifies potential defects or low-quality regions in each frame. Combining the contextual information of that frame with the temporal information of preceding and following frames, pixel-level content completion and enhancement are performed, ultimately outputting visually consistent restored image data. The restored image data maintains the same timestamp sequence as the original image data, ensuring that the spatiotemporal synchronization with data such as location and vibration is not disrupted.

[0067] During continuous compaction, the road surface condition is the spatiotemporal evolution result of the deep coupling between visual appearance and physical field. Single-modal data cannot fully reflect this complex process. To address this issue, this application constructs a cross-modal spatiotemporal cross-attention model. The model inputs repaired image data and temperature distribution data to obtain a segregated distribution heatmap. The cross-modal spatiotemporal cross-attention model combines cross-modal fusion, cross-attention, and hierarchical spatiotemporal coding techniques. It employs parallel extraction of spatiotemporal visual features from the repaired image data and spatiotemporal thermal features from the temperature distribution data before compaction. For spatiotemporal visual feature extraction, a dual-channel parallel network is used to process the restored image data. Channel 1 uses a two-dimensional convolutional neural network to extract spatial features of each frame of the restored image data, such as texture and edges. The spatial features are then stacked along the time dimension to extract temporal features, thus obtaining spatiotemporal features. Channel 2 uses an optical flow network to extract dense motion fields between consecutive frames of the restored image data, such as deformation trends. The dense motion fields are then stacked and feature extraction is performed to obtain motion evolution features. The spatiotemporal features and motion evolution features are then fused along the channel dimension to obtain spatiotemporal visual features. For spatiotemporal thermal feature extraction, a convolutional long short-term memory network combining temporal and spatial attention is employed. This extracts high-level features that simultaneously characterize the dynamic evolution of temperature over time and its non-uniform distribution pattern in the spatial plane. Specifically, a lightweight two-dimensional convolutional neural network performs spatial convolution operations along each slice of the time dimension to extract temperature change features, focusing on the fluctuations, gradients, and trends of temperature values ​​at each spatial location over time, thus obtaining the temperature change features at each spatial location. The long short-term memory network extracts the temperature distribution features at each moment. LSN, with its gating mechanism, can remember and forget historical spatial patterns, characterizing the distribution of heat across the entire plane at a given moment. Furthermore, a fully connected layer is used to assign a specific feature to each time step in the time series based on the temperature change features. Each layer generates a temporal attention weight, which automatically identifies and amplifies critical moments of drastic temperature changes while suppressing the influence of stable or irrelevant moments. Another fully connected layer generates a spatial attention weight for the temperature distribution features at each moment, focusing on temperature anomaly regions and strengthening the feature responses of key spatial locations. Finally, the temporal attention weights are used to reweight the temperature change features in the temporal dimension, and the spatial attention weights are used to reweight the temperature distribution features in the spatial dimension. The weighted temperature change features and the weighted temperature distribution features are then concatenated, achieving feature alignment and enhanced integration between key temporal events and key spatial regions. Convolution and nonlinear activation processing are applied to the concatenated features to output unified spatiotemporal thermal features rich in significant spatiotemporal information.

[0068] The cross-modal spatiotemporal cross-attention model enhances spatiotemporal visual and thermal features after extraction. Specifically, cross-attention networks are integrated into the backbone networks of both spatiotemporal visual and thermal feature extraction. The cross-attention network integrated into the spatiotemporal visual feature extraction backbone network uses the spatiotemporal visual features as the query vector and the spatiotemporal thermal features as both the key and value vectors. The dot product attention of the query vector and the key vector is calculated to obtain the first attention weight, which reflects the spatiotemporal relevance of the thermal features to the visual features. Subsequently, the first attention weight is multiplied by the spatiotemporal thermal features to obtain the first weighted context vector, which is then added to the spatiotemporal visual features through a residual connection to output the enhanced visual features. This results in the visual features being enhanced by the temperature changes of hotspots, which are indicative of the material state.

[0069] Symmetrically, the main network for spatiotemporal thermal feature extraction integrates a cross-attention network. Using spatiotemporal thermal features as the query vector and spatiotemporal visual features as the key and value vectors, a second attention weight is calculated. This second attention weight reflects the spatiotemporal correlation between visual features and thermal features. Subsequently, the second attention weight is multiplied by the spatiotemporal visual features to obtain a second weighted context vector. This vector is then added to the spatiotemporal thermal features through a residual connection, outputting the enhanced thermal features. This allows the thermal features to be corroborated and enriched by the corresponding visual texture patterns.

[0070] Furthermore, a weighted summation method is used to fuse the enhanced visual features and enhanced thermal features to form a unified multimodal fusion feature. This multimodal fusion feature already contains aligned and enhanced correlation information between the two modalities. The multimodal fusion feature is then input into a hierarchical spatiotemporal transform encoder. Each layer, through its self-attention mechanism, captures long-term dependencies between features within an increasingly larger spatiotemporal receptive field, ultimately outputting long-term dependency spatiotemporal features containing global contextual information. Finally, a lightweight decoder maps the long-term dependency spatiotemporal features to pixel-level spatial resolution, and a sigmoid function is used to generate a heatmap representing the probability of segregation at each spatial location.

[0071] In a preferred embodiment of this application, a cross-modal spatiotemporal cross-attention model is used to achieve deep, bidirectional, and adaptive fusion of visual modal data and temperature modal data in the spatiotemporal dimension. Utilizing the cross-attention mechanism, visual and thermal features guide and reinforce each other in the spatiotemporal dimension. For regions with complex textures but uniform temperature, the system reduces the interference of thermal features on visual features; for regions with visual blurring due to shadows but abnormal temperature, the system strengthens the dominance of thermal features, enabling the system to intelligently adapt to different lighting, weather, and material conditions. The hierarchical spatiotemporal transform encoder can determine whether an anomaly is isolated noise or part of a continuous separation band, which endows the cross-modal spatiotemporal cross-attention model with a global judgment capability similar to that of an expert, greatly improving accuracy.

[0072] S3. Based on the vibration distribution data, compaction distribution data is obtained, and temperature compensation is performed on the compaction distribution data based on the temperature distribution data to obtain compensated compaction distribution data. In the prior art, compaction measurement values ​​are used to quantify the stiffness of road materials. The compaction measurement value essentially transforms the dynamic response of the vibrating wheel or vibrating roller into a quantitative index reflecting the stiffness of the road material. However, stiffness is not directly equivalent to the compaction degree of the road surface. The mechanical properties of materials are highly dependent on temperature. The same asphalt mixture is a viscoplastic body at 150℃, which is easy to deform; it begins to have elasticity at 100℃; and it may become brittle and hard below 80℃. Temperature has a direct impact on the hardness of the material. The compaction measurement value obtained solely based on the vibration distribution data of the steel wheel includes the influence of temperature changes on hardness and cannot accurately reflect the true compaction degree of the road surface. In the preferred embodiment of this application, in order to eliminate the error caused by temperature changes, temperature compensation is performed on the compaction measurement value calculated by vibration data to ensure the accuracy of the compaction measurement value. The temperature distribution data used is also the temperature distribution data before compaction. Based on the vibration distribution data, compaction degree distribution data is obtained. Specifically, the acceleration signals in the X / Y / Z directions collected by the triaxial integrated electronic piezoelectric accelerometer are filtered to remove high-frequency electrical noise and low-frequency machine translation interference, retaining the frequency band related to the excitation frequency. A fast Fourier transform is performed on the filtered signal to obtain a spectrum, which shows the amplitude of each frequency component in the signal segment. Further, the first amplitude corresponding to the set excitation frequency of the road roller and the second amplitude corresponding to the first harmonic frequency are extracted from the spectrum. The compaction degree measurement value is calculated based on the first and second amplitudes, as shown in the following formula:

[0073] ;

[0074] in, This indicates the measured compaction degree. This represents the coefficient calibrated based on the parameters of the road roller. This represents the second amplitude corresponding to the first harmonic frequency. This represents the first amplitude corresponding to the set excitation frequency of the road roller. Compaction degree distribution data is generated based on the compaction degree measurements of each grid cell.

[0075] On the same homogeneous material, the measured compaction degree is negatively correlated with temperature. Compaction degree test values ​​calculated from vibration data at different temperatures on the target road section at the construction site are collected, along with actual compaction degree values ​​measured using a nuclear density meter. The actual compaction degree value measured by the nuclear density meter is equivalent to the true compaction degree of the road surface. In a preferred embodiment of this application, a training dataset is constructed based on temperature, compaction degree test values, and actual compaction degree values. This training dataset is then used to train a selected feedforward neural network to obtain a compaction degree compensation model. Small batches of training datasets are collected from the road surface to be constructed at different construction sites. These small batches of training datasets are input into the compaction degree compensation model. The compaction degree compensation model automatically calculates and updates its internal compensation coefficients or parameters based on this training data to adapt to the specific material and temperature conditions of the current road surface, thereby generating a calibrated compaction degree compensation model. In subsequent continuous compaction operations, the temperature values ​​corresponding to each grid cell in the temperature distribution data and the compaction measurement values ​​corresponding to each grid cell in the compaction measurement distribution data are input into the calibrated compaction compensation model. The model will output the corrected and compensated compaction value, which is the compaction evaluation result after eliminating temperature interference.

[0076] The compensated compaction value is bound to the precise coordinates (X, Y) provided by the GNSS positioning system of the road roller at this moment, and a record with spatial attributes is output. Based on the compensated compaction value of each grid cell, the compensated compaction distribution data is generated.

[0077] In a preferred embodiment of this application, the compaction distribution data calculated from the vibration distribution data is compensated by the temperature distribution data to eliminate the interference of temperature on the road compaction, thereby providing a real and consistent decision-making basis for real-time assessment and closed-loop control of road compaction quality.

[0078] S4. The segregation distribution heatmap, temperature distribution data, moisture content distribution data, and compensated compaction degree distribution data are used as real-time feedback data and input into a pre-constructed dynamic coupling model to obtain optimal rolling operation parameters. In a preferred embodiment of this application, the segregation distribution heatmap, temperature distribution data, moisture content distribution data, and compensated compaction degree distribution data are spatiotemporally aligned and multi-source data integrated to generate a multi-attribute fusion data layer. The multi-attribute fusion data layer is a data cube or a multi-layer feature stack, where each spatial location corresponds to a set of structured attribute vectors. The multi-attribute fusion data layer is essentially a structured, flexibly decomposable data service, indexed by location and time, and with segregation index, temperature, moisture content, and compensated compaction degree values ​​as attribute fields. Further, the multi-attribute fusion data layer is input into a pre-constructed dynamic coupling model, and the dynamic coupling model performs the following processing:

[0079] First, based on the moisture content distribution data in the multi-attribute fusion data layer and the preset optimal moisture content, a first set of correction coefficients for the roller's operating parameters is generated. During roller rolling, different processing methods are required for different moisture contents. When the moisture content is close to the preset optimal moisture content, it can be remedied by adjusting the roller's vibration parameters. For example, if the moisture content is greater than the optimal moisture content, high-frequency low-amplitude compaction is used to avoid excessive disturbance; if the moisture content is less than the optimal moisture content, low-frequency high-amplitude compaction is used to enhance the interlocking effect between particles. However, in traditional operations, these processing methods rely solely on the experience of the operators, resulting in a high rate of error. In the preferred embodiment of this application, a first set of correction coefficients is generated based on the relationship between the moisture content distribution data and the preset optimal moisture content. This first set of correction coefficients includes a first vibration frequency correction coefficient, a first amplitude correction coefficient, and a first velocity correction coefficient, enabling accurate response strategies based on different moisture contents.

[0080] For the first combination of correction coefficients, this application employs a gradient variation method, dividing the moisture difference between the current moisture content and the optimum moisture content of the mixture into different gradients. A corresponding combination of moisture correction coefficients is set for each gradient to construct a moisture correction coefficient gradient model. The moisture correction coefficient combination for each gradient is calibrated using experimental data and stored in a database. For example, if the current moisture content of the mixture in a certain grid is greater than the optimum moisture content, the moisture correction coefficient combination determined based on the moisture difference is... ,in, This represents the correction factor for the first vibration frequency. This represents the first magnitude correction factor. This represents the first travel speed correction factor.

[0081] Furthermore, based on temperature distribution data and a preset optimal operating temperature, a second set of correction coefficients for the roller's operating parameters is generated. Temperature is a key factor affecting compaction. At high temperatures, the mixture has good fluidity, is easy to compact, and can quickly achieve the required density. However, excessively high temperatures may cause the mixture to age, the binder to oxidize, reducing pavement durability, and the roller to easily adhere to asphalt, affecting smoothness. At excessively low temperatures, the mixture hardens, making it difficult to compact, easily leading to excessive pavement porosity, water seepage, and a shortened construction window. This necessitates increasing the compaction speed and the number of compaction passes, but it may still fail to achieve the designed density and may even cause cracks. However, in traditional operations, relying solely on the experience of the operators for related processing cannot achieve precise control, error correction is delayed, and construction quality consistency is poor. In the preferred embodiment of this application, a second set of correction coefficients is generated based on the relationship between temperature distribution data and a preset optimal operating temperature. The temperature distribution data used is the temperature distribution data before compaction. The second set of correction coefficients includes a second vibration frequency correction coefficient, a second amplitude correction coefficient, and a second travel speed correction coefficient. Similarly, a tiered variation method is used to divide the temperature difference between the current temperature of the mixture and the optimal operating temperature into different gradients. For each gradient, a corresponding combination of temperature correction coefficients is set to construct a temperature correction coefficient gradient model. Each gradient's temperature correction coefficient combination is also calibrated using experimental data and stored in the database. If the current temperature of the mixture in a certain grid is greater than the optimal operating temperature, the temperature difference is calculated, and the temperature correction coefficient combination determined based on this temperature difference is... ,in, This represents the correction factor for the second vibration frequency. This represents the second magnitude correction factor. This represents the second travel speed correction factor.

[0082] Furthermore, based on the segregation distribution heatmap, the coupling relationship between segregation degree and energy is simulated to generate a third set of correction coefficients. The aggravation and mitigation of segregation degree are positively correlated with the effective work done by the vibration energy applied by the roller in the segregation area. The same vibration energy input has different effects on areas with different segregation degrees and aggregate types. In a preferred embodiment of this application, the current segregation degree index of each grid cell can be directly identified based on the segregation distribution heatmap, and the segregation degree improvement requirement of each grid cell is obtained based on the segregation degree difference between the current segregation degree index and the preset segregation degree threshold. Based on the segregation degree improvement requirement, a pre-constructed energy segregation degree improvement mapping model is used to obtain the first total energy requirement. The energy segregation degree improvement mapping model is constructed based on physical mechanics and experimental data. This application uses a power-law relationship to construct the physical model, and the formula is expressed as follows:

[0083] ;

[0084] in, This indicates the demand for improved separation. This represents the first total energy demand per unit grid. and The material correlation coefficient is determined through indoor compaction tests on segregated specimens. Typically less than 1, indicating that the improvement in efficiency decreases with increasing energy.

[0085] Based on the remaining compaction passes and the preset energy allocation principle, the first total energy demand is allocated to obtain the first compaction energy demand for this compaction. The energy allocation principle adopts the principle of heavy allocation at the beginning and light allocation at the end, with a proportionally decreasing sequence. More energy is allocated in the early passes to quickly improve the structure, and the energy is used for finishing in the later passes. The first compaction energy demand is mapped to a third set of correction coefficients for the roller operating parameters. The third set of correction coefficients includes a third vibration frequency correction coefficient, a third amplitude correction coefficient, and a third travel speed correction coefficient. In a preferred embodiment of this application, a first optimization objective function is constructed based on the first compaction energy demand. The first optimization objective function is solved to obtain the third set of correction coefficients for the roller operating parameters. The first optimization objective function is expressed as follows:

[0086] ;

[0087] in, This represents the third magnitude correction factor. This represents the correction factor for the third vibration frequency. This represents the third travel speed correction factor. This indicates the set preselected amplitude. This indicates the set pre-selected vibration frequency. This indicates the preset travel speed. This indicates the energy required for the first compaction. The pre-selected amplitude, pre-selected vibration frequency, and pre-selected travel speed are the vibration values, vibration frequencies, and travel speeds of the vibratory drum that are pre-selected and set by the construction technicians on the control panel of the roller before the compaction operation begins, based on the type of material being compacted, the layer thickness, and the compaction target.

[0088] Using the upper and lower limits of amplitude, vibration frequency, and travel speed as constraints, the first objective function is optimized to obtain the third vibration frequency correction coefficient, the third amplitude correction coefficient, and the third travel speed correction coefficient.

[0089] Furthermore, based on the compensated compaction distribution data and the preset target compaction value, the coupling relationship between compaction and energy is simulated to generate a fourth set of correction coefficients for the roller's operating parameters; the final compaction is positively correlated with the effective work done by the vibration energy applied by the roller in each region. The compensated compaction of each grid cell can be directly identified based on the compensated compaction distribution data, and the compaction requirement for each grid cell is obtained based on the difference between the compensated compaction and the preset target compaction value. Based on the compaction requirement, a pre-constructed energy compaction requirement mapping model is used to obtain the second total energy requirement. This energy compaction requirement mapping model is also based on physical mechanics and experimental data, and the formula is as follows:

[0090] ;

[0091] in, This indicates the required compaction level. This represents the second total energy demand per unit grid. and The material correlation coefficient is determined through indoor specimen compaction tests. Typically less than 1, indicating that compaction efficiency decreases with increasing energy.

[0092] Based on the remaining compaction passes and the preset energy allocation principle, the second total energy demand is allocated to obtain the second compaction energy demand for this compaction. The energy allocation principle also follows the principle of heavier energy allocation at the beginning and lighter energy allocation at the end, with a proportionally decreasing sequence. More energy is allocated in the early passes to quickly improve the structure, and the remaining energy is used for finishing later. The second compaction energy demand is mapped to a fourth set of correction coefficients for the roller's operating parameters. This fourth set of correction coefficients includes a fourth vibration frequency correction coefficient, a fourth amplitude correction coefficient, and a fourth travel speed correction coefficient. In a preferred embodiment of this application, a second optimization objective function is constructed based on the second compaction energy demand. Solving the second optimization objective function yields the fourth set of correction coefficients for the roller's operating parameters. The second optimization objective function is expressed as follows:

[0093] ;

[0094] in, This represents the fourth magnitude correction factor. This represents the correction factor for the fourth vibration frequency. This represents the fourth travel speed correction factor. This indicates the energy required for the second crushing operation.

[0095] Using the upper and lower limits of amplitude, vibration frequency, and travel speed as constraints, the second objective function is optimized to obtain the fourth vibration frequency correction coefficient, the fourth amplitude correction coefficient, and the fourth travel speed correction coefficient.

[0096] Furthermore, the operating parameters of the road roller are adjusted according to the first, second, third, and fourth correction coefficient combinations to obtain the optimal rolling operating parameters. In a preferred embodiment of this application, the amplitude correction coefficients in the first, second, third, and fourth correction coefficient combinations are weighted and summed to obtain the comprehensive amplitude correction coefficient, as expressed by the following formula:

[0097] ;

[0098] in, This represents the overall magnitude correction factor. This indicates the first magnitude correction weight corresponding to the first magnitude correction coefficient. This indicates the second magnitude correction weight corresponding to the second magnitude correction coefficient. This represents the third magnitude correction weight corresponding to the third magnitude correction coefficient. This represents the fourth magnitude correction weight corresponding to the fourth magnitude correction coefficient, for , , and The value of is determined by indoor specimen compaction tests based on the influence of moisture content, temperature, segregation degree repair, and compaction degree requirements on the amplitude.

[0099] The frequency correction coefficients in the first, second, third, and fourth correction coefficient combinations are weighted and summed to obtain the comprehensive frequency correction coefficient, as shown in the following formula:

[0100] ;

[0101] in, This represents the overall frequency correction factor. This represents the first frequency correction weight corresponding to the first frequency correction coefficient. This represents the second frequency correction weight corresponding to the second frequency correction coefficient. This represents the third frequency correction weight corresponding to the third frequency correction coefficient. This represents the fourth frequency correction weight corresponding to the fourth frequency correction coefficient, for , , and The value of is determined by indoor specimen compaction tests based on the influence of moisture content, temperature, segregation degree repair, and compaction degree requirements on the frequency.

[0102] The speed correction coefficients in the first, second, third, and fourth correction coefficient combinations are weighted and summed to obtain the comprehensive speed correction coefficient, as shown in the following formula:

[0103] ;

[0104] in, This represents the overall travel speed correction factor. This indicates the first travel speed correction weight corresponding to the first travel speed correction factor. This indicates the second travel speed correction weight corresponding to the second travel speed correction factor. This indicates the third travel speed correction weight corresponding to the third travel speed correction factor. This represents the fourth travel speed correction weight corresponding to the fourth travel speed correction factor. , , and The value of is determined by indoor specimen compaction tests based on the influence of moisture content, temperature, segregation degree repair, and compaction degree requirements on the travel speed.

[0105] Finally, the set pre-selected amplitude is adjusted according to the comprehensive amplitude correction coefficient, the set pre-selected vibration frequency is adjusted according to the comprehensive frequency correction coefficient, and the set pre-selected travel speed is adjusted according to the comprehensive travel speed correction coefficient. The adjusted amplitude, vibration frequency and travel speed are obtained accordingly. The adjusted amplitude, vibration frequency and travel speed are sent to the road roller control system to execute the adjusted amplitude, vibration frequency and travel speed, so as to realize the dynamic adjustment of the road roller's operating parameters.

[0106] In the preferred embodiment of this application, multi-physics field coupling analysis is performed based on the influence of moisture content, temperature, segregation degree and real-time road compaction degree on the roller operating parameters to achieve real-time, adaptive and refined control of the roller operating parameters, eliminate hidden defects, prevent over-compaction and under-compaction, and the entire compaction process is digitally recorded throughout, so that the compaction quality at any point can be traced, transforming post-event quality control into in-process process assurance.

[0107] Accordingly, such as Figure 2 As shown, based on a neural network-based intelligent road roller parameter adjustment method, this embodiment of the invention also provides a neural network-based intelligent road roller parameter adjustment method system, used to implement the neural network-based intelligent road roller parameter adjustment method disclosed in this embodiment of the invention. The system includes: a multi-source data acquisition module 1, a segregation degree identification module 2, a compaction degree compensation module 3, and a dynamic coupling module 4.

[0108] The multi-source data acquisition module 1 is used to acquire multi-source data of the road surface to be constructed during the operation of the road roller. The multi-source data includes image data, temperature distribution data, vibration distribution data and moisture content distribution data.

[0109] The separation degree identification module 2 is used to identify the separation degree based on the image data and the temperature distribution data through a pre-constructed cross-modal spatiotemporal cross-attention model to obtain a separation distribution heatmap. The cross-modal spatiotemporal cross-attention model is set to perform cross-attention fusion of spatiotemporal visual features and spatiotemporal thermal features to generate the separation distribution heatmap. The spatiotemporal visual features are obtained by extracting dual-channel features from the image data using a two-dimensional convolutional neural network and an optical flow network. The spatiotemporal thermal features are obtained by extracting and fusing features from the temperature distribution data sequentially using a convolutional long short-term memory network based on an attention mechanism.

[0110] The compaction compensation module 3 is used to obtain compaction distribution data based on the vibration distribution data, and to perform temperature compensation on the compaction distribution data based on the temperature distribution data to obtain compensated compaction distribution data.

[0111] The dynamic coupling module 4 is used to input the segregation distribution heat map, the temperature distribution data, the moisture content distribution data and the compensated compaction degree distribution data as real-time feedback data into the pre-constructed dynamic coupling model to obtain the optimal rolling operation parameters.

[0112] For specific limitations regarding the neural network-based intelligent road roller parameter adjustment method system, please refer to the above-described limitations of the neural network-based intelligent road roller parameter adjustment method, which will not be repeated here. Those skilled in the art will recognize that the various modules and steps described in conjunction with the embodiments disclosed in this invention can be implemented in hardware, software, or a combination of both. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0113] In summary, the present application provides a neural network-based intelligent road roller parameter adjustment method, which solves the technical problems of improving the accuracy of segregation detection and compaction degree measurement during road roller operations, as well as the matching degree between compaction operation strategies and real-time road surface conditions. The method includes: acquiring multi-source data of the road surface to be constructed during road roller operation, including image data, temperature distribution data, vibration distribution data, and moisture content distribution data; identifying the degree of segregation based on the image data and temperature distribution data using a pre-constructed cross-modal spatiotemporal cross-attention model to obtain a segregation distribution heatmap; the cross-modal spatiotemporal cross-attention model is configured to cross-inject spatiotemporal visual features and spatiotemporal thermal features. The system integrates intention and force to generate a segregation distribution heat map. Spatiotemporal visual features are obtained by extracting dual-channel features from image data using a two-dimensional convolutional neural network and an optical flow network. Spatiotemporal thermal features are obtained by extracting and fusing features from temperature distribution data sequentially using a convolutional long short-term memory network based on an attention mechanism. Based on vibration distribution data, compaction distribution data is obtained, and temperature compensation is performed on the compaction distribution data to obtain compensated compaction distribution data. The segregation distribution heat map, temperature distribution data, moisture content distribution data, and compensated compaction distribution data are used as real-time feedback data input into a pre-constructed dynamic coupling model to obtain optimal rolling operation parameters. This application discloses a neural network-based intelligent road roller parameter adjustment method, which constructs a multimodal pavement state perception system. Through multi-source information complementarity and collaborative verification, it achieves a comprehensive and in-depth understanding of the microscopic mechanism and evolution trend of the compaction process, possessing strong causal reasoning ability and anti-interference capability. It achieves deep, bidirectional, and adaptive fusion of visual modal data and temperature modal data in the spatiotemporal dimension through a cross-modal spatiotemporal cross-attention model. Utilizing the cross-attention mechanism, visual features and thermal features guide and reinforce each other in the spatiotemporal dimension. A hierarchical spatiotemporal transformation encoder determines whether an anomaly is isolated noise or part of a continuous segregation zone, thus endowing it with cross-modal... The modal spatiotemporal cross-attention model, similar to the global judgment capability of experts, greatly improves accuracy. Temperature distribution data compensates for the compaction distribution data calculated from vibration distribution data to eliminate the interference of temperature on pavement compaction, providing a real and consistent decision-making basis for real-time assessment and closed-loop control of pavement compaction quality. Through multi-physics coupling analysis of the influence of moisture content, temperature, segregation degree, and real-time pavement compaction on roller operating parameters, real-time, adaptive, and refined adjustment of roller operating parameters is realized. The entire rolling process is digitally recorded, and the compaction quality at any point is traceable, transforming post-event quality control into in-process process assurance.

[0114] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0115] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this application, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for adjusting the parameters of an intelligent road roller based on a neural network, characterized in that, The method includes: Acquire multi-source data of the road surface to be constructed during roller operation, including image data, temperature distribution data, vibration distribution data, and moisture content distribution data; Based on the image data and the temperature distribution data, the degree of segregation is identified through a pre-constructed cross-modal spatiotemporal cross-attention model to obtain a segregation distribution heatmap. The cross-modal spatiotemporal cross-attention model is set to perform cross-attention fusion of spatiotemporal visual features and spatiotemporal thermal features to generate the segregation distribution heatmap. The spatiotemporal visual features are obtained by extracting dual-channel features from the image data using a two-dimensional convolutional neural network and an optical flow network. The spatiotemporal thermal features are obtained by extracting and fusing features from the temperature distribution data sequentially using a convolutional long short-term memory network based on an attention mechanism. Based on the vibration distribution data, compaction degree distribution data is obtained, and temperature compensation is performed on the compaction degree distribution data based on the temperature distribution data to obtain compensated compaction degree distribution data. The segregation distribution heat map, the temperature distribution data, the moisture content distribution data, and the compensated compaction degree distribution data are used as real-time feedback data input into the pre-constructed dynamic coupling model to obtain the optimal rolling operation parameters; The cross-modal spatiotemporal cross-attention model also performs the following operations: Using the spatiotemporal visual features as the query vector, and using the spatiotemporal hot features as both the key vector and the value vector, attention weights are calculated to obtain a first attention weight. Based on the first attention weight, the spatiotemporal visual features are enhanced to obtain enhanced visual features. Using the spatiotemporal thermal features as the query, and using the spatiotemporal visual features as both the key vector and the value vector, attention weights are calculated to obtain a second attention weight. Based on the second attention weight, the spatiotemporal thermal features are enhanced to obtain enhanced thermal features. The enhanced visual features and the enhanced thermal features are fused to obtain multimodal fusion features. A hierarchical spatiotemporal transform encoder is used to extract contextual spatiotemporal features from the multimodal fusion features to obtain long-term dependent spatiotemporal features. The long-term dependent spatiotemporal features are mapped to pixel-level spatial resolution and a dissociated distribution heatmap is generated.

2. The method for adjusting the parameters of an intelligent road roller based on a neural network as described in claim 1, characterized in that, The acquisition of multi-source data on the road surface to be constructed during roller operation includes: Receive the second pulse signal and time message output by the GNSS timing module, and use the second pulse signal and time message as the absolute time reference to synchronize the master clock; Based on the master clock, continuous sampling is performed through the positioning module and the piezoelectric accelerometer to obtain position timing data and vibration distribution data with the same timestamp. Based on the location timing data, corresponding trigger signals are generated according to different preset spatial intervals, and each trigger signal is sent to the image acquisition module, temperature acquisition module, and near-infrared spectrometer to control the image acquisition module to perform triggered acquisition to obtain image data, control the temperature acquisition module to perform triggered acquisition to obtain temperature distribution data, and control the near-infrared spectrometer to perform triggered acquisition to obtain water content distribution data. All data obtained by triggered acquisition are stamped with a unified timestamp based on the master clock.

3. The method for adjusting the parameters of an intelligent road roller based on a neural network as described in claim 1, characterized in that, The method employs a two-dimensional convolutional neural network and an optical flow network to perform dual-channel feature extraction on the image data, including: The image data is repaired by using a pre-constructed image inpainting generative adversarial network to obtain repaired image data; A two-dimensional convolutional neural network is used to extract the spatial features of each frame of the repaired image data, and the spatial features are stacked along the time dimension to extract temporal features, thus obtaining spatiotemporal features. An optical flow network is used to extract dense motion fields between consecutive frames of the repaired image data. The dense motion fields are stacked and then feature extraction is performed to obtain motion evolution features. The spatiotemporal features and the motion evolution features are fused in the channel dimension to obtain spatiotemporal visual features.

4. The method for adjusting the parameters of an intelligent road roller based on a neural network as described in claim 1, characterized in that, The convolutional long short-term memory network based on an attention mechanism is used to sequentially extract and fuse features from the temperature distribution data, including: A lightweight two-dimensional convolutional neural network is used to extract the temperature change features of the temperature distribution data, and a long short-term memory network is used to extract the temperature distribution features of the temperature distribution data. Based on the temperature change characteristics, a temporal attention weight is generated; based on the temperature distribution characteristics, a spatial attention weight is generated; and based on the spatial attention weight and the temporal attention weight, the temperature change characteristics and the temperature distribution characteristics are fused to obtain spatiotemporal thermal characteristics.

5. The method for adjusting the parameters of an intelligent road roller based on a neural network as described in claim 1, characterized in that, The step of performing temperature compensation on the compaction distribution data based on the temperature distribution data to obtain compensated compaction distribution data includes: The pre-built compaction compensation model was calibrated using a small batch of training datasets collected on the road surface to be constructed, resulting in a calibrated compaction compensation model. Based on the temperature distribution data and the compaction distribution data, the calibrated compaction compensation model is used to obtain the compensated compaction distribution data.

6. The method for adjusting the parameters of an intelligent road roller based on a neural network as described in claim 1, characterized in that, The process of inputting the segregation distribution thermogram, temperature distribution data, moisture content distribution data, and compensated compaction degree distribution data as real-time feedback data into a pre-constructed dynamic coupling model to obtain optimal rolling operation parameters includes: Based on the moisture content distribution data and the preset optimal moisture content, a first set of correction coefficients for the roller operating parameters is generated. Based on the temperature distribution data and the preset optimal operating temperature, a second set of correction coefficients for the roller operating parameters is generated. Based on the segregation distribution heat map and the preset segregation degree threshold, the coupling relationship between segregation degree and energy is simulated to generate a third set of correction coefficients for the roller operating parameters. Based on the compensated compaction degree distribution data and the preset target compaction degree value, the coupling relationship between compaction degree and energy is simulated to generate a fourth set of correction coefficients for the roller operating parameters. The roller operating parameters are adjusted according to the first combination of correction coefficients, the second combination of correction coefficients, the third combination of correction coefficients, and the fourth combination of correction coefficients to obtain the optimal roller operating parameters.

7. The method for adjusting the parameters of an intelligent road roller based on a neural network as described in claim 6, characterized in that, The first set of correction coefficients for generating the roller operating parameters based on the moisture content distribution data and the preset optimal moisture content includes: Based on the moisture content distribution data, the current moisture content of each pre-divided grid cell is obtained. Based on the pre-calibrated humidity correction coefficient gradient model, the gradient corresponding to the humidity difference between the current moisture content and the preset optimal moisture content is determined. The humidity correction coefficient gradient model is set to map the humidity difference to a combination of humidity correction coefficients with different gradients. Based on the attribution gradient, the first set of correction coefficients for the roller operation parameters is obtained.

8. The method for adjusting the parameters of an intelligent road roller based on a neural network as described in claim 6, characterized in that, The step of simulating the coupling relationship between the degree of segregation and energy based on the segregation distribution heat map and a preset segregation degree threshold, and generating a third set of correction coefficients for the roller operating parameters, includes: Based on the segregation distribution heatmap, the current segregation index of each pre-divided grid cell is obtained, and based on the current segregation index and the preset segregation threshold, the segregation improvement requirement of each grid cell is obtained. Based on the required segregation degree improvement, the first total required energy is obtained using a pre-constructed energy segregation degree improvement mapping model; Based on the remaining number of compaction passes and the preset energy allocation principle, the first total energy demand is allocated to obtain the first compaction energy demand for this compaction, and the first compaction energy demand is mapped to the third correction coefficient combination of the roller operation parameters.

9. A neural network-based intelligent road roller parameter adjustment system, used to implement the neural network-based intelligent road roller parameter adjustment method according to any one of claims 1-8, characterized in that, The system includes: a multi-source data acquisition module, a segregation degree identification module, a compaction degree compensation module, and a dynamic coupling module; The multi-source data acquisition module is used to acquire multi-source data of the road surface to be constructed during the operation of the road roller. The multi-source data includes image data, temperature distribution data, vibration distribution data and moisture content distribution data. The segregation degree recognition module is used to identify the segregation degree based on the image data and the temperature distribution data through a pre-constructed cross-modal spatiotemporal cross-attention model to obtain a segregation distribution heatmap. The cross-modal spatiotemporal cross-attention model is set to perform cross-attention fusion of spatiotemporal visual features and spatiotemporal thermal features to generate the segregation distribution heatmap. The spatiotemporal visual features are obtained by extracting dual-channel features from the image data using a two-dimensional convolutional neural network and an optical flow network. The spatiotemporal thermal features are obtained by extracting and fusing features from the temperature distribution data sequentially using a convolutional long short-term memory network based on an attention mechanism. The compaction compensation module is used to obtain compaction distribution data based on the vibration distribution data, and to perform temperature compensation on the compaction distribution data based on the temperature distribution data to obtain compensated compaction distribution data. The dynamic coupling module is used to input the segregation distribution heat map, the temperature distribution data, the moisture content distribution data and the compensated compaction degree distribution data as real-time feedback data into the pre-constructed dynamic coupling model to obtain the optimal rolling operation parameters; The cross-modal spatiotemporal cross-attention model also performs the following operations: Using the spatiotemporal visual features as the query vector, and using the spatiotemporal hot features as both the key vector and the value vector, attention weights are calculated to obtain a first attention weight. Based on the first attention weight, the spatiotemporal visual features are enhanced to obtain enhanced visual features. Using the spatiotemporal thermal features as the query, and using the spatiotemporal visual features as both the key vector and the value vector, attention weights are calculated to obtain a second attention weight. Based on the second attention weight, the spatiotemporal thermal features are enhanced to obtain enhanced thermal features. The enhanced visual features and the enhanced thermal features are fused to obtain multimodal fusion features. A hierarchical spatiotemporal transform encoder is used to extract contextual spatiotemporal features from the multimodal fusion features to obtain long-term dependent spatiotemporal features. Based on the long-term dependent spatiotemporal features, a dissociated distribution heatmap is obtained.

Images