Internet-of-things detection method and system for compaction degree of asphalt mixture

Abstract

The invention relates to the technical field of road engineering quality detection, and discloses an Internet of Things detection method and system for the compaction degree of an asphalt mixture, and the method comprises the steps: obtaining a working state signal of a road roller, and driving a plurality of types of sensors to carry out the cooperative collection; performing intelligent preprocessing on the original data flow of the multi-source heterogeneous sensor in an edge computing unit; intelligent recognition of roadbed rigidity, mixture temperature state and rolling stage is carried out on the multi-dimensional comprehensive feature vector, and compensation and correction are carried out; constructing and training a lightweight space-time convolutional neural network model, and performing real-time prediction of the compactness through transfer learning and online optimization; carrying out three-dimensional visualization processing and intelligent analysis; safe and reliable transmission and cloud deep analysis are carried out; by constructing the intelligent detection system based on the Internet of Things technology and multi-sensor fusion, real-time, continuous and full-coverage monitoring of the asphalt mixture compaction process is realized.

Description

Technical Field

[0001] This invention relates to the field of road engineering quality testing technology, and more specifically, to an Internet of Things (IoT) method and system for detecting the compaction degree of asphalt mixtures. Background Technology

[0002] The compaction degree of asphalt mixture is a core indicator for evaluating the construction quality of asphalt pavement, directly determining the pavement's load-bearing capacity, deformation resistance, durability, and service life. In my country's tens of thousands of kilometers of newly built and reconstructed expressways and high-grade highways each year, asphalt pavement accounts for over 90%, making compaction quality control a consistently key and challenging aspect of construction quality management. Insufficient compaction leads to premature rutting, potholes, cracks, and other pavement defects, severely impacting road performance and driving safety, causing significant economic losses and social impact.

[0003] Traditional methods for testing the compaction degree of asphalt mixtures mainly include core sampling, nuclear density metering, and falling weight deflectometers. Core sampling is considered the most accurate method, but it requires damaging the pavement structure, and samples need to be sent to a laboratory for density testing, resulting in testing cycles that can take hours or even days. The results are significantly delayed compared to the construction schedule; by the time substandard compaction is detected, the mixture has already cooled and solidified, requiring milling and rework, leading to material waste and project delays. Nuclear density metering uses radioactive isotopes to measure pavement density, allowing for rapid on-site testing, but it poses radiation safety risks, has limited testing depth, and the test points are scattered, making it difficult to reflect the compaction quality distribution across the entire construction area. Falling weight deflectometers indirectly assess compaction quality by measuring pavement deflection, but the results are greatly affected by subgrade conditions and lack a direct correlation with compaction degree. These traditional methods generally suffer from low testing efficiency, poor real-time performance, limited coverage, and significant human influence, making them unsuitable for the refined quality control requirements of modern high-grade highway construction.

[0004] With the rapid development of IoT, sensor, wireless communication, edge computing, and artificial intelligence technologies, new technological approaches have been provided for the real-time intelligent detection of asphalt mixture compaction. By deploying various types of sensors on construction equipment such as road rollers, multi-dimensional data such as vibration, pressure, temperature, and displacement during the compaction process can be collected in real time. This data is then transmitted to a computing platform via a wireless network for intelligent analysis and processing, enabling full-coverage, all-weather, and real-time dynamic monitoring of the construction process. However, existing technologies still have many shortcomings in areas such as the accuracy and stability of data acquisition, the fusion of multi-source data and compensation for environmental interference factors, real-time computing capabilities at the edge, and the generalization and adaptive optimization capabilities of the model. These limitations restrict the widespread application of IoT-based intelligent detection technology in the quality control of asphalt pavement construction.

[0005] Therefore, it is urgent to conduct in-depth research, break through key technological bottlenecks, and develop a practical Internet of Things (IoT) system for detecting the compaction degree of asphalt mixtures, so as to provide more accurate, efficient, and intelligent technical means for the quality control of asphalt pavement construction. Summary of the Invention

[0006] This invention provides an IoT-based method and system for detecting the compaction degree of asphalt mixtures, solving the technical problems of poor real-time performance, low coverage, and weak environmental adaptability of traditional detection methods in related technologies.

[0007] This invention provides an IoT-based method for detecting the compaction degree of asphalt mixtures, comprising: Acquire the operating status signal of the road roller, drive multiple types of sensors to collect data collaboratively, and obtain the raw data stream of multi-source heterogeneous sensors; The raw data streams from multi-source heterogeneous sensors are intelligently preprocessed in the edge computing unit to obtain multi-dimensional comprehensive feature vectors. Intelligent identification and compensation correction are performed on the multi-dimensional comprehensive feature vector to obtain the feature vector after adaptive compensation, based on the subgrade stiffness, mixture temperature state and compaction stage. Based on the adaptively compensated feature vectors, a lightweight spatiotemporal convolutional neural network model is constructed and trained. Through transfer learning and online optimization, the compaction degree is predicted in real time, and a compaction degree data package is obtained. The compaction data package is subjected to three-dimensional visualization and intelligent analysis to obtain multi-level quality distribution maps and intelligent construction adjustment schemes. Based on the compaction data package and the raw data stream from multi-source heterogeneous sensors, secure and reliable transmission and in-depth cloud analysis are performed to obtain quality analysis reports and optimization solutions.

[0008] In a preferred embodiment, the driving of multi-type sensor collaborative acquisition includes: A distributed sensor network topology is adopted, and the triaxial accelerometer is fixed to the inside of the bearing seat of the vibrating wheel of the road roller using a special damping rubber bracket; Piezoelectric dynamic pressure sensors are evenly distributed along the width of the vibrating wheel to form a pressure distribution measurement array, which is installed at the interface between the vibrating wheel and the road surface. A non-contact infrared temperature sensor array is used to achieve two-dimensional distribution measurement of the temperature field on the surface of the mixture; The BeiDou-3 high-precision RTK positioning module is used to output PPS second pulse signals as a unified time reference.

[0009] In a preferred embodiment, obtaining the multi-dimensional comprehensive feature vector includes: The triaxial acceleration signal is denoised and decomposed into three layers of wavelet. The dominant frequency and high-frequency attenuation coefficient of the Z-axis vertical acceleration signal are extracted to obtain the vibration time-frequency feature vector. The coefficient of variation of the spatial distribution of pressure is calculated for the pressure sensor array, and the time-domain fluctuation characteristics and power spectrum peak characteristics of pressure are calculated to obtain the pressure distribution and dynamic feature vector. A high-resolution temperature field is reconstructed using a bicubic interpolation algorithm on the temperature data, and the statistical features, spatial gradient features, and temporal dynamic features of the temperature field are extracted to obtain the temperature field feature vector. The elevation difference before and after compaction is calculated from the data of the front and rear displacement sensors to obtain the deformation feature vector; The vibration time-frequency feature vector, pressure distribution and dynamic feature vector, temperature field feature vector, and deformation feature vector are concatenated and Z-score normalized to obtain a multi-dimensional comprehensive feature vector.

[0010] In a preferred embodiment, the intelligent identification and compensation correction of subgrade stiffness, mixture temperature state, and compaction stage of the multi-dimensional comprehensive feature vector includes: Three features—vibration high-frequency attenuation coefficient, deformation recovery rate, and cumulative deformation—are extracted. A pre-trained radial basis function kernel is used to identify the roadbed stiffness state and classify the roadbed stiffness. Four features were extracted from the temperature field feature vector: average temperature, standard deviation, proportion of high-temperature zone, and proportion of low-temperature zone. Fuzzy C-means clustering algorithm was used to identify the temperature state of the mixture. A long short-term memory network model was used to identify the crushing stage. Based on the recognition results, a three-dimensional environmental state combination index is formed. The corresponding compensation gain coefficient and compensation bias are read from the pre-established environmental compensation parameter lookup table, and the normalized feature vector is linearly corrected.

[0011] In a preferred embodiment, the construction and training of a lightweight spatiotemporal convolutional neural network model for real-time prediction of compaction through transfer learning and online optimization includes: Construct a lightweight spatiotemporal convolutional neural network model; The physical constraint layer establishes the physical constraint conditions of the compaction process based on the Burgers four-element viscoelastic constitutive model of the asphalt mixture compaction process, and transforms the upper limit constraint of compaction degree and the compaction degree growth constraint into the penalty term of the neural network layer; The knowledge from the large teacher model is transferred to the lightweight student model, and uncertainty is estimated.

[0012] In a preferred embodiment, the three-dimensional visualization and intelligent analysis of the compaction data package includes: A digital twin virtual environment for the construction site is built based on the Unity3D real-time 3D rendering engine, including a 3D terrain model, a road geometry model, and a road roller equipment model; The Kriging space interpolation algorithm is used to generate a continuous distribution field of compaction degree for the entire road section. The gridded compaction degree field is mapped onto the surface of the three-dimensional road model and visualized using a heat map color coding method. Automatic region identification and labeling are performed, and connected component analysis algorithms are used to extract connected outlier regions; Based on real-time compaction degree prediction, current number of compaction passes, mixture temperature, and roller speed, a fuzzy control algorithm is used to generate suggestions for adjusting construction parameters. Design differentiated visual interfaces for different user roles.

[0013] In a preferred embodiment, the multi-sensor collaborative data acquisition further includes: The hardware-level synchronous controller based on FPGA receives the PPS second pulse signal as the master clock reference, multiplies the PPS signal to 100 MHz as the system master clock, and adopts multi-channel synchronous trigger control logic to control the delay time after each acquisition unit receives the trigger signal. The FPGA integrates a hardware timestamp generation module to automatically append a UTC timestamp to each set of sampled data. The FPGA reads the current RTK positioning coordinates and binds the spatial coordinates to the sensor data.

[0014] In a preferred embodiment, the real-time prediction of compaction degree further includes: Based on the received adaptively compensated feature vector, input it into a lightweight spatiotemporal convolutional neural network model for forward inference calculation; Outlier detection is performed to remove outliers that deviate from the mean by more than three times the standard deviation. The mean and standard deviation of the remaining valid predictions are recalculated. The mean is used as the final compaction degree prediction value, and the standard deviation is used as the prediction confidence level. Based on the assumption of normal distribution, construct confidence intervals; Calculate the physical constraint satisfaction index. If the predicted compaction degree is less than or equal to the maximum compaction degree and the compaction degree increment is less than or equal to the maximum increment, the physical constraint satisfaction is one; otherwise, it is zero.

[0015] In a preferred embodiment, the automatic region identification and labeling includes: Statistical analysis was conducted based on the compaction data of the entire road section to calculate the average, standard deviation, maximum, minimum, and pass rate of compaction. The pass rate was defined as the proportion of measuring points with compaction greater than or equal to the pass threshold. Draw a histogram of compaction degree frequency distribution and statistically analyze the distribution of the number of measuring points in different compaction degree intervals; A connected component analysis algorithm is used to mark grid cells with a compaction degree less than the qualified threshold as abnormal cells. Connectivity analysis is performed on the abnormal cells to extract connected abnormal regions. The area, perimeter, centroid coordinates, and minimum compaction degree of each abnormal region are calculated. In the 3D visualization interface, abnormal areas are marked with a red semi-transparent overlay, and an abnormal marker icon is placed at the centroid, generating a compaction quality analysis report.

[0016] In a preferred embodiment, an IoT-based asphalt mixture compaction degree detection system is used to perform the above-described IoT-based asphalt mixture compaction degree detection method, including: The multi-source data acquisition module is used to acquire the working status signal of the road roller, drive multiple types of sensors to collect data collaboratively, and obtain the raw data stream of the multi-source heterogeneous sensors; The intelligent preprocessing module is used to perform intelligent preprocessing on the raw data streams from multi-source heterogeneous sensors in the edge computing unit to obtain multi-dimensional comprehensive feature vectors; The environmental compensation module is used to intelligently identify and compensate for the roadbed stiffness, mixture temperature state and compaction stage of the multi-dimensional comprehensive feature vector, and obtain the adaptively compensated feature vector. The compaction degree prediction module constructs and trains a lightweight spatiotemporal convolutional neural network model based on the adaptively compensated feature vectors. It then performs real-time compaction degree prediction through transfer learning and online optimization to obtain a compaction degree data package. The visualization analysis module is used to perform three-dimensional visualization processing and intelligent analysis on compaction data packages to obtain multi-level quality distribution maps and intelligent construction adjustment schemes. The cloud-based analytics module is used for secure and reliable transmission and in-depth cloud-based analysis of compaction data packets and raw data streams from multi-source heterogeneous sensors, resulting in quality analysis reports and optimization solutions.

[0017] The beneficial effects of this invention are as follows: By constructing an intelligent detection system based on Internet of Things (IoT) technology and multi-sensor fusion, this invention achieves real-time, continuous, and comprehensive monitoring of the asphalt mixture compaction process. Through multi-sensor data fusion and edge computing technologies, continuous monitoring of the compaction quality across the entire road section is achieved, increasing the detection coverage compared to traditional sampling methods. This system can accurately identify areas with weak compaction quality and effectively avoid potential quality problems. Through low-latency data transmission technology and a real-time feedback mechanism, the roller operator can adjust construction parameters based on real-time feedback within the suitable temperature range of the mixture, thereby achieving dynamic optimization of the compaction process. Through high-precision prediction algorithms and error control technology, a high degree of consistency between predicted and measured compaction degrees is achieved, meeting engineering accuracy requirements. Through intelligent construction guidance technology, the compaction degree qualification rate and the uniformity of compaction quality are improved. By enhancing the level of construction quality control, the service life of the pavement is effectively extended, and maintenance costs are reduced.

[0018] By proposing a deep learning model technique embedding physical constraints, the physical laws of the asphalt mixture compaction process, such as the Burgers viscoelastic constitutive relation, are embedded into the neural network structure as constraints. This allows the model output to both fit the statistical laws of the training data and conform to the physical mechanism of the compaction process, improving the model's generalization ability, prediction reliability, and interpretability. Through physical constraint modeling, the model maintains stable prediction accuracy under different construction conditions, different mixture types, and different equipment parameters, overcoming the problem of decreased prediction accuracy in traditional pure data-driven deep learning models when the training data distribution is outside the expected range. Through transfer learning, good performance can be achieved with only a small amount of on-site calibration data when deploying new projects, reducing system deployment costs and time compared to traditional models. Through an online incremental learning mechanism, the model continuously learns and optimizes from construction data, thereby achieving dynamic performance improvement. Attached Figure Description

[0019] Figure 1 This is a flowchart of an IoT-based method for detecting the compaction degree of asphalt mixtures according to the present invention; Figure 2 This is a module diagram of an Internet of Things (IoT) detection system for asphalt mixture compaction degree in this invention. Detailed Implementation

[0020] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0021] At least one embodiment of the present invention discloses an IoT-based method for detecting the compaction degree of asphalt mixtures, such as... Figure 1 As shown, it includes: S1, acquire the working status signal of the road roller, drive multiple types of sensors to collect data collaboratively, and obtain the raw data stream of multi-source heterogeneous sensors; S11, Sensor System Deployment and Hardware Configuration. Based on the structural characteristics and construction operation features of the road roller, a distributed sensor network topology is adopted. A MEMS triaxial accelerometer is installed inside the bearing housing of the road roller's vibratory drum, at a horizontal distance of 150mm and a vertical distance of 80mm from the center of mass of the vibratory drum. It is fixed by a dedicated damping rubber bracket made of Shore 60A nitrile rubber, which effectively attenuates high-frequency vibration interference transmitted from the installation interface while maintaining sensitivity to compaction vibration signals. Eight piezoelectric dynamic pressure sensors are evenly distributed along the width of the vibratory drum to form a pressure distribution measurement array. Each sensor has a range of 0-2MPa, a sampling frequency of 1000Hz, and a sensitivity of 2.5pC / N. They are installed at the interface between the vibratory drum and the road surface, and are waterproofed and dustproofed by a thin protective shell. A 16×12 array of non-contact infrared temperature sensors (192 points in total) is installed on the road roller body at a height of 1 meter above the road surface. At a position of 0.2m, with an array field of view of 45°×35°, it can cover a 3m×2m road surface area in front of the vibrating roller, realizing two-dimensional distribution measurement of the temperature field on the surface of the mixture. The sensor response time is less than 100ms, the temperature measurement range is 80-180℃, and the accuracy is ±2℃. Two laser triangulation displacement sensors are installed on the front and rear sides of the vibrating roller respectively, with a measurement range of 0-50mm, a resolution of 0.01mm, and a sampling frequency of 500Hz, used to detect changes in road surface elevation before and after compaction. A Beidou-3 high-precision RTK positioning module is installed on the top of the road roller by bolt fixing. The spatial offset between the antenna phase center and the geometric center of the road roller is precisely measured and recorded by a total station. The positioning module outputs a positioning accuracy of ±2cm for plane and ±5cm for elevation, with a positioning frequency of 10Hz. It also outputs a PPS second pulse signal as a unified time reference for the system, with pulse rise edge jitter of less than 20ns.

[0022] S12, hardware-level spatiotemporal synchronization trigger control. The hardware-level synchronization controller, implemented using an FPGA, receives the PPS (pixel per second) signal output from the BeiDou positioning module as the master clock reference. The FPGA is a Xilinx FPGA. The Artix-7 series chips feature an internal high-precision phase-locked loop circuit that multiplies the PPS signal to 100MHz as the system's main clock. Employing multi-channel synchronous trigger control logic, trigger signals are synchronously sent to the accelerometer, pressure sensor, temperature sensor, and displacement sensor acquisition units at the exact 10ms mark after the rising edge of each PPS pulse. The delay time for each acquisition unit after receiving the trigger signal is controlled within 1μs, ensuring microsecond-level time synchronization accuracy of the multi-source sensor data and obtaining strictly aligned multi-source data. The FPGA integrates a hardware timestamp generation module, automatically appending a UTC timestamp to each set of sampled data. The timestamp format is year-month-day-hour-minute-second-millisecond-microsecond, achieving microsecond-level accuracy. Simultaneously, the FPGA reads the current RTK positioning coordinates from the BeiDou positioning module, in WGS84 geodetic coordinate system (longitude, latitude, and elevation), binding the spatial coordinates to the sensor data to achieve spatial labeling. The spatial positioning accuracy is ±2cm for horizontal plane and ±5cm for vertical plane, meeting the accuracy requirements for drawing compaction spatial distribution maps.

[0023] S13, Multi-source heterogeneous data acquisition and packaging. Based on a synchronous trigger signal, a triaxial accelerometer collects acceleration time-series data of the vibrating wheel in the X-direction, Y-direction, and Z-direction directions at a sampling frequency of 1000Hz. Each sampling time window is 10ms, yielding an acceleration vector sequence of 10 sampling points in floating-point format. An 8-channel pressure sensor array synchronously collects pressure distribution data at the interface between the vibrating wheel and the road surface at a sampling frequency of 1000Hz, with a single sampling time window of 10ms, resulting in an 8×10 pressure data matrix in floating-point format. A 16×12 infrared temperature sensor array collects the temperature field distribution on the surface of the mixture at a sampling frequency of 100Hz. Due to the relatively slow rate of temperature change, a high sampling frequency is unnecessary; a single sampling yields 192 temperature data points. The data is in floating-point format. Two displacement sensors, one before and one after, collect the road surface elevation before and after compaction at a sampling frequency of 500Hz. The single sampling time window is 10ms, obtaining the elevation data of 5 sampling points on the front and 5 sampling points on the back. The data format is floating-point. The FPGA encapsulates the above multi-source data along with UTC timestamps, RTK positioning coordinates, and roller operating parameters to form a unified data frame structure. The data frame includes a frame header, timestamp, coordinates, acceleration data, pressure data, temperature data, displacement data, operating parameters, checksum, and frame tail. The size of a single data frame is approximately 3KB, resulting in the raw data stream from the multi-source heterogeneous sensors. The data stream is output to the edge computing unit at a frequency of 100Hz. Data transmission uses a gigabit Ethernet interface with a transmission delay of less than 1ms.

[0024] S2 intelligently preprocesses the raw data streams from multi-source heterogeneous sensors in the edge computing unit to obtain multi-dimensional comprehensive feature vectors. S21, Adaptive Filtering and Time-Frequency Feature Extraction of Acceleration Signal. Based on the received triaxial acceleration time-series data, an Adaptive Kalman Filter (AKF) algorithm is used for noise reduction. The filter estimates the true acceleration signal by establishing a system state model and an observation model. A state transition model is established, representing the current acceleration state vector as the result of the previous state vector transformed by the state transition matrix, plus the product of the control input matrix and the control input, and superimposed with process noise. An observation model is then established, representing the observed acceleration value as the result of the true state vector transformed by the observation matrix and then superimposed with observation noise. Both process noise and observation noise are assumed to be Gaussian white noise, and the initial state covariance matrix is ​​set as the identity matrix. To enable the filter to adapt to different noise environments, an adaptive estimation method is used to update the process noise covariance matrix and the observation noise covariance matrix in real time. The values ​​of these two matrices are dynamically adjusted by calculating the statistical characteristics of the filter residuals, allowing the filter to automatically adjust its parameters according to the actual noise environment to obtain the denoised acceleration signal.

[0025] The denoised acceleration signal was decomposed using a Daubechies 4th-order wavelet basis with three levels of wavelet decomposition, yielding three levels of detail coefficients and one level of approximation coefficients. The energy distribution of the detail coefficients in different frequency bands was then calculated. Specifically, the detail coefficients were divided into three frequency bands: 0.5–5 Hz (low frequency), 5–25 Hz (mid frequency), and 25–50 Hz (high frequency). The energy value for each band was calculated by squaring the absolute values ​​of all detail coefficients within that band and then summing the squared values ​​to obtain the total energy for that band.

[0026] The dominant frequency of the Z-axis vertical acceleration signal is extracted, and the power spectral density is calculated using Fast Fourier Transform (FFT). The dominant frequency is the frequency corresponding to the maximum power spectral density value. The high-frequency attenuation coefficient is calculated, defined as the ratio of high-frequency energy to mid-frequency energy, obtained by dividing the high-frequency energy by the mid-frequency energy. This coefficient reflects the propagation and attenuation characteristics of vibration energy in the roadbed, indirectly characterizing the roadbed stiffness. The root mean square (RMS), peak value, and waveform factor of the triaxial acceleration are used as statistical features. The waveform factor is defined as the ratio of the peak value to the RMS value, resulting in a 12-dimensional vibration time-frequency feature vector, including energy in the three frequency bands, dominant frequency, attenuation coefficient, RMS values ​​of the X, Y, and Z axes, peak values ​​of the X, Y, and Z axes, and waveform factor.

[0027] S22, Extraction of Spatial Distribution and Dynamic Features of Pressure Signals. Based on the received 8-channel pressure sensor array, the uniformity index of the spatial distribution of pressure is calculated, and the coefficient of variation is used to characterize the uniformity of the pressure distribution. The calculation steps are as follows: calculate the average value of the pressure values ​​of the 8 channels, then calculate the standard deviation of the pressure values ​​of the 8 channels, and divide the standard deviation by the average value to obtain the coefficient of variation. The smaller the coefficient of variation value, the more uniform the pressure distribution and the better the compaction effect. The temporal fluctuation characteristics of pressure are calculated. The standard deviation of 10 pressure sampling points within a 10ms time window for each channel is calculated, and the average of the standard deviations of the 8 channels is used to obtain the temporal fluctuation index of pressure, which reflects the dynamic stability of the compaction process. FFT transformation is performed on the pressure time series data of the 8 channels respectively to calculate the power spectral density. The peak value of the power spectral density of each channel is extracted, and the average of the peak values ​​of the 8 channels is used to obtain the pressure power spectrum peak feature, which reflects the intensity of pressure pulsation. The maximum and minimum pressure values ​​of the eight channels are calculated. The pressure distribution range is obtained by subtracting the minimum pressure value from the maximum pressure value. This feature reflects the spatial concentration of compaction contact stress. The average pressure of the eight channels is used as the overall contact pressure feature to obtain an eight-dimensional pressure distribution and dynamic feature vector, including the coefficient of variation, time-domain fluctuation, power spectrum peak value, pressure range, maximum pressure, minimum pressure, average pressure, and pressure distribution skewness coefficient. The skewness coefficient is calculated using the third-order central moment to reflect the asymmetry of the pressure distribution.

[0028] S23, Temperature Field Reconstruction and Temperature Feature Extraction. Based on the received 192 temperature data points, a bicubic interpolation algorithm is used to reconstruct a high-resolution temperature field. The temperature field grid is refined to 64×48, totaling 3072 points. The interpolation function is a piecewise cubic polynomial, which has a continuous second-order derivative within each original grid, ensuring the smoothness of the reconstructed temperature field. The specific implementation of the bicubic interpolation algorithm is as follows: For the point to be interpolated, its position in the original 16×12 grid is determined, and 4×4 neighboring grid points containing that point are found, obtaining the temperature values ​​of these 16 grid points. A bicubic polynomial is constructed, containing 16 coefficient terms, each corresponding to a different power combination of x and y. These 16 coefficients are determined by solving a system of linear equations, ensuring that the function values, first-order partial derivatives, and second-order mixed partial derivatives of the interpolation function at the four corner points are consistent with the original data. The temperature values ​​of the interpolation points are calculated. This method ensures the continuity and smoothness of the reconstructed temperature field.

[0029] Based on the reconstructed high-resolution temperature field, a 10-dimensional feature vector is extracted hierarchically according to the physical meaning of the temperature field characteristics and the requirements of engineering applications. Statistical features of the temperature field are extracted, specifically: the average temperature of the reconstructed temperature field is calculated as the overall temperature state characteristic of the mixture; the standard deviation of the temperature is calculated to reflect the dispersion of the temperature field; a larger standard deviation indicates a more uneven temperature distribution; and the highest and lowest temperatures of the temperature field are extracted. The temperature range is obtained by subtracting the lowest temperature from the highest temperature, and this feature reflects the overall distribution range of the temperature field.

[0030] The spatial gradient features of the temperature field are extracted, specifically by using the Sobel operator to calculate the gradient field, where the gradient vector contains the partial derivatives of temperature in the x and y directions. The magnitude of the gradient field is calculated by squaring the partial derivatives in the x and y directions, summing them, and then taking the square root. The maximum value of the gradient magnitude is extracted as the maximum temperature gradient feature, which reflects abrupt temperature changes, typically corresponding to uneven cooling or segregation of the mixture.

[0031] Extracting the time-domain dynamic characteristics of the temperature field involves calculating the rate of change of the average temperature between two adjacent sampling points, i.e., the cooling rate. The calculation method is as follows: subtract the average temperature of the previous moment from the current average temperature to obtain the temperature change; divide the temperature change by the sampling time interval of 10 ms to obtain the cooling rate. The cooling rate is a negative value; the larger its absolute value, the faster the cooling. This characteristic reflects the change in the thermodynamic state of the mixture.

[0032] The temperature field distribution structure characteristics are extracted, specifically by dividing the temperature field into three regions: a high-temperature zone (>140℃), a suitable-temperature zone (110-140℃), and a low-temperature zone (<110℃). The area ratio of each region is calculated as the temperature distribution structure characteristic. The proportion of the high-temperature zone reflects the proportion of freshly laid mixture, the proportion of the suitable-temperature zone reflects the proportion of the area with the optimal compaction temperature, and the proportion of the low-temperature zone reflects the proportion of the area requiring accelerated construction.

[0033] A 10-dimensional temperature field feature vector is obtained, including average temperature, standard deviation, maximum temperature, minimum temperature, temperature range, maximum gradient, cooling rate, proportion of high-temperature zone, proportion of suitable-temperature zone, and proportion of low-temperature zone.

[0034] S24, Displacement Signal Differential Analysis and Deformation Feature Extraction. Based on the received displacement sensor data from the front and rear sides, the average elevation of the five sampling points on the front side and the average elevation of the five sampling points on the rear side are calculated. The elevation difference before and after compaction is obtained by subtracting the average elevation of the rear side from the average elevation of the front side. This value represents the road settlement caused by vibratory roller compaction, i.e., the instantaneous compaction deformation. Ten sets of elevation difference data collected within a historical 100ms time window are accumulated to obtain the cumulative permanent deformation, which reflects the cumulative compaction effect of continuous rolling. The rate of change of elevation difference, i.e., the deformation rate, is calculated as follows: the elevation difference at the current moment is subtracted from the elevation difference at the previous moment to obtain the elevation difference change. The elevation difference change is divided by the time interval of 10ms to obtain the deformation rate. The deformation rate reflects the dynamic characteristics of the compaction process. As the number of rolling passes increases, the deformation rate gradually decreases. The standard deviation of the 10 sets of elevation difference data is calculated to reflect the fluctuation of deformation. The smaller the standard deviation, the more stable the compaction process. The deformation recovery rate is calculated as the ratio of the rebound of the road surface elevation after the vibratory roller is lifted to the total deformation. The rebound is obtained by comparing the current measurement value of the front sensor with the historical measurement value of the rear sensor. The deformation recovery rate reflects the elastic properties of the mixture and is closely related to the mixture temperature and viscoelastic state. The maximum and minimum values ​​of the deformation are extracted to obtain a 6-dimensional deformation feature vector, including instantaneous deformation, cumulative deformation, deformation rate, deformation standard deviation, deformation recovery rate, and maximum deformation.

[0035] S25, Multi-source Feature Vector Fusion and Normalization. Based on the extracted 12-dimensional vibration time-frequency feature vector, 8-dimensional pressure distribution and dynamic feature vector, 10-dimensional temperature field feature vector, and 6-dimensional deformation feature vector, feature vectors are concatenated to obtain a multi-dimensional comprehensive feature vector. Features 1 to 12 are vibration features, features 13 to 20 are pressure features, features 21 to 30 are temperature features, and features 31 to 36 are deformation features. To eliminate the influence of differences in feature dimensions on subsequent models, the Z-score normalization method is used to normalize the feature vector. The normalization method is as follows: for each feature value, first subtract the mean of that feature on the historical dataset, then divide by the standard deviation of that feature to obtain the normalized feature value. The normalized feature value has a mean of 0 and a standard deviation of 1, resulting in the normalized multi-dimensional comprehensive feature vector, while retaining the corresponding UTC timestamp and RTK spatial coordinates.

[0036] S3 intelligently identifies and compensates for the subgrade stiffness, mixture temperature state and compaction stage of the multi-dimensional comprehensive feature vector to obtain the adaptively compensated feature vector. S31, Intelligent Identification of Subgrade Stiffness State. Based on the received normalized multi-dimensional comprehensive feature vector, three features—vibration high-frequency attenuation coefficient, deformation recovery rate, and cumulative deformation—are extracted as input features for subgrade stiffness identification. These three features are strongly correlated with subgrade stiffness. The vibration high-frequency attenuation coefficient is a dimensionless ratio, typically ranging from 0.1 to 2.0; the deformation recovery rate is in percentage form, ranging from 0% to 100%; and the cumulative deformation is in millimeters, ranging from 0 to 50 mm. To ensure equal weights for the three features in the support vector machine classification, a Min-Max normalization method is used to uniformly scale the three features to the range of 0 to 1. The normalization method is as follows: for each feature value, first subtract the minimum value of that feature in the historical dataset, then divide by the difference between the maximum and minimum values ​​of that feature to obtain the normalized feature value.

[0037] A pre-trained Radial Basis Function (RBF) support vector machine (SVM) three-classifier was used to identify the roadbed stiffness state. The SVM kernel function uses a radial basis function, which is calculated by multiplying the square of the Euclidean distance between two feature vectors by a negative kernel function parameter, and finally taking the exponent. The optimal kernel function parameter was determined to be 0.5 through grid search and cross-validation. The SVM achieves classification by finding the optimal hyperplane in the feature space. The hyperplane is defined by a linear combination of Lagrange multipliers, class labels, kernel function values, and bias terms. The classifier identifies roadbed stiffness into three categories: weak, moderate, and hard. The classification decision function outputs a three-dimensional probability vector, and the category with the highest probability is selected as the identification result. Weak roadbeds correspond to newly filled roadbeds or areas with high water content; their stiffness modulus is generally less than 30 MPa. These roadbeds exhibit slow vibration decay, a large high-frequency attenuation coefficient, and a high deformation recovery rate. Medium-strength subgrades correspond to normally compacted subgrades with a stiffness modulus of 30-80 MPa and moderate vibration attenuation. Hard subgrades correspond to older subgrades or very stiff bridge / culvert transition sections with a stiffness modulus greater than 80 MPa, exhibiting rapid vibration attenuation, a low high-frequency attenuation coefficient, and a low deformation recovery rate. Based on the identification results, a subgrade stiffness category label is output, which includes three possible values: weak, medium, and hard.

[0038] S32, Intelligent Identification of Mixture Temperature State. Based on four features from the received temperature field feature vector—average temperature, standard deviation, proportion of high-temperature zone, and proportion of low-temperature zone—a fuzzy C-means clustering algorithm is used to identify the mixture temperature state. Fuzzy C-means clustering achieves clustering by minimizing an objective function, which is composed of the weighted sum of squared distances of the membership degrees of all samples. Membership degree represents the degree to which a sample belongs to a certain category, the fuzzy weighting exponent is 2, the sample feature vector is the input four-dimensional feature vector, and the cluster center is the center point of each category. The objective function is minimized through an iterative optimization algorithm. The iterative process includes two steps: updating membership degrees—for each sample and each category, the ratio of the distance from the sample to the center of that category to the distance to all category centers is calculated, and the membership degree is calculated based on these ratios; updating cluster centers—for each category, the weighted feature vectors of the membership degrees of all samples are summed, and then divided by the sum of the membership degrees to obtain the new cluster centers. The iteration termination condition is that the change in cluster centers between two consecutive iterations is less than a threshold of 0.001. The clustering algorithm identifies three temperature states: high-temperature zone, moderate-temperature zone, and low-temperature zone. The high-temperature zone corresponds to freshly laid mixtures with an average temperature greater than 140℃. The mixture has strong fluidity and high compaction efficiency but is prone to displacement. The moderate-temperature zone corresponds to mixtures with temperatures between 110-140℃, representing the optimal compaction temperature range. The mixture has moderate viscoelasticity and optimal compaction effect. The low-temperature zone corresponds to mixtures with temperatures below 110℃, where viscosity increases and compaction efficiency decreases. Based on the clustering results, a temperature state category label is output, containing three possible values: high-temperature zone, moderate-temperature zone, and low-temperature zone. Simultaneously, a membership vector of the sample to each of the three categories is output. The membership vector contains three elements, representing the membership degree to the high-temperature zone, moderate-temperature zone, and low-temperature zone, respectively.

[0039] S33, Intelligent Recognition of Crushing Stage. Based on the temporal characteristics of 50 sets of pressure data within a historical 5-second time window, a Long Short-Term Memory (LSTM) network model is used for crushing stage recognition. The LSTM network contains three gating units: input gate, forget gate, and output gate. The network input is a pressure feature sequence of 50 time steps, including average pressure, pressure fluctuation, and pressure variation coefficient, forming a 50×3 input matrix. The LSTM unit state update process includes six steps: calculating the forget gate, concatenating the previous hidden state with the current input, performing a linear transformation through the forget gate weight matrix, adding a forget gate bias, and then passing it through the sigmoid activation function to obtain the forget gate output; calculating the input gate, concatenating the previous hidden state with the current input, performing a linear transformation through the input gate weight matrix, adding an input gate bias, and then passing it through the sigmoid activation function to obtain the input gate output; calculating the candidate cell state, concatenating the previous hidden state with the current input, performing a linear transformation through the candidate state weight matrix, adding a candidate state bias, and then passing it through the sigmoid activation function to obtain the input gate output; calculating the candidate cell state, concatenating the previous hidden state with the current input, performing a linear transformation through the candidate state weight matrix, and adding a candidate state bias. The hidden state is biased and then the candidate cell state is obtained through the tanh activation function. The cell state is updated by multiplying the forget gate output element-wise with the cell state from the previous time step, and then adding the element-wise product of the input gate output and the candidate cell state to obtain the current cell state. The output gate is calculated by concatenating the previous hidden state with the current input, performing a linear transformation through the output gate weight matrix, adding the output gate bias, and then passing the sigmoid activation function to obtain the output gate output. The hidden state is updated by multiplying the output gate output element-wise with the result of the current cell state after the tanh activation function to obtain the current hidden state. The hidden state of the LSTM network at the last time step is input to the fully connected layer and the softmax activation function, outputting the probability distribution of three rolling stages: the initial rolling stage, the intermediate rolling stage, and the final rolling stage. The initial rolling stage corresponds to 1-2 rolling passes, with large pressure fluctuations and a high coefficient of variation; the intermediate rolling stage corresponds to 3-5 rolling passes, with gradually stabilizing pressure and reduced fluctuations; the final rolling stage corresponds to 6 or more rolling passes, with basically stable pressure and minimal fluctuations. Based on the identification results, a compaction stage category label is output. This label includes three possible values: initial compaction, intermediate compaction, and final compaction. This label reflects the current degree of compaction and provides important prior information for compaction degree prediction.

[0040] S34, Environmental Compensation Parameter Lookup and Feature Vector Correction. Based on the identified subgrade stiffness category labels, temperature state category labels, and compaction stage category labels, a three-dimensional environmental state combination index is formed, with 27 possible combinations. An environmental compensation parameter lookup table is pre-established, obtained by analyzing the statistical patterns of 1000 kilometers of historical construction data. For each environmental state combination, the table stores the corresponding compensation gain coefficient vector and compensation bias vector. The compensation gain coefficient vector contains 36 elements, corresponding to each dimension of the 36-dimensional feature vector; the compensation bias vector also contains 36 elements, corresponding one-to-one with the gain coefficient vector. The compensation parameters are obtained through multiple linear regression fitting, with the regression objective being to make the compensated feature vector consistent with the feature vector distribution under standard environment conditions for medium-type subgrade, suitable temperature zone, and mixed compaction stage.

[0041] When performing multiple linear regression fitting, the feature data under different environmental conditions need to be preprocessed: Statistical analysis is performed on the feature data of 27 combinations of environmental conditions, calculating the mean and standard deviation of the 36-dimensional features for each condition; using the feature distribution of the standard environment (medium-sized roadbed, suitable temperature zone, and mixed-pressure stage) as a benchmark, the feature offsets and scaling factors between the other 26 environmental conditions and the standard environment are calculated; due to the large differences in the numerical ranges of different feature dimensions (e.g., the numerical range of vibration frequency features is 0-50Hz, while the numerical range of pressure features is 0-2MPa), standardization is used to ensure the stability of the regression fitting. The standardization method is as follows: for each feature value, first subtract the overall mean of that feature under all environmental conditions, then divide by the overall standard deviation of that feature.

[0042] Based on the current environmental state combined index, the corresponding compensation gain coefficient and compensation bias are read from the lookup table, and the normalized feature vector is linearly corrected. The correction method is as follows: each element of the normalized feature vector is multiplied element-wise by the corresponding compensation gain coefficient, and the corresponding compensation bias is added to obtain the adaptively compensated feature vector. This eliminates the interference of environmental factors, making the distribution of data from different environmental conditions in the feature space more concentrated and regular, thereby improving the generalization ability and prediction accuracy of the subsequent compaction degree prediction model. The compensated feature vector, along with the timestamp, spatial coordinates, and environmental state label, is encapsulated into a data packet and output to the compaction degree prediction module. The data packet size is approximately 250 bytes.

[0043] S4, based on the adaptively compensated feature vectors, constructs and trains a lightweight spatiotemporal convolutional neural network model, and performs real-time prediction of compaction degree through transfer learning and online optimization to obtain compaction degree data packets; S41, Lightweight Spatiotemporal Convolutional Neural Network Model Construction. Based on model compression technology, a lightweight spatiotemporal convolutional neural network model with only 1.2M parameters is constructed. The model input is a 36-dimensional feature vector. The model structure includes: the first layer is a one-dimensional temporal convolutional layer with 32 kernels, a kernel size of 3, a stride of 1, same padding, and ReLU activation function, which performs shallow temporal feature extraction on the input features; the second layer is a max pooling layer with a pooling window size of 2 to reduce the feature dimension; the third layer is a one-dimensional temporal convolutional layer with a kernel size of... The first layer is a 64-dimensional temporal convolutional layer with a kernel size of 3, a stride of 1, and a ReLU activation function, extracting mid-level temporal features. The fourth layer is a max pooling layer with a pooling window size of 2. The fifth layer is a one-dimensional temporal convolutional layer with 128 kernels, a kernel size of 3, a stride of 1, and a ReLU activation function, extracting high-level temporal features. The sixth layer is a global average pooling layer, compressing the temporal features into a 128-dimensional feature vector. The seventh layer is a multi-head self-attention layer with 4 attention heads, each with a dimension of 32. The self-attention mechanism is calculated as follows: the input features are linearly transformed to obtain a query matrix, a key matrix, and a value matrix. The matrix product of the query matrix and the transpose of the key matrix is ​​calculated, then divided by the square root of the key vector dimension for scaling. The attention weights are normalized using the softmax function, and the attention weights are multiplied by the value matrix to obtain the output. The self-attention mechanism can adaptively learn the correlation between different feature dimensions, increasing the model's attention weight to key features; the eighth layer is a fully connected layer with 64 neurons and the activation function is ReLU; the ninth layer is a Dropout layer with a dropout rate of 0.3 to prevent overfitting; the tenth layer is a physical constraint layer, which is the core innovation of this invention and will be described in detail in S42; the eleventh layer is a fully connected output layer with 1 neuron and the activation function is sigmoid, outputting a compaction degree prediction value with an output value range of 0-1, corresponding to a compaction degree percentage of 0% to 100%.

[0044] S42, Physical Constraint Layer Design and Embedding. Based on the Burgers four-element viscoelastic constitutive model of the asphalt mixture compaction process, physical constraints for the compaction process are established. The Burgers model consists of a Maxwell element and a Kelvin element connected in series. The stress-strain relationship of the Burgers model is described by a differential equation, which establishes the relationship between stress and its time derivative and strain and its first and second time derivatives. This equation includes four model parameters, which are related to the elastic modulus and viscosity coefficient of the mixture. By numerically solving the differential equation, the theoretical relationship between compaction degree and the number of compaction passes, compaction pressure, and mixture temperature is obtained. The theoretical relationship shows that the compaction degree increases according to an exponential saturation law, that is, the compaction degree is equal to the maximum achievable compaction degree multiplied by 1 minus an exponential function with a negative compaction rate coefficient and the number of compaction passes as the exponent. The compaction rate coefficient is positively correlated with the rolling pressure and the mixture temperature; there is an upper limit to the compaction increment of a single rolling pass, which is determined by the yield stress of the mixture; the compaction degree cannot exceed the theoretical maximum compaction degree, which is determined by the aggregate gradation and asphalt content of the mixture, and is generally 99%.

[0045] The physical constraints described above are transformed into neural network layers. The input to the physical constraint layer is the output feature of the Dropout layer. The physical constraint layer obtains the compaction degree prediction value through a fully connected transformation. The fully connected transformation process is as follows: the input feature is multiplied by the weight matrix, a bias is added, and then the sigmoid activation function is used to obtain the compaction degree prediction value between 0 and 1. A physical constraint penalty term is calculated, which consists of two parts: the first part is the upper limit constraint on compaction degree, calculated as follows: if the predicted compaction degree exceeds the maximum compaction degree of 0.99, the penalty term equals the penalty weight coefficient 10 multiplied by the excess; otherwise, the penalty term is 0. The second part is the compaction degree growth constraint, which needs to be calculated in conjunction with the compaction degree prediction sequence within the historical time window. For the compaction degree at the current moment and the compaction degree at the previous moment, the compaction degree increment is calculated, i.e., the compaction degree at the current moment is subtracted from the compaction degree at the previous moment. The constraint increment cannot exceed the theoretical maximum increment, which is determined according to the compaction stage: the maximum increment is 0.03 in the initial compaction stage, 0.02 in the intermediate compaction stage, and 0.01 in the final compaction stage. The penalty term is calculated as follows: if the compaction increment exceeds the theoretical maximum increment, the penalty term equals the penalty weight coefficient 10 multiplied by the excess; otherwise, the penalty term is 0. The total penalty term is the sum of the penalty term for the upper limit constraint on compaction and the penalty term for the compaction growth constraint. This penalty term is added to the loss function during model training. The training loss function is the sum of the mean squared error loss and the physical constraint penalty term. The mean squared error loss is the square of the difference between the predicted compaction value and the true compaction label. By minimizing the total loss function during training, the model output not only fits the training data but also satisfies the physical constraints, thereby improving the model's generalization ability and prediction reliability.

[0046] S43, Model Lightweighting and Acceleration Optimization. Based on knowledge distillation, knowledge from a large teacher model with 10M parameters is transferred to a lightweight student model with 1.2M parameters. The knowledge distillation loss function consists of two parts: the first part is the cross-entropy loss between the student model output and the true label, with a weighting coefficient of 0.3; the second part is the cross-entropy loss between the student model output and the teacher model output, with a weighting coefficient of 0.7. When calculating the second part of the loss, the logits of the teacher and student model outputs need to be softened by dividing by a temperature parameter of 3, and then converted into a probability distribution through the softmax function. Finally, the cross-entropy loss between the two probability distributions is calculated. By optimizing the knowledge distillation loss, the student model learns the output distribution characteristics of the teacher model, maintaining prediction accuracy while reducing the number of parameters. Model pruning is used to evaluate the importance of convolutional kernels in the convolutional layers, calculate the contribution of each convolutional kernel to the output, and remove convolutional kernels with a contribution below a threshold. The pruning contribution threshold is set to 0.1, and the pruning rate is set to 30%. After pruning, the number of model parameters is further reduced to 0.9M. Model quantization technology is employed to quantize model parameters from 32-bit floating-point numbers (FP32) to 8-bit integers (INT8). Symmetric quantization is used. The quantization process involves calculating the quantization step size by dividing the maximum absolute value of the floating-point number by 127; then dividing the floating-point number by the quantization step size and rounding it to the nearest integer to obtain the quantized integer. Quantization reduces the model size and improves inference speed on the NVIDIA Jetson Xavier NX edge computing platform, meeting real-time requirements. The TensorRT inference engine is used to perform graph optimization and layer fusion on the model, further improving inference speed and ensuring that end-to-end latency is kept within a reasonable range.

[0047] S44, Real-time Prediction and Uncertainty Estimation of Compaction Degree. Based on the received adaptively compensated feature vector, a lightweight spatiotemporal convolutional neural network model is input for forward inference calculation. The model's single inference time on the Jetson Xavier NX platform is 12ms, yielding a compaction degree prediction value. This value is a floating-point number between 0 and 1, which, when multiplied by 100, gives the compaction degree in percentage form. To evaluate the reliability of the prediction results, the Dropout Monte Carlo method is used for uncertainty estimation. Specifically, the Dropout layer is kept active during the inference phase, and 20 forward inferences are performed to obtain 20 compaction degree prediction values.

[0048] Because outliers may exist in the 20 predicted values, data preprocessing is required to improve the reliability of uncertainty estimation: outlier detection is performed on the 20 predicted values, and the 3σ criterion is used to remove outliers that deviate from the mean by more than 3 times the standard deviation, i.e., predicted values ​​whose absolute difference from the initial mean is greater than 3 times the initial standard deviation are removed; the mean and standard deviation of the remaining valid predicted values ​​are recalculated, with the mean used as the predicted compaction degree and the standard deviation used as the prediction confidence level to quantify the uncertainty of the model prediction. Based on the normal distribution assumption, a 95% confidence interval is constructed, with the lower limit of the interval being the mean minus 1.96 times the standard deviation and the upper limit being the mean plus 1.96 times the standard deviation. The width of the confidence interval reflects the uncertainty of the prediction; the narrower the confidence interval, the higher the prediction confidence. At the same time, whether the prediction results meet the physical constraints is evaluated, and the physical constraint satisfaction index is calculated: if the predicted compaction degree is less than or equal to the maximum compaction degree and the compaction degree increment is less than or equal to the maximum increment, the physical constraint satisfaction is 1; otherwise, it is 0. The system obtains the compaction degree prediction value and prediction confidence level, and outputs the compaction degree prediction value, prediction confidence level, confidence interval, and physical constraint satisfaction flag. These are then encapsulated together with the timestamp and spatial coordinates into a compaction degree data packet. The data packet is output to the visualization and decision support module at a frequency of 100Hz, with each data packet being approximately 100 bytes in size.

[0049] S45, Basic Model Pre-training and Knowledge Representation Learning. Based on construction data from 50 completed historical highway projects, the total data volume is 5000 kilometers and approximately 50 million labeled samples. The labeled samples include sensor feature data and corresponding core sampling compaction ground truth values. The data covers different regions, seasons, mix types, and construction techniques, demonstrating broad representativeness. The spatiotemporal convolutional neural network model architecture described in S41-S42 is used for pre-training on the 50 million samples. Training employs the Adam optimizer with an initial learning rate of 0.001, a batch size of 256, and 100 training epochs. A learning rate decay strategy is used, reducing the learning rate to 0.5 times every 20 epochs. The training loss function consists of three parts: mean squared error loss, physical constraint penalty term, and L2 regularization term. The L2 regularization term has a regularization coefficient of 0.0001, used to prevent overfitting. An early stopping strategy is employed during training: training stops when the loss on the validation set does not decrease for 10 consecutive epochs, achieving good predictive performance on the validation set. The pre-trained base model learns general knowledge representations of the asphalt mixture compaction process. The first three convolutional layers extract low-level temporal features essential to the compaction process, which are universal across projects. Subsequent layers extract high-level features specific to each project, requiring fine-tuning for new projects. The pre-trained base model parameters are saved as the initialization model for new project deployments.

[0050] S46, rapid fine-tuning and adaptation through transfer learning. When the system is deployed in a new engineering project, 30-50 sets of on-site core sampling calibration data are collected. Core sampling is conducted according to JTG E20-2011 Test Procedures for Asphalt and Asphalt Mixtures in Highway Engineering. The sampling locations are jointly determined by the construction and supervision parties, and the sampling points are evenly distributed throughout the construction area. Core samples are drilled at each sampling point, and the bulk density of the core samples is measured to calculate the true compaction value. The calculation method for the true compaction value is as follows: divide the bulk density of the core sample by the maximum theoretical density determined by the standard compaction test. For each core sampling point, the 36-dimensional multi-scale comprehensive feature vector obtained in step 2 and the environmentally adaptively compensated feature vector obtained in step 3 corresponding to that spatial location are retrieved from the system database. These preprocessed and environmentally compensated feature data are paired with the true compaction value to form 30-50 sets of calibration datasets for the new project. A pre-trained base model is loaded, and the parameters of the first three convolutional layers are frozen. Fine-tuning is then performed only on the parameters of the fourth convolutional layer, the fifth global average pooling layer, the sixth self-attention layer, the seventh fully connected layer, the eighth Dropout layer, the ninth physical constraint layer, and the tenth output layer. Freezing the bottom-level parameters preserves general knowledge representations and avoids overfitting under small sample conditions. Fine-tuning uses the Adam optimizer with a learning rate of 0.0001, a batch size of 8, and 100 training epochs. Data augmentation is employed, adding Gaussian noise to the feature vectors with a standard deviation of 0.05. Data augmentation increases the training sample size by 5 times, alleviating the small sample problem. K-fold cross-validation is used, with K set to 5. 30-50 samples are randomly divided into 5 parts. Four parts are selected as the training set and one part as the validation set in each round of training and validation. The average performance of the five validation sets is used as the model performance evaluation. The fine-tuned model showed improvements in mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) on the new project validation set, meeting engineering accuracy requirements. The fine-tuned model parameters were saved, deployed to an edge computing device, and the original model was replaced, completing model adaptation.

[0051] S47, Online Incremental Learning Mechanism Design. To enable the model to continuously learn and optimize from construction process data, an online incremental learning mechanism is designed. Specifically, during system operation, each set of compaction degree prediction results and the corresponding feature data obtained from steps 1-3 are temporarily stored in the edge database. The temporarily stored data includes the feature vector after environmental adaptive compensation output in step 3, the predicted compaction degree and prediction confidence, timestamp, and spatial coordinates. Following specifications, the construction team conducts core sampling every 500m of construction. The sampling results are entered into the system. Based on the spatial coordinates of the sampling location, the system retrieves the corresponding prediction data from the database, pairs the predicted compaction degree with the measured compaction degree, and forms a delayed verification sample. This delayed verification sample is used to evaluate the model's actual prediction performance. When the accumulated number of deferred validation samples reaches 500 sets, an online incremental learning round is automatically triggered. Incremental learning employs an experience replay strategy, randomly selecting 20% ​​of historical samples from the historical database and mixing them with the 500 new samples to form an incremental training dataset. Mixing historical samples aims to prevent catastrophic forgetting, i.e., avoiding the model forgetting previously learned knowledge when learning new data. Incremental learning still uses a fine-tuning strategy, freezing the parameters of the first three layers of the model and updating the parameters of subsequent layers. The learning rate is set to 0.00005, the batch size to 32, and the number of training epochs to 20. After incremental learning is complete, the performance of the updated model is evaluated on an independent test set. If the performance metrics are all better than the unupdated model, the new model parameters are saved and deployed; otherwise, the original model parameters are retained. Through online incremental learning, the model can gradually adapt to the specific characteristics of the current project and changes in construction conditions, achieving continuous improvement in model performance.

[0052] S5 performs three-dimensional visualization and intelligent analysis on compaction data packets to obtain multi-level quality distribution maps and intelligent construction adjustment schemes. S51, Digital Twin Construction Scene Construction. A digital twin virtual environment for the construction site is built using the Unity3D real-time 3D rendering engine. This virtual environment includes virtual objects such as a 3D terrain model, road geometry model, road roller equipment model, and construction personnel model. The 3D terrain model is generated by importing the Digital Elevation Model (DEM) of the construction site. The DEM data originates from pre-construction surveying results, with a grid resolution of 0.5m × 0.5m. A Triangulated Irregular Network (TIN) algorithm is used for 3D surface reconstruction. The road geometry model is generated based on the horizontal, vertical, and cross-sectional design parameters from the design drawings. The road centerline is fitted using a 3D Bezier curve, and parameters such as road width, cross slope, and superelevation dynamically change along the centerline. The road roller equipment model is pre-made using 3D modeling software. The model includes components such as the machine body, vibrating drum, and cab. The model accuracy is LOD2, and the number of facets is controlled within 5000 to ensure a high real-time rendering frame rate. The position and attitude of the road roller in the virtual environment are driven by real-time received BeiDou RTK positioning data. The positioning data undergoes coordinate transformation, converting from the WGS84 geodetic coordinate system to the local coordinate system of the virtual scene. The coordinate transformation process is as follows: subtracting the scene origin coordinates from the WGS84 coordinates to obtain the relative coordinates, and then rotating the coordinate system using a rotation matrix to obtain the coordinates in the local coordinate system. The road roller's attitude is calculated from data from a three-axis accelerometer. A complementary filtering algorithm is used to fuse acceleration and gyroscope data to estimate the pitch and roll angles.

[0053] The complementary filtering algorithm is implemented as follows: Attitude angles in the gravity direction are calculated using accelerometer data. The pitch angle is obtained by dividing the x-axis acceleration by the square root of the sum of the squares of the y-axis and z-axis accelerations, and then taking the arctangent. The roll angle is obtained by dividing the y-axis acceleration by the z-axis acceleration, and then taking the arctangent. Attitude angle integration is performed using gyroscope data. The current pitch angle equals the previous pitch angle plus the y-axis angular velocity multiplied by the sampling interval; the current roll angle equals the previous roll angle plus the x-axis angular velocity multiplied by the sampling interval. The two estimation results are fused using a complementary filter. The pitch angle equals the filter coefficient 0.98 multiplied by the gyroscope integration result plus 0.02 multiplied by the accelerometer calculation result; the roll angle equals the filter coefficient 0.98 multiplied by the gyroscope integration result plus 0.02 multiplied by the accelerometer calculation result. This method combines the short-term accuracy of the gyroscope with the long-term stability of the accelerometer. The virtual scene is rendered in real-time at 60 FPS, achieving real-time synchronous mapping between the physical construction process and the digital space.

[0054] S52, Visualization of 3D Heatmap of Compaction Degree. Based on the real-time received predicted compaction degree values ​​and their corresponding spatial coordinates, a continuous distribution field of compaction degree across the entire road section is generated using a spatial interpolation algorithm. The spatial interpolation employs the Kriging method, an optimal linear unbiased estimation method suitable for interpolating spatially correlated data. The Kriging interpolation method represents the compaction degree of the point to be estimated as a weighted linear combination of the compaction degrees of known points, with the weighting coefficients determined by minimizing the estimation variance. To determine the weighting coefficients, a semi-variogram model is needed to describe spatial correlation. The semi-variogram is defined as: calculating the square of the difference in compaction degree between two points at a distance h, taking the expected value, and finally multiplying by 0.5. This function describes the variance of the difference in compaction degree between two points at a distance h. An exponential semi-variogram model is used, which includes three parameters: nugget effect, sill value, and range. The model expression is: the semi-variogram value equals the nugget effect plus the sill value multiplied by 1 minus an exponential function with the negative distance divided by the range as the exponent. Model parameters were determined by fitting measured data. Kriging interpolation was used to interpolate discrete compaction measurement points into a 0.5m × 0.5m gridded compaction field, with interpolation accuracy meeting engineering requirements. The gridded compaction field was mapped onto the surface of the 3D road model and visualized using a heatmap color-coding method. The color mapping followed a piecewise linear mapping rule: compaction less than 94% was displayed in red, 94% to 95% in orange, 95% to 96% in yellow, 96% to 97% in light green, and above 97% in dark green. Heatmap rendering was implemented using vertex shaders, with the color value of each vertex calculated in parallel on the GPU, resulting in high rendering efficiency and good real-time performance. The heatmap supports 3D view rotation, scaling, and translation, allowing users to observe the compaction mass distribution from different angles and obtain an intuitive visualization of the 3D spatial distribution of compaction.

[0055] S53, Compaction Quality Statistical Analysis and Anomaly Area Identification. Based on the compaction degree data of the entire road section, statistical analysis is performed to calculate the average, standard deviation, maximum, minimum, and pass rate of compaction degree. The pass rate is defined as the proportion of measuring points with a compaction degree greater than or equal to the design requirement value, which is generally 96%. The pass rate is calculated as follows: count the number of measuring points with a compaction degree greater than or equal to 96%, divide by the total number of measuring points, and then multiply by 100% to obtain the pass rate in percentage form; a compaction degree frequency distribution histogram is plotted to statistically analyze the distribution of the number of measuring points in different compaction degree intervals. The horizontal axis of the histogram represents the compaction degree interval with a width of 0.5%, and the vertical axis represents the frequency. A compaction degree cumulative distribution curve (CDF) is plotted, with the horizontal axis representing compaction degree and the vertical axis representing cumulative probability. The CDF curve can be used to evaluate the overall level of compaction quality. Areas with compaction degrees lower than design requirements are automatically identified and marked. A connected component analysis algorithm is used to mark grid cells with compaction degrees less than 96% as anomalous cells. Connectivity analysis is performed on these anomalous cells to extract connected anomalous regions, and the area, perimeter, centroid coordinates, and minimum compaction degree of each anomalous region are calculated. In the 3D visualization interface, anomalous regions are marked with a red semi-transparent overlay, and an anomalous marker icon is placed at the centroid. Clicking the icon displays detailed information about the anomalous region, including its area, minimum compaction degree, and recommended remedial measures. A compaction quality analysis report is generated, including: construction date, construction mileage, average compaction degree, standard deviation, pass rate, number and distribution of anomalous regions, and construction efficiency assessment. The report is exported in PDF format for archiving and submission by construction management personnel.

[0056] S54, Intelligent Construction Adjustment Scheme Generation. Based on real-time compaction degree prediction, current number of compaction passes, mixture temperature, and roller travel speed, a fuzzy control algorithm is used to generate construction parameter adjustment suggestions.

[0057] Since the input variables have different dimensions and numerical ranges, data preprocessing is required. To ensure comparability of different input variables in fuzzy inference, a linear transformation is used to map each variable to a unified universe of discourse from 0 to 10. The transformation method is as follows: subtract the minimum value of the variable from its value, divide by the difference between the maximum and minimum values ​​of the variable, and then multiply by 10.

[0058] The fuzzy control system takes compaction deviation as input, calculated as follows: subtract the current predicted compaction value from 96% of the target compaction value. The output is the speed adjustment. The fuzzy control rule base includes nine rules: if the compaction deviation is negative and the temperature is high, the speed adjustment is positive and large; if the compaction deviation is negative and the temperature is medium, the speed adjustment is positive and medium; if the compaction deviation is negative and the temperature is low, the speed adjustment is positive and small; if the compaction deviation is negative and the temperature is high, the speed adjustment is positive and medium; if the compaction deviation is negative and the temperature is medium, the speed adjustment is positive and small; if the compaction deviation is negative and the temperature is low, the speed adjustment is zero; if the compaction deviation is zero, the speed adjustment is zero; if the compaction deviation is positive and small, the speed adjustment is negative and small; if the compaction deviation is positive and large, the speed adjustment is negative and medium. The fuzzy set membership function adopts the triangular membership function. The fuzzy set for compaction deviation is divided into five categories: negative large, negative small, zero, positive small, and positive large. The fuzzy set for temperature is divided into three categories: low, medium, and high. The fuzzy set for speed adjustment is divided into six categories: negative medium, negative small, zero, positive small, positive medium, and positive large. Fuzzy inference uses the Mamdani inference method for fuzzification. It calculates the membership degree of the input variables to each fuzzy set, performs rule matching, calculates the activation degree of each rule (the minimum value of the input membership degree), performs fuzzy inference, and performs a minimum operation between the activation degree and the output fuzzy set. Finally, it performs defuzzification and uses the centroid method to calculate the precise output value. The centroid method calculation method is as follows: multiply the membership degree of each output fuzzy set by the corresponding universe value, sum them, and divide by the sum of all membership degrees. After obtaining the speed adjustment suggestion, it is determined whether the current speed and the suggested speed are within the allowable range. The allowable speed range for the road roller is 2 to 6 km / h. If it exceeds the range, it is limited to the boundary value. For areas with a compaction degree below 94%, if the current number of compaction passes is less than 6 and the mixture temperature is greater than 110℃, a suggestion to increase the number of compaction passes is generated, with an increase of 2 passes recommended. If the number of compaction passes is greater than or equal to 6 or the temperature is less than 110℃, it is marked as a key area for re-inspection, and it is recommended to conduct core sampling for re-inspection after construction is completed. The construction adjustment suggestions are pushed to the on-board display screen in the roller cab in both text and voice formats. The display screen shows the current compaction degree, target compaction degree, speed adjustment suggestions, and compaction pass suggestions in real time. The voice broadcast uses text-to-speech (TTS) technology to remind the driver to adjust parameters and obtain an intelligent construction adjustment plan.

[0059] S55 features a multi-level visualization interface design. Different visualization interfaces are designed for different user roles. For construction operators, a vehicle-mounted display screen is designed with a simple layout. The main display area shows the current compaction degree in numerical form with large font for easy identification. The secondary display area provides speed adjustment suggestions using a combination of arrows and text to indicate whether to accelerate or decelerate. At the bottom of the interface is a compaction pass counter, displaying the cumulative number of compaction passes for the current area in real time. When the compaction degree falls below the alarm threshold (94% in this embodiment), the interface background turns red and an alarm sounds. A tablet computer interface is designed for site management personnel. The layout employs a multi-window design. The main window displays a 3D compaction heatmap, supporting multi-touch gestures for rotation, zoom, and panning. The left window is a compaction quality statistics dashboard, displaying real-time values ​​for average compaction, pass rate, and the number of abnormal areas using a combination of circular progress bars and numbers. The right window is a list of abnormal areas, with each item including its number, location mileage, area, and minimum compaction. Clicking on an item in the list allows you to locate the corresponding area in the main window. The bottom window displays a construction progress curve, with time on the horizontal axis and cumulative construction mileage on the vertical axis. Updated in real time to monitor construction efficiency; a web interface is designed for remote supervisors, using a dashboard layout. The top displays a project overview card showing the project name, construction unit, current construction mileage, and overall pass rate. The middle section features a multi-chart linked area, including a histogram of compaction frequency distribution, a curve of compaction as a function of mileage, a bar chart comparing pass rates for different road sections, and a line graph of construction efficiency trends. Charts support time range filtering and data export. The bottom displays a two-dimensional heat map of compaction across the entire road section, displayed using a Geographic Information System (GIS). Based on the overlay of thermal layers, the system supports map zooming and panning, allowing users to view detailed compaction information at any location. The right side of the interface features a data report generation tool, where users can select report type, time range, and statistical indicators to generate PDF or Excel reports with a single click. The multi-level interface is connected through a unified data interface, ensuring data consistency and real-time performance across different terminals. This results in multi-level quality distribution maps tailored to different user roles, including real-time compaction display maps on vehicle-mounted terminals, 3D compaction heat maps and statistical dashboards on tablet terminals, and compaction distribution maps and charts for the entire road section on the web terminal.

[0060] S6, based on compaction data packets and raw data streams from multi-source heterogeneous sensors, performs secure and reliable transmission and in-depth cloud analysis to obtain quality analysis reports and optimization solutions; S61, Hybrid Network Architecture Design and Network Selection Strategy. Based on the characteristics of the construction site network environment, a hybrid 5G and LoRa network architecture is designed. An intelligent network selection strategy is implemented: for data requiring real-time feedback, such as compaction degree predictions, alarm information, and GPS positioning data, 5G network transmission is prioritized. Data encapsulation uses Message Queuing Telemetry Transport (MQTT), with a QoS level of 1 to ensure messages are delivered at least once. For data with lower real-time requirements, such as raw sensor data, environmental monitoring data, and equipment status data, LoRa network transmission is selected to reduce communication costs and energy consumption. LoRa transmission uses an acknowledgment mode; after sending data, the terminal node waits for an ACK confirmation from the gateway. If no acknowledgment is received, the data is retransmitted, up to a maximum of three times. Edge computing devices monitor the 5G network signal strength RSRP and signal quality SINR in real time. When RSRP is less than -110dBm or SINR is less than 0dB, the 5G network quality is considered poor, and the system automatically switches to the LoRa network to transmit critical data, ensuring data transmission reliability. When both 5G and LoRa networks are unavailable, a local data caching mechanism is activated. The edge computing device is equipped with 128GB eMMC storage, which can cache about 24 hours of data. After the network is restored, the cached data is automatically retransmitted. The retransmission adopts a breakpoint resume mechanism to avoid data duplication.

[0061] S62, Data Encryption and Secure Transmission. To ensure the security and privacy of construction data, the AES256 advanced encryption standard is used to encrypt transmitted data. AES256 uses a 256-bit key, offering extremely high security and strong resistance to brute-force attacks. The encryption process is as follows: During system initialization, the edge computing device and the cloud server exchange keys using the RSA asymmetric encryption algorithm. The edge device generates a random 256-bit AES symmetric key, encrypts the AES key using the cloud server's RSA public key, and sends the encrypted AES key to the server. The server decrypts the AES key using its RSA private key. After the negotiation is completed, all subsequent data transmissions use AES symmetric encryption, resulting in high encryption efficiency. To prevent data tampering during transmission, a hash-based message authentication code (HMAC) is used. The sender calculates the HMAC value of the data. The HMAC calculation process is as follows: the key is XORed with an external padding constant; the result is XORed with the internal padding constant and the key, then hashed with the result of message concatenation; finally, the entire concatenated result is hashed to obtain the HMAC value. The hash function uses SHA256. The HMAC value is appended to the ciphertext and sent together. The receiver recalculates the HMAC value and compares it with the received HMAC value. If they match, the data is intact and has not been tampered with; otherwise, the data is discarded.

[0062] The hash function is specifically implemented using the SHA256 algorithm: The input message is preprocessed, including message padding (adding one '1' and several '0' bits to the end of the message to make the message length modulo 512 equal to 448) and length appending (appending the original message length as a 64-bit binary number to the end of the message). The preprocessed message is divided into 512-bit blocks, and each block undergoes 64 rounds of iterative computation, using different constants and logical functions in each round. A 256-bit hash value is output as the message digest. Encryption and authentication mechanisms ensure the confidentiality, integrity, and authenticity of data transmission.

[0063] S63 is a cloud-based big data storage and management system. The cloud servers utilize cloud computing services provided by Alibaba Cloud or Huawei Cloud, deployed in a data center in East China. The servers are configured with an 8-core CPU, 32GB of RAM, and 2TB of SSD storage, running Ubuntu 20.04 LTS. Data storage employs the MongoDB distributed document database. MongoDB offers advantages such as a flexible document model, high performance, high availability, and easy scalability, making it suitable for storing massive amounts of time-series sensor data. The database architecture is designed, creating multiple Collections, including: RawSensorData for storing raw sensor data, with document structures including timestamps, GPS coordinates, acceleration data, pressure data, temperature data, and displacement data; CompactionData for storing compaction prediction data, with document structures including timestamps, GPS coordinates, predicted compaction degree, confidence level, and environmental status labels; AlarmData for storing alarm data, with document structures including timestamps, GPS coordinates, alarm type, alarm level, and handling status; and ProjectInfo for storing project information, including project name, construction unit, start and end mileage, and design parameters. Indexes are created for timestamp and GPS coordinate fields to accelerate data retrieval. A time-sharing strategy is adopted, storing data shards on different physical nodes by date, with the size of a single shard kept under 100GB to avoid impacting query performance due to excessively large shards. A MongoDB replica set is configured to achieve redundant storage of data across multiple replicas. The replica set contains three nodes: one master node for read and write operations and two slave nodes for data backup. In the event of a master node failure, a new master node is automatically elected to ensure high service availability. An automatic backup strategy is configured to perform a full backup daily at 2:00 AM. Backup data is stored in Alibaba Cloud OSS object storage service, retaining the most recent 30 days of backup data for disaster recovery.

[0064] S64, Cloud-based Big Data Analytics and Quality Assessment. Based on massive amounts of stored construction data, batch processing analysis is performed using the Apache Spark big data analytics framework. The Spark cluster is deployed on a cloud server, consisting of one Master node and three Worker nodes. Each Worker node is configured with four CPU cores and 16GB of memory. Spark job programs are written to read compaction data for a specified time range from a MongoDB database. The data is loaded into Spark memory in DataFrame format, and parallel analysis is performed using Spark's distributed computing capabilities. The analysis tasks include: calculating the statistical indicators of compaction degree for the entire project, including the mean, standard deviation, maximum, minimum, median, and quartiles, using aggregate functions in Spark SQL; calculating the compaction degree qualification rate for different road sections, dividing the road sections by 100m mileage, statistically analyzing the average compaction degree and qualification rate for each section, and identifying weak road sections with a qualification rate below 95%; analyzing the trend of compaction degree over time, statistically analyzing the average compaction degree by hour or day, plotting time-series trend curves, and assessing the stability of construction quality; and analyzing the spatial distribution pattern of compaction degree, using spatial autocorrelation analysis to calculate Moran's I index, specifically: calculating the difference between each observation and the mean, calculating the product of the differences between adjacent observations, multiplying by the corresponding spatial weight, summing over all adjacent pairs, and finally dividing by the sum of squares of all observation differences, then multiplying by the coefficient of the sample size divided by the total weight. A Moran's I index greater than 0 indicates a positive spatial correlation, meaning that adjacent areas have similar compaction degrees; less than 0 indicates a negative spatial correlation; and equal to 0 indicates a random spatial distribution. To analyze the impact of different construction conditions on compaction quality, multiple linear regression or random forest models are used to establish a correlation model between compaction degree and factors such as mixture temperature, number of compaction passes, roller speed, ambient temperature, and humidity, identifying key influencing factors. The analysis results are presented in visual charts, including histograms, box plots, scatter plots, heat maps, and trend curves, generating a quality analysis report that includes project overview, construction progress, quality statistics, pass rate analysis, identification of weak sections, trend analysis, influencing factor analysis, and improvement suggestions.

[0065] S65 features cloud-based global model training and OTA updates. The cloud leverages massive amounts of data accumulated from multiple projects for continuous global model training and optimization. The training data originates from all projects deployed within the system, reaching hundreds of millions of records. This data covers different regions, seasons, feed mix types, and equipment models, exhibiting exceptional diversity and representativeness. The cloud utilizes the TensorFlow deep learning framework, with model training performed on GPU servers configured with NVIDIA V100 x 4 GPUs and 32GB x 4 GPU memory, providing powerful computing capabilities. The training process is as follows: Training samples are extracted from the MongoDB database, including feature vectors and true compaction labels, and divided into training, validation, and test sets in an 8:1:1 ratio; the current global model parameters are loaded as initialization, and a transfer learning strategy is adopted for incremental training on new data; a distributed data parallelism strategy is used for training, with four GPUs simultaneously training different batches of data, and gradients are synchronized using the AllReduce algorithm to improve training efficiency; training hyperparameters are set as follows: batch size of 1024, learning rate of 0.0005, optimizer AdamW, weight decay coefficient of 0.0001, and training epochs of 50; during training, the validation set loss and accuracy are monitored, and an early stopping strategy is adopted, stopping training if the validation set loss does not decrease for five consecutive epochs; after training, the model performance is evaluated on an independent test set. Yes, if the test set performance is better than the current online model, the new model is considered to have improved performance and can be deployed for update. Model updates use OTA (Over-The-Air) technology, which pushes the update to edge computing devices via MQTT. After receiving the model update package, the edge device verifies the digital signature to confirm the source is trustworthy, saves the new model to local storage, and loads the new model the next time the system starts. To ensure the security of the update process, an A / B dual-partition update strategy is adopted. The system maintains two model partitions, A and B. The currently running model is stored in partition A, and the new model is downloaded to partition B. After the update is completed, the startup partition is switched to B. If the model in partition B malfunctions, it is rolled back to partition A to avoid system unavailability due to update failure. Through continuous optimization of the global model in the cloud and OTA updates, the performance of the edge model is continuously improved, and the overall intelligence level of the system is constantly enhanced.

[0066] An IoT-based system for detecting the compaction degree of asphalt mixtures, such as Figure 2 As shown, an IoT-based method for detecting the compaction degree of asphalt mixtures, as described above, includes: The multi-source data acquisition module is used to acquire the working status signal of the road roller, drive multiple types of sensors to collect data collaboratively, and obtain the raw data stream of the multi-source heterogeneous sensors; The intelligent preprocessing module is used to perform intelligent preprocessing on the raw data streams from multi-source heterogeneous sensors in the edge computing unit to obtain multi-dimensional comprehensive feature vectors; The environmental compensation module is used to intelligently identify and compensate for the roadbed stiffness, mixture temperature state and compaction stage of the multi-dimensional comprehensive feature vector, and obtain the adaptively compensated feature vector. The compaction degree prediction module constructs and trains a lightweight spatiotemporal convolutional neural network model based on the adaptively compensated feature vectors. It then performs real-time compaction degree prediction through transfer learning and online optimization to obtain a compaction degree data package. The visualization analysis module is used to perform three-dimensional visualization processing and intelligent analysis on compaction data packages to obtain multi-level quality distribution maps and intelligent construction adjustment schemes. The cloud-based analytics module is used for secure and reliable transmission and in-depth cloud-based analysis of compaction data packets and raw data streams from multi-source heterogeneous sensors, resulting in quality analysis reports and optimization solutions.

[0067] In one embodiment of the present invention, a specific example is provided: This invention underwent a three-month field application verification in a highway reconstruction and expansion project in a certain province. The construction section was 15 kilometers long, using AC-20C type asphalt mixture with a target compaction degree of ≥96%. Construction took place daily from 22:00 to 6:00 the next day, with an ambient temperature of 12-25℃ and humidity of 60%-80%. The system was deployed on five vibratory rollers, accumulating over 5 million data sets, and 200 core samples were taken for verification.

[0068] Table 1 shows some data examples for the road section from K10+000 to K10+500 on a certain day under construction: Table 1: Example of partial data for the road section from K10+000 to K10+500 on a certain day;

[0069] The data in the table show that the absolute error between the system's predicted compaction degree and the measured compaction degree is controlled within 0.4%, with an average absolute error of 0.3%, and the prediction accuracy meets the engineering requirements. After applying this system, the compaction degree qualification rate of the construction section increased from 89.3% using the traditional method to 97.8%, the standard deviation of compaction degree decreased from 2.8% to 1.5%, the uniformity of the compacted material improved, and the average end-to-end response delay of the system is 185ms, meeting the real-time requirements.

[0070] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for detecting the compaction degree of asphalt mixtures using the Internet of Things (IoT), characterized in that, include: Acquire the operating status signal of the road roller, drive multiple types of sensors to collect data collaboratively, and obtain the raw data stream of multi-source heterogeneous sensors; The raw data streams from multi-source heterogeneous sensors are intelligently preprocessed in the edge computing unit to obtain multi-dimensional comprehensive feature vectors. Intelligent identification and compensation correction are performed on the multi-dimensional comprehensive feature vector to obtain the feature vector after adaptive compensation, based on the subgrade stiffness, mixture temperature state and compaction stage. Based on the adaptively compensated feature vectors, a lightweight spatiotemporal convolutional neural network model is constructed and trained. Through transfer learning and online optimization, the compaction degree is predicted in real time, and a compaction degree data package is obtained. The compaction data package is subjected to three-dimensional visualization and intelligent analysis to obtain multi-level quality distribution maps and intelligent construction adjustment schemes. Based on the compaction data package and the raw data stream from multi-source heterogeneous sensors, secure and reliable transmission and in-depth cloud analysis are performed to obtain quality analysis reports and optimization solutions.

2. The IoT-based method for detecting the compaction degree of asphalt mixtures according to claim 1, characterized in that, The driving of multi-type sensor collaborative acquisition includes: A distributed sensor network topology is adopted, and the triaxial accelerometer is fixed to the inside of the bearing seat of the vibrating wheel of the road roller using a special damping rubber bracket; Piezoelectric dynamic pressure sensors are evenly distributed along the width of the vibrating wheel to form a pressure distribution measurement array, which is installed at the interface between the vibrating wheel and the road surface. A non-contact infrared temperature sensor array is used to achieve two-dimensional distribution measurement of the temperature field on the surface of the mixture; The BeiDou-3 high-precision RTK positioning module is used to output PPS second pulse signals as a unified time reference.

3. The IoT-based method for detecting the compaction degree of asphalt mixtures according to claim 1, characterized in that, The acquisition of the multi-dimensional comprehensive feature vector includes: The triaxial acceleration signal is denoised and decomposed into three layers of wavelet. The dominant frequency and high-frequency attenuation coefficient of the Z-axis vertical acceleration signal are extracted to obtain the vibration time-frequency feature vector. The coefficient of variation of the spatial distribution of pressure is calculated for the pressure sensor array, and the time-domain fluctuation characteristics and power spectrum peak characteristics of pressure are calculated to obtain the pressure distribution and dynamic feature vector. A high-resolution temperature field is reconstructed using a bicubic interpolation algorithm on the temperature data, and the statistical features, spatial gradient features, and temporal dynamic features of the temperature field are extracted to obtain the temperature field feature vector. The elevation difference before and after compaction is calculated from the data of the front and rear displacement sensors to obtain the deformation feature vector; The vibration time-frequency feature vector, pressure distribution and dynamic feature vector, temperature field feature vector, and deformation feature vector are concatenated and Z-score normalized to obtain a multi-dimensional comprehensive feature vector.

4. The IoT-based method for detecting the compaction degree of asphalt mixtures according to claim 1, characterized in that, The intelligent identification and compensation correction of subgrade stiffness, mixture temperature state, and compaction stage of the multi-dimensional comprehensive feature vector includes: Three features—vibration high-frequency attenuation coefficient, deformation recovery rate, and cumulative deformation—are extracted. A pre-trained radial basis function kernel is used to identify the roadbed stiffness state and classify the roadbed stiffness. Four features were extracted from the temperature field feature vector: average temperature, standard deviation, proportion of high-temperature zone, and proportion of low-temperature zone. Fuzzy C-means clustering algorithm was used to identify the temperature state of the mixture. A long short-term memory network model was used to identify the crushing stage. Based on the recognition results, a three-dimensional environmental state combination index is formed. The corresponding compensation gain coefficient and compensation bias are read from the pre-established environmental compensation parameter lookup table, and the normalized feature vector is linearly corrected.

5. The IoT-based method for detecting the compaction degree of asphalt mixtures according to claim 1, characterized in that, The construction and training of a lightweight spatiotemporal convolutional neural network model, and the real-time prediction of compaction through transfer learning and online optimization, include: Construct a lightweight spatiotemporal convolutional neural network model; The physical constraint layer establishes the physical constraint conditions of the compaction process based on the Burgers four-element viscoelastic constitutive model of the asphalt mixture compaction process, and transforms the upper limit constraint of compaction degree and the compaction degree growth constraint into the penalty term of the neural network layer; The knowledge from the large teacher model is transferred to the lightweight student model, and uncertainty is estimated.

6. The IoT-based method for detecting the compaction degree of asphalt mixtures according to claim 1, characterized in that, The three-dimensional visualization and intelligent analysis of the compaction data package includes: A digital twin virtual environment for the construction site is built based on the Unity3D real-time 3D rendering engine, including a 3D terrain model, a road geometry model, and a road roller equipment model; The Kriging space interpolation algorithm is used to generate a continuous distribution field of compaction degree for the entire road section. The gridded compaction degree field is mapped onto the surface of the three-dimensional road model and visualized using a heat map color coding method. Automatic region identification and labeling are performed, and connected component analysis algorithms are used to extract connected outlier regions; Based on real-time compaction degree prediction, current number of compaction passes, mixture temperature, and roller speed, a fuzzy control algorithm is used to generate suggestions for adjusting construction parameters. Design differentiated visual interfaces for different user roles.

7. The IoT-based method for detecting the compaction degree of asphalt mixtures according to claim 2, characterized in that, The multi-sensor collaborative data acquisition also includes: The hardware-level synchronous controller based on FPGA receives the PPS second pulse signal as the master clock reference, multiplies the PPS signal to 100 MHz as the system master clock, and adopts multi-channel synchronous trigger control logic to control the delay time after each acquisition unit receives the trigger signal. The FPGA integrates a hardware timestamp generation module to automatically append a UTC timestamp to each set of sampled data. The FPGA reads the current RTK positioning coordinates and binds the spatial coordinates to the sensor data.

8. The IoT-based method for detecting the compaction degree of asphalt mixtures according to claim 5, characterized in that, The real-time prediction of compaction degree also includes: Based on the received adaptively compensated feature vector, input it into a lightweight spatiotemporal convolutional neural network model for forward inference calculation; Outlier detection is performed, and outliers that deviate from the mean by more than three times the standard deviation are removed. The mean and standard deviation of the remaining valid predictions are recalculated. The mean is used as the final compaction degree prediction value, and the standard deviation is used as the prediction confidence level. Based on the assumption of normal distribution, construct confidence intervals; Calculate the physical constraint satisfaction index. If the predicted compaction degree is less than or equal to the maximum compaction degree and the compaction degree increment is less than or equal to the maximum increment, the physical constraint satisfaction is one; otherwise, it is zero.

9. The Internet of Things (IoT) method for detecting the compaction degree of asphalt mixtures according to claim 6, characterized in that, The automatic region identification and labeling includes: Statistical analysis was conducted based on the compaction data of the entire road section to calculate the average, standard deviation, maximum, minimum, and pass rate of compaction. The pass rate was defined as the proportion of measuring points with compaction greater than or equal to the pass threshold. Draw a histogram of compaction degree frequency distribution and statistically analyze the distribution of the number of measuring points in different compaction degree intervals; A connected component analysis algorithm is used to mark grid cells with a compaction degree less than the qualified threshold as abnormal cells. Connectivity analysis is performed on the abnormal cells to extract connected abnormal regions. The area, perimeter, centroid coordinates, and minimum compaction degree of each abnormal region are calculated. In the 3D visualization interface, abnormal areas are marked with a red semi-transparent overlay, and an abnormal marker icon is placed at the centroid, generating a compaction quality analysis report.

10. An Internet of Things (IoT) system for detecting the compaction degree of asphalt mixtures, characterized in that, A method for detecting the compaction degree of asphalt mixtures using the Internet of Things (IoT) according to any one of claims 1-9 includes: The multi-source data acquisition module is used to acquire the working status signal of the road roller, drive multiple types of sensors to collect data collaboratively, and obtain the raw data stream of the multi-source heterogeneous sensors; The intelligent preprocessing module is used to perform intelligent preprocessing on the raw data streams from multi-source heterogeneous sensors in the edge computing unit to obtain multi-dimensional comprehensive feature vectors; The environmental compensation module is used to intelligently identify and compensate for the roadbed stiffness, mixture temperature state and compaction stage of the multi-dimensional comprehensive feature vector, and obtain the adaptively compensated feature vector. The compaction degree prediction module constructs and trains a lightweight spatiotemporal convolutional neural network model based on the adaptively compensated feature vectors. It then performs real-time compaction degree prediction through transfer learning and online optimization to obtain a compaction degree data package. The visualization analysis module is used to perform three-dimensional visualization processing and intelligent analysis on compaction data packages to obtain multi-level quality distribution maps and intelligent construction adjustment schemes. The cloud-based analytics module is used for secure and reliable transmission and in-depth cloud-based analysis of compaction data packets and raw data streams from multi-source heterogeneous sensors, resulting in quality analysis reports and optimization solutions.

Images