A highway pavement rut depth nondestructive testing and data analysis device

Abstract

The application discloses a kind of highway pavement rut depth nondestructive testing and data analysis device, comprising: basic detection module: using multiple laser array non-contact measurement road surface vertical distance, through cubic spline interpolation to generate deformation track and solve rut depth, output original data;Dynamic environment adaptive acquisition module: integrated illumination and dust sensor, dynamically adjust laser power, pulse interval and beam diameter, ensure that signal-to-noise ratio is greater than 30dB, output optimized measurement value;Mobile multi-dimensional scanning module: carry double-shaft platform and SLAM algorithm to realize autonomous navigation, receive optimized data and output rut depth, transverse and longitudinal slope and other multi-dimensional topographic data;Field data grading processing module: through dynamic threshold rejection, moving average filtering and spatial verification to realize data cleaning;Data analysis and prediction module: based on LSTM model, rut development trend is predicted in combination with historical data, to support preventive maintenance decision.

Description

Technical Field

[0001] This invention relates to the field of highway maintenance technology, and in particular to a non-destructive testing and data analysis device for rut depth of highway pavement. Background Technology

[0002] Rutting data is one of the core indicators for assessing the technical condition of expressways and first-class highways. Its depth changes directly reflect the health status of the pavement, profoundly impacting the scientific judgment of maintenance needs and the efficient allocation of maintenance funds. In the highway maintenance and management system, accurate rutting data detection is not only crucial for the precision of daily operations and maintenance, but also plays an irreplaceable role in preventative maintenance decisions, ensuring traffic safety, extending pavement life, and optimizing life-cycle costs. By accurately reflecting rutting trends, it provides key support for the scientific formulation of maintenance strategies, the rational allocation of maintenance resources, the reduction of accident risks, and the reduction of long-term maintenance costs, serving as a vital cornerstone for building a safe, economical, and sustainable highway operation system.

[0003] However, the inefficiency of traditional rut detection and analysis is becoming increasingly prominent. For example, at the exit section of a highway tunnel in a certain province, the laser signal of traditional rut detection equipment is severely attenuated due to strong sunlight reflection and excessive construction dust concentration. The detection personnel need to use a double process of ruler measurement and leveling. First, the rut position is manually located, and then the cross-sectional elevation is measured by setting up a leveling rod with a level. The measurement needs to be repeated 3 times per kilometer to offset the dust interference, which takes up to 3 hours. Moreover, the error of manual reading causes the rut depth measurement deviation to reach 4mm, and the cross slope data is missing, making it impossible to assess the impact of the road surface's cross and longitudinal slopes on driving safety. Therefore, a non-destructive detection and data analysis device for highway pavement rut depth is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a non-destructive testing and data analysis device for rut depth on highway surfaces.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A non-destructive testing and data analysis device for rut depth of highway pavement, comprising:

[0007] Basic detection module: It adopts a multi-laser ranging array to measure the vertical distance of the road surface in a non-contact manner. The laser unit is fixed by the crossbeam steel body and a pulsed laser emitter is integrated. The vertical distance from each point to the crossbeam is calculated by measuring the time difference of reflected light, generating the road surface deformation trajectory curve and directly solving the basic data of rut depth. The raw distance data is output to the dynamic environment adaptive acquisition module.

[0008] Dynamic environment adaptive acquisition module: integrates light intensity sensor and dust concentration monitor to sense environmental parameters in real time, dynamically adjust laser emission power and receiving sensitivity, shorten laser pulse interval and increase laser beam diameter, perform stable measurement with signal-to-noise ratio greater than 30dB, receive raw data and environmental parameters from basic detection module, and output effective measurement values ​​optimized for environmental adaptation to mobile multi-dimensional scanning module.

[0009] Mobile multi-dimensional scanning module: Equipped with a dual-axis mobile platform and an adjustable-angle laser ranging unit array, combined with a force control sensor to sense the road reaction force, it performs autonomous navigation through the SLAM algorithm, receives optimized data from the adaptive acquisition module, and outputs multi-dimensional topographic data to the on-site data classification and processing module.

[0010] On-site data classification and processing module: performs hardware-level triple real-time processing: dynamic threshold removal, moving average filtering, and spatial correlation verification. It reduces the amount of effective data transmission through logic circuits, improves the backend processing speed, receives multi-dimensional data from the scanning module, and outputs cleaned effective data to the data analysis and prediction module.

[0011] Data Analysis and Prediction Module: Integrates an LSTM neural network model, combines historical data, and performs trend prediction on the effective data after on-site processing. It supports preventive maintenance decisions by predicting the development trend of ruts. It receives cleaning data from the graded processing module and outputs prediction results to the user terminal.

[0012] The above technical solution further includes:

[0013] Furthermore, the non-contact measurement of the vertical distance to the road surface using a multi-laser ranging array includes the following steps:

[0014] 13-25 laser ranging units are arranged at equal intervals along the transverse steel beam (spacing accuracy ≤0.5mm). Each laser unit integrates a pulsed laser emitter and a high-sensitivity photoelectric receiver. The steel beam is fixed to the front of the detection vehicle by a rigid bracket to ensure that the laser array maintains a constant vertical distance from the road surface (usually 300-500mm), forming a non-contact measurement infrastructure.

[0015] Each laser unit emits nanosecond-level pulsed lasers (wavelength 905nm, pulse width ≤20ns) sequentially according to a preset time sequence. The laser beam is projected vertically onto the road surface to form a light spot (default diameter 2mm). The reflected light from the road surface is captured by a photodetector, and the optical signal is converted into an electrical signal by an avalanche photodiode (APD). The time difference Δt between laser emission and reception is recorded.

[0016] Based on the speed of light c (299792458m / s) and the time difference Δt, the vertical distance d from each laser point to the crossbeam is calculated using the formula d = c × Δt / 2. When the ambient temperature deviates from 20℃, the speed of light is calibrated (the speed of light is adjusted by approximately ±0.01ppm for every 1℃ change in temperature), and multiple measurements are taken to average the values ​​(usually 5 times) to reduce the influence of random noise.

[0017] The vertical distance data of 13-25 laser points are sorted by their horizontal positions. A continuous road deformation trajectory curve is generated by a cubic spline interpolation algorithm. By comparing the peak and valley values ​​of the trajectory curve (i.e., the height difference between the deepest rut point and the reference points on both sides), the rut depth value is directly calculated. For example, when the trajectory curve shows that there is a depression with a depth > 5mm in a certain area, the system automatically marks the area as a suspected rut area and triggers the subsequent dynamic environment adaptive acquisition module to perform environmental adaptation optimization.

[0018] The raw distance data of each laser point (including timestamp, temperature-compensated light speed value, and average value of multiple measurements) is bound to environmental parameters (illuminance, dust concentration) and output to the dynamic environment adaptive acquisition module through a standardized data interface.

[0019] Furthermore, the process of calculating the vertical distance from each point to the crossbeam by measuring the time difference of reflected light, generating the road surface deformation trajectory curve, and directly solving the basic data of rut depth includes the following steps:

[0020] The lateral positions of 13-25 laser points (e.g., x0=0mm, x1=200mm, x2=400mm, ..., x) are determined. n =3200mm) and the corresponding vertical distances (y0=d0, y1=d1,…, y n =d n Arrange them in order to form a discrete set of points;

[0021] At two adjacent laser points ( arrive Between ), fit the curve segment using a cubic polynomial, in the form:

[0022]

[0023] in, , , , The coefficients to be determined must satisfy the following conditions: ( ) = (The curve passes precisely through each measurement point), and the first derivative (slope) and second derivative (curvature) of adjacent curve segments are continuous at the connection point, while the second derivative is zero at the beginning and end endpoints (to avoid excessive curvature at the endpoints).

[0024] After determining all coefficients by solving the system of linear equations, the curve segments are spliced ​​together to form a continuous and smooth road surface deformation trajectory curve, which can accurately reflect subtle deformations such as ruts and waves.

[0025] Locate the feature points of the rut. The valley point is the lowest point of the curve (the first derivative is zero and the second derivative is positive), corresponding to the deepest position of the rut. The reference point is a point on a relatively flat area on both sides of the rut (usually the highest point within the rut width range). Calculate the depth value:

[0026] The depth of the rut is calculated as follows: [(height of the left reference point - height of the valley point) + (height of the right reference point - height of the valley point)] / 2. For example, if the height of the left reference point is 5.2mm, the right reference point is 5.0mm, and the valley point is 3.0mm, then the depth is calculated as [(5.2-3.0) + (5.0-3.0)] ÷ 2 = 2.1mm.

[0027] Furthermore, the dynamic adjustment of laser emission power and receiving sensitivity includes the following steps:

[0028] The basic detection module outputs a real-time vertical distance sequence from a multi-laser ranging array (distance values ​​of 13-25 laser points, including timestamps and initial measurements). Environmental parameters are measured using an illuminance sensor (0-20000 lux, 1 lux resolution) and a dust concentration monitor (0-100 mg / m³) integrated into the laser device. 3 Resolution 0.1 mg / m 3 Real-time data acquisition via temperature sensors;

[0029] Based on a preset environment-laser parameter mapping model, the FPGA control module analyzes environmental parameters in real time and triggers adjustment strategies.

[0030] Strong light environment adaptation (illuminance > 10000 lux): shorten the laser pulse interval to 100-200 ns (reduce photon scattering), increase the receiving sensitivity to 150% of the original value (suppress background noise), or increase the transmission power to 15-20 mW (enhance signal strength).

[0031] Dust environment suitability (dust concentration > 50 mg / m³) 3 ): Expand the laser beam diameter to 5mm (to improve penetration), reduce the emission power to 8-10mW (to avoid signal oversaturation), and adjust the receiver bandwidth to 5-10MHz (to suppress high-frequency noise);

[0032] Temperature compensation (temperature < -20℃ or > 60℃): The laser wavelength is maintained at 905nm ± 1nm by the built-in TEC temperature control module of the laser emitter, and the power compensation algorithm is triggered (power is adjusted ±0.5% for every 1℃ change in temperature) to ensure measurement stability;

[0033] The FPGA sends adjustment commands to the laser transmitter and receiver to adjust the transmission power, pulse interval, beam diameter and receiver sensitivity in real time. After adjustment, the effect is verified by the signal-to-noise ratio (SNR) monitoring module. If SNR < 30dB, a second adjustment is triggered (such as further shortening the pulse interval or increasing the power). If the standard is not met for 3 consecutive times, the system switches to the backup laser unit for remeasurement.

[0034] Dynamic weighted filtering is applied to the original distance data based on environmental parameters: Gaussian filtering (3×3 window) is used to suppress scattering noise in strong light environment, median filtering (5×5 window) is used to eliminate random jump points in dust environment, and the light speed value is corrected by temperature compensation algorithm when the temperature fluctuates (adjusted by ±0.01ppm per ℃).

[0035] The corrected measurements were verified for signal-to-noise ratio (SNR ≥ 30 dB to be valid), and environmental parameter labels were attached (e.g., "illuminance 12000 lux - dust concentration 60 mg / m³"). 3 This generates effective measurement values ​​with environmental characteristics, ensuring data traceability.

[0036] Furthermore, the autonomous navigation using the SLAM algorithm, receiving optimized data from the adaptive acquisition module, and outputting multi-dimensional topographic data includes the following steps:

[0037] Equipped with a dual-axis motion platform (horizontal / vertical movement accuracy of 0.1mm), it integrates a lidar (scanning angle 270°, angular resolution 0.5°), an inertial measurement unit (IMU, triaxial acceleration ±16g, angular velocity ±2000° / s), and a wheeled odometer (encoder accuracy 0.1mm). The SLAM algorithm operates based on a tightly coupled lidar-IMU framework.

[0038] The lidar scans the current road surface environment to generate point cloud data, and the IMU collects acceleration and angular velocity information. The two types of data are fused by extended Kalman filter (EKF) to perform real-time pose estimation of the detection equipment in the road surface coordinate system (position error ≤2cm, attitude angle error ≤0.5°).

[0039] Based on the pose estimation results, the laser point cloud data is stitched into a local 3D map of the road surface. Combining the real-time point cloud data of the road surface with the preset detection area boundary, the optimal scanning path is generated using the A* algorithm. The dual-axis moving platform is then controlled to move autonomously along the path to perform continuous and complete detection at a scanning speed of 20 km / h.

[0040] The preset detection area (e.g., a road section 100m long × 5m wide) is discretized into a two-dimensional grid. Each grid has a side length of 0.5m (to match the movement accuracy of the dual-axis platform). The grid node coordinates are (i,j), where i is the vertical (driving direction) index and j is the horizontal index. The starting point is set as the center of the entrance grid of the detection area (e.g., (0,0)) and the ending point is set as the center of the exit grid (e.g., (200,0)). The path must cover all grids to be scanned (to avoid omissions).

[0041] Given the cost g(n): the actual distance traveled from the starting point to the current node n (such as Euclidean distance or Manhattan distance), with an additional cost of 0.5m for each grid cell moved;

[0042] Heuristic cost h(n): Use Euclidean distance to estimate the remaining distance from n to the endpoint, or use diagonal distance (which better fits the motion characteristics of a dual-axis platform).

[0043] Total cost f(n) = g(n) + h(n): The node with the smallest f(n) is dynamically selected for expansion through a priority queue (such as a min-heap) to ensure that the path is globally optimal;

[0044] The system directly calls the basic rut depth data generated by the basic detection module (the deformation trajectory curve generated by cubic spline interpolation and the result of the valley-reference point height difference calculation, such as the example value of 2.1mm), only marking its spatial coordinates in the SLAM map. Based on the SLAM-generated road surface transverse (X-axis) and longitudinal (Y-axis) point cloud data, the system fits the transverse / longitudinal slope curves respectively using the least squares method, and calculates the slope angle between the reference points on both sides of the rut area (such as transverse slope tanθ1=Δh1 / Δx1, longitudinal slope tanθ2=Δh2 / Δy2):

[0045] Cross slope calculation:

[0046] The lateral height z and the lateral coordinate x satisfy a linear relationship:

[0047]

[0048] Among them, a x The tangent of the transverse slope angle (tanθ1) reflects the steepness of the transverse slope. x The intercept represents the height reference at the lateral zero point;

[0049] By adjusting parameter a x b x To minimize the sum of the squares of the differences between the predicted and actual values ​​for all data points:

[0050] Minimize objective:

[0051] Where Σ represents the summation of all N data points within the rut area;

[0052] Find a for the sum of squared residuals S. x b x The partial derivatives of , and setting the derivatives to zero, are rearranged to obtain:

[0053]

[0054]

[0055] Convert slope to angle using the arctangent function:

[0056]

[0057] Longitudinal slope calculation:

[0058] The vertical height z and the vertical coordinate y satisfy a linear relationship:

[0059] z = aᵧ · y + bᵧ

[0060] Where aᵧ is the tangent of the longitudinal slope angle (tanθ2), reflecting the steepness of the longitudinal slope; bᵧ is the intercept, representing the height reference at the longitudinal zero point;

[0061] Minimize objective:

[0062] Take the partial derivatives with respect to aᵧ and bᵧ respectively:

[0063]

[0064]

[0065] Calculate the longitudinal slope angle:

[0066]

[0067] By integrating rut depth, cross slope, and longitudinal slope data into a SLAM 3D map, a complete road surface model containing geometric shape and spatial location information is formed, supporting visualization and analysis.

[0068] Furthermore, the hardware-level triple real-time processing—dynamic threshold removal, moving average filtering, and spatial correlation verification—includes the following steps:

[0069] Dynamic threshold elimination: The input multi-dimensional topography data (rut depth, cross slope, longitudinal slope) is scanned in real time. Each data point is compared with adjacent points and historical benchmark values ​​to determine whether it exceeds the threshold range.

[0070] Based on the characteristics of road materials (such as the deformation threshold of asphalt concrete pavement is usually ≤ ±8mm) and the detection accuracy requirements (system error ≤ ±2mm), the threshold range is adjusted. For example, the threshold is appropriately widened to ±12mm in areas with dense rutting to avoid false rejection. If a data point exceeds the threshold range, the system marks it as an "abnormal point" and temporarily removes it. At the same time, the timestamp and location information of the point are recorded for subsequent manual review or algorithm optimization.

[0071] Moving average filtering: Applying a moving average filter to a data sequence after removing outliers to suppress high-frequency random noise and smooth data fluctuations;

[0072] Based on the data sampling rate (e.g., 100Hz) and noise characteristics (e.g., abrupt noise caused by dust), the size of the filtering window is adjusted (usually 3-7 data points) to ensure that detailed features of rut deformation are preserved while effectively filtering out noise. A weighted moving average algorithm (e.g., Hanning window) is used to assign different weights to the data points within the window (the center point has the highest weight, and the weights of the edge points decrease). The weighted average is calculated as the output value after filtering to improve the data smoothing effect. Millisecond-level filtering processing is implemented through FPGA hardware logic circuits to ensure that the filtering delay is ≤10ms, which meets the requirements of real-time detection and avoids scanning omissions or data distortion caused by processing delays.

[0073] Spatial correlation verification: Perform spatial correlation verification on the filtered data to ensure the logical consistency of the data in the spatial dimension;

[0074] Check whether the height difference between adjacent data points meets the preset gradient threshold (e.g., height difference between adjacent points ≤ 5mm / m). If it exceeds the threshold, it is marked as a "suspicious point" and a secondary verification is triggered. Based on the 3D road map generated by SLAM, verify whether the data in the current scanning area is consistent with the global topography features (e.g., rut direction, slope change). If there is a significant difference, a local rescan is triggered. The verification result is fed back to the dynamic threshold model and the filter parameter adjustment module, forming a closed-loop mechanism of "processing-verification-optimization" to continuously improve the accuracy and reliability of data processing.

[0075] The three-stage processing steps form a tightly linked technical chain through a progressive logic of anomaly removal, noise suppression, and consistency verification. Dynamic threshold removal serves as the first line of defense, quickly identifying and removing obviously abnormal data points to provide a clean data foundation for subsequent processing. Moving average filtering serves as the second line of defense, further smoothing the data after removing outliers and suppressing the impact of random noise on data quality. Spatial correlation verification serves as the final checkpoint, ensuring that the processed data conforms to the logical characteristics of the actual road surface in the spatial dimension, avoiding data distortion caused by local noise or algorithm errors.

[0076] Furthermore, the method of reducing the amount of effective data transmission and improving backend processing speed through logic circuits includes the following steps:

[0077] Before the data enters the logic circuit, the preprocessing module built into the FPGA performs preliminary screening and feature extraction on the raw data. Based on a preset dynamic threshold (such as the rate of change of rut depth > 0.5 mm / m), it quickly identifies and removes data from flat areas without significant deformation, retaining only the laser point data of suspected rut areas (such as depth change > 2 mm), directly reducing the amount of raw data by more than 60%. Feature labels (such as "rut edge point" and "reference point") are added to the retained data, and key data points are marked before transmission through bit operations (such as setting data packet header flag bits), which facilitates the back-end processing module to quickly locate core information.

[0078] FPGAs perform efficient data compression using dedicated logic circuits, specifically employing a combined strategy of differential coding and entropy compression.

[0079] Differential encoding: Perform differential calculations on the vertical distance data of consecutive laser points (e.g., Δd = - The absolute distance value is converted into a relative change, and the high correlation between adjacent point data is used to reduce the amount of data stored. Huffman coding is used to perform lossless compression on the differential data, and variable-length codewords are assigned according to the frequency of data occurrence (short code for high-frequency data and long code for low-frequency data) to further compress the data volume. Hardware acceleration of the compression algorithm is performed through the parallel computing unit of FPGA to ensure that the compression latency is ≤5ms, which meets the real-time processing requirements.

[0080] Furthermore, the integrated LSTM neural network model, combined with historical data, performs trend prediction on the effective data after on-site processing, including the following steps:

[0081] Build a multi-source heterogeneous data fusion library to integrate the following data:

[0082] Historical data includes monitoring data on rut depth, cross slope, and longitudinal slope over the past 3-5 years (sampling frequency 1Hz), traffic load data (such as average daily traffic volume and proportion of heavy vehicles), environmental data (such as temperature, humidity, rainfall, illuminance, and dust concentration), and pavement material characteristics data (such as asphalt type, thickness, and degree of aging).

[0083] Effective on-site data: Cleaned data output by the on-site data classification and processing module, including real-time rut depth, cross slope, longitudinal slope and multi-dimensional topographic data, as well as corresponding environmental parameters and laser parameter status;

[0084] Based on the fused dataset, construct and train an LSTM neural network model:

[0085] A two-layer LSTM structure (64 hidden units per layer) is used, combined with a Dropout layer (scale 0.2) to prevent overfitting. The output layer uses a fully connected layer to predict the trend of rut depth changes over the next 72 hours. Key features (such as rut ​​depth change rate, cross slope change gradient, and traffic load fluctuation rate) are extracted as inputs, and an attention mechanism is used to enhance the influence of important features on the prediction results. The Adam optimizer (learning rate 0.001) is used with mean squared error (MSE) as the loss function. The model is trained for 50 rounds on the training set (80% of the data) and its performance is evaluated on the validation set (20% of the data) (prediction error ≤10%). Cross-validation and leave-one-out method are used to verify the model's generalization ability, ensuring stable prediction accuracy under different environmental conditions and traffic load scenarios.

[0086] The processed data is input into the trained LSTM model to perform trend prediction.

[0087] Based on historical data from the previous 72 hours, the model predicts the trend of rut depth changes over the next 72 hours and outputs a prediction curve and confidence interval (such as the prediction range at a 95% confidence level). As new data is input in real time, the model dynamically updates the prediction results through a sliding window mechanism to ensure that the prediction is always based on the latest road conditions and environmental conditions. The prediction results are displayed on user terminals (such as in-vehicle displays or cloud platforms) in the form of curves, heat maps, etc., to intuitively present the spatiotemporal evolution trend of rut depth, making it easy for maintenance personnel to understand quickly.

[0088] Based on the prediction results, preventative maintenance decision recommendations are generated:

[0089] When the predicted rut depth exceeds a preset threshold (e.g., 5mm for asphalt pavement, 8mm for concrete pavement), the system automatically triggers a three-level warning (yellow, orange, and red) and recommends maintenance measures (e.g., local repairs, repaving, traffic control). It prioritizes areas requiring maintenance based on factors such as the rate of change of rut depth, traffic flow, and remaining pavement life. Based on the prediction results, the system recommends the optimal maintenance time (e.g., during periods of low traffic flow), maintenance scope (e.g., repair areas accurate to the meter), and maintenance method (e.g., micro-surfacing, thin overlay) to reduce maintenance costs and extend pavement life. After maintenance is performed, it collects maintenance effect data (e.g., changes in rut depth, user feedback) and feeds it back into the LSTM model for parameter updates.

[0090] The present invention has the following beneficial effects:

[0091] In this invention, a dynamic environment adaptive acquisition module integrates an illuminance sensor and a dust concentration monitor to dynamically adjust the laser emission power, receiving sensitivity, and pulse interval, achieving a stable measurement with a signal-to-noise ratio greater than 30dB to solve environmental interference problems. A mobile multi-dimensional scanning module, equipped with an adjustable-angle laser ranging unit array and SLAM algorithm, outputs multi-dimensional topographic data including rut depth and transverse and longitudinal slopes to overcome the limitations of single-dimensional analysis. A field data hierarchical processing module performs dynamic threshold removal, moving average filtering, and spatial correlation verification in real time, reducing the amount of effective data transmitted and improving processing speed. Finally, a data analysis and prediction module integrates an LSTM neural network model and combines historical data to predict trends in the cleaned effective data, supporting preventative maintenance decisions and effectively solving the problem of low efficiency in traditional rut detection and analysis. Attached Figure Description

[0092] Figure 1 This is a system block diagram of a non-destructive testing and data analysis device for rut depth of highway pavement proposed in this invention. Detailed Implementation

[0093] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0094] Please see Figure 1 As shown, the present invention is a non-destructive testing and data analysis device for rut depth of highway pavement, comprising:

[0095] Basic detection module: It adopts a multi-laser ranging array to measure the vertical distance of the road surface in a non-contact manner. The laser unit is fixed by the crossbeam steel body and a pulsed laser emitter is integrated. The vertical distance from each point to the crossbeam is calculated by measuring the time difference of reflected light, generating the road surface deformation trajectory curve and directly solving the basic data of rut depth. The raw distance data is output to the dynamic environment adaptive acquisition module.

[0096] Dynamic environment adaptive acquisition module: integrates light intensity sensor and dust concentration monitor to sense environmental parameters in real time, dynamically adjust laser emission power and receiving sensitivity, shorten laser pulse interval and increase laser beam diameter, perform stable measurement with signal-to-noise ratio greater than 30dB, receive raw data and environmental parameters from basic detection module, and output effective measurement values ​​optimized for environmental adaptation to mobile multi-dimensional scanning module.

[0097] Mobile multi-dimensional scanning module: Equipped with a dual-axis mobile platform and an adjustable-angle laser ranging unit array, combined with a force control sensor to sense the road reaction force, it performs autonomous navigation through the SLAM algorithm, receives optimized data from the adaptive acquisition module, and outputs multi-dimensional topographic data to the on-site data classification and processing module.

[0098] On-site data classification and processing module: performs hardware-level triple real-time processing: dynamic threshold removal, moving average filtering, and spatial correlation verification. It reduces the amount of effective data transmission through logic circuits, improves the backend processing speed, receives multi-dimensional data from the scanning module, and outputs cleaned effective data to the data analysis and prediction module.

[0099] Data Analysis and Prediction Module: Integrates an LSTM neural network model, combines historical data, and performs trend prediction on the effective data after on-site processing. It supports preventive maintenance decisions by predicting the development trend of ruts. It receives cleaning data from the graded processing module and outputs prediction results to the user terminal.

[0100] In one embodiment, the non-contact measurement of the vertical distance to the road surface using a multi-laser ranging array includes the following steps:

[0101] 13-25 laser ranging units are arranged at equal intervals along the transverse steel beam (spacing accuracy ≤0.5mm). Each laser unit integrates a pulsed laser emitter and a high-sensitivity photoelectric receiver. The steel beam is fixed to the front of the detection vehicle by a rigid bracket to ensure that the laser array maintains a constant vertical distance from the road surface (usually 300-500mm), forming a non-contact measurement infrastructure.

[0102] Each laser unit emits nanosecond-level pulsed lasers (wavelength 905nm, pulse width ≤20ns) sequentially according to a preset time sequence. The laser beam is projected vertically onto the road surface to form a light spot (default diameter 2mm). The reflected light from the road surface is captured by a photodetector, and the optical signal is converted into an electrical signal by an avalanche photodiode (APD). The time difference Δt between laser emission and reception is recorded.

[0103] Based on the speed of light c (299792458m / s) and the time difference Δt, the vertical distance d from each laser point to the crossbeam is calculated using the formula d = c × Δt / 2. When the ambient temperature deviates from 20℃, the speed of light is calibrated (the speed of light is adjusted by approximately ±0.01ppm for every 1℃ change in temperature), and multiple measurements are taken to average the values ​​(usually 5 times) to reduce the influence of random noise.

[0104] The vertical distance data of 13-25 laser points are sorted by their horizontal positions. A continuous road deformation trajectory curve is generated by a cubic spline interpolation algorithm. By comparing the peak and valley values ​​of the trajectory curve (i.e., the height difference between the deepest rut point and the reference points on both sides), the rut depth value is directly calculated. For example, when the trajectory curve shows that there is a depression with a depth > 5mm in a certain area, the system automatically marks the area as a suspected rut area and triggers the subsequent dynamic environment adaptive acquisition module to perform environmental adaptation optimization.

[0105] The raw distance data of each laser point (including timestamp, temperature-compensated light speed value, and average value of multiple measurements) is bound to environmental parameters (illuminance, dust concentration) and output to the dynamic environment adaptive acquisition module through a standardized data interface.

[0106] In one embodiment, the step of calculating the vertical distance from each point to the crossbeam by measuring the time difference of reflected light, generating the road surface deformation trajectory curve, and directly solving the basic data of rut depth includes the following steps:

[0107] The lateral positions of 13-25 laser points (e.g., x0=0mm, x1=200mm, x2=400mm, ..., x) are determined. n =3200mm) and the corresponding vertical distances (y0=d0, y1=d1,…, y n =d n Arrange them in order to form a discrete set of points;

[0108] At two adjacent laser points ( arrive Between ), fit the curve segment using a cubic polynomial, in the form:

[0109]

[0110] in, , , , The coefficients to be determined must satisfy the following conditions: ( ) = (The curve passes precisely through each measurement point), and the first derivative (slope) and second derivative (curvature) of adjacent curve segments are continuous at the connection point, while the second derivative is zero at the beginning and end endpoints (to avoid excessive curvature at the endpoints).

[0111] After determining all coefficients by solving the system of linear equations, the curve segments are spliced ​​together to form a continuous and smooth road surface deformation trajectory curve, which can accurately reflect subtle deformations such as ruts and waves.

[0112] Locate the feature points of the rut. The valley point is the lowest point of the curve (the first derivative is zero and the second derivative is positive), corresponding to the deepest position of the rut. The reference point is a point on a relatively flat area on both sides of the rut (usually the highest point within the rut width range). Calculate the depth value:

[0113] The depth of the rut is calculated as follows: [(height of the left reference point - height of the valley point) + (height of the right reference point - height of the valley point)] / 2. For example, if the height of the left reference point is 5.2mm, the right reference point is 5.0mm, and the valley point is 3.0mm, then the depth is calculated as [(5.2-3.0) + (5.0-3.0)] ÷ 2 = 2.1mm.

[0114] In one embodiment, the dynamic adjustment of laser emission power and receiving sensitivity includes the following steps:

[0115] The basic detection module outputs a real-time vertical distance sequence from a multi-laser ranging array (distance values ​​of 13-25 laser points, including timestamps and initial measurements). Environmental parameters are measured using an illuminance sensor (0-20000 lux, 1 lux resolution) and a dust concentration monitor (0-100 mg / m³) integrated into the laser device. 3 Resolution 0.1 mg / m 3 Real-time data acquisition via temperature sensors;

[0116] Based on a preset environment-laser parameter mapping model, the FPGA control module analyzes environmental parameters in real time and triggers adjustment strategies.

[0117] Strong light environment adaptation (illuminance > 10000 lux): shorten the laser pulse interval to 100-200 ns (reduce photon scattering), increase the receiving sensitivity to 150% of the original value (suppress background noise), or increase the transmission power to 15-20 mW (enhance signal strength).

[0118] Dust environment suitability (dust concentration > 50 mg / m³) 3 ): Expand the laser beam diameter to 5mm (to improve penetration), reduce the emission power to 8-10mW (to avoid signal oversaturation), and adjust the receiver bandwidth to 5-10MHz (to suppress high-frequency noise);

[0119] Temperature compensation (temperature < -20℃ or > 60℃): The laser wavelength is maintained at 905nm ± 1nm by the built-in TEC temperature control module of the laser emitter, and the power compensation algorithm is triggered (power is adjusted ±0.5% for every 1℃ change in temperature) to ensure measurement stability;

[0120] The FPGA sends adjustment commands to the laser transmitter and receiver to adjust the transmission power, pulse interval, beam diameter and receiver sensitivity in real time. After adjustment, the effect is verified by the signal-to-noise ratio (SNR) monitoring module. If SNR < 30dB, a second adjustment is triggered (such as further shortening the pulse interval or increasing the power). If the standard is not met for 3 consecutive times, the system switches to the backup laser unit for remeasurement.

[0121] Dynamic weighted filtering is applied to the original distance data based on environmental parameters: Gaussian filtering (3×3 window) is used to suppress scattering noise in strong light environment, median filtering (5×5 window) is used to eliminate random jump points in dust environment, and the light speed value is corrected by temperature compensation algorithm when the temperature fluctuates (adjusted by ±0.01ppm per ℃).

[0122] The corrected measurements were verified for signal-to-noise ratio (SNR ≥ 30 dB to be valid), and environmental parameter labels were attached (e.g., "illuminance 12000 lux - dust concentration 60 mg / m³"). 3 This generates effective measurement values ​​with environmental characteristics, ensuring data traceability.

[0123] In one embodiment, the autonomous navigation using the SLAM algorithm, receiving optimized data from the adaptive acquisition module, and outputting multi-dimensional topographic data includes the following steps:

[0124] Equipped with a dual-axis motion platform (horizontal / vertical movement accuracy of 0.1mm), it integrates a lidar (scanning angle 270°, angular resolution 0.5°), an inertial measurement unit (IMU, triaxial acceleration ±16g, angular velocity ±2000° / s), and a wheeled odometer (encoder accuracy 0.1mm). The SLAM algorithm operates based on a tightly coupled lidar-IMU framework.

[0125] The lidar scans the current road surface environment to generate point cloud data, and the IMU collects acceleration and angular velocity information. The two types of data are fused by extended Kalman filter (EKF) to perform real-time pose estimation of the detection equipment in the road surface coordinate system (position error ≤2cm, attitude angle error ≤0.5°).

[0126] Based on the pose estimation results, the laser point cloud data is stitched into a local 3D map of the road surface. Combining the real-time point cloud data of the road surface with the preset detection area boundary, the optimal scanning path is generated using the A* algorithm. The dual-axis moving platform is then controlled to move autonomously along the path to perform continuous and complete detection at a scanning speed of 20 km / h.

[0127] The preset detection area (e.g., a road section 100m long × 5m wide) is discretized into a two-dimensional grid. Each grid has a side length of 0.5m (to match the movement accuracy of the dual-axis platform). The grid node coordinates are (i,j), where i is the vertical (driving direction) index and j is the horizontal index. The starting point is set as the center of the entrance grid of the detection area (e.g., (0,0)) and the ending point is set as the center of the exit grid (e.g., (200,0)). The path must cover all grids to be scanned (to avoid omissions).

[0128] Given the cost g(n): the actual distance traveled from the starting point to the current node n (such as Euclidean distance or Manhattan distance), with an additional cost of 0.5m for each grid cell moved;

[0129] Heuristic cost h(n): Use Euclidean distance to estimate the remaining distance from n to the endpoint, or use diagonal distance (which better fits the motion characteristics of a dual-axis platform).

[0130] Total cost f(n) = g(n) + h(n): The node with the smallest f(n) is dynamically selected for expansion through a priority queue (such as a min-heap) to ensure that the path is globally optimal;

[0131] The system directly calls the basic rut depth data generated by the basic detection module (the deformation trajectory curve generated by cubic spline interpolation and the result of the valley-reference point height difference calculation, such as the example value of 2.1mm), only marking its spatial coordinates in the SLAM map. Based on the SLAM-generated road surface transverse (X-axis) and longitudinal (Y-axis) point cloud data, the system fits the transverse / longitudinal slope curves respectively using the least squares method, and calculates the slope angle between the reference points on both sides of the rut area (such as transverse slope tanθ1=Δh1 / Δx1, longitudinal slope tanθ2=Δh2 / Δy2):

[0132] Cross slope calculation:

[0133] The lateral height z and the lateral coordinate x satisfy a linear relationship:

[0134]

[0135] Among them, a x The tangent of the transverse slope angle (tanθ1) reflects the steepness of the transverse slope. x The intercept represents the height reference at the lateral zero point;

[0136] By adjusting parameter a x b x To minimize the sum of the squares of the differences between the predicted and actual values ​​for all data points:

[0137] Minimize objective:

[0138] Where Σ represents the summation of all N data points within the rut area;

[0139] Find a for the sum of squared residuals S. x b x The partial derivatives of , and setting the derivatives to zero, are rearranged to obtain:

[0140]

[0141]

[0142] Convert slope to angle using the arctangent function:

[0143]

[0144] Longitudinal slope calculation:

[0145] The vertical height z and the vertical coordinate y satisfy a linear relationship:

[0146] z = aᵧ · y + bᵧ

[0147] Where aᵧ is the tangent of the longitudinal slope angle (tanθ2), reflecting the steepness of the longitudinal slope; bᵧ is the intercept, representing the height reference at the longitudinal zero point;

[0148] Minimize objective:

[0149] Take the partial derivatives with respect to aᵧ and bᵧ respectively:

[0150]

[0151]

[0152] Calculate the longitudinal slope angle:

[0153]

[0154] By integrating rut depth, cross slope, and longitudinal slope data into a SLAM 3D map, a complete road surface model containing geometric shape and spatial location information is formed, supporting visualization and analysis.

[0155] In one embodiment, the hardware-level triple real-time processing—dynamic threshold removal, moving average filtering, and spatial correlation verification—includes the following steps:

[0156] Dynamic threshold elimination: The input multi-dimensional topography data (rut depth, cross slope, longitudinal slope) is scanned in real time. Each data point is compared with adjacent points and historical benchmark values ​​to determine whether it exceeds the threshold range.

[0157] Based on the characteristics of road materials (such as the deformation threshold of asphalt concrete pavement is usually ≤ ±8mm) and the detection accuracy requirements (system error ≤ ±2mm), the threshold range is adjusted. For example, the threshold is appropriately widened to ±12mm in areas with dense rutting to avoid false rejection. If a data point exceeds the threshold range, the system marks it as an "abnormal point" and temporarily removes it. At the same time, the timestamp and location information of the point are recorded for subsequent manual review or algorithm optimization.

[0158] Moving average filtering: Applying a moving average filter to a data sequence after removing outliers to suppress high-frequency random noise and smooth data fluctuations;

[0159] Based on the data sampling rate (e.g., 100Hz) and noise characteristics (e.g., abrupt noise caused by dust), the size of the filtering window is adjusted (usually 3-7 data points) to ensure that detailed features of rut deformation are preserved while effectively filtering out noise. A weighted moving average algorithm (e.g., Hanning window) is used to assign different weights to the data points within the window (the center point has the highest weight, and the weights of the edge points decrease). The weighted average is calculated as the output value after filtering to improve the data smoothing effect. Millisecond-level filtering processing is implemented through FPGA hardware logic circuits to ensure that the filtering delay is ≤10ms, which meets the requirements of real-time detection and avoids scanning omissions or data distortion caused by processing delays.

[0160] Spatial correlation verification: Perform spatial correlation verification on the filtered data to ensure the logical consistency of the data in the spatial dimension;

[0161] Check whether the height difference between adjacent data points meets the preset gradient threshold (e.g., height difference between adjacent points ≤ 5mm / m). If it exceeds the threshold, it is marked as a "suspicious point" and a secondary verification is triggered. Based on the 3D road map generated by SLAM, verify whether the data in the current scanning area is consistent with the global topography features (e.g., rut direction, slope change). If there is a significant difference, a local rescan is triggered. The verification result is fed back to the dynamic threshold model and the filter parameter adjustment module, forming a closed-loop mechanism of "processing-verification-optimization" to continuously improve the accuracy and reliability of data processing.

[0162] The three-stage processing steps form a tightly linked technical chain through a progressive logic of anomaly removal, noise suppression, and consistency verification. Dynamic threshold removal serves as the first line of defense, quickly identifying and removing obviously abnormal data points to provide a clean data foundation for subsequent processing. Moving average filtering serves as the second line of defense, further smoothing the data after removing outliers and suppressing the impact of random noise on data quality. Spatial correlation verification serves as the final checkpoint, ensuring that the processed data conforms to the logical characteristics of the actual road surface in the spatial dimension, avoiding data distortion caused by local noise or algorithm errors.

[0163] In one embodiment, reducing the amount of effective data transmission and improving backend processing speed through logic circuitry includes the following steps:

[0164] Before the data enters the logic circuit, the preprocessing module built into the FPGA performs preliminary screening and feature extraction on the raw data. Based on a preset dynamic threshold (such as the rate of change of rut depth > 0.5 mm / m), it quickly identifies and removes data from flat areas without significant deformation, retaining only the laser point data of suspected rut areas (such as depth changes > 2 mm), directly reducing the amount of raw data by more than 60%. Feature tags (such as "rut edge points") are added to the retained data, and key data points are marked before transmission through bit operations (such as setting data packet header flag bits), which facilitates the back-end processing module to quickly locate core information.

[0165] FPGAs perform efficient data compression using dedicated logic circuits, specifically employing a combined strategy of differential coding and entropy compression.

[0166] Differential encoding: Perform differential calculations on the vertical distance data of consecutive laser points (e.g., Δd = - The absolute distance value is converted into a relative change, and the high correlation between adjacent point data is used to reduce the amount of data stored. Huffman coding is used to perform lossless compression on the differential data, and variable-length codewords are assigned according to the frequency of data occurrence (short code for high-frequency data and long code for low-frequency data) to further compress the data volume. Hardware acceleration of the compression algorithm is performed through the parallel computing unit of FPGA to ensure that the compression latency is ≤5ms, which meets the real-time processing requirements.

[0167] In one embodiment, the integrated LSTM neural network model, combined with historical data, performs trend prediction on the effective data after on-site processing, including the following steps:

[0168] Build a multi-source heterogeneous data fusion library to integrate the following data:

[0169] Historical data includes monitoring data on rut depth, cross slope, and longitudinal slope over the past 3-5 years (sampling frequency 1Hz), traffic load data (such as average daily traffic volume and proportion of heavy vehicles), environmental data (such as temperature, humidity, rainfall, illuminance, and dust concentration), and pavement material characteristics data (such as asphalt type, thickness, and degree of aging).

[0170] Effective on-site data: Cleaned data output by the on-site data classification and processing module, including real-time rut depth, cross slope, longitudinal slope and multi-dimensional topographic data, as well as corresponding environmental parameters and laser parameter status;

[0171] Based on the fused dataset, construct and train an LSTM neural network model:

[0172] A two-layer LSTM structure (64 hidden units per layer) is used, combined with a Dropout layer (scale 0.2) to prevent overfitting. The output layer uses a fully connected layer to predict the trend of rut depth changes over the next 72 hours. Key features (such as rut ​​depth change rate, cross slope change gradient, and traffic load fluctuation rate) are extracted as inputs, and an attention mechanism is used to enhance the influence of important features on the prediction results. The Adam optimizer (learning rate 0.001) is used with mean squared error (MSE) as the loss function. The model is trained for 50 rounds on the training set (80% of the data) and its performance is evaluated on the validation set (20% of the data) (prediction error ≤10%). Cross-validation and leave-one-out method are used to verify the model's generalization ability, ensuring stable prediction accuracy under different environmental conditions and traffic load scenarios.

[0173] The processed data is input into the trained LSTM model to perform trend prediction.

[0174] Based on historical data from the previous 72 hours, the model predicts the trend of rut depth changes over the next 72 hours and outputs a prediction curve and confidence interval (such as the prediction range at a 95% confidence level). As new data is input in real time, the model dynamically updates the prediction results through a sliding window mechanism to ensure that the prediction is always based on the latest road conditions and environmental conditions. The prediction results are displayed on user terminals (such as in-vehicle displays or cloud platforms) in the form of curves, heat maps, etc., to intuitively present the spatiotemporal evolution trend of rut depth, making it easy for maintenance personnel to understand quickly.

[0175] Based on the prediction results, preventative maintenance decision recommendations are generated:

[0176] When the predicted rut depth exceeds a preset threshold (e.g., 5mm for asphalt pavement, 8mm for concrete pavement), the system automatically triggers a three-level warning (yellow, orange, and red) and recommends maintenance measures (e.g., local repairs, repaving, traffic control). It prioritizes areas requiring maintenance based on factors such as the rate of change of rut depth, traffic flow, and remaining pavement life. Based on the prediction results, the system recommends the optimal maintenance time (e.g., during periods of low traffic flow), maintenance scope (e.g., repair areas accurate to the meter), and maintenance method (e.g., micro-surfacing, thin overlay) to reduce maintenance costs and extend pavement life. After maintenance is performed, it collects maintenance effect data (e.g., changes in rut depth, user feedback) and feeds it back into the LSTM model for parameter updates.

[0177] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A non-destructive testing and data analysis device for rut depth of highway pavement, characterized in that, include: Basic detection module: It adopts a multi-laser ranging array to measure the vertical distance of the road surface in a non-contact manner. The laser unit is fixed by the crossbeam steel body and a pulsed laser emitter is integrated. The vertical distance from each point to the crossbeam is calculated by measuring the time difference of reflected light, generating the road surface deformation trajectory curve and directly solving the basic data of rut depth. The raw distance data is output to the dynamic environment adaptive acquisition module. Dynamic environment adaptive acquisition module: integrates light intensity sensor and dust concentration monitor to sense environmental parameters in real time, dynamically adjust laser emission power and receiving sensitivity, shorten laser pulse interval and increase laser beam diameter, perform stable measurement with signal-to-noise ratio greater than 30dB, receive raw data and environmental parameters from basic detection module, and output effective measurement values ​​optimized for environmental adaptation to mobile multi-dimensional scanning module. Mobile multi-dimensional scanning module: Equipped with a dual-axis mobile platform and an adjustable-angle laser ranging unit array, combined with a force control sensor to sense the road reaction force, it performs autonomous navigation through the SLAM algorithm, receives optimized data from the adaptive acquisition module, and outputs multi-dimensional topographic data to the on-site data classification and processing module. On-site data classification and processing module: performs hardware-level triple real-time processing: dynamic threshold removal, moving average filtering, and spatial correlation verification. It reduces the amount of effective data transmission through logic circuits, improves the backend processing speed, receives multi-dimensional data from the scanning module, and outputs cleaned effective data to the data analysis and prediction module. Data Analysis and Prediction Module: Integrates an LSTM neural network model, combines historical data, and performs trend prediction on the effective data after on-site processing. It supports preventive maintenance decisions by predicting the development trend of ruts. It receives cleaning data from the graded processing module and outputs prediction results to the user terminal.

2. The non-destructive testing and data analysis device for rut depth of highway pavement according to claim 1, characterized in that, The method of non-contact measurement of vertical distance to the road surface using a multi-laser ranging array includes the following steps: Laser ranging units are arranged at equal intervals along the transverse steel body of the beam. Each laser unit integrates a pulsed laser emitter and a high-sensitivity photoelectric receiver. The beam steel body is fixed to the front of the detection vehicle by a rigid bracket, so that the laser array maintains a constant vertical distance from the road surface. Each laser unit emits nanosecond-level pulsed lasers sequentially according to a preset time sequence. The laser beam is projected vertically onto the road surface to form a light spot. The reflected light from the road surface is captured by a photodetector, and the optical signal is converted into an electrical signal by an avalanche photodiode. The time difference between laser emission and reception is recorded. The vertical distance from each laser point to the crossbeam is calculated based on the speed of light and the time difference. When the ambient temperature deviates, the speed of light value is calibrated. The vertical distance data of the laser points are sorted by their horizontal positions. A continuous road surface deformation trajectory curve is generated by a spline interpolation algorithm. By comparing the peak and valley values ​​of the trajectory curve, the rut depth value is calculated. The original distance data of each laser point is bound to environmental parameters and output to the dynamic environment adaptive acquisition module through a standardized data interface.

3. The non-destructive testing and data analysis device for rut depth of highway pavement according to claim 1, characterized in that, The process of calculating the vertical distance from each point to the crossbeam by measuring the time difference of reflected light, generating the road surface deformation trajectory curve, and directly solving the basic data of rut depth includes the following steps: Arrange the lateral positions and corresponding vertical distances of the laser points in order to form a discrete set of points. Between two adjacent laser points, fit a curve segment with a cubic polynomial. After determining all coefficients by solving the linear equation system, stitch the curve segments together to form a continuous and smooth road surface deformation trajectory curve. Locate the feature points of the ruts. The valley point is the lowest point of the curve, corresponding to the deepest position of the rut. The reference point is the point in the relatively flat area on both sides of the rut. Calculate the depth value.

4. The non-destructive testing and data analysis device for rut depth of highway pavement according to claim 1, characterized in that, The dynamic adjustment of laser emission power and receiving sensitivity includes the following steps: The basic detection module outputs a real-time vertical distance sequence from the multi-laser ranging array, while environmental parameters are collected in real time by a light intensity sensor, a dust concentration monitor, and a temperature sensor integrated into the laser device. The FPGA control module analyzes environmental parameters in real time and triggers adjustment strategies, including strong light environment adaptation, dust environment adaptation, and temperature compensation. The FPGA sends adjustment commands to the laser transmitter and receiver to adjust the transmission power, pulse interval, beam diameter and receiver sensitivity in real time. After adjustment, the effect is verified by the signal-to-noise ratio (SNR). If the SNR is less than 30dB, a second adjustment is triggered. If the standard is not met for a continuous period, the system switches to the backup laser unit for remeasurement. The original distance data is dynamically weighted and filtered based on environmental parameters. Gaussian filtering is used in strong light environments, median filtering is used in dusty environments, and the speed of light is corrected through a temperature compensation algorithm when the temperature fluctuates.

5. The non-destructive testing and data analysis device for rut depth of highway pavement according to claim 1, characterized in that, The autonomous navigation using the SLAM algorithm, receiving optimized data from the adaptive acquisition module, and outputting multi-dimensional topographic data includes the following steps: Equipped with a dual-axis mobile platform, integrating LiDAR, inertial measurement unit, and wheeled odometry, the SLAM algorithm operates based on a tightly coupled LiDAR-IMU framework. The lidar scans the current road surface environment to generate point cloud data, while the IMU collects acceleration and angular velocity information to perform real-time pose estimation of the detection equipment in the road coordinate system. Based on the pose estimation results, the lidar point cloud data is stitched together to form a local 3D map of the road surface. Combining the real-time point cloud data of the road surface with the preset detection area boundary, the optimal scanning path is generated using the A* algorithm. The dual-axis moving platform is then controlled to move autonomously along the path for continuous and comprehensive detection. The basic data of rut depth generated by the basic detection module is directly called, and only its spatial coordinates in the SLAM map are marked. Based on the lateral and longitudinal point cloud data of the road surface generated by SLAM, the lateral / longitudinal slope curves are fitted by the least squares method to calculate the slope angle between the reference points on both sides of the rut area. By integrating rut depth, cross slope, and longitudinal slope data into a SLAM 3D map, a complete road surface model containing geometric shape and spatial location information is formed, supporting visualization and analysis.

6. The non-destructive testing and data analysis device for rut depth of highway pavement according to claim 1, characterized in that, The hardware-level triple real-time processing includes dynamic threshold removal, moving average filtering, and spatial correlation verification, comprising the following steps: Dynamic threshold rejection: The system scans the input multi-dimensional topography data in real time. Each data point is compared with neighboring points and historical benchmark values ​​to determine whether it exceeds the threshold range. The threshold range is adjusted in combination with the characteristics of the road material and the detection accuracy requirements. If a data point exceeds the threshold range, the system marks it as an abnormal point and temporarily rejects it. At the same time, the timestamp and location information of the point are recorded. Moving average filtering: The data sequence after removing outliers is processed by moving average filtering. The size of the filtering window is adjusted according to the data sampling rate and noise characteristics. A weighted moving average algorithm is used to assign different weights to the data points in the window and calculate the weighted average as the filtered output value. Spatial correlation verification: Spatial correlation verification is performed on the filtered data to check whether the height difference between adjacent data points meets the preset gradient threshold. If it exceeds the threshold, it is marked as a suspicious point and a secondary verification is triggered. Based on the 3D road map generated by SLAM, it verifies whether the data in the current scanning area is consistent with the global topographic features. If there is a significant difference, a local rescan is triggered. The verification results are fed back to the dynamic threshold model and the filter parameter adjustment module.

7. The non-destructive testing and data analysis device for rut depth of highway pavement according to claim 1, characterized in that, The method of reducing the amount of effective data transmission and improving backend processing speed through logic circuits includes the following steps: Before the data enters the logic circuit, the preprocessing module built into the FPGA performs preliminary screening and feature extraction on the raw data. Based on the preset threshold, it identifies and removes flat area data without significant deformation, retains only the laser point data of suspected rut areas, adds feature labels to the retained data, and marks key data points before transmission. FPGAs perform efficient data compression using dedicated logic circuits, specifically employing a composite strategy of differential coding and entropy compression. Differential calculation is performed on the vertical distance data of continuous laser points to convert absolute distance values ​​into relative changes. The high correlation between adjacent point data is used to reduce the amount of data stored. Huffman coding is used to perform lossless compression on the differential data, and variable-length codewords are allocated according to the frequency of data occurrence to further compress the data volume.

8. The non-destructive testing and data analysis device for rut depth of highway pavement according to claim 1, characterized in that, The integrated LSTM neural network model, combined with historical data, performs trend prediction on the effective data after on-site processing, including the following steps: A multi-source heterogeneous data fusion library is constructed to integrate historical data and effective field data. Based on the fused dataset, an LSTM neural network model is built and trained. The effective field data is input into the trained LSTM model to perform trend prediction. The model predicts the future trend of rut depth change based on historical data before the current moment and outputs the prediction curve and confidence interval. With the real-time input of new data, the model dynamically updates the prediction results through a sliding window mechanism. Based on the prediction results, preventive maintenance decision suggestions are generated. When the predicted rut depth exceeds a preset threshold, an early warning is triggered and maintenance measures are recommended. Combining factors including the rate of change of rut depth, traffic flow, and remaining pavement life, the areas requiring maintenance are prioritized. Based on the prediction results, the optimal maintenance time, maintenance scope, and maintenance method are recommended.

Images