A method for real-time evaluation and threshold correction of paving thickness and flatness

Abstract

The application discloses a kind of paving thickness and flatness real-time evaluation and subthreshold correction method on machine, it is related to data processing technical field.The method comprises the following steps: step one: the sensor height data of the local working area behind the screed is obtained to generate observation data containing virtual paving thickness deviation information and multi-scale flatness index information;Step two: based on the observation data, a pre-set sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm is called, and combined with short-term prediction and conditional risk assessment, a subthreshold correction instruction representing the target correction level is generated;Step three: based on the subthreshold correction instruction, the correction action corresponding to the subthreshold correction instruction is executed by the actuator.The application can maintain stable, continuous and forward-looking control effect under complex disturbance conditions.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for real-time on-machine evaluation and subthreshold correction of paving thickness and flatness. Background Technology

[0002] In road paving construction, the thickness and smoothness of the paved surface are core indicators for evaluating paving quality. To ensure the final pavement meets design elevation and smoothness requirements, construction equipment typically relies on sensors such as lasers, ultrasonic sensors, inertial units, or contact slippers to acquire height information near the screed, and then adjusts the thickness by controlling the raising or lowering of the screed. In traditional construction methods, these sensor outputs are often simply filtered and averaged, and the control strategy is usually based on proportional control or proportional-integral control to directly correct for paved surface thickness deviations. This approach provides basic correction capabilities when thickness deviations are large, but it is easily affected by factors such as road surface roughness, material supply fluctuations, mechanical vibration, and temperature changes under complex working conditions, leading to problems such as response lag, over-following, or frequent fine-tuning in the control system. Some technologies have attempted to introduce surface fitting, signal smoothing, or trend analysis to improve the stability of thickness assessment. For example, some technologies use moving averages, Kalman filtering, or polynomial surface fitting to smooth sensor height data, improving the stability of paved surface thickness calculation by reducing measurement noise. However, these techniques often exhibit estimation biases when there are rapid changes or non-uniform noise distributions in localized areas because they cannot adjust for estimation uncertainties based on real-time environmental changes or reflect differences in data reliability across different regions. Furthermore, some methods attempt to extract smoothness features from height signals using multi-scale filtering structures, but these often rely on fixed-length windows, resulting in insufficient response speed to sudden changes in operating conditions and difficulty in accurately reflecting short-term local disturbances caused by material accumulation and changes in screed posture. Summary of the Invention

[0003] In view of this, the present invention provides an on-machine real-time assessment and sub-threshold correction method for paving thickness and smoothness. This method enables real-time assessment of the paving thickness deviation and smoothness of the local working area behind the screed, and allows for sub-threshold correction before the deviation reaches the upper limit of the specification. By modeling the uncertainty of height data, extracting multi-scale smoothness features, identifying sudden changes in working conditions, and predicting future deviation risks, the present invention can maintain stable, continuous, and forward-looking control effects under complex disturbance conditions. Combined with an adaptive thickness controller and soft dead zone control, the actuator reduces actions at small deviations and intervenes promptly before the deviation expands, significantly improving the accuracy of thickness control, smoothness assurance capability, and equipment operation stability, thereby improving the final pavement quality consistency and construction reliability.

[0004] The technical solution adopted in this invention is as follows:

[0005] A method for real-time on-machine evaluation and subthreshold correction of paving thickness and smoothness includes the following steps:

[0006] Step 1: Obtain sensor height data of the local working area behind the ironing plate, perform sparse Gaussian process surface reconstruction on the local working area corresponding to the obtained sensor height data, and perform multi-scale wavelet roughness calculation based on the reconstruction results to generate observation data containing paving thickness deviation information and multi-scale flatness index information.

[0007] Step 2: Based on the observation data, call the preset sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm, determine the current system state by performing Bayesian online change point detection, and generate a subthreshold correction instruction characterizing the target correction level by combining short-term prediction and conditional risk assessment.

[0008] Step 3: Based on the subthreshold correction command and combined with the Lyapunov function approximation evaluation of the system stability state, the actuator performs the correction action corresponding to the subthreshold correction command through the synergistic effect of the adaptive thickness controller and the soft dead zone control module.

[0009] Furthermore, in step one, the operations of acquiring sensor height data and constructing a local working area include: fixing multiple laser displacement sensors at predetermined lateral intervals along the horizontal direction of the ironing board and setting a preset sampling frequency; collecting height data and corresponding lateral positions of multiple laser displacement sensors in each sampling cycle; establishing a device coordinate system with the center of the ironing board's trailing edge as the origin, and projecting the height data within a predetermined longitudinal range adjacent to the trailing edge of the ironing board onto the device coordinate system; dividing the predetermined longitudinal range into multiple equal-length units along the longitudinal direction, and dividing the predetermined lateral range into multiple equal-width units along the lateral direction, thereby forming a local working area containing multiple grid units.

[0010] Furthermore, the sparse Gaussian process surface reconstruction in step one includes: continuously collecting height data within a preset initial time after paving begins, accumulating horizontal, vertical, and height values ​​that are at least equal to the total number of grid cells in the local working area; using the center position of each grid cell as a candidate induced point position, selecting a preset number of induced points based on the principle of uniform vertical and horizontal distribution; constructing a sparse Gaussian process surface reconstruction model, using a preset kernel function and Gaussian noise assumption, and solving the initial sparse Gaussian process surface reconstruction model by performing matrix decomposition on the covariance matrix between induced points.

[0011] Furthermore, the sparse Gaussian process surface reconstruction model in step one is updated in each subsequent sampling period. The update operation includes: mapping the currently collected height data to the corresponding grid cell center position, calculating the Euclidean distance between the grid cell center position and the preset number of induced point positions; when the minimum distance of a data point is greater than a preset distance threshold, adding the data point position and corresponding height value to the candidate induced point set; when the number of points in the candidate induced point set reaches a preset number, sorting the candidate induced points from highest to lowest according to the number of times they are hit by height data in the most recent preset time period, selecting a preset number of candidate induced points with higher hit counts to replace the original induced point set; and recalculating the covariance matrix of the new induced point set and performing a matrix factorization to update the sparse Gaussian process surface reconstruction model.

[0012] Further, in step one, the operation of generating observation data includes: within each sampling period, using the latest sparse Gaussian process surface reconstruction model, predicting the height of the center positions of all grid cells within the local working area, outputting a height estimate and a height estimate fluctuation range value for each grid cell, wherein the height estimate fluctuation range value is obtained by taking the square root of the diagonal elements of the Gaussian process covariance matrix and truncated between the preset lower and upper limits of the fluctuation range; subtracting the height estimate value of each grid cell from the corresponding design elevation in the pre-stored road design model to obtain the virtual paving thickness deviation value of the grid cell; within each sampling period, selecting the row of grid cells closest to the trailing edge of the screed in the longitudinal direction, and taking the arithmetic mean of the virtual paving thickness deviation values ​​of this row of grid cells to obtain the virtual paving thickness deviation sampling value for the current sampling period; continuously sampling... The virtual tiling thickness deviation sampled values ​​obtained from the sampling period are arranged in chronological order to form a virtual tiling thickness deviation sequence. Multi-scale wavelet decomposition is performed on the most recent preset number of sampled values ​​in the virtual tiling thickness deviation sequence, using a preset wavelet basis and a preset number of decomposition levels to obtain the level 1, level 2, and level 3 detail coefficients. The square of each level detail coefficient is calculated and its square root is taken to obtain the shortwave smoothness index, midwave smoothness index, and longwave smoothness index values. These are then combined in a preset order to form a multi-scale smoothness index vector. Finally, the current virtual tiling thickness deviation sampled value, the multi-scale smoothness index vector, and the arithmetic mean of the height estimation fluctuation range of all grid cells within the current sampling period are combined as a set of observation data.

[0013] Furthermore, in step two, the sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm includes: setting the observation buffer length to a preset value; storing a set of observation data in the observation buffer during each sampling period; discarding the earliest set of observation data in chronological order when the number of observation data in the observation buffer exceeds the preset length; performing Bayesian online change point detection based on binomial statistics during each sampling period; Bayesian online change point detection includes: selecting the most recent preset number of observation data from the observation buffer; and performing anomaly marking judgment on each set of observation data; the anomaly marking judgment process is as follows: reading the absolute value of the virtual tiling thickness deviation sampling value and the average value of the height estimation fluctuation range from a set of observation data, calculating the ratio between the two, and marking the anomaly of the current set of observation data when the calculated ratio is greater than a preset ratio threshold. The flag is set to 1. When the calculated ratio is within the range of 0 to the preset ratio threshold, the anomaly flag for this group is set to 0. At the same time, the shortwave smoothness index value and the midwave smoothness index value are read from the observation data of this group. When either the shortwave smoothness index value or the midwave smoothness index value is greater than the preset smoothness threshold, the anomaly flag of the currently judged observation data is directly set to 1. The number of observation data with anomaly flag of 1 in the most recent preset number of observation data is summed to obtain the anomaly count value. The anomaly count value is increased by 1 and then divided by the most recent preset number plus 2 to obtain the change point probability value of the current sampling period. When the change point probability value is within the range of the preset change point probability threshold to 1.0, the current sampling period is considered to be in a change point state. When the change point probability value is within the range of 0 to the preset change point probability threshold, the current sampling period is considered to be in a stable state.

[0014] Furthermore, the sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm in step two further includes a short-term prediction sub-step, which includes: reading the most recent preset number of virtual pavement thickness deviation sample values ​​from the virtual pavement thickness deviation sequence; calculating the arithmetic mean of these virtual pavement thickness deviation sample values ​​as the average virtual pavement thickness deviation; calculating the difference between the earliest and latest sample values ​​in this batch of virtual pavement thickness deviation sample values; dividing the difference by the most recent preset number minus 1 to obtain the average change in virtual pavement thickness deviation; setting the prediction time window length to a preset number of prediction sampling periods; and for the k-th prediction sampling period, comparing the average virtual pavement thickness deviation with the average change in virtual pavement thickness deviation. The k-th prediction center deviation is obtained by adding k times the amount of data, and so on until the prediction center deviation of all prediction sampling periods is obtained; k starts from 1; within the same sampling period, the average value of the height estimation fluctuation range is read from the current observation data, and the average value of the height estimation fluctuation range is multiplied by a preset multiple to obtain the prediction deviation fluctuation range; for each prediction sampling period, the prediction center deviation plus 1 times the prediction deviation fluctuation range, the prediction center deviation plus 2 times the prediction deviation fluctuation range, and the prediction center deviation plus 3 times the prediction deviation fluctuation range are calculated respectively to obtain 3 prediction deviation candidate values. The total of 3 prediction deviation candidate values ​​multiplied by the number of prediction sampling periods is obtained for all prediction sampling periods.

[0015] Furthermore, the sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm in step two further includes a sub-threshold correction triggering sub-step based on conditional risk values, which includes: pre-storing preset allowable deviation values ​​for paving thickness in the control system; calculating the absolute value of all candidate prediction deviation values ​​generated by the short-term prediction sub-step to obtain multiple absolute prediction deviation values; for each absolute prediction deviation value, calculating the difference between the absolute prediction deviation value and the preset allowable deviation value for paving thickness, treating the difference as 0 when the difference is less than 0, and using the difference as the loss value of the candidate prediction deviation value when the difference is greater than or equal to 0; sorting all loss values ​​from largest to smallest, and selecting the top preset number of values ​​from the sorting results. The loss value is calculated as the arithmetic mean of a preset number of loss values ​​to obtain the conditional risk measure for the current sampling period. When the conditional risk is in a stable state, the conditional risk trigger threshold is set to the first trigger threshold, and when the conditional risk is in a changing state, the conditional risk trigger threshold is set to the second trigger threshold. The first trigger threshold is greater than the second trigger threshold. According to the different numerical ranges of the conditional risk measure, the sub-threshold correction instruction level for the current sampling period is set to level 0, level 1, level 2, or level 3, and the sub-threshold correction instruction level and the corresponding target dummy thickness correction amount are output to the control execution module. The level 1, level 2, and level 3 sub-threshold correction instructions correspond to progressively increasing target dummy thickness correction amounts, respectively.

[0016] Furthermore, in step three, the thickness controller for approximate evaluation and adaptive adjustment of the system stability state using the Lyapunov function includes: configuring a thickness controller containing proportional and integral elements in the paver control system, and configuring an actuator to lift or lower the screed; within each sampling period, reading the most recent preset number of virtual paving thickness deviation sample values ​​from the virtual paving thickness deviation sequence, calculating the squared average of these sample values ​​to obtain the first stability index, and subtracting the current first stability index from the first stability index of the previous sampling period to obtain the second stability index; when the first stability index is within the range of 0 to a preset stability threshold and the second stability index is within a preset negative range... When the current control state is within the range, it is considered a stable state. In the stable state, the amplification factors of the proportional and integral components in the thickness controller are multiplied by a preset first adjustment factor at a preset adjustment period. The first adjustment factor is less than 1, until the amplification factor is reduced to the preset lower limit of the initial amplification factor. When the first stability index is greater than the preset stability threshold and the second stability index is within the preset positive value range, the current control state is considered a deviation amplification state. In the deviation amplification state, the amplification factors of the proportional and integral components in the thickness controller are multiplied by a preset second adjustment factor at a preset adjustment period. The second adjustment factor is greater than 1, until the amplification factor is increased to the preset upper limit of the initial amplification factor.

[0017] Furthermore, the synergistic effect of the soft dead zone control module in step three includes: configuring a soft dead zone control module in the actuator control channel; in the soft dead zone control module, dividing the control interval according to the absolute value of the current virtual pavement thickness deviation sampling value: when the absolute value of the virtual pavement thickness deviation is less than the first dead zone threshold, the output of the thickness controller is set to 0 after processing by the soft dead zone control module; when the absolute value of the virtual pavement thickness deviation is within the range of the first dead zone threshold to the second dead zone threshold, the output of the thickness controller is continuously mapped between 0 and the original output according to the linear proportion of the absolute value of the virtual pavement thickness deviation between the two dead zone thresholds; when the absolute value of the virtual pavement thickness deviation is greater than or equal to the second dead zone threshold, the output of the thickness controller is directly transmitted to the actuator; based on the soft dead zone control, the target virtual pavement thickness correction amount corresponding to the sub-threshold correction command level is superimposed on the input of the thickness controller, so that the actuator executes the sub-threshold correction command in a graded, smooth and jitter-suppressing manner.

[0018] By employing the above technical solutions, this invention achieves the following beneficial effects: This invention utilizes a sparse Gaussian process to reconstruct the surface of the local working area behind the screed, simultaneously outputting height estimates and height estimation fluctuation ranges. This provides fundamental data with both accuracy and uncertainty expression for the analysis of paving thickness deviation and smoothness, enabling the system to maintain reliability even under conditions of high noise or complex pavement structures. Multi-scale wavelet roughness calculations are performed based on the sparse Gaussian process reconstruction results, separating the short-wave, medium-wave, and long-wave characteristics of the paving thickness deviation sequence. This accurately identifies the sources of smoothness fluctuations at different scales, providing more structured information input for subsequent change point detection and risk assessment. By performing Bayesian online change point detection based on observation data, this invention can promptly identify the system state when construction conditions undergo abrupt changes, allowing the subthreshold correction logic and control strategy to be adjusted in real time according to local stability or changing trends. By making short-term predictions of future deviations and assessing conditional risks, this invention can make small corrections before the deviation reaches the specification upper limit, keeping the thickness deviation within a safer operating range and avoiding over-reliance on large-scale corrections. Furthermore, by incorporating an adaptive thickness control strategy based on Lyapunov function approximation, this invention can automatically adjust the control gain according to system stability. This avoids high-frequency jitter caused by excessive gain during the stable phase and improves response capability during the deviation amplification phase, resulting in a smoother and more efficient thickness control process. Combined with the suppression effect of soft dead zone control on small deviation phases, this invention significantly reduces the operating frequency of the actuators without affecting the correction capability, thereby reducing mechanical wear and improving construction continuity. In summary, this invention achieves synergistic optimization in thickness estimation accuracy, evenness feature extraction, sudden change in working conditions identification, future risk prediction, and control stability, making automated corrections during the paving process more timely and reliable, and effectively improving the quality consistency of the formed pavement. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the sparse Gaussian process surface reconstruction effect in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the three-layer wavelet decomposition coefficients in an embodiment of the present invention;

[0021] Figure 3 This is a timing distribution diagram of the subthreshold correction instruction level in an embodiment of the present invention. Detailed Implementation

[0022] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0023] Any feature disclosed in this specification, unless otherwise stated, may be replaced by other equivalent or similar features. That is, unless otherwise stated, each feature is merely one example of a series of equivalent or similar features.

[0024] A method for real-time on-machine evaluation and subthreshold correction of paving thickness and smoothness includes the following steps:

[0025] Step 1: Obtain sensor height data of the local working area behind the ironing plate, perform sparse Gaussian process surface reconstruction on the local working area corresponding to the obtained sensor height data, and perform multi-scale wavelet roughness calculation based on the reconstruction results to generate observation data containing paving thickness deviation information and multi-scale flatness index information.

[0026] In step one, the sensor height data of the local working area behind the ironing plate can be obtained in the following way: sparse Gaussian process surface reconstruction is performed on the local working area corresponding to the obtained sensor height data, and multi-scale wavelet roughness calculation is performed based on the reconstruction results to generate observation data containing paving thickness deviation information and multi-scale flatness index information.

[0027] In one implementation, multiple laser displacement sensors are arranged laterally along the screed. For example, five laser displacement sensors can be fixedly installed at approximately 0.3-meter intervals along the trailing edge of the screed, covering the main area of ​​the paving width. Using multiple laser displacement sensors instead of a single sensor allows for the simultaneous acquisition of height information from multiple lateral positions, reducing interpolation errors in the lateral direction and making the subsequent sparse Gaussian process surface reconstruction more stable and reliable. To ensure a sufficiently dense set of longitudinal data points even at the paver's normal travel speed, the sampling frequency can be set to 100 times per second, simultaneously reading the height data from the five laser displacement sensors and the corresponding lateral position for each sensor in each sampling cycle.

[0028] To standardize the description of these height data, a device coordinate system is established at the center of the screed's trailing edge, defining the area behind the screed's trailing edge along the forward direction as the local working area. For example, a predetermined longitudinal range can be selected within 2 meters immediately behind the screed's trailing edge, while a predetermined lateral range can be selected corresponding to the coverage area of ​​the five laser displacement sensors. The height data obtained in each sampling cycle is converted into lateral and longitudinal positions with the screed's trailing edge center as the origin using the geometric relationship between the sensor's fixed installation position and the device coordinate system, thus obtaining sensor height data that matches the device coordinate system. Using this device coordinate system ensures that even with minor changes in the paver's posture, height data from different sampling cycles can still be compared within a unified coordinate system, avoiding height misjudgments caused by directly using vehicle body coordinates.

[0029] To facilitate the reconstruction of the sparse Gaussian process surface, the predetermined longitudinal range is divided into several equal-length units, and the predetermined transverse range is divided into several equal-width units, thus forming a local working area containing multiple grid units. For example, a 2-meter longitudinal range can be divided into 40 equal-length units, each with a length of 0.05 meters; and the transverse range can be divided into 10 equal-width units, each with a width of approximately 0.3 meters, resulting in a total of 400 grid units. Each grid unit corresponds to a unique grid unit center position. With this division method, the longitudinal resolution is 0.05 meters, which can reflect thickness changes caused by minor vibrations of the ironing plate or uneven material feeding. In the transverse direction, observation data is provided by five laser displacement sensors. Combined with the grid division, the height distribution of the entire local working area can be continuously estimated by reconstructing the sparse Gaussian process surface for grid units that were not directly measured.

[0030] Within a preset initial time period after paving begins, such as 1 second, the height data from the aforementioned sensors is continuously collected. With a sampling frequency of 100 times per second and using 5 laser displacement sensors, 500 data records containing lateral, longitudinal, and height values ​​can be accumulated within 1 second. This data volume is no less than the total number of 400 grid cells, covering the entire local working area. For each data record, its grid cell can be determined based on its lateral and longitudinal positions, and the center of that grid cell is used as a candidate guide point. The closer the candidate guide point is to the actual sensor acquisition point, the closer the height estimate obtained from the subsequent sparse Gaussian process surface reconstruction will be to the true height. To maintain model accuracy while controlling computational complexity, 64 guide points can be selected from all candidate guide point locations, following the principle of as even a distribution as possible in the longitudinal and lateral directions, forming a guide point set. The choice of the number of guide points is between the total number of grid cells and the number of sensors, balancing reconstruction accuracy and computation time, enabling the sparse Gaussian process surface reconstruction to run in real time during the paver's movement.

[0031] Based on the aforementioned set of induced points and the corresponding height values ​​for each induced point, a sparse Gaussian process surface reconstruction model can be constructed. During the construction process, the spatial correlation between any two induced points is described using a predefined kernel function, which can be implemented with higher correlation for closer spatial distances and lower correlation for farther distances. By performing matrix decomposition on the covariance matrix between the induced points, the height estimate and height fluctuation range for the center of any grid cell can be quickly calculated during subsequent prediction. Compared to simple interpolation methods, sparse Gaussian process surface reconstruction not only provides height estimates but also the height fluctuation range for each location, clearly identifying which regions have more reliable estimates. This facilitates the differentiation between high-confidence and low-confidence regions in subsequent thickness deviation and flatness analysis.

[0032] After constructing the initial sparse Gaussian process surface reconstruction model, the model input is updated using the currently acquired sensor height data in each sampling period. This data is mapped to the corresponding grid cell center positions to ensure that the latest height information can participate in surface reconstruction. Subsequently, the latest sparse Gaussian process surface reconstruction model is used to predict the height of the center positions of all 400 grid cells within the local working area. For each grid cell, a height estimate and a height estimate fluctuation range are obtained. The smaller the height estimate fluctuation range, the more stable the height estimate result of that grid cell, equivalent to having sufficient and reasonably distributed sensor data to support that location. The larger the height estimate fluctuation range, the less reliable the estimate result at that location, indicating insufficient sensor coverage or complex height variations in that area. To avoid extreme values ​​affecting subsequent calculations, the height estimate fluctuation range can be truncated in the actual implementation to, for example, between 0 mm and 20 mm. Values ​​exceeding this range are treated as boundary values, thus preventing a few outliers from raising the overall fluctuation level.

[0033] refer to Figure 1 , Figure 1 A schematic diagram of the surface reconstruction effect based on the sparse Gaussian process is shown, which illustrates how to continuously estimate the height distribution of the entire area in a local working area behind the ironing plate using limited sensor measurement data. Figure 1 A two-dimensional coordinate system is used, where the horizontal axis represents the longitudinal distance in meters (m), starting from 0 meters, corresponding to the tail edge of the screed, and extending backward along the paver's forward direction to 2.0 meters. This 2.0-meter range is the aforementioned predetermined longitudinal range. The vertical axis represents the height deviation in millimeters (mm), with values ​​ranging from -10 mm to 15 mm. Positive values ​​indicate that the actual measured height is higher than the design elevation, negative values ​​indicate that the actual measured height is lower than the design elevation, and zero corresponds to the design elevation. Figure 1 In the diagram, the solid black dots represent the actual height data points acquired by the laser displacement sensors during the acquisition period. These sensor measurement points exhibit an irregular but relatively uniform distribution in both the horizontal and vertical directions. Specifically, within a preset initial time, such as 1 second, when the sampling frequency is 100 times per second and 5 laser displacement sensors are used, 500 data records can be accumulated. As the paver continues to move forward, these data points are distributed longitudinally within the range of 0 meters to 2.0 meters; simultaneously, because the 5 laser displacement sensors are arranged laterally along the screed, these data points form multiple longitudinal trajectories at their corresponding lateral positions. Figure 1It can be observed that the height deviation values ​​of the sensor measurement points exhibit certain fluctuations. For example, within a longitudinal distance of approximately 0.2 to 0.5 meters, the height deviation of some measurement points is close to -5 mm to 0 mm; within a longitudinal distance of approximately 0.8 to 1.2 meters, the height deviation reaches a relatively high level of 5 to 10 mm; and within a longitudinal distance of approximately 1.4 to 1.8 meters, the height deviation of the measurement points falls back to near 0 mm. This fluctuation reflects the actual change in the loose paving thickness during the paving process, which may be affected by various factors such as material supply fluctuations, screed posture adjustments, and uneven roadbed.

[0034] Figure 1 Points marked with white boxes are designated as induction points. These induction points are representative locations selected from all sensor measurement points according to specific rules. In this embodiment, after dividing a predetermined longitudinal and lateral range into multiple grid cells, the center of each grid cell is used as a candidate induction point location. To maintain model accuracy while controlling computational complexity, for example, 64 induction points can be selected from all candidate induction point locations, following the principle of as even a distribution as possible in the longitudinal and lateral directions. Figure 1 As can be seen, the induced points are roughly uniformly distributed longitudinally within the range of 0 to 2.0 meters, with a longitudinal spacing of approximately 0.3 to 0.4 meters between adjacent induced points. The height deviation value corresponding to each induced point is determined based on data from nearby sensor measurement points. For example, an induced point located at a longitudinal distance of approximately 0.4 meters has a height deviation of approximately -2 millimeters; an induced point located at a longitudinal distance of approximately 1.0 meter has a height deviation of approximately 8 millimeters; and an induced point located at a longitudinal distance of approximately 2.0 meters has a height deviation of approximately 3 millimeters. These induced points constitute the basic dataset for the sparse Gaussian process surface reconstruction model. Figure 1The smooth curve drawn with a thick solid line represents the height deviation estimation surface obtained after reconstruction using a sparse Gaussian process surface. This reconstructed surface is based on a set of induced points and the corresponding height values ​​for each point. By constructing a sparse Gaussian process model and utilizing a pre-defined kernel function to describe the spatial correlation between different locations, a continuous estimation result is obtained by predicting the height of the center positions of all grid cells within the local working area. The reconstructed surface exhibits continuous and smooth characteristics, effectively describing the overall trend of the tiling thickness deviation in the vertical direction. Specifically, observing the trend of the reconstructed surface reveals the following: In the longitudinal distance range of 0 to 0.5 meters, the height deviation of the reconstructed surface gradually decreases from approximately 0 millimeters to approximately -3 millimeters, indicating that the paving thickness in this area is slightly lower than the design thickness; In the longitudinal distance range of 0.5 to 1.2 meters, the reconstructed surface rises significantly, from approximately -3 millimeters to approximately 9 millimeters, indicating a significant over-paving thickness in this area; In the longitudinal distance range of 1.2 to 1.8 meters, the reconstructed surface reaches its peak and then begins to decline, gradually falling back from approximately 9 millimeters to approximately 3 millimeters; In the longitudinal distance range of 1.8 to 2.0 meters, the reconstructed surface tends to flatten out, with the height deviation remaining at approximately 3 millimeters.

[0035] The reconstructed surface not only uses the locations directly measured by sensors but also makes reasonable inferences about locations not directly measured. For example, in the region between two adjacent sensor measurement points, the reconstructed surface provides a smooth transition estimate based on the spatial correlation assumption, avoiding the discontinuities or overly harshness that may result from simple linear interpolation. This approach allows the reconstructed surface to more realistically reflect the actual height distribution of the paved road surface. Figure 1 Two curves, drawn with thin dashed lines, are located above and below the reconstructed surface, respectively. These curves define the range of height estimation fluctuation. The upper dashed line represents the reconstructed surface plus a certain height estimation fluctuation range value, while the lower dashed line represents the reconstructed surface minus a certain height estimation fluctuation range value. The height estimation fluctuation range value quantifies the degree of uncertainty in the height estimation result at each location. Figure 1It can be observed that the height estimation fluctuation range is not constant but varies with the longitudinal position. In areas with denser sensor measurement points, such as the longitudinal distances of approximately 0.3 to 0.6 meters and approximately 1.0 to 1.3 meters, the distance between the reconstructed surface and the upper and lower fluctuation range lines is relatively small, indicating that the height estimation fluctuation range at these locations is small and the estimation results are more reliable. Conversely, in areas with relatively sparse sensor measurement points, such as the longitudinal distances of approximately 0.8 to 0.9 meters and approximately 1.6 to 1.7 meters, the distance between the reconstructed surface and the upper and lower fluctuation range lines is relatively large, indicating that the height estimation fluctuation range at these locations is large and the estimation results are relatively less reliable. For a specific example, at a longitudinal distance of approximately 1.0 meter, the height deviation estimate given by the reconstructed surface is approximately 8 millimeters, while the height estimation fluctuation range at this location is approximately 2 millimeters. Therefore, the height deviation at this location may fall within the range of 6 to 10 millimeters. At a longitudinal distance of approximately 1.6 meters, the reconstructed surface gives an estimated height deviation of about 5 millimeters, while the estimated height at this location fluctuates by about 4 millimeters. Therefore, the height deviation at this location may fall within a wider range of 1 to 9 millimeters. This difference indicates that the former estimate is more certain, while the latter has greater uncertainty.

[0036] In practical implementation, to avoid extreme values ​​affecting subsequent calculations, the height estimation fluctuation range can be truncated to, for example, between 0 mm and 20 mm. Figure 1 Looking at the fluctuation range line, the fluctuation range values ​​at all locations did not exceed the upper limit of the cutoff, indicating that the sparse Gaussian process surface reconstruction provided a reasonable uncertainty estimate in this example. By comparing... Figure 1 The sensor measurement points, induced points, reconstructed surfaces, and fluctuation ranges clearly demonstrate the technical effectiveness of the sparse Gaussian process surface reconstruction method. First, although the number of sensor measurement points is limited and their distribution is not entirely uniform, by selecting representative induced points and constructing a sparse Gaussian process model, continuous height estimation can be achieved for the entire local working area, filling data gaps in locations not directly measured. Second, the reconstructed surface does not simply connect the measurement points but considers spatial correlation, resulting in a smoother and more reasonable curve that can filter out minor abrupt changes caused by sensor noise or local particles. Third, by providing the height estimation fluctuation range value for each location, it clearly identifies which areas have more reliable estimation results and which have less certain results, providing a quantitative basis for differentiating between high-confidence and low-confidence areas in subsequent thickness deviation and flatness analysis.

[0037] After obtaining the height estimate for each grid cell, it is compared with the corresponding design elevation in the pre-stored road design model. For each grid cell, the difference between the height estimate and the design elevation is calculated, and this difference is defined as the virtual paving thickness deviation value for that grid cell. A positive virtual paving thickness deviation value indicates that the actual virtual paving thickness is greater than the design thickness, while a negative value indicates that the actual virtual paving thickness is less than the design thickness. Since the virtual paving thickness deviation near the screed's trailing edge better reflects the immediate effect of the current control command, in each sampling period, the row of grid cells closest to the screed's trailing edge in the longitudinal direction can be selected, and the virtual paving thickness deviation values ​​of this row of grid cells can be arithmetically averaged to obtain the virtual paving thickness deviation sampled value for the current sampling period. In this way, small abrupt changes caused by local particles can be filtered out to a certain extent, making the virtual paving thickness deviation sampled value more representative of the overall thickness deviation level in the current lateral range, while maintaining sensitivity to short-distance changes.

[0038] In multiple consecutive sampling periods, the virtual paving thickness deviation samples obtained in each sampling period are arranged sequentially according to the sampling time order to form a virtual paving thickness deviation sequence. This virtual paving thickness deviation sequence records the evolution of the thickness deviation of the local working area in the paving direction over time, including both slow changes during stable operation and abrupt changes caused by material supply fluctuations, mechanical posture changes, etc. In order to extract short-wavelength unevenness, medium-wavelength unevenness, and long-wavelength unevenness from the virtual paving thickness deviation sequence, multi-scale wavelet roughness calculation can be performed based on the virtual paving thickness deviation sequence.

[0039] Specifically, in each sampling period, the most recent 128 sampled values ​​of the paving thickness deviation can be extracted from the paving thickness deviation sequence and used as the current input for multi-scale wavelet roughness calculation. Using the most recent 128 sampled values ​​instead of a shorter length ensures coverage of a certain road segment while controlling the computational load, allowing the multi-scale wavelet roughness calculation to be completed within each sampling period. To balance time-domain and frequency-domain resolution, the Daubechies4 wavelet basis can be selected as the wavelet decomposition basis function, and the number of decomposition layers can be set to 3. In this configuration, the first layer of detail coefficients mainly reflects the fluctuations of shorter wavelengths, corresponding to short-wavelength irregularities on the paving surface; the second layer of detail coefficients mainly reflects the fluctuations of medium wavelengths, corresponding to medium-wavelength irregularities caused by screed posture, material supply fluctuations, etc.; and the third layer of detail coefficients reflects the trend changes of longer wavelengths, corresponding to the overall road surface undulations and slow thickness shifts.

[0040] After completing the three-level wavelet decomposition, energy statistics can be performed on the level 1, level 2, and level 3 detail coefficients respectively. Specifically, the level coefficients of each level can be squared and summed, and the square root of the summation result can be taken to obtain the energy index value corresponding to each level. A larger energy index value indicates a more severe deviation in pavement thickness at that scale. For example, a larger level 1 energy index value indicates higher short-wave roughness and more minor surface bumps; a larger level 3 energy index value indicates significant long-wave unevenness, such as a gradual thickening or thinning to one side. The level 1 energy index value is defined as the short-wave smoothness index value, the level 2 energy index value as the mid-wave smoothness index value, and the level 3 energy index value as the long-wave smoothness index value. These three index values ​​are then combined in a fixed order to form a multi-scale smoothness index vector. This method not only provides a single roughness measure but also allows for the evaluation of pavement smoothness at different scales, which is beneficial for subsequent processing of short-wave defects and long-wave trends.

[0041] Within the same sampling period, the height estimation fluctuation range values ​​of all grid cells within a local working area can also be statistically analyzed. For example, the height estimation fluctuation range values ​​of 400 grid cells can be read in the current sampling period, and the arithmetic mean of these height estimation fluctuation range values ​​can be calculated. This arithmetic mean is used as the average height estimation fluctuation range value for the current sampling period. The smaller the average height estimation fluctuation range value, the more uniform the sensor coverage distribution is over the current period, and the more reliable the reconstruction results of the sparse Gaussian process surface are. When the average height estimation fluctuation range value increases, it indicates that some areas lack sufficient sampling support, and the overall assessment confidence decreases. Adding the average height estimation fluctuation range value to the observation data can distinguish between high-confidence and low-confidence states in subsequent decision-making processes, thereby avoiding overly aggressive correction actions when data is insufficient.

[0042] Therefore, within each sampling period, the current tiling thickness deviation sample value, the multi-scale flatness index vector, and the average height estimation fluctuation range within the current sampling period can be combined to form a set of observation data. Each set of observation data simultaneously carries tiling thickness deviation information and multi-scale flatness index information, along with a quantitative description of the confidence level of the current surface reconstruction, providing a unified and well-structured input for subsequent Bayesian online change point detection and conditional risk assessment.

[0043] In another alternative implementation, the above numerical configurations can be adjusted according to different construction scenarios. For example, when the paving speed is low and the target accuracy requirement is high, the predetermined longitudinal range can be reduced to 1 meter, and the longitudinal unit length can be reduced to 0.025 meters, thereby increasing the number of grid units and enhancing the ability to resolve local details. Alternatively, the number of laser displacement sensors can be appropriately increased, for example, using 7 laser displacement sensors, to further reduce interpolation errors in the lateral direction. In multi-scale wavelet roughness calculation, the number of the most recent paving thickness deviation samples can be adjusted from 128 to 256 to obtain the smoothness statistics for longer road sections. While maintaining the overall process of sparse Gaussian process surface reconstruction and multi-scale wavelet roughness calculation unchanged, by adjusting the above specific parameters, it is possible to adapt to different structural layer thicknesses, different construction speeds, and different smoothness requirements, and to achieve accurate acquisition of paving thickness deviation information and multi-scale smoothness index information for the local working area behind the screed.

[0044] Step 2: Based on the observation data, the preset sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm is invoked. The current system state is determined by performing Bayesian online change point detection. Combined with short-term prediction and conditional risk assessment, a subthreshold correction instruction representing the target correction level is generated.

[0045] In step two, a pre-defined sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm can be invoked based on the observed data as follows: The current system state is determined by performing Bayesian online change point detection, and a subthreshold correction instruction characterizing the target correction level is generated by combining short-term prediction and conditional risk assessment. In one implementation, the paver control system receives a set of observed data from step one in each sampling cycle. This observed data includes at least the current sampled value of the paving thickness deviation, the multi-scale smoothness index vector, and the average value of the height estimation fluctuation range within the current sampling cycle. To continuously evaluate the paving state over time, an observation buffer can be maintained in the control system to store the most recent observation data in chronological order. For example, the length of the observation buffer can be preset to 200. When new observation data enters, if there are already 200 sets of data in the buffer, the oldest set of observation data is discarded, ensuring that the buffer always stores the latest 200 sets of observation data. In this way, the sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm can see the latest state each time it processes data without causing excessive computational load due to infinite data accumulation.

[0046] In the Bayesian online change point detection process, the sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm first uses binomial statistics to identify and count anomalies in the most recent observation data. For example, the most recent 50 sets of observation data can be selected from the observation buffer, and anomaly marking is performed on each set. The reason for selecting the most recent 50 sets is that, with a sampling frequency of 100 times per second, 50 sets of observation data typically cover about 0.5 seconds of the paving process. This time length is sufficient to capture rapid changes caused by material supply fluctuations or attitude changes, without being masked by long-term trends. For each set of observation data, the absolute value of the sampled value of the virtual paving thickness deviation and the average value of the height estimation fluctuation range are first read. The absolute value of the sampled value of the virtual paving thickness deviation is divided by the average value of the height estimation fluctuation range to obtain a ratio. If this ratio is significantly greater than 1, it indicates that the current virtual paving thickness deviation has exceeded the natural fluctuation level estimated by the sparse Gaussian process surface reconstruction. It can be considered that the deviation is no longer just a measurement uncertainty or an acceptable small fluctuation, but is more likely to represent a change in the construction state. Based on this consideration, the ratio threshold can be set to, for example, 2. When the calculated ratio is greater than 2, the anomaly marker of the current set of observation data is set to 1. When the calculated ratio is in the range of 0 to 2, the anomaly marker of the current set of observation data is set to 0.

[0047] In the same anomaly labeling process, both short-wave and medium-wave smoothness index values ​​from the multi-scale smoothness index vector can be considered simultaneously. When either the short-wave or medium-wave smoothness index value exceeds a pre-set smoothness threshold, even if the ratio of the sampled value of the paving thickness deviation to the average value of the height estimation fluctuation range is not high, the observed data can be directly considered anomaly. For example, the smoothness threshold can be set to 3 mm. When the short-wave or medium-wave smoothness index value is greater than 3 mm, it indicates that the short-wave or medium-wave roughness of the pavement has increased significantly in the recent period. This usually means that there is a large disturbance in material supply, paving speed, or screed posture. To capture this change as early as possible, the anomaly label of the currently judged observed data can be directly set to 1. By comprehensively using the ratio of the paving thickness deviation to the height estimation fluctuation range and the multi-scale smoothness index for anomaly labeling, the changes in thickness deviation and smoothness at different scales can be taken into account simultaneously, making the anomaly judgment more comprehensive.

[0048] refer to Figure 2This figure illustrates the variation of multi-scale smoothness index energy distribution with time window index. The horizontal axis represents the time window index, ranging from 0 to 50, and the vertical axis represents the smoothness energy index, in millimeters, ranging from 0 to 4.0 millimeters. The figure includes three curves corresponding to the short-wave smoothness index, medium-wave smoothness index, and long-wave smoothness index. The short-wave smoothness index is represented by a solid line, reflecting the intensity of short-wavelength undulations on the paved surface. This curve has a relatively large fluctuation range and can sensitively capture subtle local surface roughness changes. The medium-wave smoothness index is represented by medium-spaced dashed lines, reflecting medium-wavelength undulations caused by factors such as screed posture changes and material supply fluctuations. The value of this curve is typically between short-wave and long-wave. The long-wave smoothness index is represented by short-spaced dashed lines, reflecting the overall undulation and slow thickness shift trend of the pavement. This curve is relatively flat and mainly reflects long-term trend changes. A horizontal dashed line is also drawn in the figure, representing the smoothness threshold, set at 3 millimeters. When the smoothness energy index at any scale exceeds the threshold, the system marks the observed data as an anomaly for subsequent Bayesian online change point detection. The characteristics of three typical stages can be observed from the figure. In the initial stage (time window index 0-15), the smoothness indices at all three scales are at low levels, with the short-wave smoothness index at approximately 1.5-2.0 mm and the medium-wave and long-wave smoothness indices even lower, indicating that the paving process is in a stable state. In the middle stage (time window index 15-30), the smoothness indices at all three scales increase significantly, with the short-wave smoothness index reaching 2.5-3.3 mm and the medium-wave smoothness index reaching 2.0-2.6 mm. Some values ​​exceed the smoothness threshold of 3 mm, indicating significant disturbances in the paving process, possibly caused by uneven material feeding or changes in mechanical posture. In the later stage (time window index 30-50), the smoothness indices at all three scales fall back to low levels, indicating that the system has returned to stability after subthreshold correction. By simultaneously monitoring the smoothness energy index at three scales, the smoothness quality of the road surface can be comprehensively assessed in different wavelength ranges, providing multi-dimensional observational basis for subsequent anomaly detection and risk assessment.

[0049] After identifying anomalies in the most recent 50 sets of observations, the number of observations marked as anomaly 1 can be summed to obtain an anomaly count. To convert the anomaly count into a continuous change-point probability value, the anomaly count can be incremented by 1 and then divided by the result of the most recent preset number plus 2. For example, if the most recent preset number is 50, the anomaly count is incremented by 1 and then divided by 52 to obtain a change-point probability value between 0 and 1. The advantage of this approach is that when all observations are marked as normal, the change-point probability value will not become strictly 0, but rather a small value close to 0; when all observations are marked as anomalies, the change-point probability value will not become strictly 1, but rather a large value close to 1, thus reserving a certain buffer space numerically for subsequent state determination. Subsequently, a change-point probability threshold can be set, for example, 0.6. When the change-point probability value is in the range of 0.6 to 1.0, the current sampling period is considered to be in a change-point state; when the change-point probability value is in the range of 0 to 0.6, the current sampling period is considered to be in a stable state. In this way, as the proportion of recent abnormal observation data gradually increases, the probability value of the change point will smoothly transition from a lower value to a higher value. The control system can promptly detect whether the state is about to change or has already changed, rather than only reacting in extreme cases.

[0050] After completing Bayesian online change point detection and obtaining the current system state, the sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm continues to perform short-term predictions to estimate the potential level of illusory tiling thickness deviation over a future period. The goal of short-term prediction is to answer the question, "Given that the current deviation trend will not suddenly change, what is the possible magnitude of the illusory tiling thickness deviation over the next few sampling periods?", thus providing a predictive basis for conditional risk assessment. Specifically, a certain number of recent illusory tiling thickness deviation samples can be read from the illusory tiling thickness deviation sequence, for example, the most recent 10 samples. The arithmetic mean of these 10 samples is calculated to obtain the average illusory tiling thickness deviation. The average illusory tiling thickness deviation reflects the overall deviation level over a short period. Simultaneously, the difference between the earliest and latest samples among these 10 samples can be divided by 9 to obtain the average change in illusory tiling thickness deviation. This calculation method is equivalent to using a simple slope to approximate the trend of deviation over time in the recent period; if the deviation is increasing, the average change in illusory tiling thickness deviation is positive; if the deviation is decreasing, the average change in illusory tiling thickness deviation is negative. After obtaining the average virtual pavement thickness deviation and the average change in virtual pavement thickness deviation, a prediction time window can be set, specifying how many sampling periods the deviation will be predicted after. For example, the prediction time window length can be set to 20 prediction sampling periods. For the first prediction sampling period, the average virtual pavement thickness deviation and the average change in virtual pavement thickness deviation can be added to obtain the first predicted center deviation; for the second prediction sampling period, the average virtual pavement thickness deviation can be added to twice the average change in virtual pavement thickness deviation to obtain the second predicted center deviation; and so on, for the 20th prediction sampling period, the average virtual pavement thickness deviation can be added to 20 times the average change in virtual pavement thickness deviation to obtain the 20th predicted center deviation. Through this linear extrapolation method, an easily implemented estimate of the deviation trend within a relatively short future time range can be given without introducing a complex model, and this estimate is directly derived from the actual deviation changes in the recent period, which is more consistent with the current working conditions.

[0051] Within the same sampling period, the average height estimation fluctuation range in the current observation data can also be used to construct the prediction bias fluctuation range. For example, the average height estimation fluctuation range can be multiplied by a preset factor to obtain the prediction bias fluctuation range. The preset factor can be set to 1.5. This way, when the height estimation fluctuation range given by the sparse Gaussian process surface reconstruction is large, the prediction bias fluctuation range also increases accordingly, reflecting higher uncertainty in future biases; when the height estimation fluctuation range is small, the prediction bias fluctuation range is small, reflecting more controllable future biases. For each prediction sampling period, prediction center deviation plus 1 times the prediction bias fluctuation range, prediction center deviation plus 2 times the prediction bias fluctuation range, and prediction center deviation plus 3 times the prediction bias fluctuation range can be calculated respectively, resulting in three prediction bias candidate values. This approach considers that the prediction center deviation can be regarded as the most likely deviation, while prediction center deviation plus 1, 2, and 3 times the prediction bias fluctuation range covers different situations from common fluctuations to extreme fluctuations. By generating three prediction bias candidate values ​​for each prediction sampling period, a total of 60 prediction bias candidate values ​​can be obtained when the prediction time window length is 20 prediction sampling periods.

[0052] After completing short-term predictions, the sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm performs conditional risk assessments based on these candidate prediction deviations to form a quantitative description of the risk of future deviations exceeding construction specifications. First, the control system pre-stores permissible deviation values ​​for paving thickness, for example, set to 10 mm. This permissible deviation value comes from relevant construction specifications or engineering requirements, indicating that deviations in paving thickness within this range are considered acceptable. For each candidate prediction deviation generated from the short-term predictions, its absolute value is taken first to obtain the absolute value of the prediction deviation. Then, the absolute value of the prediction deviation is compared with the permissible deviation value for paving thickness, and the difference between the two is calculated. If the difference is less than 0, it means that the absolute value of the prediction deviation has not exceeded the permissible range, and the difference can be treated as 0, indicating no risk of exceeding the standard; if the difference is greater than or equal to 0, the difference is used as the loss value for the corresponding candidate prediction deviation, representing the severity of exceeding the specifications in the future prediction scenario. Through this approach, the loss value only reflects the portion exceeding the permissible range, allowing focus to be placed on situations that are truly likely to lead to rework or a decline in construction quality.

[0053] To simultaneously consider both the magnitude and probability of future deviation risks, all loss values ​​can be sorted from largest to smallest, and the top 10 loss values ​​in the sorted results can be averaged. For example, when the prediction time window length is 20 and 3 candidate prediction deviation values ​​are generated per prediction sampling period, there are a total of 60 loss values. The top 10 loss values ​​in the sorted results can be selected, and their arithmetic mean can be calculated. This average is used as the conditional risk measure for the current sampling period. The reason for selecting the top few largest loss values ​​in the sorted results is that these loss values ​​represent the most unfavorable scenarios in the future. By averaging these scenarios, a risk index that comprehensively considers the "worst-case scenarios" can be obtained. If the conditional risk measure is large, it indicates that some prediction scenarios will produce significant thickness overshoot deviations in the future, and the impact of these scenarios cannot be ignored. In this case, it is necessary to perform sub-threshold corrections in advance. If the conditional risk measure is close to 0, it indicates that most prediction scenarios are within the allowable range, and the current control strategy can be maintained.

[0054] After obtaining the conditional risk measurement, different conditional risk trigger thresholds can be set depending on whether the current system is in a stable or changing state. For example, when Bayesian online change point detection determines that the current sampling period is in a stable state, the conditional risk trigger threshold can be set to the first trigger threshold, such as 6 mm; when the current sampling period is determined to be in a changing state, the conditional risk trigger threshold can be set to the second trigger threshold, such as 4 mm, and the first trigger threshold is greater than the second trigger threshold. This design allows the system to have a higher tolerance for conditional risk measurements in a stable state, triggering sub-threshold corrections only when the risk of exceeding the limit increases significantly in the future, thus avoiding frequent adjustments that affect construction efficiency; while in a changing state, the system has already detected many recent anomalies, so setting the conditional risk trigger threshold lower allows for earlier issuance of sub-threshold correction commands to prevent deviations from accumulating and developing into serious problems requiring downtime and rework.

[0055] When generating subthreshold correction instructions, the subthreshold correction instruction level for the current sampling period can be set to level 0, 1, 2, or 3 based on the different numerical ranges of the conditional risk metric. For example, when the conditional risk metric is in the range of 0 to 4 mm, the subthreshold correction instruction level can be set to level 0, corresponding to no subthreshold correction; when the conditional risk metric is in the range of 4 to 6 mm, the subthreshold correction instruction level can be set to level 1, corresponding to a small correction in piec thickness; when the conditional risk metric is in the range of 6 to 8 mm, the subthreshold correction instruction level can be set to level 2, corresponding to a medium correction; and when the conditional risk metric is greater than or equal to 8 mm, the subthreshold correction instruction level can be set to level 3, corresponding to a large correction. Simultaneously with generating the subthreshold correction instruction level, the target piec thickness correction amount corresponding to each level can be pre-set to progressively increasing values, for example, level 1 corresponds to 1 mm, level 2 to 2 mm, and level 3 to 3 mm. In this way, the larger the conditional risk metric, the larger the corresponding target piec thickness correction amount, thus achieving a direct correspondence between risk magnitude and correction intensity. Finally, the subthreshold correction command level and the corresponding target dummy thickness correction amount are provided to the control execution module for subsequent specific adjustments of the actuator.

[0056] In another alternative implementation, the aforementioned numerical parameters can be adjusted to better suit different construction conditions. For example, when the sampling frequency is low, the most recent preset quantity can be adjusted from 50 to 30 to keep the time window length for Bayesian online variable point detection close to 0.5 seconds. When the road design is particularly sensitive to thickness deviation, the allowable deviation value for paving thickness can be reduced from 10 mm to 8 mm, while the first and second trigger thresholds are correspondingly lowered, so that the system issues a sub-threshold correction command when the deviation is just close to the upper limit of the specification. In the calculation of conditional risk measurement, the number of loss values ​​participating in the average can also be changed, for example, from the first 10 loss values ​​to the first 8 loss values, so that the conditional risk measurement focuses more on the most extreme prediction scenarios. While keeping the overall process of the sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm unchanged, by adjusting these parameters, the sensitivity and stability of the sub-threshold correction command generation can be further improved, thereby better meeting the requirements for paving thickness and smoothness control in different engineering scenarios.

[0057] Step 3: Based on the subthreshold correction command and combined with the Lyapunov function approximation evaluation of the system stability state, the actuator performs the correction action corresponding to the subthreshold correction command through the synergistic effect of the adaptively adjusted thickness controller and the soft dead zone control module.

[0058] In step three, the actuator can perform the correction action corresponding to the subthreshold correction command by means of the subthreshold correction command, based on the subthreshold correction command and combined with the Lyapunov function approximation evaluation of the system stability state, through the synergistic effect of the adaptively adjusted thickness controller and the soft dead zone control module.

[0059] In one implementation, a thickness controller comprising proportional and integral elements is configured in the paver control system. The output of the thickness controller is connected to the actuator that raises or lowers the screed. The input to the thickness controller is a composite of the sampled value of the paving thickness deviation and the target paving thickness correction amount corresponding to the sub-threshold correction command level. The output of the thickness controller is a control quantity used to control the actuator's movement. In this way, the thickness controller can perform conventional tracking control based on the current sampled value of the paving thickness deviation, and can also add a small-amplitude offset based on the sub-threshold correction command level generated in step two, thereby correcting the deviation in advance before it exceeds the upper limit of the construction specifications.

[0060] To approximate the system's stability using the Lyapunov function, a predetermined number of recent dummy thickness deviation samples are read from the dummy thickness deviation sequence within each sampling period, for example, the most recent 200 samples. The squares of these 200 samples are then calculated, and their arithmetic mean is taken to obtain the first stability index. A larger first stability index indicates a larger overall amplitude of the dummy thickness deviation over a recent period, which can be interpreted as higher dummy thickness error energy. A smaller first stability index indicates that the overall dummy thickness deviation is within a smaller range, and the system is close to the desired state. To observe the trend of error energy over time, the first stability index of the current sampling period can be subtracted from the first stability index of the previous sampling period to obtain the second stability index. When the second stability index is negative, it indicates that the first stability index is gradually decreasing, the dummy thickness error energy is gradually decaying, and the system is converging towards stability. When the second stability index is positive, it indicates that the dummy thickness error energy is increasing, and the system is trending towards instability. By combining the first and second stability indices, the magnitude and variation of the Lyapunov function can be approximated without explicitly constructing the mathematical form, thereby identifying whether the system is in a convergent state or a state of amplified deviation.

[0061] In practical implementation, a preset stability threshold range can be used to determine whether the magnitude of the first stability index is within an acceptable range. For example, the stability threshold can be set to the square of 4 mm. When the first stability index is within the range of 0 to 4, the overall deviation of the paving thickness is considered to be small in the recent period. For the second stability index, negative and positive ranges can be set. For example, the negative range can be set to −0.1 to 0, and the positive range can be set to 0 to 0.1. When the second stability index falls into the negative range, it indicates that the error energy is slowly decreasing; when it falls into the positive range, it indicates that the error energy is slowly increasing; and when it exceeds the above range, it indicates that the error energy is changing drastically. By checking the first and second stability indices simultaneously, it is possible to determine whether the current control state is a stable state or a deviation amplification state in each sampling period. For example, when the first stability index is within the range of 0 to 4 and the second stability index is within the range of −0.1 to 0, the current control state can be considered a stable state; when the first stability index is greater than 4 and the second stability index is within the range of 0 to 0.1, the current control state can be considered a deviation amplification state.

[0062] After assessing the system's stability, the amplification factors of the proportional and integral components of the thickness controller can be adaptively adjusted according to a preset adjustment period. For example, the preset adjustment period can be set to 1 second. For each adjustment period, if the system is currently in a stable state, the amplification factors of the proportional and integral components are multiplied by a first adjustment factor, which can be set to 0.95. Multiplying by 0.95 each time gradually reduces the control gain, making the controller output change more smoothly. This approach is based on the consideration that when the system is already in a stable state and the error energy is still slowly decreasing, maintaining a large control gain may amplify sensor noise and minor road surface irregularities, causing the actuator to make frequent small movements, introducing unnecessary vibration and mechanical wear. By gradually reducing the control gain in a stable state, the system can minimize high-frequency movements while meeting the flatness and thickness requirements, thereby extending the life of the actuator and improving construction comfort. To prevent the control gain from being repeatedly adjusted to be too small, a lower limit can be set. For example, the amplification factor of the proportional and integral components should not be lower than 0.5 of the initial amplification factor. When this lower limit is reached, even if the system remains in a stable state, the amplification factor will not be reduced further.

[0063] Conversely, when the system is in a state of amplified deviation, the amplification factors of the proportional and integral components can be multiplied by a second adjustment factor, which can be set to 1.05, according to a preset adjustment cycle. Multiplying by 1.05 each time gradually increases the control gain, making the thickness controller more sensitive to changes in the sampled values ​​of the paved thickness deviation. This approach is based on the consideration that when the first stability index is large and the second stability index is positive, it indicates that the overall paved thickness deviation is large and still increasing. If a smaller control gain is used, the controller's response to the deviation will be too slow, and the thickness error may not be effectively suppressed for a long time, even failing to quickly return to a safe range after reaching the limit value of the construction specifications. By gradually increasing the control gain during the amplified deviation state, the system can more actively correct the thickness deviation, allowing the sampled values ​​of the paved thickness deviation to fall back to an acceptable range as quickly as possible. To prevent excessive increases in control gain that could cause oscillations, an upper limit can be set, for example, the amplification factors of the proportional and integral components should not exceed 1.5 times the initial amplification factor. When this upper limit is reached, even if the system continues to be in a state of amplified deviation, the amplification factor will not be increased further.

[0064] Within the thickness controller, the proportional element directly outputs a control quantity proportional to the current input, while the integral element outputs a control quantity based on the cumulative trend of the input over time. Since the input includes the sampled value of the paved thickness deviation and the target paved thickness correction corresponding to the sub-threshold correction command level, when the sub-threshold correction command level is 0, the thickness controller primarily responds to the sampled value of the paved thickness deviation. When the sub-threshold correction command level is 1, 2, or 3, the thickness controller adds a corresponding target paved thickness correction offset for the same sampled paved thickness deviation, slightly biasing the controller's output target value towards mitigating future risks. Because this offset is only a small adjustment within the range of 1 to 3 millimeters, it does not disrupt the overall stability of thickness control under normal construction conditions. Instead, when a significant risk of exceeding the limit is detected in the future, small offsets over multiple sampling cycles gradually pull the paved thickness deviation back to a safer range.

[0065] To reduce frequent actuator movements under small deviations, a soft dead zone control module is configured in the actuator control channel to perform nonlinear mapping on the thickness controller output. In one implementation, the control interval can be divided based on the absolute value of the current sampled virtual pavement thickness deviation. For example, the first dead zone threshold can be set to 1 mm, and the second dead zone threshold to 3 mm. When the absolute value of the virtual pavement thickness deviation is less than 1 mm, its impact on the final pavement quality can be considered negligible. If the small output of the thickness controller is allowed to directly act on the actuator, it will cause the actuator to repeatedly make up and down adjustments under extremely small deviations, increasing mechanical fatigue and potentially generating fine ripples in local areas. Therefore, when the absolute value of the virtual pavement thickness deviation is less than the first dead zone threshold, the output of the thickness controller can be directly set to 0 after processing by the soft dead zone control module, keeping the actuator stationary within this deviation range.

[0066] When the absolute value of the virtual thickness deviation is between the first and second dead zone thresholds, a linear proportional method can be used to continuously map the thickness controller output between 0 and the original output. Specifically, when the absolute value of the virtual thickness deviation equals the first dead zone threshold, the mapped output is 0; when the absolute value of the virtual thickness deviation equals the second dead zone threshold, the mapped output equals the original output of the thickness controller; when the absolute value of the virtual thickness deviation is between the two, the mapped output gradually increases proportionally between 0 and the original output. The effect of this is that, in the stage of slight deviation, the actuator's movement amplitude is significantly compressed, avoiding too many small adjustments in a short period; as the deviation gradually increases, the actuator's movement amplitude also increases linearly, allowing the system to smoothly transition from a state of "minimizing movement" to one of "requiring active correction." Through this soft dead zone handling method, the number of mechanical movements under small deviations can be significantly reduced while ensuring a rapid response to large deviations.

[0067] When the absolute value of the loose pavement thickness deviation is greater than or equal to the second dead zone threshold, it indicates that the current deviation has reached a significant level. To prevent the deviation from continuing to increase, the output of the thickness controller can be directly transmitted to the actuator after processing by the soft dead zone control module, without further compression. In this way, when the pavement thickness is significantly too thick or too thin, the actuator can quickly perform lifting or lowering operations according to the calculation results of the thickness controller, and bring the sampled value of the loose pavement thickness deviation back within the second dead zone threshold as soon as possible.

[0068] Based on the aforementioned soft dead-zone control, the target paving thickness correction amount corresponding to the sub-threshold correction command level is superimposed on the thickness controller input. For example, when the sub-threshold correction command level is 1, a 1 mm paving thickness target offset can be added to the thickness controller input, causing the thickness controller to exhibit a slightly stronger upward or downward trend over a period of time; when the sub-threshold correction command level is 2, a 2 mm paving thickness target offset is added; and when the sub-threshold correction command level is 3, a 3 mm paving thickness target offset is added. Because soft dead-zone control compresses the control quantity under small deviations, even under high-level sub-threshold correction commands, the actuator's action remains relatively mild when the paving thickness deviation is still within a small range, and the controller output is only fully released when the deviation reaches a certain magnitude. Therefore, the adaptive thickness controller, soft dead zone control, and subthreshold correction command form a mutually restraining relationship: when the system is stable, the control gain is gradually reduced through approximate evaluation using the Lyapunov function, and the control gain is gradually increased when the deviation is amplified; unnecessary actions are suppressed through soft dead zone control when the deviation is small, and the control response is amplified when the deviation is large; when there is a large future risk, a small target offset is applied in advance through the subthreshold correction command to guide the system's operating state to a safer region.

[0069] refer to Figure 3 , Figure 3The figure shows the time-series distribution curve of the subthreshold correction command level as a function of the sampling period. The graph uses a stepped curve, with the horizontal axis representing the sampling period number, ranging from 0 to 160, and the vertical axis representing the subthreshold correction command level, divided into four discrete levels: level 0, level 1, level 2, and level 3. On the vertical axis, each level corresponds to a different target paving thickness correction amount. Level 0 indicates no subthreshold correction is performed, with a target paving thickness correction of 0 mm. In this case, the control system only performs routine tracking control based on the current paving thickness deviation sample value, without additional correction bias. Level 1 corresponds to a target paving thickness correction of 1 mm, indicating the need for a small-scale pre-correction. Level 2 corresponds to a target paving thickness correction of 2 mm, indicating the need for a medium-scale correction. Level 3 corresponds to a target paving thickness correction of 3 mm, indicating the need for a large-scale correction to address a higher risk of future exceedances. The graph uses different depths of shading for each level region, with deeper shading for higher levels, visually reflecting the increasing relationship of correction intensity. From a temporal evolution perspective, the curve can be divided into four typical stages. The first stage corresponds to sampling period numbers 0 to 30, during which the sub-threshold correction command level remains at level 0. This indicates that during the initial stage of paving or normal stable operation, the conditional risk metric is at a low level, and the risk of the paving thickness deviation exceeding the allowable value of the construction specification is small in the future. The system does not need to execute sub-threshold correction, and the paving quality can be maintained solely by the conventional thickness controller. The second stage corresponds to sampling period numbers 30 to 50, where the sub-threshold correction command level begins to fluctuate between level 0 and level 1. This fluctuation reflects that the conditional risk metric is gradually approaching the first or second trigger threshold, indicating that short-term predictions suggest a possible slight exceedance of the thickness deviation in the future. In a stable state, a level 1 correction is triggered when the conditional risk metric exceeds 6 mm; in a changing state, a level 1 correction is triggered when the conditional risk metric exceeds 4 mm. The intermittent level 1 correction command in this stage causes the thickness controller to add a target offset of 1 mm to the original control value, making small adjustments in advance to the potential deviation trend and preventing the deviation from accumulating further. The third stage corresponds to sampling period numbers 50 to 90. The subthreshold correction command level frequently changes between levels 1, 2, and 3, with an overall trend towards higher levels. This indicates that during this period, Bayesian online change point detection identified more abnormal observation data, suggesting the system may be in a change point state. Simultaneously, the conditional risk metric increases significantly, with a considerable proportion of predicted future thickness deviation candidate values ​​exceeding the allowable range. When the conditional risk metric is in the range of 6 to 8 mm, the command level is set to level 2; when the conditional risk metric is greater than or equal to 8 mm, the command level is raised to level 3. Frequent high-level correction commands reflect instability in the paving state caused by factors such as material supply fluctuations, paving speed changes, or screed attitude disturbances, requiring substantial subthreshold corrections to suppress deviation development.The fourth stage corresponds to sampling period numbers 90 to 120. The subthreshold correction command level mainly remains at level 2 and 3, occasionally briefly dropping to level 1. This stage is a high-risk period, with the conditional risk metric remaining at a high level, indicating that the accumulated deviations or persistent external disturbances have not been completely eliminated, and the system still needs to maintain a strong correction force. By continuously applying a target paving thickness correction of 2 to 3 mm, combined with the adaptive thickness controller gain increase and the flexible response of the soft dead zone control, the paving thickness deviation is gradually pulled back to the safe range. The fifth stage corresponds to sampling period numbers 120 to 160. The subthreshold correction command level gradually decreases, mainly distributed between level 0 and 1, occasionally appearing at level 2. This indicates that after the previous active correction, the conditional risk metric begins to decrease, the future risk of exceeding the standard is effectively controlled, and the system re-enters a relatively stable operating state. At this time, although there is still a small need for correction, the overall correction intensity is significantly weakened, and the paving quality tends to normalize.

[0070] In another alternative implementation, the above numerical settings can be adjusted according to the construction object and equipment characteristics. For example, when the construction object is an ultra-thin surface layer, it is more sensitive to thickness deviations. The first dead zone threshold can be reduced from 1 mm to 0.5 mm, and the second dead zone threshold from 3 mm to 2 mm, so that the actuator can respond more actively to the same magnitude of paving thickness deviation. At the same time, the stability threshold and the first adjustment factor can be appropriately reduced so that the system starts to reduce the control gain when the error decreases slightly, thus avoiding surface undulations caused by high gain on ultra-thin layers. When the construction object is a thicker base layer, the stability threshold and the second dead zone threshold can be appropriately increased so that the system maintains stable operation under larger deviations, avoiding frequent adjustments to the base layer structure due to oversensitivity. While maintaining the overall process of "based on subthreshold correction commands, combined with the Lyapunov function approximation of the system stability state, and through the synergistic effect of the adaptive thickness controller and the soft dead zone control module, the actuator can perform the correction action corresponding to the subthreshold correction command," by adjusting the above thresholds and factors, this step can be adapted to paving conditions with different levels, thicknesses, and construction accuracy requirements.

[0071] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for real-time in-machine evaluation and sub-threshold correction of paving thickness and flatness, characterized in that, The method comprises the following steps: Step one: obtaining sensor height data of a local working area at the rear of the screed, performing sparse Gaussian process surface reconstruction on the local working area corresponding to the obtained sensor height data, and performing multi-scale wavelet roughness calculation based on the reconstruction result to generate observation data containing virtual paving thickness deviation information and multi-scale flatness index information; Step two: based on the observation data, calling a preset sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm, determining the current system state by performing Bayesian online change point detection, and generating a subthreshold correction instruction representing the target correction level in combination with short-term prediction and conditional risk assessment; Step three: based on the subthreshold correction instruction and in combination with Lyapunov function approximation evaluation of the system stability state, the thickness controller and the soft dead zone control module are cooperatively used to make the actuator execute the correction action corresponding to the subthreshold correction instruction; The sparse Gaussian process wavelet Bayesian conditional risk assessment algorithm in step two further comprises a short-term prediction sub-step, which comprises: reading the latest preset number of virtual paving thickness deviation sample values from the virtual paving thickness deviation sequence, calculating the arithmetic mean of the virtual paving thickness deviation sample values as the average virtual paving thickness deviation, and calculating the difference between the earliest sample value and the latest sample value in the virtual paving thickness deviation sample values, and then dividing by the latest preset number minus 1 to obtain the average change of the virtual paving thickness deviation; the prediction time window length is set to the preset prediction sample period number, and for the kth prediction sample period, the average virtual paving thickness deviation and the k times average change of the virtual paving thickness deviation are added to obtain the kth prediction center deviation, until the prediction center deviations of all prediction sample periods are obtained; k starts from 1; in the same sample period, the height estimation fluctuation range average value is read from the current observation data, and the height estimation fluctuation range average value is multiplied by a preset multiple to obtain the prediction deviation fluctuation range; for each prediction sample period, the prediction center deviation plus 1 times the prediction deviation fluctuation range, the prediction center deviation plus 2 times the prediction deviation fluctuation range, and the prediction center deviation plus 3 times the prediction deviation fluctuation range are calculated respectively to obtain 3 prediction deviation candidate values, and 3 times the prediction sample period number prediction deviation candidate values are obtained in total; The sparse Gaussian process wavelet Bayesian conditional risk evaluation algorithm in step two further comprises a threshold correction trigger sub-step based on the conditional risk value, which comprises: storing a preset paving thickness allowable deviation value in the control system in advance; taking the absolute value of each prediction deviation candidate value generated by the short-term prediction sub-step to obtain a plurality of prediction deviation absolute values; for each prediction deviation absolute value, calculating the difference between the prediction deviation absolute value and the preset paving thickness allowable deviation value, and when the difference is less than 0, processing the difference as 0, and when the difference is greater than or equal to 0, taking the difference as a loss value of the prediction deviation candidate value; sorting all loss values from large to small, selecting the first preset number of loss values in the sorting result, calculating the arithmetic mean of the first preset number of loss values to obtain the conditional risk measure of the current sampling period; setting the conditional risk trigger threshold to a first trigger threshold when in a stable state and to a second trigger threshold when in a variable point state, the first trigger threshold being greater than the second trigger threshold; according to different numerical intervals of the conditional risk measure, setting the threshold correction instruction level of the current sampling period to level 0, level 1, level 2 or level 3, and outputting the threshold correction instruction level and the corresponding target virtual paving thickness correction amount to the control execution module, wherein the level 1, level 2 and level 3 threshold correction instructions correspond to target virtual paving thickness correction amounts that increase step by step.

2. The method of claim 1, wherein, In step one, the operation of obtaining sensor height data and constructing a local working area comprises: fixedly installing a plurality of laser displacement sensors at predetermined lateral intervals along the transverse direction of the screed, and setting a preset sampling frequency; in each sampling period, collecting height data and corresponding lateral positions of the plurality of laser displacement sensors; establishing a device coordinate system with the center of the trailing edge of the screed as the origin, and projecting the height data within a predetermined longitudinal range immediately behind the trailing edge of the screed to the device coordinate system; dividing the predetermined longitudinal range into a plurality of equal-length units along the longitudinal direction, and dividing the predetermined lateral range into a plurality of equal-width units along the lateral direction, thereby forming a local working area comprising a plurality of grid units.

3. The method of claim 2, wherein, The sparse Gaussian process surface reconstruction in step one comprises: continuously collecting height data within a preset initial time after paving begins, and cumulatively obtaining at least the total number of grid units in the local working area in terms of lateral position, longitudinal position and height value data; taking the center position of each grid unit where the data is located as a candidate induced point position, and selecting a preset number of induced points therefrom according to the principle of uniform distribution in the longitudinal and lateral directions; constructing a sparse Gaussian process surface reconstruction model, adopting a preset kernel function and Gaussian noise assumption, and solving the initial sparse Gaussian process surface reconstruction model by performing matrix decomposition on the covariance matrix between the induced points.

4. The method of claim 3, wherein, The sparse Gaussian process surface reconstruction model in step one is updated in each subsequent sampling period, and the updating operation includes: mapping the currently collected height data to the corresponding grid cell center positions, calculating the Euclidean distance between the grid cell center positions and a preset number of induced point positions; when the minimum distance of a certain data point is greater than a preset distance threshold, the data point position and the corresponding height value are added to the candidate induced point set; when the number of points in the candidate induced point set reaches the preset number, the candidate induced points in the sorting result are sorted from high to low according to the number of times the candidate induced points are hit by height data in the last preset time period, and the candidate induced points located in the front preset number are selected to replace the original induced point set; and the covariance matrix of the new induced point set is recalculated and matrix decomposition is performed once to update the sparse Gaussian process surface reconstruction model.

5. The method of claim 2, wherein, In step one, the operation of generating observation data includes: in each sampling period, using the latest sparse Gaussian process surface reconstruction model to predict the height of all grid cell center positions in the local working area, outputting a height estimate value and a height estimate fluctuation range value for each grid cell, wherein the height estimate fluctuation range value is obtained by square root of the diagonal element of the Gaussian process covariance matrix and is truncated between the preset fluctuation range lower limit and upper limit value; the height estimate value of each grid cell is subtracted from the corresponding design elevation in the pre-stored road design model to obtain the virtual paving thickness deviation value of the grid cell; in each sampling period, select a row of grid cells closest to the trailing edge of the screed in the longitudinal direction, and take the arithmetic mean of the virtual paving thickness deviation values of the row of grid cells to obtain the virtual paving thickness deviation sampling value of the current sampling period; arrange the virtual paving thickness deviation sampling values obtained in the continuous sampling periods in the order of sampling time to form a virtual paving thickness deviation sequence; perform multi-scale wavelet decomposition on the last preset number of sampling values in the virtual paving thickness deviation sequence, use a preset wavelet basis, set the decomposition level to a preset level, and obtain the 1st layer detail coefficient, the 2nd layer detail coefficient and the 3rd layer detail coefficient respectively; calculate the coefficient square sum and square root of each layer of detail coefficients respectively to obtain the short wave flatness index value, the medium wave flatness index value and the long wave flatness index value in turn, and combine the short wave flatness index value, the medium wave flatness index value and the long wave flatness index value into a multi-scale flatness index vector in a preset order; and combine the current virtual paving thickness deviation sampling value, the multi-scale flatness index vector, and the arithmetic mean of the height estimate fluctuation range of all grid cells in the current sampling period as a group of observation data.

6. The method of claim 1, wherein, In step two, the sparse Gaussian process wavelet Bayesian conditional risk evaluation algorithm comprises: setting the length of the observation buffer to a preset value, storing a group of observation data obtained in each sampling period in the observation buffer, and discarding the earliest group of observation data in time sequence when the observation data in the observation buffer exceeds the set observation buffer length preset value; in each sampling period, performing binomial statistical Bayesian online change point detection, which comprises: selecting the latest preset number of observation data from the observation buffer, and performing abnormal label determination on each group of observation data; the abnormal label determination process is: reading the absolute value of the virtual paving thickness deviation sampling value and the average value of the height estimation fluctuation range from a group of observation data, calculating the ratio of the two, setting the abnormal label of the observation data in the group to 1 when the calculated ratio is greater than a preset ratio threshold, setting the abnormal label of the observation data in the group to 0 when the calculated ratio is within the range of 0 to the preset ratio threshold, simultaneously reading the short wave flatness index value and the medium wave flatness index value from the observation data in the group, and setting the abnormal label of the observation data in the current determination directly to 1 when any one of the short wave flatness index value or the medium wave flatness index value is greater than a preset flatness threshold; summing the number of observation data with abnormal label 1 in the latest preset number of observation data to obtain an abnormal count value; increasing the abnormal count value by 1 and then dividing by the latest preset number plus 2 to obtain a change point probability value of the current sampling period; when the change point probability value is within the range of a preset change point probability threshold to 1.0, it is determined that the current sampling period is in a change point state; when the change point probability value is within the range of 0 to the preset change point probability threshold, it is determined that the current sampling period is in a stable state.

7. The method of claim 1, wherein, The approximate evaluation of the Lyapunov function of the system stability state and the self-adaptive thickness controller in step three comprises: configuring a thickness controller comprising a proportional link and an integral link in a paver control system, and configuring an actuator for executing the lifting or lowering of the screed; in each sampling period, reading the latest preset number of virtual paving thickness deviation sample values from the virtual paving thickness deviation sequence, calculating the square average of the sample values to obtain a first stability index, and obtaining a second stability index by subtracting the first stability index of the previous sampling period from the current first stability index; when the first stability index is within a preset stability threshold range and the second stability index is within a preset negative value interval range, the current control state is identified as a stable state, and in the stable state, the amplification multiples of the proportional link and the integral link in the thickness controller are multiplied by a preset first adjustment factor, which is less than 1, at a preset adjustment period, until the amplification multiples are reduced to a preset lower limit of the initial amplification multiples; when the first stability index is greater than the preset stability threshold and the second stability index is within a preset positive value interval range, the current control state is identified as a deviation amplification state, and in the deviation amplification state, the amplification multiples of the proportional link and the integral link in the thickness controller are multiplied by a preset second adjustment factor, which is greater than 1, at a preset adjustment period, until the amplification multiples are increased to a preset upper limit of the initial amplification multiples.

8. The method of claim 7, wherein, The synergistic effect of the soft dead zone control module in step three comprises: configuring a soft dead zone control module in the actuator control channel; in the soft dead zone control module, control intervals are divided according to the absolute value of the current virtual paving thickness deviation sample value: when the virtual paving thickness deviation absolute value is less than the first dead zone threshold, the output of the thickness controller is set to 0 after being processed by the soft dead zone control module; when the virtual paving thickness deviation absolute value is within the range of the first dead zone threshold to the second dead zone threshold, the thickness controller output is continuously mapped between 0 and the original output in a linear proportion according to the virtual paving thickness deviation absolute value between the two dead zone thresholds; when the virtual paving thickness deviation absolute value is greater than or equal to the second dead zone threshold, the thickness controller output is directly transmitted to the actuator; on the basis of the soft dead zone control, the target virtual paving thickness correction amount corresponding to the threshold correction instruction level is superimposed into the input end of the thickness controller, so that the actuator executes the threshold correction instruction in a hierarchical, smooth and damped manner.

Images