Method for measuring thickness of hot-in-place recycled pavement based on intelligent sensing technology

Abstract

The application discloses a thickness measurement method for hot in-place recycling pavement based on intelligent sensing technology, which comprises the following steps: laying multiple-source sensing units to collect construction process data, executing preprocessing to generate a construction state data set; extracting temperature variation and temperature transmission characteristics to generate a multi-source pavement state feature sequence; constructing a thickness evolution graph and establishing a multi-class connection relationship to identify a dominant evolution path sequence; synchronously inputting the multi-source features and stage information into an improved model, embedding constraints and inverting layer output thickness preliminary measurement sequence; constructing a thickness residual field to calculate a deviation amount and propagate, executing reverse compensation to generate a correction sequence; smoothing and consistency checking the thickness correction sequence, outputting continuous thickness results and generating an abnormal section record. Through multi-source intelligent sensing data fusion and improved physical consistency constraints of the NBEATSx model, the application realizes real-time, continuous and non-destructive high-precision measurement of the pavement thickness in the hot in-place recycling construction process.

Description

Technical Field

[0001] This invention relates to the field of road maintenance construction and testing technology, and in particular to a method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology. Background Technology

[0002] In-situ thermal recycling technology is widely used in asphalt pavement maintenance. It involves heating, loosening, recycling, paving, and compacting the existing pavement to reuse the old asphalt mixture on-site. The thickness of the recycled layer directly affects the pavement's structural load-bearing capacity, recycling uniformity, and long-term service life. Therefore, accurately controlling the recycled layer thickness during construction is a crucial step in ensuring the quality of in-situ thermal recycling.

[0003] Current methods for detecting the thickness of recycled layers mostly rely on post-construction core sampling, manual measurement, or a small number of single-point tests. These methods can usually only obtain thickness results for local locations. The detection process is delayed and destructive, and cannot reflect the thickness changes along the construction direction in real time during the continuous operation of the hot recycling equipment. It is also difficult to detect problems such as local recycled layers being too shallow, too deep, or having sudden thickness changes in a timely manner.

[0004] During in-situ thermal recycling, pavement materials undergo various state changes, including heating and softening, loosening and disturbance, material migration, paving and reorganization, and compaction. Single-sensor data is easily affected by temperature transmission, material migration, compaction status, and echo interface changes, leading to discontinuous and unstable thickness measurement results. Existing technologies lack a method that can combine multi-source state characteristics, thickness evolution process, and thermal conduction consistency constraints to perform real-time continuous measurement and dynamic correction of recycled layer thickness.

[0005] Therefore, how to provide a method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology. This invention fully utilizes multi-source state sensing technology, thickness evolution analysis methods, and an improved NBEAT's x model, combined with temperature-thickness coupling inversion, heat conduction consistency constraints, and error backtracking compensation mechanisms, to achieve real-time continuous measurement and dynamic correction of the recycled layer thickness during in-situ thermal recycling construction. It has the advantages of high measurement accuracy, good continuity, strong anti-interference ability, non-destructive testing, and high construction adaptability.

[0007] The method for measuring the thickness of in-situ thermal regeneration pavement based on intelligent sensing technology according to embodiments of the present invention includes:

[0008] Multi-source state sensing units are deployed on the on-site thermal recycling construction equipment to collect multi-source state data during the road heating, loosening, paving and compaction processes, and perform preprocessing to generate a construction state data set.

[0009] Based on the construction status data set, temperature change features and temperature transmission features are extracted, and displacement, depth, pressure and echo response features are also extracted to generate a multi-source pavement status feature sequence.

[0010] Based on the multi-source pavement state feature sequence, a thickness evolution map of the recycled layer is established. The pavement state at each sampling time is taken as the evolution node, and temperature transfer behavior, material migration behavior and compaction behavior are taken as the connection relationship to generate the thickness evolution path. The dominant evolution path is identified and the corresponding thickness evolution state sequence is generated.

[0011] The multi-source pavement state feature sequence, construction stage identification sequence and temperature transfer feature sequence are simultaneously input into the improved NBEATSx model with a differentiable heat conduction consistency constraint layer. A temperature-thickness coupling inversion layer is inserted after each level of neural base expansion block to output the initial thickness measurement sequence that meets the heat conduction consistency condition.

[0012] A thickness residual field is constructed based on the initial thickness measurement sequence. The deviation between each sampling point and the thickness evolution state sequence is calculated. The error backtracking relationship is formed by propagating along the construction path. Error decomposition is performed on abnormal sampling points. The initial thickness measurement sequence is then compensated in reverse to generate a thickness correction sequence.

[0013] The thickness correction sequence is smoothed and its consistency is checked. The pavement thickness measurement results distributed along the construction direction are output, and the corresponding thickness anomaly section records are generated.

[0014] Optionally, the multi-source state sensing unit includes a temperature sensor, a displacement sensor, a depth sensor, a pressure sensor, and an echo detection sensor.

[0015] Optionally, the multi-source state data includes road surface temperature data, displacement data, depth data, pressure data, and echo response data.

[0016] Optionally, generating the multi-source road surface state feature sequence includes:

[0017] Road surface temperature data is read from the construction status dataset, and the road surface temperature data is arranged according to the sampling time and construction mileage location to generate a basic temperature sequence;

[0018] Based on the temperature base sequence, extract the temperature change amplitude, temperature change direction, temperature decay state, and temperature difference between different sampling locations between adjacent sampling times to generate a temperature change feature set.

[0019] Based on the temperature change feature set, the heat transfer state between the road surface layer and the regenerated layer is identified, and the temperature transfer direction, temperature transfer intensity and temperature attenuation continuity are extracted to generate a temperature transfer feature sequence.

[0020] Displacement data, depth data, pressure data, and echo response data are read from the construction status data set and aligned with the sampling time and construction mileage location corresponding to the temperature transfer characteristic sequence to generate a structural response characteristic set.

[0021] The temperature change feature set, temperature transfer feature sequence, and structural response feature set are combined and sorted according to sampling time, construction mileage location, and construction stage identifier, and a corresponding feature record is established for each sampling location to generate a multi-source pavement state feature sequence.

[0022] Optionally, the step of generating a thickness evolution path, identifying an active chemical path, and generating a corresponding thickness evolution state sequence includes:

[0023] Read the temperature transfer characteristics, displacement characteristics, depth characteristics, pressure characteristics, echo response characteristics and construction stage identifiers corresponding to each sampling time in the multi-source pavement state characteristic sequence, and establish pavement state nodes according to the sampling time and construction mileage location;

[0024] The temperature transfer state, material position change state, compression forming state and echo interface state corresponding to each road surface state node are written into the node attributes to generate an evolution node set used to characterize the formation process of the recycled layer thickness.

[0025] Temperature transfer connection is established based on temperature transfer changes between adjacent evolution nodes; material migration connection is established based on displacement changes, depth changes, and echo response changes between adjacent evolution nodes; and compaction connection is established based on pressure changes, depth changes, and echo interface changes between adjacent evolution nodes.

[0026] Evolution nodes with temperature transfer, material migration, and compaction connections are connected in ascending order of construction mileage to form a thickness evolution diagram of the recycled layer. A set of thickness evolution paths is generated based on the continuous connection status between the evolution nodes.

[0027] The thickness evolution path with complete connection relationship, longest continuous sampling length and consistent echo interface change with depth change is selected from the thickness evolution path set as the master evolution path. The temperature transfer state, material migration state, compaction state and construction stage identifier of each evolution node in the master evolution path are extracted to generate the thickness evolution state sequence.

[0028] Optionally, the output of the initial thickness measurement sequence that satisfies the thermal conductivity consistency condition includes:

[0029] Using the NBEATSx model as the basic model, a feature synchronization input structure is set at the input end of the NBEATSx model to synchronously input the multi-source pavement state feature sequence, construction stage identification sequence and temperature transfer feature sequence according to the sampling time and construction mileage location.

[0030] In the neural basis expansion block of the NBEATSx model, the trend term mapping structure and residual backpropagation structure are preserved, and the construction stage identifier sequence is used as stage constraint information to extract the stage thickness contribution features corresponding to each construction stage.

[0031] A temperature-thickness coupling inversion layer is embedded after each level of neural basis expansion block. The temperature-thickness coupling inversion layer reads the stage thickness contribution feature and temperature transfer feature sequence to generate the corresponding learnable thermal diffusivity.

[0032] The longitudinal temperature field from the surface layer to the interior of the recycled layer is solved based on the learnable thermal diffusivity coefficient and temperature transfer characteristic sequence. The isothermal depth of the recycled layer interface is identified based on the continuous position of temperature change in the longitudinal temperature field.

[0033] In the differentiable heat conduction consistency constraint layer, the difference between the interface isothermal depth and the stage thickness contribution characteristics corresponding to each construction stage is compared to construct a physical consistency loss for constraining the thickness inversion results.

[0034] The stage weights of each construction stage are adaptively adjusted based on the physical consistency loss, and the stage thickness contribution characteristics are iteratively optimized through the residual backpropagation structure, and the initial thickness measurement sequence that meets the thermal conduction consistency condition is fused to generate.

[0035] A training dataset is constructed based on multi-source pavement state feature sequences, initial thickness measurement sequences, and known thickness reference values. A comprehensive loss function consisting of physical consistency loss and prediction error weighting is used to backpropagate and update parameters of the improved NBEATSx model in batches until the comprehensive loss converges, and the trained model parameters are saved.

[0036] Optionally, generating the corresponding learnable thermal diffusivity includes:

[0037] Within the temperature-thickness coupled inversion layer, the thermal diffusivity parameter is initialized for each construction stage, and the thermal diffusivity parameter, along with the stage thickness contribution characteristics and temperature transfer characteristic sequence, is written into the coupled inversion calculation unit.

[0038] The coupled inversion calculation unit calculates the temperature response difference based on the heat transfer relationship between the stage thickness contribution characteristics and the temperature transfer characteristic sequence, and feeds the temperature response difference back to the thermal diffusivity parameter as an update signal;

[0039] The thermal diffusivity parameter is iteratively adjusted using the update signal at a set learning rate, so that the temperature response difference tends to be minimized in continuous iterations.

[0040] When the temperature response difference meets the convergence criteria, the converged thermal diffusivity parameter value is recorded as the learnable thermal diffusivity for the corresponding construction stage.

[0041] Optionally, the step of performing reverse compensation on the initial thickness measurement sequence to generate a thickness correction sequence includes:

[0042] Read the initial thickness measurement sequence and the thickness evolution state sequence, match them according to the sampling time, construction mileage location and construction stage identifier, and establish the correspondence between the initial thickness measurement value and the thickness evolution state of each sampling point;

[0043] Write the initial thickness measurement, temperature transfer status, material migration status, compaction status, and construction stage identifier of each sampling point into the same residual record. Construct a path stage joint index according to the construction mileage location and construction stage sequence to generate a thickness residual field.

[0044] Based on the thickness residual field, the initial thickness measurement value and thickness evolution state corresponding to each sampling point are read. The thickness formation trend represented by the initial thickness measurement value and the thickness evolution state are compared to determine the deviation direction, deviation magnitude and deviation duration of each sampling point.

[0045] The deviation direction, deviation magnitude, and deviation duration of each sampling point are transmitted to adjacent sampling points along the construction path, and continuous deviation sampling points and abrupt deviation sampling points are associated to form an error backtracking relationship.

[0046] Based on the error backtracking relationship, the error source decomposition is performed on the abnormal sampling points. The deviation state corresponding to the abnormal sampling points is divided into temperature transmission deviation, material migration deviation and compaction state deviation, and an error source record is generated.

[0047] Based on the error source record, the compensation direction of abnormal sampling points and related sampling points is determined. Reverse compensation is performed on the initial thickness measurement sequence, and the sequence is rearranged according to the sampling time and construction mileage to generate a thickness correction sequence.

[0048] Optionally, the output includes pavement thickness measurement results distributed along the construction direction, and the generation of corresponding thickness anomaly section records, including:

[0049] Read the thickness correction sequence, sort the corrected thickness values ​​of each sampling point according to the sampling time, construction mileage location and construction stage identifier, and generate a construction path thickness arrangement sequence;

[0050] Based on the thickness arrangement sequence of the construction path, the continuous thickness change state between adjacent sampling points is identified, and neighborhood constraint smoothing is performed on isolated abrupt thickness values ​​to generate a continuous thickness distribution sequence.

[0051] The continuous thickness distribution sequence is matched with the thickness evolution state sequence to determine whether the thickness change direction, thickness change amplitude and construction stage identification of each sampling point are consistent, and a thickness consistency verification result is generated.

[0052] Based on the thickness consistency verification results, abnormal thickness sampling points are identified. Sampling points with consecutive abnormalities are merged into thickness abnormal sections according to the construction mileage location, and the start and end positions and abnormality types of abnormal sections are recorded.

[0053] The continuous thickness distribution sequence is used to generate pavement thickness measurement results according to the construction direction, and the sampling time, construction mileage location, construction stage identifier and correction thickness value corresponding to each sampling point are written into the thickness measurement results.

[0054] The thickness measurement results, thickness consistency verification results, and thickness anomaly sections are associated and stored to generate a thickness anomaly section record for construction process control and quality evaluation.

[0055] The beneficial effects of this invention are:

[0056] This invention constructs an in-situ thermal recycling pavement thickness measurement method based on intelligent sensing technology. During construction, it collects multi-source state data such as pavement temperature, displacement, depth, pressure, and echo response in real time. Combined with the multi-source pavement state characteristic sequence, it establishes a recycled layer thickness evolution diagram, which can continuously characterize the formation process of recycled layer thickness. This method overcomes the limitations of traditional post-construction core sampling, manual measurement, and single-point detection, which can only obtain local thickness information. It realizes real-time, continuous, and non-destructive measurement of recycled layer thickness, and improves the thickness perception capability and construction quality control capability during construction.

[0057] This invention introduces a temperature-thickness coupled inversion layer and a differentiable thermal conduction consistency constraint layer into the improved NBEAT's x model. It utilizes a learnable thermal diffusivity coefficient to construct the longitudinal temperature field of the road surface and applies physical consistency constraints to the thickness inversion results based on the isothermal depth of the interface. This enables the model to not only learn the mapping relationship between multi-source state characteristics and thickness but also to integrate the thermal conduction mechanism, thereby improving the stability and reliability of thickness prediction under complex thermal recycling construction environments and reducing the interference of temperature changes, material migration, and compaction state changes on the measurement results.

[0058] This invention further constructs a thickness residual field and establishes an error backtracking relationship by combining the thickness evolution state sequence. It performs error source decomposition and reverse compensation on abnormal sampling points to achieve dynamic correction of thickness measurement results. At the same time, it generates pavement thickness measurement results and thickness abnormal section records that are continuously distributed along the construction direction. This can provide continuous data support for adjusting construction equipment parameters, controlling recycling depth, and evaluating construction quality, thereby improving the thickness control accuracy and consistency of construction quality in in-situ thermal recycling. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of the in-situ thermal regeneration pavement thickness measurement method based on intelligent sensing technology proposed in this invention.

[0061] Figure 2 This is a graph showing the training loss variation of the improved NBEATSx model for the in-situ thermal regeneration pavement thickness measurement method based on intelligent sensing technology proposed in this invention.

[0062] Figure 3 This is a map showing the longitudinal temperature field and interface isothermal depth of the pavement in the in-situ thermal regeneration pavement thickness measurement method based on intelligent sensing technology proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figures 1-3 A method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology includes:

[0065] Multi-source state sensing units are deployed on the on-site thermal recycling construction equipment to collect multi-source state data during the road heating, loosening, paving and compaction processes, and perform preprocessing to generate a construction state data set.

[0066] Based on the construction status data set, temperature change features and temperature transmission features are extracted, and displacement, depth, pressure and echo response features are also extracted to generate a multi-source pavement status feature sequence.

[0067] Based on the multi-source pavement state feature sequence, a thickness evolution map of the recycled layer is established. The pavement state at each sampling time is taken as the evolution node, and temperature transfer behavior, material migration behavior and compaction behavior are taken as the connection relationship to generate the thickness evolution path. The dominant evolution path is identified and the corresponding thickness evolution state sequence is generated.

[0068] The multi-source pavement state feature sequence, construction stage identification sequence and temperature transfer feature sequence are simultaneously input into the improved NBEATSx model with a differentiable heat conduction consistency constraint layer. A temperature-thickness coupling inversion layer is inserted after each level of neural base expansion block to output the initial thickness measurement sequence that meets the heat conduction consistency condition.

[0069] A thickness residual field is constructed based on the initial thickness measurement sequence. The deviation between each sampling point and the thickness evolution state sequence is calculated. The error backtracking relationship is formed by propagating along the construction path. Error decomposition is performed on abnormal sampling points. The initial thickness measurement sequence is then compensated in reverse to generate a thickness correction sequence.

[0070] The thickness correction sequence is smoothed and its consistency is checked. The pavement thickness measurement results distributed along the construction direction are output, and the corresponding thickness anomaly section records are generated.

[0071] In this embodiment, the multi-source state sensing unit includes a temperature sensor, a displacement sensor, a depth sensor, a pressure sensor, and an echo detection sensor.

[0072] In this embodiment, the multi-source state data includes road surface temperature data, displacement data, depth data, pressure data, and echo response data.

[0073] In this embodiment, generating the multi-source road surface state feature sequence includes:

[0074] Road surface temperature data is read from the construction status dataset, and the road surface temperature data is arranged according to the sampling time and construction mileage location to generate a basic temperature sequence;

[0075] Based on the temperature base sequence, extract the temperature change amplitude, temperature change direction, temperature decay state, and temperature difference between different sampling locations between adjacent sampling times to generate a temperature change feature set.

[0076] Based on the temperature change feature set, the heat transfer state between the road surface layer and the recycled layer is identified. The direction of temperature transfer, the intensity of temperature transfer, and the continuity of temperature attenuation are extracted to generate a temperature transfer feature sequence. Specifically, the generated temperature transfer feature sequence is as follows:

[0077] Read the surface temperature and deep temperature one by one at the sampling time, calculate the difference between the two. If the difference is positive and shows an increasing trend within three consecutive sampling times, it is determined that the heat is transferred from the surface to the interior of the regeneration layer. If the difference is negative and shows an increasing trend within three consecutive sampling times, it is determined that the heat is transferred from the interior of the regeneration layer to the surface. All other cases are marked as non-directional transfer.

[0078] Using five sampling times as a sliding window, the absolute values ​​of the temperature differences within the window are averaged. An average value greater than 15 degrees Celsius is marked as high-intensity transmission, between 5 and 15 degrees Celsius is marked as medium-intensity transmission, and below 5 degrees Celsius is marked as low-intensity transmission.

[0079] Within the same sliding window, compare adjacent changes in temperature difference. If the change always decreases and the window length is no less than 5 sampling times, it is determined to be a continuous temperature decay state. If there is one or more rebounds during the decrease, it is determined to be an interrupted temperature decay state.

[0080] The temperature transmission direction, temperature transmission intensity level, and temperature decay state corresponding to each sampling time are combined to form a temperature transmission feature record. The sampling time and construction mileage location index are added to the record. All temperature transmission feature records are connected in series according to the sampling time order to generate a temperature transmission feature sequence covering the entire construction path.

[0081] Displacement, depth, pressure, and echo response data are read from the construction status dataset and aligned with the sampling time and construction mileage location corresponding to the temperature transfer characteristic sequence to generate a structural response feature set. Specifically, the generation of the structural response feature set involves:

[0082] Read the displacement sequence, depth sequence, pressure sequence and echo response sequence in the order of sampling time, and write the corresponding sampling time and construction mileage position into the same index;

[0083] Calculate the displacement increment for the displacement sequence between adjacent sampling times, calculate the depth increment for the depth sequence between adjacent sampling times, calculate the pressure gradient for the pressure sequence between adjacent sampling times, and calculate the echo change for the echo response sequence between adjacent sampling times.

[0084] Based on the standard that the absolute value of the displacement increment is greater than 4 mm, intermediate between 1 and 4 mm, and less than 1 mm, the displacement increment is divided into three levels of displacement response: high, medium, and low. Based on the standard that the absolute value of the depth increment is greater than 3 mm, intermediate between 1 and 3 mm, and less than 1 mm, the depth increment is divided into three levels of depth response: high, medium, and low. Based on the standard that the absolute value of the pressure gradient is greater than 200 kPa, intermediate between 50 and 200 kPa, and less than 50 kPa, the pressure gradient is divided into three levels of pressure response: high, medium, and low. Based on the standard that the absolute value of the echo change is greater than 15, between 5 and 15, and less than 5, the echo change is divided into three levels of echo response: high, medium, and low.

[0085] At each sampling moment, the displacement response level, depth response level, pressure response level, and echo response level are combined with the sampling moment and construction mileage location to generate a structural response characteristic record.

[0086] All structural response feature records are concatenated according to the sampling time sequence to generate a structural response feature set.

[0087] The temperature change feature set, temperature transfer feature sequence, and structural response feature set are combined and sorted according to sampling time, construction mileage location, and construction stage identifier. A corresponding feature record is established for each sampling location to generate a multi-source pavement state feature sequence. Specifically, the generation of the multi-source pavement state feature sequence is as follows:

[0088] Temperature change characteristic records, temperature transfer characteristic records, and structural response characteristic records are read in ascending order of sampling time. The time index of the three types of characteristic records is compared with the construction mileage location index, and records with consistent time index and location index are selected as combination candidates.

[0089] Read the construction stage identifier from the combined candidate records. For multiple records with the same time index and location index, sort them first according to the order of the construction stage identifier, and then arrange the temperature change characteristics, temperature transfer characteristics and structural response characteristics in sequence.

[0090] Assign a unique sampling location identifier to each group of sorted feature records, write temperature change features, temperature transmission features, displacement response level, depth response level, pressure response level and echo response level into the feature field, and write sampling time, construction mileage location and construction stage identifier into the index field to generate multi-source feature records.

[0091] All multi-source feature records are sorted in ascending order of sampling time. When records with the same sampling time but different construction mileage locations appear, they are arranged in ascending order of construction mileage location.

[0092] The sorted multi-source feature records are sequentially written into the multi-source road surface state feature sequence.

[0093] In this embodiment, the step of generating a thickness evolution path, identifying the dominant evolutionary path, and generating a corresponding thickness evolution state sequence includes:

[0094] Read the temperature transfer characteristics, displacement characteristics, depth characteristics, pressure characteristics, echo response characteristics and construction stage identifiers corresponding to each sampling time in the multi-source pavement state characteristic sequence, and establish pavement state nodes according to the sampling time and construction mileage location;

[0095] The temperature transfer state, material position change state, compression forming state and echo interface state corresponding to each road surface state node are written into the node attributes to generate an evolution node set used to characterize the formation process of the recycled layer thickness.

[0096] Temperature transfer connections are established based on temperature transfer changes between adjacent evolution nodes; material migration connections are established based on displacement, depth, and echo response changes between adjacent evolution nodes; and compaction connections are established based on pressure, depth, and echo interface changes between adjacent evolution nodes.

[0097] Temperature transfer connections are established based on temperature transfer changes between adjacent evolution nodes, specifically as follows:

[0098] Read the temperature transfer direction of adjacent evolution nodes in sequence. The high-intensity downward transfer direction is recorded as -1, the high-intensity upward transfer direction is recorded as 1, and the other directions are recorded as 0. If the direction values ​​of two nodes are the same, the directions are determined to be consistent.

[0099] The temperature transfer intensity of adjacent evolution nodes is compared, and the high, medium and low levels are assigned values ​​of 3, 2 and 1 respectively. If the intensity difference between two nodes does not exceed 1 and the directions are consistent, the temperature transfer change is determined to be continuous.

[0100] Write a temperature transfer connection identifier for node pairs that are determined to have continuous temperature transfer changes, add a connection edge to the evolution graph, and set the attribute of the connection edge to temperature transfer connection.

[0101] When three or more consecutive node pairs are determined to be continuous temperature transmission changes, the corresponding connecting edges are merged to form a continuous temperature transmission segment, and the starting node number and ending node number of the continuous segment are recorded.

[0102] The temperature transfer connection edge and continuous segment information are written into the connection relationship table of the regeneration layer thickness evolution diagram to complete the establishment of the temperature transfer connection relationship between adjacent evolution nodes;

[0103] Material migration connectivity is established based on displacement changes, depth changes, and echo response changes between adjacent evolution nodes, specifically as follows:

[0104] Read the displacement response level, depth response level and echo response level in adjacent evolution nodes in sequence, and assign the high, medium and low levels to 3, 2 and 1 respectively, while recording the displacement increment direction and depth increment direction.

[0105] Compare the difference between the displacement response levels of adjacent nodes and the difference between the depth response levels. If both differences do not exceed 1 and the direction of displacement increment is the same as the direction of depth increment, the material position change is determined to be continuous.

[0106] Compare the echo response level difference between adjacent nodes. If the difference does not exceed 1 and is simultaneously satisfied with the node pair where the material position change is continuous, the material migration state is determined to be consistent.

[0107] Write a material migration connection identifier to the node pairs that meet the material migration state consistency, add a connection edge to the evolution graph, and set the connection edge attribute to material migration connection.

[0108] When three or more consecutive node pairs are determined to be in the same material migration state, the corresponding connecting edges are merged into a continuous material migration segment. The starting node number and ending node number of the segment are recorded and written into the connection relationship table of the thickness evolution diagram.

[0109] The compaction connection relationship is established based on pressure changes, depth changes, and echo interface changes between adjacent evolution nodes, specifically:

[0110] Read the pressure response level, depth response level, and echo interface change level of adjacent evolution nodes in sequence, and assign the high, medium, and low levels to 3, 2, and 1 respectively, while recording the pressure change direction and depth change direction.

[0111] Compare the pressure response level difference with the depth response level difference. When both differences do not exceed 1 and the direction of pressure change is opposite to the direction of depth change, the compaction trend is determined to be consistent.

[0112] Compare the difference in the change level of the echo interface between adjacent nodes. When the difference does not exceed 1 and the direction of change is consistent with the compaction trend, the compaction state is determined to be continuous.

[0113] Write a compaction connection identifier for node pairs that satisfy the continuous compaction state, add a connection edge in the evolution diagram, and set the attribute of the connection edge to compaction connection;

[0114] When three or more consecutive node pairs are determined to be in a continuous compaction state, the corresponding connecting edges are merged to form a continuous compaction section, and the starting node number and ending node number of the section are recorded and written into the connection relationship table of the thickness evolution diagram.

[0115] Evolution nodes with temperature transfer, material migration, and compaction connections are connected in ascending order of construction mileage to form a recycled layer thickness evolution diagram. A set of thickness evolution paths is generated based on the continuous connection states between these nodes, forming the recycled layer thickness evolution diagram, specifically:

[0116] Read the set of evolution nodes in ascending order of construction mileage, merge the temperature transfer connections, material migration connections and compaction connections established between adjacent nodes, and construct an initial directed graph containing the three types of connection edges;

[0117] In the initial directed graph, if adjacent nodes simultaneously possess temperature transfer connection, material migration connection, and compaction connection, then a comprehensive connection edge is established between the two nodes, and the comprehensive connection edge is assigned the thickness evolution connection attribute.

[0118] Traverse the integrated connection edges from smallest to largest along the construction mileage. When encountering consecutive thickness evolution connections, merge the corresponding nodes into the same thickness evolution sub-path in sequence, and record the starting node number and ending node number of the sub-path.

[0119] Scan the set of thickness evolution sub-paths. If the distance between the end node and the start node of two sub-paths does not exceed one sampling interval, and the combined connection direction of the two sub-paths is consistent, then merge the two sub-paths into a longer thickness evolution path.

[0120] Sort all merged thickness evolution paths in ascending order of starting mileage position to generate a thickness evolution path set, and write the thickness evolution path set into the path index table of the regenerated layer thickness evolution diagram;

[0121] The thickness evolution path with complete connection relationship, longest continuous sampling length and consistent echo interface change with depth change is selected from the thickness evolution path set as the master evolution path. The temperature transfer state, material migration state, compaction state and construction stage identifier of each evolution node in the master evolution path are extracted to generate the thickness evolution state sequence.

[0122] In this embodiment, the output of the initial thickness measurement sequence that satisfies the thermal conductivity consistency condition includes:

[0123] Using the NBEATSx model as the base model, a feature synchronization input structure is set at the input end of the NBEATSx model to synchronously input the multi-source pavement state feature sequence, construction stage identifier sequence, and temperature transfer feature sequence according to the sampling time and construction mileage location. The feature synchronization input structure refers to the input component located at the input end of the improved NBEATSx model, which is used to simultaneously receive the multi-source pavement state feature sequence, construction stage identifier sequence, and temperature transfer feature sequence according to a unified index. It consists of the following units:

[0124] Time-mileage index generation unit: Reads the original sampling time and construction mileage location of each sequence to generate a unified time-mileage dual index;

[0125] Alignment Buffer Unit: To address sampling time errors or missing samples in different sequences, a circular buffer is set up to temporarily store unaligned data until the data for the same time mileage index is complete;

[0126] Gap filling unit: When a single-channel data gap occurs at the index, it is filled by interpolation based on the most recent valid data, and the placeholder is marked with a hold value to avoid timing gaps;

[0127] Channel splicing unit: After the time and mileage dual indexes are consistent and all channel data are complete, the multi-source road surface condition features, construction stage identifiers and temperature transfer features are spliced ​​in the preset channel order to form a single frame synchronous input tensor.

[0128] Batch processing unit: The continuous frame synchronous input tensors are processed into batches according to the order of increasing time and increasing distance, and then fed into the neural basis expansion block of the NBEATSx model.

[0129] The trend term mapping structure and residual backpropagation structure are preserved in the neural basis expansion block of the NBEATSx model, and the construction stage identifier sequence is used as stage constraint information to extract the stage thickness contribution features corresponding to each construction stage. The trend term mapping structure refers to the neural mapping component embedded in the neural basis expansion block of the improved NBEATSx model, which is used to characterize the overall trend of thickness change with time and mileage. It consists of the following sub-units:

[0130] Feature expansion unit: Performs fully connected mapping and nonlinear activation on the synchronous input tensor to expand the original channel information into high-dimensional trend candidate features;

[0131] Coefficient generation unit: Projects weights onto the expanded trend candidate features and outputs a trend coefficient vector representing the trend magnitude and direction;

[0132] Trend basis function unit: Generates multi-order polynomial basis vectors on the dual index dimensions of time and mileage, used to establish the basic pattern of overall increasing, decreasing or gradual trend;

[0133] Trend Recombination Unit: Combines the trend coefficient vector with the trend basis vector element by element to obtain the trend component output of the current neural basis expansion block;

[0134] Residual output interface: Removes the trend component from the input tensor to form a residual signal for subsequent residual backpropagation structure processing, and sends it along with the stage thickness contribution feature into the next expansion block;

[0135] The residual backpropagation structure refers to a feedback component embedded in the same neural base expansion block of the improved NBEAT's x model, used to retain and progressively correct thickness prediction errors layer by layer, and includes the following sub-units:

[0136] Residual buffer unit: After completing the calculation of trend components and stage thickness contribution features, it temporarily stores the feature differences in the input tensor that have not yet been interpreted as residual signals;

[0137] Error mapping unit: performs fully connected mapping and nonlinear compression on the residual signal, converting the residual characteristics into a compressed residual representation suitable for further processing by subsequent expansion blocks;

[0138] Feedback Gating Unit: Introducing gating weights, which proportionally determine the participation of the current residual in the next expansion block, enabling the improved NBEATSx model to dynamically adjust the intensity of residual information flow;

[0139] Residual accumulation unit: The gated compressed residual representation is added to the new input tensor of the next expansion block after being aligned by channel, so as to realize cross-layer error progressive compensation;

[0140] Output synchronization interface: When the calculation of this layer ends, the updated residual signal is output, and the time, mileage index and construction stage identifier remain unchanged, so as to ensure that the expanded block can continue to iterate and optimize under the same index system;

[0141] Using the construction stage identifier sequence as stage constraint information, the stage thickness contribution features corresponding to each construction stage are extracted, specifically:

[0142] Read the construction stage identifier sequence corresponding to the synchronous input tensor, and map the four stages of heating, loosening, spreading and compaction into four mutually exclusive stage masks respectively;

[0143] Channel expansion is performed on the mask for each stage to ensure that the mask for each stage is fully aligned with the synchronous input tensor in terms of time, mileage, and feature channel dimensions, and a subset of data belonging to that stage is selected by bitwise AND operation.

[0144] On the selected subset of stage data, average pooling is performed along the time dimension to obtain the stage mean feature, while max pooling is performed along the mileage dimension to obtain the stage peak feature. The two types of features are concatenated to form the thickness contribution candidate vector for that stage.

[0145] The thickness contribution candidate vector is input into the fully connected mapping and nonlinear activation unit to obtain the normalized stage thickness contribution feature, and the corresponding stage mask index is appended to the end of the feature.

[0146] The thickness contribution characteristics of all stages are arranged in the order of heating, loosening, spreading and compacting to form a stage thickness contribution characteristic channel, which is then sent into the temperature-thickness coupled inversion layer.

[0147] A temperature-thickness coupling inversion layer is embedded after each level of neural basis expansion block. The temperature-thickness coupling inversion layer reads the stage thickness contribution feature and temperature transfer feature sequence to generate the corresponding learnable thermal diffusivity.

[0148] The longitudinal temperature field from the pavement surface to the interior of the recycled layer is solved based on a learnable thermal diffusivity coefficient and temperature transfer characteristic sequence. The isothermal depth of the recycled layer interface is identified based on the continuous locations of temperature changes within the longitudinal temperature field. Specifically, the longitudinal temperature field from the pavement surface to the interior of the recycled layer is solved based on the learnable thermal diffusivity coefficient and temperature transfer characteristic sequence.

[0149] The road surface thickness information was read segment by segment according to the construction mileage location. The 0cm to 10cm depth was divided into 11 depth units with 1cm intervals. Each depth unit was numbered and a temperature node vector was established.

[0150] At the same sampling time, the surface temperature value and the deep temperature value are read. The surface temperature is written to the 0cm depth node, and the deep temperature is written to the 10cm depth node by linear interpolation. The initial temperature of the other depth nodes is the average of the two values.

[0151] Read the temperature transfer direction and intensity level corresponding to the current sampling time in the temperature transfer feature sequence. If the direction is downward, the heat is transferred in the direction of increasing depth. If the direction is upward, the heat is transferred in the direction of decreasing depth. The intensity levels of high, medium and low correspond to heat diffusion time steps of 0.3s, 0.2s and 0.1s, respectively.

[0152] The learnable thermal diffusivity table generated by the temperature-thickness coupled inversion layer is called, and the coefficients are mapped to 0.35, 0.28, 0.22, and 0.15 according to the four construction stages: heating, loosening, spreading, and compaction. The coefficients corresponding to the construction stages are written into all depth nodes.

[0153] Explicit thermal diffusion iteration is performed in groups of 20 time steps: In each iteration, the temperature difference between adjacent depth nodes with the same direction is calculated, and the temperature difference is multiplied by the thermal diffusion coefficient and the time step size to update the temperature of the lower layer node. At the same time, the temperature of the upper layer node is called back. If the temperature change of each depth node is less than 0.2℃ after the iteration ends, the temperature field is determined to be stable.

[0154] The temperature values ​​of the 11 stable depth nodes are written into the longitudinal temperature field matrix in depth order. The depth position where the temperature value first drops continuously by more than 2℃ is marked in the longitudinal temperature field matrix and recorded as the isothermal depth of the regeneration layer interface. The temperature field matrix and the isothermal depth of the interface are output to the next step for physical consistency constraints.

[0155] In the differentiable heat conduction consistency constraint layer, the difference between the interface isothermal depth and the stage thickness contribution characteristics corresponding to each construction stage is compared to construct a physical consistency loss for constraining the thickness inversion results. Specifically, the physical consistency loss for constraining the thickness inversion results is as follows:

[0156] Read the isothermal depth value of the regenerated layer interface corresponding to the current sampling time, as well as the stage thickness contribution characteristics of the four stages of heating, loosening, spreading and compaction, and convert the thickness contribution values ​​of the four stages into the stage thickness boundary depth in an cumulative manner.

[0157] The interface isothermal depth is compared with the thickness boundary depth of the four stages one by one. If the interface isothermal depth falls between the thickness boundary depth of a certain stage and the thickness boundary depth of the next stage, then the target stage to which the interface isothermal depth should belong is confirmed.

[0158] Calculate the distance difference between the interface isothermal depth and the target stage thickness boundary depth. If the distance difference exceeds 5mm, it is recorded as a depth deviation. At the same time, detect the distance difference between the interface isothermal depth and the previous stage thickness boundary depth. If both distance differences are less than 2mm, they are considered to be consistent in depth.

[0159] Weighting coefficients were assigned to the four stages, with the heating stage having a weight of 0.3, the loosening stage having a weight of 0.25, the paving stage having a weight of 0.25, and the compaction stage having a weight of 0.2. The depth deviation of each stage was multiplied by its corresponding weight and then summed to obtain the physical consistency loss value at the current sampling time.

[0160] The stage weights of each construction stage are adaptively adjusted based on the physical consistency loss, and the stage thickness contribution characteristics are iteratively optimized through the residual backpropagation structure. This results in the generation of a preliminary thickness measurement sequence that satisfies the thermal conductivity consistency condition. Specifically, the adaptive adjustment of the stage weights based on the physical consistency loss is as follows:

[0161] After a forward inference is completed, the physical consistency loss values ​​corresponding to the heating stage, loosening stage, paving stage and compaction stage are recorded respectively, and the average value of the loss values ​​of the four stages is calculated. If the loss value of a stage is more than 5% higher than the average value, it is recorded as the deviation exceeding the limit stage. If the loss value of a stage is more than 5% lower than the average value, it is recorded as the deviation lacking stage. The remaining stages are recorded as the deviation normal stage.

[0162] The weights in the deviation exceeding the limit stage are decreased with a decrease step size of 0.05; the weights in the deviation insufficient stage are increased with an increase step size of 0.05; the weights in the deviation normal stage remain unchanged.

[0163] After the weight adjustment, the sum of the weights of the four stages is calculated. If the sum of the weights is not equal to 1, the weights of all stages are scaled proportionally to restore the sum to 1.

[0164] The scaled stage weights are written to the stage weight cache table for use in the next temperature-thickness coupling inversion.

[0165] In subsequent iterations, loss statistics, stage classification, weight adjustment and scaling are repeated until the change in weight of each stage is less than 0.02 in three consecutive iterations. This is considered as the stage weights having converged and the current weight configuration being locked.

[0166] A training dataset is constructed based on multi-source pavement state feature sequences, initial thickness measurement sequences, and known thickness reference values. A comprehensive loss function consisting of physical consistency loss and prediction error weighting is used to backpropagate and update parameters of the improved NBEATSx model in batches until the comprehensive loss converges. The trained model parameters are then saved. The prediction error comes from the difference between the initial thickness measurement value output by the improved NBEATSx model during forward inference in each training batch and the known thickness reference value in the same batch of samples. First, the multi-source pavement state feature sequences are fed into the improved NBEATSx model in batches to obtain the corresponding initial thickness measurement sequences. These sequences are then aligned point by point with the pre-labeled thickness reference values ​​in the training dataset. The residual is obtained by subtracting the reference thickness from the initial thickness measurement. The absolute value or square value of the residual is then taken as the single-point prediction error. The single-point prediction errors of all sampling points in a batch are averaged to obtain the prediction error index for that batch.

[0167] This invention structurally improves the traditional NBEATSx model, proposing a thickness inversion model suitable for in-situ thermal recycling pavement thickness measurement scenarios. A feature synchronization input structure is set at the model input end, synchronously inputting multi-source pavement state feature sequences, construction stage identifier sequences, and temperature transfer feature sequences according to sampling time and construction mileage location. This enables the model to simultaneously sense pavement temperature changes, structural response changes, and construction stage changes, avoiding the problem of thickness inversion results caused by misalignment of different sensor data in time and location. It also improves the model's ability to fuse multi-source state data in continuous construction scenarios, making the initial thickness measurement results more consistent with the actual construction conditions.

[0168] To address the significant impact of temperature transfer on thickness measurement during in-situ thermal recycling, this invention embeds a temperature-thickness coupled inversion layer after each neural basis expansion block of the NBEATSx model. By reading the thickness contribution characteristics and temperature transfer characteristic sequences of each stage, a corresponding learnable thermal diffusivity coefficient is generated, and the longitudinal temperature field from the pavement surface to the interior of the recycled layer is solved. The isothermal depth at the recycled layer interface is identified through the longitudinal temperature field, and then the isothermal depth at the interface is compared with the thickness contribution characteristics of each construction stage to construct a physical consistency loss. This ensures that the model not only relies on data mapping relationships for prediction but is also constrained by the heat conduction mechanism, reducing the interference of temperature gradients, material softening, and heat attenuation on the thickness inversion results, and improving the physical reliability and stability of the initial thickness measurement sequence.

[0169] To enhance the model's adaptability to thickness variations across different construction stages, this invention incorporates the construction stage identifier sequence as stage constraint information into the neural basis expansion block. It then adaptively adjusts the stage weights for heating, loosening, paving, and compaction stages using physical consistency loss, and iteratively optimizes the stage thickness contribution characteristics using a residual backpropagation structure. In this way, the model can dynamically adjust the thickness contribution ratio based on the heat transfer, material migration, and compaction states at different construction stages, improving its ability to express the evolution of recycled layer thickness. Ultimately, it outputs a preliminary thickness measurement sequence that satisfies the heat conduction consistency condition, thereby improving the continuity, accuracy, and anti-interference capability of in-situ thermal recycled pavement thickness measurement.

[0170] In this embodiment, generating the corresponding learnable thermal diffusivity includes:

[0171] Within the temperature-thickness coupled inversion layer, the thermal diffusivity parameter is initialized for each construction stage, and the thermal diffusivity parameter, along with the stage thickness contribution characteristics and temperature transfer characteristic sequence, is written into the coupled inversion calculation unit.

[0172] The coupled inversion calculation unit calculates the temperature response difference based on the heat transfer relationship between the stage thickness contribution characteristics and the temperature transfer characteristic sequence, and feeds the temperature response difference back to the thermal diffusivity parameter as an update signal;

[0173] The thermal diffusivity parameter is iteratively adjusted using the update signal at a set learning rate, so that the temperature response difference tends to be minimized in continuous iterations.

[0174] When the temperature response difference meets the convergence criterion, the converged thermal diffusivity parameter value is recorded as the learnable thermal diffusivity for the corresponding construction stage. The convergence criterion is that during the temperature-thickness coupled inversion iteration process of the same construction stage, when the temperature response difference of each depth unit is less than 0.2℃ in three consecutive iterations, and the relative change rate of the corresponding thickness contribution correction amount relative to the previous round does not exceed 2%, and the update amplitude of the stage thermal diffusivity is less than 1% of its current value, it is considered that the temperature field has entered the stable region, and the thermal diffusivity of this stage is determined to have converged and can be recorded as the learnable thermal diffusivity for the corresponding construction stage.

[0175] In this embodiment, the step of performing reverse compensation on the initial thickness measurement sequence to generate a thickness correction sequence includes:

[0176] Read the initial thickness measurement sequence and the thickness evolution state sequence, match them according to the sampling time, construction mileage location and construction stage identifier, and establish the correspondence between the initial thickness measurement value and the thickness evolution state of each sampling point;

[0177] The initial thickness measurement, temperature transfer status, material migration status, compaction status, and construction stage identifier of each sampling point are written into the same residual record. A path stage joint index is constructed according to the construction mileage location and construction stage sequence to generate a thickness residual field. The generation of the thickness residual field is specifically as follows:

[0178] Each sampling point is read sequentially, and the five pieces of information, namely the initial thickness measurement, temperature transfer status, material migration status, compaction and molding status and construction stage identification, are written into a 1×5 feature vector. All feature vectors are arranged in the order of sampling time to form the original residual vector group.

[0179] A one-dimensional mileage index table is established based on the construction mileage location of the sampling points, and a one-dimensional stage index table is established based on the construction stage identifier. The two index tables are combined in Cartesian order according to the rule of mileage increment priority and stage order second, to generate a path-stage joint index matrix.

[0180] The original residual vector group is rearranged according to the order of the path-stage joint index matrix. During the rearrangement process, the four stages (heating, loosening, spreading, and compaction) at the same path position are assigned consecutive stage numbers so that adjacent stage records remain adjacent in the vertical direction.

[0181] In the rearranged residual vector group, the initial thickness measurement value is extracted and the corresponding thickness evolution state value is subtracted. The difference is written into the residual column and copied to the mileage index dimension. The number of copies is equal to the total number of rows of the stage index, forming a residual matrix extending along the construction mileage.

[0182] Map the row index of the residual matrix to the stage number and the column index to the mileage number. Concatenate the residual value in each cell of the matrix with the three columns of data vectors of temperature transfer state, material migration state, and compaction state, fill the last dimension of the four-dimensional tensor, and output the four-dimensional tensor as the thickness residual field distributed along the construction path.

[0183] Based on the thickness residual field, the initial thickness measurement value and thickness evolution state corresponding to each sampling point are read. The thickness formation trend represented by the initial thickness measurement value and the thickness evolution state are compared to determine the deviation direction, deviation magnitude, and deviation duration of each sampling point. Specifically, determining the deviation direction, deviation magnitude, and deviation duration of each sampling point is as follows:

[0184] In the thickness residual field, each sampling point is traversed sequentially, the initial thickness measurement value and the matched thickness evolution state value are read, and the difference between the two is written into the residual record.

[0185] If the residual record is positive, it means that the initial thickness measurement is greater than the evolved state, and it is recorded as a positive deviation; if the residual record is negative, it means that the initial thickness measurement is less than the evolved state, and it is recorded as a negative deviation; if the absolute value of the residual record is less than 1mm, it is recorded as no significant deviation.

[0186] Calculate the absolute value of the residual records. Residuals greater than 5 mm are recorded as high amplitude deviation, residuals between 2 mm and 5 mm are recorded as medium amplitude deviation, and residuals less than 2 mm are recorded as low amplitude deviation.

[0187] The deviation direction of continuous sampling times at the same distance position is compared. If the direction is consistent and lasts for 3 or more sampling times, it is marked as continuous deviation. If the direction changes within 3 sampling times, it is marked as intermittent deviation.

[0188] The deviation direction result, deviation magnitude level, and deviation duration status are written into the deviation feature record of the corresponding sampling point and stored by path stage joint index;

[0189] The deviation direction, deviation magnitude, and deviation duration of each sampling point are transmitted to adjacent sampling points along the construction path, and continuous deviation sampling points and abrupt deviation sampling points are associated to form an error backtracking relationship. Specifically, the error backtracking relationship is formed as follows:

[0190] Traverse the thickness residual field in ascending order of mileage index, and read the deviation direction, deviation magnitude and deviation duration at consecutive sampling times under the same mileage index;

[0191] If the deviation directions of two adjacent sampling times are the same and the difference in deviation magnitude level is no more than 1, then the latter is marked as a continuous deviation node of the former, and a continuous connection label is written between the two nodes.

[0192] If the deviation directions of two adjacent sampling times are opposite, or the deviation magnitude levels differ by 2 levels or more, then the latter is marked as a mutation deviation node, and a mutation connection label is written between the two nodes.

[0193] After completing the node connection within the same mileage index, continue to compare the first and last nodes between adjacent mileage indices. If the first and last nodes deviate in the same direction and both have a continuous deviation status, then write a cross-mileage continuous connection label between the two nodes for cross-mileage segment error backtracking.

[0194] Write all continuous connections, mutation connections, and cross-continuous connections into the error backtracking index table, using the node number as the key and the connection label as the value, and output the complete error backtracking relationship.

[0195] Based on the error backtracking relationship, error source decomposition is performed on abnormal sampling points. The deviation states corresponding to abnormal sampling points are divided into temperature transmission deviation, material migration deviation, and compaction state deviation, and error source records are generated. Specifically, the deviation states corresponding to abnormal sampling points are divided into temperature transmission deviation, material migration deviation, and compaction state deviation as follows:

[0196] Locate the abnormal sampling points marked as abrupt connections in the error backtracking index table, read the temperature transfer state, material migration state, and compaction state, and write the three states into the abnormal feature vector.

[0197] Compare the temperature transmission direction in the abnormal feature vector with the temperature transmission direction of the previous normal node. If the two are inconsistent, or the difference in temperature transmission intensity level is greater than 1, it is determined that there is a temperature transmission deviation in the abnormal sampling point.

[0198] Compare the displacement response level and the depth response level in the abnormal feature vector. If the difference between the two levels and the previous normal node is greater than 1, or the displacement increment direction is inconsistent with the depth increment direction, then it is determined that there is a material migration deviation in the abnormal sampling point.

[0199] Compare the pressure response level and the echo interface change level in the abnormal feature vector. If the difference between the pressure response level and the previous normal node is greater than 1, and the direction of the echo interface change is inconsistent with the direction of the pressure change, then it is determined that there is a compaction state deviation at the abnormal sampling point.

[0200] The results of the determination of temperature transmission deviation, material migration deviation and compaction state deviation, along with the abnormal sampling point number, are written into the error source record;

[0201] Based on the error source record, the compensation direction for abnormal sampling points and related sampling points is determined. Reverse compensation is performed on the initial thickness measurement sequence, and the sequence is rearranged according to the sampling time and construction mileage to generate a thickness correction sequence. Specifically, the generation of the thickness correction sequence is as follows:

[0202] Read the abnormal sampling point number and the corresponding temperature transmission deviation, material migration deviation and compaction state deviation from the error source record one by one. Mark the abnormal sampling points with two or more types of deviations as priority compensation points, and mark the rest as ordinary compensation points.

[0203] Based on the error backtracking relationship, extend two sampling times forward and backward from each priority compensation point, collect associated sampling points with continuous or cross-connection labels to the priority compensation point, integrate them into the same compensation area, and record the original thickness initial measurement value of all sampling points in the area.

[0204] The compensation direction is determined based on the type of deviation: if there is a temperature transmission deviation, the initial thickness measurement value is adjusted by decreasing or increasing along the temperature gradient direction; if there is a material migration deviation, the initial thickness measurement value is finely adjusted in the same direction along the displacement increment direction; if there is a compaction state deviation, the initial thickness measurement value is finely adjusted in the opposite direction along the pressure change direction.

[0205] Apply corresponding directional adjustments to the sampling points within the compensation area in sequence. The adjustment range is capped at 20% of the maximum residual in the compensation area, so that the compensated thickness value transitions smoothly point by point and avoids new abrupt changes.

[0206] After all compensation areas have been adjusted, the updated thickness values ​​will be rearranged in ascending order of sampling time and construction mileage location, and written into the thickness correction sequence.

[0207] In this embodiment, the step of outputting the pavement thickness measurement results distributed along the construction direction and generating corresponding thickness anomaly section records includes:

[0208] Read the thickness correction sequence, sort the corrected thickness values ​​of each sampling point according to the sampling time, construction mileage location, and construction stage identifier, and generate a construction path thickness arrangement sequence. Specifically, the generation of the construction path thickness arrangement sequence is as follows:

[0209] Read each record in the thickness correction sequence sequentially and write the sampling time, construction mileage location, construction stage identifier and correction thickness value into a temporary list;

[0210] The temporary list is sorted from earliest to latest sampling time to ensure that multiple data points within the same time slice are in adjacent positions.

[0211] Within the same sampling time range, a second sorting is performed according to the construction mileage location from smallest to largest, so that the list is arranged in ascending order along the construction path;

[0212] Within the same mileage range, the construction stages are sequentially identified as heating stage, loosening stage, paving stage, and compaction stage, ensuring the logical order of the thickness values ​​of the four stages.

[0213] After completing the double-layer time-mileage sorting and the internal sorting within the stage, the thickness records are sequentially written into the construction path thickness arrangement sequence, and a path order index is attached to each record as input for subsequent smoothing and consistency verification.

[0214] Based on the thickness arrangement sequence of the construction path, the continuous thickness variation state between adjacent sampling points is identified. Neighborhood constraint smoothing is performed on isolated abrupt thickness values ​​to generate a continuous thickness distribution sequence. Specifically, the generation of the continuous thickness distribution sequence is as follows:

[0215] The thickness arrangement sequence of the construction path is sequentially traversed, and the difference between the corrected thickness values ​​of two adjacent sampling points is calculated. If the absolute value of the difference does not exceed 3mm, it is judged as a continuous change; if the absolute value of the difference is between 3mm and 7mm, it is judged as a gradual change; and if the absolute value of the difference exceeds 7mm, it is judged as a sudden change.

[0216] For the sampling point identified as a mutation, two points are read before and after it to form a neighborhood window with a window length of 5. If the difference between the other four points in the window, except for the center point, is a continuous or gradual change, then the center point is confirmed as the isolated mutation thickness value.

[0217] Neighborhood constraint smoothing is applied to isolated abrupt thickness values: the average thickness values ​​of the other four points in the window are used to replace the thickness value of the center point, and the number of smoothing replacements is recorded.

[0218] If the same location on the same route appears as an isolated mutation in three consecutive time slices and has been smoothed more than twice, then the location is marked as a recurring mutation point to suppress subsequent iterations from replacing it again, in order to prevent over-smoothing.

[0219] After completing all window scans and necessary neighborhood constraints, the updated thickness values ​​are reordered according to sampling time, construction mileage location, and construction stage identifier, and written into the continuous thickness distribution sequence.

[0220] The continuous thickness distribution sequence is matched with the thickness evolution state sequence to determine whether the thickness change direction, thickness change amplitude, and construction stage identifier are consistent at each sampling point, thus generating a thickness consistency verification result. Specifically, the generation of the thickness consistency verification result involves:

[0221] Read the continuous thickness distribution sequence and the thickness evolution state sequence line by line. At the same sampling time and construction mileage position, compare the thickness change direction in the continuous thickness distribution sequence with the thickness evolution direction in the thickness evolution state sequence. If the two directions are consistent, they are recorded as consistent; otherwise, they are recorded as inconsistent.

[0222] Among sampling points with the same direction, compare the thickness change magnitude level of the continuous thickness distribution sequence with the thickness evolution magnitude level of the thickness evolution state sequence; if the difference between the two magnitude levels does not exceed one level, they are recorded as having the same magnitude; otherwise, they are recorded as having different magnitudes.

[0223] Compare the construction stage identifiers in the continuous thickness distribution sequence and the thickness evolution state sequence. If the identifiers are completely consistent, they are recorded as consistent stages; otherwise, they are recorded as inconsistent stages.

[0224] When the direction, amplitude, and stage are all consistent, the sampling point is marked as having consistent thickness; if any one of the three conditions is inconsistent, it is marked as having inconsistent thickness, and the inconsistency type is recorded.

[0225] Write the thickness consistency or inconsistency of all sampling points into the thickness consistency verification table, along with inconsistency type information, and output the thickness consistency verification result.

[0226] Based on the thickness consistency verification results, abnormal thickness sampling points are identified. Sampling points with consecutive abnormalities are merged into thickness abnormal sections according to the construction mileage location, and the start and end positions and abnormality types of abnormal sections are recorded.

[0227] The continuous thickness distribution sequence is used to generate pavement thickness measurement results according to the construction direction, and the sampling time, construction mileage location, construction stage identifier and correction thickness value corresponding to each sampling point are written into the thickness measurement results.

[0228] The thickness measurement results, thickness consistency verification results, and thickness anomaly sections are associated and stored to generate a thickness anomaly section record for construction process control and quality evaluation.

[0229] Example 1: In a simulation of continuous construction of in-situ thermal recycling, a 120m long asphalt pavement section was selected as the test object. The designed recycled layer thickness was 50mm, with an allowable deviation of ±5mm. The construction equipment operated continuously along the increasing mileage direction, sequentially experiencing the stages of heating, loosening, paving, and compaction. To simulate real construction problems, insufficient material loosening was simulated in the 36m to 48m section, and discontinuous temperature decay and compaction fluctuations were simulated in the 82m to 94m section, resulting in localized areas of thinner thickness that were not visually obvious. Traditional testing methods selected 12 core drilling points after construction, testing once every 10m, which could not continuously reflect the thickness changes in these two abnormal sections. Multi-source state sensing units were deployed on the construction equipment, with a sampling interval of 2s and a construction mileage interval of 0.5m, resulting in 241 mileage sampling locations and 964 stage sampling records. Each record only contained pavement temperature data, displacement data, depth data, pressure data, and echo response data. Taking the loosening stage at 38m as an example, the collected data showed a surface temperature of 143.1℃, a deep temperature of 112.6℃, a displacement increment of 5.2mm, a depth increment of 4.1mm, a pressure gradient of 76kPa, and an echo change of 18.3. Taking the compaction stage at 68m as an example, the surface temperature was 121.6℃, the deep temperature was 108.1℃, the displacement increment was 0.8mm, the depth increment was 1.2mm, the pressure gradient was 188kPa, and the echo change was 6.9. The system aligned the above data according to sampling time and construction mileage, removed 3 records of abrupt echo jumps and 2 records of missing pressure data, and supplemented the data to form a construction status data set, increasing the data integrity rate from the original 98.9% to 100%. The system then generated a multi-source pavement status feature sequence. When processing road surface temperature data, if the temperature difference between the surface and deep layers shows an increasing trend over three consecutive sampling times, the transmission direction is recorded. The average temperature difference is calculated using five sampling times as a window. A difference exceeding 15℃ is recorded as a level 3 transmission, 5℃ to 15℃ as a level 2 transmission, and below 5℃ as a level 1 transmission. At 38m, the temperature difference is 30.5℃, with differences of 27.8℃, 28.9℃, 30.5℃, 31.2℃, and 30.7℃ within the consecutive window, marking it as a downward level 3 transmission with interrupted attenuation. At 68m, the temperature difference is 13.5℃, with the window difference continuously decreasing, marking it as a downward level 2 transmission with continuous attenuation. When processing structural response data, displacement increments exceeding 4mm are recorded as level 3, depth increments exceeding 3mm as level 3, pressure gradients between 50kPa and 200kPa as level 2, and echo changes exceeding 15 are recorded as level 3. Therefore, the feature record generated at 38m is temperature transmission level 3, displacement level 3, depth level 3, pressure level 2, and echo level 3, while the feature record generated at 68m is temperature transmission level 2, displacement level 1, depth level 2, pressure level 2, and echo level 2.When establishing the thickness evolution map of the recycled layer, the system writes the pavement state at each sampling time as an evolution node, forming a total of 964 evolution nodes. Adjacent nodes are connected based on temperature transfer, material migration, and compaction behavior. Taking the 36m to 48m section as an example, 19 out of 25 consecutive sampling locations have a displacement response level and a depth response level of 3, with the echo response level varying between 2 and 3. The system writes these nodes into the material migration connection relationship. Taking the 82m to 94m section as an example, 16 out of 23 consecutive sampling locations experience temperature decay interruptions, with the pressure response level jumping between 2 and 3. The system writes these nodes into the temperature transfer connection and compaction connection. After all connections are completed, the system generates 7 thickness evolution paths, with the 3rd path covering 36m to 48m and a path length of 25 nodes, and the 6th path covering 82m to 94m and a path length of 23 nodes. The dominant optimization path was determined based on connectivity integrity, continuous sampling length, and consistency between echo interface changes and depth changes. Ultimately, paths 3 and 6 were marked as anomalous dominant optimization paths. During the model training phase, the system prepared 2880 samples, of which 2304 were used for training and 576 for validation. No additional data types were introduced into the samples; only pavement temperature data, displacement data, depth data, pressure data, echo response data, and the feature records generated from them were used. In the training samples, sample A corresponds to the 0.5m heating stage, with a surface temperature of 164.2℃, a deep temperature of 91.5℃, a displacement of 0.4mm, a depth of 0.3mm, a pressure of 28kPa, an echo of 3.2, and a thickness reference value of 49.6mm; sample B corresponds to the 38m loosening stage, with a surface temperature of 143.1℃, a deep temperature of 112.6℃, a displacement of 5.2mm, a depth of 4.1mm, a pressure of 76kPa, an echo of 18.3, and a thickness reference value of 44.7mm; sample C corresponds to the 92m compaction stage, with a surface temperature of 119.4℃, a deep temperature of 111.7℃, a displacement of 1.1mm, a depth of 3.9mm, a pressure of 246kPa, an echo of 18.9, and a thickness reference value of 45.6mm. The improved NBEAT's x model converged after 42 training iterations. In three consecutive iterations, the temperature response difference between depth units was less than 0.2℃, the relative change rate of the thickness contribution correction did not exceed 2%, and the update amplitude of the thermal diffusivity was less than 1%. The learnable thermal diffusivity for the heating, loosening, spreading, and compaction stages converged from 0.35, 0.28, 0.22, and 0.15 cm² / s to 0.348, 0.276, 0.219, and 0.153 cm² / s, respectively. During the initial thickness measurement, the model divided the 0cm to 10cm depth range into 11 depth units of 1cm each. Time steps of 0.3s, 0.2s, or 0.1s were selected based on the temperature transfer intensity, and the longitudinal temperature field was solved in sets of 20 time steps. Once the temperature change at the depth nodes fell below 0.2℃, the system recorded the stable temperature field.The temperatures at the 11 depth nodes after stabilization at 38m were 143.1℃, 139.4℃, 135.0℃, 129.6℃, 123.2℃, 116.4℃, 113.1℃, 111.6℃, 110.9℃, 110.5℃, and 110.3℃, respectively. The first consecutive temperature drop exceeding 2℃ occurred between 4cm and 5cm, with the interface isothermal depth identified as 4.4cm, and the initial thickness measured by the model output as 44.9mm. At 68m, the interface isothermal depth was 5.1cm, and the initial thickness measured as 50.6mm. The differentiable thermal conductivity consistency constraint layer compared the interface isothermal depth with the contribution characteristics of the stage thickness. The initial stage weights of 0.30, 0.25, 0.25, and 0.20 were adaptively adjusted to 0.27, 0.28, 0.23, and 0.22, thereby enhancing the influence of the loosening and compaction stages on the abnormal thickness. During the thickness correction stage, the system constructs a thickness residual field. At 38m, the initial thickness measurement is 44.9mm, the corresponding value after thickness evolution is 46.3mm, and the residual is -1.4mm, recorded as a first-level deviation in the reverse direction. At 44m, the initial thickness measurement is 45.0mm, the corresponding value after thickness evolution is 48.2mm, and the residual is -3.2mm, recorded as a second-level deviation in the reverse direction. At 46m, the initial thickness measurement is 45.7mm, the corresponding value after thickness evolution is 49.1mm, and the residual is -3.4mm, recorded as a second-level deviation in the reverse direction. After the error propagates along the construction path, the system associates 38m to 46m as the same error backtracking area and determines it as a material migration deviation based on displacement (level 3), depth (level 3), and echo (level 3), and determines the existence of a compaction state deviation based on pressure (level 2) and depth (level 3). During compensation, 20% of the maximum residual of 3.4mm in the region was used as the upper limit, and the maximum compensation range for a single point was 0.68mm. After compensation, the thicknesses at 38m, 44m, and 46m were corrected from 44.9mm, 45.0mm, and 45.7mm to 45.4mm, 45.6mm, and 46.2mm, respectively. Similarly, an error backtracking region was generated for the 82m to 94m section, and the thickness at 92m was corrected from 45.8mm to 46.3mm. Finally, the system performed continuity processing on the thickness correction sequence. Differences between adjacent sampling points not exceeding 3mm were considered continuous changes, 3mm to 7mm were considered gradual changes, and differences exceeding 7mm were considered abrupt changes. A total of 6 isolated abrupt change points were identified, of which 5 were smoothed out by the neighborhood window, and 1 was retained as an outlier due to its repeated occurrence in 3 consecutive time slices. The final output included 241 continuous thickness results, with the thickness anomaly sections being 36m to 48m and 82m to 94m. Compared with traditional methods, the average absolute error of 12 core samples is 3.8 mm, the maximum error is 8.6 mm, and the abnormal section identification rate is 58.3%; the average absolute error of 24 manual single-point measurement is 5.1 mm, the maximum error is 11.4 mm, and the abnormal section identification rate is 45.8%; the average absolute error of 241 points using the method of this invention is 1.6 mm, the maximum error is 4.3 mm, the abnormal section identification rate is 91.7%, and it does not damage the road surface.

[0230] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology, characterized in that, include: Multi-source state sensing units are deployed on the on-site thermal recycling construction equipment to collect multi-source state data during the road heating, loosening, paving and compaction processes, and perform preprocessing to generate a construction state data set. Based on the construction status data set, temperature change features and temperature transmission features are extracted, and displacement, depth, pressure and echo response features are also extracted to generate a multi-source pavement status feature sequence. Based on the multi-source pavement state feature sequence, a thickness evolution map of the recycled layer is established. The pavement state at each sampling time is taken as the evolution node, and temperature transfer behavior, material migration behavior and compaction behavior are taken as the connection relationship to generate the thickness evolution path. The dominant evolution path is identified and the corresponding thickness evolution state sequence is generated. The multi-source pavement state feature sequence, construction stage identification sequence and temperature transfer feature sequence are simultaneously input into the improved NBEATSx model with a differentiable heat conduction consistency constraint layer. A temperature-thickness coupling inversion layer is inserted after each level of neural base expansion block to output the initial thickness measurement sequence that meets the heat conduction consistency condition. A thickness residual field is constructed based on the initial thickness measurement sequence. The deviation between each sampling point and the thickness evolution state sequence is calculated. The error backtracking relationship is formed by propagating along the construction path. Error decomposition is performed on abnormal sampling points. The initial thickness measurement sequence is then compensated in reverse to generate a thickness correction sequence. The thickness correction sequence is smoothed and its consistency is checked. The pavement thickness measurement results distributed along the construction direction are output, and the corresponding thickness anomaly section records are generated.

2. The method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology according to claim 1, characterized in that, The multi-source state sensing unit includes a temperature sensor, a displacement sensor, a depth sensor, a pressure sensor, and an echo detection sensor.

3. The method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology according to claim 1, characterized in that, The multi-source state data includes road surface temperature data, displacement data, depth data, pressure data, and echo response data.

4. The method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology according to claim 1, characterized in that, The generation of the multi-source road surface state feature sequence includes: Road surface temperature data is read from the construction status dataset, and the road surface temperature data is arranged according to the sampling time and construction mileage location to generate a basic temperature sequence; Based on the temperature base sequence, extract the temperature change amplitude, temperature change direction, temperature decay state, and temperature difference between different sampling locations between adjacent sampling times to generate a temperature change feature set. Based on the temperature change feature set, the heat transfer state between the road surface layer and the regenerated layer is identified, and the temperature transfer direction, temperature transfer intensity and temperature attenuation continuity are extracted to generate a temperature transfer feature sequence. Displacement data, depth data, pressure data, and echo response data are read from the construction status data set and aligned with the sampling time and construction mileage location corresponding to the temperature transfer characteristic sequence to generate a structural response characteristic set. The temperature change feature set, temperature transfer feature sequence, and structural response feature set are combined and sorted according to sampling time, construction mileage location, and construction stage identifier, and a corresponding feature record is established for each sampling location to generate a multi-source pavement state feature sequence.

5. The method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology according to claim 1, characterized in that, The generation of thickness evolution paths, identifying the active evolutionary path and generating the corresponding thickness evolution state sequence, includes: Read the temperature transfer characteristics, displacement characteristics, depth characteristics, pressure characteristics, echo response characteristics and construction stage identifiers corresponding to each sampling time in the multi-source pavement state characteristic sequence, and establish pavement state nodes according to the sampling time and construction mileage location; The temperature transfer state, material position change state, compression forming state and echo interface state corresponding to each road surface state node are written into the node attributes to generate an evolution node set used to characterize the formation process of the recycled layer thickness. Temperature transfer connection is established based on temperature transfer changes between adjacent evolution nodes; material migration connection is established based on displacement changes, depth changes, and echo response changes between adjacent evolution nodes; and compaction connection is established based on pressure changes, depth changes, and echo interface changes between adjacent evolution nodes. Evolution nodes with temperature transfer, material migration, and compaction connections are connected in ascending order of construction mileage to form a thickness evolution diagram of the recycled layer. A set of thickness evolution paths is generated based on the continuous connection status between the evolution nodes. The thickness evolution path with complete connection relationship, longest continuous sampling length and consistent echo interface change with depth change is selected from the thickness evolution path set as the master evolution path. The temperature transfer state, material migration state, compaction state and construction stage identifier of each evolution node in the master evolution path are extracted to generate the thickness evolution state sequence.

6. The method for measuring the thickness of in-situ thermal regeneration pavement based on intelligent sensing technology according to claim 1, characterized in that, The output, which satisfies the thermal conductivity consistency condition, includes the following initial thickness measurement sequence: Using the NBEATSx model as the basic model, a feature synchronization input structure is set at the input end of the NBEATSx model to synchronously input the multi-source pavement state feature sequence, construction stage identification sequence and temperature transfer feature sequence according to the sampling time and construction mileage location. In the neural basis expansion block of the NBEATSx model, the trend term mapping structure and residual backpropagation structure are preserved, and the construction stage identifier sequence is used as stage constraint information to extract the stage thickness contribution features corresponding to each construction stage. A temperature-thickness coupling inversion layer is embedded after each level of neural basis expansion block. The temperature-thickness coupling inversion layer reads the stage thickness contribution feature and temperature transfer feature sequence to generate the corresponding learnable thermal diffusivity. The longitudinal temperature field from the surface layer to the interior of the recycled layer is solved based on the learnable thermal diffusivity coefficient and temperature transfer characteristic sequence. The isothermal depth of the recycled layer interface is identified based on the continuous position of temperature change in the longitudinal temperature field. In the differentiable heat conduction consistency constraint layer, the difference between the interface isothermal depth and the stage thickness contribution characteristics corresponding to each construction stage is compared to construct a physical consistency loss for constraining the thickness inversion results. The stage weights of each construction stage are adaptively adjusted based on the physical consistency loss, and the stage thickness contribution characteristics are iteratively optimized through the residual backpropagation structure, and the initial thickness measurement sequence that meets the thermal conduction consistency condition is fused to generate. A training dataset is constructed based on multi-source pavement state feature sequences, initial thickness measurement sequences, and known thickness reference values. A comprehensive loss function consisting of physical consistency loss and prediction error weighting is used to backpropagate and update parameters of the improved NBEATSx model in batches until the comprehensive loss converges, and the trained model parameters are saved.

7. The method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology according to claim 6, characterized in that, The generation of the corresponding learnable thermal diffusivity includes: Within the temperature-thickness coupled inversion layer, the thermal diffusivity parameter is initialized for each construction stage, and the thermal diffusivity parameter, along with the stage thickness contribution characteristics and temperature transfer characteristic sequence, is written into the coupled inversion calculation unit. The coupled inversion calculation unit calculates the temperature response difference based on the heat transfer relationship between the stage thickness contribution characteristics and the temperature transfer characteristic sequence, and feeds the temperature response difference back to the thermal diffusivity parameter as an update signal; The thermal diffusivity parameter is iteratively adjusted using the update signal at a set learning rate, so that the temperature response difference tends to be minimized in continuous iterations. When the temperature response difference meets the convergence criteria, the converged thermal diffusivity parameter value is recorded as the learnable thermal diffusivity for the corresponding construction stage.

8. The method for measuring the thickness of in-situ thermally recycled pavement based on intelligent sensing technology according to claim 1, characterized in that, The step of performing reverse compensation on the initial thickness measurement sequence to generate a thickness correction sequence includes: Read the initial thickness measurement sequence and the thickness evolution state sequence, match them according to the sampling time, construction mileage location and construction stage identifier, and establish the correspondence between the initial thickness measurement value and the thickness evolution state of each sampling point; Write the initial thickness measurement, temperature transfer status, material migration status, compaction status, and construction stage identifier of each sampling point into the same residual record. Construct a path stage joint index according to the construction mileage location and construction stage sequence to generate a thickness residual field. Based on the thickness residual field, the initial thickness measurement value and thickness evolution state corresponding to each sampling point are read. The thickness formation trend represented by the initial thickness measurement value and the thickness evolution state are compared to determine the deviation direction, deviation magnitude and deviation duration of each sampling point. The deviation direction, deviation magnitude, and deviation duration of each sampling point are transmitted to adjacent sampling points along the construction path, and continuous deviation sampling points and abrupt deviation sampling points are associated to form an error backtracking relationship. Based on the error backtracking relationship, the error source decomposition is performed on the abnormal sampling points. The deviation state corresponding to the abnormal sampling points is divided into temperature transmission deviation, material migration deviation and compaction state deviation, and an error source record is generated. Based on the error source record, the compensation direction of abnormal sampling points and related sampling points is determined. Reverse compensation is performed on the initial thickness measurement sequence, and the sequence is rearranged according to the sampling time and construction mileage to generate a thickness correction sequence.

9. The method for measuring the thickness of in-situ thermal regeneration pavement based on intelligent sensing technology according to claim 1, characterized in that, The output includes pavement thickness measurement results distributed along the construction direction, and generates corresponding thickness anomaly section records, including: Read the thickness correction sequence, sort the corrected thickness values ​​of each sampling point according to the sampling time, construction mileage location and construction stage identifier, and generate a construction path thickness arrangement sequence; Based on the thickness arrangement sequence of the construction path, the continuous thickness change state between adjacent sampling points is identified, and neighborhood constraint smoothing is performed on isolated abrupt thickness values ​​to generate a continuous thickness distribution sequence. The continuous thickness distribution sequence is matched with the thickness evolution state sequence to determine whether the thickness change direction, thickness change amplitude and construction stage identification of each sampling point are consistent, and a thickness consistency verification result is generated. Based on the thickness consistency verification results, abnormal thickness sampling points are identified. Sampling points with consecutive abnormalities are merged into thickness abnormal sections according to the construction mileage location, and the start and end positions and abnormality types of abnormal sections are recorded. The continuous thickness distribution sequence is used to generate pavement thickness measurement results according to the construction direction, and the sampling time, construction mileage location, construction stage identifier and correction thickness value corresponding to each sampling point are written into the thickness measurement results. The thickness measurement results, thickness consistency verification results, and thickness anomaly sections are associated and stored to generate a thickness anomaly section record for construction process control and quality evaluation.

Images