Method for estimating asphalt pavement segregation layer thickness in road and bridge construction based on machine learning

Abstract

The application discloses a kind of road bridge construction asphalt pavement segregation layer thickness estimation method based on machine learning, S1. get aligned multi-source non-destructive testing data set;S2. generate complex spectral domain tensor;S3. get initial fourier neural operator model;S4. get optimized fourier neural operator model;S5. generate thickness heat map, and mark super threshold region as suspected segregation area in thickness heat map;S6. when core drilling measurement result is obtained, measurement result is compared with suspected segregation area, and henry gas solubility optimization algorithm is triggered again optimization, updates fourier neural operator model, when updated fourier neural operator model converges, output latest segregation layer thickness continuous field, latest thickness heat map and suspected segregation area list, complete asphalt pavement segregation layer thickness estimation.The application effectively improves the accuracy and spatial positioning ability of thickness abnormal area identification, provides reliable technical foundation for pavement rapid inspection and accurate reinforcement.

Description

Technical Field

[0001] This invention relates to the field of construction technology, and in particular to a method for estimating the thickness of asphalt pavement segregation layer in road and bridge construction based on machine learning. Background Technology

[0002] In recent years, with the continuous expansion of highway construction and asphalt pavement maintenance, the road engineering field has put forward higher requirements for non-destructive testing technology of pavement structure quality. As an important indicator for measuring paving uniformity, structural durability and service performance, the thickness of asphalt pavement segregation layer directly affects the later service life of the pavement and traffic safety. At present, the detection of asphalt pavement segregation layer thickness on the engineering site mainly relies on destructive sampling methods such as core drilling and joint cutting. Although it can obtain accurate point thickness information, it has drawbacks such as inconsistency between points and surfaces, long detection cycle, high manpower consumption and obvious damage to pavement structure, making it difficult to achieve large-scale, real-time and continuous pavement structure health assessment.

[0003] However, existing nondestructive testing technologies still have many shortcomings in actual thickness inversion and spatial distribution reconstruction. On the one hand, the accuracy of segregation layer thickness estimation by single signal paths of ground-penetrating radar and infrared detection is limited by external conditions such as wave velocity assumptions, temperature drift, and changes in material parameters, which easily leads to significant systematic errors. Typically, the standard deviation of their thickness estimation is difficult to exceed ±8 mm, failing to meet the requirements of high-precision engineering applications. On the other hand, the fusion of multimodal detection data and high-dimensional inversion modeling techniques are still imperfect. Mainstream methods mostly rely on traditional empirical formulas, machine learning regression, or static parameter optimization, making it difficult to maintain good generalization ability under complex climates, different material ratios, and varied construction processes. In addition, although existing deep learning methods have shown good results in some specific scenarios, their adaptability and interpretability to large-scale heterogeneous data, strong environmental disturbances, and physical constraints in engineering fields are limited. Model transfer and fine-tuning efficiency is low, and they are prone to overfitting or sharp drops in accuracy.

[0004] Therefore, there is an urgent need to develop a new method for estimating the thickness of the pavement segregation layer that is oriented towards engineering scenarios and has physical-driven, data fusion, and adaptive optimization capabilities. Summary of the Invention

[0005] One objective of this invention is to propose a machine learning-based method for estimating the thickness of asphalt pavement segregation layers in road and bridge construction. This invention effectively improves the accuracy of identifying areas with abnormal thickness and enhances spatial positioning capabilities, providing a reliable technical foundation for rapid pavement inspection and precise reinforcement.

[0006] A method for estimating the thickness of asphalt pavement segregation layer in road and bridge construction based on machine learning, according to an embodiment of the present invention, includes the following steps:

[0007] S1. Synchronously collect asphalt pavement data, construct a multi-source non-destructive testing dataset, and perform preprocessing to obtain an aligned multi-source non-destructive testing dataset;

[0008] S2. Perform a Fourier transform on the aligned multi-source nondestructive testing dataset to generate a complex spectral domain tensor;

[0009] S3. Construct a Fourier neural operator model and perform end-to-end training on the complex spectral domain tensor to obtain the initial Fourier neural operator model.

[0010] S4. Initialize the Henry's gas solubility optimization algorithm population, define the model parameter optimization space and the Fourier neural operator model optimization variables, use the complex spectral tensor as the fitness index, execute the Henry's gas solubility optimization algorithm, and obtain the optimized Fourier neural operator model.

[0011] S5. Apply the optimized Fourier neural operator model to the complex spectral domain tensor to generate a continuous field of segregation layer thickness, generate a thickness heatmap, and mark the region exceeding the threshold in the thickness heatmap as a suspected segregation region;

[0012] S6. After the core drilling measurement results are obtained, the measurement results are compared with the suspected segregation areas, and the Henry gas solubility optimization algorithm is triggered to optimize again, and the Fourier neural operator model is updated. When the updated Fourier neural operator model converges, the latest segregation layer thickness continuous field, the latest thickness heat map and the list of suspected segregation areas are output to complete the estimation of the asphalt pavement segregation layer thickness.

[0013] Optionally, S1 includes the following steps:

[0014] S11. Collect sets of ground-penetrating radar echo signals, line scan image data, environmental parameter data, and location information;

[0015] S12. Perform wavelet domain filtering and reflection waveform denoising on the set of ground-penetrating radar echo signals to obtain the denoised set of ground-penetrating radar echo signals.

[0016] S13. Perform image preprocessing on the line scan image dataset, including enhancing the contrast of image colors, correcting geometric distortion of image edges, and standardizing brightness consistency to obtain a standardized image dataset;

[0017] S14. Perform normalization processing on the ambient temperature data to obtain a set of normalized temperature values;

[0018] S15. Combine the denoised ground-penetrating radar echo signal set, standardized image set, normalized temperature value set, original ambient air pressure data, material viscosity data, and location information. Using a unified time label as an index, point-by-point registration is performed to construct an aligned multi-source non-destructive testing dataset;

[0019] S16. Set a unified reference time base in the aligned multi-source non-destructive testing dataset, and perform synchronous correction on all time labels so that the time difference between the timestamps of all sampling points and the reference time base is less than the set maximum allowable time difference threshold. This completes the synchronous correction of multi-source data in the time dimension and yields the final multi-source non-destructive testing dataset.

[0020] Optionally, S2 includes the following steps:

[0021] S21. Perform a one-dimensional Fourier transform operation on the ground radar echo signal of each measurement point in the denoised ground radar echo signal set to obtain the ground radar complex spectral domain signal corresponding to each measurement point.

[0022] S22. Perform a two-dimensional Fourier transform operation on the line scan image of each measurement point in the standardized image set of the final multi-source nondestructive testing dataset to obtain the complex spectral domain tensor of the line scan image corresponding to each measurement point;

[0023] S23. Perform a one-dimensional Fourier transform operation on the normalized temperature value sequence of each measurement point in the normalized temperature value set in the final multi-source non-destructive testing dataset to obtain the temperature complex spectral domain signal corresponding to each measurement point;

[0024] S24. The complex spectral domain signals of ground-penetrating radar, the complex spectral domain tensors of line scan images, and the complex spectral domain signals of temperature are integrated with the original environmental air pressure data, material viscosity data, and location information of the measurement points to construct a complex spectral domain tensor dataset.

[0025] Optionally, S3 includes the following steps:

[0026] S31. Using the complex spectral domain tensor dataset as input, call the subband divider to perform interval division on the frequency dimension of the complex spectral domain signal at each measurement point, generating a set of M non-overlapping subbands including low-frequency subband, mid-frequency subband and high-frequency subband.

[0027] S32. Construct an independent sub-band convolution path for each sub-band set, and initialize the sub-band convolution kernel weight matrix, bias vector and activation function parameters respectively to form a multi-subband parallel convolutional engine;

[0028] S33. Based on the impedance ratio coefficient, normalized temperature value, ambient air pressure value and material viscosity value of each measurement point, call the physical control gate to generate the corresponding sub-band gate coefficient, and couple the sub-band gate coefficient with the multi-sub-band parallel convolutional unit sub-band by sub-band to obtain the physically controlled gated sub-band convolution output.

[0029] S34. Perform inverse Fourier transform on the convolutional output of the physically controlled gated subbands to obtain the temporal characteristics of each subband;

[0030] S35. Call the residual merger to sum the time-domain features of each sub-band according to the gating coefficients, and perform residual connection with the lifted representation of the original complex spectral domain input features to generate a fused feature vector;

[0031] S36. Construct training samples using fused feature vectors and the true values ​​of the separation layer thickness at measurement points, and iteratively optimize the physical control type multi-subband Fourier neural operator using a joint loss function until the joint loss function converges;

[0032] S37. Solidify the converged weight parameters and output an initial physical control type multi-subband Fourier neural operator model containing a subband divider, a multi-subband parallel convolutional unit, a physical control gate, and a residual merger.

[0033] Optionally, S4 includes the following steps:

[0034] S41. Define the set of parameters to be optimized for the physical control type multi-subband Fourier neural operator model. The set of parameters to be optimized includes the spectral depth, subband frequency cutoff threshold, regularization factor, and channel weight coefficient set of multi-source non-destructive testing signals. The parameters to be optimized are uniformly represented as a parameter vector to be optimized.

[0035] S42. Initialize the candidate solution population of Henry's gas solubility optimization algorithm. Each candidate solution in the candidate solution population contains a set of parameters to be optimized. Associate each candidate solution in the population with its corresponding set of physical environment parameters. The set of physical environment parameters includes normalized temperature value, ambient air pressure value and material viscosity value.

[0036] S43. Construct a solubility-partial pressure control mechanism. In each evolution of the Henry's gas solubility optimization algorithm, calculate the solubility coefficient and partial pressure factor of the candidate solution based on the set of physical environment parameters corresponding to each candidate solution. Map the solubility coefficient to the global search probability of the candidate solution and map the partial pressure factor to the local exploitation probability of the candidate solution.

[0037] S44. The spectral depth, subband frequency cutoff threshold, regularization factor and channel weight parameters contained in each candidate solution are applied to the physical control type multi-subband Fourier neural operator model. The multi-subband Fourier neural operator model is used to predict and infer the thickness field of the asphalt pavement segregation layer on the complex spectral domain tensor dataset. After the inference is completed, the root mean square error index of the thickness estimation residual, the single batch inference time index and the memory consumption index corresponding to the candidate solution are evaluated. The candidate solutions are then weighted and combined according to the set weight coefficients to generate the comprehensive fitness index corresponding to the candidate solutions.

[0038] S45. Based on the comprehensive fitness index of each candidate solution and the corresponding solubility coefficient and partial pressure factor, perform the evolutionary iteration operation of Henry's gas solubility optimization algorithm. Guide candidate solutions with fitness higher than the threshold to a new search area according to the solubility coefficient to optimize the global search capability. Perform perturbation and local mutation operations on candidate solutions with fitness lower than the threshold according to the partial pressure factor. During each iteration, dynamically adjust the search range and step size in combination with the environmental parameters of the current candidate solution.

[0039] S46. After each iteration, determine whether the comprehensive fitness index corresponding to the best candidate solution in the current population meets the preset convergence criterion. If the convergence criterion is met, select the candidate solution as the best solution; if the convergence criterion is not met, continue to execute the next round of population evolution until the convergence criterion is met.

[0040] S47. The spectral depth, subband frequency cutoff threshold, regularization factor, and multi-source signal channel weight configuration obtained from the optimal candidate solution are applied to the physical control type multi-subband Fourier neural operator model to form an optimized Fourier neural operator model.

[0041] Optionally, the root mean square error index of the thickness estimation residual is used to calculate the root mean square error between the predicted value and the actual value of the segregation layer thickness.

[0042] The single-batch inference time consumption index is the average time required for the Fourier neural operator model to perform a single round of thickness prediction inference.

[0043] The aforementioned memory consumption metric is a statistical measure of the amount of video memory used by the Fourier neural operator model during execution.

[0044] Optionally, S5 includes the following steps:

[0045] S51. Load the optimized Fourier neural operator model into the segregation layer thickness estimation task. With the complex spectral domain tensor dataset as input, call the optimized Fourier neural operator model to perform thickness prediction operation on each measurement point in the complex spectral domain tensor dataset to obtain the segregation layer thickness prediction value corresponding to the measurement point. The segregation layer thickness prediction value is used to estimate the actual asphalt segregation layer thickness at the measurement point.

[0046] S52. Arrange the predicted segregation layer thickness values ​​of all measurement points in order according to the measurement point number to construct a segregation layer thickness prediction sequence;

[0047] S53. Bind each predicted thickness value in the segregation layer thickness prediction sequence to the spatial location information of the corresponding measurement point, and construct a segregation layer thickness field mapping function. The segregation layer thickness field mapping function is used to describe the predicted thickness value corresponding to each location in two-dimensional space.

[0048] S54. Visualize the segregation layer thickness field mapping function in a two-dimensional spatial coordinate system to generate a thickness heat map image. The thickness heat map image expresses the predicted value of the segregation layer thickness at each spatial location in the form of color depth.

[0049] S55. Set a threshold range for judging the thickness of the segregation layer. The threshold range for judging the thickness of the segregation layer includes a lower threshold and an upper threshold. Extract the spatial coordinates corresponding to the positions of all thickness prediction values ​​that are less than the lower threshold or greater than the upper threshold to form a set of coordinates of suspected segregation areas.

[0050] S56. Graphically mark all location points in the coordinate set of suspected segregation areas to form a suspected segregation area annotation layer;

[0051] S57. Overlay the thickness heat map image with the labeled layer of the suspected segregation area to generate the final visualization image of the segregation layer thickness.

[0052] Optionally, S6 includes the following steps:

[0053] S61. Extract the true value of the segregation layer thickness of each core point in the core measurement results, and extract the spatial location information corresponding to each core point. Perform spatial matching of the spatial location information of all core points with the coordinate set of the suspected segregation area, and filter out the core points that fall within the range of the suspected segregation area to obtain the set of core points located in the suspected segregation area.

[0054] S62. Combine the true value of the segregation layer thickness of each core point in the set of core points located in the suspected segregation region with the corresponding complex spectral domain input feature of the core point in the complex spectral domain tensor dataset to construct an incremental supervision sample set;

[0055] S63. Add the incremental supervised sample set to the original training sample set to form a new training sample set;

[0056] S64. Using the new training sample set as the model training input, reinitialize the candidate solution population of the Henry's gas solubility optimization algorithm, and calculate the comprehensive fitness index of each candidate solution under the new training sample set in each evolution process.

[0057] S65. Form an updated Fourier neural operator model, predict the thickness of all measurement points in the complex spectral domain tensor dataset, and obtain a new continuous field for the thickness of the segregation layer.

[0058] S66. Based on the new segregation layer thickness continuous field and the spatial location information of each measurement point, generate the latest thickness thermal image and form the latest list of suspected segregation areas;

[0059] S67. Output the updated segregation layer thickness continuous field, the latest generated thickness heatmap image, and the updated list of suspected segregation areas as the segregation layer thickness estimation results for asphalt pavement.

[0060] The beneficial effects of this invention are:

[0061] This invention introduces a physically controlled multi-subband Fourier neural operator structure to achieve dynamic gating and adaptive modulation of the spectrum modeling process based on environmental physical variables. The complex spectral domain tensor dataset is divided into multiple sub-bands according to the frequency dimension, and an independent convolution path and gating mechanism are established for each sub-band to achieve multi-scale response to different physical environments and material properties. This enhances the model's adaptability and robustness to the changing working conditions of asphalt pavement and significantly improves spatial resolution and thickness estimation accuracy.

[0062] This invention proposes a multi-objective model parameter adaptive optimization mechanism based on the Henry's gas solubility optimization algorithm. By utilizing the solubility-partial pressure control principle, environmental temperature, gas pressure, and material viscosity parameters are directly mapped to the global search probability and local exploitation probability in the optimization algorithm. This effectively achieves dynamic self-optimization of high-dimensional hyperparameters such as spectral depth, subband frequency cutoff threshold, regularization factor, and multi-source channel weights. It realizes the physical drive and engineering adaptation of parameter optimization, ensuring that the model can obtain optimal inference accuracy and real-time inference efficiency under different regional, climatic, and construction mix conditions.

[0063] This invention establishes a complete engineering closed-loop process: complex spectral domain data → physically controlled multi-subband Fourier neural operator → dynamic optimization → thickness heatmap and visualization of suspected areas → core drilling incremental closed-loop self-learning. By combining the actual thickness values ​​of the core drilling points with the features of the complex spectral domain tensor dataset as incremental supervision samples and continuously incorporating them into the training system, and combining the re-optimization mechanism of the Henry's gas solubility optimization algorithm, the model can achieve online adaptive updates under the background of continuous changes in road network, season, climate and materials. It maintains high-precision generalization ability for heterogeneous data and complex construction scenarios. The results of actual engineering applications show that it effectively improves the accuracy of thickness anomaly area identification and spatial positioning ability, providing a reliable technical foundation for rapid road inspection and precise reinforcement. Attached Figure Description

[0064] 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:

[0065] Figure 1 This is a flowchart of a machine learning-based method for estimating the thickness of asphalt segregation layer in road and bridge construction, as proposed in this invention.

[0066] Figure 2This is a modular schematic diagram of the physical control type multi-subband Fourier neural operator structure in the machine learning-based method for estimating the thickness of asphalt pavement segregation layer in road and bridge construction proposed in this invention. Detailed Implementation

[0067] 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.

[0068] refer to Figures 1-2 A machine learning-based method for estimating the thickness of asphalt pavement segregation layer in road and bridge construction includes the following steps:

[0069] S1. Simultaneously acquire ground radar signals, line scan visual images, location information and environmental parameters of asphalt pavement, construct a multi-source non-destructive testing dataset, perform denoising, geometric correction, temperature normalization and spatiotemporal registration operations on the multi-source non-destructive testing dataset, and complete synchronization alignment based on a unified clock to obtain an aligned multi-source non-destructive testing dataset.

[0070] S2. Perform a Fourier transform on the aligned multi-source nondestructive testing dataset to generate a complex spectral tensor and save the coordinate index of the complex spectral tensor;

[0071] S3. Construct a Fourier neural operator model and perform end-to-end training on the complex spectral domain tensor to obtain the initial Fourier neural operator model.

[0072] S4. Initialize the Henry's gas solubility optimization algorithm population, set the solubility-partial pressure parameter to map ambient temperature, partial pressure and material viscosity, define the model parameter optimization space, use the spectral depth, frequency cutoff threshold, regularization factor and multi-source weight of the Fourier neural operator model as optimization variables, use the thickness reconstruction residual of the complex spectral domain tensor, inference time and memory consumption as fitness indicators, execute the Henry's gas solubility optimization algorithm, and obtain the optimized Fourier neural operator model;

[0073] S5. The optimized Fourier neural operator model is applied to the complex spectral domain tensor to generate a continuous field of segregation layer thickness. The continuous field of segregation layer thickness is bound to the location information to generate a thickness heatmap. The region exceeding the threshold in the thickness heatmap is marked as a suspected segregation region.

[0074] S6. After the core drilling measurement results are obtained, the measurement results are compared with the suspected segregation areas to generate an incremental supervision sample set. The incremental supervision sample set is added to the multi-source non-destructive testing dataset, triggering the Henry's gas solubility optimization algorithm to optimize again and update the Fourier neural operator model. When the updated Fourier neural operator model converges, the latest segregation layer thickness continuous field, the latest thickness heat map, and the list of suspected segregation areas are output to complete the estimation of the asphalt pavement segregation layer thickness.

[0075] In this embodiment, S1 includes the following steps:

[0076] S11. Collect sets of ground-penetrating radar echo signals, line scan image data, environmental parameter data, and location information;

[0077] The set of ground-penetrating radar echo signals is used to characterize the reflected waveforms of each measurement point; the set of line scan image data represents the road surface thermal imaging image at the corresponding location; the set of environmental parameter data includes ambient temperature data, ambient air pressure data, and material viscosity data; the set of location information represents the latitude and longitude coordinates of the corresponding location; and each measurement point has a unique number.

[0078] S12. Perform wavelet domain filtering and reflection waveform denoising on the set of ground-penetrating radar echo signals to obtain a denoised set of ground-penetrating radar echo signals. Each denoised ground-penetrating radar echo signal is obtained by filtering based on the energy change characteristics of each point in the original signal.

[0079] S13. Perform image preprocessing on the line scan image dataset, including enhancing the contrast of image colors, correcting geometric distortion of image edges, and standardizing brightness consistency to obtain a standardized image dataset;

[0080] S14. Perform normalization processing on the ambient temperature data to obtain a set of normalized temperature values. Each normalized temperature value is obtained by subtracting the minimum temperature value in the entire sample sequence from the original temperature value at that location and then dividing by the difference between the maximum and minimum temperature values ​​in the sample sequence.

[0081] S15. Combine the denoised ground-penetrating radar echo signal set, standardized image set, normalized temperature value set, original ambient air pressure data, material viscosity data, and location information. Using a unified time label as an index, point-by-point registration is performed to construct an aligned multi-source non-destructive testing dataset;

[0082] S16. Set a unified reference time base in the aligned multi-source non-destructive testing dataset, and perform synchronous correction on all time labels so that the time difference between the timestamps of all sampling points and the reference time base is less than the set maximum allowable time difference threshold. This completes the synchronous correction of multi-source data in the time dimension and yields the final multi-source non-destructive testing dataset.

[0083] In this embodiment, S2 includes the following steps:

[0084] S21. Perform a one-dimensional Fourier transform operation on the ground radar echo signal of each measurement point in the denoised ground radar echo signal set to obtain the ground radar complex spectral domain signal corresponding to each measurement point. The ground radar complex spectral domain signal is used to represent the spectral characteristics of the radar echo at different frequencies of the measurement point.

[0085] The calculation of the complex spectral domain signal of the ground-penetrating radar is to perform a Fourier integral operation on the denoised radar waveform signal corresponding to the measurement point. The Fourier integral operation constructs a frequency mapping by introducing a complex exponential kernel function with imaginary units and frequency factors, and completely converts the time domain signal into a frequency domain representation.

[0086] S22. Perform a two-dimensional Fourier transform operation on the line scan image of each measurement point in the standardized image set of the final multi-source nondestructive testing dataset to obtain the complex spectral tensor of the line scan image corresponding to each measurement point. The complex spectral tensor of the line scan image is used to characterize the spectral distribution characteristics of the measurement point image in different spatial frequency directions.

[0087] The calculation of the complex spectral domain tensor of the line scan image involves performing Fourier integral operations on the standardized image of the measurement points in both the horizontal and vertical spatial directions. The Fourier integral operation completes the frequency mapping in two-dimensional space by constructing a complex exponential kernel function for the two-dimensional frequency variable.

[0088] S23. Perform a one-dimensional Fourier transform operation on the normalized temperature value sequence of each measurement point in the normalized temperature value set in the final multi-source non-destructive testing dataset to obtain the temperature complex spectral domain signal corresponding to each measurement point. The temperature complex spectral domain signal is used to characterize the dynamic response characteristics of the temperature sequence of the measurement point in the frequency domain.

[0089] The calculation of the complex temperature spectrum domain signal involves performing a Fourier integral operation on the normalized temperature value of the measurement point along the time series direction. The Fourier integral operation is achieved by multiplying the normalized temperature signal with the complex exponential kernel function of the frequency factor and integrating it over the entire time domain.

[0090] S24. The complex spectral domain signals of ground-penetrating radar, the complex spectral domain tensors of line scan images, and the complex spectral domain signals of temperature are integrated with the original environmental air pressure data, material viscosity data, and location information of the measurement points to construct a complex spectral domain tensor dataset. The complex spectral domain tensor dataset maintains the same measurement point number and location information index as the final multi-source non-destructive testing dataset.

[0091] In this embodiment, S3 includes the following steps:

[0092] S31. Using the complex spectral domain tensor dataset as input, call the subband divider to perform interval division on the frequency dimension of the complex spectral domain signal at each measurement point, generating a set of M non-overlapping subbands including low-frequency subband, mid-frequency subband and high-frequency subband.

[0093] S32. Construct an independent sub-band convolution path for each sub-band set, and initialize the sub-band convolution kernel weight matrix, bias vector and activation function parameters respectively to form a multi-subband parallel convolutional engine;

[0094] S33. Based on the impedance ratio coefficient, normalized temperature value, ambient air pressure value and material viscosity value of each measurement point, call the physical control gate to generate the corresponding sub-band gate coefficient, and couple the sub-band gate coefficient with the multi-sub-band parallel convolutional unit sub-band by sub-band to obtain the physically controlled gated sub-band convolution output.

[0095] S34. Perform inverse Fourier transform on the convolutional output of the physically controlled gated subbands to obtain the temporal characteristics of each subband;

[0096] S35. Call the residual merger to sum the time-domain features of each sub-band according to the gating coefficients, and perform residual connection with the lifted representation of the original complex spectral domain input features to generate a fused feature vector;

[0097] S36. Construct training samples using fused feature vectors and the true values ​​of the separation layer thickness at measurement points, and iteratively optimize the physical control type multi-subband Fourier neural operator using a joint loss function until the joint loss function converges;

[0098] In this invention, the joint loss function is constructed by a weighted combination of thickness prediction error and physical prior error. The joint loss function takes the mean square error between the predicted and actual thickness values ​​of the segregation layer as the main loss term, and introduces the prediction error of the impedance ratio physical quantity calculated based on the characteristics of the ground-penetrating radar spectral domain as an auxiliary loss term. By setting weight coefficients to balance each loss term, the joint loss function can simultaneously constrain the thickness estimation accuracy and the consistency of physical characteristics during the model optimization process, effectively improving the model's generalization ability and adaptability to engineering practice.

[0099] S37. Solidify the converged weight parameters and output an initial physical control type multi-subband Fourier neural operator model containing a subband divider, a multi-subband parallel convolutional unit, a physical control gate, and a residual merger.

[0100] In this embodiment, S4 includes the following steps:

[0101] S41. Define the set of parameters to be optimized for the physical control type multi-subband Fourier neural operator model. The set of parameters to be optimized includes spectral depth, used to control the network hierarchy structure of the Fourier neural operator model; subband frequency truncation threshold, used to limit the maximum frequency retained in different frequency bands; regularization factor, used to control the penalty strength of the model parameters; and a set of channel weight coefficients for multi-source non-destructive testing signals, corresponding to the channel weights of ground-penetrating radar, line scan image, and temperature channels, respectively. The parameters to be optimized are uniformly represented as a parameter vector to be optimized.

[0102] S42. Initialize the candidate solution population of the Henry's gas solubility optimization algorithm. Each candidate solution in the candidate solution population contains a set of parameter vectors to be optimized. At the same time, assign a set of physical environment parameters to each candidate solution in the population. The set of physical environment parameters includes normalized temperature value, which is used to characterize the thermal state of the measurement point, ambient air pressure value, which is used to characterize the current external pressure, and material viscosity value, which is used to reflect the flowability characteristics of the asphalt mixture.

[0103] S43. Construct a solubility-partial pressure control mechanism. In each evolution of the Henry's gas solubility optimization algorithm, calculate the solubility coefficient and partial pressure factor of each candidate solution based on the normalized temperature value, ambient gas pressure value, and material viscosity value corresponding to each candidate solution. Map the solubility coefficient to the global search probability of the candidate solution and map the partial pressure factor to the local exploitation probability of the candidate solution. This is used to dynamically adjust the search jump scale and mutation control capability of the candidate solution.

[0104] S44. The spectral depth, subband frequency cutoff threshold, regularization factor, and channel weight parameters contained in each candidate solution are applied to the physical control type multi-subband Fourier neural operator model. The multi-subband Fourier neural operator model is used to predict and infer the thickness field of the asphalt pavement segregation layer on the complex spectral domain tensor dataset. After the inference is completed, the root mean square error index of the thickness estimation residual, the single batch inference time index, and the memory consumption index corresponding to the candidate solution are evaluated. The candidate solutions are weighted and combined according to the set weight coefficients to generate the comprehensive fitness index corresponding to the candidate solutions, which is used to evaluate the overall performance of the candidate solutions in terms of accuracy, speed, and resources.

[0105] S45. Based on the comprehensive fitness index of each candidate solution and the corresponding solubility coefficient and partial pressure factor, perform the evolutionary iteration operation of Henry's gas solubility optimization algorithm. Guide candidate solutions with fitness higher than the threshold to a new search area according to the solubility coefficient to optimize the global search capability. Perform perturbation and local mutation operations on candidate solutions with fitness lower than the threshold according to the partial pressure factor to optimize the local search accuracy. During each iteration, dynamically adjust the search range and step size in combination with the environmental parameters of the current candidate solution.

[0106] S46. After each iteration, determine whether the comprehensive fitness index corresponding to the best candidate solution in the current population meets the preset convergence criterion. If the convergence criterion is met, select the candidate solution as the best solution; if the convergence criterion is not met, continue to execute the next round of population evolution until the convergence criterion is met.

[0107] S47. The spectral depth, subband frequency cutoff threshold, regularization factor, and multi-source signal channel weight configuration obtained from the optimal candidate solution are applied to the physical control type multi-subband Fourier neural operator model to form an optimized Fourier neural operator model.

[0108] In this embodiment, the root mean square error index of thickness estimation residual is used to calculate the root mean square error between the predicted value and the actual value of the segregation layer thickness, and is used to quantify the accuracy error of the model.

[0109] The single-batch inference time metric is the average time required for the Fourier neural operator model to perform a single round of thickness prediction inference, used to measure the model's execution efficiency.

[0110] The video memory consumption metric measures the amount of video memory used by the Fourier neural operator model during execution, thus measuring the model's resource consumption.

[0111] In this embodiment, S5 includes the following steps:

[0112] S51. Load the optimized Fourier neural operator model into the segregation layer thickness estimation task. With the complex spectral domain tensor dataset as input, call the optimized Fourier neural operator model to perform thickness prediction operation on each measurement point in the complex spectral domain tensor dataset to obtain the segregation layer thickness prediction value corresponding to the measurement point. The segregation layer thickness prediction value is used to estimate the actual asphalt segregation layer thickness at the measurement point.

[0113] S52. Arrange the predicted segregation layer thickness values ​​of all measurement points in order according to the measurement point number to construct a segregation layer thickness prediction sequence. The segregation layer thickness prediction sequence represents the set of predicted segregation layer thickness values ​​corresponding to each measurement point after prediction. The thickness prediction values ​​correspond one-to-one with the measurement point number.

[0114] S53. Bind each predicted thickness value in the segregation layer thickness prediction sequence to the spatial location information of the corresponding measurement point. The spatial location information includes the horizontal coordinate value and the vertical coordinate value of the measurement point. The binding relationship is used to construct the segregation layer thickness field mapping function. The segregation layer thickness field mapping function is used to describe the predicted thickness value corresponding to each position in two-dimensional space.

[0115] The construction of the segregation layer thickness field mapping function is based on the optimized Fourier neural operator model to infer the complex spectral domain tensor dataset point by point, obtain the predicted value of the segregation layer thickness at each measurement point, and bind it with the spatial location information (horizontal and vertical coordinates) of the measurement point to achieve a one-to-one correspondence between the predicted thickness and the spatial location. The spatial coordinates and thickness prediction values ​​of all measurement points together define the segregation layer thickness field mapping function. The segregation layer thickness field mapping function covers all measurement points in the detection area and reflects the continuous distribution characteristics of the segregation layer thickness in two-dimensional space.

[0116] S54. Visualize the segregation layer thickness field mapping function in a two-dimensional spatial coordinate system to generate a thickness heat map image. The thickness heat map image expresses the predicted value of the segregation layer thickness at each spatial location in the form of color depth. When generating the thickness heat map image, use spatial interpolation method to smooth the transition of thickness values ​​between discontinuous measurement points to obtain a smooth thickness heat map image.

[0117] S55. Set a threshold range for judging the thickness of the segregation layer. The threshold range for judging the thickness of the segregation layer includes a lower threshold and an upper threshold. It is used to determine whether there is a thickness anomaly at a certain location. Extract the spatial coordinates corresponding to the locations where the thickness prediction values ​​are less than the lower threshold or greater than the upper threshold to form a set of coordinates of suspected segregation areas. The set of coordinates of suspected segregation areas is used to represent the spatial locations where segregation anomalies may exist in the thickness prediction results.

[0118] S56. Graphically mark all locations in the coordinate set of suspected segregation areas to form a suspected segregation area annotation layer. The suspected segregation area annotation layer is used to mark locations that exceed the threshold range on the original heat map. During the graphic marking process, the spatial clustering analysis method is used to merge adjacent abnormal locations and close the boundaries to construct a spatially connected boundary of the suspected segregation area, and make the boundary have practical construction reference significance.

[0119] S57. Overlay the thickness heat map image with the suspected segregation area annotation layer to generate the final segregation layer thickness visualization image. The visualization image is used to simultaneously present the thickness distribution in continuous space and the suspected segregation anomaly section, and is used by construction management personnel for interpretation and point-to-point handling.

[0120] In this embodiment, S6 includes the following steps:

[0121] S61. Extract the true value of the segregation layer thickness of each core point in the core measurement results, and extract the spatial location information corresponding to each core point. The spatial location information includes the horizontal and vertical coordinates of the core point. Perform spatial matching of the spatial location information of all core points with the coordinate set of the suspected segregation area, and filter out the core points that fall within the range of the suspected segregation area to obtain the set of core points located in the suspected segregation area.

[0122] S62. Combine the true value of the segregation layer thickness of each core point in the set of core points located in the suspected segregation region with the complex spectral domain input feature corresponding to the core point in the complex spectral domain tensor dataset to construct an incremental supervision sample set. The incremental supervision sample set is used to represent the pairing relationship between the spectral domain feature of the core point in the complex spectral domain tensor dataset and the true thickness label corresponding to the point.

[0123] S63. Add the incremental supervision sample set to the original training sample set to form a new training sample set. The new training sample set is composed of the original training sample set and the incremental supervision sample set, representing a joint sample library containing historical data and newly acquired core data.

[0124] S64. Using the new training sample set as the model training input, reinitialize the candidate solution population of the Henry's gas solubility optimization algorithm, and calculate the comprehensive fitness index of each candidate solution under the new training sample set in each evolution process.

[0125] S65. Determine whether the candidate solution with the best comprehensive fitness index in the current population satisfies the set convergence criterion. If it satisfies the convergence criterion, form an updated Fourier neural operator model. Call the updated Fourier neural operator model to predict the thickness of all measurement points in the complex spectral domain tensor dataset, and obtain a new continuous field of segregation layer thickness. The continuous field of segregation layer thickness is used to describe the estimated value of segregation layer thickness at each measurement point in the spatial range.

[0126] S66. Based on the new segregation layer thickness continuous field and the spatial location information of each measurement point, generate the latest thickness heat map image. The thickness heat map image is used to visualize the changes in the segregation layer thickness value at each location in space. Reset the segregation layer thickness judgment threshold range on the thickness heat map image, and identify all spatial locations exceeding the judgment threshold range as new suspected segregation areas, forming the latest list of suspected segregation areas.

[0127] S67. Output the updated segregation layer thickness continuous field, the latest generated thickness heatmap image, and the updated list of suspected segregation areas as the segregation layer thickness estimation results for asphalt pavement.

[0128] The segregation layer thickness estimation results of asphalt pavement are described by average, extreme value, or distribution. The system obtains the segregation layer thickness estimate at each measurement point and forms a continuous field of segregation layer thickness covering the entire detection area through spatial interpolation. The continuous field of segregation layer thickness provides the numerical distribution of the segregation layer thickness of asphalt pavement at each location in two-dimensional space. The continuous field of segregation layer thickness is rendered as a thickness heatmap image in the form of color gradient. In the thickness heatmap image, the color depth corresponds to different thickness intervals, and the high thickness area and low thickness area in space are intuitively distinguished. By performing discriminant analysis on the thickness heatmap image and the thickness continuous field, combined with the preset thickness judgment threshold, the system automatically identifies all spatial locations where the thickness exceeds the upper limit or is lower than the lower limit, forming a list of suspected segregation areas.

[0129] The example is as follows: The optimized Fourier neural operator model performs full inference on the complex spectral domain tensor to obtain the continuous field of the segregation layer thickness. The thickness heatmap image shows that the overall thickness of this road surface section is distributed between 5mm and 9mm, with a stable color transition. Combined with a set threshold... mm A total of 7 segregation anomalies were detected: red clusters with a thickness of more than 10 mm appeared at mileages 0+975m, 1+235m, and 1+410m; blue patches with a thickness of less than 5 mm appeared at mileages 2+120m and 2+450m, thus obtaining the final estimated thickness of the asphalt pavement segregation layer.

[0130] Example 1: This example uses the maintenance and inspection of section A of the expressway (1 kilometer long and 3.75 meters wide) as a real-world scenario to fully demonstrate the application process and achievements of this invention in actual engineering.

[0131] In this maintenance inspection, the project team used a self-developed intelligent inspection vehicle. The vehicle integrates a multi-band ground-penetrating radar array, a high-resolution line-scanning infrared imager, a laser positioning system, and temperature and humidity environmental monitoring instruments. It is also equipped with the intelligent computing module of this invention (including a GPU edge inference unit). The vehicle travels at a constant speed of 30 km / h and collects ground-penetrating radar signals, line-scanning images, and multi-modal data on environmental temperature and humidity and pavement material viscosity prediction in real time through an integrated data synchronization system. This forms a multi-source non-destructive testing dataset. A total of 12,800 detection points are deployed on a single kilometer section, with an average point spacing of about 7.8 centimeters. The spatial coverage resolution is higher than the current industry standard.

[0132] During data acquisition, all ground-penetrating radar echo signals underwent median filtering and wavelet domain denoising; infrared images were contrast-stretched and distortion-corrected; temperature data was normalized to the interval [0,1]; material viscosity and air pressure data were written in real time by synchronously acquired sensors; and all data points were precisely aligned in time and space based on GPS positioning and vehicle odometer data, ultimately forming a complete aligned multi-source non-destructive testing dataset. The data structure for each test point is as follows:

[0133]

[0134] The aligned dataset is first fed into the physically controlled multi-subband Fourier neural operator model described in this invention. The model divides the frequency domain into three subbands: low frequency (0–8MHz), mid frequency (8–25MHz), and high frequency (25–60MHz). Each subband has an independently set convolution kernel and weight matrix, and the contribution weight of each subband to thickness estimation is dynamically adjusted based on real-time temperature, viscosity, and impedance parameters. Taking measurement point number 5310 as an example, with a normalized temperature of 0.47, viscosity of 1312 mPa·s, and radar peak / baseline impedance ratio of 1.16, the model automatically reduces the gating weight of its high-frequency subband by 12% to suppress spurious segregation reactions caused by abnormal fluctuations.

[0135] In the parameter optimization phase, the Henry's Law gas solubility optimization algorithm was used to initialize 32 candidate parameter vectors, covering spectral depths of 6-12 layers, frequency cutoff thresholds of 5-55MHz, regularization factors of 0.01-0.2, and radar / infrared / temperature signal weight ranges [0.2, 0.7]. Physical environment parameters (average temperature across the entire road section 34.2℃, air pressure 100.1-100.8kPa, average viscosity 1285mPa·s) were directly mapped to solubility-partial pressure modulation coefficients. The optimization process converged in approximately 84 seconds. The optimal parameter combination was finally selected: spectral depth of 10 layers, frequency thresholds of 8 / 25 / 54MHz for low / medium / high subbands, regularization factor of 0.08, and signal weight ratios of 0.56 / 0.31 / 0.13.

[0136] The optimized Fourier neural operator model performs point-by-point inference on a 12,800-point complex spectral domain tensor dataset. The continuous field T(x,y) of the segregation layer thickness is generated synchronously within one minute on the vehicle inspection terminal. The thickness field and point coordinates are automatically fused to generate a thickness heatmap. The main heatmap has a resolution of 3cm × 3cm and rich spatial details. A thickness anomaly threshold range of 30-62mm is set on the heatmap. Areas exceeding this range are identified as suspected segregation areas. The detection results show that the area of ​​suspected segregation sections accounts for approximately 2.9% of the total area, concentrated near the central lane. The largest continuous anomaly area is 18 meters long and 0.6 meters wide.

[0137] To verify the effectiveness of the model inference of this invention, the project team randomly selected 20 core samples from typical suspected areas and normal areas (core numbers and coordinates are shown in Table 1 below):

[0138] Table 1. Measured actual segregation layer thickness (mm)

[0139] Core number x-coordinate (m) Vertical coordinate (m) Model prediction (mm) Actual thickness (mm) Traditional GPR method prediction (mm) C-03 41.22 1.88 57.2 59.0 65.3 C-07 46.81 0.95 35.9 36.4 43.1 C-11 58.36 2.31 49.6 48.2 56.8 C-15 65.14 1.27 38.3 40.6 45.2 C-18 76.54 0.73 31.8 32.0 37.8 C-21 84.90 1.02 44.5 45.1 50.7 C-26 100.00 1.41 60.6 63.3 70.5 C-34 125.41 2.12 39.7 38.9 44.6 C-47 163.27 1.23 54.2 56.1 61.3 C-58 200.73 1.95 34.5 33.7 40.9 C-61 221.46 1.61 50.1 49.3 56.4 C-72 251.94 2.09 41.8 43.6 48.5 C-83 284.27 1.15 62.9 65.0 72.1 C-88 310.16 0.91 59.5 62.2 68.0 C-97 344.72 1.97 37.6 36.8 42.9 C-109 383.24 2.02 49.3 48.6 54.8 C-115 399.60 1.21 53.4 55.8 60.7 C-120 420.31 2.35 32.2 33.1 39.8 C-127 444.62 0.94 47.1 46.7 53.2 C-135 468.59 1.16 38.7 39.2 44.0

[0140] As shown in Table 1 above, based on the statistical data from 20 core drilling points, the root mean square error (RMSE) between the method of this invention and the actual thickness is 2.57 mm, the maximum absolute error is 3.6 mm, and the standard deviation is 2.81 mm. The RMSE of the traditional GPR thickness inversion method is 7.74 mm, the maximum error is 12.6 mm, and the standard deviation is 7.83 mm. The accuracy of this invention in identifying all segregation anomaly zones exceeding the threshold is 95.8%, while that of the traditional method is only 78.3%. The spatial resolution of the segregation thickness field heat map is improved by about 4 times. The overall data processing and output time on site is less than 90 seconds. Compared with the single-point detection of the manual core drilling method (single-point detection takes about 20 minutes, and 100 points take 33 hours) and the post-processing process of the traditional GPR method (7 hours per kilometer), the efficiency is improved by more than 20 times.

[0141] In the field application of Example 1, the generalization ability of the model under different climate and material ratios was compared. Training and validation were conducted for four sets of conditions: autumn (average temperature 17.6℃), hot summer (average temperature 34.2℃), high humidity with rainfall (material viscosity decreased to 1024 mPa·s), and normal conditions (viscosity 1340 mPa·s). The results showed that after rapid fine-tuning with incremental samples, the RMSE fluctuation of the method of this invention did not exceed 0.7 mm under all environments, while the RMSE fluctuation of the traditional GPR algorithm and conventional deep networks was between 4.3 and 8.9 mm.

[0142] This invention not only significantly improves the spatial resolution and accuracy of asphalt pavement segregation layer thickness estimation and greatly shortens the detection cycle, but also realizes full life cycle adaptive closed-loop learning, solving the problem of insufficient accuracy and stability of thickness field modeling under heterogeneous environments, materials and working conditions in existing technologies.

[0143] This invention introduces a physically controlled multi-subband Fourier neural operator structure to achieve dynamic gating and adaptive modulation of the spectrum modeling process based on environmental physical variables. The complex spectral domain tensor dataset is divided into multiple sub-bands according to the frequency dimension, and an independent convolution path and gating mechanism are established for each sub-band to achieve multi-scale response to different physical environments and material properties. This enhances the model's adaptability and robustness to the changing working conditions of asphalt pavement and significantly improves spatial resolution and thickness estimation accuracy.

[0144] This invention proposes a multi-objective model parameter adaptive optimization mechanism based on the Henry's gas solubility optimization algorithm. By utilizing the solubility-partial pressure control principle, environmental temperature, gas pressure, and material viscosity parameters are directly mapped to the global search probability and local exploitation probability in the optimization algorithm. This effectively achieves dynamic self-optimization of high-dimensional hyperparameters such as spectral depth, subband frequency cutoff threshold, regularization factor, and multi-source channel weights. It realizes the physical drive and engineering adaptation of parameter optimization, ensuring that the model can obtain optimal inference accuracy and real-time inference efficiency under different regional, climatic, and construction mix conditions.

[0145] This invention establishes a complete engineering closed-loop process: complex spectral domain data → physically controlled multi-subband Fourier neural operator → dynamic optimization → thickness heatmap and visualization of suspected areas → core drilling incremental closed-loop self-learning. By combining the actual thickness values ​​of the core drilling points with the features of the complex spectral domain tensor dataset as incremental supervision samples and continuously incorporating them into the training system, and combining the re-optimization mechanism of the Henry's gas solubility optimization algorithm, the model can achieve online adaptive updates under the background of continuous changes in road network, season, climate and materials. It maintains high-precision generalization ability for heterogeneous data and complex construction scenarios. The results of actual engineering applications show that it effectively improves the accuracy of thickness anomaly area identification and spatial positioning ability, providing a reliable technical foundation for rapid road inspection and precise reinforcement.

[0146] 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 machine learning-based method for estimating the thickness of asphalt pavement segregation layer in road and bridge construction, characterized in that, Includes the following steps: S1. Synchronously collect asphalt pavement data, construct a multi-source non-destructive testing dataset, and perform preprocessing to obtain an aligned multi-source non-destructive testing dataset; S2. Perform a Fourier transform on the aligned multi-source nondestructive testing dataset to generate a complex spectral domain tensor; S3. Construct a Fourier neural operator model and perform end-to-end training on the complex spectral domain tensor to obtain the initial Fourier neural operator model. S4. Initialize the Henry's gas solubility optimization algorithm population, define the model parameter optimization space and the Fourier neural operator model optimization variables, use the complex spectral tensor as the fitness index, execute the Henry's gas solubility optimization algorithm, and obtain the optimized Fourier neural operator model. S5. Apply the optimized Fourier neural operator model to the complex spectral domain tensor to generate a continuous field of segregation layer thickness, generate a thickness heatmap, and mark the region exceeding the threshold in the thickness heatmap as a suspected segregation region; S6. After the core drilling measurement results are obtained, the measurement results are compared with the suspected segregation areas, and the Henry gas solubility optimization algorithm is triggered to optimize again, update the Fourier neural operator model, and when the updated Fourier neural operator model converges, the latest segregation layer thickness continuous field, the latest thickness heat map and the list of suspected segregation areas are output to complete the estimation of the asphalt pavement segregation layer thickness. S2 includes the following steps: S21. Perform a one-dimensional Fourier transform operation on the ground radar echo signal of each measurement point in the denoised ground radar echo signal set to obtain the ground radar complex spectral domain signal corresponding to each measurement point. S22. Perform a two-dimensional Fourier transform operation on the line scan image of each measurement point in the standardized image set of the final multi-source nondestructive testing dataset to obtain the complex spectral domain tensor of the line scan image corresponding to each measurement point; S23. Perform a one-dimensional Fourier transform operation on the normalized temperature value sequence of each measurement point in the normalized temperature value set in the final multi-source non-destructive testing dataset to obtain the temperature complex spectral domain signal corresponding to each measurement point; S24. The complex spectral domain signals of ground-penetrating radar, the complex spectral domain tensors of line scan images, and the complex spectral domain signals of temperature are integrated with the original environmental air pressure data, material viscosity data, and location information of the measurement points to construct a complex spectral domain tensor dataset. S3 includes the following steps: S31. Using the complex spectral domain tensor dataset as input, call the subband divider to perform interval division on the frequency dimension of the complex spectral domain signal at each measurement point, generating a set of M non-overlapping subbands including low-frequency subband, mid-frequency subband and high-frequency subband. S32. For each sub-band set, construct an independent sub-band convolution path, and initialize the sub-band convolution kernel weight matrix, bias vector and activation function parameters respectively to form a multi-subband parallel convolutional engine; S33. Based on the impedance ratio coefficient, normalized temperature value, ambient air pressure value and material viscosity value of each measurement point, call the physical control gate to generate the corresponding sub-band gate coefficient, and couple the sub-band gate coefficient with the multi-sub-band parallel convolutional unit sub-band by sub-band to obtain the physically controlled gated sub-band convolution output. S34. Perform inverse Fourier transform on the convolutional output of the physically controlled gated subbands to obtain the temporal characteristics of each subband; S35. Call the residual merger to sum the time-domain features of each sub-band according to the gating coefficients, and perform residual connection with the lifted representation of the original complex spectral domain input features to generate a fused feature vector; S36. Construct training samples using fused feature vectors and the true values ​​of the separation layer thickness at measurement points, and iteratively optimize the physical control type multi-subband Fourier neural operator using a joint loss function until the joint loss function converges; S37. Solidify the converged weight parameters and output an initial physical control type multi-subband Fourier neural operator model containing a subband divider, a multi-subband parallel convolutional unit, a physical control gate, and a residual merger. S5 includes the following steps: S51. Load the optimized Fourier neural operator model into the segregation layer thickness estimation task. With the complex spectral domain tensor dataset as input, call the optimized Fourier neural operator model to perform thickness prediction operation on each measurement point in the complex spectral domain tensor dataset to obtain the segregation layer thickness prediction value corresponding to the measurement point. The segregation layer thickness prediction value is used to estimate the actual asphalt segregation layer thickness at the measurement point. S52. Arrange the predicted segregation layer thickness values ​​of all measurement points in order according to the measurement point number to construct a segregation layer thickness prediction sequence; S53. Bind each predicted thickness value in the segregation layer thickness prediction sequence to the spatial location information of the corresponding measurement point, and construct a segregation layer thickness field mapping function. The segregation layer thickness field mapping function is used to describe the predicted thickness value corresponding to each location in two-dimensional space. S54. Visualize the segregation layer thickness field mapping function in a two-dimensional spatial coordinate system to generate a thickness heat map image. The thickness heat map image expresses the predicted value of the segregation layer thickness at each spatial location in the form of color depth. S55. Set a threshold range for judging the thickness of the segregation layer. The threshold range for judging the thickness of the segregation layer includes a lower threshold and an upper threshold. Extract the spatial coordinates corresponding to the positions of all thickness prediction values ​​that are less than the lower threshold or greater than the upper threshold to form a set of coordinates of suspected segregation areas. S56. Graphically mark all location points in the coordinate set of suspected segregation areas to form a suspected segregation area annotation layer; S57. Overlay the thickness heat map image with the labeled layer of the suspected segregation area to generate the final visualization image of the segregation layer thickness.

2. The method for estimating the thickness of asphalt pavement segregation layer in road and bridge construction based on machine learning according to claim 1, characterized in that, S1 includes the following steps: S11. Collect sets of ground-penetrating radar echo signals, line scan image data, environmental parameter data, and location information; S12. Perform wavelet domain filtering and reflection waveform denoising on the set of ground-penetrating radar echo signals to obtain the denoised set of ground-penetrating radar echo signals. S13. Perform image preprocessing on the line scan image dataset, including enhancing the contrast of image colors, correcting geometric distortion of image edges, and standardizing brightness consistency to obtain a standardized image dataset; S14. Perform normalization processing on the ambient temperature data to obtain a set of normalized temperature values; S15. Combine the denoised ground-penetrating radar echo signal set, standardized image set, normalized temperature value set, original ambient air pressure data, material viscosity data, and location information. Using a unified time label as an index, point-by-point registration is performed to construct an aligned multi-source non-destructive testing dataset; S16. Set a unified reference time base in the aligned multi-source non-destructive testing dataset, and perform synchronous correction on all time labels so that the time difference between the timestamps of all sampling points and the reference time base is less than the set maximum allowable time difference threshold. This completes the synchronous correction of multi-source data in the time dimension and yields the final multi-source non-destructive testing dataset.

3. The method for estimating the thickness of asphalt pavement segregation layer in road and bridge construction based on machine learning according to claim 1, characterized in that, S4 includes the following steps: S41. Define the set of parameters to be optimized for the physical control type multi-subband Fourier neural operator model. The set of parameters to be optimized includes the spectral depth, subband frequency cutoff threshold, regularization factor, and channel weight coefficient set of multi-source non-destructive testing signals. The parameters to be optimized are uniformly represented as a parameter vector to be optimized. S42. Initialize the candidate solution population of Henry's gas solubility optimization algorithm. Each candidate solution in the candidate solution population contains a set of parameters to be optimized. Associate each candidate solution in the population with its corresponding set of physical environment parameters. The set of physical environment parameters includes normalized temperature value, ambient air pressure value and material viscosity value. S43. Construct a solubility-partial pressure control mechanism. In each evolution of the Henry's gas solubility optimization algorithm, calculate the solubility coefficient and partial pressure factor of the candidate solution based on the set of physical environment parameters corresponding to each candidate solution. Map the solubility coefficient to the global search probability of the candidate solution and map the partial pressure factor to the local exploitation probability of the candidate solution. S44. The spectral depth, subband frequency cutoff threshold, regularization factor and channel weight parameters contained in each candidate solution are applied to the physical control type multi-subband Fourier neural operator model. The multi-subband Fourier neural operator model is used to predict and infer the thickness field of the asphalt pavement segregation layer on the complex spectral domain tensor dataset. After the inference is completed, the root mean square error index of the thickness estimation residual, the single batch inference time index and the memory consumption index corresponding to the candidate solution are evaluated. The candidate solutions are then weighted and combined according to the set weight coefficients to generate the comprehensive fitness index corresponding to the candidate solutions. S45. Based on the comprehensive fitness index of each candidate solution and the corresponding solubility coefficient and partial pressure factor, perform the evolutionary iteration operation of Henry's gas solubility optimization algorithm. Guide candidate solutions with fitness higher than the threshold to a new search area according to the solubility coefficient to optimize the global search capability. Perform perturbation and local mutation operations on candidate solutions with fitness lower than the threshold according to the partial pressure factor. During each iteration, dynamically adjust the search range and step size in combination with the environmental parameters of the current candidate solution. S46. After each iteration, determine whether the comprehensive fitness index corresponding to the best candidate solution in the current population meets the preset convergence criterion. If the convergence criterion is met, select the candidate solution as the best solution; if the convergence criterion is not met, continue to execute the next round of population evolution until the convergence criterion is met. S47. The spectral depth, subband frequency cutoff threshold, regularization factor, and multi-source signal channel weight configuration obtained from the optimal candidate solution are applied to the physical control type multi-subband Fourier neural operator model to form an optimized Fourier neural operator model.

4. The method for estimating the thickness of asphalt pavement segregation layer in road and bridge construction based on machine learning according to claim 3, characterized in that, The root mean square error index of the thickness estimation residual is used to calculate the root mean square error between the predicted value and the actual value of the segregation layer thickness. The single-batch inference time consumption index is the average time required for the Fourier neural operator model to perform a single round of thickness prediction inference. The aforementioned memory consumption metric is a statistical measure of the amount of video memory used by the Fourier neural operator model during execution.

5. The method for estimating the thickness of asphalt pavement segregation layer in road and bridge construction based on machine learning according to claim 1, characterized in that, S6 includes the following steps: S61. Extract the true value of the segregation layer thickness of each core point in the core measurement results, and extract the spatial location information corresponding to each core point. Perform spatial matching of the spatial location information of all core points with the coordinate set of the suspected segregation area, and filter out the core points that fall within the range of the suspected segregation area to obtain the set of core points located in the suspected segregation area. S62. Combine the true value of the segregation layer thickness of each core point in the set of core points located in the suspected segregation region with the corresponding complex spectral domain input feature of the core point in the complex spectral domain tensor dataset to construct an incremental supervision sample set; S63. Add the incremental supervised sample set to the original training sample set to form a new training sample set; S64. Using the new training sample set as the model training input, reinitialize the candidate solution population of the Henry's gas solubility optimization algorithm, and calculate the comprehensive fitness index of each candidate solution under the new training sample set in each evolution process. S65. Form an updated Fourier neural operator model, predict the thickness of all measurement points in the complex spectral domain tensor dataset, and obtain a new continuous field for the thickness of the segregation layer. S66. Based on the new segregation layer thickness continuous field and the spatial location information of each measurement point, generate the latest thickness thermal image and form the latest list of suspected segregation areas; S67. Output the updated segregation layer thickness continuous field, the latest generated thickness heatmap image, and the updated list of suspected segregation areas as the segregation layer thickness estimation results for asphalt pavement.

Images