SiC modified asphalt interface identification method and system based on deep learning

Abstract

This invention belongs to the field of materials testing technology and discloses a deep learning-based method and system for identifying the interface of SiC modified asphalt. It acquires multi-dimensional interface images through a layered time-series acquisition method using four devices: optical microscope, SEM, EDS-Mapping, and Raman spectrometer. Adaptive algorithms are used for targeted denoising, normalization, and registration, while attention mechanisms and weighted fusion strategies are combined to enhance interface features. Simultaneously, a multi-condition dataset containing a high proportion of minor defects is constructed to achieve deep fusion and accurate identification of multi-modal features, eliminating subjective errors from manual judgment and improving the fine-grained accuracy and detection efficiency of SiC modified asphalt interface identification. The U-Net model is improved for specific scenarios and designed to be lightweight. By optimizing the encoder-decoder structure, introducing a dual-channel attention mechanism and a condition-adaptive branch, and combining a loss function, the feature extraction capability of interface boundaries and minor defects is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of materials testing technology, specifically a method and system for identifying the interface of SiC modified asphalt based on deep learning. Background Technology

[0002] SiC-modified asphalt, with its excellent high-temperature stability, anti-aging properties, and mechanical strength, is widely used in transportation infrastructure such as high-grade highways and bridge pavements. The core of its performance improvement depends on the interfacial bonding state between SiC and the asphalt matrix. The tightness of the interface and the distribution of defects directly determine key properties of the modified asphalt, such as fatigue resistance and water damage resistance. Current technologies suffer from bottlenecks such as ineffective fusion of multimodal features, insufficient fine-grained recognition accuracy, and the lack of a complete process optimization closed loop. Existing deep learning models are mostly general-purpose and not adapted to the interface scenarios of SiC-modified asphalt, failing to meet the needs of intelligent engineering inspection and process optimization.

[0003] Currently, interface recognition mainly relies on traditional methods such as optical microscopy, SEM, and FTIR. These methods depend on manual judgment, are inefficient, and highly subjective. They cannot achieve automated quantification of interface features and are difficult to intuitively present microscopic defects. Existing deep learning models such as CNN and the original U-Net mostly use single-modal image input, which cannot comprehensively capture multi-dimensional features of the interface. They have insufficient ability to identify minute defects, are prone to missed or false identifications, and cannot accurately quantify key interface parameters. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for identifying the interface of SiC modified asphalt based on deep learning, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for identifying the interface of SiC-modified asphalt, comprising the following specific steps:

[0006] Preferably, in the sample preparation stage, SiC particles are modified with silane coupling agent KH-550 according to the preset SiC doping amount and particle size. The SiC particles are mixed with KH-550 ethanol solution with a mass fraction of 1%-3%, stirred at 60-80℃ for 2-3 hours, centrifuged, and then dried in an oven at 105℃ for 24 hours to remove surface impurities and oxide layer.

[0007] Modified SiC microparticles were mixed with base asphalt and modifier, and SiC modified asphalt matrices with different ratios were prepared by shearing at 160-180℃ for 30-60 minutes using a high-speed shearing machine. The prepared modified asphalt matrices were subjected to accelerated aging treatment simulating actual engineering service conditions to obtain SiC modified asphalt samples under multiple working conditions.

[0008] Using cryo-ultrathin sectioning technology combined with vacuum coating treatment, an ultra-thin section sample with a thickness of 50 - 100 nm is prepared to retain the original microscopic morphology of the interface. Finally, the section sample is pre-treated by cleaning the residual impurities on the surface with absolute ethanol and removing the surface oxide layer by plasma treatment.

[0009] Preferably, in the image processing stage, four devices including an optical microscope, a scanning electron microscope, energy-dispersive X-ray spectroscopy surface distribution (EDS-Mapping), and a Raman spectrometer are linked together. First, the optical microscope is used to collect macroscopic interface morphology images with a resolution of 100× - 500× to locate the interface area and macroscopic defects and define the acquisition ranges of SEM, EDS-Mapping, and Raman spectroscopy; then the SEM is used to collect microscopic interface images with a resolution of 1000× - 5000× to capture tiny interface defects and the dispersion state of SiC particles; subsequently, the EDS-Mapping is used to collect element surface distribution images to capture the spatial distribution of Si and C elements and clarify the interface boundary between SiC particles and the asphalt matrix; finally, the Raman spectrometer is used to collect Raman spectroscopy images of the interface area to capture the chemical bonding characteristics between SiC and asphalt molecules.

[0010] An adaptive image preprocessing algorithm is introduced. Gaussian filtering is used to remove Gaussian noise in the optical microscope images, median filtering is used to remove salt-and-pepper noise in the SEM images, wavelet threshold denoising algorithm is used to remove electronic noise in the EDS-Mapping images, and adaptive threshold denoising algorithm is used to remove fluorescence noise in the Raman spectroscopy images. The preprocessed images are normalized, and the image gray values are normalized to the range of [0, 1] to eliminate the gray differences caused by different devices and different acquisition parameters. An image registration algorithm based on feature point matching and element distribution calibration is used to align the multimodal images.

[0011] An ECA channel attention mechanism and a feature interaction module are introduced. First, the features of each modality are grouped. The binary grouping attention mechanism is an attention mechanism that groups multimodal features by type and then strengthens key features and suppresses irrelevant features through binary weights. Then, an adaptive weighted fusion strategy is adopted to automatically adjust the fusion weights of the features of each modality according to different aging conditions, adaptively strengthen the features of the interface area and tiny defects, and improve the recognition rate of the interface boundary and tiny defects.

[0012] Preferably, in the dataset construction stage, a SiC-modified asphalt interface image dataset with multiple working conditions, multiple ratios, multiple defect types, and multiple modification parameters is constructed. Multimodal images under different SiC dosages, different particle sizes, different surface modification degrees, different aging conditions, and different shear parameters are collected, including 5 types of interface states such as tight interface combination, interface voids, interface cracks, interface separation caused by SiC agglomeration, and poor interface chemical bonding. The proportion of tiny defect samples is not less than 45%.

[0013] A labeling team composed of three or more professionals in materials engineering, image processing, and road engineering collaborates on labeling using multimodal features. Specifically, the team determines interface boundaries through elemental distribution analysis using EDS-Mapping, identifies defect types through microscopic morphology analysis using SEM, and determines chemical bonding states through Raman spectroscopy. The team labels interface boundaries, defect types, defect locations, SiC particle locations, and chemical bonding strength levels, and generates standard labeling files. A semi-supervised learning algorithm is used to train an initial model based on labeled samples. Unlabeled samples are automatically labeled and then manually corrected.

[0014] A variety of data augmentation methods, such as random flipping, rotation, brightness and contrast adjustment, noise addition, image stitching, and scaling, were used to expand the sample size of the dataset. The dataset was divided into training, validation, and test sets in a ratio of 7:2:1. The training set was used for model parameter training, the validation set was used for model hyperparameter optimization, and the test set was used for model performance verification. The multimodal features in the dataset were deeply fused and labeled. The microscopic defect features of SEM images, the elemental distribution features of EDS-Mapping images, the macroscopic features of optical microscope images, and the chemical binding features of Raman spectroscopy images were associated and labeled to construct a multi-scale, multi-dimensional feature labeling system.

[0015] Preferably, the model construction stage is based on the U-Net model, and an improved U-Net model is designed in combination with the specific scenario of SiC modified asphalt interface recognition. The encoder is composed of a VGG16 pre-trained network and a multi-scale dilated convolution module. The VGG16 pre-trained weights are used to improve the interface feature extraction capability and shorten the model training cycle. The receptive field is expanded by the multi-scale dilated convolution module to capture interface defects of different sizes. At the same time, a dual-channel attention mechanism composed of channel attention and spatial attention is added. Spatial attention focuses on the interface edge and small defect area, while channel attention focuses on the key feature channel after multimodal fusion, thus doubly strengthening the feature extraction of interface boundary and small defects.

[0016] The decoder incorporates a cross-scale feature fusion module and an ASPPDeformable module. The ASPPDeformable module is a hollow spatial pyramid pooling module with deformable convolution, used to capture irregular features of the interface boundary. The cross-scale feature fusion module fuses the multi-scale features extracted by the encoder. The ASPPDeformable module captures irregular features of the interface boundary through deformable convolution, optimizing the segmentation accuracy of the interface boundary. Upsampling is performed using a combination of transposed convolution and interpolation. Branches are set for four core aging conditions: high and low temperature cycling, ultraviolet aging, water damage, and freeze-thaw cycles. Feature extraction parameters are preset for different aging conditions, and the model can automatically match the branch corresponding to the working condition for feature extraction. The loss function is composed of FocalLoss, DiceLoss, and ContourLoss. FocalLoss is used to improve the recognition accuracy of minor defects and poor chemical bonding defects. DiceLoss is used to optimize the overlap of interface segmentation. ContourLoss is used to constrain the continuity and accuracy of the interface boundary, further improving the interface segmentation accuracy. The parameters are adjusted by referring to boundary loss optimization technology and adapting to the SiC modified asphalt interface scenario.

[0017] The model is lightweight by using a quantization-aware training scheme, and the model is quantized using INT8 based on the Paddle-Lite quantization toolchain.

[0018] Preferably, during the model training phase, the constructed multi-scale interface feature dataset is input into the improved lightweight U-Net model, initial hyperparameters are set, and the AdamW optimizer is used for model training. During the training process, the loss value, recognition accuracy, intersection-over-union ratio, boundary intersection-over-union ratio, and other indicators of the training set and validation set are monitored in real time. An adaptive learning rate adjustment algorithm is introduced, and cosine annealing and warm-up strategies are adopted. The learning rate is automatically adjusted according to the changes in validation set indicators during the model training process. The K-fold cross-validation method is used to validate the model multiple times.

[0019] An interface model of SiC and asphalt molecules was constructed using MaterialsStudio software. The interface bonding energy and diffusion coefficient of SiC and asphalt under different working conditions and different degrees of SiC surface modification were simulated using COMPASSII force field. Combined with chemical bonding characteristic data collected by Raman spectroscopy, a correlation model between the microscopic parameters of the interface and the macroscopic identification results was constructed. The simulated interface characteristic parameters and Raman chemical bonding data were input into the model as optimization constraints.

[0020] To validate the model's performance, the test set was input into the optimized model to verify the model's recognition accuracy, mIoU, BIoU, small defect omission rate, and chemically poor bonding recognition accuracy, as well as the inference speed of the quantized model.

[0021] Preferably, in the identification and positioning stage, the SiC modified asphalt sample to be tested acquires multimodal interface images according to the image processing stage method, performs preprocessing and registration, and then inputs them into the optimized improved lightweight U-Net model. The model automatically completes the segmentation of interface boundaries, the positioning of SiC particles, the identification and classification of interface defects through multimodal feature collaborative identification and working condition adaptive branch switching, and outputs interface identification result images. The interface area, SiC particles, and different types of defects are marked with different colors to present the interface state, and the chemical bonding strength level is marked.

[0022] Based on the segmentation results output by the model, the system automatically calculates parameters such as the center coordinates, boundary contour, defect area, defect perimeter, and defect depth of each defect to locate the defect and generate a defect location report. For suspected defect areas that appear during the identification process, the system automatically calls on the detailed features of multimodal images for secondary verification to eliminate misidentified areas, optimizes the batch identification function, supports simultaneous detection of multiple groups of samples, and automatically generates interface identification results, defect location list, and chemical bonding status evaluation for each group of samples.

[0023] Preferably, the quantitative optimization stage automatically quantifies the key characteristic parameters of the SiC modified asphalt interface based on the identification results output by the model, including interface thickness, interface bonding rate, defect density, defect size and depth distribution, SiC particle dispersion uniformity, and interface chemical bonding strength grade. It also establishes a correlation model between these parameters and the high-temperature stability, fatigue resistance, and water damage resistance of the modified asphalt. Combined with dynamic shear rheological test, microhardness test, and tensile test data, it clarifies the correlation between interface characteristic parameters and the performance of modified asphalt, introduces quantitative evaluation indicators, divides the interface quality into four levels: excellent, good, qualified, and unqualified, and sets a quality evaluation standard that includes the chemical bonding strength grade.

[0024] The NSGA-Ⅲ multi-objective collaborative optimization algorithm is adopted to automatically generate optimization suggestions for SiC modification process with the goals of improving interface quality, reducing energy consumption, and controlling costs. It adjusts SiC doping, particle size, surface modification and shear parameters, and optimizes preparation temperature, aging protection and other conditions. It constructs a closed-loop system covering the entire life cycle of interface recognition, feature quantification, process optimization, on-site verification and model iteration. The optimized process parameters are applied to preparation, and the effect is re-tested and verified. If the effect does not reach the excellent level, the data is fed back to adjust the model and process. The generated process optimization parameters can be fed back to the sample preparation stage.

[0025] This invention also provides a deep learning-based interface recognition system for SiC-modified asphalt, which, based on the above method, includes:

[0026] The sample preparation module is used to complete the standardized preparation of SiC modified asphalt samples, realize SiC microparticle surface modification, asphalt matrix preparation, multi-condition accelerated aging and slice pretreatment, and use cryogenic ultrathin sectioning technology to preserve the original micro morphology of the interface.

[0027] The multimodal image acquisition and processing module integrates four devices: optical microscope, SEM, EDS-Mapping and Raman spectrometer. It adopts a hierarchical time-series collaborative acquisition mode to acquire multi-dimensional images, processes images through adaptive denoising, normalization and registration algorithms, introduces an attention mechanism to enhance interface features, and outputs fused images.

[0028] The dataset management module is used to build, label, and optimize interface image datasets with multiple working conditions and ratios. It adopts a combination of manual and semi-supervised automatic labeling, and constructs a multi-scale feature labeling system through data augmentation, SMOTE-ENN algorithm, and dataset partitioning.

[0029] The model building and training module designs a lightweight improved recognition model based on the U-Net model, optimizes the encoder and decoder structure and loss function, adds a dual-channel attention mechanism and a working condition adaptive branch, and adopts quantized perception training to achieve model lightweighting to adapt to portable devices; through adaptive learning rate adjustment, K-fold cross-validation and molecular dynamics simulation-assisted optimization, the model training and performance verification are completed.

[0030] The identification and positioning module receives pre-processed multimodal images, completes interface segmentation, SiC particle positioning and defect classification through an optimized model, adopts a secondary verification algorithm to improve accuracy, supports batch sample testing, and automatically generates identification results, defect positioning list and chemical bonding state evaluation.

[0031] The quantitative optimization module automatically quantifies key feature parameters of the interface, establishes a correlation model with the performance of modified asphalt, sets quality evaluation standards, generates process optimization suggestions through a multi-objective collaborative optimization algorithm, constructs a closed-loop system for the entire life cycle, and feeds back into the sample preparation module.

[0032] The beneficial effects of this invention are as follows:

[0033] 1. This invention acquires multi-dimensional interface images through a layered time-series acquisition method involving four devices: optical microscope, SEM, EDS-Mapping, and Raman spectrometer. Adaptive algorithms are used for targeted denoising, normalization, and registration, while attention mechanisms and weighted fusion strategies are employed to enhance interface features. Simultaneously, a multi-condition dataset containing a high proportion of minute defects is constructed to achieve deep fusion and accurate identification of multi-modal features. This replaces manual automating of interface feature quantification, eliminates subjective errors in manual judgment, and improves the fine-grained accuracy and detection efficiency of SiC modified asphalt interface identification.

[0034] 2. This invention improves and lightweights the U-Net model by optimizing its encoder-decoder structure, introducing a dual-channel attention mechanism and a working condition adaptive branch, and combining it with a combined loss function to enhance the feature extraction capability of interface boundaries and minor defects. It also achieves model lightweighting by combining quantized perception training, making it compatible with portable detection devices. At the same time, it improves model stability by using molecular dynamics simulation to assist optimization and K-fold cross-validation, while still ensuring inference speed and recognition accuracy after quantization.

[0035] 3. This invention automatically quantifies key interface feature parameters based on model recognition results, establishes a correlation model between parameters and performance by combining modified asphalt performance test data, and sets quantitative quality evaluation standards including chemical bond strength levels. It generates process optimization suggestions through the NSGA-Ⅲ multi-objective collaborative optimization algorithm, feeding the optimized parameters back into the sample preparation stage. Test results that do not meet the standards can be fed back to adjust the model and preparation process, achieving a cyclical improvement of interface recognition, feature quantification, process optimization, and model iteration. This balances interface quality improvement, energy consumption reduction, and cost control, thereby enhancing the engineering service performance of SiC modified asphalt. Attached Figure Description

[0036] Figure 1 This is an overall flowchart of the method of the present invention;

[0037] Figure 2 This is a flowchart of the multimodal image acquisition and refinement process of the present invention;

[0038] Figure 3 This is a flowchart of the interface recognition, positioning, and result output process of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] like Figures 1 to 3 As shown, this embodiment of the invention provides a deep learning-based method for identifying the interface of SiC-modified asphalt, including the following specific steps:

[0041] The sample preparation stage is carried out according to a preset SiC doping amount (0.5%-5%, covering the commonly used range in engineering) and particle size (50nm-500nm). For high-grade highway paving scenarios, the preferred SiC doping amount is 1%-3% and the particle size is 100-300nm. For bridge paving scenarios, the preferred SiC doping amount is 2%-5% and the particle size is 200-500nm, to meet the interface bonding performance requirements of different engineering scenarios.

[0042] SiC microparticles were modified using silane coupling agent KH-550. The SiC microparticles were mixed with a KH-550 ethanol solution with a mass fraction of 1%-3% and stirred at 60-80℃ for 2-3 hours. After centrifugation, the mixture was dried in an oven at 105℃ for 24 hours to remove surface impurities and oxide layers, thereby improving the interfacial bonding ability between SiC microparticles and asphalt matrix and making the interfacial features easier to capture by multimodal images.

[0043] Modified SiC microparticles were mixed with base asphalt and modifiers (such as SBS), and SiC modified asphalt matrices with different ratios were prepared by shearing with a high-speed shearing machine at a speed of 3000-5000 r / min and 160-180℃ for 30-60 min. The prepared modified asphalt matrices were subjected to accelerated aging treatment simulating actual engineering service conditions such as high and low temperature cycling, ultraviolet aging, water damage, and freeze-thaw cycles to obtain SiC modified asphalt samples under multiple working conditions, ensuring that the samples cover the interface states that may occur in the project.

[0044] The high-speed shearing machine adopts a coaxial double shear head structure with a shear head gap of 0.5-1mm. Nitrogen gas is continuously circulated during the shearing process to prevent high-temperature oxidation of asphalt.

[0045] The cryo-ultrathroplasty technique (-196℃ liquid nitrogen freezing to avoid interfacial thermal damage) is combined with vacuum coating. The vacuum coating uses a gold film coating method, and the coating thickness is controlled at 5-10nm. The slicing advance rate of the cryo-ultrathroplasty equipment is 0.5-1nm / step to ensure the integrity of the ultrathin slices.

[0046] Ultrathin slices with a thickness of 50-100 nm were prepared to preserve the original microstructure of the interface. Finally, the slices were pretreated by cleaning the surface with anhydrous ethanol to remove residual impurities and removing the surface oxide layer by plasma treatment.

[0047] The plasma treatment uses argon plasma with a processing power of 100-200W, a processing time of 5-10min, and a processing pressure of 0.1-0.3Pa.

[0048] The image processing stage employs a four-pronged approach: optical microscope, scanning electron microscope, energy-dispersive X-ray spectroscopy, and Raman spectrometer. First, the optical microscope acquires macroscopic interface morphology images with a resolution of 100×-500× to locate interface regions and macroscopic defects, defining the acquisition range for SEM, EDS-Mapping, and Raman spectroscopy. Next, the SEM acquires microscopic interface images with a resolution of 1000×-5000× to capture minute interface defects and the dispersion state of SiC particles. Subsequently, EDS-Mapping acquires elemental surface distribution images to capture the spatial distribution of Si and C elements and clarify the interface boundary between SiC particles and the asphalt matrix. Finally, the Raman spectrometer acquires Raman spectral images of the interface region to capture the chemical bonding characteristics between SiC and asphalt molecules, such as chemical bond type and bonding strength.

[0049] An adaptive image preprocessing algorithm is introduced, which uses Gaussian filtering to remove Gaussian noise from optical microscope images, median filtering to remove salt-and-pepper noise from SEM images, wavelet threshold denoising algorithm to remove electronic noise from EDS-Mapping images, and adaptive threshold denoising algorithm to remove fluorescence noise from Raman spectroscopy images. The preprocessed images are then normalized to the grayscale value range of [0, 1], thereby eliminating grayscale differences caused by different devices and acquisition parameters. An image registration algorithm based on feature point matching and element distribution calibration is used to align multimodal images to ensure that multimodal features at the same interface position can be accurately fused.

[0050] Feature point matching was performed using the SIFT algorithm, with a matching threshold of 0.7. Element distribution calibration was performed with the characteristic peaks of Si and C elements in the EDS energy spectrum as a reference. The element distribution positions of the EDS-Mapping image were aligned with the reference, and the calibration error was controlled within 1 μm.

[0051] The threshold for adaptive threshold denoising is dynamically determined based on the mean and variance of the gray level of the Raman spectrum image. The threshold is T = μ + 2σ, where μ is the mean gray level of the image and σ is the variance of the gray level of the image. Pixels with gray level values ​​higher than T are identified as fluorescence noise and are filtered.

[0052] An ECA channel attention mechanism and feature interaction module are introduced. First, the features of each modality are grouped, including macroscopic morphology group, microscopic defect group, elemental distribution group, and chemical bonding group. The convolution kernel size of the ECA channel attention mechanism is set to 3×3, and the number of feature groups is 4, corresponding to macroscopic morphology group, microscopic defect group, elemental distribution group, and chemical bonding group, respectively. The channel attention weight of each feature group is calculated by adaptive pooling and one-dimensional convolution.

[0053] By using a binarization grouping attention mechanism to strengthen the key features of each group and suppress irrelevant interference, an adaptive weighted fusion strategy is adopted to automatically adjust the fusion weight of each modality feature according to different aging conditions. For example, the weight of SEM micro-defect features is increased under high and low temperature cyclic conditions, and the weight of EDS elemental distribution and Raman chemical binding features is increased under water damage conditions. This adaptively strengthens the interface region features and micro-defect features, thereby improving the recognition of interface boundaries and micro-defects.

[0054] The weights of the adaptive weighted fusion are determined by calculating the information entropy of each modal feature. The higher the information entropy, the greater the feature weight. The basic weight of SEM microscopic defect features under high and low temperature cyclic conditions is set to 0.35. The basic weights of EDS elemental distribution and Raman chemical binding features under water damage conditions are each set to 0.3. The weights of the remaining features are adaptively allocated according to the information entropy.

[0055] The dataset construction phase involves building a SiC-modified asphalt interface image dataset with multiple working conditions, multiple mix ratios, multiple defect types, and multiple modification parameters. It collects multimodal images under different SiC content, different particle sizes, different surface modification degrees, different aging conditions, and different shear parameters, with a total sample size of no less than 12,000 images. The dataset includes five types of interface states: tight interface bonding, interface voids, interface cracks, interface separation caused by SiC agglomeration, and poor interface chemical bonding. The proportion of micro-defect samples is no less than 45%. Micro-defect samples refer to samples with defect sizes of less than 5 μm, while samples with defects of 5-50 μm are considered conventional defect samples.

[0056] Of the 12,000 samples, 30% were tightly bonded at the interface, 15% were separated due to interface voids, interface cracks, and SiC agglomeration, and 25% were poorly chemically bonded at the interface. The sample size for each aging process was evenly distributed to ensure the model's ability to identify various interface states.

[0057] A labeling team composed of three or more professionals in materials engineering, image processing, and road engineering collaborates on labeling using multimodal features. Specifically, it determines interface boundaries through elemental distribution in EDS-Mapping, identifies defect types through microscopic morphology in SEM, and determines chemical bonding states through Raman spectroscopy. The team labels interface boundaries, defect types, defect locations, SiC particle locations, and chemical bonding strength levels, generating standard labeling files in VOC format. A semi-supervised learning algorithm is used to train an initial model based on labeled samples. Unlabeled samples are automatically labeled and then manually corrected, thereby reducing manual labeling costs and improving labeling efficiency.

[0058] The core fields of the VOC format annotation file include image name, target type (including interface, SiC particles, and defects), defect type, target coordinates, chemical bonding strength level, annotator, and annotation time. The field information corresponds one-to-one with the multi-scale feature labeling system.

[0059] Diverse data augmentation methods such as random flipping, rotation, brightness and contrast adjustment, noise addition, image stitching, and scaling are used to expand the dataset sample size and improve the model's generalization ability. The SMOTE-ENN algorithm is used to solve the model training imbalance problem caused by the small number of small defect samples. The nearest neighbor number k of the SMOTE-ENN algorithm is set to 5, the oversampling sample expansion ratio is 1:2, and the noise samples generated after oversampling are removed by the ENN algorithm with a removal threshold of 0.6.

[0060] The specific operation parameters for data augmentation are as follows: random rotation angle range 0-180°, random flipping is horizontal / vertical bidirectional flipping, brightness adjustment range ±0.2, contrast adjustment range ±0.2, scale scaling range 0.8-1.2 times, image stitching is 2×2 or 3×3 small image stitching, and noise addition intensity is 0.01-0.03.

[0061] The dataset is divided into training, validation, and test sets in a ratio of 7:2:1. The training set is used for model parameter training, the validation set is used for model hyperparameter optimization, and the test set is used for model performance verification. The multimodal features in the dataset are deeply fused and labeled. The microscopic defect features of SEM images, the elemental distribution features of EDS-Mapping images, the macroscopic features of optical microscope images, and the chemical binding features of Raman spectroscopy images are associated and labeled to construct a multi-scale, multi-dimensional feature labeling system.

[0062] The model construction phase is based on the U-Net model. In combination with the specific scenario of SiC modified asphalt interface recognition (multiple working conditions, many small defects, blurred interface edges, and need for on-site engineering deployment), an improved U-Net model is designed. The encoder is composed of a VGG16 pre-trained network and a multi-scale dilated convolution module. The VGG16 pre-trained weights are used to improve the interface feature extraction capability and shorten the model training cycle. The receptive field is expanded by the multi-scale dilated convolution module, which can capture interface defects of different sizes, especially small defects smaller than 5μm, without adding parameters.

[0063] The multi-scale dilated convolution module has a dilation rate of 1, 3, and 5, respectively. It adopts a parallel convolution structure with 64 kernels in each module. The output features are spliced ​​and then fused, which can effectively capture the feature information of micro-defects below 5μm and conventional defects of 5-50μm.

[0064] Simultaneously, a dual-channel attention mechanism consisting of channel attention and spatial attention is incorporated. Spatial attention focuses on interface edges and small defect areas, while channel attention focuses on key feature channels after multimodal fusion. This dual approach enhances feature extraction of interface boundaries and small defects and suppresses interference from irrelevant features.

[0065] The decoder introduces a cross-scale feature fusion module and an ASPP Deformable module. The cross-scale feature fusion module fuses the multi-scale features extracted by the encoder, while the ASPP Deformable module captures irregular features of the interface boundary through deformable convolution, optimizes the segmentation accuracy of the interface boundary, and avoids boundary misalignment and breakage.

[0066] The ASPPDeformable module has a deformable convolution kernel size of 3×3, an offset range of -1 to 1, and 128 kernels. It uses four parallel deformable convolution branches with different dilation rates of 1, 6, 12, and 18.

[0067] Upsampling is performed by combining transposed convolution and interpolation to reduce feature loss during the upsampling process and improve the accuracy of interface boundary recognition. An adaptive branch is set up, and feature extraction parameters are preset for different aging conditions such as high and low temperature cycling and water damage. The model can automatically match the branch corresponding to the working condition for feature extraction, thereby improving the recognition stability under multiple working conditions. There are 4 adaptive branches in total, which correspond to the 4 core aging engineering conditions of high and low temperature cycling, ultraviolet aging, water damage and freeze-thaw cycle. The number of feature extraction convolution kernels in each branch is set to 64, 64, 96 and 96 respectively, and the pooling step size is set to 2×2.

[0068] The loss function is composed of FocalLoss, DiceLoss and ContourLoss. The total loss value of the combined loss function is Loss = 0.2×FocalLoss + 0.5×DiceLoss + 0.3×ContourLoss. The focusing parameter γ of FocalLoss is set to 2 and the balancing parameter α is set to 0.75 to adapt to the identification scenario of imbalanced samples with small defects.

[0069] FocalLoss is used to solve the sample imbalance problem and improve the identification accuracy of minor defects and poor chemical bonding defects. DiceLoss is used to optimize the overlap of interface segmentation and reduce the misidentification of interface boundaries. ContourLoss is used to constrain the continuity and accuracy of interface boundaries and further improve the interface segmentation accuracy. The parameters are adjusted by referring to the boundary loss optimization technology and adapting to the SiC modified asphalt interface scenario.

[0070] The lightweight model adopts a quantization-aware training scheme and performs INT8 quantization on the model based on the Paddle-Lite quantization toolchain, which can be applied to portable detection devices.

[0071] The portable testing equipment compatible with this system has the following hardware configuration: a quad-core or higher CPU, ≥8GB of RAM, ≥128GB of storage, a mobile dedicated graphics card that supports GPU acceleration, and multi-device linkage using the TCP / IP communication protocol. Data collected by optical microscopes, SEM, EDS-Mapping, and Raman spectrometers can be transmitted to portable devices in real time at a transmission rate of ≥100Mbps.

[0072] Paddle-Lite quantization process selects 10% of the dataset as the quantization calibration set, which includes various interface states and defect types. After quantization, the model is fine-tuned and trained to optimize it. The fine-tuning learning rate is set to 0.0001, and the training epochs are 20, to ensure the recognition accuracy and inference speed of the quantized model.

[0073] In the model training phase, the constructed multi-scale interface feature dataset is input into the improved lightweight U-Net model. Initial hyperparameters are set with a learning rate of 0.001, a batch size of 16, and 100 training epochs. The AdamW optimizer is used for model training. During training, the loss value, recognition accuracy, intersection-over-union ratio (IoU), and boundary IoU of the training and validation sets are monitored in real time. An adaptive learning rate adjustment algorithm is introduced, using cosine annealing and warm-up strategies to automatically adjust the learning rate based on changes in validation set metrics during model training. This avoids the model getting stuck in local optima and accelerates model convergence. K-fold cross-validation (K=5) is used to validate the model multiple times to reduce model performance deviation caused by dataset partitioning and ensure model stability and recognition accuracy.

[0074] K-fold cross-validation uses a stratified random partitioning method. The dataset is evenly divided into 5 parts according to the sample proportion of each interface state and defect type. One part is selected as the validation set each time, and the remaining 4 parts are used as the training set. After 5 training and validation cycles, the average value of the model performance index is taken as the final result.

[0075] An interface model of SiC and asphalt molecules was constructed using MaterialsStudio software. The COMPASSII force field was used to simulate the interface bonding energy, diffusion coefficient, and other parameters of SiC and asphalt under different working conditions and different degrees of SiC surface modification. Combined with chemical bonding characteristic data collected by Raman spectroscopy, a correlation model between the microscopic parameters of the interface and the macroscopic identification results was constructed. The simulated interface characteristic parameters and Raman chemical bonding data were input into the model as optimization constraints to improve the reliability of the model in identifying the interface bonding state (especially the poor chemical bonding state).

[0076] The interface model built with MaterialsStudio software has 500-1000 atoms, a simulation temperature of 25-80℃, a simulation duration of 100ps, a time step of 1fs, and uses the Ewald summation method for electrostatic interactions in the COMPASSII force field and the interatomic cutoff method for van der Waals interactions with a cutoff radius of 12.5Å.

[0077] To validate the model's performance, the test set was input into the optimized model to verify the model's recognition accuracy, mIoU, BIoU, small defect omission rate, and chemical bonding failure recognition accuracy. At the same time, the inference speed of the quantized model was verified to ensure that it is suitable for batch sample testing needs.

[0078] The core performance indicators for model verification are: overall recognition accuracy ≥ 98%, mIoU ≥ 95%, BIoU ≥ 94%, minor defect omission rate ≤ 3%, and chemical bonding failure recognition accuracy ≥ 96%. The inference speed of the quantized model for multimodal fusion images is ≥ 20 frames / second, which meets the batch inspection requirements in engineering sites. The overall recognition includes interface segmentation, SiC particle localization, and defect recognition.

[0079] In the identification and positioning stage, the SiC modified asphalt sample to be tested is preprocessed and registered according to the method of the image processing stage, and then input into the optimized improved lightweight U-Net model. The model automatically completes the segmentation of the interface boundary, the positioning of SiC particles, and the identification and classification of interface defects (voids, cracks, interface separation, poor chemical bonding) through multimodal feature collaborative identification and working condition adaptive branch switching. The interface identification result image is output. The interface area, SiC particles, and different types of defects are marked with different colors to present the interface state, and the chemical bonding strength level is marked.

[0080] The specific color scheme for the interface is as follows: the interface area is light blue, SiC particles are red, interface voids are yellow, interface cracks are black, interface separation caused by SiC agglomeration is purple, and poor chemical bonding at the interface is green. The chemical bonding strength level is marked with numbers 1-5 next to the corresponding area, with the larger the number, the higher the bonding strength.

[0081] Based on the segmentation results output by the model, the center coordinates, boundary contour, defect area, defect perimeter, defect depth and other parameters of each defect are automatically calculated to realize the location of the defect and generate a defect location report. The defect depth is calculated by combining the gray-level gradient of the SEM image with the calibration sample. First, the correlation curve between gray-level gradient and defect depth is established through the calibration sample, and then the defect depth of the detection sample is quantitatively calculated.

[0082] For suspected defect areas appearing during the identification process, the system automatically calls upon detailed features such as multimodal image SEM microstructure, EDS elemental distribution, and Raman chemical bonding characteristics for secondary verification, eliminating misidentified areas, improving the accuracy of identification results, optimizing batch identification function, supporting simultaneous detection of multiple groups of samples, and applicable to on-site testing needs of portable devices. It automatically generates interface identification results, defect location list, and chemical bonding status evaluation for each group of samples, improving detection efficiency.

[0083] The criteria for identifying suspected defect areas are: abrupt changes in grayscale values ​​at interface boundaries, abnormal distribution of SiC particles, and areas with ambiguous markings of chemical bonding strength levels. Secondary verification is performed by comparing the consistency of SEM microstructure, EDS elemental distribution, and Raman chemical bonding characteristics. If all three characteristics meet the defect characteristics, it is determined to be a real defect; if any characteristic does not meet the criteria, it is determined to be a misidentification and is removed.

[0084] The quantitative optimization stage automatically quantifies key characteristic parameters of the SiC modified asphalt interface based on the identification results output by the model. These parameters include interface thickness, interface bonding rate, defect density, defect size and depth distribution, uniformity of SiC particle dispersion, and interface chemical bonding strength grade. A correlation model is established between these parameters and the high-temperature stability, fatigue resistance, and water damage resistance of the modified asphalt. Combined with dynamic shear rheology test (DSR), microhardness test, and tensile test data, the correlation between interface characteristic parameters and the performance of modified asphalt is clarified. Quantitative evaluation indicators are introduced, and the interface quality is divided into four levels: excellent, good, qualified, and unqualified. A quality evaluation standard including the chemical bonding strength grade is set.

[0085] The quantitative calculation methods for each characteristic parameter are as follows: the interface thickness is the actual size converted from the average pixel width of the interface region; the interface bonding rate = (area of ​​tightly bonded interface / total interface area) × 100%; the total interface area is the total interface area of ​​SiC particles in contact with the asphalt matrix, excluding the area of ​​the SiC particles themselves; the defect density = number of defects / interface detection area; the uniformity of SiC particle dispersion is the standard deviation of the spatial distribution of SiC particles, and the smaller the standard deviation, the higher the uniformity; the chemical bonding strength level is divided into 1-5 levels according to the intensity ratio of the characteristic peaks of the Raman spectrum.

[0086] The threshold for quantifying interface quality level is:

[0087] Excellent: Interface bonding rate ≥98%, defect density ≤0.5 defects / mm², chemical bonding strength grade ≥4;

[0088] Good: 95% ≤ interfacial bonding rate < 98%, 0.5 defects / mm² < defect density ≤ 1 defect / mm², Grade 3 ≤ chemical bond strength grade < Grade 4;

[0089] Acceptable: 90% ≤ interface bonding rate < 95%, 1 defect / mm² < defect density ≤ 2 defects / mm², Grade 2 ≤ chemical bond strength grade < Grade 3;

[0090] Unacceptable: Interface bonding rate <90%, defect density >2 defects / mm², chemical bonding strength grade <2.

[0091] Employing the NSGA-Ⅲ multi-objective collaborative optimization algorithm, with the goals of improving interface quality, reducing energy consumption, and controlling costs, the algorithm automatically generates optimization suggestions for SiC modification processes. It adjusts SiC dosage, particle size, surface modification, and shear parameters, and optimizes preparation temperature, aging protection, and other conditions to achieve precise process optimization and energy saving. A closed-loop system covering the entire lifecycle of interface recognition, feature quantification, process optimization, on-site verification, and model iteration is constructed. The optimized process parameters are applied to preparation, and the effect is re-tested and verified. If the result is not excellent, the data is fed back to adjust the model and process to continuously improve the performance of modified asphalt and reduce energy consumption and maintenance costs. The generated process optimization parameters can also be fed back to the sample preparation stage.

[0092] The core parameters of the NSGA-Ⅲ multi-objective collaborative optimization algorithm are: population size set to 100, number of iterations set to 200, crossover probability set to 0.9, mutation probability set to 0.01, simulated binary crossover and polynomial mutation method, with interface quality level, preparation energy consumption and raw material cost as optimization objectives, generate Pareto optimal solution and select the process parameters with the best engineering operability as optimization suggestions.

[0093] The specific operation flow of the full life cycle closed-loop system is as follows: sample interface data is obtained through interface recognition and feature quantification; process optimization suggestions are generated by the NSGA-Ⅲ algorithm and applied to sample preparation; the optimized samples are tested on-site to verify the interface quality and modified asphalt performance; if the interface quality does not reach the excellent level, the test data is fed back to the model training module to update the dataset and iteratively train the model; the iterative model is reapplied to interface recognition, and the optimization is repeated until the interface quality reaches the excellent level, while the optimal process parameters are included in the engineering process library.

[0094] This invention also provides a deep learning-based interface recognition system for SiC-modified asphalt, which, based on the above method, includes:

[0095] The sample preparation module is used to complete the standardized preparation of SiC modified asphalt samples, realize SiC microparticle surface modification, asphalt matrix preparation, multi-condition accelerated aging and slice pretreatment, and use cryogenic ultrathin sectioning technology to preserve the original micro morphology of the interface.

[0096] The multimodal image acquisition and processing module integrates four devices: optical microscope, SEM, EDS-Mapping and Raman spectrometer. It adopts a hierarchical time-series collaborative acquisition mode to acquire multi-dimensional images, processes images through adaptive denoising, normalization and registration algorithms, introduces an attention mechanism to enhance interface features, and outputs fused images.

[0097] The dataset management module is used to build, label, and optimize interface image datasets with multiple working conditions and ratios. It adopts a combination of manual and semi-supervised automatic labeling, and constructs a multi-scale feature labeling system through data augmentation, SMOTE-ENN algorithm, and dataset partitioning.

[0098] The model building and training module designs a lightweight improved recognition model based on the U-Net model, optimizes the encoder and decoder structure and loss function, adds a dual-channel attention mechanism and a working condition adaptive branch, and adopts quantized perception training to achieve model lightweighting to adapt to portable devices; through adaptive learning rate adjustment, K-fold cross-validation and molecular dynamics simulation-assisted optimization, the model training and performance verification are completed.

[0099] The identification and positioning module receives pre-processed multimodal images, completes interface segmentation, SiC particle positioning and defect classification through an optimized model, adopts a secondary verification algorithm to improve accuracy, supports batch sample testing, and automatically generates identification results, defect positioning list and chemical bonding state evaluation.

[0100] The quantitative optimization module automatically quantifies key feature parameters of the interface, establishes a correlation model with the performance of modified asphalt, sets quality evaluation standards, generates process optimization suggestions through a multi-objective collaborative optimization algorithm, constructs a closed-loop system for the entire life cycle, and feeds back into the sample preparation module.

[0101] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for identifying the interface of SiC-modified asphalt, characterized in that, It includes the following specific steps: Sample preparation stage: Modify SiC particles with a preset dosage and particle size using the silane coupling agent KH-550, mix and shear with matrix asphalt and modifier to prepare a matrix, simulate accelerated aging under multiple working conditions, and obtain qualified samples after slicing, coating, and pretreatment; Image processing stage: Use four devices to联动collect multi-modal interface images in a hierarchical time sequence, complete denoising, normalization, and registration through an adaptive algorithm, and introduce an attention mechanism and weighted fusion strategy to enhance interface features; Dataset construction stage: Collect multi-modal images under multiple working conditions and multiple ratios to construct a dataset, annotate them in various ways and perform diversified data optimization processing, divide the dataset, and construct a multi-scale feature label system; Model construction stage: Based on the U-Net model, design an improved U-Net model by combining the SiC modified asphalt interface recognition specific scenario, optimize the encoder, decoder structure, and loss function, add a dual-channel attention mechanism and a working condition adaptive branch, and use quantization-aware training to achieve model lightweighting; Model training stage: Input the constructed dataset into the improved model, first use molecular dynamics simulation to obtain interface microscopic parameters, then set hyperparameters and use an optimization algorithm to train the model, use the microscopic parameters as the constraints of training, and verify the model performance through multiple indicators; Recognition and positioning stage: Process the multi-modal images of the待检测样品according to the method in the image processing stage and input them into the optimized model, complete interface segmentation, SiC particle positioning, and defect classification through multi-modal feature collaborative recognition and working condition adaptive switching, and improve the accuracy through secondary verification and output the recognition result; Quantification and optimization stage: Automatically quantify the key feature parameters of the interface based on the recognition result, establish a correlation model between the key feature parameters and the performance of modified asphalt, set the interface quality evaluation standard, generate process optimization suggestions, and construct a full-life cycle closed-loop system.

2. The deep learning-based SiC-modified asphalt interface recognition method according to claim 1, characterized in that, In the sample preparation stage, the modification of SiC particles is carried out by mixing and stirring with KH-550 ethanol solution, centrifugal separation, and drying to remove surface impurities and oxide layers; the modified SiC particles, matrix asphalt, and modifier are prepared into a matrix by high-temperature and high-speed shearing, and accelerated aging is completed by simulating the actual engineering conditions; ultra-thin sections are prepared using cryo-ultramicrotomy and vacuum coating technology, and the original microscopic morphology of the interface is retained after pretreatment.

3. The deep learning-based SiC-modified asphalt interface recognition method according to claim 2, characterized in that, In the image processing stage, an optical microscope, a scanning electron microscope, an energy-dispersive X-ray spectroscopy area distribution, and a Raman spectrometer are used to联动collect multi-modal images in a hierarchical time sequence from macroscopic to microscopic and from morphology to chemical characteristics; an adaptive denoising algorithm is used to remove various image noises, normalization processing is carried out to eliminate equipment and parameter differences, and image registration is completed through feature point matching and element calibration; a channel attention mechanism and a feature interaction module are introduced, combined with an adaptive weighted fusion strategy, to enhance interface and micro-defect features and improve recognition distinguishability.

4. The deep learning-based SiC-modified asphalt interface recognition method according to claim 3, characterized in that, The dataset construction stage collects multi-modal images under multiple working conditions, multiple ratios, and multiple defect types, including various interface states; A labeling team composed of professionals from multiple fields collaborates on multimodal feature labeling, and adopts a combination of manual and semi-supervised automatic labeling to improve labeling accuracy. The sample size is expanded through diverse data augmentation methods, and training, validation and test sets are divided proportionally to construct a multi-scale, multi-dimensional feature labeling system.

5. The deep learning-based SiC-modified asphalt interface recognition method according to claim 4, characterized in that, In the model construction phase, the encoder of the improved U-Net model uses a combination of VGG16 pre-trained network and multi-scale dilated convolution module, and the decoder introduces cross-scale feature fusion module and deformable convolution module to improve feature extraction capability and interface segmentation accuracy; the dual-channel attention mechanism consists of spatial attention and channel attention, which focus on interface edges, small defects and key feature channels respectively. The loss function adopts a combination of multiple losses to optimize parameters for interface recognition scenarios; the model is lightweighted through quantization-based training in INT8 format, making it suitable for portable devices.

6. The deep learning-based SiC-modified asphalt interface recognition method according to claim 5, characterized in that, The model training phase employs the AdamW optimizer, combined with an adaptive learning rate adjustment algorithm and K-fold cross-validation, to monitor training and validation metrics in real time. A molecular dynamics simulation is used to construct the SiC-asphalt molecular interface model, obtaining interface microscopic parameters, which are then combined with Raman spectroscopy-chemical data as model optimization constraints. Key performance indicators, including model recognition accuracy and inference speed, are validated using a test set.

7. The deep learning-based SiC-modified asphalt interface recognition method according to claim 6, characterized in that, In the identification and positioning stage, the sample to be tested is input into the optimization model after image processing. The model completes interface segmentation, SiC particle positioning and defect classification through multimodal feature collaborative recognition and working condition adaptive branch switching. Different colors are used to mark interface areas, SiC particles, different types of defects and chemical bonding strength levels. The model automatically calculates key defect parameters and generates a defect positioning report. The model eliminates misidentified areas through secondary verification, optimizes the batch recognition function, and automatically outputs recognition results, defect list and chemical bonding status evaluation.

8. The deep learning-based interface recognition method for SiC-modified asphalt according to claim 7, characterized in that, In the quantification optimization stage, the key interface characteristic parameters quantified include interface thickness, bonding rate, defect density, defect size, defect depth distribution, SiC dispersion uniformity, and chemical bonding strength level. By combining various performance test data, a correlation model between characteristic parameters and modified asphalt performance is established, interface quality is divided into four levels and evaluation criteria are set; a multi-objective collaborative optimization algorithm is used to generate optimization suggestions for SiC modification process, and a closed-loop system for the entire life cycle of interface identification, process optimization, and verification iteration is constructed to feed back into the sample preparation stage.

9. A deep learning-based interface recognition system for SiC-modified asphalt, based on the method of claim 8, characterized in that, include: The sample preparation module is used to complete SiC microparticle modification, asphalt matrix preparation, multi-condition accelerated aging and slice pretreatment, and retain the original microstructure of the interface. The multimodal image acquisition and processing module integrates four devices to acquire multi-dimensional images, processes the images through adaptive algorithms and enhances interface features, and outputs fused images. The dataset management module is used to build, label, and optimize datasets. It adopts a combination of manual and semi-supervised automatic labeling to build a multi-scale feature labeling system. The model building and training module is used to design lightweight and improved recognition models, complete model training, optimization and performance verification, and adapt to portable devices; The identification and positioning module is used to receive pre-processed images, complete interface recognition, positioning and classification, improve accuracy through secondary verification, support batch detection and output relevant reports; The quantification and optimization module is used to quantify interface feature parameters, establish performance correlation models, generate process optimization suggestions, construct a closed-loop system for the entire life cycle, and provide feedback to sample preparation.

Images