Asphalt pavement compaction quality detection method fusing surface texture visual features and vibration response

Abstract

The invention provides an asphalt pavement compaction quality detection method fusing surface texture visual features and vibration response, and belongs to the technical field of construction quality detection based on machine learning. The method comprises the following steps: firstly, establishing a standardized data reference for strictly aligning construction working condition parameters, a pavement texture image, a road roller vibration response signal and a compaction quality true value; secondly, building a detection model based on a multi-modal fusion learning architecture, extracting texture features through a texture visual coding network, and describing road roller-road surface contact dynamic features by using a vibration response coding network to realize double-evidence complementary characterization; and meanwhile, a quality regression verification mechanism is embedded, a self-supervised feedback closed loop is constructed, and misjudgment risks caused by factors such as illumination and material segregation are inhibited. And finally, outputting a quality judgment result and a process adjustment suggestion based on confidence evaluation, positioning a low-quality risk section, forming a detection, decision and construction optimization closed loop, and greatly improving the timeliness, the coverage range and the engineering implementability of compaction quality control.

Description

Technical Field

[0001] This invention belongs to the field of construction quality inspection technology based on machine learning, and particularly relates to a method for inspecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response. Background Technology

[0002] In the fields of road engineering construction and intelligent transportation infrastructure, the compaction quality of asphalt pavement is considered a core aspect of quality control during asphalt pavement construction because it directly determines the density, load-bearing capacity, and durability of the pavement structure layers. However, in actual construction and engineering applications, the detection and control of asphalt pavement compaction quality still heavily relies on traditional experience-driven and offline sampling management paradigms. Construction personnel typically pre-set control conditions such as roller travel speed, vibration frequency and amplitude, number of compaction passes, wheel type combination, watering strategy, and temperature window based on a predetermined construction organization plan. Subsequently, during construction, the compaction status is indirectly assessed mainly through manual inspection, experience-based judgment, or a small number of random equipment checks. Common detection methods include core sampling, sand cone method, nuclear density meter testing, and deflection or rebound tests at a few points. While these methods can provide quality evidence to some extent, they are usually characterized by sparse testing points, significant operational interference, long testing cycles, and delayed feedback, making it difficult to meet the real-time quality control needs of large-scale continuous road sections. Under this management model, the optimization of compaction quality often relies on the accumulation of personal experience by construction teams and repeated pressure tests and adjustments, making it difficult to form an efficient, systematic and replicable process control path.

[0003] Furthermore, the vast amounts of construction parameters, on-site images, equipment vibration data, and quality inspection results accumulated by different construction units, equipment manufacturers, and testing institutions over long-term engineering practices are typically stored separately in their respective systems. These data often involve sensitive information such as equipment calibration parameters, construction organization experience, and quality management data, making centralized sharing and unified modeling difficult without increasing management costs or introducing additional risks. This reality objectively restricts the construction of large-scale, high-quality compaction process data benchmarks, further limiting the stable application of data-driven compaction quality testing models across construction sites, equipment, and materials. Therefore, how to achieve continuous, online compaction quality testing for construction sites without increasing on-site operational burdens, and maintain stable and reliable test results under complex changing conditions, has become a critical technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] To address the above problems, this invention provides a method for detecting the compaction quality of asphalt pavement that integrates surface texture visual features and vibration response, comprising the following steps: S1. Obtain the surface texture image of the road section, the vibration response signal of the road roller and the construction condition parameters, and preprocess them to obtain the resampled surface texture image, vibration response feature vector and condition parameter vector. Based on the condition parameter vector, perform sample index binding to construct the dataset. S2. Construct a compaction quality detection model, which includes: a texture visual feature encoding network to extract multi-scale texture feature representations from resampled road texture images; a vibration response feature encoding network to semantically embed vibration response feature vectors to obtain vibration embedding feature representations; a cross-modal fusion network to align and fuse the two types of features in a unified feature space; and finally, output the compaction quality prediction value through a compaction quality regression prediction head. S3, construct a composite objective function and perform two-stage alternating update training in a fixed order to obtain the trained compaction quality detection model; S4. The trained compaction quality detection model performs multi-window sampling inference on the same detection interval to obtain a prediction sequence, and calculates the representative value of compaction quality and confidence score based on the prediction sequence. It outputs a list of low-quality risk sections and uncertain sections according to fixed grading rules.

[0005] Preferably, the sample index binding based on the operating condition parameter vector is specifically as follows: The set of working parameters specifically includes four continuous working parameters: roller travel speed, vibration frequency, mixture surface temperature, and ambient temperature; and four discrete working parameters: amplitude level, number of compaction passes, wheel type, and water spraying status. The continuous working parameters are linearly normalized according to preset upper and lower bounds and mapped to [0,1]. At the same time, the discrete working parameters are one-hot encoded and converted into sparse vector format. Finally, the normalized continuous working parameters and the encoded discrete working parameters are concatenated to form a unique working parameter vector for this sample.

[0006] Preferably, compaction quality indicators are measured for each sample interval as a monitoring label; the indicators include on-site core sampling density, density measured by a nuclear density meter, compaction degree, resilient modulus, and surface smoothness.

[0007] Preferably, the texture visual feature encoding network adopts a pyramid feature extraction architecture with a fixed number of backbone encoding, intra-level parallel multi-scale convolutional branches, channel attention recalibration, and residual fusion output. It performs four levels of convolutional encoding on the input resampled road texture image and outputs texture visual features at four scales. The input image is uniformly a single-channel grayscale image. All convolutional layers in the network use zero padding to maintain consistent spatial dimensions before and after convolution. Each convolutional layer is sequentially connected to a batch normalization layer. Activation function, where The slope is fixed at 0; The first-level coding stage consists of two layers. Convolutional layer, in the second layer After the convolutional layer, perform a step of 2. Convolutional downsampling layer; the second-level encoding stage consists of two layers in sequence. Convolutional layer, in the second layer After the convolutional layer, perform a step of 2. Convolutional downsampling layer; the third-level encoding stage consists of two layers in sequence. Convolutional layer, in the second layer After the convolutional layer, perform a step of 2. Convolutional downsampling layer; the fourth-level encoding stage consists of two layers in sequence. Convolutional layer; In the second layer of each coding stage At the output of the convolutional layer, a set of parallel multi-scale convolutional branches are constructed. Each branch consists of four parallel convolutional operators. The number of output channels of each of the four parallel convolutional operators is fixed to one-quarter of the number of channels in the backbone of that level. After the concatenation of channels in each level of parallel multi-scale convolutional branch is completed, the input... Before the convolutional layer, a channel attention module is introduced to recalibrate the channel weights of the fused features; at the same time, a local contrast enhancement mechanism is introduced before entering the first-level encoding stage to suppress non-uniform lighting and enhance local texture edges.

[0008] Preferably, the vibration response feature encoding network performs a fixed-dimensional nonlinear mapping on the vibration response feature vector, employing fully connected cascaded layers, batch normalization, and... The activated structure outputs vibration semantic embedding features; the vibration response feature vector has a fixed dimension of 64. The vibration response feature encoding network contains four fully connected layers: the first fully connected layer maps the 64-dimensional input to 256-dimensional, the second fully connected layer maps the 256-dimensional input to 128-dimensional, the third fully connected layer maps the 128-dimensional input to 64-dimensional, and the fourth fully connected layer maps the 64-dimensional input to 32-dimensional and outputs the final vibration semantic embedding features. Each of the first to third fully connected layers is sequentially connected to a batch normalization layer. Activation function, where The slope of the negative half-axis is fixed at 0.1; a residual connection is added between the output of the third fully connected layer and the input of the fourth fully connected layer. The output of the third layer is linearly projected to 32 dimensions and then added element by element to the output of the fourth layer to obtain the final vibration semantic embedding feature.

[0009] Preferably, the cross-modal alignment fusion network first performs global average pooling on the four scale texture visual features to obtain texture channel description vectors at the four scales, and then maps the texture channel description vectors at each scale to 32-dimensional texture scale embedding vectors through two fully connected layers; then, the vibration semantic embedding features are copied into four paths and concatenated with the texture scale embedding vectors at the four scales respectively, and input into the gating fusion unit to generate fusion gating coefficients; Subsequently, bidirectional information interaction between texture and vibration is performed through an interactive attention alignment module. This module uses a fixed single-head dot product attention structure, where vibration semantic embedding features are used as the query vector, and texture scale embedding vectors after four scale gating are used as key and value vectors. Attention weights for the four scales are calculated, and the texture scale embedding vectors of the four scales are weighted and summed to obtain a vibration-aware texture aggregation vector. Simultaneously, a reverse attention calculation is performed using the texture aggregation vector as the query vector and the vibration semantic embedding features as the key and value vectors to obtain a texture-aware vibration recalibration vector. The vibration-aware texture aggregation vector and the texture-aware vibration recalibration vector are concatenated in the channel dimension to form a 64-dimensional cross-modal alignment and fusion vector.

[0010] Preferably, the composite objective function includes quality regression loss and cross-modal consistency loss; The quality regression loss: for the predicted compaction quality Compaction quality with true value The regression error between them is adopted The loss is measured; The loss uses a quadratic penalty when the error magnitude is small to ensure regression accuracy, and a linear penalty when the error magnitude is large to reduce the impact of outlier errors on the gradient. The cross-modal consistency loss involves aligning the multi-scale texture visual features and vibration embedding features before fusion. Specifically, the multi-scale texture visual features and vibration embedding features are mapped to a unified feature space of the same dimension through a linear projection layer, and the mapped feature vectors are normalized. Then, the cosine similarity between the texture feature vector and the vibration feature vector of the same sample is calculated. A contrastive consistency loss is constructed with the goal of improving the cross-modal similarity of the same sample and reducing the cross-modal similarity of different samples, so that the cross-modal representation distance of the same sample in the unified feature space converges.

[0011] Preferably, in step S4, the construction section is divided into continuous segments of 10 meters in length along the direction of travel as detection intervals, and each detection interval is assigned a unique interval number; continuous texture images and synchronous vibration features are acquired for each detection interval, using a length of... Five windows are fixedly sampled within the detection interval using a texture window and a sliding method with a window overlap ratio of 50%. Model inference is performed once for each window to obtain the predicted compaction quality value of that window. The predicted values ​​of the five windows within the detection interval are combined into a prediction sequence, and the arithmetic mean of the prediction sequence is calculated as the representative value of the compaction quality of the detection interval. At the same time, the standard deviation of the prediction sequence is calculated as the prediction volatility of the detection interval. The predicted volatility is converted into a confidence score, which is defined as the reciprocal of the predicted volatility and cropped to the 0-1 range. A higher confidence score indicates a more stable and reliable prediction for the detection interval. When the confidence score is less than 0.60, the detection interval is marked as a low-confidence segment and a retest is triggered. The retest method involves adding three fixed sampling windows within the same detection interval and recalculating the representative compaction quality value and the confidence score. If the confidence score after the retest is still less than 0.60, the detection interval is determined as an uncertain segment and a manual review prompt is output. The representative compaction quality value and the confidence score are combined and risk-classified using a fixed grading rule.

[0012] Preferably, the risk classification using the fixed classification rules specifically involves: When the representative compaction quality value is not less than 0.80 and the confidence score is not less than 0.70, it is judged as a high-quality stable section; when the representative compaction quality value is not less than 0.65 and less than 0.80 and the confidence score is not less than 0.70, it is judged as a medium-quality acceptable section; when the representative compaction quality value is less than 0.65 and the confidence score is not less than 0.70, it is judged as a low-quality risk section; when the confidence score is less than 0.70, it is judged as an uncertain section and enters the retesting process; finally, the section number, start and end mileage, and spatial location list of low-quality risk section and uncertain section are output.

[0013] Preferably, the process further includes the following steps: for detection intervals identified as low-quality risk sections and uncertain sections, an executable process adjustment suggestion is generated by combining the working condition parameter vector, compaction quality representative value and confidence score of the detection interval, and a compaction quality test report is output for on-site closed-loop decision-making; the process adjustment suggestion generation process includes rule triggering, candidate scheme generation, similar sample retrieval and evaluation, optimal scheme selection and closed-loop output.

[0014] Compared with the prior art, the present invention has the following innovative features and beneficial effects: (1) Design of a surface texture and vibration response fusion detection architecture for construction site interference scenarios: A dual evidence fusion compaction quality detection architecture is proposed, which abandons the judgment method that relies solely on a single visual or single vibration signal. Through parallel modeling and cross-modal fusion of texture visual branch and vibration response branch, the two types of evidence can be mutually verified and compensated in a unified semantic space. Among them, vibration response is used to suppress pseudo-texture changes caused by factors such as uneven illumination, surface contamination, and local water accumulation, while surface texture is used to compensate for the non-uniqueness of vibration signals under different material states and different working conditions, thereby achieving robust judgment and cross-scenario generalization of compaction quality. (2) Design of an end-to-end collaborative optimization strategy with embedded quality verification and consistency orientation: An end-to-end collaborative optimization strategy without external freezing proxy model is proposed. A quality verification mechanism is introduced into the detection model and trained together with the backbone fusion network. By constructing a composite optimization objective that includes quality regression error and cross-modal consistency constraints, the model can actively suppress unreliable representations of "conflict between texture evidence and vibration evidence" while learning the compaction quality mapping law. This strategy avoids the error accumulation problem caused by the deviation of the external calibration model in the traditional step-by-step process and enables the model to maintain more stable detection consistency under the conditions of changes in construction site working conditions, equipment differences and data distribution drift. (3) Construction of standardized benchmarks for multi-source heterogeneous data and confidence-driven construction closed-loop decision-making system: In response to the problems of complex data modes, unstable acquisition conditions, and difficulty in directly converting test results into construction actions at the construction site, a full-link system from data standardization to closed-loop decision output is established: First, multi-source heterogeneity is eliminated through a unified image standardization and signal preprocessing process, and strict index association is established between working conditions, texture, vibration and quality true values ​​to form a standardized joint data benchmark that can be expanded sustainably; Second, a confidence assessment and risk classification mechanism is designed to conduct stability assessment and anomaly removal of test results for continuous road sections, and output locatable results for low-quality risk sections; Most importantly, risk sections are associated with construction working conditions to generate executable rolling parameter adjustment suggestions, thereby realizing the engineering implementation and on-site closed-loop optimization of compaction quality test results. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the following description is only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the overall process of the asphalt pavement compaction quality testing method of the present invention.

[0017] Figure 2 This is a detailed architecture diagram of the cross-modal dual-evidence fusion network of the present invention.

[0018] Figure 3 This is a diagram illustrating the training strategy for the compaction quality detection model of this invention.

[0019] Figure 4 This is a scatter plot comparing the compaction quality detection accuracy in the examples.

[0020] Figure 5 This is a robustness attenuation curve of the detection model under complex interference conditions in the embodiment.

[0021] Figure 6 This is a diagram illustrating the construction closed-loop quality optimization effect based on confidence level assessment in the example. Detailed Implementation

[0022] To achieve online, continuous, and highly reliable detection of asphalt pavement compaction quality, this invention proposes a method for detecting asphalt pavement compaction quality that integrates surface texture visual features and vibration response. Figure 1 As shown, the method includes: S1: Acquire pavement surface texture images, roller vibration response signals, and construction condition information of the construction section, and process them to obtain resampled pavement texture images, vibration response feature vectors, and construction condition parameter vectors. Then, bind the resampled pavement texture images, vibration response feature vectors, and construction condition parameter vectors one by one according to the detection interval number, and simultaneously acquire the ground truth label of compaction quality for that detection interval to construct a joint dataset of compaction samples. Finally, the resampled pavement texture images, vibration response feature vectors, and compaction quality ground truth labels are used together as the data input for model training in S2, while the construction condition parameter vectors are retained as the data input for evaluation and suggestion generation in S4. S2 takes the resampled road texture image and vibration response feature vector output from S1 as the model input, then constructs a texture visual feature encoding network and a vibration response feature encoding network, and outputs the compaction quality prediction result through a cross-modal alignment fusion network and a compaction quality regression prediction head; at the same time, it outputs the multi-scale texture feature representation and vibration embedding feature representation before fusion as the input of the consistency constraint in S3. S3 uses the compaction quality prediction results output by S2 and the ground truth labels of compaction quality output by S1 as regression supervision inputs, and the multi-scale texture feature representation and vibration embedding feature representation before fusion output by S2 as cross-modal consistency constraint inputs. A composite objective function containing quality regression loss and cross-modal consistency loss is constructed and a two-stage alternating update training with a fixed order is performed to obtain the trained compaction quality detection model. Then, the trained compaction quality detection model is used as the deployment inference model in S4. S4 uses the deployment inference model output from S3 to perform multi-window sampling inference on the same detection interval to obtain a predicted sequence. Based on the predicted sequence, it calculates the representative value of compaction quality and the confidence score, and outputs a list of low-quality risk sections and uncertain sections according to fixed grading rules. Then, using the representative value of compaction quality and the confidence score of low-quality risk sections and uncertain sections, as well as the working condition parameter vector output from S1, as input for suggestion generation, it outputs the optimal process adjustment suggestion according to the similar sample retrieval evaluation rule. Finally, it outputs a compaction quality inspection report containing the road section compaction quality curve, the list of risk section locations, the confidence assessment results, and the process adjustment suggestions, realizing closed-loop decision output at the construction site.

[0023] The implementation process of the present invention will be further described below with reference to specific embodiments.

[0024] S1. Collaborative Acquisition of Multi-Source Heterogeneous Data and Construction of Joint Dataset for Asphalt Pavement Compaction Process This step aims to construct a standardized benchmark dataset that is strictly aligned and modally unified with surface texture visual features, vibration response features, and compaction quality ground truth values. This establishes a learnable mathematical mapping relationship between the changes in pavement surface texture under different construction batches and compaction conditions and the dynamic vibration response of the road roller, providing a supervised learning sample foundation for subsequent compaction quality detection models. The process is as follows: First, vectorization of working condition parameters and binding of sample indices are performed; second, pavement surface texture image acquisition and standardized preprocessing are performed; third, vibration response signal acquisition and time-frequency feature representation are performed in parallel; subsequently, compaction quality ground truth calibration and label normalization are performed; finally, a multi-source heterogeneous joint dataset is constructed and divided into training, validation, and test sets.

[0025] S1-1 Working Condition Parameter Vectorization and Sample Index Binding: Records the set of key working condition parameters during the construction process of the compaction zone sample being tested; the set of working condition parameters specifically includes: four continuous working condition parameters: roller travel speed, vibration frequency, mixture surface temperature and ambient temperature, and four discrete working condition parameters: amplitude level, number of compaction passes, wheel type and watering status; in order to eliminate the interference of different physical dimensions and value ranges on the model training weight allocation, the continuous working condition parameters are linearly normalized according to preset upper and lower bounds, and mapped to [0,1]; at the same time, the discrete working condition parameters are one-hot encoded and converted into sparse vector format; finally, the normalized continuous working condition parameters and the encoded discrete working condition parameters are concatenated to form a unique working condition parameter vector for the sample.

[0026] S1-2 Road Surface Texture Visual Data Acquisition and Standardized Preprocessing: Industrial cameras fixedly installed at the front and rear of the road roller continuously image the road surface during the compaction process to obtain the original road texture image sequence. Given that on-site imaging is susceptible to uneven lighting, dust pollution, lens angle changes, and scale drift, this application establishes a unified image preprocessing workflow. First, grayscale space standardization is performed, and intensity truncation and histogram equalization are applied to the original road texture image sequence to eliminate brightness bias and enhance the contrast of aggregate edges, compaction strips, and microtextures, obtaining standardized road texture images. Second, pixel scale coefficients are obtained through camera calibration, and the standardized road texture images are resampled to a unified physical field of view. Non-road areas are masked and cropped to obtain resampled road texture images. Finally, anti-motion artifact processing is performed, stabilizing and registering adjacent frames and removing blurred frames to ensure that the texture statistics between different samples are comparable at the pixel level, resulting in the final resampled road texture image.

[0027] S1-3 Vibration Response Signal Acquisition and Time-Frequency Feature Characterization: To synchronously characterize the dynamic response features during the compaction process, acceleration sensors are deployed at key locations on the roller's steel wheel bearing housing or frame to acquire the original time-series vibration response signals. Considering the signal heterogeneity caused by different equipment installation stiffness, speed drift, and construction noise, this application establishes a unified signal preprocessing and characterization process. First, detrending and bandpass filtering are performed to suppress low-frequency attitude drift and high-frequency random noise. Second, the same time window segmentation as in S1-2 is performed, and a short-time Fourier transform is calculated for each signal segment to obtain the time-frequency energy spectrum. Finally, the spectral peak amplitude, dominant frequency drift, bandwidth energy ratio, and spectral entropy features that reflect changes in compaction stiffness and contact state are extracted to form the vibration response feature vector of the sample.

[0028] S1-4 Compaction Quality True Value Calibration and Label Normalization: For each sample interval, the compaction quality index is measured as a supervision label; the index includes the core density taken on site, the density measured by the nuclear density meter, the degree of compaction, the resilient modulus, and the surface smoothness; in order to ensure that the magnitude of the multi-task loss term is controllable during subsequent model training, the compaction quality index is linearly normalized according to the preset physical upper and lower bounds to construct a standardized compaction quality label y; and y∈[0,1], the larger the value, the higher the compaction quality.

[0029] Construction of S1-5 Multi-Source Heterogeneous Joint Dataset: The resampled road texture image obtained in S1-2, the vibration response feature vector obtained in S1-3, and the compaction quality label y obtained in S1-4 together constitute a joint dataset for model training of the compaction quality detection model in S2.

[0030] S2. Construct a compaction quality detection model that integrates surface texture visual features and vibration response. First, a texture visual feature encoding network is built to extract multi-scale texture representations from resampled road surface texture images. Second, a vibration response feature encoding network is built to semantically embed vibration response feature vectors. Third, a cross-modal fusion network is constructed to align and fuse the two types of evidence in a unified feature space. Finally, a compaction quality regression prediction head is built to output the predicted compaction quality value. ,in, y represents the compaction quality predicted by the model; it corresponds to the true value label y.

[0031] S2-1 Constructing a Texture Visual Feature Encoding Network: The texture visual feature encoding network adopts a pyramid feature extraction architecture with a fixed-layer backbone encoding, intra-level parallel multi-scale convolutional branches, channel attention recalibration, and residual fusion output. It performs four levels of convolutional encoding on the input resampled road texture image and outputs texture visual features at four scales. The input image is uniformly a single-channel grayscale image with a fixed spatial size. All convolutional layers in the network use zero padding to maintain consistent spatial dimensions before and after convolution. Each convolutional layer is followed by a batch normalization layer. Activation function, where The slope is fixed at 0.

[0032] The first-level coding stage consists of two layers. The convolutional layer has 64 and 64 output channels respectively, with a stride of 1 for both; in the second layer... After the convolutional layer, perform a step of 2. The convolutional downsampling layer has a fixed number of output channels of 64; the second-level encoding stage consists of two layers. The convolutional layer has 128 and 128 output channels, with a stride of 1 for both; in the second layer... After the convolutional layer, perform a step of 2. The convolutional downsampling layer has a fixed number of output channels of 128; the third-level encoding stage consists of two layers. The convolutional layer has 256 and 256 output channels, with a stride of 1 for both; in the second layer... After the convolutional layer, perform a step of 2. The convolutional downsampling layer has a fixed number of output channels of 256; the fourth-level encoding stage consists of two layers. The convolutional layers have 512 and 512 output channels, with a stride of 1 for both. The above four-level encoding stages form backbone feature maps at four scales, which are used to progressively expand the receptive field and form pyramid texture representations.

[0033] In the second layer of each coding stage At the output of the convolutional layer, a set of parallel multi-scale convolutional branches is constructed. Each parallel multi-scale convolutional branch consists of four parallel convolution operators, with the kernel size fixed at [size missing]. , , and The output channels of the four parallel convolution operators are all fixed at one-quarter of the number of channels in the backbone of that level; the outputs of the four parallel branches are concatenated along the channel dimension, and the number of channels after concatenation is consistent with the number of channels in the backbone of that level; the input of the concatenated feature map... Convolutional layers perform cross-scale information mixing, with the first to fourth levels... The number of convolutional output channels is fixed at 64, 128, 256, and 512, respectively; The output of the convolutional layer is fused with the backbone feature map of that level through element-wise addition using a residual connection method to obtain the multi-scale fused texture feature output of that scale; the multi-scale fused texture feature outputs of the four scales together form a multi-scale texture visual feature set.

[0034] To enhance robustness to changes in construction lighting and localized contamination, after each level of parallel multi-scale convolutional branch channel is spliced, the input... Before the convolutional layer, a channel attention module is introduced to recalibrate the channel weights of the fused features; the channel attention module uses a fixed approach of "global average pooling + two fully connected layers + ... The structure involves global average pooling compressing the concatenated feature maps into channel description vectors; the first fully connected layer reduces the channel dimension from... Compress to and connect after this layer. Activation function; the second fully connected layer changes the channel dimension from Restore to and connect after this layer. The activation function yields the channel weight vector; finally, the channel weight vector is multiplied element-wise by channel with the concatenated feature map to enhance the effective texture channels and suppress interfering channels; where... The values ​​64, 128, 256, and 512 are respectively used to represent the first to fourth levels of coding.

[0035] Simultaneously, a local contrast enhancement mechanism is introduced before entering the first-level encoding stage to suppress non-uniform lighting and enhance local texture edges; the local contrast enhancement mechanism uses a fixed method. A sliding window is used to calculate the local mean and local standard deviation maps, and local normalization is performed on each pixel. The local normalization results are then superimposed on the original input image using a residual method to form an enhanced image, where the superposition weights are fixed at 0.7 and 0.3. The enhanced image is then cropped pixel by pixel. The numerical range is then used as the final input to the texture visual feature encoding network, thereby enhancing the local contrast between the aggregate edge and the compaction strip while maintaining the overall brightness trend, and improving the robustness of texture visual feature encoding to complex on-site imaging conditions.

[0036] S2-2 Constructing a Vibration Response Feature Encoding Network: The vibration response feature encoding network performs a fixed-dimensional nonlinear mapping on the vibration response feature vectors obtained in S1-3, employing fully connected cascaded layers, batch normalization, and... The activated structure outputs vibration semantic embedding features; wherein, the vibration response feature vector has a fixed dimension of 64, and the 64-dimensional vector is composed of spectral peak amplitude, dominant frequency drift, bandwidth-energy ratio, spectral entropy, and statistics related to vibration amplitude and frequency in a fixed order; the vibration response feature encoding network contains four fully connected layers, the first fully connected layer maps the 64-dimensional input to 256 dimensions, the second fully connected layer maps the 256-dimensional input to 128 dimensions, the third fully connected layer maps the 128-dimensional input to 64 dimensions, and the fourth fully connected layer maps the 64-dimensional input to 32 dimensions and outputs the final vibration semantic embedding features; after the first to third fully connected layers, a batch normalization layer is sequentially connected to... Activation function, where The negative half-axis slope is fixed at 0.1; the normalization layer and activation function are not connected after the fourth fully connected layer to maintain the linear separability of the embedded features; to suppress the embedding drift caused by individual abnormal windows, a residual connection is added between the output of the third fully connected layer and the input of the fourth fully connected layer. The output of the third layer is linearly projected to 32 dimensions and then added element-wise with the output of the fourth layer to obtain the final vibration semantic embedding features; the vibration semantic embedding features are used to characterize the equivalent stiffness change and energy dissipation characteristics of the road roller road contact system, and serve as the vibration input branch features of the cross-modal fusion network.

[0037] S2-3 constructs a cross-modal alignment and fusion network and a quality prediction head: The four-scale texture visual features output from S2-1 and the vibration semantic embedding features output from S2-2 are input into the cross-modal alignment and fusion network. A structure employing intra-scale gating fusion, interactive attention alignment, and inter-scale cascaded fusion is used to output a unified multimodal fusion feature. This feature is then used to output compaction quality prediction values ​​through a compaction quality regression prediction head. Figure 2 As shown.

[0038] Specifically, firstly, global average pooling is performed on the texture visual features at four scales to obtain texture channel description vectors at four scales. Then, each scale's texture channel description vector is mapped to a 32-dimensional texture scale embedding vector through two fully connected layers. Subsequently, the vibration semantic embedding features are copied into four paths and concatenated with the texture scale embedding vectors at the four scales respectively. These concatenated vectors are then input into a gating fusion unit to generate fusion gating coefficients, achieving vibration-guided texture filtering within the scale. The gating fusion unit uses a fixed fully connected layer and... The activation function outputs gating coefficients, which are fixed to 32 dimensions. The gating coefficients are then multiplied element-wise by the texture scale embedding vector along each dimension to suppress pseudo-texture responses caused by illumination changes or local contamination and to enhance stable texture representation consistent with the compaction state.

[0039] Subsequently, an interactive attention alignment module is constructed to realize bidirectional information interaction between texture and vibration. The interactive attention alignment module adopts a fixed single-head dot product attention structure, in which vibration semantic embedding features are used as query vectors, and texture scale embedding vectors after four scales are gated as key and value vectors. Attention weights for the four scales are calculated, and the texture scale embedding vectors of the four scales are weighted and summed accordingly to obtain a vibration-aware texture aggregation vector. At the same time, a reverse attention calculation is performed using the texture aggregation vector as query vector and vibration semantic embedding features as key and value vectors to obtain a texture-aware vibration recalibration vector. The vibration-aware texture aggregation vector and the texture-aware vibration recalibration vector are concatenated in the channel dimension to form a 64-dimensional cross-modal alignment and fusion vector.

[0040] Finally, the cross-modal aligned fusion vector is input into the compaction quality regression prediction head to output the compaction quality prediction value; the compaction quality regression prediction head fixedly contains three fully connected layers, the first layer maps 64 dimensions to 32 dimensions and connects the batch normalization layer with... The activation function, the second layer maps 32 dimensions to 16 dimensions and connects the batch normalization layer to... The activation function, in the third layer, maps 16 dimensions to 1 dimension and connects them. The activation function outputs the predicted compaction quality value in the interval between 0 and 1. .

[0041] S3, Embedded Quality Verification and Multimodal Collaborative Training Optimization This step aims to establish an end-to-end quality constraint mechanism without the need for additional proxy models. It reuses the fusion feature representation built by S2 and introduces an embedded quality verification head within the model to achieve the dual functions of feature consistency constraint and quality regression supervision. This allows for the simultaneous optimization of multimodal alignment and quality prediction accuracy within a single training iteration. First, a composite objective function containing quality regression loss and cross-modal consistency loss is constructed. Second, a fixed-order, two-stage alternating update joint training strategy is implemented: In each training iteration, the first stage only updates the compaction quality regression prediction head parameters and minimizes the quality regression loss; the second stage keeps the compaction quality regression prediction head parameters unchanged and updates the parameters of the texture visual feature encoding network, vibration response feature encoding network, and cross-modal alignment fusion network to minimize the cross-modal consistency loss. Through this fixed-order and fixed-update training strategy, the model can meet engineering truth values ​​while avoiding the risk of conflicting samples where the texture appears compacted but the vibration does not match, or the vibration shows compaction but the texture is contaminated and misjudged. Specific steps are as follows: Figure 3 As shown.

[0042] S3-1 Constructing a composite objective function based on quality verification: To quantize the training objective of the model and provide an optimizable gradient, construct a composite objective function L, which includes at least the following two types of loss terms: (1) Quality regression loss: In order to suppress the excessive pull of a small number of extreme error samples on the model parameter update and improve training stability, the quality regression loss is used to predict the compaction quality. Compaction quality with true value The regression error between them is adopted The loss is measured; The loss employs a quadratic penalty when the error magnitude is small to ensure regression accuracy, and a linear penalty when the error magnitude is large to reduce the impact of abnormal errors on the gradient, thereby achieving robust learning of noise perturbations under complex working conditions without relying on additional sample class determination rules. (2) Cross-modal consistency loss: Alignment constraints are applied to the multi-scale texture visual features and vibration embedding features before fusion. Specifically, the multi-scale texture visual features and vibration embedding features are mapped to a unified feature space of the same dimension through a linear projection layer, and the mapped feature vectors are normalized. Then, the cosine similarity between the texture feature vector and the vibration feature vector of the same sample is calculated. A contrastive consistency loss is constructed with the goal of improving the cross-modal similarity of the same sample and reducing the cross-modal similarity of different samples. This makes the cross-modal representation distance of the same sample in the unified feature space converge, thereby reducing the modal conflict caused by changes in illumination, local contamination or fluctuations in operating conditions and improving the consistency and discriminability of the fused features.

[0043] Finally, the above loss terms are summed according to preset weights to obtain the composite objective function. And used for subsequent joint training.

[0044] in, For training objective function; y and y represent the predicted compaction quality and the true compaction quality, respectively.

[0045] S3-2 performs consistency-oriented joint training and parameter updates: It abandons the fragmented process of pre-training a single modality before fusion, and adopts an end-to-end joint training strategy; in each training iteration, the model simultaneously receives texture images and vibration feature inputs from the same sample, and updates them in the composite objective function. Backpropagation updates are performed under constraints; the gradient flow generated by the cross-modal consistency loss is simultaneously backpropagated to the texture visual encoding branch and the vibration response encoding branch, driving the two modal representations to adaptively align in the fusion space; the gradient flow generated by the quality regression loss drives the fusion network and the quality regression layer to learn stable engineering mapping rules; as training progresses, the model gradually learns to output reliable compaction quality predictions under the condition of mutual verification between texture evidence and vibration evidence, thereby achieving highly robust compaction quality detection without the assistance of an external proxy model.

[0046] S4. Output of quality assessment and process adjustment recommendations based on confidence level evaluation This step deploys the trained compaction quality detection model to the construction application, performing online detection and result interpretation output for newly collected road sections. First, sliding window multi-sample inference is performed to obtain the compaction quality prediction sequence for continuous road sections. Second, confidence assessment and anomaly removal are performed to output the compaction quality judgment result within the confidence interval, which is the model's prediction result. Finally, the working condition parameter vector and prediction results are combined to generate executable construction adjustment suggestions to guide the dynamic optimization of subsequent compaction passes, speed, and vibration parameters.

[0047] S4-1 Online Inference and Continuous Road Segment Quality Curve Generation: The construction road segment is divided into continuous sample intervals of uniform length. Corresponding texture images and vibration response features are collected sequentially and input into the model. The predicted compaction quality value of each interval is output. They are then spliced ​​together in spatial order to form a road section compaction quality curve, which is used to quickly locate low-quality risk sections.

[0048] S4-2 Confidence Assessment and Low-Quality Risk Segment Determination: To avoid misjudgment caused by noise from a single measurement, a confidence scoring mechanism is constructed based on the results of multiple sampling inferences using a sliding window, and a list of low-quality risk segments and their spatial locations are output. Specifically, the construction section is divided into continuous 10-meter-long sections along the direction of travel as detection intervals, and each detection interval is assigned a unique interval number; continuous texture images and synchronous vibration features are collected for each detection interval, using a length of... Five windows are fixedly sampled within the detection interval using a texture window and a sliding method with a window overlap ratio of 50%. Model inference is performed once for each window to obtain the predicted compaction quality value of that window. The predicted values ​​of the five windows within the detection interval are combined into a prediction sequence, and the arithmetic mean of the prediction sequence is calculated as the representative value of the compaction quality of the detection interval. At the same time, the standard deviation of the prediction sequence is calculated as the prediction volatility of the detection interval. Furthermore, the predicted volatility is converted into a confidence score, which is defined as the reciprocal of the predicted volatility and cropped to the range of 0 to 1. The higher the confidence score, the more stable and reliable the prediction for that detection interval. When the confidence score is less than 0.60, the detection interval is marked as a low-confidence segment and a retest is triggered. The retest method is to add three fixed sampling windows within the same detection interval and recalculate the representative value of compaction quality and the confidence score. If the confidence score after the retest is still less than 0.60, the detection interval is determined to be an uncertain segment and a manual review prompt is output.

[0049] Simultaneously, the representative compaction quality value and confidence score are combined for risk classification, with the following fixed classification rules: when the representative compaction quality value is not less than 0.80 and the confidence score is not less than 0.70, it is judged as a high-quality stable section; when the representative compaction quality value is not less than 0.65 and less than 0.80 and the confidence score is not less than 0.70, it is judged as a medium-quality acceptable section; when the representative compaction quality value is less than 0.65 and the confidence score is not less than 0.70, it is judged as a low-quality risk section; when the confidence score is less than 0.70, it is judged as an uncertain section and enters the retesting process; finally, the interval number, start and end mileage, and spatial location list of low-quality risk sections and uncertain sections are output, which are used to generate construction adjustment suggestions in subsequent steps and realize closed-loop control of compaction quality.

[0050] S4-3 Process Adjustment Suggestion Generation and Closed-Loop Output: For the detection intervals identified as low-quality risk and uncertain intervals in S4-2, the system combines the working condition parameter vector, compaction quality representative value, and confidence score of the detection interval to generate executable process adjustment suggestions and outputs a compaction quality test report for on-site closed-loop decision-making. The process adjustment suggestion generation process includes five fixed steps: rule triggering, candidate scheme generation, similar sample retrieval and evaluation, optimal scheme selection, and closed-loop output, as detailed below.

[0051] When the detection interval is determined to be a low-quality risk area, the system directly enters the process adjustment process and outputs adjustment suggestions. When the detection interval is determined to be an uncertain area, the system first outputs a retest instruction, and enters the process adjustment process after the retest is completed and the confidence score is not less than 0.70. If the confidence score is still less than 0.70 after the retest, only the prompt message "reduce interference and manually verify" is output, and no parameter adjustment suggestions are output. When the detection interval is determined to be a medium-quality acceptable area, the system outputs a prompt message to maintain the current process and continue monitoring. When the detection interval is determined to be a high-quality stable area, the system outputs a prompt message to "maintain the current process and allow the sampling step size to be increased".

[0052] Next, candidate solutions are generated. For the detection range entering the process adjustment flow, the system generates five fixed candidate adjustment solutions: compaction pass adjustment solution, travel speed adjustment solution, vibration frequency adjustment solution, amplitude level adjustment solution, and temperature window handling solution; as detailed below: The rolling pass adjustment scheme is fixed at increasing the current rolling pass number by 2 passes; the travel speed adjustment scheme is fixed at decreasing the current travel speed by 0.3 m / s; the vibration frequency adjustment scheme is fixed at increasing the current vibration frequency by 5 Hz; the amplitude level adjustment scheme is fixed at increasing the current amplitude level by one level; the temperature window handling scheme is fixed according to the following rules: when the surface temperature of the mixture is below 110 degrees Celsius, the handling suggestion of "pausing rolling in this section and waiting for the temperature to recover or re-paving" is output, and any candidate schemes to increase the vibration frequency or increase the amplitude level are prohibited from being output; when the surface temperature of the mixture is above 160 degrees Celsius, the handling suggestion of "reducing the amplitude level by one level and decreasing the travel speed by 0.3 m / s" is output; when the surface temperature of the mixture is between 110 degrees Celsius and 160 degrees Celsius, the pause or downgrade handling is not triggered, and it only participates in the evaluation and selection of the aforementioned four types of parameter adjustment candidate schemes.

[0053] The similar sample retrieval evaluation is performed again to introduce a data-driven suggestion selection mechanism; before deployment, the system establishes an offline "operating condition-quality response retrieval library" based on the joint dataset constructed in step S1, and calls the retrieval library in real time at the construction end to score the expected benefits of each candidate scheme; specifically, the operating condition parameter vector of the current detection interval is used as the query key, the comprehensive distance of each historical sample in the retrieval library is calculated and sorted in ascending order of distance, and the 30 historical samples with the smallest distance are fixedly selected as the similar sample set; The comprehensive distance is calculated by weighted summation of the normalized Euclidean distance for continuous parameters and the Hamming distance for discrete parameters. The continuous parameters include the roller's travel speed, vibration frequency, mixture surface temperature, and ambient temperature. The discrete parameters include amplitude level, number of compaction passes, wheel type, and water spraying status. For the continuous parameter distance, the normalized difference of the four continuous parameters is calculated, and the square root of the sum of squares is taken to obtain the Euclidean distance. For the discrete parameter distance, the four discrete parameters are compared for similarity; differences are recorded as 1, and similarities as 0, and the summation is used to obtain the Hamming distance. The comprehensive distance is fixed as the sum of the continuous parameter Euclidean distance multiplied by 0.7 and the discrete parameter Hamming distance multiplied by 0.3. When historical samples with the same comprehensive distance are found, the historical sample whose true compaction quality value is closer to the representative compaction quality value of the current detection interval is selected first. If they are still the same, the historical samples are selected from the most recent to the oldest collection time.

[0054] After obtaining the similar sample set, the system calculates the expected quality gain score for each candidate adjustment scheme. Specifically, it selects a subset of historical samples from the similar sample set that corresponds to the action of the candidate adjustment scheme, and calculates the arithmetic mean of the difference in the true compaction quality before and after the action in this subset as the expected quality gain score for that action. The consistency judgment rule for the compaction pass adjustment scheme is that the number of compaction passes in the historical samples increases by 2 compared to its baseline state; the consistency judgment rule for the travel speed adjustment scheme is that the travel speed in the historical samples decreases by 0.3 m / s compared to its baseline state; vibration frequency... The consistency judgment rule for adjustment schemes is that the vibration frequency in historical samples is increased by 5 Hz relative to its baseline state; the consistency judgment rule for amplitude level adjustment schemes is that the amplitude level in historical samples is increased by one level relative to its baseline state; if a candidate adjustment scheme does not have historical samples in the similar sample set that meet the consistency judgment rule, the expected quality gain score of the candidate adjustment scheme is fixed as 0; at the same time, a confidence weighting strategy is introduced, which multiplies the expected quality gain score of the detection interval by the confidence score to obtain the final confidence weighted score, so as to reduce the risk of outputting aggressive adjustment schemes in low stability sections.

[0055] The system then selects the optimal solution and sets hard constraints. From five candidate adjustment schemes, it chooses the scheme with the highest final score as the optimal adjustment suggestion, satisfying the following hard constraints: the maximum number of compaction passes is fixed at 12; the minimum travel speed is fixed at 0.8 m / s; the maximum vibration frequency is fixed at 60 Hz; the maximum amplitude level is fixed at the highest level and the minimum level is the lowest level; when the surface temperature of the mixture is below 110 degrees Celsius, suggestions to increase the vibration frequency and amplitude level are prohibited; when water spraying is on and the ambient temperature is below 5 degrees Celsius, temperature window handling suggestions are prioritized; if construction still needs to continue, suggestions to increase the vibration frequency and amplitude level are prohibited; when the wheel type is rubber tires, suggestions to increase the vibration frequency and amplitude level are prohibited; if the optimal candidate scheme violates any hard constraint, the next scheme is selected in descending order of the final score until all hard constraints are satisfied; if all schemes fail to meet the hard constraints, only a "pause and review" handling suggestion is output.

[0056] Finally, the closed-loop output is executed. The system summarizes the interval number, start and end mileage, representative compaction quality value, confidence score, risk level, retest status, optimal adjustment suggestion type, optimal adjustment suggestion parameters, and corresponding confidence weighted score of each test interval in mileage order. It generates and outputs a compaction quality test report containing the road section compaction quality curve, a list of low-quality risk sections and uncertain sections, confidence assessment results, retest records, and process adjustment suggestions. The optimal adjustment suggestions for low-quality risk sections are pushed to the on-site construction terminal in the form of interval number and start and end mileage to guide the immediate implementation of additional compaction or parameter adjustment on-site, thereby realizing a rapid closed-loop decision-making process of compaction quality test, risk location, and process correction.

[0057] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0058] While the above description illustrates specific embodiments of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response, characterized in that, Includes the following processes: S1. Obtain the surface texture image of the road section, the vibration response signal of the road roller and the construction condition parameters, and preprocess them to obtain the resampled surface texture image, vibration response feature vector and condition parameter vector. Based on the condition parameter vector, perform sample index binding to construct the dataset. S2. Construct a compaction quality detection model, which includes: a texture visual feature encoding network to extract multi-scale texture feature representations from resampled road texture images; a vibration response feature encoding network to semantically embed vibration response feature vectors to obtain vibration embedding feature representations; a cross-modal fusion network to align and fuse the two types of features in a unified feature space; and finally, output the compaction quality prediction value through a compaction quality regression prediction head. S3, construct a composite objective function and perform two-stage alternating update training in a fixed order to obtain the trained compaction quality detection model; S4. The trained compaction quality detection model performs multi-window sampling inference on the same detection interval to obtain a prediction sequence, and calculates the representative value of compaction quality and confidence score based on the prediction sequence. It outputs a list of low-quality risk sections and uncertain sections according to fixed grading rules.

2. The method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response as described in claim 1, characterized in that: The specific steps for binding sample indexes based on operating condition parameter vectors are as follows: The set of working parameters specifically includes four continuous working parameters: roller travel speed, vibration frequency, mixture surface temperature, and ambient temperature; and four discrete working parameters: amplitude level, number of compaction passes, wheel type, and water spraying status. The continuous working parameters are linearly normalized according to preset upper and lower bounds and mapped to [0,1]. At the same time, the discrete working parameters are one-hot encoded and converted into sparse vector format. Finally, the normalized continuous working parameters and the encoded discrete working parameters are concatenated to form a unique working parameter vector for this sample.

3. The method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response as described in claim 1, characterized in that: For each sample interval, compaction quality indicators were measured as monitoring labels; the indicators included core density taken on site, density measured by nuclear density meter, compaction degree, resilient modulus, and surface smoothness.

4. The method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response as described in claim 1, characterized in that: The texture visual feature encoding network employs a pyramid feature extraction architecture with a fixed-layer backbone encoding, intra-level parallel multi-scale convolutional branches, channel attention recalibration, and residual fusion output. It performs four levels of convolutional encoding on the input resampled road texture image and outputs texture visual features at four scales. The input image is uniformly a single-channel grayscale image. All convolutional layers in the network use zero padding to maintain consistent spatial dimensions before and after convolution. Each convolutional layer is sequentially connected to a batch normalization layer. Activation function, where The slope is fixed at 0; The first-level coding stage consists of two layers. Convolutional layer, in the second layer After the convolutional layer, perform a step of 2. Convolutional downsampling layer; the second-level encoding stage consists of two layers in sequence. Convolutional layer, in the second layer After the convolutional layer, perform a step of 2. Convolutional downsampling layer; the third-level encoding stage consists of two layers in sequence. Convolutional layer, in the second layer After the convolutional layer, perform a step of 2. Convolutional downsampling layer; the fourth-level encoding stage consists of two layers in sequence. Convolutional layer; In the second layer of each coding stage At the output of the convolutional layer, a set of parallel multi-scale convolutional branches are constructed. Each branch consists of four parallel convolutional operators. The number of output channels of each of the four parallel convolutional operators is fixed to one-quarter of the number of channels in the backbone of that level. After the concatenation of channels in each level of parallel multi-scale convolutional branch is completed, the input... Before the convolutional layer, a channel attention module is introduced to recalibrate the channel weights of the fused features; at the same time, a local contrast enhancement mechanism is introduced before entering the first-level encoding stage to suppress non-uniform lighting and enhance local texture edges.

5. The method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response as described in claim 4, characterized in that: The vibration response feature encoding network performs a fixed-dimensional nonlinear mapping on the vibration response feature vector, employing fully connected cascaded layers, batch normalization, and... The activated structure outputs vibration semantic embedding features; the vibration response feature vector has a fixed dimension of 64. The vibration response feature encoding network contains four fully connected layers: the first fully connected layer maps the 64-dimensional input to 256-dimensional, the second fully connected layer maps the 256-dimensional input to 128-dimensional, the third fully connected layer maps the 128-dimensional input to 64-dimensional, and the fourth fully connected layer maps the 64-dimensional input to 32-dimensional and outputs the final vibration semantic embedding features. Each of the first to third fully connected layers is sequentially connected to a batch normalization layer. Activation function, where The slope of the negative half-axis is fixed at 0.1; a residual connection is added between the output of the third fully connected layer and the input of the fourth fully connected layer. The output of the third layer is linearly projected to 32 dimensions and then added element by element to the output of the fourth layer to obtain the final vibration semantic embedding feature.

6. The method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response as described in claim 5, characterized in that: The cross-modal alignment fusion network first performs global average pooling on the texture visual features at four scales to obtain texture channel description vectors at four scales, and then maps the texture channel description vectors at each scale to 32-dimensional texture scale embedding vectors through two fully connected layers; then, the vibration semantic embedding features are copied into four paths and concatenated with the texture scale embedding vectors at the four scales respectively, and input into the gating fusion unit to generate fusion gating coefficients. Subsequently, bidirectional information interaction between texture and vibration is performed through an interactive attention alignment module. This module uses a fixed single-head dot product attention structure, where vibration semantic embedding features are used as the query vector, and texture scale embedding vectors after four scale gating are used as key and value vectors. Attention weights for the four scales are calculated, and the texture scale embedding vectors of the four scales are weighted and summed to obtain a vibration-aware texture aggregation vector. Simultaneously, a reverse attention calculation is performed using the texture aggregation vector as the query vector and the vibration semantic embedding features as the key and value vectors to obtain a texture-aware vibration recalibration vector. The vibration-aware texture aggregation vector and the texture-aware vibration recalibration vector are concatenated in the channel dimension to form a 64-dimensional cross-modal alignment and fusion vector.

7. The method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response as described in claim 1, characterized in that: The composite objective function includes quality regression loss and cross-modal consistency loss; The quality regression loss: for the predicted compaction quality Compaction quality with true value The regression error between them is adopted The loss is measured; The loss uses a quadratic penalty when the error magnitude is small to ensure regression accuracy, and a linear penalty when the error magnitude is large to reduce the impact of outlier errors on the gradient. The cross-modal consistency loss involves aligning the multi-scale texture visual features and vibration embedding features before fusion. Specifically, the multi-scale texture visual features and vibration embedding features are mapped to a unified feature space of the same dimension through a linear projection layer, and the mapped feature vectors are normalized. Then, the cosine similarity between the texture feature vector and the vibration feature vector of the same sample is calculated. A contrastive consistency loss is constructed with the goal of improving the cross-modal similarity of the same sample and reducing the cross-modal similarity of different samples, so that the cross-modal representation distance of the same sample in the unified feature space converges.

8. The method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response as described in claim 1, characterized in that: In step S4, the construction section is divided into continuous segments of 10 meters in length along the direction of travel as detection intervals, and each detection interval is assigned a unique interval number; continuous texture images and synchronous vibration features are collected for each detection interval, using a length of... Five windows are fixedly sampled within the detection interval using a texture window and a sliding method with a window overlap ratio of 50%. Model inference is performed once for each window to obtain the predicted compaction quality value of that window. The predicted values ​​of the five windows within the detection interval are combined into a prediction sequence, and the arithmetic mean of the prediction sequence is calculated as the representative value of the compaction quality of the detection interval. At the same time, the standard deviation of the prediction sequence is calculated as the prediction volatility of the detection interval. The predicted volatility is converted into a confidence score, which is defined as the reciprocal of the predicted volatility and cropped to the range of 0 to 1. The higher the confidence score, the more stable and reliable the prediction for that detection interval. When the confidence score is less than 0.60, the detection interval is marked as a low-confidence segment and a retest is triggered. The retest method is to add three fixed sampling windows within the same detection interval and recalculate the representative value of compaction quality and the confidence score. If the confidence score after the retest is still less than 0.60, the detection interval is determined as an uncertain segment and a manual review prompt is output. Risk classification is performed by combining the representative value of compaction quality with the confidence score using a fixed grading rule.

9. The method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response as described in claim 8, characterized in that: The specific method for risk classification using the fixed classification rules is as follows: When the representative value of compaction quality is not less than 0.80 and the confidence score is not less than 0.70, it is judged as a high-quality stable section; when the representative value of compaction quality is not less than 0.65 and less than 0.80 and the confidence score is not less than 0.70, it is judged as a medium-quality acceptable section. When the representative value of compaction quality is less than 0.65 and the confidence score is not less than 0.70, it is judged as a low-quality risk section; when the confidence score is less than 0.70, it is judged as an uncertain section and enters the retest process; finally, the section number, start and end mileage and spatial location list of low-quality risk section and uncertain section are output.

10. The method for detecting the compaction quality of asphalt pavement by integrating surface texture visual features and vibration response as described in claim 1, characterized in that: It also includes the following process: for the detection intervals determined to be low-quality risk sections and uncertain sections, an executable process adjustment suggestion is generated by combining the working condition parameter vector, the representative value of compaction quality and the confidence score of the detection interval, and a compaction quality test report for on-site closed-loop decision-making is output; the process adjustment suggestion generation process includes rule triggering, candidate scheme generation, similar sample retrieval and evaluation, optimal scheme selection and closed-loop output.

Images