Method for rapidly evaluating compaction quality of slag roadbed based on multi-source and multi-modal data

Abstract

The invention discloses a slag roadbed compaction quality rapid evaluation method based on multi-source and multi-modal data, and the method comprises the steps: laying a plurality of measurement points on a roadbed, and synchronously collecting the on-site high-pixel image, PFWD dynamic modulus, DCP penetration, void ratio and supervision label of each measurement point; preprocessing the field high-pixel image of the selected measuring point to form a point-level image feature vector of the measuring point; performing supervision and dimension reduction on the point-level image features through a partial least squares regression method, and generating low-dimension image feature representation rich in physical significance; fusing the image features and the numerical features of the same measuring point after dimension reduction through a superficial cross attention mechanism, and creating multi-modal fusion features; and training a machine learning model based on the multi-modal fusion features, and outputting a cloud picture of the FWD modulus predicted value of the slag roadbed. According to the method, the real compaction state of the high-discreteness slag filler with complex components can be quickly evaluated in a high-precision and non-destructive manner.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer-aided subgrade design and relates to a method for rapidly evaluating the compaction quality of a subgrade filled with slag based on multi-source multi-modal data. BACKGROUND

[0002] With the rapid development of expressways, especially high-fill subgrade projects in mountainous areas, slag-based soil and rock mixed fillers are widely used due to their rich material sources and good economy. However, such materials have characteristics such as discrete particle gradation, high stone content, non-uniform local structure, and complex compaction mechanism, which make it difficult for traditional compaction quality evaluation methods to accurately reflect their real engineering mechanical properties. The compaction degree detection (ring knife method, sand replacement method) commonly used in engineering is limited by poor sampling representativeness, strong disturbance, low testing efficiency and other problems; while the FWD (falling weight deflectometer) structural modulus evaluation method used for construction acceptance can better reflect the overall mechanical properties, but the device is large in size, high in cost and poor in mobility, which makes it difficult to meet the needs of continuous and rapid detection and real-time quality feedback on site. Therefore, how to use portable, rapid and non-destructive data collection means to reliably evaluate the compaction quality of the slag subgrade is an urgent problem to be solved in current engineering practice.

[0003] In order to overcome the limitations of insufficient information from a single detection means, in recent years, quality evaluation research based on portable non-destructive equipment such as PFWD (light falling weight deflectometer) modulus, DCP (dynamic cone penetration instrument) penetration, and void ratio of non-nuclear density meter has gradually emerged; at the same time, high-pixel field images have also been introduced to reflect the surface structure characteristics of the fillers. However, the existing technologies generally have the following shortcomings: there are large differences in measurement mechanisms between different instruments, and there is a significant modal gap between the data; there is no direct correspondence between image information and geophysical measurement values in the physical space, making it difficult to achieve effective fusion; the deep learning method is prone to overfitting in small sample engineering scenarios, and the model stability is insufficient; in addition, there are noise points and abnormal points in data collection, which will significantly reduce the prediction accuracy and engineering applicability if not distinguished before modeling.

[0004] Therefore, there is an urgent need for a rapid compaction quality evaluation method that can fuse multi-source multi-modal data, fully utilize the advantages of image and non-destructive instrument data, and have the ability of abnormal value identification and robust modeling, so as to more accurately reflect the real state of the slag subgrade structure and meet the needs of field engineering applications. SUMMARY

[0005] To solve the above problems, the application provides a method for rapidly evaluating the compaction quality of a subgrade filled with slag based on multi-source multi-modal data, which realizes high-precision and non-destructive rapid evaluation of the real compaction state of complex composition and high-dispersity slag fillers.

[0006] The technical solution adopted in this invention is a rapid assessment method for the compaction quality of slag roadbed based on multi-source multimodal data, comprising the following steps: S1: Multiple measuring points are set up on the roadbed, and high-resolution images, PFWD dynamic modulus, DCP penetration, porosity, and supervision labels of each measuring point are collected simultaneously, including the plane coordinates and measuring point number of the measuring point, for subsequent drawing of threshold cloud maps and spatial interpolation analysis; the collected data are cleaned and filtered. S2: Preprocess the high-resolution images of the selected measurement points, extract multi-scale image features through a deep learning model, average all image features from the same measurement point, and form the point-level image feature vector of the measurement point. S3: Supervised dimensionality reduction of point-level image features is performed using partial least squares regression to generate low-dimensional but physically meaningful image feature representations. S4: Multimodal fusion features are created by fusing image features and numerical features after dimensionality reduction at the same measurement point through a shallow cross-attention mechanism; S5: Train a machine learning model based on multimodal fusion features and output a cloud map of the predicted FWD modulus of the slag roadbed.

[0007] Furthermore, in S1, the supervision label is the FWD modulus. and Choose one of the two.

[0008] Furthermore, S2 includes the following steps: S21: For each original image Perform basic preprocessing; S22: For each preprocessed image at each measurement point Scale the image according to the multi-scale scaling ratio to obtain multi-scale images at different viewing distances; then scale the image at each scale... The pre-trained convolutional neural network backbone is input sequentially to obtain the corresponding 2048-dimensional depth feature vector; then the feature vectors of all scales of the image are averaged element by element.

[0009] Furthermore, S3 includes the following steps: S31: Stack the 2048-dimensional image features of all measurement points according to the sample dimensions to obtain a matrix composed of point-level image features; S32: A partial least squares regression method is used to establish a mapping, which linearly projects the 2048-dimensional image features onto a fixed low-dimensional latent space. S33: Obtain the low-dimensional feature vector of the image for each measurement point.

[0010] Furthermore, S4 includes the following steps: S41, The three-dimensional table features of each measuring point. repeating along the dimension, and truncating to the same dimension as the low-dimensional eigenvector in S2, to obtain an extended table vector; wherein, , , ; Ei represents the local dynamic modulus of resilience measured by a light falling hammer deflection meter at the measurement point , Ei represents the penetration index measured by a dynamic cone penetration instrument at the measurement point ; Ei represents the void ratio measured by a nuclear-free density meter at the measurement point ; S42, calculate the dot product similarity of the image low-dimensional feature and the extended table feature for each sample, and perform scale normalization; S43, truncate the dot product similarity to a fixed interval, and then map it to the interval (0, 1) through the Sigmoid function to obtain a gating coefficient; thereby realizing dynamic adjustment of the contribution of different modalities when multi-source multi-modal feature fusion is performed; S44, scale the extended table feature with the gating coefficient to obtain an attention weighted feature vector, S45, splice the image low-dimensional feature, the attention weighted feature vector and the original three-dimensional table feature to construct a multi-modal fusion feature.

[0011] Further, in S41, the first value in the extended table vector of the measurement point is :

[0012] represents that one is selected in turn according to the position of in the three table features of the measurement point .

[0013] Further, in S42, the gating coefficient is determined according to the following formula:

[0014] wherein, represents the Sigmoid function, represents the natural constant, is the result of truncating the cross-modal similarity score of the measurement point .

[0015] Further, S5 includes the following steps: S51: Group the multi-modal fusion features of all samples created in S4 into a matrix and perform standardization processing; train a Ridge regularized linear regression model and a lightweight multilayer perception model in each fold training set respectively, construct out-of-fold prediction results, and select the model output with good overall prediction performance; S52: Perform spatial interpolation based on the out-of-fold prediction results to draw a FWD modulus prediction value cloud map of the subgrade; S53: Perform anomaly point identification based on the residual of the out-of-fold prediction results, output the set of abnormal measurement points as a retest list, and perform labeling or rejection.

[0016] Further, in S51, the sample set is divided using GroupKFold with the measurement point number as the grouping identifier.

[0017] Further, in S52, a preset qualified threshold is selected, and an isogram with a prediction value equal to the preset qualified threshold is drawn as a qualified line.

[0018] The beneficial effects of the present application are: (1) The present application uniformly fuses high-pixel images with PFWD, DCP, void ratio and other non-destructive detection data, breaking through the limitation of traditional single detection quantity that cannot reflect the non-uniformity of particle morphology and structure.

[0019] (2) The present application constructs an image feature extraction integrated with CLAHE, gray three channels, multi-scale enhancement and ResNet50, realizes robust identification of key structural information such as illumination difference, particle scale span and local segregation, and greatly improves the engineering adaptability of the image mode.

[0020] (3) The present application uses PLS to compress 2048-dimensional image features to a low-dimensional latent space strongly related to FWD; at the same time, the cross-attention mechanism is used to adaptively weight the image information to the table features, improving the cross-modal information collaborative modeling capability. Thus, a multi-modal fusion feature with strong discriminability, low dimension and good interpretability is formed.

[0021] (4) The present application adopts a cross-validation evaluation system based on point grouping, which significantly improves the model generalization reliability. Avoids the leakage of information of the same measurement point, making the evaluation index more real and reliable.

[0022] (5) The present application constructs a threshold cloud map visualization method to realize spatial identification of weak compaction areas. Through triangular net interpolation and segmented color scale, the continuous distribution of modulus and weak areas are clearly displayed. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only represent some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0024] Figure 1 The flowchart of the prediction method of the embodiments of the present application.

[0025] Figure 2 The flowchart of the multi-source data acquisition and sample construction of the subgrade of the embodiments of the present application.

[0026] Figure 3 The structural block diagram of the image preprocessing and multi-scale point-level feature extraction of the embodiments of the present application.

[0027] Figure 4 The structural block diagram of the image feature PLS supervised dimension reduction and shallow cross-attention multi-modal feature fusion of the embodiments of the present application.

[0028] Figure 5 The comparative effect diagram of the prediction accuracy of the multi-modal evaluation method and the traditional single-modal evaluation method of the embodiments of the present application.

[0029] Figure 6 The prediction effect diagram of the multi-source multi-modal two-class model of the embodiments of the present application.

[0030] Figure 7 The model prediction threshold cloud chart of the embodiments of the present application. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort belong to the scope of protection of the present application.

[0032] A subgrade compaction quality rapid evaluation method based on multi-source multi-modal data, as shown in Figure 1 , includes the following steps: S1: data acquisition, data cleaning and screening.

[0033] Step 1: multi-source data acquisition and sample construction, as shown in Figure 2 .

[0034] Step 1.1: In the construction section of the dam slag subgrade to be evaluated, a number of point numbers (point_id) are arranged according to the preset detection interval, and the following data are synchronously collected for each point: On-site high-pixel image , record the roadbed surface texture, particle size distribution and compaction marks; Dynamic modulus measured by light falling weight deflectometer (PFWD) ; Penetration measured by dynamic cone penetration instrument (DCP) ; Void ratio measured by non-nuclear density meter ; FWD modulus as a supervision label or solid volume fraction SVF; Coordinates of the point in the plane .

[0035] Step 1.2: Establish a multi-source data record for each point.

[0036] (1) wherein, represents a multi-source multi-modal sample record corresponding to the thpoint, which is used for subsequent modeling and evaluation; represents the number of the thpoint; represents the original on-site high-pixel digital image collected at the point , wherein the subscript raw represents the original image; represents the local dynamic modulus of resilience measured by light falling weight deflectometer (PFWD) at the point , reflecting the overall stiffness response of the shallow layer of the roadbed; represents the penetration index measured by dynamic cone penetration instrument (DCP) at the point , with the unit of mm / beat; reflecting the anti-penetration capacity of the roadbed material; represents the void ratio measured by non-nuclear density meter at the point ; represents the supervision label (output index) of the point , represents the structural equivalent modulus obtained by back calculation by falling weight deflectometer (PWD) at the point ; represents the solid volume fraction at the point , which can be used as a label in actual use, ; represents the point Coordinate position in construction plane coordinate system (unit: m); used for subsequent spatial interpolation, threshold cloud map drawing and weak area positioning.

[0037] is the state label of material structure compactness, is the performance label of structure bearing stiffness, the former emphasizes more on compaction mechanism and pore structure, and the latter emphasizes more on bearing and deformation response; with as the label, more emphasis is placed on compaction state / compactness evaluation; with as the label, more emphasis is placed on predicting acceptance performance and realizing regional weak area positioning. In Zhejiang area, gangue is often used as roadbed filler, and solid volume rate and are used as construction quality control indexes in the field construction project book, wherein is used as a construction acceptance index, so the embodiment of the present application selects as the main label to be put into the model for prediction.

[0038] Step 1.3: Aggregate to build the original data table data.csv (image-point level detection data association file) and coordinate table coords.csv (measurement point spatial coordinate file).

[0039] Build data.csv: the embodiment of the present application automatically scans the image root directory through a computer terminal, obtains the complete path of each image, and automatically identifies the corresponding measurement point number (point_id) according to the name of the upper folder. Then, the measurement point number is automatically matched and associated with the point level data (such as , , , recorded in the field detection table to form a structured data.csv file. Each record of data.csv corresponds to an image and contains: point_id, image_path (image file path), , , , fields, which provide unified and complete input data for the multi-modal model.

[0040] Build coords.csv: in order to realize spatial visualization and continuous expression of subsequent prediction results, the embodiment of the present application extracts and arranges the measurement point plane coordinates recorded in the field Excel data according to the measurement point number to generate the coords.csv file. Each record in the file corresponds to a measurement point and contains: point_id, , , which is used for subsequent drawing of threshold cloud map and spatial interpolation analysis.

[0041] Step 2: Data cleaning and measurement point screening.

[0042] Step 2.1: Integrity check on records in data.csv: if a record is missing an image path or the corresponding image file does not exist, immediately check the record to ensure data integrity.

[0043] Step 2.2: Indexing by point number (point_id), count all complete data sets P of measurement points that still have images, and only model the measurement points in set P to ensure that each sample has both numerical and image modalities.

[0044] S2: Preprocess the on-site high-pixel images of the selected measurement points, extract multi-scale image features through a deep learning model, average all image features from the same measurement point, and form a point-level image feature vector for the measurement point.

[0045] S21: Image preprocessing, as shown in Figure 3 ; perform basic preprocessing on each raw image .

[0046] Brightness equalization: convert the image from RGB space to LAB space, apply contrast-limited adaptive histogram equalization (CLAHE) to the L channel, and obtain the brightness-equalized image to weaken the influence of different shooting light conditions.

[0047] Gray three-channel transformation: convert to a single-channel grayscale image , and duplicate it three times to form a grayscale three-channel image to weaken color information interference, help emphasize texture / structure features, enhance local contrast, highlight particle and pore texture, and be suitable for cases where the subgrade filler uniformity is poor; at the same time, maintain the same number of channels as the ImageNet pre-trained network.

[0048] Size uniformity: center crop and scale to 224x224 pixels to obtain the preprocessed image .

[0049] S22: Multi-scale image feature extraction, as shown in Figure 3 .

[0050] S221: Set a multi-scale scaling ratio set : (2) S222: For each preprocessed image of each measurement point​ proportional Scaling is performed to obtain multi-scale images at different viewing distances.

[0051] S223: Image at each scale By sequentially inputting the pre-trained convolutional neural network backbone (ResNet50 with fully connected layers removed), the corresponding 2048-dimensional depth feature vectors are obtained: ; For all real numbers; The space consisting of all 2048-dimensional real vectors; The 2048-dimensional deep feature vector extracted from the backbone of ResNet50.

[0052] S224: Calculate the element-wise average of the feature vectors at all scales of the image to obtain the scale-robust features of the image. : (3) S23: Point-level image feature fusion.

[0053] S231: For multiple images acquired from the same measurement point (point_id), a set of feature vectors is obtained by following steps 3 and 4 respectively. ; k is the number of the k-th image at the same measurement point; S233: The point-level image feature vector is obtained by averaging all image features at the same measurement point according to the sample dimension. : (4) S233: Simultaneously record the three-dimensional tabular features of this measuring point. : (5) Simultaneously splice the compaction quality evaluation index data This constitutes a point-level sample triplet. That is, the point-level feature set.

[0054] S3: PLS (Partial Least Squares Regression) supervised dimensionality reduction of image features, such as Figure 4 As shown; S31: Stack the 2048-dimensional image features of all measurement points according to the sample dimensions to obtain a matrix composed of point-level image features. : (6) make Then the labels form a vector. Used to guide PLS extraction and Supervised dimensionality reduction is achieved by using the most relevant latent image features. This represents the total number of measurement points. S32: Partial Least Squares Regression (PLS Regression) is used to establish a mapping, linearly projecting the 2048-dimensional image features onto a 10-dimensional latent space: (7) in, , and By PLS The training yields a 10-dimensional latent feature explanation. The variation.

[0055] S33: Obtain a 10-dimensional low-dimensional feature vector for each measurement point: (8) Under the application constraints of S3 for evaluating the compaction quality of slag filler (fluctuations in on-site lighting, strong image heterogeneity, limited sample size, and the need for rapid deployment), low-dimensional latent variables are used to constrain the high-dimensional image features to improve the stability and generalization ability of subsequent multi-source feature fusion and prediction, and significantly reduce computational complexity. This provides usable and stable input representations for subsequent steps (such as fusion gating / spatial interpolation and weak area localization).

[0056] To balance generalization and efficiency, a smaller dimension within the platform region is preferred, and in this embodiment, it is ultimately reduced to 10 dimensions. In practical engineering, the dimension can be determined based on the performance indicators of the validation set or preset engineering constraints. For example, validation evaluation can be performed within a candidate range, and a dimension that minimizes error and reduces complexity can be selected.

[0057] S4: Shallow cross-attention multimodal feature construction. To alleviate the modal gap between image and table modalities, a shallow cross-attention gating is introduced to adaptively adjust table feature weights with cross-modal consistency without introducing additional trainable parameters.

[0058] S41: Dimensional Expansion of Table Features: Expanding the 3D Table Features of Each Measurement Point Repeatedly tile along the dimensions and truncate to 10 dimensions to obtain the extended table vector. , can be represented as: (9) Therefore ;in .

[0059] : Indicates the measuring point Extended table vector The a numerical value. The extended table vector is a 10-dimensional vector obtained by repeating and truncating the three table features (PFWD modulus, DCP penetration index, and void ratio) of the measurement point in order, so the value of must come from one of the above three features.

[0060] represents the three table features of the measurement point according to the position of . That is, every increase of 1 takes the value in a cycle between When the remainder of is 1, take (PFWD modulus); when the remainder is 2, take (DCP penetration index); and when the remainder is 3, take (void ratio).

[0061] S42: Similarity calculation and scale normalization: Calculate the dot product similarity between the image low-dimensional feature vector and the extended table feature vector for each sample, and divide by to perform scale normalization: (10) represents the cross-modal similarity score of the measurement point , which measures the consistency (correlation) between the image low-dimensional feature vector and the extended table feature vector . When is larger, it means that the image information of the measurement point is more consistent with the compaction state reflected by the table information such as PFWD / DCP / void ratio; when is smaller or negative, it means that the consistency is weak or even different. This similarity score will be used as input for subsequent gating coefficient calculation, to adaptively adjust the weight of table features in the fused features.

[0062] S43: Stable Sigmoid gating coefficient: truncate to the interval to obtain to suppress the amplification effect of abnormal similarity on the gate and improve stability under field noise; then map it to the interval (0, 1) through the Sigmoid function to obtain the gating coefficient : (11) ​​In the formula, represents a Sigmoid function (S-shaped logic function), and its standard expression is , is an independent variable; represents a natural constant; is the result of truncation of the cross-modal similarity score of the measuring point ; makes adaptively generated from the cross-modal consistency score and stably falls into the (0, 1) interval, so as to realize dynamic adjustment of the contribution of different modalities in multi-source multi-modal feature fusion: when the image features and table features of the measuring point are highly consistent, a larger value is taken to enhance the fusion contribution of effective information; when the consistency is low or there is a modal conflict, a smaller value is taken to suppress the interference of unreliable modalities on the fusion result. The truncation processing is performed on before formula (11), so that the extreme similarity score will not be amplified and passed to the fusion weight by factors such as abnormal light, shielding, particle size distribution dispersion, instrument fluctuation or local construction disturbance, avoiding excessive bias or violent shock of the fusion weight. Thus, the robustness and generalization stability of the fusion output can be improved under the conditions of field noise and small sample, and the gating weight has interpretability, which is convenient for subsequent threshold partitioning, weak area identification and visual expression.

[0063] S44: Attention feature generation: scaling the extended table features to obtain the attention weighted feature vector : (12) S45: Fusion feature splicing: splicing the image low-dimensional features , the attention weighted feature vector and the original three-dimensional table features to construct a 23-dimensional multi-modal fusion feature : (13) Under the constraint of small sample engineering scene, a shallow cross-attention gating without additional trainable parameters is proposed: the gating coefficient is constructed by the cross-modal similarity of image low-dimensional features and table features, and truncation and Sigmoid are used to realize stable mapping, so as to adaptively adjust the weight of PFWD / DCP / void ratio and other table features in the fusion vector; at the same time, the original physical quantity is retained to enhance the interpretability. The combined method effectively alleviates the modal gap, suppresses the noise modal interference and improves the prediction stability and engineering applicability under the condition of small sample.

[0064] ​In on-site testing of slag filler, images are affected by illumination and texture separation, while PFWD / DCP is affected by loading depth and local non-uniformity. Inconsistencies or conflicts may occur between the two modes at different measurement points. If the images are directly stitched together, the subsequent regressor needs to learn how to handle conflicts under small sample conditions, which can easily introduce noise and lead to unstable generalization. Conventional weighted averages usually use fixed weights or trainable weights. Fixed weights cannot adaptively adjust the modal reliability for different measurement points; trainable weights introduce additional parameters and degrees of freedom, making them more prone to overfitting in small engineering sample scenarios, and their cross-site transferability and interpretability are weak. The shallow cross-attention (gating) method employed in this invention uses cross-modal consistency (i.e., consistency between image and table features) as the gating signal. By calculating the similarity between low-dimensional image features and table features and obtaining the gating coefficients through stable mapping, the contribution of table features is adaptively adjusted point by point. When the consistency is high, the effective weight of table features in the fusion vector is increased, enhancing effective information; when the consistency is low, its influence is reduced, thereby suppressing the interference of modal conflicts and noisy modalities on prediction, improving stability and generalization under small sample conditions; suppressing the influence of conflicting or noisy modalities, thereby improving the robustness and generalization ability of the fused features; at the same time, this gating is shallow and does not require the introduction of additional trainable parameters, resulting in low computational cost, meeting the needs of rapid on-site deployment, and preserving the original physical quantities for engineering interpretation and traceability.

[0065] S5: Train a machine learning model based on multimodal fusion features and output a cloud map of the compaction quality of the slag roadbed.

[0066] S51: Modeling and Cross-Validation Evaluation.

[0067] S511: The fusion feature of all samples (i.e., the 23-dimensional multimodal fusion feature constructed in S4) ) form a matrix Standardize them to eliminate the influence of different physical quantities in terms of dimensions and magnitude: (1) Calculate the first Mean and standard deviation of the dimensional features: (14) (15) Indicates the first The first measurement point (sample) One eigenvalue; It means arranging the 23 features of each measurement point into a row, and then... Line number The number in the column.

[0068] (2) Standardization processing: Transform each sample and each feature according to the following formula: (16) (3) The final standardized matrix for: (17) To represent the number of measurement points, the 23-dimensional fused feature is composed of PLS-reduced image features (10-dimensional), attention-weighted table features (10-dimensional), and original table features (3-dimensional). The corresponding label vector... for: (18) in, This is the FWD modulus of the first measurement point.

[0069] S512: GroupKFold partitioning based on point_id. Using the measurement point number (point_id) as the group identifier, GroupKFold is used to partition the sample set into S=5 folds: (19) in, Let S be the set of measurement points contained in the S-th fold. The partitioning principle ensures that data from the same measurement point will not appear in both the training and validation sets simultaneously, thereby avoiding optimistic bias caused by information leakage from the same point.

[0070] This represents the entire sample dataset; This represents the s-th fold sample subset obtained by dividing the data using GroupKFold (grouped by measurement point number point_id), where s = 1, 2, ..., S (S = 5 in this embodiment).

[0071] Indicates the first The input feature vector of each sample, i.e., the feature vector obtained by multimodal fusion (e.g., 23-dimensional), corresponds to the feature matrix. The OK.

[0072] S513: Two types of prediction models under cross-validation; In each training set, train the following two types of models respectively: (1) Ridge regularized linear regression model: in linear regression Based on this, an L2 regularization term is introduced, and the objective function is: (20) in, The regularization coefficient is calculated using RidgeCV on a logarithmic scale grid. The optimal regularization coefficient is automatically selected by cross-validation The optimal regularization coefficient of the Ridge model is automatically selected by cross-validation .

[0073] The parameter vector (regression coefficient / feature weight) of the Ridge regression model, used to represent the contribution of each input feature to the predicted label; The true label vector of the training set samples (the corresponding ). The input feature matrix of the training set samples (obtained by stacking the fusion feature vectors of all samples in the training set in rows, and the normalized feature matrix if necessary).

[0074] (2) Lightweight multilayer perceptron (MLP) model: The feedforward neural network is constructed , the ReLU activation function and L2 regularization term are used, the learning rate is set to , and the maximum number of iterations is 30000. The candidate hidden layer structure is (64, 32), (64), (32), and the optimal hidden layer structure hidden* is selected by comparing the coefficient of determination R2 on the validation set in each fold, and the optimal network structure hidden* obtained in this test is (64); S514: Construct out-of-fold prediction (OOF) results; In the Sth fold, the remaining 4-fold samples are used as the training set to train Ridge and MLP, and the prediction of the Sth fold validation set samples is obtained: . The prediction results of each fold validation set are backfilled according to the sample number to obtain the out-of-fold prediction vector for all samples:

[0075] The out-of-fold prediction (OOF) result vector of the Ridge model for all samples is represented by (21) The out-of-fold prediction (OOF) result vector of the Ridge model for all samples is represented by The out-of-fold prediction value of the Ridge model for the Sth sample in its corresponding validation fold is represented by . The out-of-fold prediction result vector of the MLP model for all samples is represented by . The out-of-fold prediction value of the MLP model for the Sth sample is represented by

[0076] The out-of-fold prediction refers to the prediction result obtained when the sample does not participate in training of the corresponding model, and is used to objectively evaluate the generalization performance of the model and avoid information leakage.

[0077] The OOF result triplet of the first round of modeling is finally formed: (22) and exported as oof_PLS10_plus_tab_mm79.csv, which provides basic data for subsequent outlier identification and nephogram drawing.

[0078] S515: evaluation index and first round of modeling results; with the true label and the out-of-fold prediction value as the basis, the determination coefficient , root mean square error (RMSE) and mean absolute error (MAE) of Ridge and MLP are calculated respectively: (23) (24) , the average value of the true label on all samples; , the prediction value of the MLP model for the i-th sample (out-of-fold prediction value); , the prediction value of the Ridge model for the i-th sample (out-of-fold prediction value); (1) Ridge model performance:

[0079] , the determination coefficient of the Ridge model, , the root mean square error of the Ridge model, , the mean absolute error of the Ridge model.

[0080] The target modulus value of the embodiment is generally about 100Mpa, and the average absolute error is about 4.7%, which indicates that the Ridge model has high fitting accuracy under the current multi-modal modeling framework.

[0081] (2) MLP model performance:

[0082] Due to factors such as sample size (only 100 points) and large nonlinear and parameter quantity of the MLP model, the overall prediction performance ​​The performance (higher and lower RMSE / MAE) is significantly weaker than the Ridge model, as shown in Figure 6 , so the output of the Ridge model is selected as the final output .

[0083] S52: Threshold cloud map generation and weak area identification.

[0084] S521: Associate the first round of OOF files with the coordinate table coords.csv by point_id, get the coordinates and the corresponding predicted value for each test point. Based on the out-of-fold prediction (OOF) results rather than the in-training fitting values, spatial interpolation and cloud mapping are performed to avoid cloud high accuracy and improve the authenticity of weak area positioning.

[0085] S522: Based on the plane coordinates of all test points, an irregular triangular subdivision grid is constructed, and the predicted value is interpolated to obtain the continuous spatial distribution of the predicted index in the detection area.

[0086] S523: Draw a filled cloud map of the predicted value (FWD modulus) on the plane, as shown in Figure 7 , and select a preset qualified threshold T=100Mpa (modulus threshold). Draw the contour line of the predicted value equal to T as the qualified line, so that the prediction result is directly mapped into the executable treatment boundary of the qualified area or weak area.

[0087] S524: According to the color distribution and contour line position of the cloud map, the weak area where the predicted value is lower than the threshold can be directly identified. Measures can be directly taken to improve the weak area.

[0088] S53: Abnormal point detection based on residual error.

[0089] S531: Select the OOF predicted value of the model with stable performance in the first round of modeling (e.g. Ridge model) to calculate the residual error of each sample :

[0090] (25) S532: Calculate the mean and overall standard deviation of the residual error sequence based on the out-of-fold prediction; when , it is considered that the model has been almost perfectly fitted, and no abnormal point identification is performed. If abnormal point identification is performed based on the in-training residual error, it is easy to underestimate the in-training residual error and miss abnormal test points.

[0091] S533: When , the standardized residual of each sample is calculated : (26) S534: Set the outlier threshold , if , it is considered that the measuring point has significant deviation behavior, and it is marked as an abnormal point. The number of abnormal points for the Tanjia filler is 9, and the abnormal points are marked.

[0092] S535: Record the point_id of all abnormal measuring points as an abnormal set , the abnormal points can be detected. The abnormal measuring point set is output as a retest list, and is labeled or removed in cloud analysis, reducing the misleading risk of noise points on weak area determination and engineering decision.

[0093] S536: In practical application, new measuring point data only needs to perform steps S1 to S4 on the model, and the created multi-modal fusion features are input into the trained model, so that the predicted FWD modulus or deflection value of the measuring point can be output through the final model, and the threshold cloud method is used to realize the rapid and visual evaluation of the Tanjia subgrade compaction quality.

[0094] The embodiment of the present application uses PFWD dynamic modulus, DCP penetration, core-free density void ratio and high-pixel image as main input data in the field. The above-mentioned devices are portable and flexible in measuring point layout, and can realize rapid acquisition of continuous points. FWD modulus is only used as a supervision label for a small amount of calibration / training, and heavy FWD detection does not need to be carried out densely on the whole line, thereby significantly reducing the organization cost and time consumption of field detection.

[0095] The embodiment of the present application uses a pre-trained convolutional neural network (such as ResNet50 backbone) to extract image features, without training a complex deep network on a small sample of engineering; further linearly supervised dimension reduction is performed by PLS to compress 2048-dimensional image features into low-dimensional latent variables; at the same time, a shallow cross-attention gate is used to adaptively weight the table features without introducing additional trainable parameters, and the calculation amount is low; finally, the Ridge or lightweight MLP is used as the regression model to realize fast inference and output the predicted index.

[0096] The embodiment of the present application automatically associates the prediction result with the measuring point coordinates, performs spatial interpolation based on out-of-fold prediction (OOF) and generates a threshold cloud map, and directly gives the range of weak areas below the threshold; at the same time, the abnormal points are automatically identified based on the Z-score of the OOF residual, and a retest / retest list is formed, which reduces the manual screening time, realizes the rapid closed loop of prediction, positioning and treatment, and realizes rapid evaluation.

[0097] The embodiment of the present application finds the image direction most related to FWD through PLS, and then uses the direction to non-parametrically adjust the weights of PFWD / DCP / void ratio. The low-dimensional features of the image compressed by PLS are used as a gating signal to dynamically adjust the effective weights of PFWD / DCP / void ratio, which not only reflects the physical correlation between modalities, but also does not introduce a large number of parameters.

[0098] The large particle size distribution span, discrete skeleton structure and complex pore structure of the slag filling material make the slag subgrade have the characteristics of high heterogeneity and small sample, which leads to obvious differences in stiffness response of different measuring points under the same compaction process, and it is difficult to accurately represent the overall compaction state by a single measuring point or a single index. The model is easy to be distorted under the condition of small sample, and the generalization ability is poor; the heavy verification test (such as FWD) that can be carried out on the engineering site is limited in point, and the multi-source data collection is constrained by the construction period and environment, resulting in less labeled samples, large noise and unstable distribution, so that the traditional experience regression or complex model is easy to overfit / underfit and unstable. The dimensions / sensitive depths of different detection methods are different, and there are inconsistencies and conflicts between the indexes; PFWD, DCP, void ratio and image information have different sensitive depths and sensitive objects for the structure response of the subgrade: PFWD is more inclined to the overall stiffness response of the shallow layer; DCP is more inclined to the local penetration resistance; void ratio is greatly affected by water content, gradation and instrument stability; and image reflects the apparent structure and particle distribution. Therefore, on the heterogeneous material, the conflict phenomenon that one index becomes better and another index becomes worse often occurs, making it difficult for the traditional single-index threshold method or simple weighting method to be stable and applicable. Although the single-modality method (only PFWD or only DCP) is efficient, it has a high risk of misjudgment for complex slag materials; and although increasing the verification method such as FWD is accurate, it is high in cost and low in efficiency, which is not conducive to the rapid closed-loop control on site.

[0099] The embodiment of the present application uniformly correlates the image information of each measuring point with the point-level detection data such as PFWD, DCP and void ratio to form a structured sample, so that the multi-source information of the same measuring point can be checked and supplemented with each other to reduce the misjudgment of a single index; the image features and table features are fused, and the cross-modality consistency measurement / weight adjustment is introduced, so that when facing a modality with large noise or distortion, the model can reduce the interference of the modality on the result, thereby improving the prediction stability. Cross-validation is carried out by taking the measuring point number as the grouping identifier to ensure that the data of the same measuring point cannot be simultaneously used for training and verification, and to reduce the optimistic bias; and the reliability under the condition of small sample is improved by objectively evaluating and iterating the model through the out-of-fold prediction results. The model residual is statistically analyzed and the abnormal samples are identified to reduce the influence of extreme points on the model parameters and evaluation conclusion, so that the model output is more in line with the stability requirements of engineering application.

[0100] The embodiment of the present application is suitable for rapid evaluation of the compaction quality of the soil and stone mixed filler subgrade (or similar coarse-grained non-homogeneous filling material). It is suitable for construction process quality control, pre-acceptance rapid survey and weak area pre-screening, encryption review guidance under limited heavy detection resources, and regional quality partitioning and disposal decision of non-homogeneous risk sections such as high fill or mountainous road sections. The output results include predicted indicators of measurement points, threshold cloud map, weak area range and abnormal measurement point list, which can be used to guide re-rolling, process adjustment and re-measurement.

[0101] Table 1 Comparison of the embodiment of the present application and the existing fusion strategy

[0102] The OOF prediction evaluation is performed by 5-fold cross-validation according to the grouping of measurement points, and the three feature construction strategies are compared: only three-dimensional table features, direct splicing, and the gating fusion of the embodiment of the present application, and the verification is performed under two types of models of Ridge and MLP. As shown in Table 1, compared with only table, the multi-modal fusion significantly improves the prediction accuracy (taking Ridge as an example, from 0.382 to about 0.815). At the same time, under the Ridge model, the accuracy of direct splicing and gating fusion is close (0.808 vs 0.815), indicating that the linear regular model has a certain inhibitory capacity for redundant features. More importantly, under the MLP model, direct splicing has obvious performance degradation ( is negative and the error significantly increases), while the gating fusion of the embodiment of the present application still maintains the positive prediction ability ( , and the RMSE is significantly reduced), indicating that the gating mechanism can suppress the feature amplification effect caused by abnormal similarity by limiting the gating coefficient to (0, 1) and adaptively scaling the table repeated features, thereby improving the stability and engineering usability of nonlinear model training and inference. In Table 1, only three-dimensional table features: input data ( ), output data ( ); direct splicing: input data ( ), output data ( ), without gating fusion. The embodiment of the present application (gating fusion): input data ( ), output data ( ), with gating fusion.

[0103] To verify the effectiveness of the multi-source and multi-modal evaluation method of the present application, the method of the embodiment of the present application is compared with the traditional single-modal baseline method. The traditional single-modal method only takes the PFWD detection index as the input, and predicts the FWD modulus through the regression model; the method of the embodiment of the present application takes the PFWD, DCP penetration index, void ratio (pore index) and high-pixel image features as the joint input, and predicts the FWD modulus through the multi-modal feature fusion and regression model.

[0104] For both methods, GroupKFold based on point_id is used to construct out-of-fold (OOF) results, and the coefficient of determination R2, root mean square error (RMSE) and mean absolute error (MAE) are calculated respectively. Further, scatter plots of predicted values and true values are drawn, and ideal reference lines are given, as shown in Figure 5

[0105] As can be seen from Figure 5 , compared with the traditional single-mode method, the scatter distribution of the multi-modal method of the embodiment of the present application is closer to the ideal line and has smaller dispersion, indicating that the present application can significantly improve the prediction accuracy of FWD modulus and enhance the prediction stability, thereby improving the reliability and engineering usability of the rapid evaluation of subgrade compaction quality.

[0106] The above merely describes preferred embodiments of the present application but is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.​​​​

Claims

1. A rapid assessment method for the compaction quality of slag-filled roadbed based on multi-source, multi-modal data, characterized in that, Includes the following steps: S1: Set up multiple measuring points on the roadbed and simultaneously collect high-resolution images, PFWD dynamic modulus, DCP penetration, porosity, and monitoring labels for each measuring point; clean and filter the collected data. S2: Preprocess the high-resolution images of the selected measurement points, extract multi-scale image features through a deep learning model, average all image features from the same measurement point, and form the point-level image feature vector of the measurement point. S3: Supervised dimensionality reduction of point-level image features is performed using partial least squares regression to generate low-dimensional but physically meaningful image feature representations. S4: Multimodal fusion features are created by fusing image features and numerical features after dimensionality reduction at the same measurement point through a shallow cross-attention mechanism; S5: Train a machine learning model based on the multimodal fusion features and output a cloud map of the predicted FWD modulus of the slag roadbed.

2. The method for rapid evaluation of compaction quality of slag roadbed based on multi-source multimodal data as described in claim 1, characterized in that, In S1, the supervision label is the FWD modulus. and Choose one of the two.

3. The method for rapid evaluation of compaction quality of slag roadbed based on multi-source multimodal data as described in claim 1, characterized in that, S2 includes the following steps: S21: For each original image Perform basic preprocessing; S22: For each preprocessed image at each measurement point Scale the image according to the multi-scale scaling ratio to obtain multi-scale images at different viewing distances; then scale the image at each scale... The pre-trained convolutional neural network backbone is input sequentially to obtain the corresponding 2048-dimensional depth feature vector; then the feature vectors of all scales of the image are averaged element by element.

4. The method for rapid evaluation of compaction quality of slag roadbed based on multi-source multimodal data according to claim 3, characterized in that, S3 includes the following steps: S31: Stack the 2048-dimensional image features of all measurement points according to the sample dimensions to obtain a matrix composed of point-level image features; S32: A partial least squares regression method is used to establish a mapping, which linearly projects the 2048-dimensional image features onto a fixed low-dimensional latent space. S33: Obtain the low-dimensional feature vector of the image for each measurement point.

5. The method for rapid evaluation of compaction quality of slag roadbed based on multi-source multimodal data according to claim 4, characterized in that, S4 includes the following steps: S41, The three-dimensional table features of each measuring point. Repeatedly tile along the dimension and truncate to the same dimension as the low-dimensional feature vector in S2 to obtain the extended table vector; where, , , ; Indicates the measuring point The local dynamic resilient modulus measured by a light falling weight deflectometer. Indicates the measuring point The penetration index was measured by a dynamic cone penetrator. Indicates the measuring point The porosity was measured by a nucleus-free density meter. S42, calculate the dot product similarity between the low-dimensional features of the image and the extended table features for each sample, and perform scale normalization; S43, the dot product similarity is truncated to a fixed interval, and then mapped to the (0,1) interval through the Sigmoid function to obtain the gating coefficient; S44, the extended table features are scaled using gating coefficients to obtain attention-weighted feature vectors; S45 concatenates low-dimensional image features, attention-weighted feature vectors, and original three-dimensional table features to construct multimodal fusion features.

6. The method for rapid evaluation of compaction quality of slag-filled roadbed based on multi-source multimodal data as described in claim 5, characterized in that, In S41, the measuring point The first in the extended table vector The number is : ； Indicates at the measuring point Three table features According to Each position is selected in turn.

7. The rapid evaluation method for compaction quality of slag-filled roadbed based on multi-source multimodal data according to claim 5, characterized in that, In S42, the gating coefficient Determined according to the following formula: ； in, This represents the Sigmoid function. Represents the natural constant. For measuring points The result after truncating the cross-modal similarity score.

8. The method for rapid evaluation of compaction quality of slag roadbed based on multi-source multimodal data according to claim 1, characterized in that, S5 includes the following steps: S51: Combine the multimodal fusion features of all samples created in S4 into a matrix and perform standardization; train the Ridge regularized linear regression model and the lightweight multilayer perceptron model in each training set, construct the out-of-fold prediction results, and select the model with the best overall prediction performance for output. S52: Based on the extrapolation prediction results, perform spatial interpolation to draw a cloud map of the predicted FWD modulus of the slag quarry roadbed; S53: Based on the residuals of the out-of-bounds prediction results, identify outliers, output the set of outlier measurement points as a retest list, and mark or remove them.

9. The rapid assessment method for compaction quality of slag roadbed based on multi-source multimodal data according to claim 8, characterized in that, In S51, the sample set is divided using GroupKFold with the measurement point number as the group identifier.

10. The method for rapid evaluation of compaction quality of slag roadbed based on multi-source multimodal data according to claim 8, characterized in that, In step S52, a preset qualified threshold is selected, and a contour line with a predicted value equal to the preset qualified threshold is drawn as the qualified line.

Images