A Prediction Method for Bearing Layer of Cast-in-Place Piles Based on Deep Fusion of Multimodal Data

Abstract

This invention discloses a method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data. By integrating the original borehole data and multimodal data of cast-in-place piles, and leveraging deep learning to mine information, it reduces additional drilling and lowers costs. Simultaneously, through data preprocessing and rapid calculation using a deep learning model, the results are obtained simply by inputting the pile location coordinates after model training, significantly shortening the cycle and improving project progress. Furthermore, by integrating multimodal data to construct a comprehensive feature vector, the model deeply fuses and analyzes the data, learning location correlation patterns to achieve bearing stratum prediction for large-area sites, resulting in more comprehensive and continuous results, overcoming the shortcomings of traditional methods. In addition, the multimodal feature vector encompasses multiple types of information, and deep learning techniques such as cross-modal attention fusion modules consider intermodal relationships. The model comprehensively mines deep feature patterns from various factors, enabling more accurate prediction of key information such as bearing stratum type, burial depth, and thickness.

Description

Technical Field

[0001] This invention relates to the field of bearing layer prediction technology for cast-in-place piles, and in particular to a method for predicting bearing layers of cast-in-place piles based on deep fusion of multimodal data. Background Technology

[0002] In the field of construction engineering, cast-in-place piles are a common type of foundation widely used in various building structures. Accurate identification of the bearing stratum is crucial to ensuring the stability and safety of the building. The bearing stratum refers to the part of the soil that bears the load of the building, and its properties directly affect the bearing capacity and settlement characteristics of the cast-in-place piles. If the bearing stratum is not accurately identified, it may lead to insufficient bearing capacity of the cast-in-place piles, which in turn may cause serious problems such as settlement, tilting or even collapse of the building.

[0003] Traditional methods for detecting the bearing stratum of cast-in-place piles mainly include geological drilling, static cone penetration testing (CPPT), and standard penetration testing (SPT). Geological drilling involves drilling underground soil and rock samples with a drilling rig for laboratory testing and analysis to determine the properties and distribution of the strata. However, this method is costly, has a long construction period, and can only obtain discrete borehole data, limiting its ability to predict the bearing stratum over large areas. Static cone penetration testing and SPT involve in-situ testing to obtain the mechanical parameters of the strata. However, these methods are also affected by factors such as test depth, test accuracy, and operator experience, making it difficult to comprehensively and accurately reflect the condition of the bearing stratum. Summary of the Invention

[0004] In view of this, the present invention proposes a method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, which can effectively solve the defects of existing technologies such as high cost, long construction period, limitations in predicting the bearing stratum of large-area sites, and difficulty in comprehensively and accurately reflecting the condition of the bearing stratum.

[0005] The technical solution of this invention is implemented as follows:

[0006] A method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, specifically including:

[0007] Obtain the original borehole data for geological exploration and preprocess the original borehole data to obtain complete borehole soil layer data;

[0008] Based on complete borehole soil layer data, a multimodal feature vector is constructed, including spatial mode, geological mode, hydrological mode, topographic mode, and cross-modal fusion features;

[0009] A deep learning model is constructed based on multimodal feature vectors. The deep learning model includes a multimodal encoder, a cross-modal attention fusion module, a spatial graph neural network, a sequence generator, and a bearing layer predictor.

[0010] Collect the coordinates of the pile location to be predicted;

[0011] Input the coordinates of the pile location to be predicted into the deep learning model, and the deep learning model outputs the prediction results of the bearing layer of the cast-in-place pile.

[0012] As a further optional scheme of the method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, the step of acquiring the original borehole data from geological exploration and preprocessing the original borehole data to obtain complete borehole soil layer data specifically includes:

[0013] Obtain the original borehole data of the cast-in-place pile, which includes coordinate location, top and bottom elevation of soil layers, main layer number and sub-layer number information;

[0014] The raw borehole data is preliminarily processed to obtain preliminarily processed borehole data.

[0015] Based on the preset bearing stratum geological standards, the pre-processed borehole data is classified into quality categories to obtain complete borehole soil layer data.

[0016] As a further optional scheme of the method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, the method constructs a multimodal feature vector based on complete borehole soil data, including spatial modes, geological modes, hydrological modes, topographic modes, and cross-modal fusion features, specifically including:

[0017] Based on complete borehole soil layer data, spatial modal features are extracted, including coordinate features, K-nearest neighbor features, density features, spatial distribution features, and neighborhood elevation features.

[0018] Based on complete borehole soil data, geological modal features are extracted, including basic statistical features of soil layer sequences, thickness variation coefficient, soil layer type diversity, sequence change frequency, and encoding of bearing layer features.

[0019] Based on complete borehole soil data, hydrological modal features are extracted, including basic hydrological parameter features, hydraulic characteristic features, and the relative position features of the bearing layer and groundwater.

[0020] Based on complete borehole soil data, topographic modal features are extracted, including local topographic features, topographic wetting index, and topographic relief and curvature features;

[0021] Based on the extracted spatial modal features, geological modal features, hydrological modal features, and topographic modal features, the Pearson correlation coefficient, feature entropy, modal balance, and comprehensive evaluation of engineering suitability among different modal features are calculated to obtain cross-modal fusion features;

[0022] By combining spatial modal features, geological modal features, hydrological modal features, topographic modal features, and cross-modal fusion features, a multimodal feature vector is obtained.

[0023] As a further optional scheme of the method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, the step of inputting the coordinates of the pile location to be predicted into a deep learning model and the outputting the prediction result of the bearing stratum of the cast-in-place pile by the deep learning model specifically includes:

[0024] The multimodal encoder maps the coordinates of the pile location to be predicted to a pre-constructed spatial modal feature space, obtaining multimodal features including spatial modality, geological modality, hydrological modality, topographic modality, and cross-modal fusion features;

[0025] The cross-modal attention fusion module uses a multi-head attention mechanism to weightedly fuse multimodal features and generate fused features.

[0026] Spatial graph neural networks utilize the spatial adjacency relationships between boreholes to perform spatial smoothing and enhancement processing on the fused feature representation, resulting in features after spatial smoothing and enhancement.

[0027] The sequence generator processes the spatially smoothed and enhanced features based on the Transformer decoder structure, and generates a sequence of geological profiles from the surface to the bearing layer in an autoregressive manner.

[0028] The bearing stratum predictor processes geological profile sequences to predict the probability of the existence, type, burial depth, and thickness of the bearing stratum.

[0029] As a further alternative to the method for predicting the bearing layer of cast-in-place piles based on deep fusion of multimodal data, the Transformer decoder structure predicts the properties of the next soil layer sequentially by introducing position encoding and an autoregressive mechanism until the end of the sequence is predicted.

[0030] As a further optional solution to the method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, the method further includes verifying the prediction results, specifically including:

[0031] Set the first search radius and the second search radius, as well as the minimum number of verification boreholes;

[0032] If the number of reliable boreholes found within the first search radius reaches the minimum number of verification boreholes, the prediction result is verified based on the bearing layer depth range of the boreholes; otherwise, the verification is expanded to the second search radius, and the verification status is marked as normal, suspicious, or abnormal according to the degree of deviation between the prediction result and the actual value of the neighboring boreholes.

[0033] A system for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data includes:

[0034] The data acquisition and preprocessing module is used to acquire raw borehole data for geological exploration and to preprocess the raw borehole data to obtain complete borehole soil layer data.

[0035] The multimodal feature vector construction module is used to construct multimodal feature vectors, including spatial modes, geological modes, hydrological modes, topographic modes, and cross-modal fusion features, based on complete borehole soil layer data.

[0036] A deep learning model building module is used to build a deep learning model based on multimodal feature vectors. The deep learning model includes a multimodal encoder, a cross-modal attention fusion module, a spatial graph neural network, a sequence generator, and a holding layer predictor.

[0037] The pile location coordinate acquisition module is used to acquire the coordinates of the pile location to be predicted.

[0038] The prediction result output module is used to input the coordinates of the pile position to be predicted into the deep learning model, and the deep learning model outputs the prediction result of the bearing layer of the cast-in-place pile.

[0039] As a further optional solution to the cast-in-place pile bearing stratum prediction system based on multimodal data deep fusion, the system further includes a prediction result verification module for verifying the prediction results, the prediction result verification module comprising:

[0040] The setting unit is used to set the first search radius and the second search radius, as well as the minimum number of verification boreholes;

[0041] The verification unit is used to verify the prediction results based on the bearing layer depth range of the boreholes if the number of reliable boreholes found within the first search radius reaches the minimum number of verification boreholes; otherwise, it expands to the second search radius for verification and marks the verification status as normal, suspicious or abnormal according to the degree of deviation between the prediction results and the actual values ​​of neighboring boreholes.

[0042] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the pre-defined method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data as described above.

[0043] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data.

[0044] The beneficial effects of this invention are as follows: By integrating and analyzing original borehole data from geological exploration and multiple modal data, large-scale additional drilling work is unnecessary. Furthermore, through the mining and utilization of multi-source data using a deep learning model, valuable information can be extracted from existing data, reducing the need for on-site drilling and thus significantly lowering costs. Through data preprocessing and rapid calculation by the deep learning model, the bearing stratum of cast-in-place piles can be predicted in a short time. Once the deep learning model is trained, the prediction results can be quickly output by inputting the coordinates of the pile location to be predicted, greatly shortening the time cycle from data acquisition to result output and improving project progress. In addition, by integrating spatial, geological, hydrological, and topographical multimodal data, a comprehensive... Feature vectors are generated and deep learning models are used to deeply fuse and analyze this data. The deep learning model can learn the correlations and patterns between different locations, thereby enabling the prediction of the bearing layer in large-area sites and providing more comprehensive and continuous prediction results, making up for the shortcomings of traditional methods in terms of spatial coverage. In addition, the constructed multimodal feature vectors cover information from multiple aspects such as space, geology, hydrology, and topography. Through deep learning techniques such as cross-modal attention fusion modules, the intrinsic connections and interactions between different modal data are fully considered. The deep learning model can integrate various factors, mine deep features and patterns in the data, and thus more comprehensively and accurately predict key information such as the type, depth, and thickness of the bearing layer. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments 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 drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, according to the present invention.

[0047] Figure 2 This is a schematic diagram of the composition of a cast-in-place pile bearing layer prediction system based on deep fusion of multimodal data according to the present invention.

[0048] Figure 3 This is a schematic diagram of the composition of a computing device according to the present invention;

[0049] Figure 4 A schematic diagram for constructing multimodal feature vectors;

[0050] Figure 5 This is a schematic diagram of the structure of a deep learning model;

[0051] Figure 6 This is a schematic diagram of the prediction process of a deep learning model. Detailed Implementation

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] refer to Figures 1 to 6 A method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, specifically including:

[0054] Acquiring raw borehole data for geological exploration and preprocessing it to obtain complete borehole soil layer data. In some embodiments, the acquisition and preprocessing of raw borehole data to obtain complete borehole soil layer data specifically includes:

[0055] Obtain the original borehole data of the cast-in-place pile, which includes coordinate location, top and bottom elevation of soil layers, main layer number and sub-layer number information;

[0056] The raw borehole data is preliminarily processed to obtain preliminarily processed borehole data. The preliminary processing includes cleaning the raw borehole data, checking the integrity of key fields, and removing abnormal records; and performing coordinate system transformation to convert the original coordinates into a unified engineering coordinate system to facilitate subsequent calculations and feature extraction.

[0057] Based on the preset bearing stratum geological standards, the pre-processed borehole data is classified into quality categories to obtain complete borehole soil layer data. The preset bearing stratum geological standards include categories such as slightly weathered bearing stratum, moderately weathered bearing stratum, other bearing stratum, insufficient bearing stratum soil thickness, and no bearing stratum encountered.

[0058] Specifically, cleaning the raw borehole data, checking the integrity of key fields and removing abnormal records can effectively remove noise and errors from the data. In actual engineering, the raw data may contain outliers or missing values ​​due to measurement errors, recording mistakes, etc. This step can ensure the accuracy and integrity of the data.

[0059] Transforming the original coordinates into a unified engineering coordinate system solves the problem of coordinate differences that may exist between different data sources. The unified coordinate system facilitates subsequent spatial analysis and feature extraction, enabling accurate comparison and correlation between different borehole data, which helps to more accurately describe the characteristics of geological spatial distribution.

[0060] Based on a pre-defined geological standard for bearing strata, the pre-processed borehole data is classified according to quality, enabling the data to be organized in an orderly manner according to geological characteristics. The pre-defined geological standard for bearing strata covers a variety of common bearing strata types, such as slightly weathered bearing strata and moderately weathered bearing strata. This classification method meets the actual needs of engineering. Through classification, borehole data that meets specific geological conditions can be quickly screened out, improving the usability and relevance of the data. This provides more valuable information for subsequent multimodal feature construction and deep learning model training based on these reliable data, thereby enhancing the scientificity and reliability of the entire prediction scheme.

[0061] Based on complete borehole soil layer data, a multimodal feature vector is constructed, including spatial mode, geological mode, hydrological mode, topographic mode, and cross-modal fusion features. In some embodiments, the construction of the multimodal feature vector based on complete borehole soil layer data, including spatial mode, geological mode, hydrological mode, topographic mode, and cross-modal fusion features, specifically includes:

[0062] Based on complete borehole soil layer data, spatial modal features are extracted, including coordinate features (original coordinates, standardized coordinates, distance from the center point), K-nearest neighbor features (distance between the 5 nearest boreholes), density features (borehole density within 100m, 200m, and 500m radii), spatial distribution features (number of neighboring boreholes in the four quadrants, neighborhood variance, and boundary distance), and neighborhood elevation features (neighborhood elevation standard deviation, elevation range, and elevation gradient), totaling 25 dimensions.

[0063] Based on complete borehole soil data, geological modal features are extracted. These features include basic statistical features of the soil layer sequence (total number of soil layers, total drilling depth, average soil layer thickness), thickness variation coefficient, soil layer type diversity, sequence change frequency, and specialized coding of bearing layer features (existence features, type features, depth features, thickness features, and relative depth features), totaling 20 features.

[0064] Based on complete borehole soil data, hydrological modal features are extracted. These features include basic hydrological parameters (water level characteristics, underwater soil layer ratio) and simplified estimates of hydraulic properties (permeability coefficient estimation), as well as the relative positional characteristics of the bearing layer and groundwater, totaling 15 features.

[0065] Based on complete borehole soil data, topographic modal features are extracted, including the identification of local topographic features (topographic location index), the calculation of topographic moisture index, and topographic relief and curvature, totaling 15 features.

[0066] Based on the extracted spatial modal features, geological modal features, hydrological modal features, and topographic modal features, Pearson correlation coefficient, feature entropy, modal balance, and comprehensive evaluation of engineering suitability among different modal features are calculated to form a 10-dimensional cross-modal fusion feature.

[0067] Spatial modal features, geological modal features, hydrological modal features, topographic modal features, and cross-modal fusion features are combined to construct an 85-dimensional multimodal feature vector.

[0068] Specifically, rich features are extracted from four modes: spatial, geological, hydrological, and topographic. Spatial modal features encompass information such as coordinates, K-nearest neighbors, and density, accurately describing the location of the borehole and its surrounding distribution. Geological modal features include basic statistics on soil layer sequences and specific coding of the bearing layer, reflecting the soil structure and bearing layer characteristics. Hydrological modal features involve basic hydrological parameters and the relative position of the bearing layer and groundwater, demonstrating the influence of hydrological conditions on the bearing layer. Topographic modal features describe topographic features through topographic location indices. This multi-dimensional feature extraction comprehensively characterizes the geological environment of the cast-in-place piles, providing comprehensive data support for bearing layer prediction. All features are extracted based on reliable borehole data, ensuring data authenticity and accuracy, enabling the constructed feature vectors to truly reflect the actual geological conditions and avoiding prediction deviations due to data errors.

[0069] By calculating the Pearson correlation coefficient, feature entropy, modal balance, and comprehensive evaluation of engineering suitability among different modal features, cross-modal fusion features are formed. This step delves into the intrinsic connections and interactions between various modal features, such as the relationship between spatial distribution and geological structure, and between hydrological conditions and topographic features, thereby gaining a more comprehensive understanding of the complexity of the geological environment. Cross-modal fusion features integrate information from different modalities, enhancing the expressive power of feature vectors. Compared with single modal features, fusion features can more accurately describe the comprehensive characteristics of the bearing layer, helping deep learning models to better learn patterns and regularities in the data.

[0070] By combining spatial, geological, hydrological, and topographic modal features with cross-modal fusion features, an 85-dimensional multimodal feature vector is constructed, forming a complete feature system. This system covers various factors affecting the bearing layer of cast-in-place piles, providing sufficient input information for the deep learning model, enabling the model to learn and predict the bearing layer from multiple perspectives.

[0071] A deep learning model is constructed based on multimodal feature vectors. The deep learning model includes a multimodal encoder, a cross-modal attention fusion module, a spatial graph neural network, a sequence generator, and a bearing layer predictor.

[0072] Specifically, the construction of the multimodal encoder:

[0073] A parallelized 1D convolutional neural network branch structure is adopted to independently encode spatial, geological, hydrological, and topographic modal features. Each branch contains: 3 convolutional layers, batch normalization layers and LeakyReLU activation function, and max pooling layer.

[0074] The outputs of each branch are mapped to a 128-dimensional shared feature space through a fully connected layer to generate modality-specific encoding vectors.

[0075] Construction of cross-modal attention fusion module:

[0076] A Transformer-based self-attention mechanism is constructed, with input consisting of four modality encoding vectors (128×4). Feature fusion is achieved through the following operations:

[0077] Calculate the query matrix (Q), key matrix (K), and value matrix (V), each with a dimension of 128×64;

[0078] Scaling dot product attention is applied to output 1024-dimensional fused features; these features are then connected to the residuals via a feedforward neural network to generate 256-dimensional cross-modal enhanced features.

[0079] Spatial Graph Neural Network Construction:

[0080] Construct a Delaunay triangulation space graph structure based on borehole coordinates, with nodes representing borehole locations;

[0081] A graph attention network (GAT) is used to aggregate neighborhood information. Each layer includes: multi-head attention computation, LeakyReLU activation, skip connections, and layer normalization.

[0082] Output 512-dimensional spatial context-aware features.

[0083] Sequence generator construction:

[0084] The cross-modal features (256 dimensions) and spatial graph features (512 dimensions) are concatenated into a 768-dimensional vector;

[0085] The input is fed into a Transformer-based decoder structure, which includes a multi-head self-attention mechanism, position encoding, and a feedforward network, and generates geological profile sequences through an autoregressive approach.

[0086] Construction of the bearing layer predictor:

[0087] A multi-task learning framework is adopted, which includes: a classification head (Softmax outputs the bearing layer type: slightly weathered rock / moderately weathered rock / soil / unpenetrated), a regression head (outputs the burial depth and thickness of the top plate, with the loss function being Huber loss), and a confidence evaluation head (Sigmoid outputs 0-100% confidence).

[0088] Collect the coordinates of the pile location to be predicted;

[0089] The coordinates of the pile location to be predicted are input into a deep learning model, and the deep learning model outputs the prediction result of the bearing layer of the cast-in-place pile; in some embodiments, the step of inputting the coordinates of the pile location to be predicted into a deep learning model and the deep learning model outputting the prediction result of the bearing layer of the cast-in-place pile specifically includes:

[0090] The multimodal encoder maps the coordinates of the pile location to be predicted to a pre-constructed spatial modal feature space, resulting in multimodal features including spatial modes, geological modes, hydrological modes, topographic modes, and cross-modal fusion features;

[0091] The cross-modal attention fusion module uses a multi-head attention mechanism to weightedly fuse multimodal features and generate fused features.

[0092] Spatial graph neural networks utilize the spatial adjacency relationships between boreholes to perform spatial smoothing and enhancement processing on the fused feature representation, resulting in features after spatial smoothing and enhancement.

[0093] The sequence generator processes the spatially smoothed and enhanced features based on the Transformer decoder structure, and generates a geological profile sequence from the surface to the bearing layer in an autoregressive manner through a multi-head self-attention mechanism and position encoding.

[0094] The bearing stratum predictor processes geological profile sequences to predict the probability of the existence, type, burial depth, and thickness of the bearing stratum.

[0095] Specifically, the multimodal encoder maps the coordinates of the pile location to be predicted to a pre-constructed spatial modal feature space, thereby obtaining multimodal features covering space, geology, hydrology, topography and cross-modal fusion. This process can associate the pile location coordinates with rich geological information, providing a comprehensive and targeted feature foundation for subsequent predictions, enabling the model to understand the geological environment of the pile location from multiple dimensions.

[0096] The cross-modal attention fusion module uses a multi-head attention mechanism to weightedly fuse multimodal features to generate fused features. The multi-head attention mechanism can focus on the importance of different modal features from different perspectives and dynamically adjust the weight of each modal feature, thereby more effectively integrating multi-source information, mining the intrinsic relationship between different modal features, and improving the expressive power and information richness of features.

[0097] Spatial graph neural networks utilize the spatial adjacency relationships between boreholes to perform spatial smoothing and enhancement processing on the fused feature representation. By considering the spatial positional relationships between boreholes, the network can better capture the spatial continuity and correlation of geological features, making the features more reasonable and accurate in the spatial dimension, which helps the model understand the spatial distribution patterns of geological features.

[0098] The sequence generator processes the spatially smoothed and enhanced features based on the Transformer decoder structure. Through multi-head self-attention mechanism and position encoding, it autoregressively generates a geological profile sequence from the surface to the bearing layer. The Transformer decoder structure has good sequence modeling capabilities and can gradually generate reasonable geological profile sequences based on existing feature information, accurately simulating the distribution of geological layers from the surface to the bearing layer, and providing more intuitive and detailed information for bearing layer prediction.

[0099] The bearing stratum predictor processes the geological profile sequence to predict the probability of the bearing stratum's existence, type, burial depth, and thickness. Due to the effective processing of features and the reasonable generation of the geological profile sequence in the preceding steps, the bearing stratum predictor can make predictions based on more accurate information, thereby improving the accuracy of bearing stratum prediction. This provides a reliable basis for the design and construction of cast-in-place piles, reduces engineering risks, and improves engineering quality and safety.

[0100] In some embodiments, the Transformer decoder structure predicts the properties of the next soil layer sequentially by introducing positional encoding and an autoregressive mechanism until the end of the sequence is predicted or the maximum depth is reached.

[0101] Specifically, in the generation of geological profile sequences, the order and positional relationships of soil layers are crucial. Positional encoding assigns unique coding information to each soil layer location, enabling the Transformer decoder to identify and utilize the relative positions of soil layers in the sequence. This helps the model understand the spatial distribution patterns of geological layers, such as the order and relationship between upper and lower soil layers, thereby more accurately simulating the real geological structure and providing a more realistic basis for bearing layer prediction.

[0102] By using location encoding, the model can better capture the long-term dependencies of geological profile sequences. Soil properties at different locations may have some potential relationship. Location encoding helps the model pay attention to the correlation between these locations during the learning process, thereby improving the model's understanding and modeling ability of complex geological sequence structures.

[0103] The autoregressive mechanism enables the model to predict the properties of the next soil layer step by step based on the properties of the soil layers that have already been generated. This step-by-step generation method is consistent with the actual formation process of geological profiles from the surface downwards. The model can dynamically adjust subsequent predictions based on the previously predicted soil layer information to better adapt to changes in geological conditions. For example, when the predicted soil layer is of a certain type, the model can more reasonably predict the possible properties of the next soil layer based on geological laws and existing data.

[0104] The autoregressive process ensures the coherence of the geological profile sequence generation. The prediction of each soil layer depends on the previous results, making the entire sequence more logically coherent. At the same time, by continuously using the information from previous predictions to make subsequent predictions, the model can accumulate more contextual information, thereby improving the accuracy of predictions on the probability, type, depth, and thickness of the bearing layer.

[0105] In some embodiments, the method further includes validating the prediction results, specifically including:

[0106] Set the first search radius and the second search radius, as well as the minimum number of verification boreholes;

[0107] If the number of reliable boreholes found within the first search radius reaches the minimum number of verification boreholes, the prediction result is verified based on the bearing layer depth range of the boreholes; otherwise, the verification is expanded to the second search radius, and the verification status is marked as normal, suspicious, or abnormal according to the degree of deviation between the prediction result and the actual value of the neighboring boreholes.

[0108] Specifically, a first search radius, a second search radius, and a minimum number of verification boreholes are set. A phased verification method is adopted. Initial verification is carried out with a smaller first search radius. If the minimum number of verification boreholes is met, verification is carried out within this range based on the depth range of the borehole bearing layer. If not, the search radius is expanded to the second search radius to continue verification. This strategy takes into account the reliability of data within a local range and can expand the range to obtain more information when local data is insufficient, making the verification process more scientific and reasonable and adaptable to the verification needs under different site conditions.

[0109] Using actual borehole data as the basis for verification, the verification status is marked according to the degree of deviation between the predicted results and the actual values ​​of neighboring boreholes. Actual borehole data reflects the real geological conditions, and by comparing it with the actual data, the accuracy of the predicted results can be objectively evaluated.

[0110] By verifying the predictions within the first search radius based on the depth of the borehole bearing layer, predictions that match the actual borehole data in a localized area can be quickly identified and marked as reliable predictions. This helps to identify the more accurate predictions among numerous predictions, providing a more credible basis for subsequent engineering decisions. When the number of boreholes within the first search radius is insufficient, verification can be expanded to a second search radius. This allows for full utilization of borehole data over a larger area to evaluate the predictions. This approach can more comprehensively consider the geological changes and uncertainties of the site, avoiding misjudgments of predictions due to insufficient local data, thereby improving the comprehensiveness and accuracy of the prediction accuracy assessment.

[0111] Marking the verification status as normal, suspicious, or abnormal provides engineers with clear judgment criteria. A normal verification status indicates that the prediction results are in good agreement with the actual borehole data, and the prediction results can be used with confidence for engineering design and construction. A suspicious status suggests that engineers need to further analyze and verify the data. An abnormal status indicates that the prediction results may have a large deviation, and the prediction model needs to be re-examined or more field investigations need to be conducted. This clear marking helps engineers make more reasonable decisions and reduce engineering risks.

[0112] A system for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data includes:

[0113] The data acquisition and preprocessing module is used to acquire raw borehole data for geological exploration and to preprocess the raw borehole data to obtain complete borehole soil layer data.

[0114] The multimodal feature vector construction module is used to construct multimodal feature vectors, including spatial modes, geological modes, hydrological modes, topographic modes, and cross-modal fusion features, based on complete borehole soil layer data.

[0115] A deep learning model building module is used to build a deep learning model based on multimodal feature vectors. The deep learning model includes a multimodal encoder, a cross-modal attention fusion module, a spatial graph neural network, a sequence generator, and a holding layer predictor.

[0116] The pile location coordinate acquisition module is used to acquire the coordinates of the pile location to be predicted.

[0117] The prediction result output module is used to input the coordinates of the pile position to be predicted into the deep learning model, and the deep learning model outputs the prediction result of the bearing layer of the cast-in-place pile.

[0118] In some embodiments, the system further includes a prediction result verification module for verifying the prediction results, the prediction result verification module comprising:

[0119] The setting unit is used to set the first search radius and the second search radius, as well as the minimum number of verification boreholes;

[0120] The verification unit is used to verify the prediction results based on the bearing layer depth range of the boreholes if the number of reliable boreholes found within the first search radius reaches the minimum number of verification boreholes; otherwise, it expands to the second search radius for verification and marks the verification status as normal, suspicious or abnormal according to the degree of deviation between the prediction results and the actual values ​​of neighboring boreholes.

[0121] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the pre-defined method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data as described above.

[0122] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, characterized in that, Specifically, it includes: Obtain the original borehole data for geological exploration and preprocess the original borehole data to obtain complete borehole soil layer data; Based on complete borehole soil layer data, a multimodal feature vector is constructed, including spatial mode, geological mode, hydrological mode, topographic mode, and cross-modal fusion features; A deep learning model is constructed based on multimodal feature vectors. The deep learning model includes a multimodal encoder, a cross-modal attention fusion module, a spatial graph neural network, a sequence generator, and a bearing layer predictor. Collect the coordinates of the pile location to be predicted; Input the coordinates of the pile location to be predicted into the deep learning model, and the deep learning model outputs the prediction results of the bearing layer of the cast-in-place pile. Specifically, based on complete borehole soil layer data, a multimodal feature vector is constructed, including spatial modes, geological modes, hydrological modes, topographic modes, and cross-modal fusion features. Based on complete borehole soil layer data, spatial modal features are extracted, including coordinate features, K-nearest neighbor features, density features, spatial distribution features, and neighborhood elevation features. Based on complete borehole soil data, geological modal features are extracted, including basic statistical features of soil layer sequences, thickness variation coefficient, soil layer type diversity, sequence change frequency, and encoding of bearing layer features. Based on complete borehole soil data, hydrological modal features are extracted, including basic hydrological parameter features, hydraulic characteristic features, and the relative position features of the bearing layer and groundwater. Based on complete borehole soil data, topographic modal features are extracted, including local topographic features, topographic wetting index, and topographic relief and curvature features; Based on the extracted spatial modal features, geological modal features, hydrological modal features, and topographic modal features, the Pearson correlation coefficient, feature entropy, modal balance, and comprehensive evaluation of engineering suitability among different modal features are calculated to obtain cross-modal fusion features; By combining spatial modal features, geological modal features, hydrological modal features, topographic modal features, and cross-modal fusion features, a multimodal feature vector is obtained.

2. The method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data according to claim 1, characterized in that, The process of acquiring raw borehole data for geological exploration and preprocessing it to obtain complete borehole soil layer data specifically includes: Obtain the original borehole data of the cast-in-place pile, which includes coordinate location, top and bottom elevation of soil layers, main layer number and sub-layer number information; The raw borehole data is preliminarily processed to obtain preliminarily processed borehole data. Based on the preset bearing stratum geological standards, the pre-processed borehole data is classified into quality categories to obtain complete borehole soil layer data.

3. The method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data according to claim 2, characterized in that, The process of inputting the coordinates of the pile location to be predicted into the deep learning model, and the outputting the prediction result of the bearing layer of the cast-in-place pile by the deep learning model, specifically includes: The multimodal encoder maps the coordinates of the pile location to be predicted to a pre-constructed spatial modal feature space, resulting in multimodal features including spatial modes, geological modes, hydrological modes, topographic modes, and cross-modal fusion features; The cross-modal attention fusion module uses a multi-head attention mechanism to weightedly fuse multimodal features and generate fused features. Spatial graph neural networks utilize the spatial adjacency relationships between boreholes to perform spatial smoothing and enhancement processing on the fused feature representation, resulting in features after spatial smoothing and enhancement. The sequence generator processes the spatially smoothed and enhanced features based on the Transformer decoder structure, and generates a sequence of geological profiles from the surface to the bearing layer in an autoregressive manner. The bearing stratum predictor processes geological profile sequences to predict the probability of the existence, type, burial depth, and thickness of the bearing stratum.

4. The method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data according to claim 3, characterized in that, The Transformer decoder structure, by introducing positional encoding and an autoregressive mechanism, predicts the properties of the next soil layer sequentially until the end-of-sequence marker is predicted.

5. The method for predicting the bearing stratum of cast-in-place piles based on deep fusion of multimodal data according to claim 4, characterized in that, The method also includes verifying the prediction results, specifically including: Set the first search radius and the second search radius, as well as the minimum number of verification boreholes; If the number of reliable boreholes found within the first search radius reaches the minimum number of verification boreholes, the prediction result is verified based on the bearing layer depth range of the boreholes; otherwise, the verification is expanded to the second search radius, and the verification status is marked as normal, suspicious, or abnormal according to the degree of deviation between the prediction result and the actual value of the neighboring boreholes.

6. A prediction system for the bearing stratum of cast-in-place piles based on deep fusion of multimodal data, characterized in that, include: The data acquisition and preprocessing module is used to acquire raw borehole data for geological exploration and to preprocess the raw borehole data to obtain complete borehole soil layer data. The multimodal feature vector construction module is used to construct multimodal feature vectors, including spatial modes, geological modes, hydrological modes, topographic modes, and cross-modal fusion features, based on complete borehole soil layer data. A deep learning model building module is used to build a deep learning model based on multimodal feature vectors. The deep learning model includes a multimodal encoder, a cross-modal attention fusion module, a spatial graph neural network, a sequence generator, and a holding layer predictor. The pile location coordinate acquisition module is used to acquire the coordinates of the pile location to be predicted. The prediction result output module is used to input the coordinates of the pile position to be predicted into the deep learning model, and the deep learning model outputs the prediction result of the bearing layer of the cast-in-place pile. Specifically, based on complete borehole soil layer data, a multimodal feature vector is constructed, including spatial modes, geological modes, hydrological modes, topographic modes, and cross-modal fusion features. Based on complete borehole soil layer data, spatial modal features are extracted, including coordinate features, K-nearest neighbor features, density features, spatial distribution features, and neighborhood elevation features. Based on complete borehole soil data, geological modal features are extracted, including basic statistical features of soil layer sequences, thickness variation coefficient, soil layer type diversity, sequence change frequency, and encoding of bearing layer features. Based on complete borehole soil data, hydrological modal features are extracted, including basic hydrological parameter features, hydraulic characteristic features, and the relative position features of the bearing layer and groundwater. Based on complete borehole soil data, topographic modal features are extracted, including local topographic features, topographic wetting index, and topographic relief and curvature features; Based on the extracted spatial modal features, geological modal features, hydrological modal features, and topographic modal features, the Pearson correlation coefficient, feature entropy, modal balance, and comprehensive evaluation of engineering suitability among different modal features are calculated to obtain cross-modal fusion features; By combining spatial modal features, geological modal features, hydrological modal features, topographic modal features, and cross-modal fusion features, a multimodal feature vector is obtained.

7. The cast-in-place pile bearing stratum prediction system based on deep fusion of multimodal data according to claim 6, characterized in that, The system further includes a prediction result verification module for verifying the prediction results, the prediction result verification module comprising: The setting unit is used to set the first search radius and the second search radius, as well as the minimum number of verification boreholes; The verification unit is used to verify the prediction results based on the bearing layer depth range of the boreholes if the number of reliable boreholes found within the first search radius reaches the minimum number of verification boreholes; otherwise, it expands to the second search radius for verification and marks the verification status as normal, suspicious or abnormal according to the degree of deviation between the prediction results and the actual values ​​of neighboring boreholes.

8. A computing device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method for predicting the bearing layer of cast-in-place piles based on deep fusion of multimodal data as described in any one of claims 1-5.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for predicting the bearing layer of cast-in-place piles based on deep fusion of multimodal data as described in any one of claims 1-5.

Images