TBM construction geological risk prediction method based on multi-modal historical data fusion

Abstract

This invention discloses a method for predicting geological risks in TBM construction based on multimodal historical data fusion, belonging to the field of tunnel and underground engineering technology. This method includes: aligning and unifying the format of the first multimodal geological data; constructing an initial TBM construction geological risk prediction model, including a text semantic feature extraction module, a graphic structure feature extraction module, a multimodal feature fusion module, and a risk classification prediction module; training the parameters of the initial TBM construction geological risk prediction model using a sample set to obtain a final TBM construction geological risk prediction model; inputting the acquired second multimodal geological data into the final TBM construction geological risk prediction model to output the geological risk type and probability of the section to be predicted. This invention achieves advanced prediction of geological risks ahead by deeply exploring the potential value of historical geological exploration reports and design drawings.

Description

Technical Field

[0001] This invention relates to the field of tunnel and underground engineering technology, and in particular to a method for predicting geological risks in TBM construction based on multimodal historical data fusion. Background Technology

[0002] Tunnel boring machines (TBMs) are the core equipment for constructing deep, long tunnels. Their construction process faces severe challenges from complex geological hazards such as fault fracture zones and sudden water and mud inrushes. Therefore, accurate and advanced geological risk prediction is the key to ensuring the safety and efficiency of TBM construction.

[0003] Existing geological risk early warning methods for TBM construction mainly rely on two types of technologies: one is anomaly detection methods based on real-time monitoring data. These methods are post-event or pre-earthquake responses and lack prior perception capabilities when geological conditions change abruptly, resulting in significant lag in prediction results; the other is advanced detection methods based on geophysical exploration techniques such as ground-penetrating radar and seismic waves. Their interpretation process heavily relies on human experience and is easily affected by on-site electromagnetic interference and complex geological conditions, leading to large fluctuations in prediction accuracy.

[0004] Therefore, the above methods are difficult to achieve high-precision and advanced prediction of geological disasters. Risk areas cannot be identified in advance during construction. Responses are often only made when disasters are imminent or have already occurred, which can lead to major safety accidents such as machine blockages, water inrushes, and landslides, seriously threatening the safety of construction workers and causing delays and huge economic losses. Summary of the Invention

[0005] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes a TBM construction geological risk prediction method based on multimodal historical data fusion. This method can deeply integrate the textual semantics in geological exploration reports with the spatial topology in geological profile maps, achieving advanced and accurate prediction of geological hazards such as faults and water inrushes, and providing reliable risk warnings and decision support for TBM construction.

[0006] The TBM construction geological risk prediction method based on multimodal historical data fusion according to embodiments of the present invention includes: Step 1: Obtain the first multimodal geological data and location index key of the TBM historical construction section, and perform data alignment and format unification processing on the first multimodal geological data. The first multimodal geological data includes geological exploration reports and geological profile maps. Step 2: Construct an initial TBM construction geological risk prediction model. The initial TBM construction geological risk prediction model includes: The text semantic feature extraction module is used to extract text semantic feature vectors from geological exploration reports; The graphic structure feature extraction module is used to extract geological graphic structure feature vectors from geological profile maps; The multimodal feature fusion module is used to perform cross-modal interaction between text semantic feature vectors and geological graphic structural feature vectors, and generate fused feature vectors. The risk classification and prediction module is used to output the type and probability of construction geological risks based on the fused feature vector. Step 3: Construct a sample set based on the first multimodal geological data, and use the sample set to train the parameters of the initial TBM construction geological risk prediction model to obtain the final TBM construction geological risk prediction model. Step 4: Obtain the second multimodal geological data of the section to be predicted, input the second multimodal geological data into the final TBM construction geological risk prediction model, and output the geological risk type and risk probability of the section to be predicted.

[0007] This invention offers at least the following advantages: Addressing the shortcomings of existing TBM construction geological risk early warning methods, such as lag and reliance on human experience, this invention constructs a dual-stream parallel multimodal deep learning architecture based on the data heterogeneity between geological exploration reports and geological profiles. Specifically, the text semantic feature extraction module extracts text semantic feature vectors from geological exploration reports, enabling it to penetrate lengthy geological descriptions and accurately capture the semantic relationships of key geological entities; the geological graphic structure feature extraction module extracts geological graphic structure feature vectors from geological profiles, parsing drawings into graph topology and overcoming the shortcomings of traditional methods in processing vector drawings; the multimodal feature adaptive fusion module performs cross-modal interaction between the two types of feature vectors and generates a fused feature vector, simulating the cognitive process of experts comparing drawings with reports, adaptively assigning different weights to text and graphic features to achieve complementary advantages; and the risk classification and prediction module outputs the geological risk type and probability based on the fused feature vector. Therefore, this invention achieves advanced prediction of geological disasters by mining prior knowledge from historical geological data, overcoming the shortcomings of real-time monitoring methods that lack prior perception capabilities. At the same time, it automatically extracts image and text features and performs fusion decision-making through deep learning models, reducing the reliance on manual interpretation and overcoming the problems of strong subjectivity and large accuracy fluctuations in geophysical interpretation, thereby improving the accuracy and advance capability of geological disaster prediction.

[0008] According to some embodiments of the present invention, in step 1, the geological exploration report is provided in DOC or DOCX format, and the geological profile is provided in DWG or DXF format. Step 1 includes: The geological exploration report was analyzed to extract text information at the corresponding mileage locations of the historical construction sections. Based on the text information, a geological text sequence was constructed. The text information included the surrounding rock grade, rock quality indicators, uniaxial compressive strength, and groundwater level parameters. Analyze the geological profile map, extract the graphic information at the corresponding mileage location of the historical construction section, and obtain the geological topology map structure data based on the graphic information. The graphic information includes the outline of the fault fracture zone, the boundary line of the rock strata, and the mileage coordinate axis. Obtain the location index key of the historical construction section, and spatially align and aggregate the extracted geological text sequence and geological topology map structure data according to the location index key, so that the same location index key contains the corresponding geological text sequence and geological topology map structure data, forming an aligned sample arranged in the tunneling sequence.

[0009] According to some embodiments of the present invention, parsing a geological exploration report includes: performing targeted extraction of the geological exploration report through keyword matching and regular expressions to obtain surrounding rock grade, rock quality indicators, uniaxial compressive strength, and groundwater level parameters as a first geological parameter set; using a large model for semantic approximation processing, inferring and completing parameters for non-standard expressions through contextual understanding, and completing the keyword extraction for resetting, as a second geological parameter set; and integrating the first geological parameter set and the second geological parameter set to construct a standardized geological text sequence.

[0010] According to some embodiments of the present invention, step 1 further includes: obtaining historical construction logs, associating the actual geological disaster events recorded in the historical construction logs with the alignment samples as supervision labels, and completing the construction of the dataset.

[0011] According to some embodiments of the present invention, obtaining geological topological map structure data based on graphic information specifically includes: extracting layer data, filtering labels and map frames, and retaining fault fracture zone outlines, rock strata boundaries, and mileage coordinate axes; vectorizing the geological profile map along the mileage direction, taking the sampling points on the fault fracture zone outlines and rock strata boundaries as geological key points, extracting the coordinates of the geological key points as original node features, and constructing an initial node feature matrix; constructing an adjacency matrix based on the spatial connection relationship of the geological key points; and combining the original node feature matrix and the adjacency matrix to transform it into geological topological map structure data.

[0012] According to some embodiments of the present invention, the text semantic feature extraction module includes an input embedding layer, a position encoding layer, and a Transformer encoder. The working process of the text semantic feature extraction module includes: converting the geological text sequence into a token embedding vector; generating a position encoding through the position encoding layer, and superimposing the token embedding vector and the position encoding to generate an embedding representation; inputting the embedding representation into the Transformer encoder, which is a 12-layer stacked structure, wherein a multi-head self-attention sublayer, a layer normalization and residual connection, and a feedforward neural network sublayer are sequentially set inside the Transformer encoder; capturing long-distance contextual dependencies in the geological text sequence through the Transformer encoder, and outputting a text feature vector containing global deep semantics.

[0013] According to some embodiments of the present invention, the geological graphic structure feature extraction module includes a graph convolutional layer and global average pooling; the working process of the geological graphic structure feature extraction module includes: normalizing the original node features to obtain a standardized node feature matrix; inputting the node feature matrix and the adjacency matrix into the graph convolutional layer, aggregating the spatial neighborhood information of each geological key point through graph convolution operation to obtain updated node features; performing global average pooling on the updated node features, aggregating the updated node-level features into a graph-level feature vector, and outputting the graph-level feature vector as a geological graphic structure feature vector containing geological spatial topological information.

[0014] According to some embodiments of the present invention, the multimodal feature fusion module includes a cross-attention interaction unit, a semantic alignment weighting unit, and a multimodal channel splicing layer. The input of the semantic alignment weighting unit is connected to the output of the cross-attention interaction unit, and the input of the multimodal channel splicing layer is connected to the output of the semantic alignment weighting unit. The fusion of text semantic feature vectors and graphic structure feature vectors includes: the cross-attention interaction unit using the text semantic feature vector as a query vector and the geological graphic structure feature vector as a key vector and a value vector, and calculating the semantic complementary weight matrix of the geological graphic structure feature vector relative to the text semantic feature vector through dot product scaling operation; the semantic alignment weighting unit using the semantic complementary weight matrix to perform weighted summation on the geological graphic structure feature vector to generate spatially aligned graphic context features; and the multimodal channel splicing layer splicing the text semantic feature vector and the spatially aligned graphic context features in the channel dimension to generate a fused feature vector.

[0015] According to some embodiments of the present invention, the risk classification prediction module includes a fully connected layer and a Softmax normalization layer; the working process of the risk classification prediction module includes: inputting the fused feature vector into the fully connected layer for mapping to obtain a classification feature vector; inputting the classification feature vector into the Softmax normalization layer, the Softmax normalization layer maps the classification feature vector to the probability space of geological risk categories, and outputs the geological risk type and its risk probability.

[0016] According to some embodiments of the present invention, in step 3, when training the parameters of the initial TBM construction geological risk prediction model, Focal Loss is used as the loss function, and its expression is:

[0017] in, This represents the probability that the model predicts the actual geological risk category. To focus parameters, It is a balancing factor.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart illustrating the TBM construction geological risk prediction method based on multimodal historical data fusion in this specific embodiment. Figure 2 for Figure 1 A schematic diagram of the network architecture of the geological risk prediction model for TBM construction in China; Figure 3 This is a schematic diagram illustrating the data processing principle of spatiotemporal alignment of multimodal data based on mileage markers in an embodiment of the present invention. Figure 4 This is a comparison chart of the model training convergence curves of the loss function and the ordinary cross-entropy loss function in this embodiment of the invention; Figure 5 This is a comparison chart showing the iterative changes in the accuracy of geological risk prediction between the multimodal fusion model and the single-modal model in embodiments of the present invention. Detailed Implementation

[0020] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, left, right, front, back, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0021] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," "exceeding," etc. are understood to exclude the stated number, and "above," "below," "within," etc. are understood to include the stated number. If "first," "second," etc. are used in the description, they are only configured to distinguish technical features and should not be construed as indicating or implying relative importance or implicitly indicating the number of indicated technical features or the order of the indicated technical features.

[0022] In the description of this invention, unless otherwise explicitly defined, terms such as "set", "install", and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0023] Please refer to Figures 1 to 2 This embodiment discloses a method for predicting geological risks in TBM construction based on multimodal historical data fusion, including: Step 1: Obtain the first multimodal geological data and location index key of the TBM historical construction section, and perform data alignment and format unification processing on the first multimodal geological data. The first multimodal geological data includes geological exploration reports and geological profile maps. Step 2: Construct an initial TBM construction geological risk prediction model. The initial TBM construction geological risk prediction model includes: The text semantic feature extraction module is used to extract text semantic feature vectors from geological exploration reports; The graphic structure feature extraction module is used to extract geological graphic structure feature vectors from geological profile maps; The multimodal feature fusion module is used to perform cross-modal interaction between text semantic feature vectors and geological graphic structural feature vectors, and generate fused feature vectors. The risk classification and prediction module is used to output the type and probability of construction geological risks based on the fused feature vector. Step 3: Construct a sample set based on the first multimodal geological data, and use the sample set to train the parameters of the initial TBM construction geological risk prediction model to obtain the final TBM construction geological risk prediction model. Step 4: Obtain the second multimodal geological data of the section to be predicted, input the second multimodal geological data into the final TBM construction geological risk prediction model, and output the geological risk type and risk probability of the section to be predicted.

[0024] This invention addresses the shortcomings of existing TBM construction geological risk early warning methods, such as their lag and reliance on human experience. Based on the data heterogeneity between geological exploration reports and geological profiles, it constructs a dual-stream parallel multimodal deep learning architecture. Specifically, the text semantic feature extraction module extracts text semantic feature vectors from geological exploration reports, penetrating lengthy geological descriptions to accurately capture the semantic relationships of key geological entities; the geological graphic structure feature extraction module extracts geological graphic structure feature vectors from geological profiles, parsing drawings into graph topology, thus overcoming the limitations of traditional methods for handling vector graphics; the multimodal feature adaptive fusion module performs cross-modal interaction between the two types of feature vectors and generates a fused feature vector, simulating the cognitive process of experts comparing drawings with reports, adaptively assigning different weights to text and graphic features to achieve complementary advantages; and the risk classification and prediction module outputs the geological risk type and probability based on the fused feature vector. Therefore, this invention achieves advanced prediction of geological disasters by mining prior knowledge from historical geological data, overcoming the shortcomings of real-time monitoring methods that lack prior perception capabilities. At the same time, it automatically extracts image and text features and performs fusion decision-making through deep learning models, reducing the reliance on manual interpretation and overcoming the problems of strong subjectivity and large accuracy fluctuations in geophysical interpretation, thereby improving the accuracy and advance capability of geological disaster prediction.

[0025] The following is combined Figures 1 to 5 The specific embodiments shown further illustrate the present invention. The TBM construction geological risk prediction method based on multimodal historical data fusion disclosed in this embodiment includes the following specific steps.

[0026] Step 1: Obtain the first multimodal geological data and location index key of the TBM historical construction section, and perform data alignment and format unification processing on the first multimodal geological data.

[0027] The first type of multimodal geological data includes geological exploration reports and geological profile maps. The geological exploration reports are provided in DOC or DOCX format, and their textual information includes the surrounding rock grade, rock quality indicators, uniaxial compressive strength, and groundwater level parameters at the corresponding mileage locations in the historical construction sections. The geological profile maps are provided in DWG or DXF format, and their graphic information includes the outline of fault fracture zones, strata boundaries, and mileage coordinate axes.

[0028] In some specific embodiments of the present invention, the specific process of data alignment and format unification processing of the first multimodal geological data is as follows: Step 1.1: Parse the geological exploration report in DOC or DOCX format and extract the text information at the corresponding mileage locations of the historical construction sections.

[0029] Specifically, for geological exploration reports in DOC or DOCX format, OCR technology is used to identify the text content, and keyword matching and regular expressions are used to extract the surrounding rock grade, rock quality indicators, uniaxial compressive strength and groundwater level parameters, which are then used as the first set of geological parameters.

[0030] For information data that cannot be extracted through keyword matching or regular expressions, the corresponding text fragments from the geological exploration report are input into a pre-trained large model. The model's contextual understanding capability is used to identify non-standard descriptions in the text fragments and infer their corresponding standard geological parameter values ​​as the second geological parameters. The first and second geological parameter sets are then integrated to construct a standardized geological text sequence.

[0031] Step 1.2: Parse the geological profile map in DWG or DXF format, extract the graphic information at the corresponding mileage location of the historical construction section, and obtain the geological topology structure data based on the graphic information.

[0032] In some specific embodiments of the present invention, obtaining geological topological map structure data based on graphic information specifically includes: Step 1.2.1: For the geological profile map in DWG format, extract the layer data through the DXF parsing interface, filter out irrelevant labels and frames, and retain the outline of the fault fracture zone, the boundary of the rock strata and the mileage coordinate axis. Step 1.2.2: Vectorize the geological profile along the mileage direction, take the sampling points on the fault fracture zone outline and the rock strata boundary as geological key points, extract the coordinates of the geological key points as the original node features, and construct the initial node feature matrix; Step 1.2.3: Construct an adjacency matrix based on the spatial connection relationship of adjacent geological key points on the same fault fracture zone outline or the same rock stratum boundary; Step 1.2.4: Combine the original node feature matrix and adjacency matrix to transform it into geological topological map structure data containing the node feature matrix and adjacency matrix.

[0033] Step 1.3: Obtain the location index key of the historical construction section. Based on the location index key, spatially align and aggregate the extracted geological text sequence and geological topology map structure data, so that the same location index key simultaneously contains the corresponding geological text sequence and geological topology map structure data, forming an aligned sample arranged in the tunneling sequence. The location index key is the TBM construction mileage station number.

[0034] like Figure 3As shown, using the TBM construction mileage station number as a unified location index key, the geological text fragments describing the mileage segment in the text modality are vertically projected and aligned with the corresponding geological profile data in the graphic modality to form a multimodal alignment sample.

[0035] Step 1.4: Obtain historical construction logs, and associate the actual geological disaster events recorded in the historical construction logs with the aligned samples as supervisory labels to complete the dataset construction. This invention provides the model with a clear prediction target by associating the actual geological disaster events recorded in historical construction logs with the aligned samples. This allows the model to learn the mapping pattern between textual and graphical features and risk outcomes from historical experience, thereby achieving accurate prediction of geological risks in the area to be predicted.

[0036] It should be noted that when there is a lack of textual or graphical information in a certain tunneling ring number or mileage section, historical data from adjacent tunneling sections are used for filling or interpolation to reasonably estimate and complete the missing information, thus ensuring the integrity of the multimodal input samples.

[0037] Step 2: Construct an initial geological risk prediction model for TBM construction.

[0038] like Figure 2 As shown, the initial TBM construction geological risk prediction model includes a text semantic feature extraction module, a graphic structure feature extraction module, a multimodal feature fusion module, and a risk classification prediction module. The text semantic feature extraction module and the graphic structure feature extraction module are set up in parallel, with their outputs connected to the input of the multimodal feature fusion module. The output of the multimodal feature fusion module is then connected to the input of the risk classification prediction module.

[0039] In some specific embodiments of the present invention, constructing an initial TBM construction geological risk prediction model specifically includes: Step 2.1: Construct a text semantic feature extraction module adapted to the engineering field to extract text semantic feature vectors containing engineering semantics from geological exploration reports; Step 2.2: Construct a geological graphic structure feature extraction module based on graph topology analysis to extract geological graphic structure feature vectors that can characterize the spatial topological relationship of the construction section from the geological profile map; Step 2.3: Construct a multimodal feature fusion module to perform cross-modal interaction between text semantic features and graphic structural features and generate a fused feature vector; Step 2.4: Construct a risk classification and prediction module to output geological risk type and risk probability based on fused feature vectors.

[0040] The text semantic feature extraction module is built on a BERT pre-trained model based on the Transformer architecture, which specifically includes an input embedding layer, a positional encoding layer, and a Transformer encoder. Specifically, the Transformer encoder is a 12-layer stacked structure, with each layer sequentially consisting of a multi-head self-attention sub-layer, layer normalization and residual connections, and a feedforward neural network sub-layer. The multi-head self-attention sub-layer is configured with 12 attention heads, and the feedforward neural network sub-layer has a 768-dimensional dimension.

[0041] In some specific embodiments of the present invention, the working process of the text semantic feature extraction module is as follows: Step 2.1.1: Convert the input geological text sequence into a token embedding vector, generate a location code through a location coding layer, and superimpose the token embedding vector and the location code to generate an embedding representation.

[0042] Step 2.1.2: Input the embedded representation into the Transformer encoder, which captures long-distance contextual dependencies in the geological text sequence and outputs a text feature vector containing global deep semantics.

[0043] Specifically, the calculation formula is as follows:

[0044] in, , , These are the query matrix, key matrix, and value matrix, respectively, generated from the input vector through a linear transformation. is the scaling factor for the dimension of the key vector; The normalization function ultimately outputs a text semantic feature vector that includes global semantics. .

[0045] In this specific embodiment, the dimension of the text semantic feature vector is 768.

[0046] Furthermore, the geological graphic structure feature extraction module is built based on a graph convolutional neural network, using geological key points as nodes and geological structural lines as edges, and aggregating spatial neighborhood information through graph convolution operations. Specifically, the graphic structure feature extraction module includes graph convolutional layers and global average pooling layers.

[0047] In some specific embodiments of the present invention, the working process of the graphic structure feature extraction module is as follows: Step 2.2.1: Perform Min-Max normalization on the original node features to map the coordinate data to the [0,1] interval to eliminate the influence of dimensions and obtain the standardized node feature matrix; Specifically, the normalization calculation formula is as follows:

[0048] in, These are the original coordinate values. This is the minimum value in the corresponding dimension. This represents the maximum value for the corresponding dimension.

[0049] Furthermore, construct the initial node feature matrix. ,in, The number of nodes is determined by the vectorized sampling density of the geological boundary, and typically ranges from 50 to 500 nodes. This represents the initial feature dimension.

[0050] In this specific embodiment, The corresponding plane coordinates of key geological points.

[0051] Step 2.2.2: Input the node feature matrix and adjacency matrix into the stacked graph convolutional layer, and aggregate the spatial neighborhood information of each geological key point through multi-layer graph convolution operation to obtain the updated node features.

[0052] In this specific embodiment, a two-layer graph convolutional layer structure is adopted. The first graph convolutional layer has an output dimension of 64 dimensions and uses the ReLU activation function. To add non-linear features, the output dimension of the second graph convolutional layer is set to 32 dimensions, and the Dropout mechanism is introduced to prevent overfitting.

[0053] Specifically, the inter-layer feature propagation formula for graph convolutional layers is:

[0054] In the formula, , for the first The input node feature matrix of the layer; , for the first The trainable weight matrix of the layer, its output dimension Typical values ​​are 32 or 64 to match the model complexity; This is an adjacency matrix with added self-loops, used to characterize the spatial topological relationships of geological structures; for The degree matrix; As a non-linear activation function, this specific embodiment uses the ReLU function, with the formula as follows: .

[0055] Step 2.2.3: Perform global average pooling on the updated node features to aggregate node-level features into graph-level feature vectors, outputting a geological graphic structure feature vector containing geological spatial topological information. .

[0056] In this specific embodiment, the geological graphic structure feature vector The dimension is 32.

[0057] Furthermore, the multimodal feature fusion module includes a cross-attention interaction unit, a semantic alignment weighting unit, and a multimodal channel concatenation layer. The input of the semantic alignment weighting unit is connected to the output of the cross-attention interaction unit, and the input of the multimodal channel concatenation layer is connected to the output of the semantic alignment weighting unit.

[0058] In some specific embodiments of the present invention, the working process of the multimodal feature fusion module is as follows: Step 2.3.1: The cross-attention interaction unit uses the text semantic feature vector as the query vector and the geological graphic structure feature vector as the key vector and value vector, and calculates the semantic complementary weight matrix of the geological graphic structure feature vector relative to the text semantic feature vector through dot product scaling operation.

[0059] Step 2.3.2: The semantic alignment weighted unit uses the semantic complementary weight matrix to perform weighted summation on the geological graphic structure feature vector to generate spatially aligned graphic context features.

[0060] Step 2.3.3: The multimodal channel concatenation layer concatenates the text semantic feature vector with the spatially aligned graphical context features along the channel dimension to generate a fused feature vector. .

[0061] The fusion calculation formula for the multimodal feature fusion module is as follows:

[0062] In the formula, This is the fused multimodal risk feature vector; This refers to the structural feature vector of geological graphics; For channel splicing operations, this step aims to allow the model to focus on specific areas in the drawings based on the descriptions in the geological report; This is the feature scaling factor.

[0063] Through the above steps, the model can simulate the cognitive process of an expert comparing drawings with a report, adaptively focusing on the graphic areas most relevant to the text description, and achieving complementary advantages of graphic and text features.

[0064] Furthermore, the risk classification prediction module includes a fully connected layer and a Softmax normalization layer. The fused feature vector is sequentially input into the fully connected layer and the Softmax normalization layer, and after mapping and probability normalization, the probability value of each risk category is output.

[0065] Specifically, the fully connected layer consists of two layers: the first fully connected layer has a dimension of 128 and uses the ReLU activation function; the second fully connected layer has a dimension of 64. After nonlinear mapping by the fully connected layer, the classification feature vector is input into the Softmax normalization layer, which maps it to the probability space of geological risk categories, outputting the specific geological risk type and its probability.

[0066] Step 3: Construct a sample set based on the first multimodal geological data, and use the sample set to train the parameters of the initial TBM construction geological risk prediction model to obtain the final TBM construction geological risk prediction model. Furthermore, the aligned samples formed in step 1 are used to construct training samples, and the sample set is divided into a training set and a test set according to a preset ratio. The training set is used for model parameter training, and the test set is used for model prediction performance verification.

[0067] It should be noted that a test set was used for testing to obtain prediction results and evaluation metrics such as accuracy and F1 score. Among them, accuracy is used to measure the correctness of the model's prediction for the overall sample, while the F1 score is used to comprehensively evaluate the model's precision and recall for a minority of risk samples such as faults and water inrushes.

[0068] Furthermore, to address the imbalance between positive and negative samples caused by the scarcity of geological hazard samples, Focal Loss is used as the loss function during parameter training of the initial TBM construction geological risk prediction model. This loss function forces the model to focus on the minority class samples that are difficult to classify, and continuously updates the model parameters through the backpropagation algorithm until the model converges. Specifically, the expression for the loss function is:

[0069] In the formula, This represents the probability that the model predicts the actual geological risk category. Focus parameters are used to increase the model's attention to risky samples that are difficult to classify. It is a balancing factor.

[0070] To verify the effectiveness of the multimodal historical data fusion method proposed in this invention, this embodiment constructs a dataset based on real TBM construction data and conducts a comparative test with existing single-modal methods.

[0071] The experiment used historical geological data from a TBM construction section of a water diversion tunnel project to construct a dataset containing 2400 sets of text-image aligned samples. The training and test sets were divided in an 8:2 ratio. The evaluation metrics used were accuracy and F1 score. Accuracy measures the model's correct prediction of the overall sample, while the F1 score comprehensively evaluates the model's precision and recall for minority risk samples such as faults and water inrushes, better reflecting the model's actual performance under imbalanced sample conditions.

[0072] The training parameters were set as follows: the Adam optimizer was used, the initial learning rate was set to 0.001, the batch size was set to 32, and the maximum number of iterations was 100. During training, an early stopping strategy was adopted to prevent overfitting when the validation set loss no longer decreased.

[0073] To verify the advantages of multimodal fusion, the following comparison scheme was set up: (1) Baseline Scheme A: Graphical mode only, using GCN to process DWG geological profiles; (2) Baseline Scheme B: Text-only, using BERT to process geological exploration reports; (3) The present invention solution: integrates text modality and graphics modality, and adopts BERT+GCN+cross attention fusion mechanism.

[0074] Figure 5 The figure shows the accuracy variation curves of different schemes during the training process. As can be seen from the figure, the single-modality scheme is limited by the single information dimension, resulting in a slower convergence speed and a lower upper limit of accuracy; while the scheme of this invention achieves the complementary advantages of image and text features through the cross-attention mechanism, which significantly improves the accuracy in the later stages of iteration and eventually stabilizes at a high level.

[0075] Figure 4 The convergence curves of the Focal Loss loss function used in this invention and the traditional cross-entropy loss function are compared. The results show that Focal Loss can more effectively solve the training difficulties caused by the scarcity of geological disaster samples, enabling the model to converge faster and with lower loss values.

[0076] Table 1 Comparison of Prediction Performance of Different Model Schemes

[0077] Table 1 presents the final quantitative comparison results. As can be seen from Table 1, after fusing text and graphic features, the accuracy of the proposed solution is improved by 13.9% compared to the best-performing benchmark solution B, and the F1 score is improved by 0.13. This fully demonstrates that the cross-attention fusion module proposed in this invention can effectively extract and integrate complementary information from the text and graphic modalities, significantly enhancing the model's ability and robustness in identifying complex geological risks such as faults and water inrushes.

[0078] Step 4: Obtain the second multimodal geological data of the section to be predicted, input the second multimodal geological data into the final TBM construction geological risk prediction model, and output the geological risk type and risk probability of the section to be predicted.

[0079] Specifically, the geological exploration report and geological profile map corresponding to the current section to be predicted are obtained. The data is aligned and formatted according to the method in step 1. The processed multimodal data is then input into the final TBM construction geological risk prediction model after training. The model outputs the geological risk type and risk probability of the corresponding location index key, thereby assisting construction personnel in formulating reinforcement measures in advance.

[0080] In summary, this invention proposes a TBM construction geological risk prediction method based on multimodal historical data fusion. By deeply mining the potential value of historical geological exploration reports and design drawings, it achieves high-precision, intelligent, and advanced prediction of geological risks ahead. Addressing the data heterogeneity of geological exploration reports and design drawings, this invention constructs a dual-stream parallel multimodal deep learning architecture. Specifically, the text semantic feature extraction module utilizes the self-attention mechanism of the BERT model to penetrate lengthy geological descriptions and accurately capture the semantic relationships of key geological entities such as "faults" and "water inrushes." The geological graphic structure feature extraction module innovatively parses DWG drawings into a graph topology structure and uses graph convolutional networks to aggregate the spatial neighborhood information of stratigraphic nodes, compensating for the shortcomings of traditional methods in processing vector drawings. The multimodal feature adaptive fusion module, through a cross-attention mechanism, simulates the cognitive process of experts comparing drawings with reports, adaptively assigning different weights to text and graphic features to achieve complementary advantages. Furthermore, to address the problem of scarce geological hazard samples, a Focal Loss loss function is introduced, forcing the model to focus on indistinguishable risk samples, significantly improving the robustness of the prediction. Meanwhile, experimental results show that the method of the present invention achieved a prediction accuracy of 92.5% on a real TBM construction dataset, which is 13.9% higher than that of a single text modality, and the F1 score reached 0.91. This represents a leap from post-event response to proactive early warning, providing reliable technical support for geological risk prevention and control in TBM construction.

[0081] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] Of course, the present invention is not limited to the above-described embodiments. Those skilled in the art can make equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for predicting geological risks in TBM construction based on multi-modal historical data fusion, characterized in that, include: Step 1: Obtain the first multimodal geological data and location index key of the TBM historical construction section, and perform data alignment and format unification processing on the first multimodal geological data, wherein the first multimodal geological data includes geological exploration reports and geological profile maps; Step 2: Construct an initial TBM construction geological risk prediction model, which includes: The text semantic feature extraction module is used to extract text semantic feature vectors from the geological exploration report; The graphic structure feature extraction module is used to extract geological graphic structure feature vectors from the geological profile map; A multimodal feature fusion module is used to perform cross-modal interaction between the text semantic feature vector and the geological graphic structural feature vector, and generate a fused feature vector; The risk classification and prediction module is used to output the construction geological risk type and risk probability based on the fused feature vector; Step 3: Construct a sample set based on the first multimodal geological data, and use the sample set to train the parameters of the initial TBM construction geological risk prediction model to obtain the final TBM construction geological risk prediction model. Step 4: Obtain the second multimodal geological data of the section to be predicted, input the second multimodal geological data into the final TBM construction geological risk prediction model, and output the geological risk type and risk probability of the section to be predicted.

2. The method for predicting geological risks in TBM construction based on multimodal historical data fusion according to claim 1, characterized in that, In step 1, the geological exploration report is provided in DOC or DOCX format, and the geological profile is provided in DWG or DXF format. Step 1 includes: The geological exploration report is analyzed to extract text information at the corresponding mileage locations of historical construction sections. A geological text sequence is constructed based on the text information, which includes surrounding rock grade, rock quality indicators, uniaxial compressive strength, and groundwater level parameters. The geological profile is analyzed to extract graphic information at the corresponding mileage locations of historical construction sections. Based on the graphic information, geological topology data is obtained, wherein the graphic information includes the outline of fault fracture zones, rock strata boundaries, and mileage coordinate axes. Obtain the location index key of the historical construction section, and spatially align and aggregate the extracted geological text sequence and the geological topology map structure data according to the location index key, so that the same location index key simultaneously contains the corresponding geological text sequence and the geological topology map structure data, forming an aligned sample arranged in the tunneling sequence.

3. The method for predicting geological risks in TBM construction based on multimodal historical data fusion according to claim 2, characterized in that, The analysis of the geological exploration report includes: The geological exploration report was extracted in a targeted manner by keyword matching and regular expressions to obtain the surrounding rock grade, rock quality index, uniaxial compressive strength and groundwater level parameters, which were used as the first set of geological parameters. A large model is used for semantic approximation. Non-standard expressions are inferred and parameters are completed through contextual understanding. Keyword extraction is then used as a second set of geological parameters. The first set of geological parameters and the second set of geological parameters are integrated to construct a standardized geological text sequence.

4. The method for predicting geological risks in TBM construction based on multimodal historical data fusion according to claim 2, characterized in that, Step 1 also includes: obtaining historical construction logs, using the actual geological disaster events recorded in the historical construction logs as supervisory labels to associate with the aligned samples, and completing the construction of the dataset.

5. The method for predicting geological risks in TBM construction based on multimodal historical data fusion according to claim 2, characterized in that, Obtaining geological topological map structure data based on the aforementioned graphic information specifically includes: Extract layer data, filter out labels and map frames, and retain the outline of fault fracture zones, rock strata boundaries, and mileage coordinate axes; The geological profile is vectorized along the mileage direction. The sampling points on the fault fracture zone outline and the rock strata boundary are taken as geological key points. The coordinates of the geological key points are extracted as the original node features to construct an initial node feature matrix. An adjacency matrix is ​​constructed based on the spatial connectivity of the aforementioned key geological points. The original node feature matrix is ​​combined with the adjacency matrix and transformed into geological topological map structure data.

6. The method for predicting geological risks in TBM construction based on multimodal historical data fusion according to claim 1, characterized in that, The text semantic feature extraction module includes an input embedding layer, a positional encoding layer, and a Transformer encoder; The working process of the text semantic feature extraction module includes: The geological text sequence is converted into a token embedding vector; The location encoding layer generates a location code, and the token embedding vector and the location code are superimposed to generate an embedding representation; The embedded representation is input to the Transformer encoder, which is a 12-layer stacked structure. The Transformer encoder contains a multi-head self-attention sub-layer, a layer normalization and residual connection, and a feedforward neural network sub-layer. The Transformer encoder captures long-distance contextual dependencies in the geological text sequence and outputs a text feature vector containing global deep semantics.

7. The method for predicting geological risks in TBM construction based on multimodal historical data fusion according to claim 5, characterized in that, The geological graphic structure feature extraction module includes a graph convolutional layer and global average pooling; The working process of the geological graphic structure feature extraction module includes: The original node features are normalized to obtain a standardized node feature matrix; The node feature matrix and the adjacency matrix are input into the graph convolutional layer, and the spatial neighborhood information of each geological key point is aggregated through the graph convolution operation to obtain updated node features; The updated node features are subjected to global average pooling to aggregate the updated node-level features into a graph-level feature vector. The graph-level feature vector is then output as a geological graphic structure feature vector containing geological spatial topological information.

8. The method for predicting geological risks in TBM construction based on multimodal historical data fusion according to claim 1, characterized in that, The multimodal feature fusion module includes a cross-attention interaction unit, a semantic alignment weighting unit, and a multimodal channel splicing layer. The input of the semantic alignment weighting unit is connected to the output of the cross-attention interaction unit, and the input of the multimodal channel splicing layer is connected to the output of the semantic alignment weighting unit. The fusion of the text semantic feature vector and the graph structure feature vector includes: The cross-attention interaction unit uses the text semantic feature vector as the query vector and the geological graphic structure feature vector as the key vector and value vector, and calculates the semantic complementary weight matrix of the geological graphic structure feature vector relative to the text semantic feature vector through dot product scaling operation. The semantic alignment weighting unit uses the semantic complementary weight matrix to perform a weighted summation of the geological graphic structure feature vector to generate spatially aligned graphic context features. The multimodal channel splicing layer splices the text semantic feature vector with the spatially aligned graphic context features along the channel dimension to generate a fused feature vector.

9. The method for predicting geological risks in TBM construction based on multimodal historical data fusion according to claim 1, characterized in that, The risk classification prediction module includes a fully connected layer and a Softmax normalization layer; The working process of the risk classification and prediction module includes: The fused feature vector is input into the fully connected layer for mapping to obtain a classification feature vector. The classification feature vector is input into the Softmax normalization layer, which maps the classification feature vector to the probability space of geological risk categories and outputs the geological risk type and its risk probability.

10. The method for predicting geological risks in TBM construction based on multimodal historical data fusion according to claim 1, characterized in that, In step 3, when training the parameters of the initial TBM construction geological risk prediction model, FocalLoss is used as the loss function, and its expression is: in, This represents the probability that the model predicts the actual geological risk category. To focus parameters, It is a balancing factor.

Images