Intelligent tunnel surrounding rock grading prediction method and system based on multi-modal spatio-temporal fusion

Abstract

The application discloses a tunnel surrounding rock intelligent grading prediction method and system based on multi-modal space-time fusion, relates to the technical field of tunnel construction, and provides the following scheme, which comprises obtaining time-stamped surrounding rock multi-modal data, the multi-modal data comprising geological radar waveform data, drilling core images and three-dimensional laser point cloud data; a space-time fusion network architecture is adopted to generate surrounding rock feature data and perform tunnel surrounding rock grading prediction. According to the scheme, three types of multi-modal data, namely, geological radar waveform data, drilling core images and three-dimensional laser point cloud data, are fused to provide a comprehensive feature basis for surrounding rock grading prediction, a real-time feedback mechanism is constructed by introducing construction machinery vibration spectrum data, abnormal disturbance is identified by comparing a dynamic threshold library with historical data, vibration features and surrounding rock features are associated based on an LSTM network, accurate rock mass parameter correction amounts are generated, and space-time fusion network parameters are reversely optimized, so that the deviation between static prediction and actual geological conditions is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of tunnel construction technology, specifically to a method and system for intelligent classification and prediction of tunnel surrounding rock based on multimodal spatiotemporal fusion. Background Technology

[0002] In highway tunnel construction, the stability of the surrounding rock is the core factor that determines the quality of the project and the safety of construction. Accurately classifying the surrounding rock grade and fully understanding the rock mass quality characteristics are the key basis for formulating excavation methods, optimizing support schemes, and avoiding geological disasters.

[0003] While current mainstream methods for predicting surrounding rock classification have incorporated multimodal data fusion technology to construct prediction models, these methods still essentially rely on static survey data before construction and fail to incorporate mechanical disturbance signals generated during construction into the iterative system of the prediction model. This results in discrepancies between the predicted surrounding rock classification and the actual geological conditions. Therefore, we propose a method and system for intelligent classification and prediction of tunnel surrounding rock based on multimodal spatiotemporal fusion to address this issue. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides a method and system for intelligent classification and prediction of tunnel surrounding rock based on multimodal spatiotemporal fusion. This technical solution solves the problem that current tunnel surrounding rock classification and prediction methods generally adopt multimodal data fusion technology to establish a surrounding rock model, but in essence, they still belong to the prediction system based on static survey data before construction. This technical solution fails to incorporate real-time disturbance signals generated during construction into the prediction model, resulting in the surrounding rock classification results being unable to dynamically respond to changes in actual working conditions.

[0005] To achieve the above objectives, the technical solution adopted by this invention is as follows: This invention provides a method for intelligent classification and prediction of tunnel surrounding rock based on multimodal spatiotemporal fusion, the method comprising:

[0006] Acquire time-stamped multimodal data of the surrounding rock, including ground-penetrating radar waveform data, drill core images, and three-dimensional laser point cloud data;

[0007] A spatiotemporal fusion network architecture is adopted, combined with a Transformer-based cross-modal attention mechanism, to extract heterogeneous features and align them spatiotemporally from multimodal data, generate surrounding rock feature data, and perform tunnel surrounding rock classification prediction.

[0008] Based on the computing nodes deployed at the edge of the tunnel construction section, the vibration spectrum data of the construction machinery is collected in real time.

[0009] An online learning mechanism of construction disturbance feedback is established, when the vibration spectrum data is abnormal, the parameters in the spatial-temporal fusion network architecture are corrected in combination with the surrounding rock characteristic data, and the tunnel surrounding rock classification prediction result is updated.

[0010] Further, a tunnel surrounding rock intelligent classification prediction system based on multi-modal spatial-temporal fusion is proposed, which is used for implementing the multi-modal spatial-temporal fusion based tunnel surrounding rock intelligent classification prediction method in any one of the above, and comprises:

[0011] A multi-modal data acquisition module is configured to acquire timestamped surrounding rock multi-modal data, wherein the multi-modal data comprises geological radar waveform data, drilling core image and three-dimensional laser point cloud data.

[0012] A spatial-temporal fusion and classification prediction module is configured to adopt a spatial-temporal fusion network architecture, combine a cross-modal attention mechanism based on Transformer, perform heterogeneous feature extraction and spatial-temporal alignment on the multi-modal data, generate surrounding rock characteristic data, and perform tunnel surrounding rock classification prediction.

[0013] A vibration spectrum data acquisition module is configured to acquire vibration spectrum data of construction machinery in real time based on a computing node arranged at an edge of a tunnel construction section.

[0014] A learning and parameter correction module is configured to establish an online learning mechanism of construction disturbance feedback, correct parameters in the spatial-temporal fusion network architecture in combination with the surrounding rock characteristic data when the vibration spectrum data is abnormal, and update the tunnel surrounding rock classification prediction result.

[0015] Compared with the prior art, the present application has the following beneficial effects:

[0016] The present application fuses three types of multi-modal data, i.e., geological radar waveform data, drilling core image and three-dimensional laser point cloud data, combines a Transformer cross-modal attention mechanism to realize spatial-temporal alignment and dynamic correlation of heterogeneous features, completely characterizes surrounding rock properties from three dimensions of internal structure, appearance and shape, provides a comprehensive feature basis for surrounding rock classification prediction, introduces construction machinery vibration spectrum data to construct a real-time feedback mechanism, identifies abnormal disturbance through dynamic threshold value library and historical data comparison, correlates vibration features and surrounding rock features based on an LSTM network, generates accurate rock mass parameter correction amount, and reversely optimizes spatial-temporal fusion network parameters, thereby effectively reducing the deviation between static prediction and actual geological conditions. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A flowchart of the multi-modal spatial-temporal fusion based tunnel surrounding rock intelligent classification prediction method proposed in the present application;

[0018] Figure 2 The execution flow chart of the spatio-temporal fusion architecture in the present application;

[0019] Figure 3 The method flow chart of the surrounding rock grading prediction in the present application;

[0020] Figure 4 The structural block diagram of the tunnel surrounding rock intelligent grading prediction system based on multi-modal spatio-temporal fusion proposed in the present application. DETAILED DESCRIPTION

[0021] The following description is provided to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only examples, and other obvious modifications can be made by those skilled in the art.

[0022] Referring to Figures 1-3 As shown, the tunnel surrounding rock intelligent grading prediction method based on multi-modal spatio-temporal fusion comprises:

[0023] Obtaining multi-modal data associated with time stamp of surrounding rock, the multi-modal data comprising geological radar waveform data, drilling core image and three-dimensional laser point cloud data; the geological radar waveform data is used to obtain internal structure information of surrounding rock, the drilling core image is used to record surface texture and fracture distribution of surrounding rock, and the three-dimensional laser point cloud data is used to construct a spatial geometric model of surrounding rock, to completely characterize the surrounding rock from three dimensions and support stability evaluation;

[0024] Using a spatio-temporal fusion network architecture, combining a cross-modal attention mechanism based on Transformer, heterogeneous feature extraction and spatio-temporal alignment are performed on multi-modal data to obtain surrounding rock feature data, and tunnel surrounding rock grading prediction is performed;

[0025] Based on the computing nodes deployed at the edge of the tunnel construction section, vibration spectrum data of construction machinery is collected in real time;

[0026] Exemplarily, the construction machinery can be a tunnel boring machine;

[0027] An online learning mechanism of construction disturbance feedback is established, when the vibration spectrum data is abnormal, the parameters in the spatio-temporal fusion network architecture are corrected in combination with the surrounding rock feature data, and the tunnel surrounding rock grading prediction is updated.

[0028] Through the technical solution, the application collects vibration frequency spectrum data of construction machinery in real time by deploying computing nodes at the edge of the construction section, extracts the energy proportion and peak frequency of each frequency sub-section, judges whether the vibration frequency spectrum data is abnormal according to the surrounding rock characteristic vector and the dynamic threshold library, calculates the energy deviation and frequency deviation if the vibration frequency spectrum data is abnormal, constructs a vibration characteristic vector, inputs the vibration characteristic vector into a pre-trained LSTM network after splicing with the surrounding rock characteristic vector, outputs the rock mass parameter correction amount, combines the reconstruction error analysis, corrects the weight, bias and learning matrix of the spatio-temporal fusion network architecture, and finally recalculates the rock mass basic quality index with the corrected engineering parameters, updates the tunnel surrounding rock grade, and reduces the deviation between the model prediction result and the actual result.

[0029] As an embodiment of the application, as shown in Figure 2 The spatio-temporal fusion network architecture combines a cross-modal attention mechanism based on Transformer to extract heterogeneous features and spatio-temporal alignment of multi-modal data, obtain surrounding rock characteristic data, and perform intelligent classification and prediction of tunnel surrounding rock. Specifically, it includes:

[0030] The extraction layer of the spatio-temporal fusion network architecture includes a time domain feature extraction sub-channel for geological radar waveform data, a convolutional neural network feature extraction sub-channel for drilling core images, and a point cloud feature extraction sub-channel for three-dimensional laser point cloud data.

[0031] For example, the time domain feature extraction sub-channel divides the geological radar waveform data into fixed length segments, arranges them into a sequence in timestamp order after extracting the local features of each segment through a one-dimensional convolutional neural network, and arranges them into a sequence in timestamp order.

[0032] The convolutional neural network feature extraction sub-channel cuts the drilling core images into fixed interval slices along the core axis and associates them with the drilling depth coordinates, arranges them into a sequence in drilling depth order after extracting the deep features of each slice through a convolutional neural network, and arranges them into a sequence in drilling depth order. The point cloud feature extraction sub-channel divides the three-dimensional laser point cloud into a fixed size voxel grid, arranges them into a sequence in spatial coordinate ascending order after extracting the 1024-dimensional features of each voxel grid through PointNet++, and arranges them into a sequence in spatial coordinate ascending order. The fixed length is set to 1 / 10 of the tunnel boring machine cutterhead rotation period, which facilitates the capture of high frequency vibration features of the interaction between the cutter teeth and the rock mass. The fixed interval is set to 1 / 20 of the cutter diameter, which ensures that the single drilling influence depth is covered. The voxel grid size is set to 1 / 30 of the cutter diameter, which ensures that 30 voxel units are included within the single drilling range.

[0033] The multi-modal data is input into the extraction layer of the spatio-temporal fusion network architecture to generate corresponding sequence data, which is mapped to a unified dimension through a linear projection layer and is synchronously associated with time encoding. The time encoding is a sine function encoding of the timestamp associated with each modal data.

[0034] The encoded sequence data are aligned according to time encoding, concatenated into a joint matrix, and input into a Transformer architecture. The output is a global feature vector of the multimodal data, which includes the ground-penetrating radar feature vector. Drilling core feature vectors and laser point cloud feature vector ;

[0035] The learning matrix is ​​initialized using Xavier, and the global feature vectors are fused using a weighted model with a gating mechanism to generate the surrounding rock feature vector.

[0036] The surrounding rock feature vector is input into the decoding layer of the spatiotemporal fusion network architecture, mapped to generate engineering parameters, and used for tunnel surrounding rock classification prediction.

[0037] The weighted model of the gating mechanism is as follows:

[0038] ;

[0039] In the formula, The characteristic vector of the surrounding rock. This is the concatenated vector of all global feature vectors. For learning the matrix, Let i be the global feature vector corresponding to number i, where i = 1, 2, 3;

[0040] As one embodiment of the present invention, such as Figure 3 As shown, the step of inputting the surrounding rock feature vector into the decoding layer of the spatiotemporal fusion network architecture, mapping it to generate engineering parameters, and performing intelligent classification prediction of tunnel surrounding rock specifically includes:

[0041] The feature vectors of the surrounding rock are subjected to layer normalization, and then mapped to the engineering parameter space through a two-layer fully connected network in the decoding layer.

[0042] The first fully connected layer uses a weight matrix. and bias For the characteristic vector of the surrounding rock A linear transformation is performed, followed by the ReLU activation function to extract higher-order nonlinear features. The second fully connected network then uses a weight matrix... and bias Perform a linear transformation on the higher-order nonlinear characteristics to generate an output vector;

[0043] For example, the weight matrix Using a normal distribution initialization method, starting with a mean of 0 and a standard deviation of 0... Random sampling is performed from a normal distribution, with a bias. Using an all-zero initialization strategy, the weight matrix... The Xavier uniform distribution initialization method is used, and the sampling range is set to... 'n' represents the output parameter dimension being 3, and the bias is... Based on the actual range of values ​​for the engineering parameters, set it to the median value of the range;

[0044] The output vector is split and processed according to the engineering adaptation requirements to obtain engineering parameters, including rock mass quality indicators, elastic modulus and joint density.

[0045] Input the engineering parameters into the industry standard grading formula, calculate the basic quality index of the rock mass, and compare it with the rock mass grading standard table to obtain the tunnel surrounding rock grade.

[0046] It should be noted that the calculation formula for the basic quality indicators of the rock mass is as follows: ,in, , For elastic modulus, , The density is determined by joints; and the basic quality indicators of the rock mass are corrected by adjusting the formula, which is as follows: ,in, These are the revised basic quality indicators for the rock mass. These are indicators of rock mass quality.

[0047] As one embodiment of the present invention, the online learning mechanism for establishing construction disturbance feedback specifically includes:

[0048] The vibration spectrum data transmitted by the computing node is monitored in real time, divided into preset frequency sub-segments, and its characteristic parameters are extracted, including energy ratio and peak frequency.

[0049] Based on historical construction data, the distribution patterns of vibration spectrum characteristic parameters are statistically analyzed according to the tunnel surrounding rock grade, and a dynamic threshold library corresponding to each surrounding rock grade is generated.

[0050] It should be noted that the peak frequency is the frequency corresponding to the maximum power spectral density, and the energy percentage is analyzed using the energy entropy formula, which is: In the formula, Let f be the normalized power spectral density of the vibration signal at frequency f, and and These are the endpoint values ​​of the frequency sub-band range. This refers to the percentage of energy used.

[0051] Based on the current surrounding rock feature vector, the feature threshold range corresponding to the surrounding rock grade is matched from the dynamic threshold library. When any feature parameter exceeds the corresponding feature threshold range, it is judged as abnormal data.

[0052] For example, based on the coupling vibration characteristics of the tunnel and tunneling machine with the surrounding rock and industry standard references, multiple frequency sub-bands can be set to 0-50Hz, 50-150Hz, 150-300Hz and 300-500Hz; based on the characteristic parameters corresponding to each surrounding rock grade in historical construction data, their mean and standard deviation are calculated, and the characteristic threshold range is set to mean ± standard deviation.

[0053] As one embodiment of the present invention, when the vibration spectrum data is abnormal, the parameters in the spatiotemporal fusion network architecture are corrected in conjunction with the surrounding rock characteristic data, and the tunnel surrounding rock classification prediction is updated, specifically including:

[0054] Based on the current surrounding rock grade, extract the historical average values ​​of energy proportion and peak frequency from the historical construction data;

[0055] Based on the abnormal vibration spectrum data, the difference between its energy proportion and peak frequency and the corresponding historical mean is recorded as energy deviation and frequency offset, respectively, and a vibration feature vector is constructed using energy deviation and frequency offset.

[0056] The vibration feature vector and the surrounding rock feature vector are concatenated into a joint vector, which is then input into a pre-trained LSTM network model to output rock mass parameter corrections. The rock mass parameter corrections include rock mass quality index corrections, elastic modulus corrections, and joint density corrections.

[0057] It should be noted that the training data for the LSTM network model are vibration spectrum data, surrounding rock characteristic data, and actual rock mass parameter changes extracted from historical construction data. Its architecture consists of an input layer, an output layer, and two LSTM layers. Its training strategies include the MSE loss function, an optimizer, and an early stopping mechanism. The learner rate of the optimizer is 0.001, and the early stopping mechanism terminates training when the validation set loss no longer decreases.

[0058] From the formula The reconstruction error between the vibration characteristic vector and the surrounding rock characteristic vector is analyzed, where, For reconstruction error, For vibration characteristic vectors, The characteristic vector of the surrounding rock;

[0059] Calculate the partial derivatives of the rock mass parameter correction with respect to the weights and biases, as well as the partial derivatives of the reconstruction error with respect to the learning matrix. Multiply each partial derivative with the learning rate to generate the parameter update.

[0060] Subtract the corresponding parameter update amount from the current weights, biases, and learning matrices respectively to obtain new weights, biases, and learning matrices, and then feed them back into the spatiotemporal fusion network architecture to overwrite the original parameters;

[0061] The sum of the current engineering parameters and the corresponding rock mass parameter corrections is used as the new engineering parameters. The basic quality index of the rock mass is recalculated to generate a new tunnel surrounding rock grade.

[0062] Using the above technical solution, after detecting abnormal vibration spectrum data, the following sequential steps can be followed to correct the spatiotemporal fusion network architecture parameters and update the tunnel surrounding rock classification prediction results: First, retrieve historical construction data corresponding to the current preliminary predicted surrounding rock level from the dynamic threshold library to obtain the historical average energy proportion and peak frequency of the corresponding frequency sub-segment under that level. Then, collect real-time abnormal vibration data to obtain the real-time energy proportion and real-time peak frequency of the same frequency sub-segment. Subtract the real-time values ​​from the historical averages to obtain the energy deviation and frequency offset. Next, construct a vibration feature vector and a surrounding rock feature vector. Use the calculated energy deviation and frequency offset to form the vibration feature vector. Simultaneously, extract the current surrounding rock feature vector through the spatiotemporal fusion network. This vector is a comprehensive feature representation obtained after fusing multimodal data. Finally, concatenate the vibration feature vector and the surrounding rock feature vector into a joint vector and input it into the pre-trained LSTM. The network model outputs corrections for rock mass parameters, including corrections for rock mass quality indices, elastic modulus, and joint density. Next, the specific values ​​of the surrounding rock feature vector, vibration feature vector, and learnable matrix are determined. The product of the learnable matrix and the surrounding rock feature vector is calculated using a formula. This product is then subtracted from the vibration feature vector to obtain the error vector, from which the reconstruction error is calculated. Based on the learning rate, the parameter updates for the weight matrix, bias, and learning matrix are calculated via backpropagation. Finally, the corresponding update amounts are subtracted from the current weight, bias, and learning matrix to obtain the final result. The new parameters are fed back to the spatiotemporal fusion network architecture to overwrite the original parameters. Finally, the engineering parameters obtained before correction based on the preliminary prediction of the surrounding rock grade are obtained, including rock mass quality indicators, elastic modulus and joint density. These are added to the corresponding rock mass parameter correction amounts to obtain the corrected engineering parameters. Based on the corrected engineering parameters and the industry standard grading formula, the original basic rock mass quality indicators are first calculated, and then the corrected basic rock mass quality indicators are calculated. The corrected basic rock mass quality indicators are compared with the rock mass grading standard table to determine the new tunnel surrounding rock grade.

[0063] To further clarify, the above process achieves continuous optimization of surrounding rock classification through a dynamic feedback mechanism: when abnormal vibration spectrum data is detected, the system quickly identifies the differences and corrects the model based on the energy proportion and peak frequency anomalies of the tunnel boring machine vibration signal, ensuring that the output results are highly consistent with the on-site geological conditions. This mechanism is robust; even if the vibration data is subject to short-term fluctuations due to accidental interference, it only triggers small-amplitude parameter adjustments, preventing the model from deviating from a reasonable range and ensuring the smoothness of the prediction results. As construction progresses, through multiple progressive corrections, the model parameters will gradually converge to the true geological characteristics. In addition, the dynamic threshold library built by the system will continuously accumulate historical data, enabling the anomaly judgment criteria to adapt to the vibration characteristics of different geological units, effectively avoiding misjudgment problems caused by fixed thresholds, and ultimately forming a closed-loop prediction system with self-learning capabilities.

[0064] As one embodiment of the present invention, the method for obtaining historical construction data specifically includes:

[0065] Construct a historical construction database, divide data windows according to the construction progress of construction machinery, and store the multimodal data, vibration spectrum data and engineering parameters of the current data window;

[0066] All data windows are grouped according to the similarity of engineering parameters using the DBSCAN clustering algorithm to generate geological units;

[0067] Based on whether the vibration spectrum data is abnormal, all data windows are divided into normal windows and abnormal windows.

[0068] For each geological unit, normal and abnormal windows are extracted as historical construction data according to a dynamic ratio;

[0069] It should be noted that the dynamic ratio is set as follows: within a geological unit, if the proportion of abnormal windows is >15%, the number of normal windows = min(5, total number of normal windows × 15%), and the number of sampled abnormal windows = min(3, total number of abnormal windows × 30%). Otherwise, all abnormal windows are retained, and 5 normal windows are randomly selected. If the number is less than 5, the maximum value is taken. The method for extracting historical construction data is as follows: the surrounding rock grade of all geological units is calculated using engineering parameters, and the data is extracted by comparison with the current surrounding rock grade.

[0070] This scheme constructs a closed-loop technology chain of "fusion-trigger-correction" based on the Transformer cross-modal attention mechanism: First, through heterogeneous data alignment and joint characterization, multimodal data are fused into a spatiotemporally unified surrounding rock feature vector; second, the frequency segment energy ratio and peak frequency analysis associated with historical surrounding rock grades are used to replace static threshold judgment to ensure the matching of abnormal vibration spectrum data with the current geological state; finally, by correcting the underlying parameters of the spatiotemporal fusion network architecture, including the learning matrix, weights, and biases, the static exploration data and real-time disturbance signals form an inseparable synergistic effect, thereby improving the adaptability and accuracy of surrounding rock classification prediction.

[0071] See Figure 4 As shown, this scheme proposes a tunnel surrounding rock intelligent classification and prediction system based on multimodal spatiotemporal fusion, used to implement the aforementioned tunnel surrounding rock intelligent classification and prediction method based on multimodal spatiotemporal fusion, including:

[0072] A multimodal data acquisition module is used to acquire time-stamped multimodal data of the surrounding rock, including ground-penetrating radar waveform data, drill core images, and three-dimensional laser point cloud data.

[0073] The spatiotemporal fusion and hierarchical prediction module is used to extract heterogeneous features and align them spatiotemporally from multimodal data using a spatiotemporal fusion network architecture, combined with a Transformer-based cross-modal attention mechanism, to generate surrounding rock feature data and perform hierarchical prediction of tunnel surrounding rock.

[0074] A vibration spectrum data acquisition module is used to acquire vibration spectrum data of construction machinery in real time based on computing nodes deployed at the edge of the tunnel construction section.

[0075] The learning and parameter correction module is used to establish an online learning mechanism for construction disturbance feedback. When the vibration spectrum data is abnormal, it combines the surrounding rock characteristic data to correct the parameters in the spatial fusion network architecture and update the tunnel surrounding rock classification prediction results.

[0076] The spatiotemporal fusion and hierarchical prediction module specifically includes:

[0077] The multimodal feature extraction unit is used to extract the temporal features of ground-penetrating radar waveform data, the convolutional neural network features of drill core images, and the point cloud features of 3D laser point cloud data.

[0078] The Transformer cross-modal fusion unit is used to align the encoded sequence data according to time encoding, concatenate them into a joint matrix and input it into the Transformer architecture, and output the global feature vector of the multimodal data;

[0079] The gated weighted fusion unit is used to initialize the learning matrix with Xavier and fuse the global feature vectors using the gated weighted model to generate the surrounding rock feature vector.

[0080] The decoding and classification prediction unit is used to input the surrounding rock feature vector into the decoding layer of the spatiotemporal fusion network architecture, map it to generate engineering parameters, and perform intelligent classification prediction of tunnel surrounding rock.

[0081] The vibration spectrum data acquisition module specifically includes:

[0082] A distributed computing node deployment unit is used to deploy computing nodes at the edge of the tunnel construction section.

[0083] The vibration signal real-time acquisition unit is used to acquire vibration spectrum data of construction machinery in real time.

[0084] The learning and parameter correction module specifically includes:

[0085] The vibration characteristic analysis unit is used to monitor the vibration spectrum data transmitted by the computing node in real time, divide it into preset frequency sub-segments, and extract its characteristic parameters.

[0086] The dynamic threshold library management unit is used to statistically analyze the distribution patterns of vibration spectrum characteristic parameters based on historical construction data and grouped by tunnel surrounding rock grade, and generate a dynamic threshold library corresponding to each surrounding rock grade.

[0087] The LSTM model parameter correction unit is used to concatenate the vibration feature vector and the surrounding rock feature vector into a joint vector, which is then input into the pre-trained LSTM network model and outputs the rock mass parameter correction amount.

[0088] The network architecture parameter update unit is used to calculate the partial derivatives of the rock mass parameter correction with respect to the weights and biases, as well as the partial derivatives of the reconstruction error with respect to the learning matrix. The partial derivatives are multiplied by the learning rate to generate the parameter update.

[0089] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for intelligent classification and prediction of tunnel surrounding rock based on multimodal spatiotemporal fusion, characterized in that, The method includes: Acquire time-stamped multimodal data of the surrounding rock, including ground-penetrating radar waveform data, drill core images, and three-dimensional laser point cloud data; A spatiotemporal fusion network architecture is adopted, combined with a Transformer-based cross-modal attention mechanism, to extract heterogeneous features and align them spatiotemporally from multimodal data, generate surrounding rock feature data, and perform tunnel surrounding rock classification prediction. Based on the computing nodes deployed at the edge of the tunnel construction section, the vibration spectrum data of the construction machinery is collected in real time. An online learning mechanism for construction disturbance feedback is established. When the vibration spectrum data is abnormal, the parameters in the spatiotemporal fusion network architecture are corrected by combining the surrounding rock characteristic data, and the tunnel surrounding rock classification prediction results are updated. The online learning mechanism for establishing construction disturbance feedback specifically includes: The vibration spectrum data transmitted by the computing node is monitored in real time, divided into preset frequency sub-segments, and its characteristic parameters are extracted, including energy ratio and peak frequency. Based on historical construction data, the distribution patterns of vibration spectrum characteristic parameters are statistically analyzed according to the tunnel surrounding rock grade, and a dynamic threshold library corresponding to each surrounding rock grade is generated. Based on the current surrounding rock feature vector, the feature threshold range corresponding to the surrounding rock grade is matched from the dynamic threshold library. When any feature parameter exceeds the corresponding feature threshold range, it is judged as abnormal data. When the vibration spectrum data shows anomalies, the parameters in the spatiotemporal fusion network architecture are corrected in conjunction with the surrounding rock characteristic data, and the tunnel surrounding rock classification prediction results are updated, specifically including: Based on the current surrounding rock grade, extract the historical average values ​​of energy proportion and peak frequency from the historical construction data; Based on the abnormal vibration spectrum data, the difference between its energy proportion and peak frequency and the corresponding historical mean is recorded as energy deviation and frequency offset, respectively, and a vibration feature vector is constructed using energy deviation and frequency offset. The vibration feature vector and the surrounding rock feature vector are concatenated into a joint vector, which is then input into a pre-trained LSTM network model to output rock mass parameter corrections. The rock mass parameter corrections include rock mass quality index corrections, elastic modulus corrections, and joint density corrections. From the formula The reconstruction error between the vibration characteristic vector and the surrounding rock characteristic vector is analyzed, where, For reconstruction error, For vibration characteristic vectors, The characteristic vector of the surrounding rock; Calculate the partial derivatives of the rock mass parameter correction with respect to the weights and biases, as well as the partial derivatives of the reconstruction error with respect to the learning matrix. Multiply each partial derivative with the learning rate to generate the parameter update. Subtract the corresponding parameter update amount from the current weights, biases, and learning matrices respectively to obtain new weights, biases, and learning matrices, and then feed them back into the spatiotemporal fusion network architecture to overwrite the original parameters; The sum of the current engineering parameters and the corresponding rock mass parameter corrections is used as the new engineering parameters. The basic quality index of the rock mass is recalculated to generate a new tunnel surrounding rock grade.

2. The method according to claim 1, characterized in that, The method employs a spatiotemporal fusion network architecture, combined with a Transformer-based cross-modal attention mechanism, to extract heterogeneous features from multimodal data and align them spatiotemporally, generating surrounding rock feature data and performing intelligent classification and prediction of tunnel surrounding rock. Specifically, this includes: The extraction layer of the spatiotemporal fusion network architecture includes a temporal feature extraction sub-channel for geological radar waveform data, a convolutional neural network feature extraction sub-channel for drill core images, and a point cloud feature extraction sub-channel for three-dimensional laser point cloud data. Multimodal data is input into the extraction layer of the spatiotemporal fusion network architecture to generate corresponding sequence data. All sequence data are mapped to a unified dimension through a linear projection layer and synchronously associated with time encoding. The encoded sequence data are aligned according to time encoding, concatenated into a joint matrix, and input into a Transformer architecture. The output is a global feature vector of the multimodal data, which includes the ground-penetrating radar feature vector. Drilling core feature vectors and laser point cloud feature vector ; The learning matrix is ​​initialized using Xavier, and the global feature vectors are fused using a weighted model with a gating mechanism to generate the surrounding rock feature vector. The surrounding rock feature vector is input into the decoding layer of the spatiotemporal fusion network architecture, mapped to generate engineering parameters, and used for tunnel surrounding rock classification prediction. The weighted model of the gating mechanism is as follows: ； In the formula, The characteristic vector of the surrounding rock. This is the concatenated vector of all global feature vectors. For learnable matrices, Let i be the global feature vector corresponding to number i, where i = 1, 2, 3.

3. The method according to claim 2, characterized in that, The process of inputting the surrounding rock feature vector into the decoding layer of the spatiotemporal fusion network architecture, mapping it to generate engineering parameters, and performing intelligent classification prediction of tunnel surrounding rock specifically includes: The feature vectors of the surrounding rock are subjected to layer normalization, and then mapped to the engineering parameter space through a two-layer fully connected network in the decoding layer. The first fully connected layer uses a weight matrix. and bias For the characteristic vector of the surrounding rock A linear transformation is performed, followed by the ReLU activation function to extract higher-order nonlinear features. The second fully connected network then uses a weight matrix... and bias Perform a linear transformation on the higher-order nonlinear characteristics to generate an output vector; The output vector is split and processed according to the engineering adaptation requirements to obtain engineering parameters, including rock mass quality indicators, elastic modulus and joint density. By inputting engineering parameters into the industry standard grading formula, the basic quality indicators of the rock mass are calculated, and then compared with the rock mass grading standard table to obtain the tunnel surrounding rock grade.

4. The method according to claim 1, characterized in that, The specific methods for obtaining the historical construction data include: Construct a historical construction database, divide data windows according to the construction progress of construction machinery, and store the multimodal data, vibration spectrum data and engineering parameters of the current data window; All data windows are grouped according to the similarity of engineering parameters using the DBSCAN clustering algorithm to generate geological units; Based on whether the vibration spectrum data is abnormal, all data windows are divided into normal windows and abnormal windows. For each geological unit, normal and abnormal windows are extracted as historical construction data according to a dynamic ratio.

5. A tunnel surrounding rock intelligent classification and prediction system based on multimodal spatiotemporal fusion, characterized in that, The method for implementing the intelligent classification and prediction method for tunnel surrounding rock based on multimodal spatiotemporal fusion as described in any one of claims 1-4 includes: A multimodal data acquisition module is used to acquire time-stamped multimodal data of the surrounding rock, including ground-penetrating radar waveform data, drill core images, and three-dimensional laser point cloud data. The spatiotemporal fusion and hierarchical prediction module is used to extract heterogeneous features and align them spatiotemporally from multimodal data using a spatiotemporal fusion network architecture, combined with a Transformer-based cross-modal attention mechanism, to generate surrounding rock feature data and perform tunnel surrounding rock hierarchical prediction. A vibration spectrum data acquisition module is used to acquire vibration spectrum data of construction machinery in real time based on computing nodes deployed at the edge of the tunnel construction section. The learning and parameter correction module is used to establish an online learning mechanism for construction disturbance feedback. When the vibration spectrum data is abnormal, it combines the surrounding rock characteristic data to correct the parameters in the spatiotemporal fusion network architecture and update the tunnel surrounding rock classification prediction results.

6. The system according to claim 5, characterized in that, The spatiotemporal fusion and hierarchical prediction module specifically includes: The multimodal feature extraction unit is used to extract the temporal features of ground-penetrating radar waveform data, the convolutional neural network features of drill core images, and the point cloud features of 3D laser point cloud data. The Transformer cross-modal fusion unit is used to align the encoded sequence data according to time encoding, concatenate them into a joint matrix and input it into the Transformer architecture, and output the global feature vector of the multimodal data; The gated weighted fusion unit is used to initialize the learning matrix with Xavier and fuse the global feature vectors using the gated weighted model to generate the surrounding rock feature vector. The decoding and classification prediction unit is used to input the surrounding rock feature vector into the decoding layer of the spatiotemporal fusion network architecture, map it to generate engineering parameters, and perform intelligent classification prediction of tunnel surrounding rock.

7. The system according to claim 5, characterized in that, The vibration spectrum data acquisition module specifically includes: A distributed computing node deployment unit is used to deploy computing nodes at the edge of the tunnel construction section. The vibration signal real-time acquisition unit is used to acquire vibration spectrum data of construction machinery in real time.

8. The system according to claim 5, characterized in that, The learning and parameter correction module specifically includes: The vibration characteristic analysis unit is used to monitor the vibration spectrum data transmitted by the computing node in real time, divide it into preset frequency sub-segments, and extract its characteristic parameters. The dynamic threshold library management unit is used to statistically analyze the distribution patterns of vibration spectrum characteristic parameters based on historical construction data and grouped by tunnel surrounding rock grade, and generate a dynamic threshold library corresponding to each surrounding rock grade. The LSTM model parameter correction unit is used to concatenate the vibration feature vector and the surrounding rock feature vector into a joint vector, which is then input into the pre-trained LSTM network model and outputs the rock mass parameter correction amount. The network architecture parameter update unit is used to calculate the partial derivatives of the rock mass parameter correction with respect to the weights and biases, as well as the partial derivatives of the reconstruction error with respect to the learning matrix. The partial derivatives are multiplied by the learning rate to generate the parameter update.

Images