Tunnel surrounding rock grading dynamic identification system based on horizontal drilling multi-parameter fusion

Abstract

The application relates to the engineering technical field and discloses a tunnel surrounding rock grading dynamic identification system based on horizontal drilling multi-parameter fusion, which comprises a multi-source drilling data acquisition module, a feature extraction and fusion module, a surrounding rock grading intelligent identification module, a grading result dynamic visualization module, a model self-adaptive optimization module and a grading data management module. The drilling engineering parameters, drilling geophysical data and digital images of the working face are synchronously collected through the multi-source drilling data acquisition module, the drilling specific energy features, vibration frequency spectrum features, electromagnetic wave and seismic wave reflection features and fracture image features are deeply fused through the feature extraction and fusion module, and a unified high-dimensional fusion feature vector is generated, so that the feature representation one-sidedness caused by the dependence on traditional methods and single information sources can be solved, the comprehensiveness of surrounding rock state identification and the multi-dimensional information collaboration are guaranteed, and the objectivity of grading and the perception ability for complex geology are improved.

Description

Technical Field

[0001] This invention relates to the field of engineering technology, specifically to a dynamic identification system for classifying surrounding rock in tunnels based on multi-parameter fusion during horizontal drilling. Background Technology

[0002] Rock mass classification is a method in tunnel engineering to classify the stability of engineering rock masses based on the integrity and strength of the rock mass. This method provides a basis for selecting construction methods, designing support structures, managing projects, setting labor quotas, and determining material consumption standards by classifying rock masses with similar stability.

[0003] Currently, due to the hidden and complex geological conditions of tunnels, the traditional methods and single-parameter criteria relied upon for real-time classification of the surrounding rock in front of the tunnel face are difficult to integrate multi-dimensional information such as the mechanical response of the drilling process, geophysical waveforms and visual images. When only empirical formulas and static indicators are used, the classification results will deviate significantly from the actual geological conditions, and the accuracy of the support design basis cannot be guaranteed.

[0004] Therefore, a dynamic identification system for tunnel surrounding rock classification based on multi-parameter fusion of horizontal drilling is proposed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a dynamic identification system for tunnel surrounding rock classification based on multi-parameter fusion of horizontal drilling, which solves the problem mentioned in the background technology of the difficulty in fusing multi-dimensional information such as mechanical response, geophysical waveform and visual image of the drilling process.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a dynamic identification system for tunnel surrounding rock classification based on multi-parameter fusion of horizontal drilling, the system comprising: The multi-source data acquisition module for drilling is deployed on a horizontal drilling rig to simultaneously and in real time acquire drilling engineering parameters, drilling geophysical data, and digital images of the tunnel face during the advanced drilling process at the tunnel face. The feature extraction and fusion module is connected to each data acquisition unit in the multi-source drilling data acquisition module, receives the drilling engineering parameters, drilling geophysical data and digital images of the working face, extracts drilling specific energy features, vibration spectrum features, electromagnetic wave and seismic wave reflection features and fracture image features from them, and fuses the features into a unified high-dimensional fusion feature vector. The surrounding rock grading intelligent identification module is connected to the feature extraction and fusion module. It stores a pre-trained multimodal deep learning model, receives the high-dimensional fused feature vector, calls the multimodal deep learning model for processing, and outputs real-time, continuous and automated grading results of the surrounding rock in front of the tunnel face. The dynamic visualization module for grading results is connected to the intelligent identification module for surrounding rock grading. It receives the automated grading results and generates and displays a dynamic grading cloud map and curve showing the changes in the surrounding rock grading results along the tunnel axis based on the tunnel axis mileage information. The model adaptive optimization module is connected to the surrounding rock classification intelligent identification module and the multi-source drilling data acquisition module. Based on the newly added drilling data and the corresponding surrounding rock classification revealed by actual excavation, the module performs online incremental learning and parameter optimization on the multimodal deep learning model. The hierarchical data management module is connected to the multi-source drilling data acquisition module, the feature extraction and fusion module, the surrounding rock classification intelligent identification module, and the model adaptive optimization module, respectively, and stores and manages all acquired raw data, intermediate feature data during processing, model parameters, historical classification results, and optimization records.

[0007] Preferably, the multi-source drilling data acquisition module includes: The drilling engineering parameter acquisition unit is integrated into the drilling rig power head and propulsion mechanism, and collects and outputs real-time time-series data of drilling engineering parameters, including drilling pressure, rotational speed, torque, and drilling speed. The geophysical data acquisition unit is integrated inside the drill pipe and drill bit. During the drilling process, it synchronously excites and receives electromagnetic wave and seismic wave signals, and outputs geophysical data that characterizes the reflection features of the geological interface ahead. The tunnel face image acquisition unit is deployed in front of the tunnel face to acquire high-definition digital images of the tunnel face during drilling intervals and at specific mileage points; The data synchronization and preprocessing unit is connected to the drilling engineering parameter acquisition unit, the drilling geophysical data acquisition unit, and the working face image acquisition unit, respectively. It performs timestamp alignment, filtering and noise reduction, and format standardization preprocessing on the received multi-source heterogeneous data, and outputs the synchronized drilling engineering parameters, drilling geophysical data, and working face digital image to the feature extraction and fusion module.

[0008] Preferably, the feature extraction and fusion module includes: The engineering parameter feature extraction unit calculates and extracts drilling specific energy features from the drilling engineering parameters based on drilling power and drilling speed. Calculated by the following formula: ; in, Indicates drilling specific energy. Indicates drilling pressure. Indicates rotational speed. Indicates torque, This indicates the cross-sectional area of ​​the drill bit. Indicates drilling speed; The vibration spectrum features, including the dominant frequency and amplitude, are extracted from the vibration signals of the drilling parameters using Fast Fourier Transform. The geophysical data feature extraction unit performs reflection wave analysis on the drilling geophysical data to extract electromagnetic wave and seismic wave reflection features, including reflection wave amplitude, phase, and travel time. The digital image feature extraction unit performs edge detection and pattern recognition on the digital image of the working face, and extracts fracture image features including fracture density, fracture dip angle and rock mass block size; The multi-source feature fusion unit is connected to the engineering parameter feature extraction unit, the geophysical data feature extraction unit, and the digital image feature extraction unit, respectively. It normalizes and vectorizes the drilling energy characteristics, the vibration spectrum characteristics, the electromagnetic wave and seismic wave reflection characteristics, and the fracture image characteristics, and generates the unified high-dimensional fusion feature vector through a weighted fusion method of feature splicing.

[0009] Preferably, the surrounding rock classification intelligent identification module includes: A hierarchical model storage unit stores the pre-trained multimodal deep learning model, which is a network model built on a convolutional neural network architecture and obtained through supervised training using drilling data from historical tunnel projects and the surrounding rock grade labels verified by the final excavation. A real-time hierarchical inference unit, connected to the hierarchical model storage unit, receives the high-dimensional fused feature vector, calls the multimodal deep learning model for forward inference calculation, and outputs the probability distribution of surrounding rock grades corresponding to the BQ grading system. It can be expressed by the following formula: ; in, Represents a probability distribution vector. This represents the high-dimensional fusion feature vector input to the multimodal deep learning model. This represents the set of all trainable parameters of the multimodal deep learning model. This represents the mapping function from feature vectors to probability distribution vectors defined in the multimodal deep learning model; The hierarchical decision-making unit, connected to the real-time hierarchical reasoning unit, determines the final automated hierarchical result of the surrounding rock according to the probability distribution of the surrounding rock level and according to a set threshold, and adds a location tag based on the drilling mileage.

[0010] Preferably, the dynamic visualization module for the grading results includes: The spatiotemporal correlation unit receives the automated classification results and their accompanying mileage location tags, and correlates the classification results with the spatial coordinates in the tunnel's three-dimensional design model. The dynamic rendering unit is connected to the spatiotemporal association unit. Based on the associated classification results, it dynamically renders areas of different rock grades along the tunnel excavation direction with a fill pattern to generate the dynamic classification cloud map. The curve drawing unit, connected to the spatiotemporal correlation unit, uses the tunnel mileage as the horizontal axis and the surrounding rock grade value as the vertical axis to draw and update the dynamic grading curve of the surrounding rock grading result along the tunnel axis in real time. The visualization interface unit integrates the outputs of the dynamic rendering unit and the curve drawing unit, synchronously displays the dynamic hierarchical cloud map and the dynamic hierarchical curve on the display device, and provides zoom, pan, and history playback interactive functions.

[0011] Preferably, the model adaptive optimization module includes: The verification data acquisition unit collects the surrounding rock grade of the actual excavated surface after tunnel excavation, which serves as a verification label for model optimization. It also pairs the data with the corresponding mileage of advanced drilling data to form model optimization sample pairs. An incremental learning engine, connected to the verification data acquisition unit and the surrounding rock classification intelligent identification module, inputs the model optimization sample pairs into the multimodal deep learning model, uses the verification label as a supervision signal, and employs a mini-batch gradient descent algorithm to fine-tune and optimize the network parameters of the multimodal deep learning model. The model performance evaluation unit is connected to the incremental learning engine. After the model is optimized, it uses an independent test dataset to evaluate the classification accuracy and recall performance metrics of the optimized model. When the performance improvement exceeds a set threshold, it triggers a model update command. The model hot update unit is connected to both the model performance evaluation unit and the surrounding rock classification intelligent identification module. Upon receiving the model update instruction, it seamlessly replaces the original model parameters stored in the surrounding rock classification intelligent identification module with the optimized multimodal deep learning model parameters, thereby achieving hot updating of the model.

[0012] Preferably, the hierarchical data management module includes: The raw database stores the raw data packets collected by the multi-source drilling data acquisition module, categorized by tunnel engineering, borehole number, and mileage interval. The raw data packets include unprocessed drilling engineering parameters, drilling geophysical data, and digital images of the tunnel face. The feature and model library stores the high-dimensional fused feature vector generated by the feature extraction and fusion module, all version parameters of the multimodal deep learning model in the surrounding rock classification intelligent identification module, and each optimization record generated by the model adaptive optimization module. The classification result library archives and stores all historical and current automated classification results output by the surrounding rock classification intelligent identification module in chronological order and spatial mileage. The data service interface provides standardized data query, retrieval, and calling services for the feature extraction and fusion module, the surrounding rock classification intelligent identification module, the model adaptive optimization module, and the classification result dynamic visualization module.

[0013] Preferably, the multimodal deep learning model stored in the hierarchical model storage unit adopts an architecture of a multi-branch feature encoder and a feature interaction fusion unit; The multi-branch feature encoder includes a first feature coding sub-network, a second feature coding sub-network, and a third feature coding sub-network, which respectively perform deep nonlinear feature coding on the engineering parameter feature sub-vectors, geophysical data feature sub-vectors, and image feature sub-vectors separated from the high-dimensional fused feature vectors; The feature interaction fusion unit is connected to the output of each sub-network of the multi-branch feature encoder. Through tensor splicing fusion, it interacts and fuses the deep features output by the first feature encoding sub-network, the second feature encoding sub-network, and the third feature encoding sub-network to generate the fused feature representation for the final hierarchical decision.

[0014] Preferably, the multi-source feature fusion unit is specifically configured to perform cross-modal feature interaction fusion, which includes: The feature interaction attention subunit receives the drilling specific energy feature, vibration spectrum feature, electromagnetic wave and seismic wave reflection feature, and fracture image feature output by the engineering parameter feature extraction unit, the geophysical data feature extraction unit, and the digital image feature extraction unit, respectively, and calculates the correlation weight between different modal features through a cross-attention mechanism. The feature interaction attention subunit first maps each modal feature into a query vector, a key vector, and a value vector, respectively. Then it calculates the attention weights of the engineering parameter modal features to the geophysical data modal features, the attention weights of the engineering parameter modal features to the image modal features, and the attention weights of the geophysical data modal features to the image modal features, generating a series of context feature vectors enhanced with cross-modal information. The feature aggregation subunit is connected to the feature interaction attention subunit. It concatenates the context feature vectors generated by the feature interaction attention subunit with the original modal features, and performs dimensionality reduction and fusion through a multilayer perceptron to finally output the unified high-dimensional fused feature vector.

[0015] Preferably, the hierarchical data management module further includes: The automatic generation unit for the grading report is connected to the grading result library and the data service interface. According to user instructions and scheduled tasks, it extracts all grading results within a specified time range and mileage interval from the grading result library, and combines them with the corresponding original data snapshots extracted from the original database to automatically generate a structured grading report that includes the statistical distribution of surrounding rock grades, trend analysis of changes, and location information of key adverse geological sections. The data encryption and secure storage unit encrypts and stores all data in the original database, the feature and model library, and the hierarchical result library, and sets a role-based access control policy to authenticate and log data access operations performed through the data service interface.

[0016] Compared with existing technologies, this invention provides a dynamic identification system for tunnel surrounding rock classification based on multi-parameter fusion during horizontal drilling, which has the following advantages: 1. In this invention, when performing dynamic identification of the surrounding rock grading in front of the tunnel face, the multi-source drilling data acquisition module simultaneously collects drilling engineering parameters, drilling geophysical data, and digital images of the tunnel face. The feature extraction and fusion module then deeply fuses drilling specific energy characteristics, vibration spectrum characteristics, electromagnetic wave and seismic wave reflection characteristics, and fracture image characteristics to generate a unified high-dimensional fused feature vector. This solves the problem of one-sided feature representation caused by relying on traditional methods and single information sources, ensuring the comprehensiveness of surrounding rock condition identification and the synergy of multi-dimensional information, thereby improving the objectivity of grading and the ability to perceive complex geology.

[0017] 2. In this invention, when performing dynamic identification of the surrounding rock grading ahead of the tunnel face, the intelligent identification module for surrounding rock grading calls a pre-trained multimodal deep learning model to process the high-dimensional fused feature vector. The multi-source feature fusion unit and feature interaction attention subunit in the feature extraction and fusion module are used to realize the interaction and weighted fusion of cross-modal features. This enables the system to perform deep information mining and complementary arbitration on multi-source heterogeneous data, reducing the risk of misjudgment caused by data contradictions and omission of key features, and ensuring the intelligence of the grading reasoning process and the consistency of the results.

[0018] 3. In this invention, when performing dynamic identification of the surrounding rock classification in front of the tunnel face, the model adaptive optimization module performs online incremental learning and parameter optimization of the multimodal deep learning model based on the actual rock classification revealed by excavation. The classification data management module continuously accumulates case data, enabling the system to continuously iterate and evolve using feedback from engineering practice, solving the problem of insufficient generalization ability of static models. This ensures the accuracy and reliability of the system under different geological conditions in the long term, and achieves adaptive improvement of the intelligent identification capability of surrounding rock classification. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion of the present invention; Figure 2 This is a schematic diagram of the multi-source drilling data acquisition module of the present invention; Figure 3 This is a diagram illustrating the architecture of the feature extraction and fusion module of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1-3 This tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion includes: The multi-source data acquisition module for drilling is deployed on a horizontal drilling rig to simultaneously and in real time acquire drilling engineering parameters, drilling geophysical data, and digital images of the tunnel face during the advanced drilling process at the tunnel face. The feature extraction and fusion module is connected to each data acquisition unit in the multi-source drilling data acquisition module. It receives drilling engineering parameters, drilling geophysical data and digital images of the tunnel face, extracts drilling specific energy features, vibration spectrum features, electromagnetic wave and seismic wave reflection features and fracture image features, and fuses the features into a unified high-dimensional fusion feature vector. The surrounding rock classification intelligent identification module is connected to the feature extraction and fusion module. It internally stores a pre-trained multimodal deep learning model, receives high-dimensional fused feature vectors, calls the multimodal deep learning model for processing, and outputs real-time, continuous, and automated classification results of the surrounding rock in front of the tunnel face. The dynamic visualization module for grading results is connected to the intelligent identification module for surrounding rock grading. It receives automated grading results and generates and displays dynamic grading cloud maps and curves showing the changes in surrounding rock grading results along the tunnel axis based on tunnel axis mileage information. The model adaptive optimization module is connected to the surrounding rock classification intelligent identification module and the multi-source drilling data acquisition module. Based on the newly added drilling data and the corresponding surrounding rock classification revealed by actual excavation, the multimodal deep learning model is subjected to online incremental learning and parameter optimization. The hierarchical data management module is connected to the multi-source drilling data acquisition module, feature extraction and fusion module, surrounding rock classification intelligent identification module, and model adaptive optimization module, respectively. It stores and manages all acquired raw data, intermediate feature data during processing, model parameters, historical classification results, and optimization records.

[0022] The multi-source drilling data acquisition module includes: The drilling engineering parameter acquisition unit is integrated into the drilling rig power head and propulsion mechanism, and collects and outputs real-time time-series data of drilling engineering parameters, including drilling pressure, rotational speed, torque, and drilling speed. The geophysical data acquisition unit is integrated inside the drill pipe and drill bit. During the drilling process, it synchronously excites and receives electromagnetic wave and seismic wave signals, and outputs geophysical data that characterizes the reflection features of the geological interface ahead. The tunnel face image acquisition unit is deployed in front of the tunnel face to acquire high-definition digital images of the tunnel face during drilling intervals and at specific mileage points; The data synchronization and preprocessing unit is connected to the drilling engineering parameter acquisition unit, the drilling geophysical data acquisition unit, and the face image acquisition unit, respectively. It performs timestamp alignment, filtering and noise reduction, and format standardization preprocessing on the received multi-source heterogeneous data. Specifically, timestamp alignment involves: Based on a unified system reference clock, the time offset of each acquisition unit is calculated and corrected, and the aligned timestamps are then generated. It is obtained by calculation using the following formula: ; in, Indicates the first The corrected timestamp of each acquisition unit Indicates the first The original timestamps recorded locally by each acquisition unit. Indicates the system reference clock timestamp at the same moment. For indexing; Filtering and noise reduction: For the vibration signal in the drilling procedure data and the waveform signal time series data in the drilling geophysical data, a time-frequency domain filter is used for noise reduction. Let the original signal be... ,in To represent time, design a frequency response function as follows: The filter, where The angular frequency is the key factor, and this filter is designed to remove unwanted components in specific frequency bands of a signal. The filtering process can be represented as a convolution operation in the time domain: ; in, This represents the signal after filtering and noise reduction. This represents a linear convolution operation. It is the frequency response function. The corresponding time-domain impulse response; This step outputs This refers to the effective signal after filtering out the main noise. Standardized format: To eliminate the influence of data features from different sources and with different dimensions on model training and inference, the filtered data features are normalized. Let the original set of values ​​for a certain feature on the samples be... ,in Indicates the total number of original values; The min-max normalization method is used to linearly map the original eigenvalues ​​to a uniform numerical range. The calculation formula is as follows:

[0023] in, Represents the standardized first 1 eigenvalue, Represents the original first 1 eigenvalue, and Representing the original eigenvalue sets respectively The minimum and maximum values ​​in and These are the lower and upper limits within the preset target range, respectively. For indexing; After this processing, the numerical range of all features is unified to... This accelerates model convergence and improves stability; Through the above two specific operations of "filtering and noise reduction" and "format standardization", the original multi-source heterogeneous data is transformed into a clean, scale-consistent standard format data stream, laying the foundation for subsequent feature extraction and fusion. It outputs the synchronized drilling engineering parameters, drilling geophysical data, and digital images of the working face to the feature extraction and fusion module.

[0024] The feature extraction and fusion module includes: The engineering parameter feature extraction unit calculates and extracts drilling specific energy features from the drilling engineering parameters based on drilling power and drilling speed. Calculated by the following formula: ; in, Indicates drilling specific energy. Indicates drilling pressure. Indicates rotational speed. Indicates torque, This indicates the cross-sectional area of ​​the drill bit. Indicates drilling speed; The vibration spectrum features, including the dominant frequency and amplitude, were extracted from the vibration signals of the drilling parameters using Fast Fourier Transform. The specific steps are as follows: Signal sampling and preprocessing: Continuous vibration analog signals acquired from sensors Perform sampling at equal time intervals, where Representing time, after sampling, a discrete time-domain vibration signal sequence is obtained. ,in The sampling point number, This represents the total number of sampling points. Indicates the first The amplitude of the vibration signal at each sampling point; Spectrum transformation: for discrete-time signal sequences The Fast Fourier Transform (FFT) is applied to transform it from the time domain to the frequency domain to obtain its spectral information. The FFT calculation formula is as follows: ; in, The first frequency domain signal represents the second frequency domain signal. One portion, For indexing, It is the time-domain sampling point number. This represents the total number of sampling points. It is the imaginary unit. It is the twitch factor of the discrete Fourier transform.

[0025] Spectral feature extraction: Main frequency extraction: Within the main frequency band, search for the spectrum. Frequency domain index corresponding to the maximum value Then, through the formula The dominant frequency of the vibration signal was calculated. The main frequency reflects the frequency component in the signal where energy is most concentrated; Amplitude extraction: Main frequency The corresponding amplitude is It is a direct physical quantity characterizing the vibration intensity of this frequency component; Through the above specific steps, key spectral features such as the dominant frequency and amplitude can be extracted from the original vibration signal efficiently and accurately, providing important input information for subsequent multi-source feature fusion and surrounding rock classification identification; The geophysical data feature extraction unit performs reflection wave analysis on the geophysical data while drilling, extracting electromagnetic and seismic wave reflection features, including reflected wave amplitude, phase, and travel time. The specific process is as follows: Reflection wave identification and parameter extraction: First, the reflected wave groups representing different stratigraphic interfaces are identified from the received time-domain waveform signal. For each independent reflected wave event, its amplitude is directly measured from the waveform. Phase and the time from excitation to reception ; Calculation of reflection coefficient: Reflection coefficient It is a key physical quantity characterizing the difference in wave impedance and determining the lithological interface. Its calculation formula is: ; in, Represents the reflection coefficient. These represent the wave impedances of the upper and lower strata at the interface, respectively. It is the product of density and wave speed. These represent the densities of the upper and lower strata above and below the interface, respectively. These represent the wave propagation speeds above and below the interface in the underlying strata, respectively. In practice, and The precise value of the reflectance is usually unknown; therefore, the reflectance coefficient... Typically, this is achieved by extracting the amplitude of the reflected wave. With the known reference amplitude To make an estimate, that is Ultimately, the extracted feature vector contains , and ; The digital image feature extraction unit performs edge detection and pattern recognition on the digital image of the working face, and extracts fracture image features including fracture density, fracture dip angle and rock mass block size; The multi-source feature fusion unit is connected to the engineering parameter feature extraction unit, the geophysical data feature extraction unit, and the digital image feature extraction unit, respectively. It normalizes and vectorizes the drilling energy specificity feature, vibration spectrum feature, electromagnetic wave and seismic wave reflection feature, and fracture image feature. The normalization process uses the Z-score standardization method, and its calculation formula is as follows: ; in, Represents the normalized eigenvalues. Represents the original feature values. This represents the mean of the feature on the training dataset. This represents the standard deviation of the feature on the training dataset; A unified high-dimensional fusion feature vector is generated through a weighted fusion method of feature concatenation.

[0026] The intelligent identification module for surrounding rock classification includes: The hierarchical model storage unit stores pre-trained multimodal deep learning models. These models are constructed based on a convolutional neural network architecture and are obtained through supervised training using drilling data from historical tunnel projects and rock grade labels verified during final excavation. The real-time hierarchical inference unit, connected to the hierarchical model storage unit, receives high-dimensional fused feature vectors and calls a multimodal deep learning model for forward inference calculation, outputting the probability distribution of surrounding rock grades corresponding to the BQ grading system. It can be expressed by the following formula: ; in, Represents a probability distribution vector. This represents the high-dimensional fused feature vector input into a multimodal deep learning model. This represents the set of all trainable parameters for a multimodal deep learning model. This represents the mapping function from feature vectors to probability distribution vectors defined in a multimodal deep learning model; The hierarchical decision-making unit, connected to the real-time hierarchical reasoning unit, determines the final automated hierarchical classification result of the surrounding rock based on the probability distribution of the surrounding rock grade and according to a set threshold. Given by the decision function: ; in, This indicates the final rock mass grade determined by the system. Indicates taking the probability value The highest level As output; It also includes location labels based on drilling mileage.

[0027] The dynamic visualization module for grading results includes: The spatiotemporal correlation unit receives the automated classification results and their accompanying mileage location labels, and correlates the classification results with the spatial coordinates in the tunnel's 3D design model. This correlation is achieved through a mapping function. accomplish: ; in, This represents the spatial coordinates in the coordinate system of the tunnel's three-dimensional design model. This indicates the tunnel axis mileage attached to the automated classification results. This represents the mapping relationship from mileage to three-dimensional spatial coordinates; The dynamic rendering unit is connected to the spatiotemporal association unit. Based on the associated classification results, it dynamically renders areas of different rock grades along the tunnel excavation direction with a fill pattern to generate a dynamic classification cloud map. The curve plotting unit, connected to the spatiotemporal correlation unit, plots and updates a dynamic grading curve along the tunnel axis with the tunnel mileage as the horizontal axis and the surrounding rock grade value as the vertical axis. The visualization interface unit integrates the outputs of the dynamic rendering unit and the curve drawing unit, synchronously displaying dynamic hierarchical cloud maps and dynamic hierarchical curves on the display device, and providing interactive functions such as zooming, panning, and history rewinding.

[0028] The model adaptive optimization module includes: The verification data acquisition unit collects the surrounding rock grade of the actual excavated surface after tunnel excavation, which serves as a verification label for model optimization. It also pairs the data with the corresponding mileage of advanced drilling data to form model optimization sample pairs. The incremental learning engine, connected to the validation data acquisition unit and the intelligent identification module for surrounding rock grading, inputs optimized sample pairs into the multimodal deep learning model. Using validation labels as supervision signals, it employs a mini-batch gradient descent algorithm to fine-tune and optimize the network parameters of the multimodal deep learning model. Its workflow and core computational steps are as follows: Model and Data Preparation: The incremental learning engine receives model optimization sample pairs from the validation data acquisition unit. Let the multimodal deep learning model be a function. ,in This represents the set of all trainable parameters of a multimodal deep learning model, which is the set of variables to be optimized. A model's optimization sample pair contains a set of input features. That is, the high-dimensional fused feature vector from the feature extraction and fusion module, and its corresponding verification label. ; Forward propagation and loss calculation: Randomly sampled from the model optimization sample set Each sample pair constitutes a mini-batch. ,in To determine the batch size, this small batch of data... Input Model Perform forward propagation to calculate the model's predicted output. Predicted output Typically, this is a probability distribution vector, followed by the calculation of the predicted output. With real labels The difference between them, which is determined by the loss function For classification tasks, quantization often employs the cross-entropy loss function, which is used in mini-batch processing. Average loss The calculation formula is: ; in, This represents the average loss for small batches. For indexing; Backpropagation and gradient calculation: The average loss is calculated using the backpropagation algorithm. About all model parameters gradient Each component of the gradient vector indicates the change in the loss function when the corresponding parameter increases by a minimum unit. The rate of change of , and the gradient calculation formula is: ; in, This represents the average loss with respect to the model parameters. The gradient; Gradient descent: based on the calculated gradient The model parameters were adjusted according to the mini-batch gradient descent rule. Iterative updates are performed to adjust the model along the direction of the decreasing loss function, thereby reducing the model's prediction error. The parameter update formula is as follows: ; in, This indicates an assignment operation. Indicates the learning rate. Indicates the parameter update amount; Iterative optimization: Repeat the above steps, traversing all mini-batch data in the model optimization sample set to complete one complete training "round". Based on the feedback from the model performance evaluation unit, multiple rounds of iteration can be performed until the model's performance on the validation set meets the preset optimization objective and stopping condition, thereby completing the optimization of the multimodal deep learning model parameters. Fine-tuning and optimization.

[0029] The model performance evaluation unit, connected to the incremental learning engine, evaluates the classification accuracy and recall performance metrics of the optimized model using an independent test dataset after model optimization. Classification accuracy is the most important metric. The calculation formula is: ; in, For the actual number of cases, The number of true negative examples. The number of false positives The number of false negatives; And when the performance improvement exceeds a set threshold, a model update instruction is triggered; The model hot update unit is connected to the model performance evaluation unit and the surrounding rock classification intelligent identification module. After receiving the model update instruction, it seamlessly replaces the original model parameters stored in the surrounding rock classification intelligent identification module with the optimized multimodal deep learning model parameters, thereby realizing the hot update of the model.

[0030] The hierarchical data management module includes: The raw database stores raw data packets collected by the multi-source drilling data acquisition module, categorized by tunnel engineering, borehole number, and mileage interval. The raw data packets include unprocessed drilling engineering parameters, drilling geophysical data, and digital images of the tunnel face. The feature and model library stores the high-dimensional fused feature vectors generated by the feature extraction and fusion module, all version parameters of the multimodal deep learning model in the surrounding rock classification intelligent identification module, and the optimization records generated by the model adaptive optimization module. The classification results library archives and stores all historical and current automated classification results output by the surrounding rock classification intelligent identification module in chronological order and spatial mileage. The data service interface provides standardized data query, retrieval, and access services for the feature extraction and fusion module, the intelligent identification module for surrounding rock grading, the model adaptive optimization module, and the dynamic visualization module for grading results. Its query logic can be expressed as follows: ; in, This represents the query result dataset. This represents a specific data record in the database. This represents the union of the original database, the feature and model library, and the hierarchical results library. This represents the conditional decision function defined by the query request; When data records The function evaluates to true when all query conditions are met.

[0031] The multimodal deep learning models stored in the hierarchical model storage unit adopt an architecture of multi-branch feature encoder and feature interaction fusion unit; The multi-branch feature encoder includes a first feature coding sub-network, a second feature coding sub-network, and a third feature coding sub-network, which respectively perform deep nonlinear feature coding on the engineering parameter feature sub-vectors, geophysical data feature sub-vectors, and image feature sub-vectors separated from the high-dimensional fused feature vectors; The feature interaction fusion unit is connected to the outputs of each sub-network of the multi-branch feature encoder. Through tensor concatenation fusion, it interacts and fuses the deep features output from the first, second, and third feature encoding sub-networks to generate the final hierarchical decision-making fused feature representation. The fusion process is represented as follows: ; in, This represents the feature representation after fusion. These represent the deep feature vectors output by the first, second, and third feature encoding subnetworks, respectively. This represents a tensor splicing operation performed along the feature dimension.

[0032] The multi-source feature fusion unit is specifically configured to perform cross-modal feature interaction fusion, which includes: The feature interaction attention subunit receives drilling specific energy features, vibration spectrum features, electromagnetic wave and seismic wave reflection features, and fracture image features output by the engineering parameter feature extraction unit, geophysical data feature extraction unit, and digital image feature extraction unit, respectively. It then calculates the correlation weights between different modal features through a cross-attention mechanism. For any two modal features... and Its attention weight matrix The calculation formula is: ; in, Representing modes Feature-mode The attention weight matrix of the features. Indicates modality The query matrix obtained by the characteristic linear transformation Indicates modality The key matrix obtained by the characteristic linear transformation, Let be the dimension of the key vector. Scaling factor This represents the row normalized exponential function. For indexing, It is a transpose operator; The feature interaction attention subunit first maps each modal feature into a query vector, a key vector, and a value vector, respectively. Then, it calculates the attention weights of the engineering parameter modal features on the geophysical data modal features, the attention weights of the engineering parameter modal features on the image modal features, and the attention weights of the geophysical data modal features on the image modal features, generating a series of context feature vectors enhanced with cross-modal information. The specific steps are as follows: Linear transformation generates query, key, and value vectors: Let the input feature vectors be: engineering parameter modal features Geophysical data modal characteristics Image modal features First, the feature vectors of each modality are linearly mapped to query vectors, key vectors, and value vectors respectively using trainable parameter matrices; For engineering parameter modes: ; in, The query vector represents the modalities of engineering parameters. Key vectors representing engineering parameter modes. A value vector representing the modalities of engineering parameters. This represents the modal feature vector of the input engineering parameters. , , These are the trainable query, key, and value weight matrices corresponding to the engineering parameter modes, respectively. Similarly, the same operation is performed on the geophysical data modes and image modes to obtain... and ; Calculate cross-modal attention weights: Calculate the attention distribution of a query vector from one modality to a key vector from another modality to compute the attention weights of engineering parameter modal features to geophysical data modal features. For example: ; in, This represents the attention weight matrix of engineering parameters on geophysical data. The query vector represents the modalities of engineering parameters. The transpose of the key vector representing the mode of geophysical data. Scaling factor It is the dimension of the key vector. This represents the row normalization exponential function; Similarly, the attention weights of engineering parameters on the image can be calculated. Attention weights of geophysical data to images : ; in, , These are the attention weight matrices for engineering parameters on the image and geophysical data on the image, respectively. Generate context feature vectors: Using the calculated attention weight matrix, the value vectors are weighted and summed to generate context feature vectors enhanced with cross-modal information; Engineering parameter features are obtained by focusing on the context vector of geophysical data features. for:

[0033] in, A contextual feature vector that incorporates geophysical data information into engineering parameter features; Similarly, we can obtain and ; Ultimately, each mode, in addition to its own original features, also acquires contextual information from other related modes. The enhanced features of the engineering parameter mode can be represented as follows: ,in This indicates vector concatenation; Through the above steps, the feature interaction attention subunit completes the interaction and enhancement of cross-modal information, providing feature representations rich in interaction information for the subsequent feature aggregation subunit; The feature aggregation subunit, connected to the feature interaction attention subunit, concatenates the context feature vectors generated by the feature interaction attention subunit with the original modal features, and then performs dimensionality reduction and fusion through a multilayer perceptron, finally outputting a unified high-dimensional fused feature vector.

[0034] The hierarchical data management module also includes: The automatic generation unit for grading reports connects to the grading result library and data service interface. Based on user instructions and scheduled tasks, it extracts all grading results within a specified time range and mileage interval from the grading result library. Combined with the corresponding original data snapshots extracted from the original database, it automatically generates a structured grading report that includes statistical distribution of surrounding rock grades, trend analysis of changes, and location information of key adverse geological sections. The data encryption and secure storage unit encrypts and stores all data in the original database, feature and model library, and hierarchical result library, and sets up role-based access control policies to authenticate and log data access operations performed through the data service interface.

[0035] The operation steps of the tunnel surrounding rock classification and dynamic identification system based on horizontal drilling multi-parameter fusion are as follows: Step 1: Synchronous acquisition of multi-source data: The system's operation begins during horizontal advance drilling at the tunnel face. Through the "multi-source data acquisition module" integrated on the drilling rig, three types of heterogeneous data are collected synchronously and in real time. These include: "drilling engineering parameters" reflecting the drilling status obtained by the drilling rig's sensors; "drilling geophysical data" that detects the geological interface ahead, excited and received by the geophysical equipment inside the drill pipe; and "digital images of the tunnel face" collected by the camera equipment deployed at the tunnel face. These three types of data together constitute the original information basis for identifying the surrounding rock condition.

[0036] Step 2: Data Preprocessing and Alignment The collected raw data are asynchronous, noisy, and have different dimensions. The system standardizes them through the "data synchronization and preprocessing unit". The core operations include: "timestamp alignment" of data streams from different devices based on a unified clock to ensure that all data correspond to the same drilling time; "filtering and noise reduction" of vibration and waveform signals using time-frequency domain filters to remove interference; and "format standardization" of all features using a normalization method to eliminate dimensional differences. This transforms multi-source heterogeneous data into a clean, scale-consistent, and time-synchronized standard data stream, preparing it for subsequent feature extraction.

[0037] Step 3: Multimodal feature extraction and fusion: This step is crucial for information extraction and integration. The system extracts features with clear physical meaning from the preprocessed data through the "feature extraction and fusion module." Specifically, it calculates "drilling specific energy" and "vibration spectrum characteristics" from engineering parameters, analyzes and extracts reflection features such as "reflected wave amplitude, phase, and travel time" from geophysical data, and identifies and calculates structural surface features such as "fracture density, dip angle, and rock mass size" from digital images. Subsequently, the "multi-source feature fusion unit" normalizes and vectorizes these features from different modes, and performs deep fusion through feature splicing and a more advanced cross-modal attention mechanism to generate a unified "high-dimensional fused feature vector," thereby comprehensively and complementaryly representing the overall state of the surrounding rock ahead.

[0038] Step 4: Intelligent Hierarchical Reasoning and Decision Making The system uses a pre-trained "multimodal deep learning model" to intelligently analyze the above-mentioned fused feature vectors. The model outputs a "surrounding rock grade probability distribution" corresponding to the BQ grading system through forward inference calculation. That is, the model predicts the existence of surrounding rock belonging to each grade. Then, the "grading decision unit" makes a final judgment based on this probability distribution, determines a specific automated grading result of the surrounding rock, and attaches the corresponding tunnel mileage location label to the result.

[0039] Step 5: Dynamic visualization and feedback of grading results: To make the results intuitive and usable, the "Dynamic Visualization Module for Grading Results" graphically displays the automated grading results. It first associates the grading results with the spatial coordinates of the tunnel's 3D design model, and then dynamically renders a "Dynamic Grading Cloud Map" with different colors and patterns along the tunnel excavation direction. At the same time, it draws a "Dynamic Grading Curve" with mileage as the horizontal axis. These visualization results are updated in real time on the interface, providing construction personnel with advanced, continuous, and intuitive predictions of changes in the surrounding rock grade ahead of the tunnel face.

[0040] Step 6: Model self-optimization and data closed-loop management: The system possesses self-evolution capabilities. After the actual tunnel excavation, the "model adaptive optimization module" collects the actual exposed surrounding rock grades as real labels and pairs them with the drilling data used in the previous prediction to form new training samples. Subsequently, the system uses these new samples to fine-tune and optimize the parameters of the original deep learning model through incremental learning algorithms. After verifying the performance improvement, the old model is replaced through "hot update," enabling the system to continuously accumulate experience as the project progresses and improve the accuracy of classification under different geological conditions. Throughout the process, the "classification data management module" performs unified encrypted storage and management of all raw data, features, model versions, classification results, and historical optimization records, providing a data foundation for the continuous operation and optimization of the system.

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0042] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic identification system for tunnel surrounding rock classification based on multi-parameter fusion during horizontal drilling, characterized in that, The system includes: The multi-source data acquisition module for drilling is deployed on a horizontal drilling rig to simultaneously and in real time acquire drilling engineering parameters, drilling geophysical data, and digital images of the tunnel face during the advanced drilling process at the tunnel face. The feature extraction and fusion module is connected to each data acquisition unit in the multi-source drilling data acquisition module, receives the drilling engineering parameters, drilling geophysical data and digital images of the working face, extracts drilling specific energy features, vibration spectrum features, electromagnetic wave and seismic wave reflection features and fracture image features from them, and fuses the features into a unified high-dimensional fusion feature vector. The surrounding rock grading intelligent identification module is connected to the feature extraction and fusion module. It stores a pre-trained multimodal deep learning model, receives the high-dimensional fused feature vector, calls the multimodal deep learning model for processing, and outputs real-time, continuous and automated grading results of the surrounding rock in front of the tunnel face. The dynamic visualization module for grading results is connected to the intelligent identification module for surrounding rock grading. It receives the automated grading results and generates and displays a dynamic grading cloud map and curve showing the changes in the surrounding rock grading results along the tunnel axis based on the tunnel axis mileage information. The model adaptive optimization module is connected to the surrounding rock classification intelligent identification module and the multi-source drilling data acquisition module. Based on the newly added drilling data and the corresponding surrounding rock classification revealed by actual excavation, the module performs online incremental learning and parameter optimization on the multimodal deep learning model. The hierarchical data management module is connected to the multi-source drilling data acquisition module, the feature extraction and fusion module, the surrounding rock classification intelligent identification module, and the model adaptive optimization module, respectively, and stores and manages all acquired raw data, intermediate feature data during processing, model parameters, historical classification results, and optimization records.

2. The tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion according to claim 1, characterized in that, The multi-source drilling data acquisition module includes: The drilling engineering parameter acquisition unit is integrated into the drilling rig power head and propulsion mechanism, and collects and outputs real-time time-series data of drilling engineering parameters, including drilling pressure, rotational speed, torque, and drilling speed. The geophysical data acquisition unit is integrated inside the drill pipe and drill bit. During the drilling process, it synchronously excites and receives electromagnetic wave and seismic wave signals, and outputs geophysical data that characterizes the reflection features of the geological interface ahead. The tunnel face image acquisition unit is deployed in front of the tunnel face to acquire high-definition digital images of the tunnel face during drilling intervals and at specific mileage points; The data synchronization and preprocessing unit is connected to the drilling engineering parameter acquisition unit, the drilling geophysical data acquisition unit, and the working face image acquisition unit, respectively. It performs timestamp alignment, filtering and noise reduction, and format standardization preprocessing on the received multi-source heterogeneous data, and outputs the synchronized drilling engineering parameters, drilling geophysical data, and working face digital image to the feature extraction and fusion module.

3. The tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion according to claim 1, characterized in that, The feature extraction and fusion module includes: The engineering parameter feature extraction unit calculates and extracts drilling specific energy features from the drilling engineering parameters based on drilling power and drilling speed. Calculated by the following formula: ； in, Indicates drilling specific energy. Indicates drilling pressure. Indicates rotational speed. Indicates torque, This indicates the cross-sectional area of ​​the drill bit. Indicates drilling speed; The vibration spectrum features, including the dominant frequency and amplitude, are extracted from the vibration signals of the drilling parameters using Fast Fourier Transform. The geophysical data feature extraction unit performs reflection wave analysis on the drilling geophysical data to extract electromagnetic wave and seismic wave reflection features, including reflection wave amplitude, phase, and travel time. The digital image feature extraction unit performs edge detection and pattern recognition on the digital image of the working face, and extracts fracture image features including fracture density, fracture dip angle and rock mass block size; The multi-source feature fusion unit is connected to the engineering parameter feature extraction unit, the geophysical data feature extraction unit, and the digital image feature extraction unit, respectively. It normalizes and vectorizes the drilling energy characteristics, the vibration spectrum characteristics, the electromagnetic wave and seismic wave reflection characteristics, and the fracture image characteristics, and generates the unified high-dimensional fusion feature vector through a weighted fusion method of feature splicing.

4. The tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion according to claim 1, characterized in that, The surrounding rock classification intelligent identification module includes: A hierarchical model storage unit stores the pre-trained multimodal deep learning model, which is a network model built on a convolutional neural network architecture and obtained through supervised training using drilling data from historical tunnel projects and the surrounding rock grade labels verified by the final excavation. A real-time hierarchical inference unit, connected to the hierarchical model storage unit, receives the high-dimensional fused feature vector, calls the multimodal deep learning model for forward inference calculation, and outputs the probability distribution of surrounding rock grades corresponding to the BQ grading system. It can be expressed by the following formula: ； in, Represents a probability distribution vector. This represents the high-dimensional fusion feature vector input to the multimodal deep learning model. This represents the set of all trainable parameters of the multimodal deep learning model. This represents the mapping function from feature vectors to probability distribution vectors defined in the multimodal deep learning model; The hierarchical decision-making unit, connected to the real-time hierarchical reasoning unit, determines the final automated hierarchical result of the surrounding rock according to the probability distribution of the surrounding rock level and according to a set threshold, and adds a location tag based on the drilling mileage.

5. The tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion according to claim 1, characterized in that, The dynamic visualization module for the grading results includes: The spatiotemporal correlation unit receives the automated classification results and their accompanying mileage location tags, and correlates the classification results with the spatial coordinates in the tunnel's three-dimensional design model. The dynamic rendering unit is connected to the spatiotemporal association unit. Based on the associated classification results, it dynamically renders areas of different rock grades along the tunnel excavation direction with a fill pattern to generate the dynamic classification cloud map. The curve drawing unit, connected to the spatiotemporal correlation unit, uses the tunnel mileage as the horizontal axis and the surrounding rock grade value as the vertical axis to draw and update the dynamic grading curve of the surrounding rock grading result along the tunnel axis in real time. The visualization interface unit integrates the outputs of the dynamic rendering unit and the curve drawing unit, synchronously displays the dynamic hierarchical cloud map and the dynamic hierarchical curve on the display device, and provides zoom, pan, and history playback interactive functions.

6. The tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion according to claim 1, characterized in that, The model adaptive optimization module includes: The verification data acquisition unit collects the surrounding rock grade of the actual excavated surface after tunnel excavation, which serves as a verification label for model optimization. It also pairs the data with the corresponding mileage of advanced drilling data to form model optimization sample pairs. An incremental learning engine, connected to the verification data acquisition unit and the surrounding rock classification intelligent identification module, inputs the model optimization sample pairs into the multimodal deep learning model, uses the verification label as a supervision signal, and employs a mini-batch gradient descent algorithm to fine-tune and optimize the network parameters of the multimodal deep learning model. The model performance evaluation unit is connected to the incremental learning engine. After the model is optimized, it uses an independent test dataset to evaluate the classification accuracy and recall performance metrics of the optimized model. When the performance improvement exceeds a set threshold, it triggers a model update command. The model hot update unit is connected to both the model performance evaluation unit and the surrounding rock classification intelligent identification module. Upon receiving the model update instruction, it seamlessly replaces the original model parameters stored in the surrounding rock classification intelligent identification module with the optimized multimodal deep learning model parameters, thereby achieving hot updating of the model.

7. The tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion according to claim 1, characterized in that, The hierarchical data management module includes: The raw database stores the raw data packets collected by the multi-source drilling data acquisition module, categorized by tunnel engineering, borehole number, and mileage interval. The raw data packets include unprocessed drilling engineering parameters, drilling geophysical data, and digital images of the tunnel face. The feature and model library stores the high-dimensional fused feature vector generated by the feature extraction and fusion module, all version parameters of the multimodal deep learning model in the surrounding rock classification intelligent identification module, and each optimization record generated by the model adaptive optimization module. The classification result library archives and stores all historical and current automated classification results output by the surrounding rock classification intelligent identification module in chronological order and spatial mileage. The data service interface provides standardized data query, retrieval, and calling services for the feature extraction and fusion module, the surrounding rock classification intelligent identification module, the model adaptive optimization module, and the classification result dynamic visualization module.

8. The tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion according to claim 4, characterized in that, The multimodal deep learning model stored in the hierarchical model storage unit adopts an architecture of multi-branch feature encoder and feature interaction fusion. The multi-branch feature encoder includes a first feature coding sub-network, a second feature coding sub-network, and a third feature coding sub-network, which respectively perform deep nonlinear feature coding on the engineering parameter feature sub-vector, geophysical data feature sub-vector, and image feature sub-vector separated from the high-dimensional fused feature vector; The feature interaction fusion unit is connected to the output of each sub-network of the multi-branch feature encoder. Through tensor splicing fusion, it interacts and fuses the deep features output by the first feature encoding sub-network, the second feature encoding sub-network, and the third feature encoding sub-network to generate the fused feature representation for the final hierarchical decision.

9. The tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion according to claim 3, characterized in that, The multi-source feature fusion unit is specifically configured to perform cross-modal feature interaction fusion, which includes: The feature interaction attention subunit receives the drilling specific energy feature, vibration spectrum feature, electromagnetic wave and seismic wave reflection feature, and fracture image feature output by the engineering parameter feature extraction unit, the geophysical data feature extraction unit, and the digital image feature extraction unit, respectively, and calculates the correlation weight between different modal features through a cross-attention mechanism. The feature interaction attention subunit first maps each modal feature into a query vector, a key vector, and a value vector, respectively. Then it calculates the attention weights of the engineering parameter modal features to the geophysical data modal features, the attention weights of the engineering parameter modal features to the image modal features, and the attention weights of the geophysical data modal features to the image modal features, generating a series of context feature vectors enhanced with cross-modal information. The feature aggregation subunit is connected to the feature interaction attention subunit. It concatenates the context feature vectors generated by the feature interaction attention subunit with the original modal features, and performs dimensionality reduction and fusion through a multilayer perceptron to finally output the unified high-dimensional fused feature vector.

10. The tunnel surrounding rock classification dynamic identification system based on horizontal drilling multi-parameter fusion according to claim 7, characterized in that, The hierarchical data management module also includes: The automatic generation unit for the grading report is connected to the grading result library and the data service interface. According to user instructions and scheduled tasks, it extracts all grading results within a specified time range and mileage interval from the grading result library, and combines them with the corresponding original data snapshots extracted from the original database to automatically generate a structured grading report that includes the statistical distribution of surrounding rock grades, trend analysis of changes, and location information of key adverse geological sections. The data encryption and secure storage unit encrypts and stores all data in the original database, the feature and model library, and the hierarchical result library, and sets a role-based access control policy to authenticate and log data access operations performed through the data service interface.

Images