A Method and System for Predicting Blasting Vibration in Hydraulic Tunnels by Integrating Multimodal Data

Abstract

This application discloses a method and system for predicting vibration during blasting in hydraulic tunnels by fusing multimodal data, relating to the field of tunnel engineering safety technology. The method includes: acquiring multi-type heterogeneous modal data under tunnel blasting scenarios; generating standardized feature sets through purification, quantization, standardization, and spatiotemporal alignment; obtaining fused feature vectors using a dynamic attention mechanism with adaptive weighted fusion; training a multimodal prediction model to obtain a vibration prediction model; inputting the data for the predicted working conditions into the model after the same preprocessing; outputting peak particle velocity and triggering proactive safety warnings and parameter optimization suggestions. This application achieves a closed loop from vibration prediction to proactive safety management through deep fusion of multimodal data and dynamic weight adjustment, effectively avoiding risks such as surrounding rock instability, lining cracking, and damage to adjacent structures, meeting the application needs of complex environments in hydraulic tunnels.

Description

Technical Field

[0001] This application relates to the field of blasting vibration prediction technology in hydraulic tunnel engineering, specifically to a method and system for predicting blasting vibration in hydraulic tunnel excavation by integrating multimodal data. Background Technology

[0002] As the core hub of water conservancy projects, hydraulic diversion tunnels undertake important functions such as water resource allocation, power generation, and water supply. Due to its high efficiency, economy, and adaptability, the drill-and-blast method has become the mainstream construction method for excavating hydraulic diversion tunnels. However, the vibration waves generated by blasting operations propagate along the surrounding rock medium, easily inducing the expansion of joints and fissures in the surrounding rock, cracking of the lining structure, and even damage to adjacent structures, seriously threatening the safety of construction and long-term operational stability. Therefore, accurately predicting the blasting vibration response parameters is a core prerequisite for optimizing blasting parameters and controlling the stability of the surrounding rock.

[0003] Current technologies for predicting blasting vibrations in hydraulic water diversion tunnels are mainly divided into three categories: First, the traditional empirical formula method, represented by the Sadovsky formula, only considers a few parameters such as explosive consumption and detonation distance, without fully integrating key information such as geological conditions, surrounding rock mechanical properties, and blasting sequence. Under complex geological conditions, the prediction error generally exceeds 30%, making it difficult to adapt to the special engineering environment of hydraulic tunnels. Second, single-modal data prediction models rely heavily on a single type of vibration monitoring data and do not integrate multi-source information such as geological exploration, drilling sensing, and numerical simulation. The data dimension is singular, which cannot characterize the nonlinear and spatiotemporal heterogeneous characteristics of blasting vibration propagation, and the model has weak generalization ability. Third, simple multi-source data fusion models often use simple feature splicing methods, which fail to solve problems such as noise interference, dimensional differences, and spatiotemporal asynchrony of multimodal data (such as static geological data and dynamic time series data). Furthermore, they lack feature extraction and model optimization strategies for water diversion tunnel engineering scenarios, making it difficult to meet the needs of high-precision engineering prediction.

[0004] Existing technologies also attempt to optimize prediction results through machine learning models such as neural networks and support vector machines, for example, by using the XGBoost model combined with the sparrow search algorithm to optimize parameters. However, these technologies are still limited to simple splicing of single data types or multi-source data, and have not achieved deep integration of multimodal data such as geology, construction, monitoring, and simulation. At the same time, given the characteristics of deep burial, high water pressure, and complex surrounding rock grades in water diversion tunnels, there is a lack of effective signal denoising, feature alignment, and dynamic model update mechanisms. These technologies cannot adapt to the dynamic changes in geological conditions during tunneling and cannot provide comprehensive support for scientific decision-making on tunnel excavation and blasting construction safety.

[0005] Therefore, there is an urgent need for a method to predict blasting vibration during water diversion tunnel excavation that integrates multi-source heterogeneous data, adapts to complex geological conditions, has high prediction accuracy, and strong generalization ability. Summary of the Invention

[0006] This application provides a method and system for predicting blasting vibration in hydraulic tunnels by integrating multimodal data, which can solve the technical problems of insufficient prediction accuracy and inability to achieve active safety control in existing technologies.

[0007] In a first aspect, this application provides a method for predicting blasting vibration during hydraulic tunnel excavation by fusing multimodal data, comprising the following steps: Collect multi-type heterogeneous modal data in hydraulic tunnel excavation and blasting scenarios; The heterogeneous modal data are purified, quantized, and standardized respectively to obtain a standardized multimodal feature set; After aligning the standardized multimodal feature set synchronously on both temporal and spatial scales, a dynamic attention mechanism is used to adaptively weight and fuse the standardized multimodal feature set to obtain a fused feature vector. The fused feature vector is input into a preset multimodal prediction model, and after training, verification and optimization, a trained hydraulic tunnel blasting vibration prediction model is obtained. After data preprocessing and spatiotemporal registration, the multimodal data of the predicted tunnel excavation blasting conditions are input into the trained hydraulic tunnel excavation blasting vibration prediction model, outputting the peak mass velocity and triggering active safety warnings and parameter optimization suggestions.

[0008] In conjunction with the first aspect, in one implementation, the dynamic attention mechanism is implemented through a multimodal attention weight network, which includes an input layer, a hidden layer, and an output layer. The hidden layer uses a nonlinear activation function, and the output layer calculates the importance weight of each modal feature through a normalized exponential function. The value range of the importance weight is [0, 1], and the sum of the importance weights of each modal feature is 1. The importance weights are dynamically adjusted based on the prediction error feedback under different tunnel excavation conditions. The specific adjustment rules are as follows: For the t-th training iteration, the modal contribution is calculated based on the change in validation set error before and after individually masking each modal feature, and the importance weight for the next iteration is updated according to the preset weight update rule.

[0009] In conjunction with the first aspect, in one implementation, the multi-type heterogeneous modal data includes vibration time-series sensing data, tunneling and blasting construction parameter data, tunnel hydrogeological parameters, and derived feature data. The purification, quantification, and standardization processing specifically includes: The vibration time-series sensing data is subjected to wavelet threshold denoising and filtering. The tunneling and blasting construction parameter data are normalized. The hydrogeological parameters of the tunnel are classified and quantified. The derived feature data is processed by outlier removal and missing data completion.

[0010] In conjunction with the first aspect, in one implementation, the derived feature data is calculated in the following manner: The charge-to-distance ratio is calculated based on the maximum charge of a single section and the straight-line distance from the detonation point to the monitoring point. Calculate the standard proportional distance based on the charge amount and spatial propagation distance; Calculate the charge concentration based on the total charge amount and cycle advance. The tunneling adaptation coefficient is calculated based on the cyclic advance and the maximum amount of explosives used in a single section.

[0011] In conjunction with the first aspect, in one implementation, the multimodal prediction model is a CNN-LSTM hybrid model, comprising: The CNN feature extraction layer is configured to extract local features of the blast vibration time-series waveform through multi-layer convolution and pooling processing; The LSTM timing capture layer is configured to capture the temporal evolution of blasting vibration signals during tunnel excavation through timing memory units. A fully connected layer, configured to integrate features and output prediction results; The model's loss function uses a weighted combination of root mean square error and mean absolute error.

[0012] In conjunction with the first aspect, in one implementation, the training of the multimodal prediction model employs an adaptive learning rate optimization algorithm. By iteratively adjusting the model parameters until the change in the validation set loss value is less than a preset threshold, a trained hydraulic tunnel blasting vibration prediction model is obtained.

[0013] In conjunction with the first aspect, in one implementation, the importance weight adjustment rule further includes: The importance weights of different modal characteristics are adaptively adjusted based on the tunnel surrounding rock grade and the maximum amount of explosives per section, so that the weights are adapted to the current tunnel excavation and blasting conditions.

[0014] In conjunction with the first aspect, in one implementation method, the linkage triggering of proactive safety warnings and parameter optimization suggestions specifically includes: Compare the peak particle velocity with the preset safe vibration velocity threshold, and obtain the vibration risk level determination result based on the comparison result; When the judgment result is that the predicted peak velocity of the mass particles exceeds the preset safety threshold and triggers an early warning or an active safety warning of the over-limit level, the linkage outputs parameter optimization suggestions for the next cycle of blasting operation at the working face.

[0015] In conjunction with the first aspect, in one implementation, the outlier removal process uses statistical criteria to identify and remove outlier data points that deviate from the normal range, and then uses interpolation to fill in the missing data.

[0016] Secondly, this application provides a hydraulic tunnel excavation blasting vibration prediction system that integrates multimodal data, comprising: The multimodal data acquisition module is used to collect various heterogeneous modal data in hydraulic tunnel excavation and blasting scenarios. The data preprocessing module is used to clean, quantize, and standardize the multi-type heterogeneous modal data respectively to obtain a standardized multimodal feature set; The multimodal feature fusion module is used to synchronously align the standardized multimodal feature set in terms of time and space, and then use a dynamic attention mechanism to adaptively weight and fuse the standardized multimodal feature set to obtain a fused feature vector. The prediction model construction and training module is used to input the fused feature vector into the preset multimodal prediction model, and after training, verification and optimization, a trained hydraulic tunnel blasting vibration prediction model is obtained. The vibration prediction and early warning module is used to input the multimodal data of the tunnel excavation and blasting conditions to be predicted into the trained hydraulic tunnel excavation and blasting vibration prediction model after data preprocessing and spatiotemporal registration, output the peak mass velocity, and trigger active safety early warning and parameter optimization suggestions.

[0017] The beneficial effects of the technical solutions provided in this application include at least the following: By integrating multiple heterogeneous modes, and after purification, quantification, standardization, and spatiotemporal alignment, a dynamic attention mechanism is used to adaptively weighted fuse the fused feature vector. This vector is then input into a pre-set multimodal prediction model for training and optimization. The output is the peak velocity of the mass point, which triggers proactive safety warnings and parameter optimization suggestions. Compared with a single data source, the prediction error is significantly reduced. The dynamic attention mechanism adaptively adjusts the weights according to the working conditions to enhance the generalization ability. The prediction accuracy meets the requirements of hydraulic tunnel engineering, realizing a closed loop from prediction to safety control, and effectively avoiding risks such as surrounding rock instability, lining cracking, and damage to nearby buildings and structures. Attached Figure Description

[0018] Figure 1 A flowchart of the method for predicting blasting vibration in hydraulic tunnels by integrating multimodal data, provided in this application embodiment; Figure 2This is a schematic diagram of the two-stage blast hole arrangement in Class IV surrounding rock of the tunnel provided in an embodiment of this application; Figure 3 This is a flowchart of multimodal data acquisition and preprocessing for tunnel excavation and blasting provided in an embodiment of this application; Figure 4 This is a schematic diagram of the multimodal fusion structure of the dynamic attention mechanism provided in the embodiments of this application; Figure 5 This is a schematic diagram of the CNN-LSTM multimodal prediction model structure provided in the embodiments of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0020] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0021] ReLU: Rectified Linear Unit; Softmax: Softmax Function, normalized exponential function; CNN: Convolutional Neural Network; LSTM: Long Short-Term Memory; Adam: Adaptive Moment Estimation; RMSE: Root Mean Square Error; MAE: Mean Absolute Error.

[0022] Firstly, such as Figure 1 As shown in the embodiments of this application, a method for predicting blasting vibration during hydraulic tunnel excavation by fusing multimodal data is provided, including the following steps: Step S1: Collect multi-type heterogeneous modal data under the blasting scenario of hydraulic tunnel excavation; Step S2: The heterogeneous modal data of the various types are cleaned, quantized, and standardized to obtain a standardized multimodal feature set; Step S3: After synchronizing the standardized multimodal feature set in terms of time and space, an adaptive weighted fusion of the standardized multimodal feature set is performed using a dynamic attention mechanism to obtain a fused feature vector; Step S4: Input the fused feature vector into the preset multimodal prediction model, and after training, verification and optimization, obtain the trained hydraulic tunnel blasting vibration prediction model. Step S5: After data preprocessing and spatiotemporal registration, the multimodal data of the tunnel excavation blasting conditions to be predicted is input into the trained hydraulic tunnel excavation blasting vibration prediction model, outputting the peak mass velocity, and triggering active safety warnings and parameter optimization suggestions.

[0023] This embodiment integrates multiple heterogeneous modalities, and after purification, quantification, standardization, and spatiotemporal alignment, it uses a dynamic attention mechanism to adaptively weighted fuse to obtain a fused feature vector. This vector is then input into a preset multimodal prediction model for training and optimization, outputting the peak velocity of the mass point and triggering proactive safety warnings and parameter optimization suggestions. This significantly reduces the prediction error compared to a single data source. The dynamic attention mechanism adaptively adjusts weights according to working conditions to enhance generalization ability. The prediction accuracy meets the requirements of hydraulic tunnel engineering, realizing a closed loop from prediction to safety management and control, and effectively avoiding risks such as surrounding rock instability, lining cracking, and damage to nearby buildings and structures.

[0024] In one embodiment, step S1, collecting multi-type heterogeneous modal data under the blasting scenario of hydraulic tunnel excavation, specifically includes the following steps: Step S11: Based on the deployment of monitoring points and vibration monitoring instruments to obtain vibration signals, collect vibration time sequence sensing data around the tunnel face and sensitive points.

[0025] Specifically, monitoring points are set up within 50m around the tunnel face, key parts of the lining structure, and sensitive points around the tunnel, with a sampling frequency of no less than 1000Hz; the vibration monitoring instrument has a sampling rate of 10000sps and a vibration velocity range of 0.001~35.000cm / s.

[0026] Step S12: Based on the construction records and on-site measurements at the tunnel face, collect the blasting construction parameter data for the hydraulic tunnel excavation.

[0027] The construction parameters include single-hole charge, single-section charge, total charge, borehole mesh parameters, borehole depth, micro-delay interval time, and face cyclic advance.

[0028] Preferably, the tunneling blasting construction parameter data are obtained simultaneously by combining the blasting construction plan with on-site measurements at the tunnel face to ensure that the parameters are consistent with the actual construction.

[0029] Step S13: Based on the geological survey report, advanced geological forecast, and geological logging of the tunnel face, obtain the tunnel hydrogeological parameters of the current tunneling section.

[0030] Among them, the hydrogeological parameters of the tunnel include the surrounding rock grade, rock mass integrity coefficient, rock mass water-bearing characteristics, and tunnel face burial depth.

[0031] Preferably, the geological log of the tunnel face is updated before each tunneling cycle to ensure that the parameters reflect the current geological conditions.

[0032] Among them, geological logging of the tunnel face refers to the observation, recording and description of geological phenomena at the current excavation face of the tunnel, including geological data such as rock type, structure, joint and fissure development and groundwater status.

[0033] Step S14: Calculate physical correlation characteristics based on the collected basic modal data, and construct derived feature data covering geometric attenuation, energy distribution and geological damping.

[0034] More specifically, each derived feature is calculated in the following way: The charge-to-distance ratio is calculated based on the maximum charge of a single section and the straight-line distance from the detonation point to the monitoring point. Calculate the standard proportional distance based on the charge amount and spatial propagation distance; Calculate the charge concentration based on the total charge amount and cycle advance. The tunneling adaptation coefficient is calculated based on the cyclic advance and the maximum amount of explosives used in a single section.

[0035] Specifically, each derived feature is calculated using the following formula: Dosage distance ratio ,in This is the maximum amount of explosive charge for a single stage. The straight-line distance from the explosion center to the monitoring point is used for preliminary quantification of the energy decay rate. Standard proportional distance This is used to deeply reflect the comprehensive geometric attenuation law of the charge amount and propagation distance; Explosive concentration ,in Total charge For cyclic advance, it is used to characterize the longitudinal energy distribution density at the working face; Tunneling adaptation coefficient It is used to characterize the matching relationship between cyclic advance and explosive consumption.

[0036] Preferably, the derived features are calculated once per cycle to dynamically match them with the current construction parameters.

[0037] This embodiment collects four types of heterogeneous modal data: vibration time sequence sensing, tunneling and blasting construction parameters, tunnel hydrogeological parameters, and derived features, covering all dimensions of influencing factors of tunneling and blasting vibration. Among them, the derived features are constructed based on the physical and mechanical relationship between the tunnel blasting energy attenuation law and spatial geometry, which is different from the conventional empirical formula fitting scheme and effectively improves the model's ability to generalize to complex geological conditions. The four types of data, after subsequent preprocessing, dynamic attention fusion, and CNN-LSTM model training, provide reliable input for the prediction model, ensuring accurate capture of vibration patterns under different tunneling conditions, and significantly improving prediction accuracy and adaptability to hydraulic tunnel engineering.

[0038] In one embodiment, such as Figure 3 As shown, step S2 involves cleaning, quantizing, and standardizing the various heterogeneous modal data to obtain a standardized multimodal feature set. This specifically includes the following steps: Step S21: Perform noise reduction and filtering operations on the vibration time sequence sensing data to obtain purified vibration time sequence data.

[0039] Specifically, the original signal is decomposed into 5 levels using the db4 wavelet basis, and wavelet threshold noise reduction is performed with a threshold of 0.02. Then, high-frequency interference noise such as rock drilling and ventilation is further filtered out by a Butterworth filter with a cutoff frequency of 100Hz.

[0040] Step S22: Perform dimensionless processing on the tunneling and blasting construction parameter data to obtain standardized construction parameter data.

[0041] Specifically, the tunneling and blasting construction parameter data are normalized using the min-max normalization method, according to the formula... Standardized construction parameter data were calculated, and all construction parameters were mapped to the interval [0, 1] to achieve dimensionless measurement; among which These are the original parameter values. This is the minimum value of the parameter. This is the maximum value.

[0042] Step S23: Perform quantification and assignment processing on the hydrogeological parameters of the tunnel to obtain geological parameter data that can be used for calculation.

[0043] Specifically, values ​​are assigned hierarchically according to the following rules: The surrounding rock grade of the tunnel is 5 for Grade I, 4 for Grade II, 3 for Grade III, 2 for Grade IV, and 1 for Grade V; the two-stage blast hole arrangement for Grade IV surrounding rock is as follows: Figure 2 As shown; The integrity of the rock mass is rated as follows: intact (5), relatively intact (4), moderately intact (3), relatively broken (2), and broken (1). The water-bearing characteristics of the rock mass are 5 for dryness, 4 for dripping, 3 for sprinkling or linear flow, 2 for gushing, and 1 for sudden water inrush. The burial depth of the tunnel face is 5 for H<50m, 4 for 50≤H<300m, 3 for 300≤H<600m, 2 for 600≤H<1000m, and 1 for H≥1000m.

[0044] Step S24: Perform outlier removal and missing value completion processing on the derived feature data to obtain a complete derived feature dataset.

[0045] More specifically, outliers are removed from the derived feature data using the 3σ criterion; specifically, the sequence average is first calculated. with standard deviation The standard deviation is calculated using the formula Calculate, where n is the total number of samples. Let i be the i-th data point; If satisfied Then determine Outliers caused by strong construction interference or occasional sensor malfunctions are removed, and then missing data is filled in by linear interpolation.

[0046] This embodiment effectively removes vibration signal noise, unifies the dimensions of construction parameters, transforms qualitative geological information into quantitative data, and corrects outliers of derived features through targeted preprocessing of four types of heterogeneous modal data. This results in a high-quality standardized multimodal feature set, providing reliable input for subsequent feature fusion and model training, and significantly improving the accuracy and stability of the prediction model.

[0047] In one embodiment, such as Figure 4 As shown, step S3, after synchronously aligning the standardized multimodal feature set in both time and space, uses a dynamic attention mechanism to adaptively weight and fuse the standardized multimodal feature set to obtain a fused feature vector. This specifically includes the following steps: Step S31: Perform spatiotemporal synchronization alignment on the preprocessed standardized multimodal feature set to obtain the spatiotemporally aligned standardized multimodal feature set.

[0048] Specifically, based on the unified timing information of the data acquisition equipment, the vibration time sequence sensing data, tunneling blasting construction parameters, tunnel hydrogeological parameters and derived characteristic data are aligned to the same time series; at the same time, based on the relative positional relationship between the monitoring point and the blast center, the monitoring data of different spatial locations are mapped to a unified tunneling coordinate system.

[0049] Step S32: Perform multi-layer nonlinear mapping on the spatiotemporally aligned standardized multimodal feature set, calculate the importance weights of various modal features, and obtain the adaptively assigned weight vector.

[0050] Specifically, a multimodal attention weight network is constructed, which includes an input layer, a hidden layer, and an output layer; The number of hidden layer nodes is set to 32, and the ReLU function is used as the activation function; The output layer uses the Softmax function to normalize the weights to the [0, 1] interval, and the sum of the weights of the four modes—vibration timing sensing, tunneling blasting construction parameters, tunnel hydrogeological parameters, and derived features—is strictly 1.

[0051] Step S33: Perform feature weighted fusion based on the adaptively assigned weight vector to obtain the fused feature vector.

[0052] Specifically, feature vector fusion Calculated using the following formula: ; In the formula, To fuse feature vectors; , , , The weights are for four types of modes: vibration timing sensing, tunneling and blasting construction parameters, tunnel hydrogeological parameters, and derived features. , , , These are the standardized feature vectors for the four modalities.

[0053] Step S34: Dynamically adjust the weights based on the prediction error feedback to obtain the updated weight vector.

[0054] Specifically, dynamic adjustments are performed according to the error feedback rules. For the t-th training iteration, the changes in the validation set error before and after individually masking each modality feature are considered. Calculate the modal contribution and update the weights using the following formula: ; In the formula, The step size for weight updates is preset to 0.05; Let be the current base weights of the i-th modality features during the t-th training iteration; This represents the updated importance weight assignment value at the (t+1)th training iteration; is the change in error feedback for the i-th modality, which is the increase in the overall prediction error of the model after the input of this modality feature is individually masked. The larger this value is, the more significant and irreplaceable the modality feature is in the accuracy of the current prediction condition. Softmax is a normalized exponential function, which performs nonlinear mapping and proportional compression on the weight values ​​after superimposed error feedback increments to ensure that the importance weights of the updated modal features are strictly defined in the range of [0, 1], and the sum of the weights of the four modalities input to the model always remains 1.

[0055] The importance weights of different modal characteristics are adaptively adjusted according to the tunnel surrounding rock grade and the maximum amount of explosives per section, so that the weights are adapted to the current tunnel excavation and blasting conditions. Preferably, the weight allocation strategy is dynamically adjusted according to the tunnel excavation conditions: When the surrounding rock grade is IV or V, increase the weight of the tunnel hydrogeological parameter modes; When the adjustment range of the explosive charge exceeds the preset threshold, the weight of the tunneling blasting construction parameter mode is increased.

[0056] This embodiment ensures a strict correspondence between multimodal features in time series and spatial location by synchronizing and aligning time and spatial scales, eliminating fusion errors caused by spatiotemporal misalignment. A dynamic attention mechanism adaptively adjusts the weights of each modality, highlighting the influence of key modal features under current working conditions and solving the problems of feature redundancy and loss of effective information in fixed-weight models under complex geological conditions. An error feedback mechanism enables continuous self-optimization of the weights, allowing the fused feature vector to accurately reflect the vibration propagation patterns at different tunneling stages, providing a high-quality input foundation for subsequent prediction models.

[0057] In one embodiment, step S4, inputting the fused feature vector into a preset multimodal prediction model, and after training, verification, and optimization, obtaining a trained hydraulic tunnel blasting vibration prediction model, specifically includes the following steps: Step S41: Receive the fused feature vector, construct the CNN-LSTM hybrid model, and output the initial model.

[0058] More specifically, the CNN-LSTM hybrid model includes an input layer, a CNN feature extraction layer, an LSTM temporal capture layer, a fully connected layer, and an output layer.

[0059] The CNN feature extraction layer is configured to extract local features of the blast vibration time-series waveform (such as waveform segments corresponding to peak values ​​and main frequencies) through multi-layer convolution and pooling processing. The LSTM timing capture layer is configured to capture the temporal evolution of blasting vibration signals during tunnel excavation through timing memory units. A fully connected layer, configured to integrate features and output prediction results.

[0060] This model extracts local detail features of the vibration waveform through CNN and captures the temporal evolution pattern by combining LSTM. It takes into account both the local nonlinearity and temporal dependence of tunnel blasting vibration, which is different from the LSTM-SA model in comparison document 1, which does not combine convolution to extract local features.

[0061] The model training uses the Adam optimizer to ensure that the prediction accuracy meets the requirements of hydraulic tunnel engineering (R²). 2 ≥0.90, RMSE≤0.5cm / s).

[0062] Specifically, the CNN feature extraction layer has two convolutional layers, each followed by a batch normalization layer. The first convolutional kernel has a size of 3×3 and a number of 32 kernels, and the second convolutional kernel has a size of 3×3 and a number of 64 kernels. The activation function for both layers is ReLU, followed by a 2×2 max pooling layer. The LSTM timing capture layer has one hidden layer with 64 nodes, followed by a dropout layer with a dropout probability of 0.3. The fully connected layer has one hidden layer and 32 nodes; The peak vibration velocity of the output particles is output from the output layer.

[0063] Step S42: Divide the preprocessed multimodal data into training set, validation set and test set according to the proportion.

[0064] Specifically, the dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used for learning model parameters, the validation set is used for tuning hyperparameters and early stopping judgment, and the test set is used to evaluate the final model performance.

[0065] Step S43: Receive training set and validation set data, and iteratively optimize and adjust the parameters of the preset multimodal prediction model until the decrease in the validation set loss value is less than the preset threshold, and output the trained hydraulic tunnel blasting vibration prediction model.

[0066] Specifically, the Adam optimizer is used with an initial learning rate of 0.001, a decay factor of 0.9, a momentum factor of 0.99, and a weight decay factor of 1×10⁻⁶. 5 loss function The expression is: ; in, The root mean square error, , The mean absolute error, , For the true value, is the predicted value, and n is the number of samples.

[0067] The maximum number of iterations is set to 500, and the iteration termination condition is that the change in the validation set loss value is less than 1 × 10⁻⁶ after n consecutive iterations. 4 (n is preset to 10), when the early stopping mechanism is triggered, training is terminated and the optimal model weights are saved.

[0068] Step S44: Receive test set data, evaluate the prediction accuracy of the trained model, and output a model that meets engineering requirements.

[0069] Specifically, calculate the coefficient of determination R² and root mean square error RMSE on the test set to ensure that R²... 2 ≥0.90 and RMSE≤0.5cm / s.

[0070] If the target is not met, adjust the number of convolutional kernels, the number of hidden layer nodes, or the learning rate parameter, and return to step S43 to retrain; If the target is met, the model parameters are locked, and the trained hydraulic tunnel blasting vibration prediction model is output.

[0071] This embodiment constructs a CNN-LSTM hybrid model structure, combining convolutional layers to extract local features of blasting vibration waveforms with LSTM layers to capture temporal evolution patterns, effectively improving the accuracy of vibration prediction under complex working conditions. By dividing the dataset in a 7:2:1 ratio and introducing an early stopping mechanism, overfitting of the model is avoided, enhancing the generalization ability under different surrounding rock grades and blasting scales. Through training with a weighted loss function and an adaptive optimizer, the model's prediction accuracy is ensured to strictly meet the requirements of hydraulic tunnel engineering, providing reliable support for subsequent vibration prediction and safety management.

[0072] In one embodiment, step S5 involves preprocessing and spatiotemporally registering the multimodal data of the tunnel excavation blasting conditions to be predicted, then inputting it into the trained hydraulic tunnel excavation blasting vibration prediction model to output the peak mass velocity and trigger active safety warnings and parameter optimization suggestions. Specifically, this includes the following steps: Step S51: Perform preprocessing and spatiotemporal registration on the multimodal data of the tunnel excavation blasting conditions to be predicted, and obtain a standardized and spatiotemporally aligned multimodal feature set.

[0073] Specifically, the preprocessing follows the same procedure as step S2: For vibration time-series sensor data, wavelet thresholding with a 5-layer decomposition using a db4 wavelet base and a threshold of 0.02 was employed for noise reduction, combined with Butterworth filtering at a cutoff frequency of 100Hz. For tunneling blasting construction parameter data, min-max normalization was used to map it to the [0, 1] interval. For tunnel hydrogeological parameters, 1-5 level coding and quantization were used for graded assignment. For derived feature data, outliers were removed using the 3σ criterion, followed by linear interpolation for completion. Spatiotemporal alignment was achieved by matching time series using unified timing information and mapping the relative positions of monitoring points and blast centers to a unified tunneling coordinate system.

[0074] Step S52: Call the trained hydraulic tunnel blasting vibration prediction model to perform inference calculation, receive a standardized and spatiotemporally aligned multimodal feature set, and output the peak particle velocity and related vibration parameters of the working condition to be predicted through forward propagation.

[0075] Specifically, the forward propagation process of the model is as follows: the input layer receives multimodal feature vectors, the CNN feature extraction layer extracts local features of the vibration waveform, the LSTM time-series capture layer extracts the time-series evolution law, the fully connected layer integrates the features and the output layer outputs the peak velocity of the mass point, and at the same time outputs the dominant frequency and duration.

[0076] Preferably, the inference process utilizes the optimal weights saved during model training to ensure that the prediction accuracy meets engineering requirements.

[0077] Step S53: Determine the blasting vibration risk level of the working condition to be predicted, receive the peak particle velocity output by the model, and output the corresponding risk level by comparing it with the preset safe vibration velocity threshold.

[0078] Specifically, the preset safe vibration velocity threshold is 0.50 cm / s. When the peak vibration velocity of the mass point is less than 0.50 cm / s, it is determined to be at the safe level; when it is greater than or equal to 0.50 cm / s and less than 0.80 cm / s, it is determined to be at the warning level; and when it is greater than or equal to 0.80 cm / s, it is determined to be at the over-limit level.

[0079] Preferably, the risk level classification can be adjusted according to the tunnel structure type. For example, the warning threshold for weak sections of the lining structure can be reduced to 0.40 cm / s.

[0080] Step S54: Trigger proactive safety warnings and parameter optimization suggestions, receive the determined risk level, generate corresponding control measures based on the level, and output proactive safety warning information and parameter optimization adjustment suggestions.

[0081] Specifically, when the risk level is safe, only the prediction result is output; When the alert level is reached, a yellow alert is triggered, and enhanced monitoring is prompted. When the level exceeds the limit, a red alert is triggered and construction is suspended. At the same time, based on the feature sensitivity analysis, optimization suggestions are output in reverse, including reducing the maximum amount of explosive initiation in a single section, increasing the micro-delay initiation interval of the detonator, adjusting the hole mesh parameters, or adding peripheral vibration damping holes.

[0082] For example, if the PPV is predicted to exceed the limit in the next cycle, the system can automatically suggest reducing the maximum amount of explosive charge per segment of the upper step cut hole from 4.5 kg to 3.8 kg, or extending the micro-delay interval by 10 ms, thereby achieving dynamic optimization of blasting parameters.

[0083] Preferably, the optimization suggestions can be simultaneously pushed to the on-site construction terminal to guide the adjustment of blasting parameters in the next cycle.

[0084] This embodiment ensures that the data quality of the input model is consistent with the training data by standardizing and preprocessing the multimodal data of the predicted working conditions, thus avoiding prediction errors caused by data deviation. By calling the trained CNN-LSTM model for inference, it outputs accurate peak mass velocities and related vibration parameters, providing a reliable basis for risk assessment. Through hierarchical risk warning and linkage optimization suggestions, it realizes a closed loop from vibration prediction to proactive safety management, effectively avoiding risks such as surrounding rock instability, lining cracking, and damage to nearby buildings and structures, and directly supporting on-site construction decisions.

[0085] Taking the blasting excavation of a Class IV surrounding rock in a water diversion tunnel as an example, this paper details the specific implementation steps of this method. The tunnel is 312.75m long, of which 121m (38.7%) is in Class IV surrounding rock; the tunnel depth is 40-80m. A two-stage blasting method is used for excavation, with a portal-shaped cross-section. The cross-section height is 4.6-7.8m, the width is 4.2-7.48m, and the excavation area is 18-50.4m². 2 , Figure 2 This is a schematic diagram of the two-stage blast hole layout for Class IV surrounding rock in a tunnel.

[0086] Figure 3 The tunnel excavation blasting multimodal data acquisition and preprocessing process of this application involves four types of heterogeneous modal data, namely: vibration time series sensing data, tunnel excavation blasting construction parameter data, tunnel hydrogeological parameters, and derived feature data; The vibration time-series sensing data was acquired by setting up vibration monitoring points at specific locations within a 50m radius around the tunnel face. In this embodiment, there were a total of three monitoring points, corresponding to the sidewall of the lining structure, the ground surface above the tunnel, and nearby important buildings, respectively. The distances from the three monitoring points to the tunnel face were 10m, 20m, and 30m, respectively. Vibration signals were acquired using a vibration monitoring instrument, requiring a sampling rate of 10,000 sps, a sampling frequency of not less than 1,000 Hz, and a vibration velocity range of 0.001–35.000 cm / s. The vibration time-series sensing data included the blasting vibration velocity time-series waveform, from which features such as peak value, dominant frequency, and duration were extracted. The preprocessing process of the vibration time-series sensing data was as follows: the vibration signal was decomposed using a db4 wavelet basis, 5-level decomposition, and a threshold of 0.02 for wavelet threshold denoising, and then filtered using a Butterworth filter with a cutoff frequency of 100 Hz to remove environmental interference noise such as tunnel drilling and ventilation. The collection of tunneling blasting construction parameter data is based on the blasting construction plan. The two-stage blasting construction parameters for the Class IV surrounding rock tunnel section determined in this embodiment are detailed in Table 1 below. The data preprocessing process is as follows: the min-max normalization method is used to map parameters such as delay time, borehole depth, number of boreholes, single borehole charge, single section charge, and cycle advance to the [0, 1] interval to completely eliminate the influence of dimensions. Table 1. Parameters for Two-Stage Blasting Method The collection of tunnel hydrogeological parameters is based on the geological survey report and the geological logging of the tunnel face. In this embodiment, the surrounding rock grade is determined to be Class IV, the rock mass integrity grade is Class III, the water-bearing characteristics of the rock mass are water seepage or linear flow, and the burial depth of the section where the tunnel face is located is 40~80m. According to the 5-level coding quantification standard, the surrounding rock grade is assigned a value of 2, the rock mass integrity grade is assigned a value of 3, the burial depth of the tunnel face is assigned a value of 4, and the water-bearing characteristics of the rock mass are assigned a value of 3.

[0087] In order to deeply explore the physical and mechanical relationships between the basic variables, this embodiment constructs a derived feature matrix covering geometric decay, energy distribution and geological correction based on a multimodal input system.

[0088] After completing feature extraction and calculation, 3 σ The criteria remove outliers from the calculation results; In practice, each derived feature data point is calculated one by one. Compared with the average The absolute deviation, when the absolute deviation is greater than 3 times the standard deviation (i.e., meets the requirement of...) When the data point is identified as an outlier caused by strong construction interference or occasional sensor malfunction, the system automatically identifies it as such and removes it. Then, the system uses linear interpolation to fill in the missing data caused by the removal of outliers, and finally completes the preprocessing of the above four types of data to obtain a standardized multimodal feature set.

[0089] Figure 4 This is a schematic diagram of the multimodal fusion structure of the dynamic attention mechanism in an embodiment of this application. The four standardized multimodal feature sets obtained after the above preprocessing are input into the multimodal attention weight network; the specific mathematical calculation formulas corresponding to the discrimination and elimination processes are as follows: The multimodal attention weight network includes an input layer, a hidden layer, and an output layer. The number of nodes in the hidden layer is set to 32, and the activation function is ReLU. The activation function of the output layer is Softmax. The input layer receives standardized feature vectors of four modes, the hidden layer performs nonlinear mapping on the features, and the output layer calculates the importance weights of the four modes—vibration timing sensing, tunneling and blasting construction parameters, tunnel hydrogeological parameters, and derived features—through the Softmax function.

[0090] In this process, the features of various modalities are weighted and concatenated based on the calculated weights to obtain a fused feature vector.

[0091] Figure 5 This is a schematic diagram of the CNN-LSTM multimodal prediction model structure according to an embodiment of this application. The fused feature vector obtained in step S3 is input into the preset CNN-LSTM hybrid model for training and optimization; The CNN feature extraction layer extracts local features of the explosion vibration time-series waveform through two convolutional layers and one pooling layer. The first convolutional layer uses 32 3×3 convolutional kernels, the second convolutional layer uses 64 3×3 convolutional kernels, the activation function is ReLU, and the pooling layer uses 2×2 max pooling. The LSTM temporal capture layer captures the temporal evolution of tunnel blasting vibration signals through one hidden layer (64 nodes) and one dropout layer (dropout probability 0.3). The fully connected layer includes one hidden layer (32 nodes), and the output layer outputs the prediction results. The model training is preset to a maximum of 500 iterations, and an early stopping mechanism is introduced during actual training to prevent overfitting. After 156 iterations, the validation set loss value tends to converge, and the decrease in the validation set loss value over 10 consecutive iterations is less than 1 × 10⁻⁶. -4 The early stopping mechanism is triggered to meet the iteration termination condition, thereby completing the training of the prediction model for blasting vibration in hydraulic tunnel excavation, and saving the model weights of this iteration as the optimal prediction model.

[0092] The prediction method based on CNN-LSTM multimodal data fusion proposed in this invention effectively overcomes the defect of traditional methods caused by amplified noise leading to main frequency deviation. Meanwhile, this method achieves lower root mean square error (RMSE) and lower coefficient of determination (R²). 2 The performance indicators are significantly superior, and the prediction accuracy meets the requirements for tunnel excavation and blasting engineering applications. For the predicted working condition in this embodiment, the RMSE is 0.032cm / s ≤ 0.5cm / s, and R... 2 The value of 0.976 ≥ 0.90 indicates the generalization ability and prediction accuracy of the prediction model under different tunnel excavation and blasting conditions.

[0093] Secondly, this application provides a hydraulic tunnel excavation blasting vibration prediction system that integrates multimodal data, comprising: The multimodal data acquisition module is used to collect various heterogeneous modal data in hydraulic tunnel excavation and blasting scenarios. The data preprocessing module is used to clean, quantize, and standardize the multi-type heterogeneous modal data respectively to obtain a standardized multimodal feature set; The multimodal feature fusion module is used to synchronously align the standardized multimodal feature set in terms of time and space, and then use a dynamic attention mechanism to adaptively weight and fuse the standardized multimodal feature set to obtain a fused feature vector. The prediction model construction and training module is used to input the fused feature vector into the preset multimodal prediction model, and after training, verification and optimization, a trained hydraulic tunnel blasting vibration prediction model is obtained. The vibration prediction and early warning module is used to input the multimodal data of the tunnel excavation and blasting conditions to be predicted into the trained hydraulic tunnel excavation and blasting vibration prediction model after data preprocessing and spatiotemporal registration, output the peak mass velocity, and trigger active safety early warning and parameter optimization suggestions.

[0094] The functions of each module in the above-mentioned hydraulic tunnel blasting vibration prediction system that integrates multimodal data correspond to the steps in the above-mentioned hydraulic tunnel blasting vibration prediction method embodiment that integrates multimodal data. Their functions and implementation processes will not be described in detail here.

[0095] Thirdly, embodiments of this application also provide a readable storage medium.

[0096] This application stores a hydraulic tunnel blasting vibration prediction program that integrates multimodal data on a readable storage medium. When the hydraulic tunnel blasting vibration prediction program that integrates multimodal data is executed by a processor, it implements the steps of the hydraulic tunnel blasting vibration prediction method that integrates multimodal data as described above.

[0097] The method implemented when the hydraulic tunnel blasting vibration prediction program that integrates multimodal data is executed can refer to the various embodiments of the hydraulic tunnel blasting vibration prediction method that integrates multimodal data in this application, and will not be repeated here.

[0098] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0099] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the hydraulic tunnel blasting vibration prediction method that integrates multimodal data as described in the various embodiments of this application.

[0100] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for predicting blasting vibration during hydraulic tunnel excavation by integrating multimodal data, characterized in that, Includes the following steps: Collect multi-type heterogeneous modal data in hydraulic tunnel excavation and blasting scenarios; The heterogeneous modal data are purified, quantized, and standardized respectively to obtain a standardized multimodal feature set; After aligning the standardized multimodal feature set synchronously on both temporal and spatial scales, a dynamic attention mechanism is used to adaptively weight and fuse the standardized multimodal feature set to obtain a fused feature vector. The fused feature vector is input into a preset multimodal prediction model, and after training, verification and optimization, a trained hydraulic tunnel blasting vibration prediction model is obtained. After data preprocessing and spatiotemporal registration, the multimodal data of the predicted tunnel excavation blasting conditions are input into the trained hydraulic tunnel excavation blasting vibration prediction model, outputting the peak mass velocity and triggering active safety warnings and parameter optimization suggestions.

2. The method for predicting blasting vibration during hydraulic tunnel excavation by fusing multimodal data as described in claim 1, characterized in that, The dynamic attention mechanism is implemented through a multimodal attention weight network, which includes an input layer, a hidden layer, and an output layer. The hidden layer uses a nonlinear activation function, and the output layer calculates the importance weight of each modal feature through a normalized exponential function. The value range of the importance weight is [0, 1] and the sum of the importance weights of each modal feature is 1. The importance weights are dynamically adjusted based on the prediction error feedback under different tunnel excavation conditions. The specific adjustment rules are as follows: For the t-th training iteration, the modal contribution is calculated based on the change in validation set error before and after individually masking each modal feature, and the importance weight for the next iteration is updated according to the preset weight update rule.

3. The method for predicting blasting vibration during hydraulic tunnel excavation by fusing multimodal data as described in claim 1, characterized in that, The heterogeneous modal data includes vibration time-series sensing data, tunneling and blasting construction parameter data, tunnel hydrogeological parameters, and derived feature data. The purification, quantification, and standardization processing specifically includes: The vibration time-series sensing data is subjected to wavelet threshold denoising and filtering. The tunneling and blasting construction parameter data are normalized. The hydrogeological parameters of the tunnel are classified and quantified. The derived feature data is processed by outlier removal and missing data completion.

4. The method for predicting blasting vibration in hydraulic tunnel excavation by fusing multimodal data according to claim 3, characterized in that, The derived feature data is calculated in the following way: The charge-to-distance ratio is calculated based on the maximum charge of a single section and the straight-line distance from the detonation point to the monitoring point. Calculate the standard proportional distance based on the charge amount and spatial propagation distance; Calculate the charge concentration based on the total charge amount and cycle advance. The tunneling adaptation coefficient is calculated based on the cyclic advance and the maximum amount of explosives used in a single section.

5. The method for predicting blasting vibration in hydraulic tunnel excavation by fusing multimodal data according to claim 1, characterized in that, The multimodal prediction model is a CNN-LSTM hybrid model, including: The CNN feature extraction layer is configured to extract local features of the blast vibration time-series waveform through multi-layer convolution and pooling processing; The LSTM timing capture layer is configured to capture the temporal evolution of blasting vibration signals during tunnel excavation through timing memory units. A fully connected layer, configured to integrate features and output prediction results; The model's loss function uses a weighted combination of root mean square error and mean absolute error.

6. The method for predicting blasting vibration during hydraulic tunnel excavation by fusing multimodal data according to claim 5, characterized in that, The training of the multimodal prediction model adopts an adaptive learning rate optimization algorithm. The model parameters are iteratively adjusted until the change in the validation set loss value is less than a preset threshold, thus obtaining the trained hydraulic tunnel blasting vibration prediction model.

7. The method for predicting blasting vibration in hydraulic tunnel excavation by fusing multimodal data according to claim 2, characterized in that, The importance weight adjustment rules also include: The importance weights of different modal characteristics are adaptively adjusted based on the tunnel surrounding rock grade and the maximum amount of explosives per section.

8. The method for predicting blasting vibration during hydraulic tunnel excavation by fusing multimodal data according to claim 1, characterized in that, The linked triggering of proactive safety warnings and parameter optimization suggestions specifically includes: Compare the peak particle velocity with the preset safe vibration velocity threshold, and obtain the vibration risk level determination result based on the comparison result; When the judgment result is that the predicted peak velocity of the mass particles exceeds the preset safety threshold and triggers an early warning or an active safety warning of the over-limit level, the linkage outputs parameter optimization suggestions for the next cycle of blasting operation at the working face.

9. The method for predicting blasting vibration during hydraulic tunnel excavation by fusing multimodal data according to claim 3, characterized in that, The outlier removal process uses statistical criteria to identify and remove abnormal data points that deviate from the normal range, and then uses interpolation to fill in the missing data.

10. A vibration prediction system for blasting in hydraulic tunnel excavation that integrates multimodal data, characterized in that, include: The multimodal data acquisition module is used to collect various heterogeneous modal data in hydraulic tunnel excavation and blasting scenarios. The data preprocessing module is used to clean, quantize, and standardize the multi-type heterogeneous modal data respectively to obtain a standardized multimodal feature set; The multimodal feature fusion module is used to synchronously align the standardized multimodal feature set in terms of time and space, and then use a dynamic attention mechanism to adaptively weight and fuse the standardized multimodal feature set to obtain a fused feature vector. The prediction model construction and training module is used to input the fused feature vector into the preset multimodal prediction model, and after training, verification and optimization, a trained hydraulic tunnel blasting vibration prediction model is obtained. The vibration prediction and early warning module is used to input the multimodal data of the tunnel excavation and blasting conditions to be predicted into the trained hydraulic tunnel excavation and blasting vibration prediction model after data preprocessing and spatiotemporal registration, output the peak mass velocity, and trigger active safety early warning and parameter optimization suggestions.

Images