A method for predicting the vibration impact of tunnel excavation blasting on surface buildings

Abstract

The present invention relates to the field of artificial intelligence and data processing technology, and specifically to a method for predicting the vibration impact of tunnel excavation blasting on surface buildings, which is as follows: collecting vibration signals generated by blasting operations during tunnel excavation and denoising them as samples; generating meta-tasks using a spectral clustering algorithm based on the energy distribution characteristics of the signals; extracting features from the samples in each meta-task using a neural network model; using a hierarchical attention mechanism to capture the time response characteristics of the vibration signal at different time lengths, obtaining the key features of each stage and fusing them to obtain a classification feature vector, and then determining the classification prediction results of the vibration impact of blasting on surface buildings based on the classification feature vector; and optimizing the classification prediction process. The present invention processes tunnel excavation blasting vibration data based on a neural network model with an improved meta-learning strategy, which can improve the accuracy of the vibration impact on surface buildings and the safety of construction.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and data processing technology, and in particular to a method for predicting the vibration impact of tunnel excavation blasting on surface buildings. Background Art

[0002] Blasting vibrations are inevitable during tunnel excavation. These vibration signals not only reflect information such as the geological conditions and blasting parameters during the blasting process, but also contain complex noise and mixed signals from different frequency bands. Traditional vibration monitoring technologies often rely on fixed threshold methods or simple Fourier transforms. However, these methods have many limitations when dealing with non-stationary conditions, strong noise interference, and instantaneous frequency domain features. Especially in environments with complex geological conditions and fluctuating blasting parameters, traditional methods struggle to effectively extract useful features from vibration signals, resulting in low classification and prediction accuracy. Furthermore, traditional machine learning methods, such as meta-learning, often employ random approaches to task division, resulting in samples of similar vibration modes being assigned to different tasks and failing to effectively capture the correlation between geological conditions and blasting parameters. Furthermore, traditional neural network methods often employ random initialization, which can lead to slow convergence or a tendency to fall into local optima during training. Feature extraction methods are also often limited to one-dimensional extraction in the time or frequency domain, failing to fully capture the complex characteristics of vibration signals. The temporal and energy decay properties of vibration signals are also not fully utilized, making the models incapable of efficient and accurate predictions in practical applications.

[0003] The existing technologies have the following objective shortcomings: traditional Fourier transform or fixed threshold wavelet denoising methods often cannot fully retain the instantaneous features in effective vibration signals, which may lead to the loss of effective vibration signals; conventional meta-learning methods use random task division, which may cause samples of similar vibration modes to be assigned to different tasks, and cannot effectively capture the correlation between geological conditions and blasting parameters; traditional neural network initialization methods use random initialization, which easily leads to slow convergence of network training or even no convergence; traditional meta-learning methods usually use a fixed learning rate and do not consider the energy attenuation characteristics of vibration signals, resulting in the inability to effectively adapt to the timing changes of signals during training; traditional feature extraction methods may only be able to extract one aspect of the signal, such as the time domain or frequency domain, and cannot fully capture the instantaneous and long-term characteristics of the vibration signal; traditional blasting parameter adjustment usually relies on manual experience or simple rules, which is inefficient.

[0004] Therefore, the present invention proposes a method for predicting the vibration impact of tunnel excavation blasting on surface buildings to solve the above problems. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention develops a method for predicting the vibration impact of tunnel excavation blasting on surface buildings. The present invention processes tunnel excavation blasting vibration data based on a neural network model with an improved meta-learning strategy, which can improve the accuracy of the vibration impact on surface buildings and the safety of construction.

[0006] The technical solution to the technical problem solved by the present invention is a method for predicting the vibration impact of tunnel excavation blasting on surface buildings, comprising the following steps:

[0007] S1. Using a vibration sensor to collect vibration signals generated by blasting during tunneling, and then performing a denoising operation on the collected vibration signals, using the denoised vibration signals as samples;

[0008] S2. Divide the denoised vibration signal into tasks and use a spectral clustering algorithm to generate meta-tasks based on the energy distribution characteristics of the vibration signal. Each meta-task is a sample set.

[0009] S3. Extract features from samples in each meta-task using a neural network model. The feature extraction process includes initializing the parameters of the neural network model, performing dual-path feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy attenuation curve to optimize the meta-parameters in the meta-task.

[0010] S4. A hierarchical attention mechanism is used to capture the temporal response characteristics of vibration signals at different time lengths, obtaining the key features of each stage. The key features of each stage are then fused to obtain a classification feature vector. The classification prediction results of the impact of blasting on surface building vibration are then determined based on the classification feature vector.

[0011] S5. Optimize the classification prediction process, including generalization verification and dynamic adjustment of meta-tasks, optimization of closed-loop feedback control and parameters, and online optimization of blasting parameter strategies through reinforcement learning strategies.

[0012] The denoising operation uses dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy. The denoising operation is as follows:

[0013] The collected vibration signal is non-stationary, multi-band mixed and strongly noise-interfered. Dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy are used to denoise the collected vibration signal.

[0014] Specifically, the collected vibration signal is first subjected to dynamic wavelet packet decomposition to obtain the wavelet packet coefficients of each layer. Then, the local frequency band energy of each layer of coefficients is calculated through a sliding time window, and the noise floor energy of the front section of the collected vibration signal without explosion is extracted. Finally, the wavelet packet coefficients of each layer are dynamically weighted according to the ratio of the local frequency band energy to the noise floor to reconstruct the denoised signal.

[0015] When dividing the denoised vibration signal into tasks, the energy sequence of each sample frequency band is first divided by time segment. The energy difference between the time segments of any two samples is then calculated. The sample similarity matrix is then constructed based on the difference in blasting parameters. Finally, the similarity matrix is subjected to feature decomposition and clustering using a spectral clustering algorithm. The samples are then divided into different meta-tasks. Each meta-task is a set of samples with similar geological conditions and blasting parameters. The specific steps are as follows:

[0016] (1) Calculate the frequency band energy distribution difference measure of the sample and construct the energy distribution matrix;

[0017] The calculation formula for the difference measure of the frequency band energy distribution of any two samples is as follows:

[0018] ,

[0019] in, is the energy distribution matrix The element of Samples and The difference measure of frequency band energy distribution of samples; For time segments; For the The sample in The frequency band energy of the time slice; For the The sample in The frequency band energy of the time slice;

[0020] Energy distribution matrix is represented as follows:

[0021] ,

[0022] in, represents the number of samples, , ,when hour, ;

[0023] (2) Calculate the similarity between samples and construct a similarity matrix;

[0024] The formula for calculating the similarity between any two samples is as follows:

[0025] ,

[0026] in, is the similarity matrix The element of The sample and The similarity of samples; is an exponential function with a natural constant as its base; is the similarity attenuation coefficient; is an indicative function. When the parameter difference of the indicative function is less than the preset blasting parameter difference threshold Take 1 when it is, otherwise take 0; is the preset blasting parameter difference threshold; For the The sample and Quantitative value of the difference in blasting parameters of samples;

[0027] Similarity Matrix is represented as follows:

[0028] ,

[0029] Among them, when hour, ;

[0030] (3) Generate meta-tasks using spectral clustering algorithm:

[0031] Similarity matrix Perform Laplace matrix eigendecomposition and select K-means clustering is performed on the feature vectors to assign similar samples to the same meta-task, ensuring that the samples within the task have similar geological-blasting conditions.

[0032] S3 is as follows:

[0033] The neural network model is used to extract features from samples in each meta-task. The feature extraction process includes initializing the parameters of the neural network model, performing dual-path feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy attenuation curve to optimize the meta-parameters in the meta-task.

[0034] (1) Initialize the parameters of the neural network model: A neural network initialization method based on vibration energy distribution is used. The samples in each meta-task are mapped into low-dimensional embedding vectors through a pre-trained autoencoder neural network. The mean of the embedding vectors in each meta-task is calculated as the prototype of each meta-task. Then, the initial weights of the network are generated by weighted fusion based on the statistical distribution of the prototype vectors of each meta-task and the historical task parameter offsets.

[0035] (2) Perform feature fusion after dual-path feature extraction in the time domain and frequency domain: adopt a dual-path interaction structure to define the structure of the dual-path feature fusion neural network, including the time domain path neural network and the frequency domain path neural network;

[0036] 1) Time-domain pathway neural network: Perform one-dimensional convolution on samples in the meta-task to extract time-domain features, and fuse frequency-domain features through a cross-pathway attention gating mechanism to obtain the output features of the time-domain pathway neural network;

[0037] 2) Frequency-domain pathway neural network: Perform a short-time Fourier transform on the samples in the meta-task, encode the frequency-domain features using a multi-layer perceptron, and adaptively scale the frequency-band weights based on the frequency-band energy to obtain the output features of the frequency-domain pathway neural network.

[0038] 3) Dual-pathway feature fusion: The output features of the time-domain neural network and the frequency-domain neural network are cross-pathway interacted, feature cross-enhancement is performed through element-by-element multiplication, and then feature complementary fusion is achieved through element-by-element addition, ultimately generating a representation vector containing the joint time-frequency features;

[0039] (3) Meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task:

[0040] A dynamic learning rate adjustment strategy is adopted. First, the energy peak moment of each vibration signal is extracted, and the average energy attenuation ratio in the time period after the peak is calculated as the attenuation coefficient. When the inner loop parameters of the neural network model are updated, the basic learning rate is smoothly adjusted through the hyperbolic tangent function according to the signal time point and attenuation coefficient corresponding to the current training step, so that the parameter update intensity is consistent with the energy attenuation trend. During the outer loop update, the gradients of multiple tasks are aggregated and the meta-parameters are optimized.

[0041] The parameters of the neural network model are initialized as follows:

[0042] (1) Calculate the meta-task prototype vector: For all samples in each meta-task, map them into low-dimensional embedding representations through a pre-trained autoencoder neural network, and calculate the mean of the embedding representation within the task as the task prototype;

[0043] (2) Generate initialization parameters: Based on the meta-task prototype vector and historical parameter offset, generate the network initial weights through weighted fusion, and then generate the network initialization weight matrix.

[0044] The meta-parameters in the meta-task are optimized by meta-learning based on the vibration energy attenuation curve as follows:

[0045] (1) Constructing the energy attenuation coefficient: First, locate the energy peak point of each vibration signal, then count the relative energy per unit time in the time period after the peak, calculate its ratio to the peak energy, and then construct the overall energy attenuation coefficient;

[0046] (2) Dynamically adjust the inner loop learning rate:

[0047] The inner loop represents the parameter update within the meta-task, and dynamically adjusts the learning rate according to the energy decay coefficient. First, the time point corresponding to the current step is obtained, and the energy decay coefficient is input into the hyperbolic tangent function to generate a smoothed adjustment value. The smoothed adjustment value is then multiplied by the base learning rate to obtain the adjusted value of the base learning rate. The base learning rate and the base learning rate adjustment value are then added to obtain the inner loop adaptive learning rate.

[0048] (3) External loop parameter update:

[0049] The outer loop represents the optimization of meta-parameters. By performing inner loop training on multiple meta-tasks separately, the specific model parameters of each meta-task are obtained, the loss function of all meta-tasks on the support set is calculated, the neural network model parameters are back-propagated, and the meta-parameters are updated according to the optimization step size, thereby obtaining the meta-parameters to be updated in the outer loop.

[0050] S4 is as follows:

[0051] A hierarchical attention mechanism is adopted. According to the event response characteristics of different samples at different time scales, attention weights are extracted using a short-time window convolutional neural network, a medium-time window long short-term memory neural network, and a long-time window global pooling perception neural network. The attention weights are respectively focused on instantaneous, event-level, and global vibration changes. The attention vectors at each scale are weighted and superimposed on the fusion features element by element to generate a multi-scale classification feature vector. The classification feature vector is then input into a preset Softmax function to obtain a predicted class probability vector. The class corresponding to the largest class probability in the class probability vector is taken as the classification prediction result of the impact of blasting on surface building vibration.

[0052] The short-term window convolutional neural network implements the short-term window attention enhancement operation and sets , the calculation formula is as follows:

[0053] ;

[0054] The medium-time window long short-term memory neural network implements the medium-time window attention enhancement operation, setting , the calculation formula is as follows:

[0055] ;

[0056] The long-term window global pooling perception neural network implements long-term attention enhancement operation, setting , the calculation formula is as follows:

[0057] ;

[0058] The calculation formula for multi-scale feature aggregation is as follows:

[0059] ;

[0060] in, is a representation vector containing time-frequency joint features; is the time window length of the attention mechanism; is the short-term window attention weight; is the Softmax function; is the short-time window convolution kernel parameter; It is a long short-term memory neural network; Output of the long short-term memory neural network The transpose of is the attention weight of the mid-time window; is the linear transformation parameter of the medium time window; is the long-term window attention weight; is the multi-layer perceptron function; is the Sigmoid activation function; Represents element-wise multiplication; It is the global maximum pooling operation; is the classification feature vector after aggregating multi-scale attention.

[0061] The generalization verification and dynamic adjustment of meta-tasks are as follows:

[0062] The meta-task is divided into a training set and a validation set. The support set is used in the inner loop to quickly fine-tune the neural network model parameters. The meta-parameters are updated based on the query set loss in the outer loop.

[0063] During the training process, the performance of the neural network model is periodically evaluated on the validation meta-task. The loss calculation in the neural network model uses cross-entropy loss. The loss function is calculated once during each training. If the loss does not decrease after P consecutive cycles, the outer loop learning rate is reduced exponentially.

[0064] Finally, the optimal hyperparameter combination of the neural network model is determined through multi-task cross-validation, and the classification accuracy and generalization are verified on an independent test set.

[0065] The closed-loop feedback control and parameter optimization are as follows:

[0066] Combining the classification prediction results with the blasting parameter adjustment rule base to generate real-time control instructions for adjusting blasting parameters;

[0067] The blasting parameter adjustment rule base includes expert experience or physical models. The calculation formula of the adjusted blasting parameters is as follows:

[0068] ,

[0069] in, is the adjusted blasting parameter; is the benchmark parameter; is the category sensitivity coefficient fitted by historical data; The vibration category predicted by the model; This is the category corresponding to the safety vibration threshold.

[0070] The strategy for online optimization of blasting parameters through reinforcement learning strategy is as follows:

[0071] Using a reinforcement learning strategy, with classification accuracy and vibration energy control as reward functions, the blasting parameter strategy is optimized online. The state space, action space, and reward function are defined. The state space includes vibration signal characteristics and geological parameters, the action space includes blasting parameter adjustments, and the reward function is the sum of vibration energy below a threshold and classification accuracy.

[0072] The proximal policy optimization method is used to train the reinforcement learning policy network, and the model parameters are updated together with the meta-learning model.

[0073] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects:

[0074] The present invention uses dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy to accurately separate noise and effective vibration signals, and can dynamically adjust the weights of wavelet packet coefficients to avoid the loss of effective signals that may be caused by fixed threshold wavelet denoising methods.

[0075] The meta-task division method based on spectral clustering proposed in this paper generates meta-tasks based on the energy distribution characteristics of vibration signals and the differences in burst parameters. This ensures that samples of similar vibration modes are clustered together, effectively overcoming the inter-class differences caused by random task division in traditional meta-learning, and improving the rationality of task division and model training efficiency.

[0076] The present invention uses a pre-trained autoencoder neural network to map the denoised vibration signal into a low-dimensional embedding vector. The initial weights of the network are generated by weighted fusion by calculating the mean of the task prototype and the historical parameter offset. This can accelerate the convergence of the network and reduce the risk of overfitting.

[0077] The present invention uses a dual-pathway neural network structure to extract the time domain features and frequency domain features of the vibration signal respectively, and then fuses the two features through a cross-pathway attention mechanism, which can improve the signal representation ability and more comprehensively capture the instantaneous characteristics and long-term resonance characteristics of the vibration signal.

[0078] The present invention adopts a dynamic learning rate adjustment strategy based on the vibration energy decay curve. By capturing the energy decay law of the vibration signal, the learning rate is automatically adjusted to adapt to the temporal energy distribution of the signal, significantly improving learning efficiency and preventing over-updates or slow convergence.

[0079] The present invention adopts a multi-scale attention mechanism to extract the key features of vibration signals at different time scales through a short-time window convolutional neural network, a medium-time window long short-term memory neural network, and a long-time window global pooling neural network, which can enhance the response capability at different time scales.

[0080] The present invention proposes a real-time control instruction generation mechanism that combines classification prediction results with a blasting parameter adjustment rule library, and in an optional strategy, utilizes reinforcement learning to further optimize blasting parameters, thereby realizing intelligent parameter adjustment.

[0081] In summary, the present invention can improve the accuracy of vibration impact on surface buildings and the safety of construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0083] Figure 1 Schematic diagram of the method of the present invention.

[0084] Figure 2 A comparison chart of the effects of different denoising methods.

[0085] Figure 3 is the collected vibration signal.

[0086] Figure 4 The figure shows the comparison between the denoised vibration signal and the real signal.

[0087] Figure 5 Comparison of meta-learning effects for different task partitioning strategies.

[0088] Figure 6 Comparison chart of learning rate adjustment strategies. DETAILED DESCRIPTION

[0089] In order to clearly illustrate the technical features of this solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0090] Example 1

[0091] like Figure 1 As shown, a method for predicting the vibration impact of tunnel excavation blasting on surface buildings includes the following steps:

[0092] S1. Using a vibration sensor to collect vibration signals generated by blasting during tunneling, and then performing a denoising operation on the collected vibration signals, using the denoised vibration signals as samples;

[0093] S2. Divide the denoised vibration signal into tasks and use a spectral clustering algorithm to generate meta-tasks based on the energy distribution characteristics of the vibration signal. Each meta-task is a sample set.

[0094] S3. Extract features from samples in each meta-task using a neural network model. The feature extraction process includes initializing the parameters of the neural network model, performing dual-path feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy attenuation curve to optimize the meta-parameters in the meta-task.

[0095] S4. A hierarchical attention mechanism is used to capture the temporal response characteristics of vibration signals at different time lengths, obtaining the key features of each stage. The key features of each stage are then fused to obtain a classification feature vector. The classification prediction results of the impact of blasting on surface building vibration are then determined based on the classification feature vector.

[0096] S5. Optimize the classification prediction process, including generalization verification and dynamic adjustment of meta-tasks, optimization of closed-loop feedback control and parameters, and online optimization of blasting parameter strategies through reinforcement learning strategies.

[0097] In a specific embodiment, the denoising operation adopts dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy. The denoising operation is as follows:

[0098] The collected vibration signal is non-stationary, multi-band mixed and strongly noise-interfered. Dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy are used to denoise the collected vibration signal.

[0099] Specifically, the collected vibration signal is first subjected to dynamic wavelet packet decomposition to obtain the wavelet packet coefficients of each layer. Then, the local frequency band energy of each layer of coefficients is calculated through a sliding time window, and the noise floor energy of the front section of the collected vibration signal without explosion is extracted. Finally, the wavelet packet coefficients of each layer are dynamically weighted according to the ratio of the local frequency band energy to the noise floor to reconstruct the denoised signal.

[0100] The calculation formula for the denoising operation is as follows:

[0101] ,

[0102] in, is the vibration signal after denoising; For the Layer wavelet packet decomposition coefficients characterize the detailed information of vibration signals at different scales; is the total number of layers of wavelet packet decomposition; For the Layer wavelet packet at time The local frequency band energy of the blasting vibration signal is determined by calculating the energy difference, and the instantaneous energy fluctuation of the blasting vibration signal is dynamically captured, such as the high-frequency impact energy at the moment of explosive detonation. For the The noise floor energy corresponding to the layer is used to separate effective vibration features from the vibration signal. For example, the energy average of the non-burst period before the vibration signal for 0.1 seconds can be taken to represent the inherent noise and background vibration of the sensor. For the Layer wavelet packet decomposition coefficients in time The reconstruction component at ; Represents the time index of the vibration signal; For the Layer wavelet packet basis function, representing the decomposition and reconstruction of vibration signals, Characterize the wavelet packet components after time window adjustment; is the noise base energy, which is the average energy of the non-explosion section before 0.1s of the signal; To prevent division by zero, set ;

[0103] is the center point of the time window, and the calculation formula is:

[0104] ;

[0105] is the time window width, and the calculation formula is:

[0106] ;

[0107] is the local frequency band energy, and the calculation formula is:

[0108] ;

[0109] in, The time window length for local energy calculation; Represents the sliding time index within the time window; For the The decomposition coefficients of the layer wavelet packet are in the sliding window The value at

[0110] It should be noted that the time window center point and time window width are adaptively determined based on the energy distribution of wavelet packet coefficients, which can accurately locate the time-frequency locality of blasting events, such as multi-stage vibration caused by delayed explosion. Compared with the fixed time window wavelet decomposition, the dynamic adjustment of the time window can capture the frequency band characteristics of different fracture stages of the rock mass.

[0111] In a specific implementation, when performing task division on the denoised vibration signal, the energy sequence of each sample frequency band is first divided by time segment. Then, the energy difference between the time segments of any two samples is calculated. The similarity matrix of the samples is then constructed in combination with the difference in blasting parameters. Finally, the similarity matrix is subjected to feature decomposition and clustering using a spectral clustering algorithm. The samples are then divided into different meta-tasks. Each meta-task is a set of samples with similar geological conditions and blasting parameters. The specific steps are as follows:

[0112] (1) Calculate the frequency band energy distribution difference measure of the sample and construct the energy distribution matrix;

[0113] The calculation formula for the difference measure of the frequency band energy distribution of any two samples is as follows:

[0114] ,

[0115] in, is the energy distribution matrix The element of The sample and The frequency band energy distribution difference measurement of the samples characterizes the similarity of the energy distribution of the two samples and quantifies the impact of geological conditions on the vibration mode, such as the difference in energy release rate between hard rock and soft rock. is the number of time segments, representing the granularity of time segment division; For the The sample in The frequency band energy of the time slice; For the The sample in The frequency band energy of the time slice;

[0116] Energy distribution matrix is represented as follows:

[0117] ,

[0118] in, represents the number of samples, , ,when hour, ;

[0119] It should be noted that because different geological conditions and blasting parameters lead to different vibration patterns, traditional random task division methods cannot effectively capture these differences. By sequentially calculating the frequency band energy differences of vibration signals, we can better divide tasks based on the differences in geological conditions and blasting parameters, ensuring that similar samples are clustered together, which helps meta-learning quickly adapt under specific conditions.

[0120] (2) Calculate the similarity between samples and construct a similarity matrix;

[0121] The formula for calculating the similarity between any two samples is as follows:

[0122] ,

[0123] in, is the similarity matrix The element of The sample and The similarity of samples; is an exponential function with a natural constant as its base; is the similarity attenuation coefficient, set ; is an indicative function. When the parameter difference of the indicative function is less than the preset blasting parameter difference threshold Take 1 when it is, otherwise take 0; For the preset blasting parameter difference threshold, set ; For the The sample and The quantitative value of the blasting parameter difference of each sample is used to characterize the difference in blasting parameters between two samples, and to force samples in the same meta-task to have similar blasting conditions, such as the difference in blasting parameters such as explosive quantity and rock mass grade;

[0124] Similarity Matrix is represented as follows:

[0125] ,

[0126] Among them, when hour, ;

[0127] (3) Generate meta-tasks using spectral clustering algorithm:

[0128] Similarity matrix Perform Laplace matrix eigendecomposition and select K-means clustering is performed on the feature vectors to assign similar samples to the same meta-task, ensuring that the samples within the task have similar geological-blasting conditions.

[0129] It should be noted that aggregating samples of similar geological-blasting conditions into the same meta-task can ensure the consistency of data within the task, thereby improving the model's ability to quickly adapt to different tasks during the meta-learning process and providing a structured task division basis for subsequent parameter initialization.

[0130] In a specific implementation, S3 is as follows:

[0131] The neural network model is used to extract features from samples in each meta-task. The feature extraction process includes initializing the parameters of the neural network model, performing dual-path feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy attenuation curve to optimize the meta-parameters in the meta-task.

[0132] (1) Initialize the parameters of the neural network model: A neural network initialization method based on vibration energy distribution is used. The samples in each meta-task are mapped into low-dimensional embedding vectors through a pre-trained autoencoder neural network. The mean of the embedding vectors in each meta-task is calculated as the prototype of each meta-task. Then, the initial weights of the network are generated by weighted fusion based on the statistical distribution of the prototype vectors of each meta-task and the historical task parameter offsets.

[0133] (2) Perform feature fusion after dual-path feature extraction in the time domain and frequency domain: adopt a dual-path interaction structure to define the structure of the dual-path feature fusion neural network, including the time domain path neural network and the frequency domain path neural network;

[0134] 1) Time-domain pathway neural network: Perform one-dimensional convolution on samples in the meta-task to extract time-domain features, and fuse frequency-domain features through a cross-pathway attention gating mechanism to obtain the output features of the time-domain pathway neural network;

[0135] The calculation formula of the time domain pathway neural network is as follows:

[0136] ,

[0137] in, is the neural network output of the time domain pathway; Represents a one-dimensional convolution operation; is the convolution kernel parameter; It means element-by-element addition;

[0138] It is the feature generated by concatenating the time domain and frequency domain features, performing linear transformation and Sigmoid activation. The calculation formula is as follows:

[0139] ,

[0140] in, is the Sigmoid activation function; The frequency domain feature vector after multi-layer perceptron encoding of the output of the frequency domain pathway neural network; is the gate weight matrix, which controls the fusion ratio of time domain and frequency domain features; Represents vector concatenation; Represents element-wise multiplication;

[0141] 2) Frequency-domain pathway neural network: Perform a short-time Fourier transform on the samples in the meta-task, encode the frequency-domain features using a multi-layer perceptron, and adaptively scale the frequency-band weights based on the frequency-band energy to obtain the output features of the frequency-domain pathway neural network.

[0142] The calculation formula of the frequency domain pathway neural network is as follows:

[0143] ,

[0144] in, is the frequency domain pathway neural network output; is the multi-layer perceptron function; is the real part operation; is the short-time Fourier transform;

[0145] Indicates the frequency band scaling factor, which is calculated as follows:

[0146] ,

[0147] in, is the adjustment coefficient of the frequency band scaling strength; For the Energy of the frequency band; For the historical data Band energy mean; is the frequency band index;

[0148] 3) Dual-pathway feature fusion: The output features of the time-domain neural network and the frequency-domain neural network are cross-pathway interacted, feature cross-enhancement is performed through element-by-element multiplication, and then feature complementary fusion is achieved through element-by-element addition, ultimately generating a representation vector containing time-frequency joint features.

[0149] The calculation formula for dual-path feature fusion is as follows:

[0150] ,

[0151] in, is the fused feature vector.

[0152] It should be noted that in the classification task of tunnel excavation blasting vibration data, the combination of time domain and frequency domain features can fully reflect the changing pattern of blasting vibration. Dual-path feature fusion can simultaneously capture the characteristics of time domain and frequency domain, and then extract the instantaneous characteristics and long-term resonance characteristics of the signal, thereby improving the characterization ability of the vibration signal.

[0153] (3) Meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task:

[0154] A dynamic learning rate adjustment strategy is adopted. First, the energy peak moment of each vibration signal is extracted, and the average energy attenuation ratio in the time period after the peak is calculated as the attenuation coefficient. When the inner loop parameters of the neural network model are updated, the basic learning rate is smoothly adjusted through the hyperbolic tangent function according to the signal time point and attenuation coefficient corresponding to the current training step, so that the parameter update intensity is consistent with the energy attenuation trend. During the outer loop update, the gradients of multiple tasks are aggregated and the meta-parameters are optimized.

[0155] In a specific implementation, the parameters of the neural network model are initialized as follows:

[0156] (1) Calculate the meta-task prototype vector: For all samples in each meta-task, map them into low-dimensional embedding representations through a pre-trained autoencoder neural network, and calculate the mean of the embedding representation within the task as the task prototype;

[0157] The calculation formula of the meta-task prototype vector is as follows:

[0158] ,

[0159] in, For the a meta-task sample set; Embedding representation for pre-trained autoencoder neural networks; For the The denoised signal of samples; For the Prototype vectors of meta-tasks;

[0160] The prototype vector is a representative of the feature vectors of all samples in a certain category or task. In meta-learning, it is obtained by calculating the low-dimensional embedding average of all samples in the task, reflecting the overall characteristic pattern of the task. For example, assuming that a certain meta-task contains 5 vibration signal samples, each sample is mapped to a 10-dimensional vector by an autoencoder neural network, then the prototype vector is the element-by-element average of these 5 10-dimensional vectors. The meta-task is the basic training unit of meta-learning. Each task contains a group of related samples, simulating sub-problems in real scenarios. For example, through spectral clustering, samples with similar geological conditions and blasting parameters are aggregated into the same task. If the vibration signal samples of a tunnel section with a rock mass grade of Grade III and an explosives amount of 200kg are clustered into the same meta-task, the model learns how to quickly adapt under such conditions.

[0161] (2) Generate initialization parameters: Based on the meta-task prototype vector and historical parameter offset, the network initial weights are generated through weighted fusion, and then the network initialization weight matrix is generated. Compared with the single meta-parameter of MAML, weighted fusion can inherit cross-task common knowledge;

[0162] The calculation formula of the network initialization weight matrix is as follows:

[0163] ,

[0164] in, For the task The number of samples; Initialize the weight matrix for the network; is the total number of meta-tasks; is the total number of samples; For the Prototype vectors of meta-tasks; represents the Kronecker product; is the task-related orthogonal matrix; is the fusion coefficient, taking the statistical mean of the historical task parameter offsets; The offset of historical task parameters;

[0165] The historical task parameter offset is calculated by integrating the historical task parameter update trajectory to avoid overfitting of the model on the new task. For example, the difference in rock mass stiffness in different tunnel sections requires elastic parameter adjustment. The calculation formula of the historical task parameter offset is as follows:

[0166] ,

[0167] in, is the number of historical tasks; is a positive integer; For the The final training parameters of the historical tasks; For the The initial parameters of the historical task.

[0168] In a specific embodiment, meta-learning based on the vibration energy attenuation curve is performed to optimize the meta-parameters in the meta-task as follows:

[0169] (1) Constructing the energy attenuation coefficient: First, locate the energy peak point of each vibration signal, then count the relative energy per unit time in the time period after the peak, calculate its ratio to the peak energy, and then construct the overall energy attenuation coefficient, which can measure the time series energy dissipation speed of the vibration signal;

[0170] The energy attenuation coefficient is calculated as follows:

[0171] ,

[0172] in, is the energy attenuation coefficient, is the total duration of the signal; For the moment Vibration energy, vibration energy refers to the energy intensity of the vibration signal within a certain period of time, and the vibration energy is expressed by the root mean square value; It is the vibration energy at the peak moment of energy; It is the moment of peak vibration energy;

[0173] (2) Dynamically adjust the inner loop learning rate:

[0174] The inner loop represents the parameter update within the meta-task, and dynamically adjusts the learning rate according to the energy decay coefficient. First, the time point corresponding to the current step is obtained, and the energy decay coefficient is input into the hyperbolic tangent function to generate a smooth adjustment value. The smooth adjustment value is then multiplied by the basic learning rate to obtain the adjustment value of the basic learning rate. The basic learning rate and the basic learning rate adjustment value are then added to obtain the inner loop adaptive learning rate. After the parameters of the inner loop are updated, they can better adapt to the temporal distribution of the current signal energy.

[0175] The calculation formula of the inner loop adaptive learning rate is as follows:

[0176] ,

[0177] in, is the adaptive learning rate of the inner loop; As the basic learning rate, set ; is the attenuation adjustment coefficient, set ; is the number of training steps; is the hyperbolic tangent function; Indicates the smoothing adjustment value;

[0178] The inner loop adaptive learning rate can be adjusted according to the energy attenuation characteristics of the vibration signal. By adaptively adjusting the learning rate, the model can be optimized according to the vibration energy attenuation characteristics of different time periods, thereby improving learning efficiency and avoiding excessive updates or slow convergence;

[0179] (3) External loop parameter update:

[0180] The outer loop represents the optimization of meta-parameters. By executing the inner loop training on multiple meta-tasks, the specific model parameters of each meta-task are obtained. The loss function of all meta-tasks on the support set is calculated, the neural network model parameters are back-propagated, and the meta-parameters are updated according to the optimization step size. The meta-parameters to be updated in the outer loop are obtained, which can make the neural network model more adaptable to new original tasks.

[0181] The calculation formula for the outer loop parameter update is as follows:

[0182] ,

[0183] in, It is the parameter update operation; The meta parameters to be updated in the outer loop; For the outer loop learning rate, set ; represents the gradient of the parameter; is a positive integer; For the The loss function on each task; The task-specific model after the inner loop update;

[0184] The meta-parameters are the parameters optimized by the model in the outer loop, which determine the initial performance of the model on the new task. Through multi-task training, the meta-parameters enable the model to quickly adapt to new tasks with the prior knowledge.

[0185] The task-specific model updated by the inner loop is the task-specific model obtained by fine-tuning the neural network model on the support set of a single task after initialization based on the meta-parameters in the inner loop. For example, if the support set of a task has 10 samples, the inner loop adjusts the parameters through 3 steps of gradient descent to obtain a specific model adapted to the task. , whose loss is used to update the meta-parameters.

[0186] In a specific implementation manner, S4 is specifically as follows:

[0187] The hierarchical attention mechanism can more comprehensively capture the key features of each stage in the building response process. According to the event response characteristics of different samples at different time scales, a short-term window convolutional neural network, a medium-term window long short-term memory neural network, and a long-term window global pooling perception neural network are used to extract attention weights, focusing on instantaneous, event-level, and global vibration changes respectively. The attention vectors at each scale are weighted and superimposed with the fusion features element by element to generate a multi-scale classification feature vector. The classification feature vector is then input into the preset Softmax function to obtain the predicted class probability vector. The class corresponding to the largest class probability in the class probability vector is taken as the classification prediction result of the impact of blasting on surface building vibration.

[0188] The short-term window convolutional neural network implements the short-term window attention enhancement operation and sets , the calculation formula is as follows:

[0189] ;

[0190] The medium-time window long short-term memory neural network implements the medium-time window attention enhancement operation, setting , the calculation formula is as follows:

[0191] ;

[0192] The long-term window global pooling perception neural network implements long-term attention enhancement operation, setting , the calculation formula is as follows:

[0193] ;

[0194] The calculation formula for multi-scale feature aggregation is as follows:

[0195] ;

[0196] in, is the time window length of the attention mechanism; is the short-term window attention weight; is the Softmax function; is the short-time window convolution kernel parameter; It is a long short-term memory neural network; Output of the long short-term memory neural network The transpose of is the attention weight of the mid-time window; is the linear transformation parameter of the medium time window; is the long-term window attention weight; is the Sigmoid activation function; It is the global maximum pooling operation; is the classification feature vector after aggregating multi-scale attention.

[0197] The mid-time window linear transformation parameters are used to map the temporal features output by the long short-term memory neural network to the attention weight space to capture the event correlation within the mid-time window;

[0198] The short-time window attention weight focuses on instantaneous vibration changes, the medium-time window attention weight focuses on event-level vibration changes, and the long-time window attention weight focuses on global vibration changes.

[0199] In a specific implementation, generalization verification and dynamic adjustment of meta-tasks are performed as follows:

[0200] The meta-task is divided into a training set and a validation set. The support set is used in the inner loop to quickly fine-tune the neural network model parameters. The meta-parameters are updated based on the query set loss in the outer loop.

[0201] During the training process, the performance of the neural network model is periodically evaluated on the validation meta-task. The loss calculation in the neural network model uses cross-entropy loss. The loss function is calculated once during each training. If the loss does not decrease after P consecutive cycles, the outer loop learning rate is reduced exponentially.

[0202] Finally, the optimal hyperparameter combination of the neural network model is determined through multi-task cross-validation, and the classification accuracy and generalization are verified on an independent test set.

[0203] In a specific embodiment, closed-loop feedback control and parameter optimization are performed as follows:

[0204] Combining the classification prediction results with the blasting parameter adjustment rule base to generate real-time control instructions for adjusting blasting parameters;

[0205] The blasting parameter adjustment rule base includes expert experience or physical models. The calculation formula of the adjusted blasting parameters is as follows:

[0206] ,

[0207] in, is the adjusted blasting parameter; is the benchmark parameter; is the category sensitivity coefficient fitted by historical data; The vibration category predicted by the model; This is the category corresponding to the safety vibration threshold.

[0208] In a specific implementation, the strategy for online optimization of blasting parameters through reinforcement learning strategy is as follows:

[0209] Using a reinforcement learning strategy, with classification accuracy and vibration energy control as reward functions, the blasting parameter strategy is optimized online. The state space, action space, and reward function are defined. The state space includes vibration signal characteristics and geological parameters, the action space includes blasting parameter adjustments, and the reward function is the sum of vibration energy below a threshold and classification accuracy.

[0210] The proximal policy optimization method is used to train the reinforcement learning policy network, and the model parameters are updated together with the meta-learning model.

[0211] Example 2

[0212] like Figures 2 to 4 As shown, the performance of the denoising method in the present invention is verified; Figure 2 As shown in FIG, it is a comparison chart of the effects of different denoising methods, comparing the Fourier filter denoising method and the fixed prefabricated wavelet denoising method with the denoising method of the present invention. Figure 2The signal-to-noise ratio indicators of different methods are displayed through bar graphs, and the line graph reflects the accuracy of the denoised signal in the classification task. By comparing the present invention with conventional Fourier filtering and fixed threshold wavelet denoising methods, it can be intuitively seen that the dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy of the present invention have advantages in retaining effective vibration components and suppressing noise. The denoising method of the present invention is significantly better than the other two methods in terms of signal-to-noise ratio indicators, and its bar graph height is significantly higher; at the same time, the classification accuracy curve corresponding to the method of the present invention in the line graph is always at the highest position, indicating that the strategy of dynamically adjusting the weight of the wavelet packet coefficient layer can not only effectively separate noise, but also avoid the loss of effective signals caused by the fixed threshold of the traditional method.

[0213] like Figure 3 and Figure 4 As shown, Figure 3 is the collected vibration signal, Figure 4 For the comparison between the denoised vibration signal and the real signal, Figure 3 The vibration signal in the image shows a typical strong noise interference feature. There is a continuous high-frequency noise base in the front part of the signal (0-0.3 seconds), which causes the waveform details of the effective vibration component (0.5-0.8 seconds) to be submerged. The pulse vibration (around 1.2 seconds) is mixed with random noise and is difficult to identify. Figure 4 By comparing the denoised vibration signal (blue curve) with the real signal (red dotted line), it can be seen that the dynamic wavelet packet decomposition accurately separates the noise and the effective components: the noise base area is completely suppressed, and the front part of the waveform tends to be flat; the vibration details of the high-frequency resonance segment (such as the waveform peak at 0.6 seconds) are completely preserved, which is highly consistent with the real signal; the amplitude and phase of the pulse vibration component (1.2 seconds) are not distorted, and the local magnification window further shows that the edges of the transient features are clear. The denoised waveform retains the physical vibration laws while eliminating the common problems of over-smoothing or loss of details in traditional methods. The dynamic wavelet packet decomposition of the present invention solves the defect of fixed threshold wavelet denoising that is not sensitive enough to transient features by adaptively adjusting the time window width and inter-layer weights. Figure 4 The ability is verified by the degree of detail restoration of the high-frequency resonance segment in the local magnified contrast area. The local signal-to-noise ratio weighting strategy dynamically calculates the noise floor energy, suppressing the continuous noise in the previous segment while avoiding the frequency domain confusion problem of traditional Fourier filtering on non-stationary signals. The adaptive calculation of the time window parameters ensures the complete extraction of the pulse vibration component, and its peak retention rate is better than that of conventional methods, reflecting the precise modeling of the signal energy attenuation characteristics, providing high-fidelity vibration features for subsequent classification tasks.

[0214] Example 3

[0215] like Figure 5As shown in Figure 2, the meta-learning effects of different task partitioning strategies are analyzed to verify the role of the spectral clustering algorithm based on energy distribution similarity in improving the efficiency of meta-learning. The changes in the verification set accuracy of the three strategies of spectral clustering partitioning, random partitioning and traditional clustering partitioning during the training cycle are compared through three training curves of different colors. Figure 5 It can be seen that the curve using spectral clustering has the fastest convergence speed and reaches a stable high value in the middle of training, while the curve using random division has large fluctuations and slow convergence. Figure 5 The experimental results show that since spectral clustering constructs a similarity matrix by combining signal energy distribution and blasting parameter differences, the smooth upward trend of the curve indicates that the samples within the task are highly consistent, enabling the model to quickly capture the correlation between geological conditions and blasting parameters, avoiding the oscillation phenomenon caused by unreasonable task division in traditional methods.

[0216] Example 4

[0217] like Figure 6 As shown in Figure 2, the effects of fixed learning rate and dynamic adjustment strategy based on energy decay on the training process are compared by combining scatter plot and kernel density estimation. Figure 6 The horizontal axis represents the training progress, and the vertical axis represents the model performance. The scatter distribution of the dynamic strategy shows an obvious clustering trend in the upper right corner, and the kernel density colored area is more concentrated, indicating that it can maintain stable improvement in different training stages. In contrast, the scatter distribution of the fixed learning rate is scattered and there is a phenomenon of performance decline in the later stage, indicating that the fixed learning rate method does not consider the characteristic that the energy of the vibration signal decays over time; the present invention incorporates the energy decay coefficient into the learning rate adjustment through the hyperbolic tangent function. Figure 6 As shown in the figure, the gradient color arrows match the parameter update strength with the physical characteristics of the signal, thus achieving a more robust training process under complex vibration patterns.

[0218] Although the above describes the specific implementation methods of the invention in conjunction with the accompanying drawings, it does not limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting the vibration effect of tunnel excavation blasting on surface buildings, characterized by: The following steps are involved: S1. Using a vibration sensor to collect vibration signals generated by blasting during tunneling, and then performing a denoising operation on the collected vibration signals, using the denoised vibration signals as samples; S2. Divide the denoised vibration signal into tasks and use a spectral clustering algorithm to generate meta-tasks based on the energy distribution characteristics of the vibration signal. Each meta-task is a sample set. S3. Extract features from samples in each meta-task using a neural network model. The feature extraction process includes initializing the parameters of the neural network model, performing dual-path feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy attenuation curve to optimize the meta-parameters in the meta-task. S4. A hierarchical attention mechanism is used to capture the temporal response characteristics of vibration signals at different time lengths, obtaining the key features of each stage. The key features of each stage are then fused to obtain a classification feature vector. The classification prediction results of the impact of blasting on surface building vibration are then determined based on the classification feature vector. S5. Optimize the classification prediction process, including generalization verification and dynamic adjustment of meta-tasks, optimization of closed-loop feedback control and parameters, and online optimization of blasting parameter strategies through reinforcement learning strategies.

2. The method for predicting the vibration effect of tunnel excavation blasting on surface buildings according to claim 1, characterized in that: The denoising operation adopts dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy. The denoising operation is as follows: The collected vibration signal is non-stationary, multi-band mixed and strongly noise-interfered. Dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy are used to denoise the collected vibration signal. Specifically, the collected vibration signal is first subjected to dynamic wavelet packet decomposition to obtain the wavelet packet coefficients of each layer. Then, the local frequency band energy of each layer of coefficients is calculated through a sliding time window, and the noise floor energy of the front section of the collected vibration signal without explosion is extracted. Finally, the wavelet packet coefficients of each layer are dynamically weighted according to the ratio of the local frequency band energy to the noise floor to reconstruct the denoised signal.

3. The method for predicting the vibration effect of tunnel excavation blasting on surface buildings according to claim 2, characterized in that: When dividing the denoised vibration signal into tasks, the energy sequence of each sample frequency band is first divided by time segment. The energy difference between the time segments of any two samples is then calculated. The sample similarity matrix is then constructed based on the difference in blasting parameters. Finally, the similarity matrix is subjected to feature decomposition and clustering using a spectral clustering algorithm. The samples are then divided into different meta-tasks. Each meta-task is a set of samples with similar geological conditions and blasting parameters. The specific steps are as follows: (1) Calculate the frequency band energy distribution difference measure of the sample and construct the energy distribution matrix; The calculation formula for the difference measure of the frequency band energy distribution of any two samples is as follows: ， in, is the energy distribution matrix The element of Samples and The difference measure of frequency band energy distribution of samples; For time segments; For the The sample in The frequency band energy of the time slice; For the The sample in The frequency band energy of the time slice; Energy distribution matrix is represented as follows: ， in, represents the number of samples, , ,when hour, ; (2) Calculate the similarity between samples and construct a similarity matrix; The formula for calculating the similarity between any two samples is as follows: ， in, is the similarity matrix The element of Samples and The similarity of samples; is an exponential function with a natural constant as its base; is the similarity attenuation coefficient; is an indicative function. When the parameter difference of the indicative function is less than the preset blasting parameter difference threshold Take 1 when it is, otherwise take 0; is the preset blasting parameter difference threshold; For the Samples and Quantitative value of the difference in burst parameters of samples; Similarity Matrix is represented as follows: ， Among them, when hour, ; (3) Generate meta-tasks using spectral clustering algorithm: Similarity matrix Perform Laplace matrix eigendecomposition and select K-means clustering is performed on the feature vectors to assign similar samples to the same meta-task, ensuring that the samples within the task have similar geological-blasting conditions.

4. The method for predicting the vibration effect of tunnel excavation blasting on surface buildings according to claim 3, characterized in that: S3 is as follows: The neural network model is used to extract features from samples in each meta-task. The feature extraction process includes initializing the parameters of the neural network model, performing dual-path feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy attenuation curve to optimize the meta-parameters in the meta-task. (1) Initialize the parameters of the neural network model: A neural network initialization method based on vibration energy distribution is used. The samples in each meta-task are mapped into low-dimensional embedding vectors through a pre-trained autoencoder neural network. The mean of the embedding vectors in each meta-task is calculated as the prototype of each meta-task. Then, the initial weights of the network are generated by weighted fusion based on the statistical distribution of the prototype vectors of each meta-task and the historical task parameter offsets. (2) Perform feature fusion after dual-path feature extraction in the time domain and frequency domain: adopt a dual-path interaction structure to define the structure of the dual-path feature fusion neural network, including the time domain path neural network and the frequency domain path neural network; 1) Time-domain pathway neural network: Perform one-dimensional convolution on samples in the meta-task to extract time-domain features, and fuse frequency-domain features through a cross-pathway attention gating mechanism to obtain the output features of the time-domain pathway neural network; 2) Frequency-domain pathway neural network: Perform a short-time Fourier transform on the samples in the meta-task, encode the frequency-domain features using a multi-layer perceptron, and adaptively scale the frequency-band weights based on the frequency-band energy to obtain the output features of the frequency-domain pathway neural network. 3) Dual-pathway feature fusion: The output features of the time-domain neural network and the frequency-domain neural network are cross-pathway interacted, feature cross-enhancement is performed through element-by-element multiplication, and then feature complementary fusion is achieved through element-by-element addition, ultimately generating a representation vector containing the joint time-frequency features; (3) Meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task: A dynamic learning rate adjustment strategy is adopted. First, the energy peak moment of each vibration signal is extracted, and the average energy attenuation ratio in the time period after the peak is calculated as the attenuation coefficient. When the inner loop parameters of the neural network model are updated, the basic learning rate is smoothly adjusted through the hyperbolic tangent function according to the signal time point and attenuation coefficient corresponding to the current training step, so that the parameter update intensity is consistent with the energy attenuation trend. During the outer loop update, the gradients of multiple tasks are aggregated and the meta-parameters are optimized.

5. The method for predicting the vibration effect of tunnel excavation blasting on surface buildings according to claim 4, characterized in that: The parameters of the neural network model are initialized as follows: (1) Calculate the meta-task prototype vector: For all samples in each meta-task, map them into low-dimensional embedding representations through a pre-trained autoencoder neural network, and calculate the mean of the embedding representation within the task as the task prototype; (2) Generate initialization parameters: Based on the meta-task prototype vector and historical parameter offset, generate the network initial weights through weighted fusion, and then generate the network initialization weight matrix.

6. The method for predicting the vibration effect of tunnel excavation blasting on surface buildings according to claim 4, characterized in that: The meta-parameters in the meta-task are optimized by meta-learning based on the vibration energy attenuation curve as follows: (1) Constructing the energy attenuation coefficient: First, locate the energy peak point of each vibration signal, then count the relative energy per unit time in the time period after the peak, calculate its ratio to the peak energy, and then construct the overall energy attenuation coefficient; (2) Dynamically adjust the inner loop learning rate: The inner loop represents the parameter update within the meta-task, and dynamically adjusts the learning rate according to the energy decay coefficient. First, the time point corresponding to the current step is obtained, and the energy decay coefficient is input into the hyperbolic tangent function to generate a smoothed adjustment value. The smoothed adjustment value is then multiplied by the base learning rate to obtain the adjusted value of the base learning rate. The base learning rate and the base learning rate adjustment value are then added to obtain the inner loop adaptive learning rate. (3) External loop parameter update: The outer loop represents the optimization of meta-parameters. By performing inner loop training on multiple meta-tasks separately, the specific model parameters of each meta-task are obtained, the loss function of all meta-tasks on the support set is calculated, the neural network model parameters are back-propagated, and the meta-parameters are updated according to the optimization step size, thereby obtaining the meta-parameters to be updated in the outer loop.

7. The method for predicting the vibration effect of tunnel excavation blasting on surface buildings according to claim 4, characterized in that: S4 is as follows: A hierarchical attention mechanism is adopted. Based on the event response characteristics of different samples at different time scales, attention weights are extracted using a short-term convolutional neural network, a medium-term long short-term memory neural network, and a long-term global pooling perception neural network. These attention weights focus on instantaneous, event-level, and global vibration changes, respectively. The attention vectors at each scale are weighted and superimposed on the fusion features element by element to generate a multi-scale classification feature vector. The classification feature vector is then input into a preset Softmax function to obtain a predicted class probability vector. The class corresponding to the largest class probability in the class probability vector is taken as the classification prediction result of the vibration impact of blasting on surface buildings. The short-term window convolutional neural network implements the short-term window attention enhancement operation and sets , the calculation formula is as follows: ； The medium-time window long short-term memory neural network implements the medium-time window attention enhancement operation, setting , the calculation formula is as follows: ； The long-term window global pooling perception neural network implements long-term attention enhancement operation, setting , the calculation formula is as follows: ； The calculation formula for multi-scale feature aggregation is as follows: ； in, is a representation vector containing time-frequency joint features; is the time window length of the attention mechanism; is the short-term window attention weight; is the Softmax function; is the short-time window convolution kernel parameter; It is a long short-term memory neural network; Output of the long short-term memory neural network The transpose of is the attention weight of the mid-time window; is the linear transformation parameter of the medium time window; is the long-term window attention weight; is the multi-layer perceptron function; is the Sigmoid activation function; Represents element-wise multiplication; It is the global maximum pooling operation; is the classification feature vector after aggregating multi-scale attention.

8. The method for predicting the vibration effect of tunnel excavation blasting on surface buildings according to claim 7, characterized in that: The generalization verification and dynamic adjustment of meta-tasks are as follows: The meta-task is divided into a training set and a validation set. The support set is used in the inner loop to quickly fine-tune the neural network model parameters. The meta-parameters are updated based on the query set loss in the outer loop. During the training process, the performance of the neural network model is periodically evaluated on the validation meta-task. The loss calculation in the neural network model uses cross-entropy loss. The loss function is calculated once during each training. If the loss does not decrease after P consecutive cycles, the outer loop learning rate is reduced exponentially. Finally, the optimal hyperparameter combination of the neural network model is determined through multi-task cross-validation, and the classification accuracy and generalization are verified on an independent test set.

9. The method for predicting the vibration effect of tunnel excavation blasting on surface buildings according to claim 8, characterized in that: The closed-loop feedback control and parameter optimization are as follows: Combining the classification prediction results with the blasting parameter adjustment rule base to generate real-time control instructions for adjusting blasting parameters; The blasting parameter adjustment rule base includes expert experience or physical models. The calculation formula of the adjusted blasting parameters is as follows: ， in, is the adjusted blasting parameter; is the benchmark parameter; is the category sensitivity coefficient fitted by historical data; The vibration category predicted by the model; This is the category corresponding to the safety vibration threshold.

10. A method for predicting the vibration impact of tunnel excavation blasting on surface buildings according to claim 9, characterized in that: The strategy for online optimization of blasting parameters through reinforcement learning strategy is as follows: Using a reinforcement learning strategy, with classification accuracy and vibration energy control as reward functions, the blasting parameter strategy is optimized online. The state space, action space, and reward function are defined. The state space includes vibration signal characteristics and geological parameters, the action space includes blasting parameter adjustments, and the reward function is the sum of vibration energy below a threshold and classification accuracy. The proximal policy optimization method is used to train the reinforcement learning policy network, and the model parameters are updated together with the meta-learning model.

Images