Method for predicting influence of tunneling blasting on surface building vibration

Through dynamic wavelet packet decomposition, spectral clustering and improved meta-learning neural network model, the problem of insufficient extraction of vibration signal characteristics in tunnel bore blasting is solved, and efficient, accurate prediction and construction safety improvement of the impact on the vibration of surface buildings is achieved.

CN120257096AActive Publication Date: 2025-07-04SHANDONG LUQIAO GROUP CO LTD

Patent Information

Application Number
CN202510703222.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-29
Publication Date
2025-07-04
Estimated Expiration
2045-05-29

AI Technical Summary

Technical Problem

Traditional methods are difficult to effectively extract the instantaneous and long-term characteristics of vibration signals during tunnel boring and blasting, resulting in low accuracy of vibration impact prediction, and traditional neural network training slow convergence speed or easy to fall into local optimality, and the existing technology is difficult to adapt to complex geological conditions and blasting parameters changes.

Method used

Dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy are used to denoise, combined with the spectral clustering algorithm to generate meta-tasks, feature extraction and classification prediction are performed through improved meta-learning neural network model, hierarchical attention mechanism is used to capture the time response characteristics of the vibration signal, and blasting parameters are optimized through reinforcement learning.

Benefits of technology

The prediction accuracy and construction safety of the impact on the vibration of surface buildings are improved, the wavelet packet coefficient is dynamically adjusted to avoid signal loss, spectral clustering ensures the clustering of similar samples, dual-path feature fusion captures the complex characteristics of vibration signals, multi-scale attention enhances response capabilities, and strengthens learning and optimization parameter adjustment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120257096A_ABST
    Figure CN120257096A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of artificial intelligence and data processing, in particular to a method for predicting the influence of tunneling blasting on surface building vibration, which specifically comprises the following steps of: collecting a vibration signal generated by blasting operation in a tunneling process and denoising the vibration signal as a sample; according to the energy distribution characteristics of the signals, a spectral clustering algorithm is adopted to generate meta-tasks; performing feature extraction on samples in each meta-task through a neural network model; a hierarchical attention mechanism is adopted to capture time response characteristics of the vibration signals in different time durations, key characteristics of all stages are obtained and fused, a classification feature vector is obtained, and then a classification prediction result of the influence of blasting on surface building vibration is determined according to the classification feature vector; and optimizing the classification prediction process. According to the method, the tunneling blasting vibration data is processed based on the neural network model of the improved meta-learning strategy, and the accuracy of the vibration influence on the surface building and the construction safety can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and data processing, and particularly to a prediction method for the vibration impact of tunnel excavation blasting on surface buildings. Background Art

[0002] During the process of tunnel excavation, blasting vibrations are inevitable, and these vibration signals not only reflect information such as geological conditions and blasting parameters during the blasting process, but also contain complex noise and mixed signals of different frequency bands. Traditional vibration monitoring techniques often rely on fixed threshold methods or simple Fourier transforms, but these methods have many limitations in dealing with non-stationary, strong noise interference, and instantaneous frequency domain characteristics. Especially in an environment with complex geological conditions and changing blasting parameters, it is difficult for traditional methods to effectively extract useful features from vibration signals, resulting in low classification prediction accuracy. In addition, traditional machine learning methods, such as meta-learning, usually adopt a random way in task division, resulting in samples of the same vibration pattern being possibly divided into different tasks, and unable to effectively capture the correlation between geological conditions and blasting parameters. At the same time, traditional neural network methods often adopt random initialization, which may lead to a slow convergence speed or being easily trapped in a local optimum during the training process. Feature extraction methods are also mostly limited to one-dimensional extraction in the time domain or frequency domain, and fail to comprehensively capture the complex characteristics of vibration signals. The time series characteristics, energy attenuation characteristics, etc. of vibration signals are not fully utilized, making the model unable to perform efficient and accurate predictions in practical applications.

[0003] The existing technology has the following objective drawbacks: Traditional Fourier transforms 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 result in samples of similar vibration patterns being assigned to different tasks and unable to effectively capture the correlation between geological conditions and blasting parameters; Traditional neural network initialization methods use random initialization, which easily leads to a slow convergence speed of network training or even non-convergence; Traditional meta-learning methods usually adopt a fixed learning rate and do not consider the energy attenuation characteristics of vibration signals, resulting in being unable to effectively adapt to the time series changes of signals during the training process; Traditional feature extraction methods may only extract certain aspects of signal features, such as the time domain or frequency domain, and cannot comprehensively capture the instantaneous and long-term characteristics of vibration signals; Traditional blasting parameter adjustment usually relies on artificial experience or simple rules, with low efficiency.

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

[0005] In view of the deficiencies of the prior art, the present invention develops a prediction method for the vibration impact of tunnel excavation blasting on surface buildings. The present invention processes the vibration data of tunnel excavation blasting 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 for the present invention to solve the technical problem is a prediction method for the vibration impact of tunnel excavation blasting on surface buildings, including the following steps: S1. Collect the vibration signals generated by blasting operations during tunnel excavation through vibration sensors, and then perform denoising operations on the collected vibration signals, and use the denoised vibration signals as samples; S2. Divide the denoised vibration signals into tasks. According to the energy distribution characteristics of the vibration signals, use the spectral clustering algorithm to generate meta-tasks, and each meta-task is a sample set; S3. Extract features from the samples in each meta-task through a neural network model. The feature extraction process includes initializing the parameters of the neural network model, performing dual-channel feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task; S4. Adopt a hierarchical attention mechanism to capture the time response characteristics of vibration signals at different time lengths, obtain the key features of each stage, then fuse the key features of each stage to obtain a classification feature vector, and further determine the classification prediction result of the vibration impact of blasting on surface buildings according to the classification feature vector; S5. Optimize the classification prediction process, including generalizing and validating the meta-tasks and dynamically adjusting them, optimizing the closed-loop feedback control and parameters, and online optimizing the blasting parameter strategy through a reinforcement learning strategy.

[0007] The denoising operation adopts a dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy, and the denoising operation is specifically as follows: The collected vibration signals are non-stationary, multi-band mixed, and strongly noise-interfered signals. The collected vibration signals are denoised by using a dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy; Specifically, first perform dynamic wavelet packet decomposition on the collected vibration signals to obtain wavelet packet coefficients of each layer, then calculate the local band energy of each layer of coefficients through a sliding time window, and extract the noise base energy of the non-blasting period at the front end of the collected vibration signals. Finally, dynamically weight the wavelet packet coefficients of each layer according to the ratio of the local band energy to the noise base, and reconstruct the denoised signal.

[0008] In the task division of the denoised vibration signal, first divide each sample frequency band energy sequence according to time segments, then calculate the energy difference between the time segments of any two samples, and then construct a similarity matrix of the samples by combining the blasting parameter differences. Finally, perform eigen-decomposition and clustering on the similarity matrix through the spectral clustering algorithm to divide the samples into different meta-tasks. Each meta-task is a sample set, and the samples in it have similar geological conditions and blasting parameters. The specific operation steps are as follows: (1)Calculate the difference metric of the frequency band energy distribution of the samples and construct an energy distribution matrix; The calculation formula for the difference metric of the frequency band energy distribution between any two samples is as follows: , where, is the element of the energy distribution matrix and represents the difference metric of the frequency band energy distribution between the -th sample and the -th sample; is the time segment; is the frequency band energy of the -th sample in the -th time slice; is the frequency band energy of the -th sample in the -th time slice; The energy distribution matrix is represented as follows: , where, represents the number of samples, , when , ; (2)Calculate the similarity between samples and construct a similarity matrix; The calculation formula for the similarity between any two samples is as follows: , where, is the element of the similarity matrix and represents the similarity between the -th sample and the -th sample; is the exponential function with the natural constant as the base; is the similarity decay coefficient; is the indicator function, which takes 1 when the parameter difference of the indicator function is less than the preset blasting parameter difference threshold and takes 0 otherwise; is the preset blasting parameter difference threshold; is the The quantification value of the difference in blasting parameters between the th sample and the similarity matrix is represented as follows: , where, when , ; (3) Generate meta-tasks using the spectral clustering algorithm: Perform Laplacian matrix eigen-decomposition on the similarity matrix , select the first eigenvectors for K-means clustering, and assign samples of the same class to the same meta-task to ensure that the samples within the task have similar geological-blasting conditions.

[0009] S3 is specifically as follows: Extract features from the samples in each meta-task through a neural network model. The feature extraction process includes initializing the parameters of the neural network model, performing dual-channel feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task; (1) Initialize the parameters of the neural network model: Adopt a neural network initialization method based on vibration energy distribution. Map the samples in each meta-task to low-dimensional embedding vectors through a pre-trained autoencoder neural network, calculate the mean of the embedding vectors within each meta-task as the prototype of each meta-task, and then, according to the statistical distribution of the prototype vectors of each meta-task and the historical task parameter offset, generate the initial network weights through weighted fusion; (2) Perform dual-channel feature extraction in the time domain and frequency domain and then perform feature fusion: Adopt a dual-channel interaction structure, and define the structure of the dual-channel feature fusion neural network, including a time-domain path neural network and a frequency-domain path neural network; 1) Time-domain path neural network: Perform one-dimensional convolution on the samples in the meta-task to extract time-domain features, and fuse the frequency-domain features through a cross-path attention gating mechanism to obtain the output features of the time-domain path neural network; 2) Frequency-domain path neural network: Perform short-time Fourier transform on the samples in the meta-task, encode the frequency-domain features through a multi-layer perceptron, and adaptively scale the band weights based on the band energy to obtain the output features of the frequency-domain path neural network; 3) Dual-channel feature fusion: Perform cross-path interaction on the output features of the time-domain path neural network and the output features of the frequency-domain path neural network, perform feature cross-enhancement through element-wise multiplication, and then achieve feature complementary fusion through element-wise addition to finally generate a representation vector containing time-frequency joint features; (3) Optimize the meta-parameters in the meta-task based on the vibration energy decay curve through meta-learning: Adopt a dynamic learning rate adjustment strategy. First, extract the energy peak moments of each segment of the vibration signal, calculate the average energy decay ratio within the time period after the peak as the decay coefficient. When updating the loop parameters in the neural network model, according to the signal time point corresponding to the current training step and the decay coefficient, smoothly adjust the base learning rate through the hyperbolic tangent function, so that the parameter update intensity is consistent with the energy decay trend. When updating the outer loop, aggregate the gradients of multiple tasks and optimize the meta-parameters.

[0010] Initialize the parameters of the neural network model as follows: (1) Calculate the meta-task prototype vector: For all samples in each meta-task, map them to a low-dimensional embedded representation through a pre-trained autoencoder neural network, and calculate the mean of the in-task embedded representations as the task prototype; (2) Generate the initial parameters: According to the meta-task prototype vector and the historical parameter offset, generate the initial network weights through weighted fusion, and then generate the network initial weight matrix.

[0011] Optimize the meta-parameters in the meta-task based on the vibration energy decay curve through meta-learning as follows: (1) Construct the energy decay coefficient: First, locate the energy peak points of each segment of the vibration signal, then within the time period after the peak, count the relative energy per unit time, calculate its ratio to the peak energy, and then construct the overall energy decay coefficient; (2) Dynamically adjust the inner-loop learning rate: The inner loop represents the parameter update within the meta-task. Dynamically adjust the learning rate according to the energy decay coefficient. First, obtain the time point corresponding to the current step, input the energy decay coefficient into the hyperbolic tangent function to generate a smooth adjustment value, then multiply the smooth adjustment value by the base learning rate to obtain the adjustment value of the base learning rate, and then add the base learning rate and the base learning rate adjustment value to obtain the inner-loop adaptive learning rate; (3) Outer-loop parameter update: The outer loop represents the optimization of the meta-parameters. After performing inner-loop training on multiple meta-tasks respectively, obtaining the specific model parameters of each meta-task, calculate the loss function of all meta-tasks on the support set, perform backpropagation on the neural network model parameters, and update the meta-parameters according to the optimization step size, so as to obtain the meta-parameters to be updated in the outer loop.

[0012] S4 is as follows: Adopt a hierarchical attention mechanism. According to the event response characteristics of different samples at different time scales, 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 are respectively used to extract attention weights, focusing on instantaneous, event-level, and global vibration changes. The attention vectors at each scale are respectively weighted element-wise with the fusion features and then superimposed to generate a multi-scale classification feature vector. Then, the classification feature vector is input into a preset Softmax function to obtain a predicted class probability vector, and the class corresponding to the maximum class probability in the class probability vector is taken as the classification prediction result of the vibration impact of the blasting on surface buildings; The short-time window convolutional neural network performs a short-time window attention enhancement operation, setting , and the calculation formula is as follows: ; The medium-time window long short-term memory neural network performs a medium-time window attention enhancement operation, setting , and the calculation formula is as follows: ; The long-time window global pooling perception neural network performs a long-time window attention enhancement operation, setting , and the calculation formula is as follows: ; The calculation formula for multi-scale feature aggregation is as follows: ; Among them, is a representation vector containing time-frequency joint features; is the time window length of the attention mechanism; is the short-time window attention weight; is the Softmax function; is the short-time window convolution kernel parameter; is the long short-term memory neural network; is the output of the long short-term memory neural network transpose; is the medium-time window attention weight; is the medium-time window linear transformation parameter; is the long-time window attention weight; is the multi-layer perceptron function; is the Sigmoid activation function; represents element-wise multiplication; is the global maximum pooling operation; is the classification feature vector after aggregating multi-scale attention.

[0013] The generalization verification and dynamic adjustment of the meta-task are as follows: Divide the meta-task into a training set and a validation set, and use the support set to quickly fine-tune the neural network model parameters in the inner loop, and update the meta-parameters based on the query set loss in the outer loop; During the training process, periodically evaluate the performance of the neural network model on the validation meta-task. The loss calculation in the neural network model uses cross-entropy loss, and the loss function is calculated once in each training. If the loss does not decrease for consecutive P periods, the learning rate of the outer loop is decreased exponentially; Finally, determine the optimal hyperparameter combination of the neural network model through multi-task cross-validation, and verify the classification accuracy and generalization on an independent test set.

[0014] The optimization of the closed-loop feedback control and parameters is as follows: Combine the classification prediction results with the blasting parameter adjustment rule base to generate real-time control instructions for adjusting the blasting parameters; The blasting parameter adjustment rule base includes expert experience or physical models. The calculation formula for the adjusted blasting parameters is as follows: , where is the adjusted blasting parameter; is the reference parameter; is the class-sensitive coefficient fitted through historical data; is the vibration class predicted by the model; is the class corresponding to the safe vibration threshold.

[0015] The online optimization of the blasting parameter strategy through the reinforcement learning strategy is as follows: Adopt the reinforcement learning strategy, with the classification accuracy and vibration energy control as the reward function, and online optimize the blasting parameter strategy. Define the state space, action space, and reward function. The state space includes vibration signal features and geological parameters, the action space includes the adjustment amount of blasting parameters, and the reward function adopts the sum of the vibration energy below the threshold and the classification accuracy; And use the proximal policy optimization method to train the reinforcement learning strategy network, and jointly update the model parameters with the meta-learning model.

[0016] The effects provided in the invention content are only the effects of the embodiments, rather than all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: The present invention can accurately separate noise and effective vibration signals through dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy, and can dynamically adjust the weights of wavelet packet coefficients, avoiding the loss of effective signals that may be caused by the fixed threshold wavelet denoising method; The meta-task division method based on spectral clustering proposed by the present invention generates meta-tasks through the energy distribution characteristics of vibration signals and the differences in blasting parameters, which can ensure that samples with the same vibration mode 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 the model training efficiency; The present invention uses a pre-trained autoencoder neural network to map the denoised vibration signal into a low-dimensional embedding vector, and generates the initial weight of the network by calculating the mean of the task prototype and the historical parameter offset for weighted fusion, which can accelerate the convergence of the network and reduce the risk of overfitting; The present invention extracts the time-domain features and frequency-domain features of the vibration signal through a dual-path neural network structure, and then fuses the two features through a cross-path attention mechanism, which can improve the signal representation ability and can more comprehensively capture the instantaneous characteristics and long-term resonance characteristics of the vibration signal; 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 and automatically adjusting the learning rate, it can adapt to the temporal energy distribution of the signal, significantly improve the learning efficiency, and prevent over-updating or slow convergence; The present invention adopts a multi-scale attention mechanism. By 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 neural network to extract the key features of the vibration signal at different time scales, it can enhance the response ability to different time scales; The present invention proposes a real-time control instruction generation mechanism that combines the classification prediction results with the blasting parameter adjustment rule base. And in the optional strategy, the blasting parameters are further optimized using reinforcement learning, which can realize intelligent parameter adjustment.

[0017] In summary, the present invention can improve the accuracy of the impact on surface buildings and the safety of construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention.

[0019] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0020] Figure 2 It is a comparison diagram of the effects of different denoising methods.

[0021] Figure 3 It is the collected vibration signal.

[0022] Figure 4 It is a comparison diagram of the denoised vibration signal and the real signal.

[0023] Figure 5 Comparison chart of the meta - learning effects of different task - division strategies.

[0024] Figure 6 Comparison chart of the learning - rate adjustment strategies. Detailed implementation manners

[0025] In order to clearly illustrate the technical features of this solution, the present invention will be elaborated in detail below through specific implementation manners and in combination with its accompanying drawings.

[0026] Embodiment 1 As Figure 1 shown, a prediction method for the vibration influence of tunnel - boring blasting on surface buildings includes the following steps: S1. Collect vibration signals generated by blasting operations during tunnel boring through vibration sensors, and then perform denoising operations on the collected vibration signals. The denoised vibration signals are used as samples. S2. Perform task division on the denoised vibration signals. According to the energy - distribution characteristics of the vibration signals, use the spectral clustering algorithm to generate meta - tasks, and each meta - task is a sample set. S3. Extract features from the samples in each meta - task through 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 and then fusing the extracted features, and performing meta - learning based on the vibration - energy decay curve to optimize the meta - parameters in the meta - tasks. S4. Adopt a hierarchical attention mechanism to capture the time - response characteristics of vibration signals at different time lengths, obtain the key features at each stage, then fuse the key features at each stage to obtain a classification feature vector, and further determine the classification prediction result of the vibration influence of blasting on surface buildings according to the classification feature vector. S5. Optimize the classification - prediction process, including performing generalization verification and dynamic adjustment on the meta - tasks, optimizing the closed - loop feedback control and parameters, and online - optimizing the blasting - parameter strategy through a reinforcement - learning strategy.

[0027] In the specific implementation manner, the denoising operation adopts a dynamic wavelet - packet decomposition and local signal - to - noise - ratio weighting strategy, and the denoising operation is as follows: The collected vibration signals are non - stationary, multi - band - mixed, and strongly noise - interfered signals. The dynamic wavelet - packet decomposition and local signal - to - noise - ratio weighting strategy are used to perform denoising operations on the collected vibration signals. Specifically, first perform dynamic wavelet - packet decomposition on the collected vibration signals to obtain wavelet - packet coefficients at each layer. Then calculate the local - band energy of each layer of coefficients through a sliding time window, and extract the noise - floor energy of the non - blasting period at the front end of the collected vibration signals. Finally, dynamically weight the wavelet - packet coefficients at each layer according to the ratio of the local - band energy to the noise - floor energy, and reconstruct the denoised signal.

[0028] The calculation formula for the denoising operation is as follows: , where is the vibration signal after denoising; is the wavelet packet decomposition coefficient of the th layer, which characterizes the detailed information of the vibration signal at different scales; is the total number of layers of wavelet packet decomposition; is the local band energy of the th layer of wavelet packet at time . By calculating the energy difference, the main components of the vibration signal are determined, 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 initiation; is the noise floor energy corresponding to the th layer, which is used to separate the effective vibration characteristics from the vibration signal. For example, the average energy of the first 0.1 seconds of the vibration signal without blasting can be taken to characterize the inherent noise of the sensor and the background vibration; is the reconstruction component of the wavelet packet decomposition coefficient of the th layer at time ; represents the time index of the vibration signal; is the wavelet packet basis function of the th layer, which represents the decomposition and reconstruction of the vibration signal, characterizes the wavelet packet component after time window adjustment; is the noise floor energy, which is the average energy of the non-blasting section in the first 0.1 s of the signal; is the anti-zero constant, set to ; is the center point of the time window, and the calculation formula is: ; is the time window width, and the calculation formula is: ; is the local band energy, and the calculation formula is: ; where is the time window length for local energy calculation; represents the sliding time index within the time window; is the value of the wavelet packet decomposition coefficient of the th layer at time within the sliding time window; It should be noted that the center point and width of the time window are adaptively determined based on the energy distribution of wavelet packet coefficients, which can accurately locate the time-frequency locality of blasting events. For example, for multi-stage vibrations caused by delayed explosions, compared with fixed-time-window wavelet decomposition, dynamically adjusting the time window can capture the frequency band characteristics of different rock fracture stages.

[0029] In the specific implementation, when dividing the denoised vibration signals into tasks, first divide the frequency band energy sequence of each sample according to time segments, then calculate the energy difference between the time segments of any two samples, and then construct a similarity matrix of the samples by combining the differences in blasting parameters. Finally, perform eigen-decomposition and clustering on the similarity matrix through the spectral clustering algorithm to divide the samples into different meta-tasks. Each meta-task is a sample set, and the samples in it have similar geological conditions and blasting parameters. The specific operation steps are as follows: (1)Calculate the difference metric of the frequency band energy distribution of the samples and construct an energy distribution matrix; The calculation formula for the difference metric of the frequency band energy distribution between any two samples is as follows: , where, is the element of the energy distribution matrix indicating the difference metric of the frequency band energy distribution between the th sample and the th sample, which characterizes the similarity of the two samples in energy distribution and quantifies the impact of geological condition differences on the vibration mode, such as the difference in energy release rates between hard rock and soft rock; is the number of time segments, which characterizes the granularity of the time period division; is the frequency band energy of the th sample in the th time slice; is the frequency band energy of the th sample in the th time slice; The energy distribution matrix is represented as follows: , where, represents the number of samples, , when , ; It should be noted that because different geological conditions and blasting parameters will result in different vibration modes, traditional random task division methods cannot effectively capture these differences. By calculating the frequency band energy differences of vibration signals in sequence, it is possible to better divide tasks according to the differences in geological conditions and blasting parameters, ensure that similar samples are clustered together, and help meta-learning to quickly adapt under specific conditions; (2) Calculate the similarity between samples and construct a similarity matrix; The calculation formula for the similarity between any two samples is as follows: , where, is an element of the similarity matrix , representing the similarity between the -th sample and the -th sample; is the exponential function with the natural constant as the base; is the similarity attenuation coefficient, set ; is the indicator function, which takes 1 when the difference between the parameters of the indicator function is less than the preset blasting parameter difference threshold , and takes 0 otherwise; is the preset blasting parameter difference threshold, set ; is the quantification value of the blasting parameter difference between the -th sample and the -th sample, which characterizes the difference in blasting parameters between the two samples and forces the samples within the same meta-task to have similar blasting conditions. For example, the difference in blasting parameters such as the amount of explosive and the rock mass grade; The similarity matrix is represented as follows: , where, when , ; (3) Use the spectral clustering algorithm to generate meta-tasks: Perform Laplacian matrix eigenvalue decomposition on the similarity matrix , select the first eigenvectors for K-means clustering, and assign samples of the same class to the same meta-task to ensure that the samples within the task have similar geological-blasting conditions.

[0030] It should be noted that aggregating samples with the same geological-blasting conditions into the same meta-task can ensure the consistency of the data within the task, thereby improving the ability of the model to quickly adapt to different tasks during the meta-learning process and providing a structured task division basis for subsequent parameter initialization.

[0031] In the specific implementation manner, S3 is specifically as follows: Extract features from the samples in each meta-task through a neural network model. The feature extraction process includes initializing the parameters of the neural network model, performing dual-channel feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task; (1) Initialize the parameters of the neural network model: Adopt a neural network initialization method based on vibration energy distribution. Map the samples in each meta-task to low-dimensional embedding vectors through a pre-trained autoencoder neural network, calculate the mean of the embedding vectors within each meta-task as the prototype of each meta-task, and then, based on the statistical distribution of the prototype vectors of each meta-task and the historical task parameter offsets, generate the initial network weights through weighted fusion. (2) After performing dual-channel feature extraction in the time domain and frequency domain, perform feature fusion: Adopt a dual-channel interaction structure, and define the structure of the dual-channel feature fusion neural network, including a time-domain channel neural network and a frequency-domain channel neural network. 1) Time-domain channel neural network: Perform one-dimensional convolution on the samples in the meta-task to extract time-domain features, and fuse the frequency-domain features through a cross-channel attention gating mechanism to obtain the output features of the time-domain channel neural network. The calculation formula of the time-domain channel neural network is as follows: , where, is the output of the time-domain channel neural network; represents the one-dimensional convolution operation; are the convolution kernel parameters; represents element-wise addition; is the feature generated by linear transformation and Sigmoid activation after concatenating the time-domain and frequency-domain features, and its calculation formula is as follows: , where, is the Sigmoid activation function; is the frequency-domain feature vector encoded by the multi-layer perceptron of the output of the frequency-domain channel neural network; is the gating weight matrix, which controls the fusion ratio of the time-domain and frequency-domain features; represents vector concatenation; represents element-wise multiplication; 2) Frequency-domain channel neural network: Perform short-time Fourier transform on the samples in the meta-task, encode the frequency-domain features through a multi-layer perceptron, and adaptively scale the band weights based on the band energy to obtain the output features of the frequency-domain channel neural network. The calculation formula of the frequency-domain channel neural network is as follows: , where, is the output of the frequency-domain channel neural network; is the multi-layer perceptron function; is the operation of taking the real part; is the short-time Fourier transform; Indicates the frequency band scaling factor, and the calculation formula is as follows: , where is the adjustment coefficient of the frequency band scaling intensity; is the energy of the frequency band; is the average energy of the frequency band in the historical data; is the frequency band index; 3) Dual-channel feature fusion: Cross-channel interaction is performed on the output features of the time-domain channel neural network and the output features of the frequency-domain channel neural network. Feature cross-enhancement is performed through element-wise multiplication, and then feature complementary fusion is achieved through element-wise addition, finally generating a representation vector containing time-frequency joint features.

[0032] The calculation formula for dual-channel feature fusion is as follows: , where is the fused feature vector.

[0033] It should be noted that in the classification task of tunnel boring blast vibration data, the combination of time-domain and frequency-domain features can comprehensively reflect the change pattern of blast vibration. Dual-channel feature fusion can capture both time-domain and frequency-domain features, thereby extracting the instantaneous characteristics and long-term resonance characteristics of the signal, and improving the representation ability of the vibration signal.

[0034] (3) Optimize the meta-parameters in the meta-task through meta-learning based on the vibration energy decay curve: Adopt a dynamic learning rate adjustment strategy. First, extract the energy peak moment of each segment of vibration signal, calculate the average energy decay ratio within the time period after the peak as the decay coefficient. When updating the internal loop parameters of the neural network model, according to the signal time point corresponding to the current training step and the decay coefficient, smoothly adjust the base learning rate through the hyperbolic tangent function, so that the parameter update intensity is consistent with the energy decay trend. When updating the external loop, aggregate the gradients of multiple tasks and optimize the meta-parameters.

[0035] In the specific implementation manner, the parameters of the neural network model are initialized as follows: (1) Calculate the meta-task prototype vector: Map all samples in each meta-task to a low-dimensional embedded representation through a pre-trained autoencoder neural network, and calculate the mean of the in-task embedded representations as the task prototype; The calculation formula for the meta-task prototype vector is as follows: , where is the th meta-task sample set; Is the embedded representation of the pre-trained auto-encoding neural network; Is the denoised signal of the th sample; Is the prototype vector of the th meta-task; (2) Generate initialization parameters: According to the meta-task prototype vector and the historical parameter offset, generate the initial network weights through weighted fusion, and then generate the network initialization weight matrix. Compared with the single meta-parameter of MAML, weighted fusion can inherit the common knowledge across tasks; The calculation formula of the network initialization weight matrix is as follows: , where is the number of samples of task ; is the network initialization weight matrix; is the total number of meta-tasks; is the total number of all samples; is the rd prototype vector of the th meta-task; represents the Kronecker product; is the task-related orthogonal matrix; is the fusion coefficient, taking the statistical mean of the historical task parameter offsets; is the historical task parameter offset; , where is the number of historical tasks; is a positive integer; is the th final training parameter of the For the initial parameters of the nth historical task.

[0036] In the specific implementation manner, meta-learning is performed based on the vibration energy decay curve to optimize the meta-parameters in the meta-task as follows: (1) Construct the energy decay coefficient: First, locate the energy peak point of each segment of the vibration signal. Then, within the time period after the peak, count the relative energy per unit time, calculate its ratio to the peak energy, and thereby construct the overall energy decay coefficient, which can measure the temporal energy dissipation speed of the vibration signal; The calculation formula of the energy decay coefficient is as follows: , where is the energy decay coefficient, is the total signal duration; is the vibration energy at time . The vibration energy refers to the energy intensity of the vibration signal within a certain period of time, and the root mean square value is used to represent the vibration energy; is the vibration energy at the energy peak time; is the vibration energy peak time; (2) Dynamically adjust the inner loop learning rate: The inner loop represents the parameter update within the meta-task. The learning rate is dynamically adjusted according to the energy decay coefficient. First, obtain the time point corresponding to the current step, input the energy decay coefficient into the hyperbolic tangent function to generate a smooth adjustment value, then multiply the smooth adjustment value by the base learning rate to obtain the adjustment value for the base learning rate, and then add the base learning rate and the base learning rate adjustment value to obtain the inner loop adaptive learning rate. After the parameters in the inner loop are updated, they can better adapt to the temporal distribution of the current signal energy; The calculation formula of the inner loop adaptive learning rate is as follows: , where is the inner loop adaptive learning rate; is the base learning rate, set ; is the decay adjustment coefficient, set ; is the number of training steps; is the hyperbolic tangent function; represents the smooth adjustment value; The inner loop adaptive learning rate can be adjusted according to the energy decay characteristics of the vibration signal. By adaptively adjusting the learning rate, the model can be optimized according to the energy decay characteristics of the vibration energy in different time periods, thereby improving the learning efficiency and avoiding over-updating or slow convergence; (3) Outer loop parameter update: The outer loop represents the optimization of meta-parameters. By separately performing inner loop training on multiple meta-tasks, after obtaining the specific model parameters of each meta-task, calculate the loss function of all meta-tasks on the support set, perform backpropagation on the neural network model parameters, and update the meta-parameters according to the optimization step size. Then, the meta-parameters to be updated in the outer loop can be obtained, which can enable the neural network model to have a faster adaptation ability on new original tasks; The calculation formula for updating the outer loop parameters is as follows: , where, is the parameter update operation; is the meta-parameter to be updated in the outer loop; is the outer loop learning rate, set ; represents the gradient of the parameter; is a positive integer; is the th loss function on the task; is the task-specific model after inner loop update; The meta-parameters are the parameters optimized by the model in the outer loop, which determine the initial performance of the model on new tasks. Through multi-task training, the meta-parameters enable the model to have prior knowledge for quickly adapting to new tasks; The task-specific model after inner loop update is the task-specific model after fine-tuning on the support set of a single task in the inner loop based on the initialization of the neural network model with meta-parameters. For example, if the support set of a certain task has 10 samples, the inner loop adjusts the parameters through 3 steps of gradient descent to obtain a task-specific model suitable for this task , and its loss is used to update the meta-parameters.

[0037] In the specific implementation manner, S4 is as follows: By adopting a hierarchical attention mechanism, it can more comprehensively capture the key features at each stage of the building response process. According to the event response characteristics of different samples at different time scales, 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 are respectively used to extract attention weights, focusing on instantaneous, event-level, and global vibration changes respectively. The attention vectors at each scale are weighted element-wise with the fusion features and then superimposed to generate a multi-scale classification feature vector. Then, the classification feature vector is input into a preset Softmax function to obtain a predicted class probability vector, and the class corresponding to the maximum class probability in the class probability vector is taken as the classification prediction result of the impact of blasting on the vibration of surface buildings; The short-time window convolutional neural network implements short-time window attention enhancement operation, set , and the calculation formula is as follows: ; The medium-time-window long short-term memory neural network implements medium-time-window attention enhancement operation, and sets , and the calculation formula is as follows: ; The long-time-window global pooling perception neural network implements long-term attention enhancement operation, and sets , and the calculation formula is as follows: ; The calculation formula of multi-scale feature aggregation is as follows: ; Among them, is the time window length of the attention mechanism; is the short-time-window attention weight; is the Softmax function; is the short-time-window convolution kernel parameter; is the long short-term memory neural network; is the output of the long short-term memory neural network transpose; is the medium-time-window attention weight; is the medium-time-window linear transformation parameter; is the long-time-window attention weight; is the Sigmoid activation function; is the global maximum pooling operation; is the classification feature vector after aggregating multi-scale attention.

[0038] The medium-time-window linear transformation parameter is 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 medium-time window; 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.

[0039] In the specific implementation manner, the generalization verification and dynamic adjustment of the meta-task are as follows: The meta-task is divided into a training set and a validation set. In the inner loop, the neural network model parameters are quickly fine-tuned using the support set, and in the outer loop, the meta-parameters are updated based on the query set loss; 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, and the loss function is calculated once in each training. If the loss does not decrease for P consecutive periods, the outer loop learning rate is decreased 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.

[0040] In the specific implementation manner, the closed-loop feedback control and parameters are optimized as follows: Combine the classification prediction results with the blasting parameter adjustment rule base to generate real-time control instructions for adjusting the blasting parameters; The blasting parameter adjustment rule base includes expert experience or physical models. The calculation formula for the adjusted blasting parameters is as follows: , where, is the adjusted blasting parameter; is the reference parameter; is the class-sensitive coefficient fitted through historical data; is the vibration class predicted by the model; is the class corresponding to the safe vibration threshold.

[0041] In the specific implementation manner, the blasting parameter strategy is optimized online through a reinforcement learning strategy as follows: Adopt a reinforcement learning strategy, with the classification accuracy and vibration energy control as the reward functions, to optimize the blasting parameter strategy online. Define the state space, action space, and reward function. The state space includes vibration signal features and geological parameters, the action space includes the adjustment amount of blasting parameters, and the reward function uses the sum of the vibration energy below the threshold and the classification accuracy; And use the proximal policy optimization method to train the reinforcement learning policy network and jointly update the model parameters with the meta-learning model.

[0042] Example 2 As Figures 2 to 4 shown, verify the performance of the denoising method in the present invention; as Figure 2 shown, it is a comparison chart of the effects of different denoising methods. Compare the Fourier filtering denoising method and the fixed prefabricated wavelet denoising method with the denoising method of the present invention. Figure 2 The signal-to-noise ratio indicators of different methods are shown by a bar chart in , and the line chart reflects the accuracy of the signal in the classification task after denoising. By comparing with the 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 in the present invention is significantly superior to the other two methods in terms of the signal-to-noise ratio index, and its bar chart is significantly higher; at the same time, the classification accuracy curve corresponding to the method of the present invention in the line chart is always at the highest position, indicating that the strategy of dynamically adjusting the weights of the wavelet packet coefficient layers can not only effectively separate noise but also avoid the loss of effective signals caused by fixed thresholds in traditional methods.

[0043] As Figure 3 and Figure 4 shown, Figure 3 is the collected vibration signal, Figure 4 is the comparison between the vibration signal after denoising and the true signal. The vibration signal in Figure 3 exhibits typical strong noise interference characteristics. There is a continuous high-frequency noise base in the front section of the signal (0 - 0.3 seconds), resulting in the waveform details of the effective vibration component (0.5 - 0.8 seconds) being submerged, and it is difficult to distinguish the pulse vibration (near 1.2 seconds) from the random noise; from Figure 4 it can be seen from the comparison between the vibration signal after denoising (blue curve) and the true signal (red dashed line) in Figure 4 that the dynamic wavelet packet decomposition accurately separates the noise and the effective components: the noise base area is completely suppressed, and the front section of the waveform tends to be flat; the vibration details in the high-frequency resonance section (such as the waveform peak at 0.6 seconds) are completely retained, highly consistent with the true 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 characteristics are clear. While retaining the physical vibration law, the denoised waveform eliminates the common problems of over-smoothing or detail loss in traditional methods. The dynamic wavelet packet decomposition of the present invention solves the defect of insufficient sensitivity of fixed-threshold wavelet denoising to transient characteristics by adaptively adjusting the time window width and the inter-layer weight,

[0044] Example 3 As Figure 5 shown, 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 meta-learning efficiency. Through three training curves of different colors, the changes in the validation set accuracy of the spectral clustering partition, random partition, and traditional clustering partition strategies during the training cycle are compared. From Figure 5 it can be seen that the curve of the spectral clustering partition converges the fastest and reaches a stable high value in the middle of the training, while the curve of the random partition fluctuates greatly and converges slowly, Figure 5 and the experimental results in show that since the spectral clustering constructs a similarity matrix by combining the signal energy distribution and the difference in blasting parameters, the smooth rising trend of the curve indicates that the samples within the task have high consistency, enabling the model to quickly capture the correlation law between geological conditions and blasting parameters, and avoiding the oscillation phenomenon caused by unreasonable task partitioning in traditional methods.

[0045] Example 4 As Figure 6 shown, by combining scatter plots with kernel density estimation, the effects of a fixed learning rate and a dynamic adjustment strategy based on energy decay on the training process were compared. Figure 6 In the figure, 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 trend of aggregation in the upper right, and the kernel density coloring area is more concentrated, indicating that it can maintain stable improvement at 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. In the present invention, the energy decay coefficient is incorporated into the learning rate adjustment through the hyperbolic tangent function, Figure 6 which is indicated by a gradually changing color arrow in the figure, so that the parameter update intensity matches the physical characteristics of the signal, thereby achieving a more robust training process under complex vibration patterns.

[0046] Although the specific implementation manners of the invention have been described above in conjunction with the accompanying drawings, they are not limitations on the protection scope of the invention. Based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A prediction method for the vibration impact of tunnel excavation blasting on surface buildings, characterized in that It includes the following steps: S1. Collect the vibration signals generated by blasting operations during tunnel boring through vibration sensors, and then perform denoising operations on the collected vibration signals. The denoised vibration signals are used as samples; S2. Divide the tasks of the denoised vibration signals. According to the energy distribution characteristics of the vibration signals, use the spectral clustering algorithm to generate meta-tasks, and each meta-task is a sample set; S3. Extract features from the samples in each meta-task through a neural network model. The feature extraction process includes initializing the parameters of the neural network model, performing dual-channel feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task; S4. Adopt a hierarchical attention mechanism to capture the time response characteristics of the vibration signals at different time lengths, obtain the key features of each stage, and then fuse the key features of each stage to obtain a classification feature vector. Furthermore, determine the classification prediction result of the vibration impact of blasting on surface buildings according to the classification feature vector; S5. Optimize the classification prediction process, including generalizing and validating the meta-tasks and dynamically adjusting them, optimizing the closed-loop feedback control and parameters, and online optimizing the blasting parameter strategy through reinforcement learning strategies; 2. The prediction method for the vibration influence of tunnel excavation blasting on surface buildings according to claim 1, characterized in that, The denoising operation adopts the dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy. The specific denoising operation is as follows: The collected vibration signals are non-stationary, multi-band mixed, and strongly noise-interfered signals. The dynamic wavelet packet decomposition and local signal-to-noise ratio weighting strategy are used to perform denoising operations on the collected vibration signals; Specifically, first perform dynamic wavelet packet decomposition on the collected vibration signals to obtain wavelet packet coefficients of each layer. Then calculate the local band energy of each layer of coefficients through a sliding time window, and extract the noise base energy of the non-blasting period at the front end of the collected vibration signals. Finally, dynamically weight the wavelet packet coefficients of each layer according to the ratio of the local band energy to the noise base, and reconstruct the denoised signal; 3. The prediction method for the vibration influence of tunnel excavation blasting on surface buildings according to claim 2, characterized in that, When dividing the tasks of the denoised vibration signals, first divide each sample frequency band energy sequence according to time segments, then calculate the energy difference between any two sample time segments, and then construct a similarity matrix of the samples in combination with the blasting parameter differences. Finally, perform eigenvalue decomposition and clustering on the similarity matrix through the spectral clustering algorithm, and divide the samples into different meta-tasks. Each meta-task is a sample set, and the samples in it have similar geological conditions and blasting parameters. The specific operation steps are as follows: (1) Calculate the difference metric of the frequency band energy distribution of the samples and construct an energy distribution matrix; The calculation formula for the difference metric of the frequency band energy distribution between any two samples is as follows: , Among them, is the energy distribution matrix of the element, indicating the difference measure of the band energy distribution between the th sample and the th sample; is the time segment; is the band energy of the th sample in the th time slice; is the band energy of the th sample in the th time slice; Energy distribution matrix is represented as follows: , Among them, represents the number of samples, , , when , ; (2) Calculate the similarity between samples and construct a similarity matrix; The calculation formula for the similarity between any two samples is as follows: , Among them, is the similarity matrix element, representing the similarity between the th sample and the th sample; is the exponential function with the natural constant as the base; is the similarity attenuation coefficient; is the indicator function, which takes 1 when the difference of the parameters of the indicator function is less than the preset blasting parameter difference threshold and takes 0 otherwise; is the preset blasting parameter difference threshold; is the quantification value of the blasting parameter difference between the th sample and the th sample; Similarity matrix is represented as follows: , Among them, when , ; (3) Use the spectral clustering algorithm to generate meta-tasks: For the similarity matrix perform Laplacian matrix eigenvalue decomposition, and select the first eigenvectors for K-means clustering. Assign samples of the same class to the same meta-task to ensure that samples within the task have similar geological-blasting conditions.

4. The prediction method for the vibration influence of tunnel excavation blasting on surface buildings according to claim 3, characterized in that S3 is specifically as follows: Extract features from the samples in each meta-task through a neural network model. The feature extraction process includes initializing the parameters of the neural network model, performing dual-channel feature extraction in the time domain and frequency domain, fusing the extracted features, and performing meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task; (1)Initialize the parameters of the neural network model: Adopt a neural network initialization method based on vibration energy distribution. Map the samples in each meta-task to low-dimensional embedding vectors through a pre-trained autoencoder neural network, calculate the mean of the embedding vectors within each meta-task as the prototype of each meta-task, and then, based on the statistical distribution of the prototype vectors of each meta-task and the historical task parameter offsets, generate the initial network weights through weighted fusion; (2)Perform dual-channel feature extraction in the time domain and frequency domain and then conduct feature fusion: Adopt a structure of dual-channel interaction, and define the structure of the dual-channel feature fusion neural network, including the time-domain path neural network and the frequency-domain path neural network; 1) Time-domain path neural network: Perform one-dimensional convolution on the samples in the meta-task to extract time-domain features, and fuse the frequency-domain features through a cross-path attention gating mechanism to obtain the output features of the time-domain path neural network; 2) Frequency-domain path neural network: Perform short-time Fourier transform on the samples in the meta-task, encode the frequency-domain features through 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 path neural network; 3) Dual-channel feature fusion: Perform cross-path interaction on the output features of the time-domain path neural network and the output features of the frequency-domain path neural network, enhance the features through element-wise multiplication, and then achieve complementary feature fusion through element-wise addition to finally generate a representation vector containing time-frequency joint features; (3)Conduct meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task: Adopt a dynamic learning rate adjustment strategy. First, extract the energy peak moments of each segment of the vibration signal, calculate the average energy decay ratio within the time period after the peak as the decay coefficient. When updating the loop parameters in the neural network model, according to the signal time point corresponding to the current training step and the decay coefficient, smoothly adjust the base learning rate through the hyperbolic tangent function to make the parameter update intensity consistent with the energy decay trend. When updating the outer loop, aggregate the gradients of multiple tasks and optimize the meta-parameters.

5. The prediction method for the vibration influence of tunnel excavation blasting on surface buildings according to claim 4, characterized in that, The initialization of the parameters of the neural network model is specifically as follows: (1)Calculate the prototype vectors of the meta-tasks: Map all the samples in each meta-task to low-dimensional embedding representations through a pre-trained autoencoder neural network, and calculate the mean of the in-task embedding representations as the task prototype; (2)Generate the initialization parameters: Generate the initial network weights through weighted fusion based on the meta-task prototype vectors and the historical parameter offsets, and then generate the network initialization weight matrix.

6. A prediction method for the vibration influence of tunnel excavation blasting on surface buildings according to claim 4, characterized in that The meta-learning based on the vibration energy decay curve to optimize the meta-parameters in the meta-task is specifically as follows: (1)Construct the energy decay coefficient: First, locate the energy peak points of each segment of the vibration signal, and then, within the time period after the peak, count the relative energy per unit time, calculate its ratio to the peak energy, and then construct the overall energy decay coefficient; (2)Dynamically adjust the in-loop learning rate: The inner loop represents the parameter update within a meta-task. The learning rate is dynamically adjusted according to the energy decay coefficient. First, the time point corresponding to the current step is obtained. The energy decay coefficient is input into the hyperbolic tangent function to generate a smooth adjustment value. Then, the smooth adjustment value is multiplied by the base learning rate to obtain the adjustment value for the base learning rate. Next, the base learning rate and the adjustment value of the base learning rate are added together to obtain the adaptive learning rate for the inner loop. (3) Outer loop parameter update: The outer loop represents the optimization of meta-parameters. After performing inner loop training on multiple meta-tasks respectively and obtaining the specific model parameters of each meta-task, the loss function of all meta-tasks on the support set is calculated. The neural network model parameters are backpropagated, 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. A prediction method for the vibration influence of tunnel excavation blasting on surface buildings according to claim 4, characterized in that S4 is specifically as follows: The hierarchical attention mechanism is adopted. According to the event response characteristics of different samples at different time scales, the short-time window convolutional neural network, the medium-time window long short-term memory neural network, and the long-time window global pooling perception neural network are respectively used to extract attention weights, focusing on instantaneous, event-level, and global vibration changes respectively. The attention vectors at each scale are weighted element-wise with the fusion features and then superimposed to generate a multi-scale classification feature vector. Then, the classification feature vector is input into the preset Softmax function to obtain the predicted class probability vector. The class corresponding to the maximum class probability in the class probability vector is taken as the classification prediction result of the impact of blasting on surface building vibration. The short-time window convolutional neural network performs short-time window attention enhancement operations and sets , and the calculation formula is as follows: ; The medium-time window long short-term memory neural network implements the medium-time window attention enhancement operation and sets , and the calculation formula is as follows: ; The long-time window global pooling-aware neural network performs a long-term attention enhancement operation and sets , and the calculation formula is as follows: ; The calculation formula for multi-scale feature aggregation is as follows: ; Among them, is a representation vector containing time-frequency joint features; is the time window length of the attention mechanism; is the short-time window attention weight; is the Softmax function; is the short-time window convolution kernel parameter; is the long short-term memory neural network; is the output of the long short-term memory neural network transpose; is the medium-time window attention weight; is the medium-time window linear transformation parameter; is the long-time window attention weight; is the multi-layer perceptron function; is the Sigmoid activation function; represents element-wise multiplication; is the global max pooling operation; is the classification feature vector after aggregating multi-scale attention.

8. A prediction method for the vibration influence of tunnel excavation blasting on surface buildings according to claim 7, characterized in that The generalization verification and dynamic adjustment of meta-tasks are specifically as follows: The meta-tasks are divided into a training set and a validation set. In the inner loop, the neural network model parameters are quickly fine-tuned using the support set, and in the outer loop, the meta-parameters are updated based on the loss of the query set. During the training process, the performance of the neural network model is evaluated periodically on the validation meta-tasks. The loss calculation in the neural network model uses cross-entropy loss, and the loss function is calculated once in each training. If the loss does not decrease for consecutive P periods, the learning rate of the outer loop is decreased 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 the independent test set.

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

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

Citation Information

Patent Citations

  • Vibration damage experiment system for large foundation pit near subway area

    CN118862655A

  • Building health monitoring and evaluation method and system based on physical neural network

    CN119249073A

  • Object-level infrared-and-visible-light image fusion method based on fully convolutional neural network

    WO2024174488A1

Cited By

  • Intelligent identification method and device for poor lubrication of equipment for equipment integrity management

    CN120744587A

  • Device lubrication poor intelligent identification method and device for device integrity management

    CN120744587B

  • Shallow-buried tunnel blasting surface vibration waveform prediction method and system

    CN120833016A

  • Knock vibration detection method, device, and storage medium

    CN122594890A