CN-121234719-B - Blasting vibration speed waveform simulation method and system based on deep learning

CN121234719BCN 121234719 BCN121234719 BCN 121234719BCN-121234719-B

Abstract

The invention provides a blasting vibration speed waveform simulation method and system based on deep learning, and relates to the technical field of engineering blasting, wherein the method comprises the steps of collecting blasting vibration speed waveform data of a plurality of different monitoring points and corresponding blasting parameters for a plurality of times, and preprocessing to construct a database for model training; the method comprises the steps of constructing a Unet-VAE-GAN model, training the model to learn a nonlinear mapping relation between blasting parameters and blasting vibration speed waveforms and generate potential vectors, constructing a multi-layer perceptron model, training the model to build the nonlinear mapping relation between different blasting parameters and the potential vectors of the Unet-VAE-GAN model, inputting actual blasting parameters into the trained multi-layer perceptron model to output corresponding potential vectors, and inputting the output potential vectors into the trained Unet-VAE-GAN model to generate simulated blasting vibration speed waveforms.

Inventors

XU CHEN
QIU LANG
ZHU YUJIE
WANG CHAO
REN GAOFENG
HU YINGGUO
LIU XIAOLI

Assignees

武汉理工大学

Dates

Publication Date: 20260508
Application Date: 20250901

Claims (9)

1. The blasting vibration speed waveform simulation method based on deep learning is characterized by comprising the following steps of: the method comprises the steps of acquiring blasting vibration speed waveform data of a plurality of different monitoring points and corresponding blasting parameters for a plurality of times, and preprocessing to construct a database for model training; constructing Unet-VAE-GAN model, training to enable the model to learn nonlinear mapping relation between blasting parameters and blasting vibration speed waveforms, and generating potential vectors; The Unet-VAE-GAN model comprises an encoder, a decoder and a discriminator; The encoder adopts a U-Net downsampling path structure and comprises three convolution blocks for feature extraction and downsampling, each convolution block comprises two convolution layers and a ReLU activation function, the two convolution layers are connected with a maximum pooling layer for downsampling, and potential vectors are finally output through a full connection layer; The decoder is provided with two parallel up-sampling paths, wherein one path is connected with the feature map and the potential vector of each layer of the fusion encoder through jumping, the other path carries out up-sampling reconstruction by using the potential vector only, and finally, reconstruction signals are output through convolution and a Tanh activation function; The discriminator adopts a convolutional neural network structure to discriminate the authenticity, extracts the characteristics through three convolutional blocks, and finally activates and outputs the true probability through the full connection layer and Sigmoid; establishing a multi-layer perceptron model, and establishing a nonlinear mapping relation between different blasting parameters and potential vectors of the Unet-VAE-GAN model through training; inputting the actual blasting parameters into the trained multi-layer perceptron model to output corresponding potential vectors, and inputting the output potential vectors into the trained Unet-VAE-GAN model to generate a simulated blasting vibration velocity waveform.
2. The deep learning based blasting vibration velocity waveform simulation method according to claim 1, wherein the preprocessing step specifically comprises: performing alignment treatment on the blasting vibration speed waveform data to obtain an initial blasting vibration speed waveform; optimizing parameter combinations of variation modal decomposition through a genetic algorithm, and performing variation modal decomposition on the initial blasting vibration speed waveform by utilizing the optimized parameters to obtain a plurality of modal components; And calculating the contribution rate of each modal component through the analysis of the nuclear principal component, and reserving the first plurality of modal components with the accumulated contribution rate exceeding a preset threshold value to be combined into the blasting vibration velocity waveform after noise reduction.
3. The deep learning based blasting vibration velocity waveform simulation method according to claim 2, wherein the step of optimizing the parameter combination of the variation modal decomposition by the genetic algorithm specifically comprises: Randomly generating an initial population, wherein the individuals of the population are parameter combinations of the modal number K and the penalty factor alpha, calculating the fitness value of each individual in the population through a fitness function, and storing the individuals with the highest fitness; screening high-quality individuals through a roulette selection mechanism, generating new individuals by adopting arithmetic crossover operation with random weight, and introducing Gaussian random disturbance to implement mutation operation; And iteratively executing the operations of selection, crossing and mutation until the preset termination condition is met, and outputting the optimal mode number K and penalty factor alpha parameter combination.
4. The deep learning-based blasting vibration velocity waveform simulation method according to claim 2, wherein the step of performing a variation modal decomposition on the initial blasting vibration velocity waveform by using the optimized parameters, and obtaining a plurality of modal components specifically comprises: Constructing an objective function comprising modal component bandwidth minimization and original signal reconstruction constraint; the method adopts an alternate direction multiplier method to carry out iterative solution on the objective function, and specifically comprises the following steps: a Lagrange multiplier and a penalty term are introduced to construct a Lagrange function; When the modal component is updated, maintaining the center frequency and the Lagrangian multiplier unchanged, and minimizing the Lagrangian function; When the center frequency is updated, the mode components are kept unchanged, and the center frequency of each mode is updated through the calculation of the spectrum moment; when updating the Lagrange multiplier, updating the Lagrange multiplier by a gradient rising method; and alternately updating the modal components, the center frequency and the Lagrange multipliers until a preset convergence condition is met, and outputting a plurality of modal components with specific center frequencies.
5. The deep learning based blasting vibration velocity waveform simulation method according to claim 1, wherein the training step of Unet-VAE-GAN model specifically comprises: Performing joint training on the Unet-VAE-GAN model through a multi-objective loss function by adopting a staged optimization strategy; When training the discriminator, calculating the discrimination loss of the real signal and the discrimination loss of the generated signal, and back-propagating and updating the parameters of the discriminator to optimize the discriminator so as to distinguish the real blasting vibration signal from the signal synthesized by the decoder; when the decoder and the encoder are jointly trained, the encoder maps real signals into mean and variance parameters of potential space, potential vectors are obtained through sampling by a random variable re-parameterization method, the decoder fuses the potential vectors and jump connection characteristics provided by the encoder to reconstruct signals, reconstruction loss, KL divergence loss and counterloss are calculated, weighted summation is carried out to serve as total loss, and the decoder and the encoder parameters are optimized through back propagation; The whole training cycle is repeatedly carried out in a preset period, and the reconstructed signal is gradually optimized through dynamic reconstruction precision, distribution regularization and countermeasure training intensity.
6. The deep learning-based blasting vibration velocity waveform simulation method according to claim 1, wherein the training step of the multi-layer perceptron specifically comprises: taking the blasting parameters subjected to standardized processing in the database as data input, and taking potential vectors as supervision labels to form training sample pairs; initializing multi-layer perceptron network weights, and mapping blasting parameters into predicted potential vectors through forward propagation; calculating the difference between the predicted potential vector and the actual potential vector by adopting a mean square error loss function; and iteratively updating weight parameters of the multi-layer perceptron network by adopting a back propagation algorithm, and minimizing the mean square error loss function to establish a nonlinear mapping relation from blasting parameters to potential vectors.
7. A blasting vibration velocity waveform simulation system based on deep learning, comprising: the database construction module is used for constructing a database for model training by acquiring the blasting vibration speed waveform data of a plurality of different monitoring points and corresponding blasting parameters for a plurality of times and preprocessing the blasting vibration speed waveform data; Unet-VAE-GAN module for constructing Unet-VAE-GAN model, training to make model learn nonlinear mapping relation between blasting parameter and blasting vibration speed waveform, and generating potential vector; The Unet-VAE-GAN model comprises an encoder, a decoder and a discriminator; The encoder adopts a U-Net downsampling path structure and comprises three convolution blocks for feature extraction and downsampling, each convolution block comprises two convolution layers and a ReLU activation function, the two convolution layers are connected with a maximum pooling layer for downsampling, and potential vectors are finally output through a full connection layer; The decoder is provided with two parallel up-sampling paths, wherein one path is connected with the feature map and the potential vector of each layer of the fusion encoder through jumping, the other path carries out up-sampling reconstruction by using the potential vector only, and finally, reconstruction signals are output through convolution and a Tanh activation function; The discriminator adopts a convolutional neural network structure to discriminate the authenticity, extracts the characteristics through three convolutional blocks, and finally activates and outputs the true probability through the full connection layer and Sigmoid; the multi-layer perceptron module is used for establishing a multi-layer perceptron model, and establishing a nonlinear mapping relation between different blasting parameters and potential vectors of the Unet-VAE-GAN model through training; the simulation waveform generation module is used for inputting actual blasting parameters into the trained multi-layer perceptron model to output corresponding potential vectors, and inputting the output potential vectors into the trained Unet-VAE-GAN model to generate a simulation blasting vibration speed waveform.
8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the steps of the deep learning based blast vibration velocity waveform simulation method of any one of claims 1 to 6.
9. A computer readable storage medium having stored thereon a computer program, characterized in that the computer program, when executed by a processor, implements the steps of the deep learning based blasting vibration velocity waveform simulation method of any one of claims 1 to 6.

Description

Blasting vibration speed waveform simulation method and system based on deep learning Technical Field The invention relates to the technical field of engineering blasting, in particular to a blasting vibration speed waveform simulation method and system based on deep learning. Background In blasting engineering practices in the wide fields of mining, civil engineering, transportation hub construction, building demolition and the like, blasting operation points often inevitably adjoin sensitive areas such as residential areas, key infrastructure and the like. At this time, the energy of the vibration wave generated by the explosion is carefully designed and controlled, but the vibration wave may cause a risk of potential accumulated damage to the adjacent structure and possibly interfere with the life of the residents. This is a core problem in engineering safety and environmental protection that must be carefully managed. At this time, the method has important significance in analyzing the dynamic load and structural response characteristics of the whole blasting vibration process. The waveform simulation can predict the whole process of blasting vibration generated by engineering blasting and comprehensively characterize the time domain and the frequency domain of the blasting vibration. The most common blasting vibration waveform simulation method at present is a time domain linear superposition method, and the basic principle is based on a seismic wave superposition theory, wherein a single-hole blasting vibration waveform is taken as a wavelet, and the Cheng Diejia algorithm is used for carrying out superposition calculation on blasting vibration of seismic waves in a field medium under inter-hole and inter-section delay time. Many scholars have conducted a great deal of research on their principles and applications over the years. Although the superposition calculation improves the accuracy and the comprehensiveness of the blasting vibration prediction result to a certain extent, the problem still exists in the current waveform simulation process that firstly, most blasting vibration speed waveform simulation methods are based on a time domain linear superposition method, the method cannot achieve the purpose that the porous blasting vibration speed waveform is obtained through the input of various blasting parameters, and the complex mapping relation between the various blasting parameters and the porous blasting vibration speed waveform cannot be established, so that the influence of different blasting parameters on the porous blasting vibration speed waveform is difficult to analyze. Secondly, the implementation process of the time-course superposition algorithm requires that a professional has deeper knowledge on the signal processing method, and the method has poor visualization and man-machine interaction effects, so that the implementation of the method is difficult for people without signal processing knowledge, thereby influencing engineering application. Disclosure of Invention The invention aims to provide a blasting vibration speed waveform simulation method and system based on deep learning, which are used for solving the problems that the current blasting vibration speed waveform simulation method adopting a time domain linear superposition method is difficult to analyze the influence of different blasting parameters on a porous blasting vibration speed waveform and has poor visualization and man-machine interaction effects. The technical scheme includes that the method comprises the steps of collecting explosion vibration speed waveform data of a plurality of different monitoring points and corresponding explosion parameters for a plurality of times, preprocessing the explosion vibration speed waveform data to construct a model training database, constructing Unet-VAE-GAN models, enabling the models to learn a nonlinear mapping relation between explosion parameters and explosion vibration speed waveforms through training to generate potential vectors, building a multi-layer perceptron model, enabling the different explosion parameters and the potential vectors of the Unet-VAE-GAN model to build the nonlinear mapping relation through training, inputting actual explosion parameters into the trained multi-layer perceptron model, outputting the corresponding potential vectors, and inputting the output potential vectors into the trained Unet-VAE-GAN model to generate simulated explosion vibration speed waveforms. Optionally, the preprocessing step specifically includes performing alignment processing on the blasting vibration speed waveform data to obtain an initial blasting vibration speed waveform, optimizing parameter combinations of variation modal decomposition through a genetic algorithm, performing variation modal decomposition on the initial blasting vibration speed waveform by utilizing the optimized parameters to obtain a plurality of modal components, calculating contribution rates