Blasting vibration velocity waveform simulation method and system based on deep learning

Abstract

The application provides a blasting vibration velocity waveform simulation method and system based on deep learning, and relates to the technical field of engineering blasting. The method steps include: collecting blasting vibration velocity waveform data and corresponding blasting parameters of a plurality of different monitoring points for a plurality of times, and pre-processing to construct a database for model training; constructing an Unet-VAE-GAN model, training the model to learn the nonlinear mapping relationship between the blasting parameters and the blasting vibration velocity waveform, and generating a latent vector; establishing a multilayer perceptron model, training the model to establish a nonlinear mapping relationship between different blasting parameters and the latent vector of the Unet-VAE-GAN model; inputting the actual blasting parameters into the trained multilayer perceptron model to output the corresponding latent vector, and inputting the output latent vector into the trained Unet-VAE-GAN model to generate a simulated blasting vibration velocity waveform.

Description

Technical Field

[0001] This invention relates to the field of engineering blasting technology, and in particular to a method and system for simulating blasting vibration velocity waveforms based on deep learning. Background Technology

[0002] In blasting engineering practices across a wide range of fields, including mining, civil engineering, transportation hub construction, and building demolition, blasting sites are often unavoidably located near residential areas, critical infrastructure, and other sensitive locations. In such cases, although the energy of the vibration waves generated by the blast is carefully designed and controlled, the potential cumulative damage risk to adjacent structures and the possible disruption to residents' lives pose a significant challenge. This becomes a core issue that cannot be ignored and requires meticulous management in engineering safety and environmental protection. Therefore, analyzing the dynamic load and structural response characteristics of the entire blasting vibration process is of great importance. Waveform simulation can predict the entire blasting vibration process generated by engineering blasting and comprehensively characterize its time and frequency domain features.

[0003] The most common method for simulating blasting vibration waveforms is the time-domain linear superposition method. Its basic principle is based on seismic wave superposition theory. Using the single-hole blasting vibration waveform as a wavelet, a time-history superposition algorithm is used to superimpose and calculate the blasting vibrations in the site medium under inter-hole and inter-segment delay times. Over the years, many scholars have conducted extensive research on its principles and applications. Although this superposition calculation improves the accuracy and comprehensiveness of blasting vibration prediction results to some extent, the following problems still exist in the waveform simulation process: First, most blasting vibration velocity waveform simulation methods are based on the time-domain linear superposition method. This method cannot obtain multi-hole blasting vibration velocity waveforms by inputting multiple blasting parameters, and it cannot establish a complex mapping relationship between multiple blasting parameters and multi-hole blasting vibration velocity waveforms, making it difficult to analyze the influence of different blasting parameters on multi-hole blasting vibration velocity waveforms. Second, the implementation of the time-history superposition algorithm requires professionals with a deep understanding of signal processing methods, and the method has poor visualization and human-computer interaction effects. This makes the implementation of this method difficult for personnel without signal processing knowledge, thus affecting engineering applications. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method and system for simulating blasting vibration velocity waveforms, in order to solve the problems mentioned in the background art, such as the difficulty in analyzing the influence of different blasting parameters on the vibration velocity waveform of multi-hole blasting, as well as the poor visualization and human-computer interaction effects of the current blasting vibration waveform simulation method using the time-domain linear superposition method.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for simulating blasting vibration velocity waveforms, comprising the following steps: collecting blasting vibration velocity waveform data and corresponding blasting parameters from several different monitoring points multiple times, and preprocessing them to construct a database for model training; constructing a Unet-VAE-GAN model, training the model to learn the nonlinear mapping relationship between blasting parameters and blasting vibration velocity waveforms, and generating latent vectors; establishing a multilayer perceptron model, training it to establish a nonlinear mapping relationship between different blasting parameters and the latent vectors of the Unet-VAE-GAN model; inputting actual blasting parameters into the trained multilayer perceptron model to output the corresponding latent vectors, and inputting the output latent vectors into the trained Unet-VAE-GAN model to generate simulated blasting vibration velocity waveforms.

[0006] Optionally, the preprocessing steps specifically include: aligning the blasting vibration velocity waveform data to obtain an initial blasting vibration velocity waveform; optimizing the parameter combination of variational mode decomposition using a genetic algorithm, and using the optimized parameters to perform variational mode decomposition on the initial blasting vibration velocity waveform to obtain multiple modal components; calculating the contribution rate of each modal component through kernel principal component analysis, and retaining the first few modal components whose cumulative contribution rate exceeds a preset threshold and merging them into a denoised blasting vibration velocity waveform.

[0007] Optionally, the step of optimizing the parameter combination of variational mode decomposition using a genetic algorithm specifically includes: randomly generating an initial population, wherein the individuals in the population are a parameter combination of the number of modes K and the penalty factor α; calculating the fitness value of each individual in the population using a fitness function and saving the individual with the highest fitness; selecting high-quality individuals using a roulette wheel selection mechanism; generating new individuals using arithmetic crossover with random weights and introducing Gaussian random perturbation to implement mutation operations; iteratively executing selection, crossover, and mutation operations until a preset termination condition is met, and outputting the optimal parameter combination of the number of modes K and the penalty factor α.

[0008] Optionally, the step of performing variational mode decomposition on the initial blasting vibration velocity waveform using optimized parameters to obtain multiple modal components specifically includes: constructing an objective function containing constraints on minimizing modal component bandwidth and reconstructing the original signal; iteratively solving the objective function using the alternating direction multiplier method, specifically including: constructing a Lagrange function by introducing Lagrange multipliers and penalty terms; minimizing the Lagrange function while keeping the center frequency and Lagrange multipliers constant when updating modal components; updating the center frequency of each mode while keeping the modal components constant by calculating the spectral moments; updating the Lagrange multipliers by using the gradient ascent method when updating the Lagrange multipliers; alternately updating the modal components, center frequency, and Lagrange multipliers until a preset convergence condition is met, and outputting multiple modal components with specific center frequencies.

[0009] Optionally, the Unet-VAE-GAN model includes an encoder, a decoder, and a discriminator. The encoder adopts a U-Net downsampling path structure, containing three convolutional blocks for feature extraction and downsampling. Each convolutional block includes two convolutional layers and a ReLU activation function, followed by a max-pooling layer for downsampling, and finally outputs a latent vector through a fully connected layer. The decoder has two parallel upsampling paths. One path fuses the feature maps and latent vectors from each layer of the encoder through skip connections, while the other path uses only the latent vectors for upsampling reconstruction, and finally outputs the reconstructed signal through convolution and a Tanh activation function. The discriminator adopts a convolutional neural network structure for authenticity judgment, extracts features through three convolutional blocks, and finally outputs the true probability through a fully connected layer and a Sigmoid activation function.

[0010] Optionally, the training steps of the Unet-VAE-GAN model specifically include: employing a phased optimization strategy to jointly train the Unet-VAE-GAN model using a multi-objective loss function; when training the discriminator, calculating the discrimination loss for real signals and the discrimination loss for generated signals, and updating the discriminator parameters through backpropagation to optimize the discriminator so that it can distinguish between real blasting vibration signals and signals synthesized by the decoder; when jointly training the decoder and encoder, the encoder maps the real signal to the mean and variance parameters of the latent space, and obtains the latent vector by sampling through a random variable reparameterization method; the decoder fuses the latent vector with the jump connection features provided by the encoder to reconstruct the signal, and simultaneously calculates the reconstruction loss, KL divergence loss, and adversarial loss and performs a weighted summation as the total loss; the parameters of the decoder and encoder are simultaneously optimized through backpropagation; the entire training cycle is repeated within a preset period, and the reconstructed signal is gradually optimized by dynamically balancing reconstruction accuracy, distribution regularization, and adversarial training intensity.

[0011] Optionally, the training steps of the multilayer perceptron specifically include: using standardized blasting parameters from the database as data input and latent vectors as supervision labels to form training sample pairs; initializing the weights of the multilayer perceptron network and mapping the blasting parameters to predicted latent vectors through forward propagation; calculating the difference between the predicted latent vectors and the true latent vectors using the mean squared error loss function; and iteratively updating the weight parameters of the multilayer perceptron network using the backpropagation algorithm to minimize the mean squared error loss function, thereby establishing a nonlinear mapping relationship from blasting parameters to latent vectors.

[0012] On the other hand, the present invention also provides a deep learning-based blasting vibration velocity waveform simulation system, comprising: a database construction module, used to collect blasting vibration velocity waveform data and corresponding blasting parameters from several different monitoring points multiple times and perform preprocessing to construct a database for model training; a Unet-VAE-GAN module, used to construct a Unet-VAE-GAN model, which learns the nonlinear mapping relationship between blasting parameters and blasting vibration velocity waveforms through training and generates latent vectors; a multilayer perceptron module, used to establish a multilayer perceptron model, which establishes a nonlinear mapping relationship between different blasting parameters and the latent vectors of the Unet-VAE-GAN model through training; and a simulation waveform generation module, used to input actual blasting parameters into the trained multilayer perceptron model to output the corresponding latent vectors, and input the output latent vectors into the trained Unet-VAE-GAN model to generate simulated blasting vibration velocity waveforms.

[0013] On the other hand, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described deep learning-based blasting vibration velocity waveform simulation method.

[0014] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described deep learning-based blasting vibration velocity waveform simulation method.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] This application achieves comprehensive coverage and efficient utilization of the actual engineering blasting vibration characteristics by collecting blasting vibration velocity waveform data and corresponding blasting parameters from different monitoring points multiple times and constructing a database. It eliminates the need to collect single-hole blasting vibration test data required by traditional methods, breaking through the dependence of traditional methods on single-hole test data. This provides a rich and diverse source of samples for model training, significantly improving data utilization efficiency and engineering applicability.

[0017] This application constructs the Unet-VAE-GAN model. Unlike the time-domain linear superposition method that relies on fixed physical superposition rules, the Unet-VAE-GAN model possesses powerful nonlinear fitting and feature extraction capabilities. Through a deep neural network architecture, it automatically captures the end-to-end, high-dimensional nonlinear relationship between multidimensional blasting parameters and complex waveform features, thereby directly generating multi-hole blasting vibration velocity waveforms that match actual engineering conditions. This overcomes the limitation of traditional methods that cannot flexibly handle multiple parameter combinations and directly output complete waveforms. Because the model learns the comprehensive influence of different blasting parameters on waveform morphology and amplitude characteristics during training, it can not only accurately reconstruct the overall waveform morphology but also effectively capture subtle fluctuation details and phase changes, making it possible to analyze the fine-grained effects of different blasting parameters on waveform morphology, amplitude, spectrum, and duration.

[0018] This application establishes a multilayer perceptron model to map blasting parameters to latent vectors, thereby constructing an intelligent conversion bridge between engineering parameters and waveform feature space. By transforming intuitive and adjustable engineering parameters into a latent space representation that can be understood by a deep generative model, even non-experts can obtain professional-grade waveform prediction results with simple parameter input, significantly reducing the technical threshold and improving the ease of use of the method. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0020] Figure 2 This is a flowchart of the method of the present invention.

[0021] Figure 3 This is a deep learning block diagram of the method of the present invention.

[0022] Figure 4 This is a system architecture block diagram of the present invention.

[0023] Figure 5 This is a schematic diagram illustrating the simulation effect of the method of the present invention.

[0024] Figure 6 This is a schematic diagram of the system structure of the present invention.

[0025] In the diagram: 10 - Database construction module, 20 - Unet-VAE-GAN module, 30 - Multilayer perceptron module, 40 - Simulation waveform generation module. Detailed Implementation

[0026] The present invention will now be clearly and completely described in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be used interchangeably where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0029] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0030] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0031] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] Please refer to Figures 1-5This invention discloses a deep learning-based method for simulating blasting vibration velocity waveforms, comprising the following steps: collecting blasting vibration velocity waveform data and corresponding blasting parameters from several different monitoring points multiple times, and preprocessing them to construct a database for model training; constructing a Unet-VAE-GAN model, training it to learn the nonlinear mapping relationship between blasting parameters and blasting vibration velocity waveforms, and generating latent vectors; establishing a multilayer perceptron model, training it to establish a nonlinear mapping relationship between different blasting parameters and the latent vectors of the Unet-VAE-GAN model; inputting actual blasting parameters into the trained multilayer perceptron model to output the corresponding latent vectors, and inputting the output latent vectors into the trained Unet-VAE-GAN model to generate simulated blasting vibration velocity waveforms.

[0033] Specifically, this application achieves comprehensive coverage and efficient utilization of the actual engineering blasting vibration characteristics by collecting blasting vibration velocity waveform data and corresponding blasting parameters from different monitoring points multiple times and constructing a database. It eliminates the need to collect single-hole blasting vibration test data required by traditional methods, breaks through the dependence of traditional methods on single-hole test data, provides a rich and diverse source of samples for model training, and significantly improves data utilization efficiency and engineering applicability.

[0034] This application constructs the Unet-VAE-GAN model. Unlike the time-domain linear superposition method that relies on fixed physical superposition rules, the Unet-VAE-GAN model possesses powerful nonlinear fitting and feature extraction capabilities. Through a deep neural network architecture, it automatically captures the end-to-end, high-dimensional nonlinear relationships between multi-dimensional blasting parameters (such as detonation distance, maximum single-segment charge, elevation difference, hole depth, hole inclination, row spacing, explosive consumption per unit volume, charge per hole, plugging length, and number of holes) and complex waveform features. This allows it to directly generate multi-hole blasting vibration velocity waveforms that match actual engineering conditions. This overcomes the limitations of traditional methods that cannot flexibly handle multiple parameter combinations and directly output complete waveforms. Because the model learns the comprehensive influence of different blasting parameters on waveform morphology and amplitude characteristics during training, it can not only accurately reconstruct the overall waveform morphology but also effectively capture subtle fluctuation details and phase changes, making it possible to analyze the fine-grained effects of different blasting parameters on waveform morphology, amplitude, spectrum, and duration.

[0035] This application establishes a multilayer perceptron model to map blasting parameters to latent vectors, thereby constructing an intelligent conversion bridge between engineering parameters and waveform feature space. By transforming intuitive and adjustable engineering parameters into a latent space representation that can be understood by a deep generative model, even non-experts can obtain professional-grade waveform prediction results with simple parameter input, significantly reducing the technical threshold and improving the ease of use of the method.

[0036] In some embodiments, the preprocessing step specifically includes: aligning the blasting vibration velocity waveform data to obtain an initial blasting vibration velocity waveform; optimizing the parameter combination of variational mode decomposition using a genetic algorithm, and using the optimized parameters to perform variational mode decomposition on the initial blasting vibration velocity waveform to obtain multiple modal components; calculating the contribution rate of each modal component using kernel principal component analysis, and retaining the first few modal components whose cumulative contribution rate exceeds a preset threshold and merging them into a denoised blasting vibration velocity waveform.

[0037] Specifically, probes are deployed according to the blasting vibration monitoring scheme. A blasting vibration testing instrument monitors the three-dimensional blasting vibration velocity waveforms at different measuring points. For this waveform signal, the initial blasting vibration velocity waveform is taken 1 second before and after the peak value. This processing method is to align the lengths of all data. Because blasting vibration is affected by various factors, the blasting vibration velocity waveforms obtained by the blasting vibration testing instrument will have length differences. In addition, the corresponding blasting parameters of the blasting vibration velocity waveforms also need to be recorded. These blasting parameters include the detonation center distance, maximum single-stage charge, elevation difference, hole depth, hole inclination, row spacing, explosive consumption per unit area, charge per hole, plugging length, and number of holes.

[0038] Furthermore, the parameter combination of variational mode decomposition is optimized by a genetic algorithm, and the optimized parameters are used to perform variational mode decomposition on the initial blasting vibration velocity waveform to obtain multiple modal components.

[0039] Furthermore, kernel principal component analysis was used to calculate the contribution rate and cumulative contribution rate of each modal component IMF. The top m modal component IMFs with a cumulative contribution exceeding 85% were retained, and the remaining modal component IMFs were discarded. Then, the retained modal component IMFs were merged into a noise-reduced blasting vibration velocity waveform.

[0040] This application employs variational mode decomposition driven by a genetic algorithm to denoise blasting vibration signals. It adaptively separates effective modal components from noise interference, overcoming the limitations of traditional filtering methods that rely on empirically set parameters. Combined with a modal screening mechanism based on kernel principal component analysis, it retains over 85% of the effective information while eliminating redundant components, providing a standardized data foundation with a high signal-to-noise ratio for subsequent model training and fundamentally improving the reliability of waveform simulation.

[0041] In some embodiments, the step of optimizing the parameter combination of variational mode decomposition using a genetic algorithm specifically includes: randomly generating an initial population, wherein the individuals in the population are a parameter combination of the number of modes K and the penalty factor α; calculating the fitness value of each individual in the population using a fitness function and saving the individual with the highest fitness; selecting high-quality individuals using a roulette wheel selection mechanism; generating new individuals using arithmetic crossover with random weights and introducing Gaussian random perturbation to implement mutation operations; iteratively executing selection, crossover, and mutation operations until a preset termination condition is met, and outputting the optimal parameter combination of the number of modes K and the penalty factor α.

[0042] Specifically, the parameters of variational mode decomposition (VMD) are optimized using a genetic algorithm: the number of modes K and the penalty factor α. After finding the optimal parameter combination, this is used as the parameters for VMD. Then, VMD is used to decompose the initial blasting vibration velocity waveform to obtain K modal components (IMFs). The basic steps of the genetic algorithm are as follows: An initial population is randomly generated, where each individual... Represented using binary or real number encoding. Population size set to... And define the fitness function. To assess the quality of individuals, the fitness value of each individual in the population is calculated. The individual with the highest fitness is preserved. Roulette wheel selection is used to select individuals based on their fitness values, calculated using the following formula:

[0043] In the formula: To select the probability, For individuals in a population fitness value, For population size, The summation index is used to iterate through every individual in the population. For individuals in a population The fitness value.

[0044] Furthermore, a crossover operation is performed on the selected parent individuals to generate new individuals. (Parent individuals) and Crossover is performed, and the new individuals generated are and Where α is a random number between 0 and 1.

[0045] ; ;

[0046] In the formula: and There are two parent individuals. and There are two new individuals, and α is a random number between 0 and 1.

[0047] Furthermore, a Gaussian random perturbation δ is introduced to mutate the genes of the new individuals, forming a new population from which the individuals with the highest fitness in the current generation are retained. If the fitness of the new individual is better than that of the previous generation, the optimal solution is updated. The process then checks if a termination condition is met, such as the number of iterations reaching a preset value or the fitness of an individual in the population reaching a preset value. If the condition is met, the optimal solution is output; otherwise, the iteration continues.

[0048] This application endows variational mode decomposition with adaptability through the parameter optimization process of a genetic algorithm, automatically locking the optimal combination of decomposition parameters by minimizing the envelope entropy. The synergistic effect of roulette wheel selection and Gaussian perturbation ensures both the genetic efficiency of high-quality genes and the ability to explore the solution space, avoiding the time consumption of traditional trial and error methods. This keeps VMD decomposition in an optimal working state, providing stable and consistent decomposition quality for signal denoising.

[0049] In some embodiments, the step of performing variational mode decomposition on the initial blasting vibration velocity waveform using optimized parameters to obtain multiple modal components specifically includes: constructing an objective function containing constraints on minimizing modal component bandwidth and reconstructing the original signal; iteratively solving the objective function using the alternating direction multiplier method, specifically including: constructing a Lagrange function by introducing Lagrange multipliers and penalty terms; minimizing the Lagrange function while keeping the center frequency and Lagrange multipliers constant when updating modal components; updating the center frequency of each mode while keeping the modal components constant by calculating the spectral moments; updating the Lagrange multipliers by using the gradient ascent method when updating the Lagrange multipliers; alternately updating the modal components, center frequency, and Lagrange multipliers until a preset convergence condition is met, and outputting multiple modal components with specific center frequencies.

[0050] Specifically, the goal of Variational Mode Decomposition (VMD) is to transform the signal... The system is decomposed into K modes, each with a specific center frequency. These modes are determined by minimizing the bandwidth. The objective function can be defined as:

[0051] ;

[0052] In the formula: This is the k-th modal component; The center frequency of the k-th mode; This represents the convolution operation; It is the Dirac delta function; For partial derivative operators; The complex unit is K, where K is the number of modes; signal It can be used in all modes The sum is used to represent the total, therefore the following constraints apply:

[0053] In the formula: The signal is a variational mode decomposition signal. Let K be the k-th modal component, and K be the number of modes.

[0054] Furthermore, to solve constrained optimization problems, Lagrange multipliers are introduced. Construct the Lagrangian function with the penalty term α:

[0055] ;

[0056] In the formula: For Lagrange functions, For Lagrange multipliers, This is the k-th modal component; The center frequency of the k-th mode; The signal is a variational mode decomposition signal; For partial derivative operators; This represents the convolution operation; It is the Dirac delta function; Let K be the complex unit, K be the number of modes, and α be the penalty term.

[0057] Furthermore, Variational Mode Decomposition (VMD) solves the optimization problem using the alternating direction multiplier method, updating the modal components. At the same time, maintain the center frequency. and Lagrange multipliers Unchanged, minimize about The Lagrangian function. The update formula for the modal components is as follows:

[0058] In the formula: This represents the frequency domain representation of the mode in the (n+1)th iteration; This is the summation of the frequency domain representations of all modes except the current k-th mode in the current iteration. This is the Fourier transform of the signal; For the Fourier transform of the Lagrange multipliers, It is the frequency variable; α is the penalty parameter; Let be the center frequency of the k-th mode in the n-th iteration.

[0059] Furthermore, after updating the modal components, while keeping the modal components unchanged, the center frequency of each mode is updated. The formula is as follows:

[0060] In the formula: Let be the frequency domain Fourier transform of the mode in the (n+1)th iteration, and d be the dimension of the latent vector. For frequency variables.

[0061] Finally, update the Lagrange multipliers. This ensures that the constraints are met. The formula is as follows:

[0062] In the formula: For step size parameters, For the (n+1)th iteration, the Lagrange multiplier is used. For the Lagrange multipliers in the nth iteration, The signal is a variational mode decomposition, where K is the number of modes. For the (n+1)th iteration, there are modal components.

[0063] The modal components, center frequencies, and Lagrange multipliers are updated alternately until the convergence condition is met. Finally, the algorithm outputs K modal components and their corresponding center frequencies.

[0064] This application utilizes an alternating direction optimization algorithm based on Lagrange multipliers to achieve accurate real-time decomposition of blasting vibration signals in the time-frequency domain. By balancing the constraint of minimizing modal bandwidth with signal reconstruction error, it ensures that each modal component (IMF) has a clear physical meaning. The dynamic update mechanism of the center frequency enables the decomposition results to adapt to changes in signal characteristics, making it particularly suitable for the analysis of non-stationary transient signals such as blasting vibration, and establishing a rigorous mathematical foundation for subsequent feature extraction.

[0065] In some embodiments, the Unet-VAE-GAN model includes an encoder, a decoder, and a discriminator. The encoder adopts a U-Net downsampling path structure, containing three convolutional blocks for feature extraction and downsampling. Each convolutional block includes two convolutional layers and a ReLU activation function, followed by a max-pooling layer for downsampling, and finally outputs a latent vector through a fully connected layer. The decoder has two parallel upsampling paths. One path fuses the feature maps and latent vectors from each layer of the encoder through skip connections, while the other path uses only the latent vectors for upsampling reconstruction, and finally outputs the reconstructed signal through convolution and a Tanh activation function. The discriminator adopts a convolutional neural network structure for authenticity judgment, extracts features through three convolutional blocks, and finally outputs the true probability through a fully connected layer and a Sigmoid activation function.

[0066] Specifically, the Unet-VAE-GAN model is the core innovative architecture, which uses deep learning methods to achieve high-precision, end-to-end simulation of blasting vibration velocity waveforms. The multilayer perceptron model enables the establishment of a nonlinear mapping relationship between different blasting parameters and the latent vectors in Unet-VAE-GAN.

[0067] Furthermore, the network architecture of this model consists of three core components: an encoder, a decoder, and a discriminator. The encoder employs a downsampling path similar to U-Net, containing three convolutional blocks. Each block consists of two convolutional layers and a ReLU activation function, followed by a max-pooling layer for downsampling. Finally, a fully connected layer outputs the mean and log-variance parameters of the latent space. The decoder, as the decoder part of U-Net, has two parallel paths with skip connections that fuse the feature maps of each layer of the encoder with the latent vectors through transposed convolutions for upsampling. The path without skip connections only uses the latent vectors for upsampling and reconstruction. Finally, convolutions and Tanh activations are used to output the reconstructed signal. The discriminator uses a traditional convolutional neural network structure, extracting features through three convolutional blocks. Finally, a fully connected layer and a Sigmoid activation are used to output the true probability. This design enables the encoder-decoder to form a U-Net structure with skip connections, achieving accurate reconstruction, while simultaneously improving the generation quality through adversarial training mechanisms.

[0068] Furthermore, the model's loss function includes discriminator loss, VAE loss, and generative adversarial loss. The VAE loss includes reconstruction loss and KL divergence loss.

[0069] The discriminator loss is calculated using the following formula:

[0070] In the formula: The discriminator outputs the true signal. The output of the discriminator to the generated signal, This is the binary cross-entropy loss.

[0071] The formula for calculating reconstruction loss in VAE loss is as follows:

[0072] In the formula: The original signal, For reconstructing the signal, n is the total number of samples.

[0073] The formula for calculating the KL divergence loss in VAE loss is as follows:

[0074] In the formula: Let be the mean of the potential vectors. Let be the standard deviation of the latent vector, and d be the dimension of the latent vector.

[0075] The formula for calculating the decoder adversarial loss is as follows:

[0076] In the formula: For binary cross-entropy loss, This is the output of the discriminator to the generated signal.

[0077] The total decoder loss is calculated using the following formula:

[0078] In the formula: This refers to the reconstruction loss in VAE loss. These are the weighting coefficients for the KL divergence loss. For the KL divergence loss in VAE loss, Weighting coefficients to counteract losses, To combat loss in the decoder.

[0079] The Unet-VAE-GAN triad architecture in this application integrates the advantages of multi-scale feature extraction, latent spatial constraints, and adversarial training. The encoder's U-Net structure preserves waveform details through skip connections, the decoder's dual-path design balances global reconstruction and local refinement, and the discriminator's adversarial mechanism forces the generated waveform to approximate the real data distribution. This collaborative design enables the model to generate high-fidelity waveforms while maintaining the physical plausibility of the output, solving the problem of artifacts easily generated by traditional generative models.

[0080] In some embodiments, the training steps of the Unet-VAE-GAN model specifically include: employing a phased optimization strategy to jointly train the Unet-VAE-GAN model using a multi-objective loss function; when training the discriminator, calculating the real signal discrimination loss and the generated signal discrimination loss, and backpropagating to update the discriminator parameters to optimize the discriminator so that it can distinguish between real blasting vibration signals and signals synthesized by the decoder; when jointly training the decoder and encoder, the encoder maps the real signal to the mean and variance parameters of the latent space, and obtains the latent vector by sampling through a random variable reparameterization method; the decoder fuses the latent vector and the jump connection features provided by the encoder to reconstruct the signal, and simultaneously calculates the reconstruction loss, KL divergence loss, and adversarial loss and performs a weighted summation as the total loss; and optimizes the decoder and encoder parameters simultaneously through backpropagation; the entire training cycle is repeated within a preset period, and the reconstructed signal is gradually optimized by dynamically balancing reconstruction accuracy, distribution regularization, and adversarial training intensity.

[0081] Specifically, the discriminator is first trained to distinguish between real signals and signals generated by the decoder. The discriminator loss for real signals and the discriminator loss for generated signals are calculated and backpropagated to update the discriminator parameters. Then, the decoder and encoder are trained. The encoder encodes the real signal into a latent spatial distribution (mean and variance), samples the latent vector using a reparameterization technique, and the decoder reconstructs the signal by combining the skip connection features provided by the encoder. Simultaneously, the reconstruction loss, KL divergence loss, and adversarial loss are calculated, and the weighted sum of these three losses is used as the total loss for backpropagation, simultaneously optimizing the decoder and encoder parameters. The entire training cycle is repeated within a specified period. By dynamically balancing reconstruction accuracy, distribution regularization, and adversarial training intensity, the model gradually improves the signal reconstruction quality.

[0082] This application employs a phased training strategy, achieving synergistic optimization of model parameters through dynamic balancing of reconstruction loss, KL divergence, and adversarial loss. Adversarial training of the discriminator enhances the realism of the generated waveforms. The use of VAEs to introduce latent spatial distribution constraints improves the diversity and stability of generated samples, and the adversarial training mechanism of GANs further enhances the realism and fidelity of the generated waveforms. Compared with the traditional time-domain linear superposition method, this method has significant advantages in simulation accuracy, adaptability, and visualization effects. It allows for real-time linkage between blasting parameter adjustment and waveform prediction through a user-friendly interface, lowering the barrier to entry for non-professionals. This joint optimization mechanism enables the model to learn the common laws of blasting vibrations while adapting to parameter variations in different engineering scenarios, forming an intelligent simulation system with both generalization ability and specificity.

[0083] In some embodiments, the training steps of the multilayer perceptron specifically include: using standardized blasting parameters from the database as data input and latent vectors as supervision labels to form training sample pairs; initializing the weights of the multilayer perceptron network and mapping the blasting parameters to predicted latent vectors through forward propagation; calculating the difference between the predicted latent vectors and the true latent vectors using a mean squared error loss function; and iteratively updating the weight parameters of the multilayer perceptron network using a backpropagation algorithm to minimize the mean squared error loss function, thereby establishing a nonlinear mapping relationship from blasting parameters to latent vectors.

[0084] Specifically, the standardized explosion parameters in the database are used as input data, and the latent vectors generated by the encoder of the trained Unet-VAE-GAN model for the corresponding waveform samples are used as supervision labels to form training sample pairs. Then, the weights of the MLP network are initialized, and the explosion parameters are mapped to predicted latent vectors through forward propagation. The difference between the predicted vectors and the true vectors is calculated using the mean squared error (MSE) loss function. Finally, the backpropagation algorithm is used to iteratively update the weight parameters of the MLP and minimize the loss function until the model can stably learn the nonlinear mapping relationship from the explosion parameters to the latent space.

[0085] Furthermore, the input data, output data, and data change process in the above model are as follows:

[0086] The input data consists of various blasting parameters (such as blast center distance, hole depth, and explosive consumption), which serve as feature variables. The model output is the multi-hole blasting vibration velocity waveform. The denoised real waveform is input into the encoder and progressively downsampled through three layers of convolution to extract multi-scale features, then outputs latent spatial parameters. The encoder achieves feature compression from waveform signal to high-dimensional features and then to latent spatial distribution parameters. The standardized blasting parameters (such as blast center distance and hole depth) are then input into a multilayer perceptron model. A fully connected network establishes a mapping relationship between the blasting parameters and the latent spatial distribution parameters. Inputting the blasting parameters into the multilayer perceptron model yields the latent spatial distribution parameters. These parameters are then input into the decoder, which fuses the multi-scale features to output a reconstructed time-domain waveform. Both the denoised real waveform and the decoder output waveform are input into a discriminator for discrimination and verification. The discriminator outputs a discrimination probability scalar, which quantifies the authenticity of the generated waveform. When the probability scalar output by the discriminator approaches 0, the decoder loss increases, forcing the decoder to improve its model.

[0087] This application transforms engineering experience into a computable mathematical expression through a parameter-vector mapping relationship established by a multilayer perceptron. By fitting high-dimensional nonlinear relationships using a deep neural network, it achieves an accurate conversion of blasting design parameters into potential waveform characteristics. This data-driven modeling approach overcomes the limitations of traditional empirical formulas, enabling waveform prediction under complex geological conditions or special charge structures to automatically adapt through parameter adjustments, significantly improving the method's engineering applicability.

[0088] Please refer to Figure 6 On the other hand, the present invention also provides a deep learning-based blasting vibration velocity waveform simulation system, comprising: a database construction module, used to collect blasting vibration velocity waveform data and corresponding blasting parameters from several different monitoring points multiple times and perform preprocessing to construct a database for model training; a Unet-VAE-GAN module, used to construct a Unet-VAE-GAN model, which learns the nonlinear mapping relationship between blasting parameters and blasting vibration velocity waveforms through training and generates latent vectors; a multilayer perceptron module, used to establish a multilayer perceptron model, which establishes a nonlinear mapping relationship between different blasting parameters and the latent vectors of the Unet-VAE-GAN model through training; and a simulation waveform generation module, used to input actual blasting parameters into the trained multilayer perceptron model to output the corresponding latent vectors, and input the output latent vectors into the trained Unet-VAE-GAN model to generate simulated blasting vibration velocity waveforms.

[0089] Specifically, six measuring points were set up around the blasting zone, each equipped with a three-axis blasting vibration tester (sampling frequency 10000 Hz). Waveform segments one second before and after the peak blasting vibration velocity were captured as initial data, with a uniform length of up to 20,000 sampling points. Key parameters such as blast center distance, maximum single-stage charge, and elevation difference were recorded simultaneously. This process was repeated multiple times to collect as much peak blasting vibration velocity waveform data and corresponding blasting parameter data as possible. The blasting parameter data included blast center distance, maximum single-stage charge, elevation difference, hole depth, hole inclination, row spacing, explosive consumption per unit area, charge per hole, plugging length, and number of holes.

[0090] Furthermore, a genetic algorithm was used to optimize the parameter combination of variational mode decomposition (VMD). The population size was set to 50, the number of iterations to 100, the crossover probability to 0.8, and the mutation probability to 0.01. The optimal parameter combination was obtained using envelope entropy as the fitness function. Then, these parameters were used to perform VMD decomposition on the monitored 3-axis initial waveform to generate multiple modal component IMFs. Next, a Gaussian kernel KPCA (kernel parameter σ=1.0) was applied to the modal component IMFs. After calculating the cumulative contribution rate, the top n modal component IMFs with a cumulative contribution rate exceeding 85% were retained and merged to reconstruct a denoised waveform.

[0091] Furthermore, a database was constructed, and the aforementioned steps were repeated to collect peak velocity waveform data of blasting vibrations from multiple blasts in the project. This data was then associated with blasting parameters and stored in an SQL database. Next, 80% of the data was divided into a training set and 20% into a test set, and the blasting parameters were subjected to Min-Max standardization.

[0092] Furthermore, the Unet-VAE-GAN model is trained. The encoder employs a three-layer convolutional structure, with each layer containing two convolutional layers using 5×5 kernels and 2-pixel padding. Each convolutional layer is followed by a ReLU activation function, then a max-pooling layer with a stride of 2 for downsampling. The first layer outputs a 32-channel feature map, the second layer outputs a 64-channel feature map, and the third layer outputs a 128-channel feature map. The encoding process ultimately outputs a 32-dimensional latent vector through a fully connected layer. The decoder contains two parallel paths. A skip-connected path receives feature maps from the encoder's three layers, with sizes of 1 / 2, 1 / 4, and 1 / 8 of the original length, respectively. This path upsamples the latent vector using transposed convolutions with a kernel size of 5, a stride of 2, padding of 2, and output padding of 1. The non-skipped-connected path upsamples the latent vector only. The outputs of the two paths are finally merged and processed through a convolutional layer and a Tanh activation function to generate a reconstructed signal. The discriminator also employs a three-layer convolutional structure with 32, 64, and 128 channels respectively. Each layer contains a convolutional layer using a 5×5 kernel with 2-pixel padding, followed by a LeakyReLU activation function with a negative slope of 0.2 and a max-pooling layer with a stride of 2. The discrimination process ultimately outputs the discrimination probability through a fully connected layer and a sigmoid activation function. In the loss function settings, the weight coefficient for KL divergence is 0.0001, and the weight coefficient for adversarial loss is 0.5. The model is trained using the Adam optimizer with a learning rate of 0.0002, a batch size of 64, and a total training duration of 500 epochs.

[0093] Furthermore, an MLP network with three hidden layers was constructed, using standardized blasting parameters as input and latent vectors generated by the encoder as labels. The mapping model was trained by minimizing the MSE loss. The blasting parameter data included blast center distance, maximum single-stage charge, elevation difference, hole depth, hole inclination, row spacing, explosive consumption per unit, charge per hole, plugging length, and number of holes.

[0094] Furthermore, the input test set explosion parameters are used to generate latent vectors through an MLP, which are then decoded to output simulated waveforms. Quantitative evaluation displays the waveform simulation effect, and the measured waveform, denoised waveform, and simulated waveform are displayed side-by-side through a visual interface, comprehensively verifying the reliability of the model in waveform morphology restoration and frequency domain reconstruction.

[0095] On the other hand, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described deep learning-based blasting vibration velocity waveform simulation method.

[0096] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described deep learning-based blasting vibration velocity waveform simulation method.

[0097] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0098] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, database, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0099] The above are merely embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention's specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A deep learning-based method for simulating blasting vibration velocity waveforms, characterized by the following steps: include: By collecting blasting vibration velocity waveform data and corresponding blasting parameters from several different monitoring points multiple times and preprocessing them, a database for model training is constructed. A Unet-VAE-GAN model is constructed, and through training, the model learns the nonlinear mapping relationship between blasting parameters and blasting vibration velocity waveform, and generates potential vectors. The Unet-VAE-GAN model includes an encoder, a decoder, and a discriminator; The encoder adopts a U-Net downsampling path structure, which includes three convolutional blocks for feature extraction and downsampling. Each convolutional block includes two convolutional layers and a ReLU activation function, followed by a max pooling layer for downsampling, and finally outputs the latent vector through a fully connected layer. The decoder has two parallel upsampling paths. One path connects the feature maps and latent vectors of each layer of the encoder by skipping connections, while the other path uses only the latent vectors for upsampling reconstruction. Finally, the reconstructed signal is output through convolution and Tanh activation function. The discriminator uses a convolutional neural network structure to determine authenticity. It extracts features through three convolutional blocks and finally outputs the true probability by a fully connected layer and a sigmoid activation layer. A multilayer perceptron model is established, and through training, a nonlinear mapping relationship is established between different blasting parameters and the latent vectors of the Unet-VAE-GAN model. The actual blasting parameters are input into the trained multilayer perceptron model to output the corresponding latent vector, and the output latent vector is input into the trained Unet-VAE-GAN model to generate the simulated blasting vibration velocity waveform.

2. The deep learning-based simulation method for blasting vibration velocity waveforms according to claim 1, characterized in that, The preprocessing steps specifically include: The blasting vibration velocity waveform data is aligned to obtain the initial blasting vibration velocity waveform; The parameter combination of variational mode decomposition is optimized by a genetic algorithm, and the initial blasting vibration velocity waveform is subjected to variational mode decomposition using the optimized parameters to obtain multiple modal components. The contribution rate of each modal component is calculated by kernel principal component analysis, and the first few modal components whose cumulative contribution rate exceeds the preset threshold are retained and merged into the noise-reduced blasting vibration velocity waveform.

3. The deep learning-based blasting vibration velocity waveform simulation method according to claim 2, characterized in that, The steps of optimizing the parameter combination of variational mode decomposition using a genetic algorithm specifically include: An initial population is randomly generated, where the individuals in the population are a combination of the number of modes K and the penalty factor α. The fitness value of each individual in the population is calculated using a fitness function, and the individual with the highest fitness is saved. High-quality individuals are selected through a roulette wheel selection mechanism, new individuals are generated by arithmetic crossover with random weights, and Gaussian random perturbation is introduced to implement mutation operations. Iteratively perform selection, crossover, and mutation operations until a preset termination condition is met, and output the optimal combination of the number of modes K and the penalty factor α.

4. The deep learning-based blasting vibration velocity waveform simulation method according to claim 2, characterized in that, The step of performing variational mode decomposition on the initial blasting vibration velocity waveform using optimized parameters to obtain multiple modal components specifically includes: Construct an objective function that includes constraints on minimizing modal component bandwidth and reconstructing the original signal; The objective function is solved iteratively using the alternating direction multiplier method, specifically including: Construct the Lagrange function by introducing Lagrange multipliers and a penalty term; When updating the modal components, keep the center frequency and Lagrange multipliers constant, and minimize the Lagrange function; When updating the center frequency, the modal components are kept constant, and the center frequency of each mode is updated by calculating the spectral moments. When updating the Lagrange multipliers, the gradient ascent method is used. Alternately update the modal components, center frequency, and Lagrange multipliers until the preset convergence condition is met, and output multiple modal components with specific center frequencies.

5. The deep learning-based blasting vibration velocity waveform simulation method according to claim 1, characterized in that, The training steps of the Unet-VAE-GAN model specifically include: A phased optimization strategy is adopted, and the Unet-VAE-GAN model is jointly trained using a multi-objective loss function; During the training of the discriminator, the discrimination loss of the real signal and the discrimination loss of the generated signal are calculated, and the discriminator parameters are updated by backpropagation to optimize the discriminator so that it can distinguish between the real blasting vibration signal and the signal synthesized by the decoder. When jointly training the decoder and encoder, the encoder maps the real signal to the mean and variance parameters of the latent space and obtains the latent vector by sampling through the random variable reparameterization method. The decoder fuses the latent vector with the jump connection features provided by the encoder to reconstruct the signal. At the same time, it calculates the reconstruction loss, KL divergence loss and adversarial loss and performs a weighted sum as the total loss. The parameters of the decoder and encoder are optimized simultaneously through backpropagation. The entire training cycle is repeated within a preset period, and the reconstructed signal is gradually optimized by dynamically balancing reconstruction accuracy, distribution regularization, and adversarial training intensity.

6. The deep learning-based blasting vibration velocity waveform simulation method according to claim 1, characterized in that, The training steps of the multilayer perceptron specifically include: The standardized blasting parameters in the database are used as data input, and the latent vectors are used as supervision labels to form training sample pairs; Initialize the weights of the multilayer perceptron network and map the blasting parameters to the predicted latent vectors through forward propagation; The difference between the predicted latent vector and the true latent vector is calculated using the mean squared error loss function; The weight parameters of the multilayer perceptron network are iteratively updated using the backpropagation algorithm to minimize the mean squared error loss function, thereby establishing a nonlinear mapping relationship from the blasting parameters to the potential vector.

7. A deep learning-based simulation system for blasting vibration velocity waveforms, characterized in that, include: The database construction module is used to collect blasting vibration velocity waveform data and corresponding blasting parameters from several different monitoring points multiple times, and perform preprocessing to build a database for model training. The Unet-VAE-GAN module is used to build the Unet-VAE-GAN model. Through training, the model learns the nonlinear mapping relationship between blasting parameters and blasting vibration velocity waveforms and generates latent vectors. The Unet-VAE-GAN model includes an encoder, a decoder, and a discriminator; The encoder adopts a U-Net downsampling path structure, which includes three convolutional blocks for feature extraction and downsampling. Each convolutional block includes two convolutional layers and a ReLU activation function, followed by a max pooling layer for downsampling, and finally outputs the latent vector through a fully connected layer. The decoder has two parallel upsampling paths. One path connects the feature maps and latent vectors of each layer of the encoder by skipping connections, while the other path uses only the latent vectors for upsampling reconstruction. Finally, the reconstructed signal is output through convolution and Tanh activation function. The discriminator uses a convolutional neural network structure to determine authenticity. It extracts features through three convolutional blocks and finally outputs the true probability by a fully connected layer and a sigmoid activation layer. The multilayer perceptron module is used to build a multilayer perceptron model and, through training, establish a nonlinear mapping relationship between different blasting parameters and the latent vectors of the Unet-VAE-GAN model. The simulation waveform generation module is used to input the actual blasting parameters into the trained multilayer perceptron model to output the corresponding latent vector, and then input the output latent vector into the trained Unet-VAE-GAN model to generate the simulated blasting vibration velocity waveform.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based blasting vibration velocity waveform simulation method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based blasting vibration velocity waveform simulation method according to any one of claims 1 to 6.

Images