Multi-objective collaborative optimization parameter setting method for blasting vibration and dust control

Abstract

The application discloses a multi-target cooperative optimization parameter setting method for blasting vibration and dust control, which comprises the following steps: based on training samples containing geological conditions, blasting parameters and actual blasting effects, a neural network model capable of predicting complete vibration time curve and three-dimensional dust concentration field is trained; by using the model, an evolution algorithm is used to perform multi-target optimization on a target function containing four indexes of vibration, dust, cost and regulation compliance; a Pareto optimal solution set composed of multiple non-dominant parameter schemes is obtained; a preset explanation algorithm is used to analyze the schemes in the optimal solution set, and the influence and contribution of each blasting parameter on the prediction result are quantified, so that the "black box" model is converted into an interpretable "white box" decision; the application solves the problems that the traditional method is difficult to balance multiple conflicting targets and the decision-making process is not transparent, and provides an optimal blasting parameter scheme for blasting operation, which takes into account safety, environmental protection and cost efficiency.

Description

A Multi-Objective Collaborative Optimization Parameter Tuning Method for Blasting Vibration and Dust Control Technical Field

[0001] This invention relates to the field of intelligent strategy optimization technology in blasting engineering, and in particular to a multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control. Background Technology

[0002] Blasting is widely used as a highly efficient rock breaking technology in geotechnical engineering projects such as tunnel excavation, mining, and infrastructure construction. However, vibration waves and dust pollution generated by blasting operations are unavoidable byproducts, posing potential threats to the structural safety of surrounding buildings, the stable operation of precision instruments, the ecological environment, and human health. Therefore, the precise design and optimization of blasting parameters to synergistically control vibration and dust effects has long been a core technical challenge in this field. To address this challenge, related control technologies have evolved from empirical formulas and numerical simulations to data-driven models. Initially, vibration prediction relied mainly on empirical regression models such as the Sadovsky formula. While these methods are simple, their applicability is limited and they cannot reflect complex propagation mechanisms. With the rise of numerical simulation methods such as the finite element method and the discrete element method, these methods can further reproduce the blasting process from a mechanical perspective. However, they often suffer from inherent drawbacks such as complex modeling and high computational costs, making it difficult to meet the needs of rapid decision-making in engineering projects.

[0003] In recent years, with the development of artificial intelligence technology, data-driven models such as support vector machines and traditional neural networks have been introduced to establish nonlinear mapping relationships between blasting parameters and effect indicators, and to optimize parameters using intelligent methods such as genetic algorithms. Although existing technologies have made some progress, significant limitations still exist. First, most data-driven models simplify the prediction target to a single scalar indicator such as peak particle velocity (PPV) and average dust concentration, which is essentially a dimensionality reduction fit of complex physical processes. This loses the time-domain and frequency-domain characteristics of the vibration waveform and the spatiotemporal distribution details of dust cloud diffusion, resulting in insufficient model fidelity and difficulty in supporting refined control. Second, existing optimization methods often treat vibration and dust control as independent single-objective problems for decoupling or perform simple weighted combinations, thus ignoring the complex nonlinear coupling and antagonistic relationship between the two, making it difficult to obtain a globally optimal cooperative control scheme. More importantly, most current intelligent optimization methods based on deep learning and evolutionary algorithms are essentially "black box" processes. The optimal parameter combinations they provide lack clear physical meaning and decision-making logic explanations, making it difficult to gain the trust of on-site engineers. This greatly limits the practical adoption and application of such intelligent technologies in blasting engineering. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control, to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control, comprising:

[0007] Based on training samples containing geological condition data, blasting parameter data, and corresponding actual blasting effect data, a neural network model is trained to establish a mapping relationship from the geological condition data and blasting parameter data to the predicted blasting effect data.

[0008] Using the neural network model, a preset evolutionary algorithm is used to optimize an objective function containing four types of indicators: vibration, dust, cost, and regulatory compliance, with blasting parameters as variables, until a Pareto optimal solution set consisting of multiple non-dominated parameter schemes is obtained.

[0009] For at least one non-dominated parameter scheme in the Pareto optimal solution set, a preset interpretation algorithm is used to analyze the prediction process of the neural network model to determine the quantitative impact of each blasting parameter in the scheme on the predicted blasting effect data.

[0010] As a preferred embodiment of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control described in this invention, the construction steps of the neural network model specifically include:

[0011] First and second encoder networks are used to extract geological condition feature vectors and blasting parameter feature vectors from the geological condition data and blasting parameter data, respectively.

[0012] The encoder network is trained using a preset loss function, so that the distance between the geological condition feature vector and the blasting parameter feature vector originating from the same blasting event in the latent space converges.

[0013] As a preferred embodiment of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control described in this invention, the construction step further includes:

[0014] The geological condition feature vector and blasting parameter feature vector obtained after distance convergence training are used as conditional inputs to train the neural network model.

[0015] The neural network model learns the process of gradually denoising random noise to generate target data, thereby realizing the generation operation from the conditional input to the predicted explosion effect data.

[0016] As a preferred embodiment of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control described in this invention, the method of introducing the conditional input into the model during training of the neural network model includes:

[0017] An attention calculation module is set in the network structure of the neural network model. The module receives the intermediate layer features of the model as queries and the condition inputs as keys and values. It calculates the correlation between the query and the key to weight the value and applies the weighting result to the intermediate layer features.

[0018] As a preferred embodiment of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control described in this invention, the predicted blasting effect data includes:

[0019] Complete velocity-time history curve data representing the vibration process;

[0020] Three-dimensional concentration field data representing the spatial distribution of dust.

[0021] As a preferred embodiment of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control described in this invention, the four types of indicators in the objective function are quantified in the following manner:

[0022] The vibration index is calculated from the vibration intensity or frequency domain characteristics in the predicted blasting effect data;

[0023] The dust index is calculated from the dust concentration or the volume exceeding the standard in the predicted blasting effect data.

[0024] The cost index is calculated based on the resource consumption corresponding to the blasting parameters;

[0025] The regulatory compliance indicators are obtained by comparing the vibration and dust indicators with regulatory limits.

[0026] As a preferred embodiment of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control described in this invention, the specific process of the evolution algorithm includes:

[0027] Initialize a population consisting of multiple candidate blasting parameter schemes;

[0028] In each iteration of the algorithm, for each scheme in the population, the neural network model is invoked to calculate its corresponding target indicators;

[0029] Based on the target index, selection, crossover, and mutation operations are performed on the population through non-dominated sorting and crowding calculation to generate a new generation of population.

[0030] As a preferred embodiment of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control described in this invention, the regulatory compliance index is calculated using a penalty function:

[0031] When the calculated value of the vibration or dust index does not exceed the regulatory limit, the output value of the penalty function is less than a preset first threshold.

[0032] When the calculated value of the indicator exceeds the regulatory limit, the output value of the penalty function increases as the excess increases.

[0033] As a preferred embodiment of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control described in this invention, the interpretation algorithm used is a method for quantifying the contribution of each input feature to the model output. This method is used to calculate the specific contribution value of each parameter in the parameter scheme to the predicted blasting effect data.

[0034] As a preferred embodiment of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control described in this invention, the interpretation algorithm used is a gradient calculation method, which determines the sensitivity of each blasting parameter by calculating the gradient of the predicted blasting effect data with respect to the input blasting parameters.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1. By constructing a neural network model that can directly generate complete vibration time history and three-dimensional dust concentration field, this invention overcomes the limitation of traditional methods that only predict a single scalar index and achieves high-fidelity reproduction of the physical process of explosion effect.

[0037] 2. By placing the four interdependent objectives of vibration, dust, cost, and compliance under a unified multi-objective optimization framework, and using the NSGA-II algorithm to find the Pareto optimal solution set, the one-sidedness of single-objective optimization or simple weighting method is avoided, and a collaborative control scheme for blasting parameters that achieves the best trade-off among multiple conflicting objectives can be obtained.

[0038] 3. In addition, this invention introduces interpretability algorithms such as SHAP and gradient sensitivity analysis, transforming the "black box" optimization model into a "white box" decision-making process, which enhances engineers' confidence in the intelligent optimization results. At the same time, it can minimize the safety risks of blasting operations to the environment and surrounding facilities while ensuring the economic efficiency of the project. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0040] Figure 1 is a flowchart of the multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control according to an embodiment of the present invention. Detailed Implementation

[0041] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0043] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0044] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0045] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0046] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0047] Example 1

[0048] Referring to Figure 1, which illustrates the first embodiment of the present invention, this embodiment provides a multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control, comprising:

[0049] S1. Based on training samples containing geological condition data, blasting parameter data, and corresponding actual blasting effect data, train a neural network model to establish a mapping relationship from geological condition data and blasting parameter data to predicted blasting effect data.

[0050] It should be noted that the goal of this step is to build a high-fidelity neural network model that can accurately learn and reproduce the blasting effect produced by the combined action of blasting parameters under specific geological conditions.

[0051] Preferably, the neural network model is a surrogate model because real blasting tests are costly, risky, and time-consuming. Therefore, we train this surrogate model to replace real tests. The advantage of doing so is that the model can be quickly invoked at a very low cost to predict the results of any set of blasting parameters.

[0052] Specifically, since the neural network model is composed of neural network modules, the proxy model in this invention is composed of a combination of various different types of neural network modules, including but not limited to:

[0053] Convolutional neural networks (CNNs) are used to process geological images.

[0054] Multilayer perceptron (MLP) for processing quantitative geological indicators.

[0055] Transformer encoder used to capture complex relationships between blasting parameters;

[0056] And, as the core generator, the conditional diffusion model based on the U-Net architecture;

[0057] Furthermore, a comprehensive database containing hundreds (e.g., 500) on-site blasting events is constructed as the dataset for this proxy model;

[0058] Specifically, the dataset includes geological condition data, blasting parameter data, and corresponding actual blasting effect data. The data collection methods for each dataset are as follows:

[0059] For geological condition data, a set of structured data describing the macroscopic mechanical properties of the surrounding rock for each blast is obtained through laboratory rock sample tests and in-situ field tests, using quantitative geological indicators. This data constitutes a... Dimensional geological index vector Its vector representation can include: rock saturated uniaxial compressive strength (MPa), elastic modulus E (GPa), longitudinal wave velocity (m / s), rock mass integrity factor The basic quality (BQ) grading of the rock mass is also included. Furthermore, before each blasting operation, an industrial-grade high-resolution camera (e.g., with a resolution of at least 4096×4096 pixels) is used to photograph the tunnel excavation face directly, acquiring digital images that clearly reflect the distribution, density, and weathering degree of the rock mass structural surfaces (such as joints, fissures, and bedding). ;

[0060] For blasting parameter data, all controllable parameters of each blasting design are recorded in detail and organized into a... Dimensional blasting parameter vector This vector includes not only the total charge Q (kg) and the maximum single-shot charge. In addition to scalar values ​​such as (kg) and borehole diameter d (mm), it also includes a delay time series representing the timing structure of the detonation network. For example, for a design containing K delay segments, this delay time series can be expressed as: ,in, It is the millisecond delay time of the k-th detonator;

[0061] For actual blasting effect data, it is divided into full-time vibration velocity data and three-dimensional dust concentration field. For full-time vibration velocity data, at key protected objects at different locations from the blast source center, three-component (radial R, tangential T, and vertical V) velocity sensors are used to simultaneously record the entire process of the blasting event's vibration velocity time history curve at a sampling rate of no less than 5kHz. The data obtained from the three-component velocity sensors are then stitched together to form a... matrix L represents the number of sampling points (e.g., L=4096). For a three-dimensional dust concentration field, within a fixed post-explosion time window (e.g., 30 to 45 seconds post-explosion), a three-dimensional scanning lidar (LiDAR) is used to rapidly scan the space in front of the tunnel face to obtain a three-dimensional point cloud of dust aerosol clouds. The sparse point cloud data is then reconstructed into a regular point cloud using methods such as kriging interpolation or voxel density estimation. (For example, 64×64×64) The value of each voxel represents the dust mass concentration (mg / m³) at that spatial location. 3 );

[0062] Specifically, for all structured numerical vectors ( and The model employs Min-MaxNormalization to linearly scale the feature values ​​in each numerical vector to the range of [0, 1] or [-1, 1] to eliminate dimensional differences and stabilize model training.

[0063] It should be noted that, in order to consider the quality of the collected dataset and the model performance, it is usually necessary to augment some of the collected data.

[0064] Specifically, for digital images The size of the model was uniformly adjusted to 256×256 pixels, and data augmentation was performed, including random horizontal flipping, small angle rotation (-10 to +10°), and small random perturbations of contrast and brightness, to improve the robustness of the model.

[0065] Specifically, for the vibration velocity time history matrix A fourth-order Butterworth bandpass filter is applied to filter out high-frequency noise and low-frequency trend drift that are irrelevant to the focus of the blasting project (e.g., retaining the 1-100Hz frequency band).

[0066] Specifically, for the three-dimensional dust concentration field Similar to the normalization operation of the numerical vectors mentioned above, its value range is scaled to [0, 1] or [-1, 1].

[0067] Furthermore, in order to enable the model to understand the deep semantic relationship between geological condition data and blasting parameter data, this invention designs a multi-stream encoder architecture and forces the architecture to perform semantic alignment in a shared latent space through a contrastive learning loss function.

[0068] Furthermore, the multi-stream encoder architecture consists of a geological condition encoder and a blasting parameter encoder, and the encoders operate in parallel mode.

[0069] Furthermore, the geological condition encoder consists of an image branch, an index branch, and a feature fusion module, and the encoder has a sequential architecture.

[0070] Specifically, this image branch employs a convolutional neural network (CNN) pre-trained on a large image dataset (such as ImageNet), using ResNet-50 as its backbone network, and taking pre-processed digital images as input. The image is processed sequentially through multiple convolutional layers, ReLU activation layers, batch normalization layers, and residual connections of a ResNet-50 network. It's important to note that the original fully connected layers used for classification are not used in this process. After the last convolutional block of the network, a high-dimensional feature map is obtained. Global average pooling is then applied to this feature map to compress it spatially, ultimately resulting in a fixed-length image feature vector. ;

[0071] It should be noted that pre-training is used here to leverage the advantages of transfer learning, transferring the powerful visual feature extraction capabilities learned by convolutional neural networks in general image recognition tasks to the specific task of recognizing geological condition features, thereby achieving better performance and faster convergence speed on a limited blasting engineering dataset.

[0072] Specifically, the indicator branch uses a multilayer perceptron (MLP) consisting of three fully connected layers, with the following network structure: After each fully connected layer, a Rectified Linear Unit (ReLU) is used as the activation function to introduce nonlinearity. Simultaneously, Layer Normalization is inserted between layers to accelerate training convergence and improve model stability. The normalized geological index vectors are then used to... By inputting this MLP, the indicator feature vector can be obtained. The dimension is 512;

[0073] Specifically, the feature fusion module combines image feature vectors and indicator eigenvectors By concatenating the two feature vectors, a high-dimensional fused feature vector is obtained. Furthermore, to enable the model to learn the optimal fusion method between the two feature vectors, the fused feature vector is mapped to a shared latent space with the same dimension as the output of the blasting parameter encoder. That is, a linear projection layer is added after the concatenation operation to linearly transform and reduce the dimensionality of the concatenated high-dimensional vector to the target dimension d (e.g., d=512), thus obtaining the geological condition feature vector. ;

[0074] Specifically, for the geological condition encoder, in one feasible implementation of the present invention, the initial learning rate of the Adam optimizer is 1e-4, the weight decay coefficient is 0.01, the batch size is 32, and the training period is 100 periods.

[0075] It should be noted that, due to the complex nonlinear interactions between blasting parameters (for example, the influence of the maximum single explosive charge in blasting engineering changes drastically with the change of the initiation delay time), this invention preferably uses a Transformer encoder model as the basic architecture of the blasting parameter encoder.

[0076] Furthermore, the Transformer encoder model consists of an embedding layer, a positional coding layer, a multi-layer Transformer coding block stacking module, and a global average pooling layer;

[0077] Specifically, since the Transformer model requires a sequence of vectors as input, the explosion parameter vectors are first processed beforehand. Convert it into an embedding matrix, and in the embedding layer, The blasting parameters of dimension are considered as a length of Given a sequence where each element is a scalar value, a learnable linear projection layer (i.e., a fully connected layer without an activation function) is set up, and this sequence is transformed into a linear projection layer of length 1. A vector of dimension is mapped to a single dimension. The embedding matrix is ​​equivalent to creating an embedding matrix for each blasting parameter. The initial eigenvectors of dimension;

[0078] Specifically, since the standard Transformer does not contain any information about the order or position of the inputs, but in the task of this invention, the position of the parameters is meaningful (e.g., the order of the delay time series). In order for the model to distinguish the relationship between delay times (e.g., distinguishing the delay times of the first and fifth segments), position information is added to the embedded parameter features. In the position encoding layer, a fixed sine-cosine position encoding method is used to generate a position matrix with the same dimension as the embedding matrix, and this matrix is ​​added to the embedding matrix. In this way, the representation vector of each parameter contains both its numerical information and position information.

[0079] Specifically, the multi-layer Transformer encoding block stacking module consists of L (e.g., L=6) identical encoding blocks stacked together. Each encoding block contains two sub-layers: a multi-head self-attention layer and a feedforward neural network. In the multi-head self-attention layer, for each parameter in the input sequence, the self-attention mechanism dynamically calculates the association weights of that parameter with all other parameters. The multi-head (e.g., h=8 heads) mechanism also allows the model to learn different types of dependencies in parallel from different representation subspaces (e.g., one head focuses on the relationship between explosive charge and aperture, while another head focuses on the interference relationship between delay times). After the multi-head self-attention layer, the output vector at each position is independently passed through a feedforward network (structured as two linear layers and a ReLU activation function) to further non-linearly transform the features, enhancing the model's expressive power. Furthermore, residual connections and layer normalization are used around each sub-layer to effectively prevent gradient vanishing.

[0080] Specifically, after stacking L coded blocks, an output matrix of the same length as the input sequence is obtained. To obtain a single feature vector representing the entire brute-force scheme, we calculate the dimension of the output matrix in the sequence length using a global average pooling layer. The average value over (dimensionality) compresses the aforementioned embedding matrix into a single... A vector of dimension, obtained through pooling. The dimension vector is the feature vector of the final explosive parameters. ;

[0081] Specifically, for the blasting parameter encoder, in one feasible process of the present invention, the Adam optimizer is also used, with the learning rate set to 1e-4; the batch size is 16; the noise scheduling table adopts a linear noise scheduling strategy, with an initial time step of 1 and a total time step of 1000. The initial time step increases linearly with a learning rate of 1e-4 until the total time step of 1000 is 0.02.

[0082] Furthermore, the two encoders mentioned above are jointly trained using the Adam optimizer, with the objective set to minimize the loss function. For a training batch of size B, the specific form of the loss function is:

[0083] ;

[0084] in, These are positive sample pairs from the same explosion event. when At that time, they were negative sample pairs from different blasting events; The cosine similarity function; This is a temperature hyperparameter, and its range is... (For example, ), used to adjust the degree of attention the loss function pays to difficult negative samples;

[0085] It should be noted that by minimizing this loss function using the Adam optimizer, the model can be driven to passively learn the mapping relationship, ensuring that the matched geological conditions-blasting parameter pairs are accurate. In the latent space of dimension, the representation vectors are brought close to each other, while those that do not match are pushed away from each other, thus completing the feature extraction operation from the original data to semantic alignment.

[0086] Furthermore, after obtaining the semantically aligned features, this invention employs Denoising Diffusion Probabilistic Models (DDPM) to learn the generation mapping of the blast effect data;

[0087] Specifically, the core of this model is a noise prediction network based on the U-Net architecture. The network contains a symmetric encoder-decoder structure with multiple resolution levels. The encoder extracts features through convolution and downsampling operations, while the decoder gradually restores the original resolution of the data through upsampling and convolution operations. Corresponding levels between the encoder and decoder are connected by skip connections to preserve multi-scale information.

[0088] It should be noted that, in order to handle the multiple outputs in the present invention (mainly for full-time vibration velocity data and three-dimensional dust concentration field), the input and output channels of this U-Net are designed to accommodate both data structures simultaneously. For example, the full-time vibration velocity data can be regarded as a 1×L "image," and... The dust field is spliced ​​together in a certain dimension or used as different channels for input;

[0089] Specifically, the information of time steps in the model is converted into a vector through sinusoidal position encoding and injected into each residual block of U-Net to inform the prediction network of its current denoising progress.

[0090] Furthermore, the semantically aligned geological feature vectors and blasting parameter eigenvectors Concatenate the vectors to form a unified condition vector. ;

[0091] Furthermore, to ensure that the conditional vectors accurately guide the generation process, this invention integrates a cross-attention module into each resolution level of the U-Net's Transformer coding block (or convolutional block). The specific operation of this module is as follows:

[0092] In each downsampled or upsampled block of U-Net, the intermediate layer features processed in that block are used as the query (Q).

[0093] The external input condition vector is transformed linearly to obtain the key (K) and value (V).

[0094] The attention weight A is obtained by calculating the scaled dot product similarity of Q and K and applying the Softmax function: T represents the transpose operation. The dimension of the key vector;

[0095] The attention weight is applied to V to obtain the weighted context information, which is then added back to the intermediate layer features of U-Net.

[0096] It should be noted that by introducing the cross-attention module, the model can adaptively focus on the most relevant part of the conditional input (geological or blasting parameters) at each step of generating blasting effect data (e.g., a segment of the vibration time history curve or a certain area of ​​the dust field), thereby greatly improving the accuracy and detail fidelity of the generated data.

[0097] Furthermore, for this neural network model, the training objective is to minimize the mean squared error (MSE) loss between the predicted noise and the actual added noise:

[0098] ;

[0099] in, It is real data on the effects of explosions. and (combination) It is the corresponding condition vector. It is the time step of random sampling. It is standard Gaussian noise. Represented as noise, It is the identity matrix. These are the coefficients in the preset noise scheduling table. This is the mean square error loss value;

[0100] Specifically, for this loss, the Adam optimizer is used for iterative training with a large batch size of tens of thousands of steps until the mean squared error loss value converges.

[0101] Specifically, in one feasible process of the present invention, the training set and the validation set are divided in an 8:2 ratio. When the mean squared error (MSE) loss of the model on the validation set no longer decreases significantly for 10 consecutive cycles (e.g., the decrease is less than 1e-5), the model is considered to have converged.

[0102] S2. Using a neural network model and a preset evolutionary algorithm, the objective function containing four types of indicators—vibration, dust, cost, and compliance—is optimized with blasting parameters as variables until a Pareto optimal solution set consisting of multiple non-dominated parameter schemes is obtained.

[0103] It should be noted that the goal of this step is to use the surrogate model trained in S1 as a virtual test platform and systematically explore the entire design space of blasting parameters through a multi-objective evolution algorithm. The aim is to find a set of blasting parameter schemes that can achieve the best trade-off among four conflicting objectives: vibration control, dust suppression, economic cost, and regulatory compliance, so as to provide engineering decision-makers with diverse and selectable optimal solutions.

[0104] Furthermore, before starting the optimization algorithm, the engineering problem must first be rigorously transformed into a mathematical optimization model;

[0105] Specifically, the optimization variables in this mathematical optimization model are blasting parameters that engineers can directly control; these parameters together constitute a... Dimensional decision vector :

[0106] ;

[0107] in, Representing the Each blasting parameter, such as total charge, maximum single-shot charge, borehole diameter, and delay time sequence, is taken within a preset upper and lower limit that conforms to engineering realities, thus satisfying the constraints. , This is the lower limit of the range of the blasting parameters themselves. This represents the upper limit of the range of the blasting parameters themselves.

[0108] Furthermore, a vector is constructed that includes four objective functions to be minimized: vibration control, dust suppression, economic cost, and regulatory compliance. :

[0109] ;

[0110] Specifically, the quantization methods for these four objective functions are as follows:

[0111] Vibration index This index is used to comprehensively evaluate the intensity and spectral characteristics of blasting vibrations. For any given blasting parameter scheme, the vibration velocity time history matrix in the neural network model is first called. Subsequently, based on Calculating this index yields:

[0112] ;

[0113] in, It is the peak particle velocity extracted from the vibration velocity time history curve, representing the vibration intensity; The dominant frequency is obtained by performing a Fast Fourier Transform (FFT) on the time history curve. The reciprocal of the dominant frequency is mainly used to evaluate the risk of low-frequency components, because low-frequency vibrations are more destructive to buildings. and These are preset weighting coefficients used to balance the importance of intensity and frequency;

[0114] Dust index This indicator quantifies the degree of dust pollution generated by blasting. Similarly, it utilizes the three-dimensional dust concentration field in a neural network model. ,get:

[0115] ;

[0116] in, This represents the volume where the dust concentration exceeds the standard. It is a collection concentration field of individual elements It is the first Dust concentration values ​​of individual elements, It is the dust concentration limit required by regulations or environmental protection requirements (e.g., 10 mg / m³). 3 ); It is an indicator function that takes the value 1 when the condition is true and 0 otherwise. It is the actual volume represented by a single voxel;

[0117] Cost indicators This indicator is calculated based on the resource consumption corresponding to the blasting parameter scheme. A typical linear cost model is as follows:

[0118] ;

[0119] in, It refers to the total charge amount in the blasting parameter scheme. This is the total number of detonators. It is the total drilling length. These are the unit costs of explosives, detonators, and drilling, respectively.

[0120] Compliance indicators This indicator is a penalty function used to measure whether the blasting parameter scheme violates mandatory safety regulations. Its calculation formula is as follows:

[0121] ;

[0122] in, and These are the mandatory safety limits for vibration velocity and dust concentration stipulated by regulations; and These represent the peak vibration velocity and peak dust concentration extracted from the model, respectively. Ensure that penalties are only incurred when limits are exceeded; and A relatively large penalty coefficient is applied; only when the blasting parameters are fully compliant... Once the limit is exceeded, the function value will increase rapidly as the amount of excess increases, thus strongly rejecting non-compliant solutions during the optimization process;

[0123] It should be noted that this invention preferably uses the Non-dominated Sorting Genetic Algorithm II (NSGA-II) with an elitist strategy as the preset evolutionary algorithm, because this algorithm can effectively converge to the true Pareto front and maintain good solution diversity when dealing with multi-objective optimization problems. Its specific process is as follows:

[0124] S201, Population Initialization: Randomly generate a population containing... Individuals (e.g., The initial population Each individual is a candidate blasting parameter scheme that satisfies the boundary constraints of the decision variables;

[0125] S202, Iterative Optimization Loop (the first...) (Generation): The algorithm enters the main loop until the termination condition is met. The operations for each generation are as follows:

[0126] (i) Fitness assessment: for the current population Each individual in The pre-trained neural network model in S1 is invoked to predict the corresponding explosion effect data. and Then, according to the formulas defined above, the values ​​of the four objective functions are calculated respectively. ;

[0127] (ii) Generate offspring population :

[0128] Selection: A binary tournament selection method will be used, from... Two individuals are randomly selected for comparison, based on non-dominated ranking and crowding distance: the individual with the better non-dominated ranking (i.e., smaller ranking value) is preferred; if the rankings are the same, the individual with the larger crowding distance (i.e., sparser surrounding solutions) is selected, and this process is repeated. This forms a parent pool for reproduction;

[0129] Crossover: Select pairs of individuals from the parent pool with a preset crossover probability. ( Perform crossover operations to generate new offspring individuals. For continuous blasting parameters (such as explosive quantity and delay time), simulated binary crossover (SBX) is preferred, with its internal distribution index set to 20 to directly control the exploration range of the offspring solution.

[0130] Mutation: For the offspring produced after crossover, a preset mutation probability is applied. When performing mutation operations, for continuous parameters, polynomial mutations are preferred. The distribution index is set to the same as that of the simulated binary crossover, which is 20, to make small perturbations around the existing value of the parameter in order to explore the neighboring solution space.

[0131] Through selection, crossover, and mutation, a subset of the same size is generated. offspring population ;

[0132] (iii) Elite preservation and the generation of new populations:

[0133] parental population and offspring population The populations are merged to form a temporary population of size 2N. ;

[0134] right Perform a fast nondominated sort on all 2N individuals, stratifying them into different nondominated fronts. , This is the current optimal solution set;

[0135] These individuals from the frontier will be sequentially added to the new generation population. From Beginning, followed by And so on, until... The scale reached ;

[0136] If filled to a certain leading edge At that time, it was impossible to All individuals were placed Then for The crowding distance is calculated for each individual, and individuals with larger distances are prioritized until... The size is exactly ;

[0137] S203. Termination Conditions and Result Output:

[0138] Set a maximum number of iterations (e.g., The loop terminates when the algorithm reaches the maximum number of generations, at which point the final generation of the population... In the middle, it is located at the first non-dominant frontier. All the individuals together constitute the Pareto optimal solution set sought in this invention. Each blasting parameter scheme in this solution set is a non-dominated solution, which means that there is no other feasible scheme that is better than it at least in one objective and not inferior to it in any other objective.

[0139] S3. For at least one non-dominated parameter scheme in the Pareto optimal solution set, a preset interpretation algorithm is used to analyze the prediction process of the neural network model in order to determine the quantitative impact of each blasting parameter in the scheme on the predicted blasting effect data.

[0140] It should be noted that, in order to solve the "black box" problem of existing intelligent optimization methods, this invention introduces Explainable Artificial Intelligence (XAI) algorithm, which aims to analyze the internal decision-making logic of the constructed neural network model, transforming the optimization results from numerical values ​​into decision-making basis that engineers can understand, trust, and fine-tune accordingly; after the engineer selects a non-dominated parameter scheme that meets the current engineering preferences (e.g., the lowest cost scheme under vibration and dust requirements) from the Pareto optimal solution set, it is analyzed through methods including but not limited to the following two methods;

[0141] Furthermore, in order to fully understand the contribution of each blasting parameter in the scheme to the final model prediction result, the present invention preferably adopts a contribution quantification method based on game theory, specifically the SHAP (SHapley Additive ex Planations) algorithm. The core idea of ​​this algorithm is to compare the prediction process of the neural network model to a cooperative game, in which each blasting parameter is a "participant", and the final prediction result of the model (e.g., the predicted peak particle velocity PPV) is the total "payout" generated by the cooperation of all "participants". The SHAP value, based on strict mathematical axioms, fairly distributes this total "payout" to each "participant", thereby quantifying the specific contribution value of each parameter to the prediction result.

[0142] Specifically, for a given nondominated solution and its corresponding model prediction results, the first Explosion parameters SHAP value It is then defined as:

[0143] ;

[0144] in, This represents the completed training of the neural network model. It is all The complete set of explosion parameters, It does not include parameters Any subset of parameters, Is the model in a known subset? and parameters The prediction was made when the value was given. The model is only available on a known subset. The prediction was made when the value was given. Indicates parameters In the known subset Marginal contribution based on A weighting coefficient used for all possible subsets of parameters. The marginal contributions are weighted and averaged.

[0145] Furthermore, since calculating the marginal contribution of all parameter subsets in a neural network is an NP-hard problem, in practical implementations, approximate estimation algorithms such as KernelSHAP or DeepSHAP can be used. For example, when using KernelSHAP, the input features are sampled and perturbed multiple times, and an interpretable linear model is used to fit the output of the neural network model corresponding to these perturbed samples, thereby obtaining the SHAP approximate values ​​of each parameter. This approach mainly ensures the interpretability of the model while greatly reducing the computational complexity, making this scheme feasible in blasting engineering.

[0146] Furthermore, through calculation The SHAP value of each parameter can generate an intuitive mechanical interpretation diagram, which includes the baseline value and the contributing force.

[0147] Specifically, the baseline value represents the average predicted value of the model when there is no blasting parameter information;

[0148] Specifically, the contribution of each blasting parameter is represented by a colored arrow; a red arrow indicates that the current value of the parameter increases the final prediction result (for example, the maximum single-explosive charge results in a positive SHAP value for PPV, represented by a red arrow that increases PPV); a blue arrow indicates that the current value of the parameter decreases the final prediction result (for example, an optimized delay time series may produce waveform interference cancellation effects, resulting in a negative SHAP value for PPV, represented by a blue arrow that decreases PPV). It is important to note that the length of the arrow is proportional to the absolute magnitude of the parameter's SHAP value and can intuitively represent the strength of the parameter's influence.

[0149] Furthermore, by adding the "boosting" and "lowering" contributions from all the factors to this baseline value, the model's effectiveness against the proposed solution can be accurately obtained. The final predicted value;

[0150] It should be noted that visual analysis allows engineers to easily identify the problems in the solution. In this context, what are the key risk factors that lead to increased vibration or dust, and what are the core control parameters that play a suppressive role?

[0151] It should be noted that after understanding the overall contribution of each parameter, engineers often need to know how to fine-tune the current blasting parameter scheme. To this end, this invention adopts a local sensitivity analysis method based on gradient calculation. At the same time, since the constructed neural network model is differentiable in nature, it can efficiently calculate the gradient of its output with respect to the input.

[0152] Furthermore, regarding the first part of the plan... Explosion parameters Its prediction output for a specific purpose Local sensitivity (e.g., PPV value, or concentration value at a key point in a dust concentration field) Defined as the partial derivative of the output with respect to the parameter:

[0153] ;

[0154] in, Indicates output Relative to input parameters rate of change, This indicates that the partial derivative is in the current non-dominated solution (scheme). The calculation is performed at this specific point;

[0155] It should be noted that this gradient value can be calculated in one step using the backpropagation algorithm of a neural network;

[0156] Specifically, when When, it means in the plan Nearby, slightly increase parameters The value will cause the predicted output The increase, and When, it indicates a slight increase. Predicted output The decrease;

[0157] Specifically, if The larger the absolute value, the better the predicted output. For parameters The more sensitive the parameter is to changes, the more it prompts engineers to pay attention to highly sensitive parameters. Precise control is essential; even minor adjustments can cause significant changes in the explosive effect. The smaller the absolute value, the better the predicted output. Regarding this parameter Insensitive; at this point, the engineer is adjusting the parameter. This may allow for greater leeway, or imply that the control precision can be appropriately relaxed if cost permits.

[0158] It should be noted that by using the above two optimization methods, engineers can not only understand which parameters contribute to the "excellence" and "inferiority" of a solution, but also grasp the scientific basis and direction for fine-tuning the current solution on-site. This will enhance the credibility and practical application value of blasting parameter solutions in blasting projects where safety is of paramount importance.

[0159] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0160] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0163] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0164] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control, characterized in that, include: Based on training samples containing geological condition data, blasting parameter data, and corresponding actual blasting effect data, a neural network model is trained to establish a mapping relationship from the geological condition data and blasting parameter data to the predicted blasting effect data. Using the neural network model, a preset evolutionary algorithm is employed to optimize an objective function containing four categories of indicators—vibration, dust, cost, and regulatory compliance—with blasting parameters as variables, until a Pareto optimal solution set composed of multiple non-dominated parameter schemes is obtained. The four categories of indicators in the objective function are quantified as follows: the vibration indicator is calculated from the vibration intensity or frequency domain characteristics in the predicted blasting effect data; the dust indicator is calculated from the dust concentration or excess volume in the predicted blasting effect data. The cost index is calculated based on the resource consumption corresponding to the blasting parameters; the regulatory compliance index is obtained by comparing the vibration and dust indexes with regulatory limits; the regulatory compliance index is calculated using a penalty function: when the calculated value of the vibration or dust index does not exceed the regulatory limit, the output value of the penalty function is less than a preset first threshold; when the calculated value of the index exceeds the regulatory limit, the output value of the penalty function increases with the increase of the excess; for at least one non-dominated parameter scheme in the Pareto optimal solution set, a preset interpretation algorithm is used to analyze the prediction process of the neural network model to determine the quantitative impact of each blasting parameter in the scheme on the predicted blasting effect data.

2. The multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control as described in claim 1, characterized in that, The specific steps for constructing the neural network model include: using first and second encoder networks to extract geological condition feature vectors and blasting parameter feature vectors from the geological condition data and blasting parameter data, respectively; and training the encoder network using a preset loss function so that the distance between the geological condition feature vectors and blasting parameter feature vectors originating from the same blasting event in the latent space converges.

3. The multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control as described in claim 2, characterized in that, The construction steps further include: using the geological condition feature vector and blasting parameter feature vector obtained after distance convergence training as conditional inputs to train a neural network model; the neural network model learns the process of gradually denoising random noise to generate target data, thereby realizing the generation operation from the conditional input to the predicted blasting effect data.

4. The multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control as described in claim 3, characterized in that, When training the neural network model, the way to introduce the conditional input into the model includes: setting an attention calculation module in the network structure of the neural network model, the module receiving the intermediate layer features of the model as a query, and receiving the conditional input as a key and a value, weighting the value by calculating the correlation between the query and the key, and applying the weighting result to the intermediate layer features.

5. The multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control as described in claim 1, characterized in that, The predicted blasting effect data includes: complete velocity-time history curve data representing the vibration process; and three-dimensional concentration field data representing the spatial distribution of dust.

6. The multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control as described in claim 1, characterized in that, The specific process of the evolutionary algorithm includes: initializing a population consisting of multiple candidate blasting parameter schemes; in each iteration of the algorithm, for each scheme in the population, calling the neural network model to calculate its corresponding target indicators; and based on the target indicators, performing selection, crossover, and mutation operations on the population through non-dominated sorting and crowding calculation to generate a new generation of population.

7. The multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control as described in claim 1, characterized in that, The interpretation algorithm used is a method for quantifying the contribution of each input feature to the model output. This method is used to calculate the specific contribution value of each parameter in the parameter scheme to the predicted explosion effect data.

8. The multi-objective collaborative optimization parameter tuning method for blasting vibration and dust control as described in claim 1, characterized in that, The interpretation algorithm used is a gradient calculation method, which determines the sensitivity of each blasting parameter by calculating the gradient of the predicted blasting effect data with respect to the input blasting parameters.

Images