Tunnel multidirectional energy-gathered hole underholing drilling tool and blasting method

Abstract

The invention provides a tunnel multidirectional energy-gathered hole slotting drilling tool and a blasting method, and relates to the technical field of tunneling engineering, the drilling tool comprises a drilling tool main body, a depth detection module, a hole diameter detection module, an electric quantity monitoring module, a wear degree detection module, a master control module and a data transmission module; the method comprises the steps that geological data are collected and preprocessed, feature extraction is carried out on the preprocessed geological data, blasting parameter prediction is carried out based on extracted features, and blasting is carried out based on blasting parameters. According to the method, drilling in the blasting process is achieved based on the intelligent drilling tool, the defects of a traditional slotting process are overcome, blasting parameters are predicted, the energy utilization rate can be effectively increased, surrounding rock damage is reduced, use is convenient, and important engineering application value and social significance are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel boring engineering, and particularly to a tunnel multi-directional energy-gathering hole cutting drill and a blasting method. Background Art

[0002] In tunnel boring engineering, cut blasting is the core link determining construction efficiency and forming quality. Its core goal is to form an effective cavity through precise drilling layout and blasting energy control to accommodate the rock throwing generated by subsequent blasting. However, there are significant defects in both drill design and blasting methods in traditional cut blasting processes, resulting in problems such as low energy utilization rate, insufficient cut depth, and large surrounding rock damage. Therefore, it is very necessary to design a tunnel multi-directional energy-gathering hole cutting drill and a blasting method. Summary of the Invention

[0003] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a tunnel multi-directional energy-gathering hole cutting drill and a blasting method.

[0004] To achieve the above purpose, the present invention provides the following solutions:

[0005] The present invention provides a tunnel multi-directional energy-gathering hole cutting drill, including: a drill body, a depth detection module, a hole diameter detection module, a power quantity monitoring module, a wear degree detection module, a total control module, and a data transmission module. The depth detection module is arranged on the outside of the drill rod propulsion mechanism of the drill body for real-time measurement of the drilling depth. The hole diameter detection module is annularly distributed behind the drill bit of the drill body for detecting the actual diameter of the drill hole. The power supply interface of the power system of the drill body is connected to the power quantity monitoring module for real-time monitoring of the power quantity of the drill body. The wear degree detection module is embedded inside the drill bit tool holder of the drill body for analyzing the wear degree of the tool head through vibration spectrum. The depth detection module, the hole diameter detection module, the power quantity monitoring module, and the wear degree detection module are connected to the total control module, and the total control module is connected to the background monitoring host through the data transmission module.

[0006] The present invention also provides a tunnel multi-directional energy-gathering hole cut blasting method, including:

[0007] Step 1: Collect geological data and perform preprocessing on it;

[0008] Step 2: Extract features from the preprocessed geological data;

[0009] Step 3: Predict blasting parameters based on the extracted features;

[0010] Step 4: Perform blasting based on the blasting parameters.

[0011] Preferably, the geological data includes three-dimensional rock mass strength data, joint fracture data, water content and groundwater data, and historical blasting data.

[0012] Preferably, in step 1, the geological data is preprocessed, specifically:

[0013] Denoising processing is performed on the three-dimensional rock mass strength data to obtain a three-dimensional multi-channel feature network;

[0014] Fracture extraction and three-dimensional reconstruction processing are performed on the joint fracture data to obtain joint topology map data;

[0015] Outlier rejection and spatial matching processing are performed on the water content and groundwater data to obtain a water content and groundwater data matrix;

[0016] Data cleaning processing is performed on the historical blasting data to obtain a historical blasting parameter matrix.

[0017] Preferably, feature extraction is performed on the preprocessed geological data, specifically:

[0018] For the three-dimensional multi-channel feature network and the water content and groundwater data matrix, feature extraction is performed on them based on the 3D-CNN network structure to obtain a spatial feature vector for characterizing the overall mechanical properties of the rock mass;

[0019] For the joint topology map data, feature extraction is performed on it based on the graph neural network structure to obtain a node feature matrix and an adjacency matrix;

[0020] Based on the historical blasting parameter matrix, feature extraction is performed on it based on the fully connected branch to obtain a dynamic feature vector;

[0021] The spatial feature vector, the node feature matrix and the adjacency matrix, and the dynamic feature vector are concatenated according to the dimension to obtain a fused feature.

[0022] Preferably, in step 3, blasting parameter prediction is performed based on the extracted features, specifically:

[0023] Construct a blasting parameter prediction model;

[0024] Train the blasting parameter prediction model based on a preset data set;

[0025] Input the obtained fused feature into the trained blasting parameter prediction model to obtain the blasting parameters.

[0026] Preferably, the blasting parameter prediction model is constructed based on the SOA-optimized BP neural network structure.

[0027] According to the specific embodiments provided by the present invention, the following technical effects are disclosed by the present invention:

[0028] The present invention provides a tunnel multi-directional energy-gathering hole cutting drill and a blasting method. The drill includes a drill body, a depth detection module, a hole diameter detection module, a power monitoring module, a wear degree detection module, a total control module and a data transmission module. The method includes: collecting geological data and preprocessing it, extracting features from the preprocessed geological data, predicting blasting parameters based on the extracted features, and performing blasting based on the blasting parameters. The present invention realizes drilling during the blasting process based on an intelligent drill, solves the defects of traditional cutting processes. The present invention predicts blasting parameters, can effectively improve energy utilization rate, reduce surrounding rock damage, is easy to use, and has important engineering application value and social significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0030] Figure 1 It is a flowchart of the method provided by the embodiment of the present invention;

[0031] Figure 2 It is a structural schematic diagram of MsEDNet;

[0032] Figure 3 It is a structural schematic diagram of Unet++;

[0033] Figure 4 It is a structural schematic diagram of VGG16. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0035] The purpose of the present invention is to provide a tunnel multi-directional energy-gathering hole cutting drill and a blasting method, which realizes drilling during the blasting process based on an intelligent drill, solves the defects of traditional cutting processes. The present invention predicts blasting parameters, can effectively improve energy utilization rate, reduce surrounding rock damage, is easy to use, and has important engineering application value and social significance.

[0036] To make the above objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] The present invention provides a tunnel multi-directional energy-gathering hole cutting drill, including: a drill body, a depth detection module, a hole diameter detection module, a power monitoring module, a wear degree detection module, a total control module, and a data transmission module;

[0038] The depth detection module is arranged outside the drill rod propulsion mechanism of the drill body for real-time measurement of the drilling depth. The depth detection module can adopt an optoelectronic encoder + laser ranging for auxiliary calibration;

[0039] The hole diameter detection module is annularly distributed behind the drill bit of the drill body for detecting the actual diameter of the drilled hole. The hole diameter detection module can adopt a micro ultrasonic sensor array (4 groups symmetrically arranged);

[0040] The power supply interface of the power system of the drill body is connected to the power monitoring module for real-time monitoring of the power of the drill body. The power monitoring module can adopt a voltage / current sensor + a coulomb meter chip;

[0041] The wear degree detection module is embedded inside the tool bit seat of the drill body for analyzing the wear degree of the tool bit through vibration spectrum. The wear degree detection module can adopt a three-axis acceleration sensor + a temperature sensor;

[0042] The present invention can set linkage calibration and adaptive adjustment according to specific requirements. Two embodiments are given as follows:

[0043] 1. Depth-hole diameter linkage calibration: During the drilling process, the hole diameter detection is triggered every 10 cm of drilling. If the detected hole diameter shrinks (such as caused by rock mass collapse), the drill bit rotation speed and propulsion pressure are automatically adjusted to ensure that the hole diameter meets the standard;

[0044] 2. Wear-lithology adaptability: When the wear degree detection is "moderate wear" and the lithology is identified as hard rock (uniaxial compressive strength ≥ 100 MPa), the drilling speed is automatically reduced by 20% to extend the service life of the drill bit.

[0045] The depth detection module, the hole diameter detection module, the power monitoring module, and the wear degree detection module are connected to the total control module, and the total control module is connected to the background monitoring host through the data transmission module.

[0046] As Figure 1 shown, the present invention also provides a tunnel multi-directional energy-gathering hole cutting blasting method, including:

[0047] Step 1: Collect geological data and preprocess it;

[0048] Step 2: Extract features from the preprocessed geological data;

[0049] Step 3: Predict blasting parameters based on the extracted features;

[0050] Step 4: Conduct blasting based on the blasting parameters.

[0051] The geological data includes three-dimensional rock mass strength data, joint fracture data, water content and groundwater data, and historical blasting data;

[0052] Introduce them respectively:

[0053] 1. Three-dimensional rock mass strength data

[0054] The acquisition equipment includes:

[0055] Ground penetrating radar (such as MALA ProEx): emits high-frequency electromagnetic waves, and inversely calculates the elastic modulus and wave velocity ratio of the rock mass through the reflected signals, and indirectly calculates the uniaxial compressive strength;

[0056] Borehole penetrometer (such as Panda Dynamic Penetrometer): measures the penetration resistance of the drill bit in real time during the drilling process and directly obtains the local rock mass strength (resolution: 5 cm interval);

[0057] The acquired data includes: ground penetrating radar signals (raw electromagnetic waveform data), penetrometer penetration resistance curve (CSV format, containing depth-resistance value correspondence table);

[0058] 2. Joint fracture data

[0059] The acquisition equipment includes:

[0060] Borehole camera (Opticlog): takes 360° high-definition images of the borehole wall (resolution 5MP, adjustable focal length);

[0061] Three-dimensional laser scanner (FARO Focus): generates tunnel face point cloud (accuracy ±1mm, point spacing 2mm);

[0062] The acquired data includes: borehole image sequence (JPEG format, about 200 - 500 images per hole), laser point cloud data (.las format, containing three-dimensional coordinates and reflection intensity);

[0063] 3. Water content and groundwater data

[0064] The acquisition equipment includes:

[0065] Transient electromagnetic instrument (TEM): detects the resistivity of the rock mass and inversely calculates the water content distribution;

[0066] Drilling sampler: extract core samples, and determine the moisture content by laboratory drying method (accuracy ±0.5%);

[0067] The collected data includes: TEM resistivity profile map (GeoTIFF format), laboratory moisture content determination report (Excel spreadsheet);

[0068] 4. Historical blasting data

[0069] The collection sources include:

[0070] Blasting monitoring system: record vibration velocity, air shock wave, cut depth (sampling rate 1 kHz);

[0071] Field construction log: manually enter hole position coordinates, charge amount, initiation timing sequence;

[0072] The collected data includes: hole position coordinates (X, Y, Z), hole diameter, hole depth, charge type (emulsion explosive, ammonium nitrate fuel oil explosive, etc.), charge density (kg / m 3 ), initiation delay (ms), cut depth (m), average rock block particle size (cm).

[0073] In step 1, preprocess the geological data, specifically:

[0074] Perform denoising processing on the three-dimensional rock mass strength data to obtain a three-dimensional multi-channel feature network (normalized rock mass strength, normalized moisture content, joint density). Among them, the denoising processing part is introduced in detail:

[0075] The present invention uses the MsEDNet denoising model to perform denoising processing on the three-dimensional rock mass strength data, and introduce it. The multi-scale echo denoising model (Multi-scale Echo Denoising Network, MsEDNet) with a multi-scale residual structure and an attention mechanism is used for denoising of echo signals, and has a complete end-to-end structure without manually extracting features from the signal;

[0076] The main task of the model is to estimate the denoised signal from the noisy signal y and remove the noise v as much as possible. Figure 2 Shows the structure diagram of MsEDNet. The network is mainly composed of a noise estimation sub-network, a multi-scale residual module, an attention module, and a non-blind denoising sub-network. The input of the noise estimation sub-network is the noisy echo signal, and the output is the estimated noise where θ E$\theta$ is the parameter of the noise estimation sub-network. The multi-scale residual module and the attention module constitute the feature extraction part, which outputs feature maps of multiple levels. These feature maps are concatenated with the estimated noise in the channel dimension and then fed into the non-blind denoising sub-network. The classic denoising model DnCNN in the field of images has verified through experiments the benefits of residual learning for improving training stability and denoising effect. Therefore, MsEDNet also connects the noisy echo signal across layers to the output, that is, what the non-blind denoising sub-network learns is the residual. $\theta$ D is the parameter of the non-blind denoising sub-network;

[0077] The flowchart of echo signal denoising based on MsEDNet is as follows:

[0078] 1. Collect echo signals and simulation signals to form a dataset, which includes original noisy signals and noise-free signals;

[0079] 2. Preprocess the dataset, linearly scale the data to between 0 and 1 using normalization, and apply the scaling coefficient of the original signal to the noise-free signal;

[0080] 3. Divide the dataset into a training set and a test set according to a certain splitting ratio, and the splitting ratio is usually 7:3;

[0081] 4. Use the training set to train the MsEDNet model, verify the well-trained model on the test set, and output the denoised signal.

[0082] Next, the multi-scale residual module will be introduced. Feature extraction is the first step in the network's learning and training process. The quality of the features has a crucial impact on the denoising result of the model. At the same time, factors such as depth, width, and convolution kernel size need to be considered in the design of the network architecture;

[0083] To more effectively learn features from echo signals, the multi-scale residual module is constructed by combining residual learning and the Inception architecture. Vertically stacking residual blocks increases the depth of the network, and the horizontally extended Inception structure increases the width of the network by parallel computing convolution kernels of different scales, obtaining multi-scale features in the echo signal to improve the network's feature extraction ability;

[0084] Before the original signal is input into the multi-scale residual module, it is converted into a feature map of 64×800 (channels×signal length) through a convolutional layer. The input and output of the multi-scale residual module used in the present invention are both 64×800, and the signal length remains unchanged during the convolution process. The identity mapping preserves the original information of the input feature map. Four parallel residual branches are designed. According to multiple experiments, convolution kernels of different scales such as 1×1, 1×5, 1×11, and 1×21 are used, and the padding numbers of the convolution are 0, 2, 5, and 10 respectively. Each branch first uses a 1×1 convolution to reduce the number of channels of the feature map to 16, which reduces a large amount of computational complexity compared to directly using a 1×N convolution kernel to reduce the channels. Then, a 1×N convolution kernel is used for feature extraction. The features of different scales are concatenated in the channel dimension, and the feature map is restored to 64 channels, thereby expanding the feature mapping space and realizing multi-scale feature fusion. Finally, the input is added to the output of the residual mapping to obtain the overall output of the network. In addition, referring to the residual block structure of ResNetV2, the output of the previous layer is preprocessed using the BN and ReLU activation functions before the one-dimensional convolution operation.

[0085] Next, the present invention introduces the attention module. The essence of the attention mechanism is to apply greater attention to the regions of the features that are more interesting. This process is very similar to the way humans observe objects. In the network, the weights are learned using the relevant feature maps, and then the learned weights are fused with the original feature maps, so that the effective features are strengthened and the useless or inefficient features are weakened. Introducing the attention mechanism can enhance the learning ability of the network and thus improve the model effect;

[0086] Currently, the Squeeze-and-Excitation Networks (SENet) and the Convolutional Block Attention Module (CBAM) are representative networks in the attention mechanism. The former explicitly models the interdependence between channels to adaptively recalibrate the feature responses in terms of channels, while the latter first performs channel-domain attention and then spatial-domain attention. SENet is lighter than CBAM and can be selectively embedded between the layers of any network, and the effect is not much different from that of CBAM. Therefore, the present invention uses SENet as the attention module;

[0087] The size of the input feature map X of the SE module is L×C, where L is the length and C is the number of channels. The squeezing operation compresses the feature map spatially to 1×C through global average pooling, enabling information aggregation; the excitation operation uses two fully connected layers to introduce non-linearity and aid generalization, that is, a dimensionality reduction layer with an output dimension of 1×C / r, a ReLU, and then a dimensionality increase layer to expand to 1×C, and finally a sigmoid activation function to scale it between 0 and 1 to form weights. Finally, each channel of X is multiplied by the corresponding weight to obtain the recalibrated feature map.

[0088] Next, the present invention introduces the loss function:

[0089] MsEDNet aims to learn the mapping from the noisy echo signal to the denoised signal, so the reconstruction loss function is defined as the main body to evaluate the denoised signal The difference from the noise-free signal x, that is:

[0090]

[0091] Considering that the echo signal exhibits smooth characteristics in mathematical formulas, that is, it is differentiable everywhere, the total variation (TV) of the signal contaminated by noise is much larger than that of the noise-free signal. Therefore, a gradient constraint loss function is introduced as a regular term to minimize the sum of the squares of the gradients of the denoised signal, that is,

[0092]

[0093] The model uses an asymmetric loss function to evaluate the quality of noise estimation. Given the noise σ(y) of the noisy signal and the estimated noise level When the estimated noise level is lower than the actual noise, a greater penalty needs to be imposed. Conversely, the penalty is reduced. The asymmetric noise estimation loss function is expressed as:

[0094]

[0095] where, when then and 0 in other cases. When α∈(0,0.5), a greater penalty can be imposed on the underestimated predicted noise;

[0096] To sum up, the objective loss function of the entire MsEDNet model is

[0097] L = L rec + λ TV L TV + λ asymm L asymm ;

[0098] Among them, λ τv and λ asymm are the proportions of the gradient constraint loss function and the asymmetric noise estimation loss function respectively.

[0099] For joint fracture data, fracture extraction and three-dimensional reconstruction processing are carried out to obtain joint topology map data (including adjacency matrix and node / edge attributes, specifically including coordinates, fracture length, strike, dip angle, connectivity, average width). Among them, this part is introduced in detail:

[0100] First, the joint fracture data is segmented, and finally the set parameters are calculated (fracture length, strike, dip angle), its joint density, fracture rate and joint occurrence are calculated. Finally, point cloud registration is carried out to align the two-dimensional fracture coordinates of borehole imaging with the three-dimensional point cloud of laser scanning. The ICP (Iterative Closest Point) algorithm is used, and the registration error ≤ 1 cm. Fracture network modeling is carried out. The Marching Cubes algorithm is used to extract the fracture surface from the point cloud to generate a triangular mesh model (STL format), and a graph structure is constructed;

[0101] The fracture segmentation part is introduced in detail. Based on the VGG_Unet++ network structure, fracture segmentation is realized. The borehole image of the joint fracture data is input, and a binary mask and a marked fracture area are output;

[0102] The present invention adopts the VGG_Unet++ network structure and introduces it as follows:

[0103] Unet++ is improved on the basis of the Unet network. As Figure 3 shown, the nine nodes of the leftmost encoder (contraction path) and the rightmost decoder (expansion path) are the architecture of the original Unet network. The Unet++ network changes the connection method between them to form a dense and nested connection segmentation structure;

[0104] The network has five layers from top to bottom (L0 - L4). The difference between Unet++ and Unet lies in the nodes in the middle part and their connections. The green arrows represent the process of upsampling the deep feature map and fusing it with the shallow feature map, and the blue arrows represent the nested connection of the features in the same layer. Therefore, each node receives the outputs of all nodes before the same jump path, as well as the upsampled output of the lower jump path. This process makes full use of the features of multiple semantic levels to form a dense jump structure, bridging the semantic gap and comprehensively and effectively capturing coarse-grained features and fine-grained features. The specific expression is as follows:

[0105]

[0106] In the formula: x i,jis a node, where i is the index along the encoder downsampling layer, taking values from 0 to 4, j is the index of the dense convolutional layer along the skip path, H(·) represents the convolutional operation, [·] is the connection layer. When j = 0, it means receiving only the output of the previous encoder layer. When j > 0, this node receives j + 1 inputs simultaneously;

[0107] The backbone network of the encoder part can be selected according to the training requirements. Commonly used ones as the backbone network include VGG16, VGG19, ResNet50, ResNet101, etc. Among them, the VGG series is often used as an image feature extraction tool because of its strong expandability. In this invention, VGG16 is selected as the backbone network, and its network structure diagram is as Figure 4 shown, Figure 3 Each node in is a feature extraction module. Each module contains a convolutional layer, a batch normalization BN layer, and an activation function layer according to the basic structure of VGG16. The size of the convolutional kernel is 3×3. When performing the convolutional operation, the moving step of the convolutional kernel is set to 1. The data features extracted by the convolution are immediately input into the batch normalization BN layer to perform normalization on the current input data, ensuring that the output of each layer is limited within a fixed range. The Relu activation function performs non-linear calculations, which can accelerate the training speed of the network on the one hand and prevent the occurrence of gradient disappearance or gradient explosion on the other hand. There will be a pooling layer with a window of 2×2 after each module, which slides on the image according to the moving rule with a step of 2 to downsample the image. In this invention, the selected image size is relatively small, so some modifications are made based on VGG16: changing the number of convolutional kernels in the first four layers, and the number of modified convolutional kernels is half of the number of convolutional kernels in the original VGG16; removing the last three fully connected layers, and the size of each layer of feature map in the Unet++ network is the same.

[0108] The VGG_Unet++ network uses the DeepSupervision method. Since the network extracts multi-level features of the image for fusion and finally obtains the segmented image, there is a corresponding loss for each level of feature map compared with the label map. The DeepSupervision method can receive the gradients passed from the intermediate part of the output and directly optimize the shallow network;

[0109] VGG Unet++ combines the binary cross-entropy loss function and the Dice loss function;

[0110] To prevent the denominator of the Dice loss function from being 0, a smoothing factor smooth is added to the denominator to ensure that the network can be optimized.

[0111]

[0112] In the formula, γ represents the sample label map, Represents the predicted segmentation image of the sample, with smooth taking le-4.

[0113] The Dice loss function has the problem of unstable training. The binary cross-entropy loss function is introduced to ensure the stability of network training. Therefore, the expression of the composite loss function is as follows:

[0114]

[0115] Since the network introduces the deep supervision strategy, the loss function receives not only the loss value of node X 0,4 but also the loss values of the first three nodes X 0,1 、X 0,2 、X 0,3 Then the expression of the final loss function is:

[0116]

[0117] In the formula, L represents the loss function of the entire network, and N represents the level.

[0118] Outlier rejection and spatial matching processing are performed on the water content and groundwater data to obtain the water content and groundwater data matrix. This part is introduced as follows:

[0119] 1. Outlier rejection: For the water content and groundwater data, the 3-Sigma criterion is adopted to delete the data points that deviate from the mean by ±3 times the standard deviation. The 3-Sigma criterion is introduced as follows:

[0120] The 3-Sigma criterion, also known as the 3-Sigma rule, can effectively diagnose abnormal data through probability and statistics knowledge. Its basic idea is that in a sample data set, usually normal data exists in the region of high probability distribution of the model, while abnormal data exists in the region of low probability distribution;

[0121] The steps for the 3-Sigma criterion to identify abnormal data are generally as follows: First, assume that the sample data set conforms to a normal distribution, and the standard deviation σ and mean μ of the data set can be obtained through calculation. Then, a confidence interval (μ - 3σ, μ + 3σ) is determined. If it is judged that the sample data does not belong to this interval, then it can be considered as a gross error, that is, abnormal data;

[0122] For the sample data set x i (i = 0, 1, 2,..., n), the calculation formula for the mean μ is:

[0123]

[0124] The absolute value v i of the difference between the data x i and the mean is calculated as:

[0125] v i = |x i - μ|;

[0126] Finally, calculate the data x i Standard deviation:

[0127]

[0128] This application can also perform spatial matching: align the moisture content data with the rock mass strength grid, assign moisture content attributes to each voxel, and form a multi-channel three-dimensional tensor (strength + moisture content);

[0129] Perform data cleaning on historical blasting data to obtain a historical blasting parameter matrix (hole position coordinates (X, Y, Z), charge density, initiation delay, cut depth, etc.). Introduce this part. The specific processing includes:

[0130] 1. Missing value processing: For fields such as charge amount and vibration velocity, use random forest regression to fill in the missing values;

[0131] 2. Data standardization: Perform one-hot encoding on non-dimensional parameters (such as explosive type), and perform Z-score standardization on continuous parameters (such as hole depth);

[0132] 3. Temporal alignment: Synchronize the drilling, charging, and initiation data of the same blasting cycle according to the timestamp to construct a temporally correlated dataset.

[0133] In step 2, feature extraction is performed on the preprocessed geological data. Specifically:

[0134] For the three-dimensional multi-channel feature network and the moisture content and groundwater data matrix, perform feature extraction on it based on the 3D-CNN network structure to obtain a spatial feature vector (extract the spatial distribution features of rock mass strength, moisture content, and joint density). Here, introduce the 3D-CNN network structure:

[0135] Its input is a three-dimensional multi-channel feature network, and the output is a spatial feature vector, which characterizes the overall mechanical properties of the rock mass. The present invention gives a specific meaning schematic table of a 32-dimensional spatial feature vector as shown in Table 1;

[0136] Table 1 Schematic table of the specific meaning of the spatial feature vector

[0137]

[0138] 3D-CNN was first proposed for human action recognition. It can extract relevant features of human actions from consecutive frames of a video, thereby identifying and classifying human actions. An introduction to the 3D-CNN network model is as follows, and its calculation formula is as follows

[0139]

[0140] In the formula, v ij xyz represents the value of the neuron at (x, y, z), i represents the i-th layer of neurons, j represents the j-th feature map, P i and Q i are the height and width of the convolutional kernel, R i is the dimensionality size of the convolutional kernel in the spectral dimension, m represents the number of connected features in the previous layer, w i jm pqr is the weight connected to the (p, q, r)-th neuron in the m-th feature, b ij is the bias value of the j-th feature map on the i-th layer of neurons, and f is the activation function. The ReLU function is used in this invention.

[0141] For the joint topological map data, feature extraction is performed based on the graph neural network structure to obtain the node feature matrix and the adjacency matrix (modeling the influence of joint connectivity on blasting energy transfer). Regarding the graph neural network structure, it will not be introduced in detail here. Its input is the adjacency matrix and node / edge attributes, specifically including coordinates, fracture length, strike, dip angle, connectivity, and average width. The output is the node feature matrix and the adjacency matrix, representing the influence of the joint network on blasting energy propagation. A specific meaning schematic table of a 48-dimensional node feature matrix and an adjacency matrix vector is shown in Table 2 of this invention;

[0142] Table 2 Schematic table of the specific meanings of the node feature matrix and the adjacency matrix vector

[0143]

[0144] Based on the historical blasting parameter matrix, feature extraction is performed based on the fully connected branch to obtain the dynamic feature vector (fusing historical experience). Regarding the fully connected branch structure, it will not be introduced in detail here. Its input is the historical blasting parameter matrix, and the output is the dynamic feature vector;

[0145] The spatial feature vector, the node feature matrix and the adjacency matrix, and the dynamic feature vector are concatenated by dimension to obtain the fused feature.

[0146] In step 3, blasting parameter prediction is performed based on the extracted features. Specifically:

[0147] Construct a blasting parameter prediction model;

[0148] Train the blasting parameter prediction model based on a preset data set;

[0149] Input the obtained fusion features into the trained blasting parameter prediction model to obtain the blasting parameters [X 主 , Y 主 , Z 主 , N 辅助 , ρ 装药 , △t 延时 , D 孔深 .

[0150] The blasting parameter prediction model is constructed based on the SOA-optimized BP neural network structure, and an introduction is given as follows:

[0151] First, introduce the improved seagull optimization algorithm:

[0152] Due to the random factor B in the traditional SOA, the optimization effect is poor, it is easy to fall into local optimum, and the development ability in the later stage is weak. Therefore, the present invention uses an improved variation factor B, as shown in the following formula

[0153] B = 2 × A 2 × rd * (1 - x / Max iteration );

[0154] To solve the deficiency of weak development ability of the algorithm in the later stage, the present invention adopts the chaotic idea, increases the diversity of particles in the later stage, enhances its search ability, and introduces the Logistic mapping. The basic formula is:

[0155]

[0156] Among them, represents the kth iteration of the i (i = 1, 2, 3,..., n)th chaotic variable, and the value of μ is generally 4;

[0157] The conversion between the chaotic variable and the original variable is as follows

[0158]

[0159] In the formula, x L,i and x U,i are respectively the upper and lower bounds of the search for the i-th dimension variable, is the i-th chaotic variable after chaotic mapping The value obtained by converting into an optimized variable;

[0160] The main steps of the algorithm based on the SOA-optimized BP neural network structure are:

[0161] 1. Initialize parameters such as the number of seagulls, and form an initial seagull population with the weights and thresholds of the BP neural network to be optimized.

[0162] 2. Initialize the positions of the seagulls and process their positions using a formula to prevent position overlap.

[0163] 3. Calculate the fitness of all current seagulls and find the best one among them as the best seagull in this iteration.

[0164] 4. Update the positions of each seagull according to the formula.

[0165] 5. Chaotic algorithm part: Use the Logistic mapping to map the individual extreme values of the particles, perform chaotic iteration within (0, 1), and after the iterative operation, inverse map the result back to the original solution space range, calculate and evaluate the fitness value of the current solution. When the new solution is better than the old solution, output the new solution.

[0166] 6. Determine whether the iteration times or the required accuracy are reached. If so, output the final position as the optimal seagull position; otherwise, return to step 3.

[0167] 7. Decode the optimal output into the initial weights and thresholds of the BP neural network, and train the neural network until the requirements are met.

[0168] In step 4, blasting is carried out based on the blasting parameters, specifically:

[0169] According to the obtained blasting parameters, drill holes and set explosives based on the drilling tools, and finally achieve blasting.

[0170] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other.

[0171] Specific examples are used in this article to elaborate on the principles and implementation methods of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A multi-directional energy-gathering hole cutting tool for a tunnel, characterized in that: include: A drilling tool body, a depth detection module, an aperture detection module, a power monitoring module, a wear detection module, a master control module and a data transmission module. The depth detection module is arranged on the outside of the drill rod propulsion mechanism of the drilling tool body, which is used to measure the drilling depth in real time. The aperture detection module is arranged in a ring-shaped distribution behind the drill bit of the drilling tool body, which is used to detect the actual diameter of the drill hole. The power supply interface of the power system of the drilling tool body is connected to the power monitoring module, which is used to monitor the power of the drilling tool body in real time. The wear detection module is embedded inside the drill bit holder of the drilling tool body, which is used to analyze the wear degree of the cutter head through the vibration spectrum. The depth detection module, the aperture detection module, the power monitoring module and the wear detection module are connected to the master control module, and the master control module is connected to the background monitoring host through the data transmission module.

2. A tunnel multi-directional energy-gathering hole cutting blasting method, characterized in that: include: Step 1: Collect geological data and preprocess them; Step 2: Extract features from the preprocessed geological data; Step 3: Predict blasting parameters based on the extracted features; Step 4: Perform blasting based on blasting parameters.

3. The method according to claim 2, characterized in that The geological data include three-dimensional rock mass strength data, joint and fissure data, water content and groundwater data and historical blasting data.

4. The method according to claim 3, characterized in that In step 1, the geological data is preprocessed, specifically: De-noising is performed on the three-dimensional rock mass strength data to obtain a three-dimensional multi-channel feature network; Perform crack extraction and three-dimensional reconstruction on the joint and crack data to obtain joint topology data; For the water content and groundwater data, outliers are eliminated and spatial matching is performed to obtain the water content and groundwater data matrix; Data cleaning is performed on historical blasting data to obtain the historical blasting parameter matrix.

5. The method according to claim 4, characterized in that The feature extraction of the preprocessed geological data is as follows: Based on the 3D-CNN network structure, the three-dimensional multi-channel feature network and the water content and groundwater data matrix are subjected to feature extraction to obtain spatial feature vectors, which are used to characterize the overall mechanical properties of the rock mass. For the joint topology data, feature extraction is performed based on the graph neural network structure to obtain the node feature matrix and adjacency matrix; Based on the historical blasting parameter matrix, feature extraction is performed based on the fully connected branch to obtain a dynamic feature vector; The spatial feature vector, node feature matrix, adjacency matrix and dynamic feature vector are concatenated by dimension to obtain fused features.

6. The method according to claim 5, characterized in that In step 3, the blasting parameters are predicted based on the extracted features, specifically: Construct a blasting parameter prediction model; Training the blasting parameter prediction model based on a preset data set; The obtained fusion features are input into the trained blasting parameter prediction model to obtain the blasting parameters.

7. The method according to claim 6, characterized in that The blasting parameter prediction model is constructed based on the SOA optimized BP neural network structure.

Images