Methd and system for predicting blasting deformation of muck pile based on 3D convolutional neural network

Abstract

The application provides a kind of based on 3D convolutional neural network's explosive pile blasting deformation prediction method, comprising: constructing 3D convolutional neural network, 3D convolutional neural network is trained by training data, obtains trained 3D convolutional neural network;Exploration data before explosive pile blasting is obtained, and exploration data is preprocessed, obtains first matrix and second matrix;First matrix and second matrix are input into trained 3D convolutional neural network and are predicted, and deformation prediction result after explosive pile blasting is obtained.The application constructs 3D convolutional neural network by the stacking mode of ordinary unit and dimension reduction unit, can obtain more varied features, the same structure can be used to the same unit, can simplify the structure of 3D convolutional neural network, reduce the difficulty of building;Through exploration data, first matrix with block segment information and second matrix with hardness information are constructed as the input of 3D convolutional neural network, fully embody the multiple characteristics of explosive pile blasting, avoid feature loss.

Description

Technical Field

[0001] This invention relates to the field of blasting prediction, and more particularly to a method and system for predicting blasting deformation based on a 3D convolutional neural network. Background Technology

[0002] Before open-pit mining, planned blast piles need to be blasted. Although blasting designs aim to minimize displacement of the blast piles, some degree of change is unavoidable. Therefore, the grade distribution of the blast piles will inevitably change to some extent after blasting. As mining processes become increasingly automated and intelligent, the accuracy requirements for various production data and indicators are becoming more stringent. Data errors directly determine the rationality of task allocation. The grade distribution of the blast piles is one of the most crucial data points. Using exploration data as a reference grade at this stage will introduce errors, which accumulate during continuous mining operations, leading to resource waste. In open-pit mining, geological exploration is typically conducted on the target blast piles, involving uniform sampling to obtain relatively accurate exploration data. Displacement after blasting means the grade distribution of the blast piles will not perfectly match the exploration data; this error is significant. Since blast piles are generally considered to consist of numerous segments, machine learning can be introduced to predict the deformation of the entire blast pile in order to predict the deformation of a large number of targets.

[0003] Current research on deformation prediction is mostly based on long-term time series forecasting, which uses the long-term morphological changes of the target as a feature to predict subsequent changes. It mainly includes adjustment theory, linear theory, nonlinear theory, and artificial intelligence. The first three are analytical methods, using the characteristics of the target itself and relevant professional theories to conduct detailed analysis. Artificial intelligence methods construct the problem as an approximate black-box model, mapping its inputs and outputs. Research on 3D data mainly includes multi-view methods, voxel-based methods, graph structure-based methods, and point set-based methods.

[0004] Due to the practical requirements of mine production, simply obtaining the surface shape changes of the blast pile is insufficient for automated production. The location and grade information of internal blocks are also necessary data for scheduling algorithms. Existing products for analyzing blast pile deformation mainly involve using positioning devices such as small balls buried in the blast pile before blasting and then locating and obtaining displacement after blasting. However, this method is extremely costly, the small balls are difficult to recover, and it can even cause some damage to the mining environment, making it unsuitable for large-scale use. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for predicting the deformation of a blasting pile based on a 3D convolutional neural network, comprising the following steps:

[0006] S1: Construct a 3D convolutional neural network, train the 3D convolutional neural network with training data, and obtain a trained 3D convolutional neural network.

[0007] S2: Obtain exploration data before blasting of the blast pile, preprocess the exploration data, and obtain the first matrix and the second matrix;

[0008] S3: Input the first and second matrices into the trained 3D convolutional neural network for prediction to obtain the deformation prediction results after the blasting of the blast pile.

[0009] Preferably, the 3D convolutional neural network includes: an input unit, an activation function unit, an eight-layer network structure, an attention module, and a fully connected layer;

[0010] The eight-layer network structure includes: the first ordinary unit, the second ordinary unit, the first dimensionality reduction unit, the third ordinary unit, the fourth ordinary unit, the second dimensionality reduction unit, the fifth ordinary unit, and the sixth ordinary unit;

[0011] The first output terminal of the input unit is connected to the input terminal of the activation function unit, and the second output terminal of the input unit is connected to the first input terminal of the first general unit;

[0012] The output of the activation function unit is connected to the second input of the first general unit and the first input of the second general unit.

[0013] The output of the first ordinary unit is connected to the second input of the second ordinary unit and the first input of the first dimension reduction unit;

[0014] The output of the second ordinary unit is connected to the second input of the first dimension reduction unit and the first input of the third ordinary unit;

[0015] The output of the first dimension reduction unit is connected to the second input of the third ordinary unit and the first input of the fourth ordinary unit;

[0016] The output of the third ordinary unit is connected to the second input of the fourth ordinary unit and the first input of the second dimension reduction unit;

[0017] The output of the fourth general unit is connected to the second input of the second dimension reduction unit and the first input of the fifth general unit;

[0018] The output of the second dimension reduction unit is connected to the second input of the fifth ordinary unit and the first input of the sixth ordinary unit;

[0019] The output of the fifth general unit is connected to the second input of the sixth general unit, and the output of the sixth general unit is connected to the input of the attention module.

[0020] The output of the attention module is connected to the input of the fully connected layer through global average pooling.

[0021] Preferably, the ordinary unit includes: a first input module c_{k-1}, a second input module c_{k-2}, a 0 node, a 1 node, a 2 node, a 3 node, and an output module c_{k};

[0022] The second input module c_{k-2} is connected to node 0 through the sep_conv_3×3 operation;

[0023] The first input module c_{k-1} is connected to node 0 through the sep_conv_5×5 operation, to node 1 through the dil_conv_3×3 operation, and to node 3 through the sep_conv_5×5 operation.

[0024] Node 0 has no operation and is connected to the output module c_{k}. It is connected to node 1 through the dil_conv_3×3 operation and to node 2 through the max_pool_3×3 operation.

[0025] Node 1 is not connected to the output module c_{k}, but is connected to Node 2 through the avg_pool_3×3 operation;

[0026] Node 2 is not connected to the output module c_{k}, but is connected to node 3 through the sep_conv_5×5 operation;

[0027] The 3-node has no operation or output module c_{k} connection.

[0028] Preferably, the dimensionality reduction unit includes: a first input module c_{k-1}, a second input module c_{k-2}, a 0 node, a 1 node, a 2 node, a 3 node, and an output module c_{k};

[0029] The second input module c_{k-2} is connected to node 0 via the skip_connect operation;

[0030] The first input module c_{k-1} is connected to node 0 through the sep_conv_3×3 operation, to node 1 through the max_pool_3×3 operation, to node 2 through the avg_pool_3×3 operation, and to node 3 through the max_pool_3×3 operation.

[0031] Node 0 has no operation and is not connected to the output module c_{k}, but is connected to node 1 through the dil_conv_5×5 operation;

[0032] Node 1 is not connected to the output module c_{k}, is connected to node 2 through the max_pool_3×3 operation, and is connected to node 3 through the sep_conv_3×3 operation;

[0033] Nodes 2 and 3 have no operation or output module c_{k} connection.

[0034] Preferably, step S2 specifically includes:

[0035] S21: Divide the blast pile into multiple segments using the exploration data before blasting, and construct a 3D matrix with 32*16*16 units;

[0036] S22: Place the explosive pile into a 3D matrix. Set the cells with segments to 1 and the cells without segments to 0 in the 3D matrix. Output the first matrix [a,x,y,z]. Obtain the hardness value of each segment and output the second matrix [b,x,y,z]. Where x,y,z correspond to the size of the 3D matrix 32*16*16, a takes the value of 1 or 0, and b is the hardness value.

[0037] Preferably, step S3 specifically includes:

[0038] S31: Input the first matrix and the second matrix into the trained 3D convolutional neural network, and perform convolution calculations on the first matrix and the second matrix in turn through each unit in the eight-layer network structure to obtain convolutional features;

[0039] S32: Input the convolutional features into the attention module, and add spatial weights and channel weights to the convolutional features in sequence to obtain weighted convolutional features;

[0040] S32: The weighted convolutional features are calculated through a fully connected layer to obtain the output vector;

[0041] S33: Construct a blasting coordinate system, obtain the coordinates of each segment in the blasting coordinate system through the output vector, and output the deformation prediction results after the blasting of the blast pile.

[0042] A blasting deformation prediction system based on a 3D convolutional neural network includes the following modules:

[0043] The network training module is used to build a 3D convolutional neural network. It trains the 3D convolutional neural network using training data to obtain a trained 3D convolutional neural network.

[0044] The data preprocessing module is used to acquire exploration data before the blasting of the blast pile, preprocess the exploration data, and obtain the first matrix and the second matrix.

[0045] The blasting prediction module is used to input the first matrix and the second matrix into a trained 3D convolutional neural network for prediction, and obtain the deformation prediction results after the blasting of the blast pile.

[0046] The present invention has the following beneficial effects:

[0047] By stacking ordinary units and dimensionality-reduced units to construct a 3D convolutional neural network, more diverse features can be obtained. Using the same structure for the same type of unit can simplify the structure of the 3D convolutional neural network and reduce the difficulty of construction. By constructing a first matrix with block information and a second matrix with hardness information from exploration data as input to the 3D convolutional neural network, the various characteristics of blasting pile blasting are fully reflected, and feature loss is avoided. Attached Figure Description

[0048] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0049] Figure 2 This is a structural diagram of a 3D convolutional neural network;

[0050] Figure 3 This is a structural diagram of a common unit;

[0051] Figure 4 This is a structural diagram of the dimensionality reduction unit;

[0052] Figure 5 Visualize the deformation prediction results;

[0053] Figure 6 A schematic diagram illustrating the linear correlation between predicted displacement and actual displacement;

[0054] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0055] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0056] Reference Figure 1 This invention provides a method for predicting the deformation of a blasting pile based on a 3D convolutional neural network, comprising the following steps:

[0057] S1: Construct a 3D convolutional neural network, train the 3D convolutional neural network with training data, and obtain a trained 3D convolutional neural network.

[0058] S2: Obtain exploration data before blasting of the blast pile, preprocess the exploration data, and obtain the first matrix and the second matrix;

[0059] S3: Input the first and second matrices into the trained 3D convolutional neural network for prediction to obtain the deformation prediction results after the blasting of the blast pile.

[0060] Furthermore, the 3D convolutional neural network includes: an input unit, an activation function unit, an eight-layer network structure, an attention module, and fully connected layers;

[0061] The eight-layer network structure includes: the first ordinary unit, the second ordinary unit, the first dimensionality reduction unit, the third ordinary unit, the fourth ordinary unit, the second dimensionality reduction unit, the fifth ordinary unit, and the sixth ordinary unit;

[0062] The first output terminal of the input unit is connected to the input terminal of the activation function unit, and the second output terminal of the input unit is connected to the first input terminal of the first general unit;

[0063] The output of the activation function unit is connected to the second input of the first general unit and the first input of the second general unit.

[0064] The output of the first ordinary unit is connected to the second input of the second ordinary unit and the first input of the first dimension reduction unit;

[0065] The output of the second ordinary unit is connected to the second input of the first dimension reduction unit and the first input of the third ordinary unit;

[0066] The output of the first dimension reduction unit is connected to the second input of the third ordinary unit and the first input of the fourth ordinary unit;

[0067] The output of the third ordinary unit is connected to the second input of the fourth ordinary unit and the first input of the second dimension reduction unit;

[0068] The output of the fourth general unit is connected to the second input of the second dimension reduction unit and the first input of the fifth general unit;

[0069] The output of the second dimension reduction unit is connected to the second input of the fifth ordinary unit and the first input of the sixth ordinary unit;

[0070] The output of the fifth general unit is connected to the second input of the sixth general unit, and the output of the sixth general unit is connected to the input of the attention module.

[0071] The output of the attention module is connected to the input of the fully connected layer through global average pooling.

[0072] Specifically, the structure of a 3D convolutional neural network is as follows: Figure 2 As shown, input is the input unit, A is the activation function unit, Cell is the unit, CBAM is the attention module, and FC layer is the fully connected layer;

[0073] Traditional network design typically involves designing each layer individually. However, this approach struggles to create diverse structures. Therefore, this invention employs a unit stacking method to construct the neural network. Instead of designing each layer separately, this method introduces two types of network units: ordinary units and dimensionality-reducing units. Ordinary units extract features without modifying the data size, while dimensionality-reducing units halve the data size and expand the output channels to eight times the number of input channels to capture more diverse features. Units of the same type use the same structure, which reduces the complexity of network construction to some extent.

[0074] After the network units are constructed, the complete network structure is built by stacking the units. At this time, due to the certain complexity of the units, and the fact that each unit is connected in a dual-input form, each unit takes the outputs of the first two units as its input. This can build a feature extraction network with rich connections and a certain width. The input of the first ordinary unit will be specially processed. One input is the output of the input unit, and the other needs to be the network input after passing through the activation function unit as the second input.

[0075] Since this invention does not require a high network depth, and excessively high network depth would lead to excessive computational complexity and deployment difficulty, the complete network of this invention also uses an eight-layer network structure. Dimensionality reduction units are used at 1 / 3 and 2 / 3 of the layers. Before flattening the features and inputting them into the fully connected layer, this network does not directly flatten the feature matrix, but instead inserts an attention module to strengthen the regions in the output of the convolutional network that have a significant impact on the result while weakening meaningless regions, thereby enhancing the accuracy of the network for this problem. Finally, global average pooling is performed on each channel to reduce the number of network parameters.

[0076] The fully connected layer has two hidden layers and one output layer. Since the network displacement can be negative, the hyperbolic tangent function (Tanh) is selected as the activation function to preserve the network's mapping effect on negative values. A dropout layer is used to randomly deactivate the neurons in the fully connected layer to avoid overfitting.

[0077] Since this invention flattens all block displacements during output, modeling the problem as a regression problem, the loss function used by the network is the mean squared error loss function, as shown in the following formula:

[0078]

[0079] During data preprocessing, this invention sets non-existent segments to 0. Similarly, in the output, in order to accommodate all segments, the output size is a 3*32*16*16 vector. To eliminate the influence of non-existent segments on the network, this invention adds a mask vector before the network output, setting the displacement data of all non-existent segments to 0, thereby making their contribution to the loss 0, and thus eliminating their influence on gradient backpropagation.

[0080] Furthermore, the parameters of this invention are set as follows: learning rate: 0.040, which gradually decreases to 0 with the number of training generations; training iterations: 100 times; batch size for batch training: 16. Setting the batch size too large will affect the distribution of data.

[0081] The network training of this invention used 2700 data points, of which 2200 were used as training data and 500 were used as validation data.

[0082] Further reference Figure 3 The ordinary unit includes: a first input module c_{k-1}, a second input module c_{k-2}, a 0 node, a 1 node, a 2 node, a 3 node, and an output module c_{k};

[0083] The second input module c_{k-2} is connected to node 0 through the sep_conv_3×3 operation;

[0084] The first input module c_{k-1} is connected to node 0 through the sep_conv_5×5 operation, to node 1 through the dil_conv_3×3 operation, and to node 3 through the sep_conv_5×5 operation.

[0085] Node 0 has no operation and is connected to the output module c_{k}. It is connected to node 1 through the dil_conv_3×3 operation and to node 2 through the max_pool_3×3 operation.

[0086] Node 1 is not connected to the output module c_{k}, but is connected to Node 2 through the avg_pool_3×3 operation;

[0087] Node 2 is not connected to the output module c_{k}, but is connected to node 3 through the sep_conv_5×5 operation;

[0088] The 3-node has no operation or output module c_{k} connection.

[0089] Further reference Figure 4 The dimensionality reduction unit includes: a first input module c_{k-1}, a second input module c_{k-2}, a 0 node, a 1 node, a 2 node, a 3 node, and an output module c_{k};

[0090] The second input module c_{k-2} is connected to node 0 via the skip_connect operation;

[0091] The first input module c_{k-1} is connected to node 0 through the sep_conv_3×3 operation, to node 1 through the max_pool_3×3 operation, to node 2 through the avg_pool_3×3 operation, and to node 3 through the max_pool_3×3 operation.

[0092] Node 0 has no operation and is not connected to the output module c_{k}, but is connected to node 1 through the dil_conv_5×5 operation;

[0093] Node 1 is not connected to the output module c_{k}, is connected to node 2 through the max_pool_3×3 operation, and is connected to node 3 through the sep_conv_3×3 operation;

[0094] Nodes 2 and 3 have no operation or output module c_{k} connection.

[0095] Specifically, in the design of ordinary and dimensionality-reduced units, each unit takes the outputs of the previous two units as input to form the outer model. Each unit has four nodes, which together form a directed acyclic graph. The parameters of each node are calculated from all its predecessor nodes using the following formula:

[0096]

[0097] Where j represents the current node number, i represents the predecessor node number, and o represents the connection operation.

[0098] The connections used in this invention mainly include the following:

[0099] Zero-connection, max pooling (3x3), average pooling (3x3), skip connections, split convolution (3x3), dilated convolution (3x3), etc.

[0100] Max pooling: max_pool, average pooling: avg_pool, skip connection: skip_connect, separating convolution: sep_conv, dilated convolution: dil_conv;

[0101] In the construction of the network, this invention avoids excessive use of skip connections and zero connections as much as possible in order to maintain the continuity of the network and the efficiency of computation.

[0102] Furthermore, step S2 specifically involves:

[0103] S21: Divide the blast pile into multiple segments using the exploration data before blasting, and construct a 3D matrix with 32*16*16 units;

[0104] S22: Place the explosive pile into a 3D matrix. Set the cells with segments to 1 and the cells without segments to 0 in the 3D matrix. Output the first matrix [a,x,y,z]. Obtain the hardness value of each segment and output the second matrix [b,x,y,z]. Where x,y,z correspond to the size of the 3D matrix 32*16*16, a takes the value of 1 or 0, and b is the hardness value.

[0105] Specifically, the actual information of the blast pile is obtained through the geological exploration work in the early stage of the mine, including the location, hardness and grade of each block of the blast pile. Here, the block refers to the smallest unit of ore block that is considered during mining operations. The actual size is designed according to the resolution required by the mine.

[0106] Finally, a 3D matrix is ​​modeled based on hardness and grade, typically limited to 32*16*16 pixels, with the actual spatial size determined by the block size. However, such neatly arranged block piles do not exist in actual mines, so another matrix is ​​needed to identify the locations of blocks, setting a value of 1 where a block exists and 0 otherwise. The first matrix [a,x,y,z] serves as the first input to the network. Furthermore, based on exploration data, approximate hardness and other information about the blocks at various locations can be obtained. This invention uses a second matrix [b,x,y,z] representing hardness and other information as the second input to the network, performing 3D spatial convolution operations. This can largely describe the determining features of the interactions between the various blocks in the model, allowing the 3D convolution to extract more spatial features.

[0107] After input, secondary analysis or ball positioning methods are used to obtain the displacement of each segment after blasting. Due to the change in the blast pile shape, deconvolution may exceed the matrix range. Therefore, the displacement data of all segments are flattened and used as the network output. During the evaluation of segment displacement data, since the data positions of each segment are fixed, we can obtain the displacement of each segment at once. This makes the grade prediction of the entire blast pile equivalent to the displacement prediction of each segment, laying the foundation for the present invention.

[0108] Furthermore, step S3 specifically includes:

[0109] S31: Input the first matrix and the second matrix into the trained 3D convolutional neural network, and perform convolution calculations on the first matrix and the second matrix in turn through each unit in the eight-layer network structure to obtain convolutional features;

[0110] S32: Input the convolutional features into the attention module, and add spatial weights and channel weights to the convolutional features in sequence to obtain weighted convolutional features;

[0111] S32: The weighted convolutional features are calculated through a fully connected layer to obtain the output vector;

[0112] S33: Construct a blasting coordinate system, obtain the coordinates of each segment in the blasting coordinate system through the output vector, and output the deformation prediction results after the blasting of the blast pile.

[0113] Specifically, since the network output of this invention uses flattened vectors, it is difficult to intuitively reflect the quality of the network and the actual shape of the blast pile after the explosion. Therefore, MATLAB is needed to process the network output and visualize the results. The visualized deformation prediction results are as follows: Figure 5 As shown.

[0114] To quantify the network results, this invention uses the correlation coefficient as an evaluation criterion to assess the degree of linear correlation between the target value and the predicted value. A schematic diagram illustrating the linear correlation between the predicted and actual displacements obtained experimentally is shown below. Figure 6 As shown in the experimental data, the linear correlation coefficient predicted by the network obtained by this invention in the simulated dataset is 0.942, indicating that this invention has a very significant prediction effect. Furthermore, the complete network model built by this invention has a parameter size of 192.9MB, which is less than 20% when using an NVIDIA RTX3090 GPU, making this invention quite easy to deploy.

[0115] A blasting deformation prediction system based on a 3D convolutional neural network includes the following modules:

[0116] The network training module is used to build a 3D convolutional neural network. It trains the 3D convolutional neural network using training data to obtain a trained 3D convolutional neural network.

[0117] The data preprocessing module is used to acquire exploration data before the blasting of the blast pile, preprocess the exploration data, and obtain the first matrix and the second matrix.

[0118] The blasting prediction module is used to input the first matrix and the second matrix into a trained 3D convolutional neural network for prediction, and obtain the deformation prediction results after the blasting of the blast pile.

[0119] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0120] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.

[0121] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting the deformation of a blasting reactor based on a 3D convolutional neural network, characterized in that, Including the following steps: S1: Construct a 3D convolutional neural network, train the 3D convolutional neural network with training data, and obtain a trained 3D convolutional neural network; S2: Obtain exploration data before the blasting of the blast pile, preprocess the exploration data, and obtain the first matrix and the second matrix; S3: Input the first and second matrices into the trained 3D convolutional neural network for prediction to obtain the deformation prediction results after the explosion of the blast pile. 3D convolutional neural networks include: input units, activation function units, an eight-layer network structure, attention modules, and fully connected layers; The eight-layer network structure includes: the first ordinary unit, the second ordinary unit, the first dimensionality reduction unit, the third ordinary unit, the fourth ordinary unit, the second dimensionality reduction unit, the fifth ordinary unit, and the sixth ordinary unit; The first to sixth ordinary units include: a first input module c_{k-1}, a second input module c_{k-2}, a 0 node, a 1 node, a 2 node, a 3 node, and an output module c_{k}; The second input module c_{k-2} is connected to node 0 through the sep_conv_3×3 operation; The first input module c_{k-1} is connected to node 0 through the sep_conv_5×5 operation, to node 1 through the dil_conv_3×3 operation, and to node 3 through the sep_conv_5×5 operation. Node 0 has no operation and is connected to the output module c_{k}. It is connected to node 1 through the dil_conv_3×3 operation and to node 2 through the max_pool_3×3 operation. Node 1 is not connected to the output module c_{k}, but is connected to Node 2 through the avg_pool_3×3 operation; Node 2 is not connected to the output module c_{k}, but is connected to node 3 through the sep_conv_5×5 operation; The 3-node has no connection to the output module c_{k}; The first and second dimensionality reduction units include: a first input module c_{k-1}`, a second input module c_{k-2}`, a 0` node, a 1` node, a 2` node, a 3` node, and an output module c_{k}`; The second input module c_{k-2}` is connected to the 0` node through the skip_connect operation; The first input module c_{k-1}` is connected to node 0` through the sep_conv_3×3 operation, to node 1` through the max_pool_3×3 operation, to node 2` through the avg_pool_3×3 operation, and to node 3` through the max_pool_3×3 operation. Node 0 is not connected to the output module c_{k}, but is connected to node 1 via the dil_conv_5×5 operation; Node 1 is not connected to the output module c_{k}, is connected to node 2 through the max_pool_3×3 operation, and is connected to node 3 through the sep_conv_3×3 operation; Nodes 2 and 3 have no operational connection to the output module c_{k}. Step S2 is as follows: S21: Using pre-blast exploration data, the blast pile is divided into multiple segments, constructing a 32... 16 A 3D matrix with 16 elements; S22: Place the explosive pile into a 3D matrix. Set the cells containing segments to 1 and the cells without segments to 0, outputting the first matrix [a,x,y,z]. Obtain the hardness value of each segment and output the second matrix [b,x,y,z]. Where x,y,z correspond to the size 32 of the 3D matrix. 16 16. The value of a is 1 or 0, and b is the hardness value.

2. The method for predicting blast deformation of explosive reactors based on 3D convolutional neural networks according to claim 1, characterized in that, The first output terminal of the input unit is connected to the input terminal of the activation function unit, and the second output terminal of the input unit is connected to the first input terminal of the first general unit; The output of the activation function unit is connected to the second input of the first ordinary unit and the first input of the second ordinary unit. The output of the first ordinary unit is connected to the second input of the second ordinary unit and the first input of the first dimension reduction unit; The output of the second ordinary unit is connected to the second input of the first dimension reduction unit and the first input of the third ordinary unit; The output of the first dimension reduction unit is connected to the second input of the third ordinary unit and the first input of the fourth ordinary unit; The output of the third ordinary unit is connected to the second input of the fourth ordinary unit and the first input of the second dimension reduction unit; The output of the fourth general unit is connected to the second input of the second dimension reduction unit and the first input of the fifth general unit; The output of the second dimension reduction unit is connected to the second input of the fifth ordinary unit and the first input of the sixth ordinary unit; The output of the fifth general unit is connected to the second input of the sixth general unit, and the output of the sixth general unit is connected to the input of the attention module. The output of the attention module is connected to the input of the fully connected layer through global average pooling.

3. The method for predicting blast deformation of explosive reactors based on 3D convolutional neural networks according to claim 2, characterized in that, Step S3 is as follows: S31: Input the first matrix and the second matrix into the trained 3D convolutional neural network, and perform convolution calculations on the first matrix and the second matrix in turn through each unit in the eight-layer network structure to obtain convolutional features; S32: Input the convolutional features into the attention module, and add spatial weights and channel weights to the convolutional features in sequence to obtain weighted convolutional features; S33: The weighted convolutional features are calculated through a fully connected layer to obtain the output vector; S34: Construct a blasting coordinate system, obtain the coordinates of each segment in the blasting coordinate system through the output vector, and output the deformation prediction results after the blasting of the blast pile.

4. A system for predicting the deformation of a blasting reactor based on a 3D convolutional neural network, characterized in that, include: The network training module is used to build a 3D convolutional neural network. It trains the 3D convolutional neural network using training data to obtain a trained 3D convolutional neural network. The data preprocessing module is used to acquire exploration data before the blasting of the blast pile, preprocess the exploration data, and obtain the first matrix and the second matrix. The blasting prediction module is used to input the first matrix and the second matrix into a trained 3D convolutional neural network for prediction, and obtain the deformation prediction result after the blasting of the blast pile. 3D convolutional neural networks include: input units, activation function units, an eight-layer network structure, attention modules, and fully connected layers; The eight-layer network structure includes: the first ordinary unit, the second ordinary unit, the first dimensionality reduction unit, the third ordinary unit, the fourth ordinary unit, the second dimensionality reduction unit, the fifth ordinary unit, and the sixth ordinary unit; The first to sixth ordinary units include: a first input module c_{k-1}, a second input module c_{k-2}, a 0 node, a 1 node, a 2 node, a 3 node, and an output module c_{k}; The second input module c_{k-2} is connected to node 0 through the sep_conv_3×3 operation; The first input module c_{k-1} is connected to node 0 through the sep_conv_5×5 operation, to node 1 through the dil_conv_3×3 operation, and to node 3 through the sep_conv_5×5 operation. Node 0 has no operation and is connected to the output module c_{k}. It is connected to node 1 through the dil_conv_3×3 operation and to node 2 through the max_pool_3×3 operation. Node 1 is not connected to the output module c_{k}, but is connected to Node 2 through the avg_pool_3×3 operation; Node 2 is not connected to the output module c_{k}, but is connected to node 3 through the sep_conv_5×5 operation; The 3-node has no connection to the output module c_{k}; The first and second dimensionality reduction units include: a first input module c_{k-1}`, a second input module c_{k-2}`, a 0` node, a 1` node, a 2` node, a 3` node, and an output module c_{k}`; The second input module c_{k-2}` is connected to the 0` node through the skip_connect operation; The first input module c_{k-1}` is connected to node 0` through the sep_conv_3×3 operation, to node 1` through the max_pool_3×3 operation, to node 2` through the avg_pool_3×3 operation, and to node 3` through the max_pool_3×3 operation. Node 0 is not connected to the output module c_{k}, but is connected to node 1 via the dil_conv_5×5 operation; Node 1 is not connected to the output module c_{k}, is connected to node 2 through the max_pool_3×3 operation, and is connected to node 3 through the sep_conv_3×3 operation; Nodes 2 and 3 have no operational connection to the output module c_{k}. The specific processing steps of the data preprocessing module are as follows: Based on the exploration data before the blasting, the blast pile was divided into multiple blocks, constructing a 32-block structure. 16 A 3D matrix with 16 elements; The explosive pile is placed in a 3D matrix. In this matrix, cells containing segments are set to 1, and cells without segments are set to 0, outputting the first matrix [a, x, y, z]. The hardness values ​​of each segment are then obtained, outputting the second matrix [b, x, y, z]. Here, x, y, and z correspond to the size 32 of the 3D matrix. 16 16. The value of a is 1 or 0, and b is the hardness value.

Images