Explosive and rock intelligent matching system based on neural network

Abstract

The invention relates to the technical field of rock blasting, in particular to an explosive and rock intelligent matching system based on a neural network, which comprises the following modules: a sample preprocessing module for training by using sample data through a certain learning rule to improve the adaptive capacity of the network; the model training module is used for constructing a tunnel blasting parameter optimization design model by adopting a three-layer neural network; the BP model training and optimizing module comprises a BP model training process and a BP model optimizing process; the RBF model training module comprises an RBF model training process and an RBF model optimization process; the beneficial effects are that the explosive and rock intelligent matching system based on the neural network employs an artificial intelligence neural network method to establish an explosive and rock matching model, the prediction result has a few errors of about 15%, other errors are less than 10%, and the average error is about 6%. And meanwhile, the matching time is within 45s, the matching efficiency is high, and the matching precision is accurate.

Description

Technical Field

[0001] The invention relates to the technical field of rock blasting, and in particular to an intelligent matching system of explosives and rocks based on a neural network. Background Art

[0002] In blasting engineering, improving the utilization rate of blasting energy is one of the key issues. Traditional blasting methods often rely on experience and trial and error to determine the type and amount of explosives, which is not only inefficient but also may lead to energy waste and safety hazards. In order to more effectively utilize blasting energy and consider factors such as blasting methods and blasting processes at the same time, a more scientific and accurate method is needed to match explosives with rocks.

[0003] In recent years, a number of authorized Chinese invention patents have disclosed the application of neural networks in explosive blasting. The specific patents are as follows: CN114186478B blasting block size prediction method based on RBF neural network; CN108399296B a vibration velocity prediction method for static blasting of foundation pits adjacent to structures; CN108007784B coupled fracture cavity volume visualization test system and fracture development analysis method; CN116992770B a wall protection and control blasting method based on GOA-DBN neural network; the application of legal and controllable neural networks in explosive blasting has become a hot research.

[0004] At present, the rapid development of artificial intelligence technology has provided new ideas for solving this problem. In particular, the artificial intelligence neural network method, with its powerful data processing and pattern recognition capabilities, has become a powerful tool for solving complex matching problems. By taking the rock property parameters obtained from rock mechanics tests and on-site blasting, the block size requirements required for blasting, and the blasting safety requirements as input, an intelligent matching system is constructed. The system can output mixed explosives with different performances required for blasting, thereby achieving the optimal matching of explosives and rocks.

[0005] In the process of realizing this intelligent matching system, the neural network model has become the first choice for system modeling because it can directly build models based on the input / output data of the object, without the need for complex knowledge and mathematical formula derivation, and can achieve high learning accuracy through appropriate training algorithms. At present, BP neural network and RBF neural network are two widely used neural network models. They have been proven to be able to approximate any nonlinear function, so they are very suitable for modeling nonlinear objects, such as the matching problem of explosives and rocks.

[0006] Based on the above background, this study uses BP neural network and RBF network to establish a blasting parameter optimization design model based on the preprocessing of raw data. This model aims to maximize the use of blasting energy through intelligent matching of explosives and rocks, while controlling the blasting hazards to the minimum range, thereby creating good economic and social benefits. Summary of the invention

[0007] The purpose of the present invention is to provide an intelligent matching system of explosives and rocks based on neural networks to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solution: an intelligent matching system of explosives and rocks based on neural network, the system comprising the following modules:

[0009] The sample preprocessing module uses sample data to train through certain learning rules to improve the adaptability of the network;

[0010] Model training module, which uses a three-layer neural network to construct a tunnel blasting parameter optimization design model;

[0011] BP model training and optimization module, including BP model training process and BP model optimization process;

[0012] The RBF model training module includes the RBF model training process and the RBF model optimization process.

[0013] Preferably, the sample preprocessing module normalizes the learning samples. The normalized data is easier for the neural network to train and learn. The normalized data is obtained by dividing the deviation between the sample value and the reference value under the corresponding conditions by the absolute value of the maximum deviation value to obtain the input variable.

[0014] Preferably, in order to provide a sufficient training set, an underground tunnel excavation blasting test is carried out, and blasting parameters, charging parameters, tunnel geometry parameters, and post-blasting vibration, over-break, under-break, and blockiness parameters are recorded before and after the test.

[0015] Preferably, the model training module comprehensively considers all factors in the construction of the model, taking into account the operation convergence speed and the possibility and reliability of obtaining these indicators on site, and determines the main variables affecting the blasting effect as input variables through analysis and screening. The designed three-layer neural network is: input layer, hidden layer and output layer.

[0016] Preferably, the BP model training process specifically includes:

[0017] The BP network propagates the input signal forward to the hidden layer nodes first, and then propagates the output information of the hidden layer nodes to the output nodes after the transfer function, and finally gives the output result. If there is an error between the output result and the expected output, it will turn to reverse propagation and return the error signal along the original connection path. By modifying the connection weights of the neurons in each layer, it will be recalculated until the error signal meets the requirements.

[0018] After the network training meets the requirements, network simulation is performed. The neural network prediction model needs to be given a certain amount of samples. The system learns on its own and obtains the optimal connection weights between the parameters. When new parameters are input, the system obtains the optimal output results based on the learning results. The prediction accuracy of the model improves with the increase in the number of samples.

[0019] Preferably, the BP model optimization process specifically includes:

[0020] The number of hidden layer units is directly related to the structure, training speed and prediction accuracy of the BP network, and the number of neurons in the hidden layer is directly related to the requirements of the problem and the number of input and output units.

[0021] Preferably, the RBF model training process is divided into two steps: the first step is unsupervised learning, determining the weights between the training input layer and the hidden layer; the second step is supervised learning, determining the weights between the training hidden layer and the output layer; before training, the input vector, the corresponding target vector and the expansion constant of the radial basis function are provided.

[0022] Preferably, the RBF model optimization process specifically includes:

[0023] Set SPREAD to 0.5, 1, and 2, and observe their effects on network performance through MATLAB programming to determine the optimal value.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The neural network-based intelligent matching system for explosives and rocks proposed in the present invention uses an artificial intelligence neural network method to establish an explosives and rock matching model. Except for a few errors of about 15%, the other errors in the prediction results are less than 10%, and the average error is about 6%. At the same time, the matching time is within 45 seconds, the matching efficiency is high, and the matching accuracy is accurate. It shows that the tunnel excavation blasting parameter optimization model established by the neural network method can be applied to actual engineering.

[0026] Both BP network and RBF network can complete effective simulation prediction in the tunnel excavation blasting parameter optimization model and achieve the expected results. BP network has more parameters to be adjusted, converges slowly, but has a small prediction error; RBF network has fewer parameters to be adjusted, converges quickly, and has a relatively large prediction error. Each has its own advantages. In this model, it can be seen that the prediction results of BP network are closer to the true value and have better effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the field test of tunnel blasting excavation of the present invention;

[0028] Figure 2 It is a schematic diagram of the block size test results of the original blasting scheme of the present invention;

[0029] Figure 3 This is a schematic diagram of the block size test results after the blasting scheme of the present invention is optimized;

[0030] Figure 4 It is a schematic diagram of the over-excavation and under-excavation test results of the original blasting scheme of the present invention;

[0031] Figure 5 This is a schematic diagram of the over-break and under-break test results after the blasting scheme of the present invention is optimized;

[0032] Figure 6 Schematic diagram of the three-layer training model of the present invention;

[0033] Figure 7 This is a flow chart of BP model training of the present invention;

[0034] Figure 8 This is a curve diagram of the network training error and the number of training samples of the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose and technical solution of the present invention clearly and completely described, and the advantages more clearly understood, the embodiments of the present invention are further described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present invention, rather than all of the embodiments, and are only used to explain the embodiments of the present invention, and are not used to limit the embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] See also Figures 1 to 8 The present invention provides a technical solution: an intelligent matching system of explosives and rocks based on a neural network, specifically comprising:

[0037] 1. Sample preprocessing module: After the network structure is determined, it is necessary to use sample data to train through certain learning rules to improve the adaptability of the network. However, the amplitude of the original data varies, and sometimes the difference is quite large. If it is put into use directly, the fluctuation of the larger sample value will monopolize the learning process of the neural network and make it unable to reflect the changes of the smaller sample value. Therefore, the learning samples need to be normalized. The normalized data is easier for the neural network to train and learn. The normalized data is obtained by dividing the deviation between the sample value and the benchmark value under the corresponding conditions by the absolute value of the maximum deviation value to obtain the input variable.

[0038] In order to provide enough training sets, dozens of underground tunnel excavation blasting tests were carried out in Kailin Qingcai Chong Mine. The blasting parameters, charging parameters, tunnel geometry parameters, vibration after blasting, over-excavation, under-excavation, and block size were recorded before and after the test. Some test results are listed in Table 1.

[0039] Table 1 Sample data of matching model obtained based on blasting field test

[0040]

[0041]

[0042] The test site is located in the north-south vein tunnel in the middle section of 600 underground. The tunnel size is 4.2m wide and 3.5m high. The number of blastholes is basically maintained between 46 and 47. The on-site test results are shown in Figure 1 Take the blasting effect obtained by part of the original blasting scheme and the blasting effect obtained after the blasting scheme is optimized, such as Figure 2 — Figure 5 As shown in the figure, the main blasting effects are blasting vibration, fragmentation and over-excavation and under-excavation.

[0043] This model is trained and optimized based on the results of a series of development tunnel blasting excavation tests carried out in the middle section of 600m in Guizhou Kailin Qingcai Chong Mine, and the blasting parameter optimization and blasting effect evaluation of rock tunnels are studied. The research results have a certain reference value for improving the blasting effect of rock tunnel excavation in similar mines, improving excavation efficiency, improving support effect and reducing the cost of rock tunnel excavation. A total of 14 blasting field tests were carried out in this project, of which 9 were conducted on the basis of the original mine plan, and 5 were blasting plans optimized according to the blasting effect obtained from the previous blasting plan. Five samples were used as samples for simulation and detection. Some data samples are shown in Table 2.

[0044] Table 2 Sample data sample

[0045]

[0046]

[0047] 2. Model training module: A three-layer neural network is used to construct a tunnel blasting parameter optimization design model. In the construction of the model, all factors must be considered comprehensively, while the convergence speed of the operation and the possibility and reliability of obtaining these indicators on site must also be considered. Through analysis and screening, the main variables that affect the blasting effect are determined as input variables. The designed three-layer neural network is: input layer, hidden layer and output layer. Through the above analysis, the tunnel excavation blasting design neural network model is as follows Figure 6 shown.

[0048] 3. BP model training and optimization module: (1) Training process: After the network is determined, it is more important to determine the network's learning and training parameters. These factors have a great impact on the performance of the network. Of course, these parameters need to be determined after debugging. The network parameters that are initially determined are shown in Table 3. The learning rate is an important factor in the training process, which determines the amount of weight change in each cycle. In general, a smaller learning rate is preferred to ensure the stability of learning. The network must also be debugged to select the best value. Here, the learning rate is taken as 0.05. The BP network first propagates the input signal forward to the hidden layer nodes, and after the transfer function, the output information of the hidden layer nodes is propagated to the output nodes, and finally the output result is given. If there is an error between the output result and the expected output, it will turn to back propagation and return the error signal along the original connection path. By modifying the connection weights of the neurons in each layer, it will be recalculated until the error signal meets the requirements.

[0049] Table 3 Network parameter settings

[0050] Training function Learning Function Performance functions Trainlm Learngdm Mse Number of training sessions Training Goals Learning Rate 1000 0.001 0.05

[0051] After the network training meets the requirements, network simulation can be performed. The neural network prediction model needs to be given a certain amount of samples. The system can learn on its own and obtain the optimal connection weights between the parameters. When new parameters are input, the system can obtain better output results based on the learning results. Generally speaking, the prediction accuracy of the model will increase with the increase in the number of samples.

[0052] (2) The number of hidden layer units in the optimization process is directly related to the structure, training speed and prediction accuracy of the BP network. It is often determined based on the designer's experience and multiple experiments. There is no ideal analytical expression to represent it. The number of neurons in the hidden layer is directly related to the requirements of the problem and the number of input and output units. Too many hidden layer units will lead to a long learning time and the error may not be optimal. It will also lead to poor fault tolerance and inability to recognize samples that have not been seen before. Therefore, there must be an optimal number of hidden units. Generally, there are two reference formulas for selecting the number of hidden layer units s:

[0053] 1) Kolmogorov theorem: that is, s = 2m + 1, where m is the number of input layer units;

[0054] 2) s = log m, where m is the number of input layer units;

[0055] The number of neurons in the input layer of the network is 14, and the number of neurons in the output layer is 14. According to the above three hidden layer design empirical formulas and considering the actual situation of this model, the number of neurons in the hidden layer to solve this problem should be between 10 and 30. The number of neurons in the hidden layer s is 5, 10, 15, 25, 35, 45, 76, and the optimal number of neurons is selected by comparing its convergence speed and training error; after programming and operation in MATLAB, the network training error can be calculated; the error between the simulation results and the target data is small, and there is a good correlation. The average error between the predicted value and the true value of the BP network model and the RBF network model is about 2%-13%. Except for the error of more than 24% caused by the small number of hidden layers, the others are all below 15%, which can meet the engineering requirements.

[0056] 4. RBF model training module: (1) If the training process achieves the same function, the number of neurons in the RBF network may be more than that in the BP network, but the training time required for the RBF network is less than that of the BP network. Since the establishment process of the RBF network is actually a learning and training process, the network model requires fewer network parameters to be adjusted. When the network is established, it will automatically select the optimal number of hidden layers to make the error meet the requirements. The only parameter that needs to be adjusted is the distribution density SPREAD of the RBF function. The larger the SPREAD, the smoother the function. Of course, it is necessary to select the best value by comparing the error.

[0057] The training process of RBF network is divided into two steps: the first step is unsupervised learning, which determines the weights between the training input layer and the hidden layer; the second step is supervised learning, which determines the weights between the training hidden layer and the output layer. Before training, the input vector, the corresponding target vector and the expansion constant of the radial basis function need to be provided; (2) Optimization process The prediction accuracy of BP network is relatively high, but the training error converges slowly and the calculation time is long. Therefore, based on the same background, we try to use RBF network for prediction. RBF network requires very few variables. The only thing that needs to be determined is the distribution density SPREAD of the radial basis function. Since the choice of SPREAD has a relatively important influence on the network performance, the larger the SPREAD, the smoother the function fitting, but too many neurons are needed to adapt to the rapid changes of the function, which may cause network calculation difficulties: if the SPREAD is set too small, fewer neurons are required, and the network performance will not be very good at that time. Therefore, here we try to use multiple different SPREAD values ​​to determine an optimal value. According to experience, SPREAD is first set to 0.5, 1, and 2. Through MATLAB programming, we observe their effects on network performance to determine the optimal value.

[0058] Depend on Figure 8 It can be seen that as the distribution density decreases, the network error decreases. When the distribution density SPREAD is 1, the network training error is the smallest. Therefore, considering the convergence speed and calculation reasons, the distribution density SPREAD is selected as 0.5. At the same time, the model prediction results are shown in Table 4.

[0059] Table 4 Model prediction results

[0060]

[0061]

[0062] Table 4 shows that: (1) The artificial intelligence neural network method is used to establish the explosive and rock matching model. Except for a few prediction results with errors of about 15%, the other errors are less than 10%, and the average error is about 6%. At the same time, the matching time is within 45 seconds, the matching efficiency is high, and the matching accuracy is accurate, indicating that the tunnel excavation blasting parameter optimization model established by the neural network method can be applied to actual engineering;

[0063] (2) Both BP network and RBF network can complete effective simulation prediction in the tunnel excavation blasting parameter optimization model and achieve the expected effect. BP network has more parameters to be adjusted, converges slowly, but has a small prediction error; RBF network has fewer parameters to be adjusted, converges quickly, and has a relatively large prediction error. Each has its own advantages. In this model, it can be seen that the prediction results of BP network are closer to the true value and have better effects.

[0064] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The intelligent matching system of explosives and rocks based on neural network is characterized by: The system includes the following modules: The sample preprocessing module uses sample data to train through certain learning rules to improve the adaptability of the network; Model training module, which uses a three-layer neural network to construct a tunnel blasting parameter optimization design model; BP model training and optimization module, including BP model training process and BP model optimization process; The RBF model training module includes the RBF model training process and the RBF model optimization process.

2. The explosive and rock intelligent matching system based on neural network according to claim 1 is characterized in that: The sample preprocessing module performs normalization processing on the learning samples. The normalized data is easier for the neural network to train and learn. The normalized data is obtained by dividing the deviation between the sample value and the reference value under the corresponding conditions by the absolute value of the maximum deviation value to obtain the input variable.

3. The neural network-based intelligent matching system for explosives and rocks according to claim 2 is characterized in that: In order to provide sufficient training sets, underground tunnel excavation blasting tests were carried out. The blasting parameters, charging parameters, tunnel geometry parameters, vibration after blasting, over-excavation, under-excavation, and block size parameters were recorded before and after the tests.

4. The neural network-based intelligent matching system for explosives and rocks according to claim 1 is characterized in that: The model training module comprehensively considers all factors in the construction of the model, taking into account the operation convergence speed and the possibility and reliability of obtaining these indicators on site. Through analysis and screening, the main variables affecting the blasting effect are determined as input variables, and the designed three-layer neural network is: input layer, hidden layer and output layer.

5. The neural network-based intelligent matching system for explosives and rocks according to claim 4 is characterized in that: The BP model training process specifically includes: The BP network propagates the input signal forward to the hidden layer nodes first, and then propagates the output information of the hidden layer nodes to the output nodes after the transfer function, and finally gives the output result. If there is an error between the output result and the expected output, it will turn to reverse propagation and return the error signal along the original connection path. By modifying the connection weights of the neurons in each layer, it will be recalculated until the error signal meets the requirements. After the network training meets the requirements, network simulation is performed. The neural network prediction model needs to be given a certain amount of samples. The system learns on its own and obtains the optimal connection weights between the parameters. When new parameters are input, the system obtains the optimal output results based on the learning results. The prediction accuracy of the model improves with the increase in the number of samples.

6. The neural network-based intelligent matching system for explosives and rocks according to claim 5 is characterized in that: The BP model optimization process specifically includes: The number of hidden layer units is directly related to the structure, training speed and prediction accuracy of the BP network, and the number of neurons in the hidden layer is directly related to the requirements of the problem and the number of input and output units.

7. The neural network-based intelligent matching system for explosives and rocks according to claim 1 is characterized in that: The RBF model training process is divided into two steps: the first step is unsupervised learning, which determines the weights between the training input layer and the hidden layer; the second step is supervised learning, which determines the weights between the training hidden layer and the output layer; before training, the input vector, the corresponding target vector and the expansion constant of the radial basis function are provided.

8. The neural network-based intelligent matching system for explosives and rocks according to claim 1 is characterized in that: The RBF model optimization process specifically includes: Set SPREAD to 0.5, 1, and 2, and use MATLAB programming to observe their effects on network performance to determine the optimal value.

Images