Underground cavern group rock drilling efficiency prediction method and equipment based on deep learning
Through a deep learning-based method, a backpropagation neural network drilling efficiency prediction model is established, and the model structure is optimized using bacterial foraging optimization algorithm, which solves the accuracy of drilling efficiency prediction in the existing technology and achieves efficient prediction under complex conditions.
Patent Information
- Application Number
- CN202510125724.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-27
- Publication Date
- 2025-05-30
AI Technical Summary
The prior art is difficult to accurately predict the efficiency of rock drilling in underground cavities, especially under complex nonlinear relationships and a variety of influencing factors.
A deep learning-based method is adopted to establish a drilling efficiency prediction model through backpropagation neural networks, and the hidden layer structure hyperparameters of the model are optimized using a bacterial foraging optimization algorithm to improve prediction accuracy.
Accurate prediction of the efficiency of drilling drilling in underground cave chamber groups under complex conditions, and improve the accuracy of construction progress control and equipment personnel allocation.
Smart Images
Figure CN120069191A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of water conservancy and hydropower construction, and particularly relates to a method and device for predicting the rock drilling efficiency of an underground cavern group based on deep learning. Background Technique
[0002] At present, the underground cavern group of the hydropower station project was first built in the 1890s of the 19th century. The Vernayaz Hydropower Station in Switzerland first adopted an underground power plant with an installed capacity of 5,100 kilowatts and was completed and put into production in 1897. Subsequently, due to the advantages of the underground cavern group not occupying ground positions, having less interference with the construction of surface hydraulic structures, and having a shorter construction period, it has been increasingly used in large-scale hydropower projects.
[0003] The construction of the underground cavern group is a cyclic process composed of various processes such as drilling, blasting, ventilation and smoke dispersion, safety inspection, and mucking.
[0004] To ensure the smooth progress of construction, structural stability, and operation safety, rock drilling construction under different conditions needs to be carried out in the underground cavern group. The exploration holes are used for detailed exploration of underground rock formations and groundwater conditions to provide necessary data support; the blasting holes are used for rock blasting to make the rock easy to excavate; the support holes are used for installing anchor bolts and cables to enhance the stability of the rock formation and prevent collapse; the drainage holes are used to drain groundwater to prevent seepage from affecting construction; the ventilation holes provide fresh air to ensure the safety of the construction environment; the observation holes are used to monitor the deformation and stress changes of the cavern to ensure long-term safe operation. The various types of drilling holes cooperate with each other to form an important technical means for the construction of the underground cavern group. For the needs of various underground cavern group rock drilling constructions, predicting the drilling efficiency under different influencing conditions can effectively control the drilling time, thereby controlling the progress of the excavation and drilling construction, so as to achieve the purpose of selecting appropriate drilling equipment and personnel allocation according to the progress.
[0005] Some studies on drilling efficiency prediction are based on establishing mathematical models. These models mostly use auxiliary parameters such as bit characteristics, bit diameter, mud properties, and rotational speed to explore the correlation between drilling-related parameters and drilling efficiency and establish a multiple regression analysis mathematical model. For example, Kahraman discovered the correlation between rock brittleness and drilling rig performance by summarizing the original data of the experimental work of other researchers; Akun and Karpuz derived an empirical correlation model that can predict the penetration rate of diamond bits on the surface of sandstone; H. Abbaspour et al. adopted the formula proposed by Hustrulid to predict the drilling efficiency in drilling operations for subsequent optimization of the operation progress of drilling and blasting processes. The above studies usually consider the "average" properties of rocks. To consider the influence of the uncertainty of rock properties, Saeidi et al. developed a non-linear multiple regression prediction stochastic model for rotary drilling rig efficiency using the Monte Carlo method. In addition, Mustafa et al. applied the response surface method to establish the mathematical relationship between controllable drilling parameters and the penetration rate of the drilling rig to address the simultaneous influence of controllable parameters (such as Weight On Bit (WOB), revolutions per minute, and Flow rates (FR)) on drilling efficiency.
[0006] However, due to the highly non-linear or other complex relationships between drilling-related parameters and drilling efficiency, mathematical models cannot comprehensively and accurately estimate drilling efficiency. In addition, most mathematical models focus on the drilling rig itself and have not quantitatively considered the influence of other factors (such as construction environment and rock properties). With the development of machine learning technology, some researchers have tried various attempts using machine learning methods.
[0007] Deep learning is a new research direction in the field of machine learning and belongs to a type of supervised learning in machine learning, which includes various algorithm models such as neural networks, deep neural networks, and deep reinforcement learning. With the booming development of deep learning, modern deep learning algorithms have surpassed the prediction and classification accuracy of traditional machine learning algorithms for data. Deep learning automatically screens data and extracts high-dimensional features in the data instead of manually extracting features. For example: using deep learning methods to optimize construction project scheduling, combining deep learning with Building Information Modeling (BIM), and using the extraction, induction, and non-linear modeling capabilities of convolutional neural networks to establish a BIM model to achieve project automation control; adopting hybrid deep learning (a hybrid of deep AE-DT and AE-ELM), deep AE, and machine learning (Decision Tree, DT, and ELM) algorithms to predict the compressive strength of nano-silica modified engineering cementitious composites at high temperatures; using the collected data to construct a deep learning model for evaluating the seismic capacity of buildings, forming a method for evaluating the seismic capacity of individual buildings based on deep learning methods; adopting an improved long short-term memory neural network to predict the cable crane operation time parameters in the high arch dam construction simulation model and analyzing its local non-linear and volatility change characteristics; establishing a deep neural network model based on transfer learning to handle the non-linear relationship between complex climate and reservoir surface water temperature and predicting the reservoir surface water temperature.
[0008] In summary, using deep learning technology can avoid the incompleteness of the mathematical linear programming model in predicting drilling efficiency. At the same time, drilling operations and environmental factors have a great impact on the prediction of drilling efficiency. For example, precipitation will affect the moisture content of the rock wall in the underground chamber and thus affect its rock mass strength, and different numbers of operators will significantly affect the time for adding or unloading drill pipes, thereby affecting drilling efficiency. Therefore, the present invention proposes a deep learning prediction model for the non-linear and complex drilling efficiency, enabling it to solve the problem of predicting the rock drilling efficiency in underground chambers under the influence of various complex factors. Summary of the Invention
[0009] The present invention provides a method and device for predicting the rock drilling efficiency of underground chamber groups based on deep learning to solve the technical problems existing in the known technology.
[0010] The technical solution adopted by the present invention to solve the technical problems existing in the known technology is:
[0011] A method for predicting the rock drilling efficiency of underground chamber groups based on deep learning includes the following steps:
[0012] Step 1, collect relevant data on the drilling efficiency of underground chamber groups and its influencing factors, and use the collected data as the original data;
[0013] Step 2: Preprocess the collected raw data and compile a sample data set.
[0014] Step 3: Establish a drilling efficiency prediction model based on a backpropagation neural network; use the sample data set to train the drilling efficiency prediction model.
[0015] Step 4: Use the BFO algorithm to optimize the hyperparameters of the hidden layer structure in the drilling efficiency prediction model.
[0016] Step 5: Monitor the drilling efficiency of the underground cavern group and its influencing factors during the construction process; preprocess the monitored data, and then use the optimized drilling efficiency prediction model to predict the drilling efficiency during the construction process.
[0017] Furthermore, in Step 1, the method for collecting relevant data on the drilling efficiency of the underground cavern group and its influencing factors includes the following method steps: Obtain the following relevant data on the drilling operation of the underground cavern group through on-site investigation and historical records: drilling time, drilling depth, bit wear condition, drilling angle, drilling diameter, drilling speed, geological conditions, equipment performance, and personnel configuration of the operators; the data sources include construction logs, equipment records, and geological exploration reports.
[0018] During the data collection process, first record the basic information, including year, month, day, and the start and end times of each drilling process. At the same time, record the number and time of adding and unloading drill pipes for each drill rig, the coordinates, elevation, and depth of the borehole, the type of rock, and the type of drill rig. Secondly, the meteorological monitoring station automatically records a set of meteorological data every 10 minutes. Match the drilling data with the meteorological data, and find the meteorological data at or near the drilling time in the meteorological database and merge it into the drilling data to obtain the weather conditions corresponding to each borehole. Finally, integrate the data from different sources, structure the independent variables affecting the drilling efficiency, and finally obtain a sample data set containing target variables and feature variables that can be directly input into the drilling efficiency prediction model.
[0019] Furthermore, Step 2 includes the following method steps:
[0020] Step 2-1: Clean the raw data, delete duplicate data, and handle outliers and missing values.
[0021] Step 2-2: Quantify the non-numerical data.
[0022] Step 2-3: Standardize all the data obtained after being processed in Step 2-1 and Step 2-2 so that each numerical column in the data set has the same scale.
[0023] Step 2-4: Extract and select features from the data processed in Step 2-3; by analyzing the relationship between each feature and the drilling efficiency, extract the key features that have a greater impact on the drilling efficiency; organize these features and the corresponding drilling efficiency into a structured sample data set as the data input for training the drilling efficiency prediction model.
[0024] Furthermore, in Step 2-2, the method for quantifying non-numerical data includes the following method steps:
[0025] The non-numerical data includes categorical feature data, and artificial dummy variables are used to mark different categories of each feature.
[0026] Furthermore, in Step 3, a drilling efficiency prediction model is constructed based on a multi-layer perceptron regression model using the backpropagation algorithm.
[0027] Furthermore, in Step 3, the method for training the drilling efficiency prediction model using the sample data set includes the following method steps:
[0028] Step 3-1: Divide the sample data set into five equal parts for five-fold cross-validation, select four of them as the training set, and one as the validation set;
[0029] Step 3-2: Initialize the parameters of the training drilling efficiency prediction model, and the hyperparameters of the drilling efficiency prediction model are set to fixed values during initialization; use the training set for fitting and adjust the parameters of the drilling efficiency prediction model to minimize the loss function;
[0030] Step 3-3: After the training of the drilling efficiency prediction model is completed, use the validation set and the cross-validation method to evaluate the model performance.
[0031] Furthermore, Step 3-2 includes the following method steps:
[0032] By calculating the error between the predicted value and the true value, adjust the weights and biases of the model; use the gradient descent method to minimize the prediction error and gradually improve the prediction accuracy of the model; during the training process, use the cross-validation method to evaluate the model performance and prevent overfitting.
[0033] Furthermore, Step 3-3 includes the following method steps: Reverse normalize the predicted value and the true value, convert them back to the range of the original data, and output the model training and validation results.
[0034] Furthermore, in Step 4, the number of hidden layers of the model and the number of neurons in each layer are optimized by the BFO algorithm to improve the prediction performance of the model on the given data set. Before applying the algorithm, the maximum number of hidden layers and the number of neurons in each layer are specified first so that the training of the model can find the optimal parameters within a certain time;
[0035] In the initialization stage of the BFO algorithm, a group of initial bacterial populations are randomly generated, and each bacterium represents a neuron in the hidden layer of a neural network model; the position of each bacterium is the position of its corresponding neuron in the hidden layer structure, and the bacterium position is randomly initialized; other optimization parameters are artificially initialized;
[0036] In the optimization process of the BFO algorithm, each bacterium undergoes multiple fine-tuning and selection-elimination operations; the fine-tuning operation refers to the operation of increasing or decreasing the number of neurons in the hidden layer structure corresponding to the bacterium, in order to expect to find a better structure; this process simulates the swimming process of bacteria in the environment to find the most suitable position to survive; after fine-tuning, the fitness of each bacterium on the training set is calculated, that is, the performance index of the model; the mean square error is used as the fitness function; if the fitness of the fine-tuned bacterium is better, its position is updated to the new fine-tuned structure;
[0037] After each iteration, the BFO algorithm sorts the bacteria according to their fitness and selects half of the bacteria with better fitness for replication to maintain the population size; the other half of the bacteria are re-initialized to new positions according to a certain elimination probability to increase the diversity of the population; after all iterations are completed, the bacterium with the smallest mean square error among all bacteria is selected as the best model, and the hidden layer structure is determined.
[0038] The present invention also provides a device for a method for predicting the rock drilling efficiency of an underground cavern group based on deep learning, including a memory and a processor, where the memory is used to store a computer program; the processor is used to execute the computer program and implement the steps of the method for predicting the rock drilling efficiency of an underground cavern group based on deep learning as described above when executing the computer program.
[0039] The advantages and positive effects of the present invention are:
[0040] 1. By analyzing the influencing factors of the drilling efficiency of the underground cavern group, characteristic variables that have a greater impact on the drilling efficiency are selected, and non-numerical influencing factors are processed by quantification to obtain a training data set available for the training model.
[0041] 2. A multi-layer perceptron regression (MLP Regressor) model that utilizes the basic principle of the BPNN model is used to build a regression analysis training model, and regression model evaluation indicators are selected to evaluate the subsequent training effect of the model.
[0042] 3. For the problem of adjusting and optimizing the model hyperparameters, the bacterial foraging optimization (BFO) algorithm is selected to optimize the model hyperparameters. After comparing and analyzing the model hyperparameters, the hidden layer structure is selected as the optimization object and the optimization algorithm is successfully embedded in the original code model to improve the training effect of the model. Description of the Drawings
[0043] Figure 1 is a schematic flow chart of a method for predicting the rock drilling efficiency of an underground cavern group based on deep learning according to the present invention.
[0044] Figure 2 is a schematic diagram of the operating principle of a backpropagation neural network.
[0045] In the figure: X 1 ~X n : Input layer data; Y 1 ~Y n : Output layer data.
[0046] The arrow direction is the transmission direction of the neural network weights and biases. Detailed Embodiment
[0047] The present invention will be described in detail below with reference to the drawings and in conjunction with embodiments. It should be understood that the preferred embodiments described herein are only for illustrating and explaining the present invention, and are not used to limit the present invention.
[0048] The following are the Chinese interpretations of the following English words, phrases and abbreviations:
[0049] MSE: Mean Square Error.
[0050] BFO: Bacterial Foraging Optimization.
[0051] bfoa_optimization: The function name of the bacterial foraging optimization algorithm in the code.
[0052] MLP Regressor: Multi-Layer Perceptron Regressor.
[0053] BPNN(Back propagation neural network): Backpropagation Neural Network.
[0054] scikit-learn: A machine learning software package for the Python programming language.
[0055] Please refer to Figures 1 to 2 , a method for predicting the rock drilling efficiency of an underground cavern group based on deep learning, comprising the following steps:
[0056] Step 1, collect relevant data on the drilling efficiency of the underground cavern group and its influencing factors, and use the collected data as the original data;
[0057] Step 2, preprocess the collected original data to compile a sample data set;
[0058] Step 3: Establish a drilling efficiency prediction model based on a backpropagation neural network; train the drilling efficiency prediction model using the sample data set;
[0059] Step 4: Optimize the hyperparameters of the hidden layer structure in the drilling efficiency prediction model using the BFO algorithm;
[0060] Step 5: Monitor the drilling efficiency of the underground cavern group and its influencing factors during the construction process; preprocess the monitored data, and then predict the drilling efficiency during the construction process through the optimized drilling efficiency prediction model.
[0061] Preferably, in Step 1, the method for collecting the relevant data of the drilling efficiency of the underground cavern group and its influencing factors may include the following method steps: Obtain the following relevant data of the underground cavern group drilling operation through on-site investigation and historical records: drilling time, drilling depth, bit wear condition, drilling angle, drilling diameter, drilling speed, geological conditions, equipment performance, and personnel allocation of the operators; the data sources include construction logs, equipment records, and geological exploration reports;
[0062] During the data collection process, first record the basic information, including year, month, day, and the start and end times of each drilling process. At the same time, record the number and time of adding and unloading drill pipes for each drill rig, the coordinates, elevation, and depth of the drill hole, the type of rock, and the type of drill rig; secondly, the meteorological monitoring station automatically records a set of meteorological data every 10 minutes, match the drilling data and meteorological data in terms of time, find the meteorological data at or near the drilling moment in the meteorological database and merge it into the drilling data to obtain the weather conditions corresponding to each drill hole; finally, integrate the data from different sources, structure the independent variables affecting the drilling efficiency, and finally obtain a sample data set containing target variables and feature variables that can be directly input into the drilling efficiency prediction model.
[0063] Preferably, Step 2 may include the following method steps:
[0064] Step 2-1: Clean the original data, delete duplicate data, and process outliers and missing values;
[0065] Step 2-2: Quantify the non-numerical data;
[0066] Step 2-3: Standardize all the data obtained after being processed in Step 2-1 and Step 2-2 so that each numerical column of the data set has the same scale;
[0067] Step 2-4: Extract and select features from the data processed in Step 2-3; by analyzing the relationship between each feature and the drilling efficiency, extract the key features that have a greater impact on the drilling efficiency; organize these features and the corresponding drilling efficiency into a structured sample data set as the data input for training the drilling efficiency prediction model.
[0068] Preferably, in Step 2-2, the method for quantifying non-numerical data may include the following method steps:
[0069] The non-numerical data includes categorical feature data, and artificial dummy variables are used to mark different categories of each feature.
[0070] Preferably, in Step 3, a drilling efficiency prediction model can be constructed based on a multi-layer perceptron regression model using the backpropagation algorithm.
[0071] Preferably, in Step 3, the method for training the drilling efficiency prediction model using the sample data set may include the following method steps:
[0072] Step 3-1: Divide the sample data set into five equal parts for five-fold cross-validation, select four of them as the training set, and one as the validation set;
[0073] Step 3-2: Initialize the parameters of the training drilling efficiency prediction model, and the hyperparameters of the drilling efficiency prediction model are set to fixed values during initialization; use the training set for fitting, and adjust the parameters of the drilling efficiency prediction model to minimize the loss function;
[0074] Step 3-3: After the training of the drilling efficiency prediction model is completed, use the validation set and the cross-validation method to evaluate the model performance.
[0075] Preferably, Step 3-2 may include the following method steps:
[0076] By calculating the error between the predicted value and the true value, adjust the weights and biases of the model; use the gradient descent method to minimize the prediction error and gradually improve the prediction accuracy of the model; during the training process, use the cross-validation method to evaluate the model performance to prevent overfitting.
[0077] Preferably, Step 3-3 includes the following method steps: Reverse normalize the predicted value and the true value, convert them back to the range of the original data, and output the model training and validation results.
[0078] Preferably, in Step 4, the number of hidden layers of the model and the number of neurons in each layer can be optimized through the BFO algorithm to improve the prediction performance of the model on the given data set. Before applying the algorithm, the maximum number of hidden layers and the number of neurons in each layer are first specified so that the training of the model can find the optimal parameters within a certain time;
[0079] In the initialization stage of the BFO algorithm, a group of initial bacterial populations can be randomly generated, and each bacterium represents a neuron in the hidden layer of a neural network model; the position of each bacterium is the position of its corresponding neuron in the hidden layer structure, and the bacterium position is randomly initialized; other optimization parameters are artificially initialized;
[0080] In the optimization process of the BFO algorithm, each bacterium undergoes multiple fine-tuning and selection-elimination operations; the fine-tuning operation refers to the operation of increasing or decreasing the number of neurons in the hidden layer structure corresponding to the bacterium, in order to expect to find a better structure; this process simulates the swimming process of bacteria in the environment to find the most suitable position for survival; after fine-tuning, calculate the fitness of each bacterium on the training set, that is, the performance index of the model; the mean square error is used as the fitness function; if the fitness of the bacterium after fine-tuning is better, update its position to the new fine-tuned structure;
[0081] After each iteration, the BFO algorithm sorts the bacteria according to their fitness and selects half of the bacteria with better fitness for replication to maintain the population size; the other half of the bacteria are re-initialized to new positions according to a certain elimination probability to increase the diversity of the population; after all iterations are completed, the bacterium with the smallest mean square error among all bacteria is selected as the best model, and the hidden layer structure is determined.
[0082] The BFO algorithm may include the following sub-steps:
[0083] Step 4-1, let: the total number of bacterial chemotaxis be N c , the maximum number of steps for a bacterium to move forward in one direction of chemotaxis be N s , the total number of bacterial replications be N re , the total number of bacterial migrations be N ed , the migration probability of bacteria be p ed , the size of the bacterial population be S, and in order to meet the need of reproduction by taking half, this value is taken as an even number;
[0084] Let: each bacterium represents a hidden layer of the drilling efficiency prediction model; the position of each bacterium is its corresponding hidden layer; i represents the bacterium serial number, j represents the bacterium chemotaxis number serial number, k represents the bacterium replication number serial number, f represents the bacterium migration number serial number; P(i, j, k, f) represents the position of the i-th bacterium at the j-th bacterial chemotaxis, the k-th bacterial replication, and the f-th bacterial migration;
[0085] Initialize the following parameters to be optimized, let j = 0, k = 0, f = 0, S = 4, fine-tuning step size = 0.01;
[0086] Step 4-2, randomly initialize the bacterium positions and calculate the initial fitness value J of the bacteria; use the following formula to generate the positions of the initial bacteria:
[0087] X = x min + rand() × (x max - x min ) (1);
[0088] Where:
[0089] X is the position vector of the bacteria, and the position of each bacterium can be regarded as a representation of a hidden layer structure;
[0090] x min is the lower bound of the position vector X, representing the minimum number of hidden layer neurons;
[0091] x max is the upper bound of the position vector X, representing the maximum number of hidden layer neurons;
[0092] rand() represents a random number uniformly distributed between [0 - 1];
[0093] Step 4 - 3, Bacterial migration operation loop: f = f + 1;
[0094] Step 4 - 4, Bacterial replication operation loop: k = k + 1;
[0095] Step 4 - 5, Bacterial chemotaxis operation loop: j = j + 1;
[0096] The bacterial chemotaxis operation includes tumbling and swimming, and the chemotaxis operation is carried out as follows:
[0097] Tumbling: Update the bacterial position according to the following formula:
[0098]
[0099] Where:
[0100] P(i, j + 1, k, f) is the position of the i - th bacterium at the (j + 1) - th bacterial chemotaxis, the k - th bacterial replication, and the f - th bacterial migration;
[0101] C(i) is the step size of the i - th bacterium swimming forward in the solution space;
[0102] Δ(t) is a random direction vector, whose dimension is the same as that of the position P(i, j, k, f), usually generated by a random number generator, and is used to simulate the direction of random movement of bacteria;
[0103] Δ x (i) is the random change value of the i - th bacterium in a certain specific dimension, and it may be a single random number, used to represent the random step size of the bacterium moving in this dimension;
[0104] Δ(i) is a unit vector in a random direction when the i-th bacterium flips;
[0105] Calculate the fitness value after flipping;
[0106] Swimming: If the fitness becomes better after flipping, swim in this direction until the fitness no longer improves or reaches the set maximum swimming step number N s ; If the fitness does not become better, stop swimming;
[0107] Calculate the fitness value after swimming;
[0108] The evaluation index of fitness selects MSE in the model evaluation index. When the MSE at the bacterium position after flipping is less than the existing position, swim; if the MSE after flipping is greater than the existing position, stop swimming;
[0109] Step 4 - 6, if j < N c , jump to Step 4 - 5; In this case, continue the chemotaxis operation of the bacterium because the life of the bacterium has not ended;
[0110] Step 4 - 7, perform the bacterium replication operation. If k < N re , jump to Step 4 - 4; In this case, the set replication quantity has not been reached. Therefore, start the next cycle of the bacterium replication operation to produce offspring; Sort the bacterium fitness values, eliminate the S / 2 bacteria with poorer fitness values before extinction, and replicate the S / 2 bacteria with better fitness values. Each bacterium divides into two identical bacteria;
[0111] Step 4 - 8, perform the bacterium migration operation. After the bacterium completes the replication operation, each bacterium is randomly redistributed into the optimization space with probability p ed ; i = 1, 2, …… S, initialize in the solution space according to formula (1); If f < N ed , jump to Step 4 - 3;
[0112] Step 4 - 9, the loop ends, and the result is output.
[0113] The present invention also provides a device for a method for predicting the rock drilling efficiency of an underground cavern group based on deep learning, including a memory and a processor. The memory is used to store a computer program; the processor is used to execute the computer program and, when executing the computer program, implement the steps of the method for predicting the rock drilling efficiency of an underground cavern group based on deep learning as described above.
[0114] The following further illustrates the working process and working principle of the present invention with a preferred embodiment of the present invention:
[0115] As an important part of the rock drilling construction technology for underground cavern groups, the prediction of drilling efficiency is of great significance. With the development of artificial intelligence technology, more and more intelligent analysis techniques have been applied in engineering simulation. As an indispensable data processing method in machine learning intelligent technology, deep learning has also been continuously developed in the field of engineering applications. Considering the instability and inaccuracy of past research using mathematical models for prediction, it is necessary to use computer methods to perform model learning and training on non-linear data and complete the prediction data task.
[0116] To achieve the above objectives, the present invention proposes a method based on a deep learning neural network for numerically predicting the drilling efficiency of underground cavern groups. The process of model training and prediction is as Figure 1 shown, and the calculation principle of the backpropagation neural network is as Figure 2 shown. The specific implementation method is as follows:
[0117] The first step: Obtain relevant data on the drilling efficiency of underground cavern groups and its influencing factors.
[0118] The drilling efficiency is affected by multiple index factors. Therefore, it is necessary to select the influencing factors and perform quantitative processing on them before constructing the drilling efficiency model. Geological conditions, operating factors, and mechanical characteristics are usually considered the main factors affecting the drilling efficiency. In addition, different construction environments, such as rainfall, etc., will also have an inestimable impact on the drilling progress.
[0119] According to the start and end times of each drilling operation, calculate the duration required for each drill hole to obtain the drilling efficiency (h / hole). Among all factors, the altitude refers to the altitude of the drilling working face, and its change may affect the air pressure, and may also affect the machine performance and personnel status. As the drilling depth increases, the bit pressure of the drill rig will also increase. Therefore, the drilling depth may affect the drilling efficiency. In addition, different temperatures and rock types may directly affect the operating efficiency of the drill rig, and the performance and efficiency of different types of drill rigs are also different. For outdoor operations, weather conditions (sunny, cloudy, or rainy, etc.) will also directly affect the operation process of the drill rig. The personnel configuration also has an impact on the drilling efficiency. According to the construction operation manual, each drill rig should be equipped with at least one operator for operation. In fact, increasing auxiliary personnel can significantly shorten the time for loading and unloading drill pipes, thereby improving the drilling efficiency.
[0120] After comprehensively analyzing the factors affecting the drilling efficiency and their impacts, the influencing factors covering four aspects of geological conditions, operation procedures, environmental factors, and machine characteristics were finally determined as the characteristic inputs of the drilling efficiency prediction model. To obtain the relevant data on the drilling efficiency of underground cavern groups and their influencing factors, systematic on-site investigations and historical records need to be collected. The data sources mainly include construction logs, equipment records, and geological exploration reports. Construction logs usually detail daily drilling activities, including time, depth, and equipment usage; equipment records provide data on the performance and usage of drilling rigs; geological exploration reports provide detailed information on geological conditions. By integrating these data, the drilling efficiency of underground cavern groups and their influencing factors can be systematically analyzed, thereby optimizing the drilling operation and improving the overall efficiency. First, through on-site investigations, the drilling operation process can be directly observed and relevant data recorded, including drilling efficiency, drilling depth, bit wear, drilling angle, drilling diameter, and drilling speed. In addition, historical records such as construction logs, equipment records, and geological exploration reports provide rich auxiliary data sources.
[0121] Step 2: Quantify the influencing factors of the working conditions and environment of the obtained underground cavern groups on drilling efficiency to obtain a sample data set for model learning.
[0122] The influencing factor characteristics in the data set obtained after integrating multi-source data cannot be directly used as the input of the model, and the feature form still needs to be standardized, that is, all non-numerical data needs to be quantified. The features of the present invention are divided into categorical features and numerical features. For categorical features, 0, 1 variables are used to mark different classifications of each feature. Different types of drilling rigs have different performances, and the drilling rig type is an unordered categorical variable. Before putting it into the model, it needs to be converted into a dummy variable. Therefore, (0,0)(0,1)(1,0)(1,1) dummy variables are used to represent different types of machines. For example, for the rock type variable of a certain project, 0 represents slate and 1 represents sandstone. For numerical features, the feature values are directly input into the model.
[0123] During the process of collecting numerical-related data, first record the basic information, including year, month, day, and the start and end times of each drilling process. At the same time, record the number and time of adding and unloading drill pipes for each drilling rig, the coordinates, elevation, and depth of the borehole, the type of rock, and the type of drilling rig. Secondly, the meteorological monitoring station automatically records a set of meteorological data every 10 minutes. Match the drilling data and meteorological data in terms of time, and find the meteorological data at or near the drilling time in the meteorological database and merge it into the drilling data to obtain the weather conditions corresponding to each borehole. Finally, integrate the data from different sources, structure the independent variables affecting the drilling efficiency, and finally obtain a sample data set containing target variables and feature variables that can be directly used as the input of the model.
[0124] Step 3: Establish a deep learning prediction model for drilling efficiency based on a backpropagation neural network.
[0125] The backpropagation neural network (BPNN) is a multi-layer feedforward network, usually consisting of an input layer, a hidden layer, and an output layer. Among them, each layer is composed of several nodes, and each node represents a neuron. The neurons in the hidden layer usually use the Sigmoid activation function to represent, while the neurons in the input layer or output layer usually use the linear transfer function to represent. The upper-layer nodes of the backpropagation neural network are connected to the lower-layer nodes through weights, and there is no connection between the nodes in the same layer.
[0126] The backpropagation neural network consists of forward propagation and backpropagation. In the forward propagation process, when a large number of samples are input into the neural network, the input signal propagates forward through the input layer, passes through the hidden layer, and after the function of the hidden layer, the output signal of the hidden layer is transmitted to the output layer, and finally the output result is obtained. In the whole forward propagation process, each layer of neurons only receives the input from the neurons in the previous layer, and the output of each layer of neurons only affects the output of the neurons in the next layer. If there is a large error between the final output result and the expected result, it will turn to the backpropagation process, and the error value will be along the original connection channel from the output layer through the hidden layer, and finally return to the input layer. It modifies the weights of each layer of neurons by backpropagating the error signal along the original connection path until the expected goal is reached. The specific algorithm logic steps are as follows:
[0127] (1) Set the step size ρ, and the initial value of the step size ρ should be a small number; set the weight vector W, and the initial value of the weight vector W should also be set to a small value;
[0128] (2) Select a training sample data <E s , C s >;
[0129] (3) Forward propagation stage: Starting from the input neurons, calculate the weighted linear sum of the weight vector and the input data for each neuron. Let the weighted linear sum of the input data be S, and calculate u using the activation function u =f(S u ); u u is the output of the output layer, and f() represents the forward propagation activation function.
[0130] (4) Backpropagation stage:
[0131] Starting from the output neurons; calculate the gradient for each neuron: f′(S u )=u u (1 - u u);f′() represents the backpropagation activation function.
[0132] If u u is the output unit, then: δ u =(C u -u u )f′(S u );
[0133] For other units: δ u =(∑ m:m>u ω m,u δ m )f′(S u );
[0134] (5) Update the weight vector:
[0135] (6) If the termination condition is satisfied, exit from the constructed neural network; otherwise, return to step (2) and continue execution.
[0136] In the above formulas:
[0137] ρ is the step size or learning rate, which controls the step of weight update;
[0138] u, v, m are neuron numbers;
[0139] E s is the characteristic variable of the sample data;
[0140] C s is the target variable of the sample data;
[0141] u u is the output of the u-th neuron;
[0142] u v is the output of the v-th neuron;
[0143] S u is the input of the u-th neuron;
[0144] δ u is the error term of the u-th neuron, indicating the transmission of the error through the derivative of the activation function, which affects the weight update;
[0145] δ m is the error term of the m-th neuron, indicating the error of the output of the m-th neuron, which affects the weight update;
[0146] C u is the true value of the u-th neuron;
[0147] is the optimized weight, the weight after backpropagation update, used to represent the optimal weight after training;
[0148] w u,v is the weight value before training or at the initial stage, representing the connection weight between the \(u\)-th neuron and the \(v\)-th neuron;
[0149] ω m,u is the weight from neuron \(m\) to neuron \(u\);
[0150] \(f()\) represents the forward propagation activation function;
[0151] \(f'()\) represents the backward propagation activation function.
[0152] When constructing the BPNN prediction model, the present invention calls the Multi-Layer Perceptron Regressor (MLP Regressor) model. MLPRegressor provides a highly encapsulated interface, making it more convenient to build and train a multi-layer perceptron model without delving into the implementation details of the underlying backpropagation algorithm. As a powerful regression model in the scikit-learn library, MLP Regressor encapsulates the BPNN algorithm and provides a simple and intuitive interface for model construction, facilitating the construction and training of the model. The prediction model only needs to provide data and parameter settings, and the MLP Regressor model can automatically handle the backpropagation algorithm during the training process, thereby learning a model suitable for the data. This way of using a highly encapsulated model enables the prediction program to focus more on higher-level tasks such as hyperparameter tuning and cross-validation of the model, without paying too much attention to the implementation details of the underlying backpropagation algorithm. Therefore, the present invention uses the MLP Regressor model to quickly build and train the BPNN model.
[0153] Based on calling the MLP Regressor model that can implement the backpropagation neural network algorithm, the specific training steps of the BPNN prediction model of the present invention are as follows:
[0154] (1) All data is preprocessed by standardization, that is, each value in this column is subtracted by the mean of this column, and then divided by the standard deviation of this column. This can make each numerical column of the dataset have the same scale.
[0155] (2) The preprocessed training set including drilling efficiency and related influencing factors is evenly divided into 5 parts for five-fold cross-validation. Four of them are selected as the training set, and one is used as the validation set.
[0156] (3) Initialize the parameters of the training model. The hyperparameters of the model (such as the size of the hidden layer, learning rate, etc.) are set to fixed values during initialization. Use the divided training set for fitting, that is, adjust the parameters of the model to minimize the loss function.
[0157] (4) After the model training is completed, use the validation set for prediction. For each sample, the model generates a predicted value, representing its estimate of the target variable.
[0158] (5) Reverse normalize the predicted value and the true value (the inverse operation of normalization), convert them back to the range of the original data, and output the model training and validation results.
[0159] (6) Output the model evaluation metrics.
[0160] Step 4: Use the bacterial foraging optimization algorithm to optimize the hyperparameter of the hidden layer structure in the deep learning prediction model.
[0161] The process of the BFO algorithm simulating a bacterial population includes three steps: chemotaxis operation, reproduction operation, and elimination-dispersal operation. The chemotaxis operation refers to the movement of bacteria in the direction that is beneficial to their survival, including tumbling and swimming. When a bacterium moves one unit step in any direction, it is called a tumbling movement; when a bacterium moves one unit step along the previous movement direction, it is called a swimming movement. Usually, in a poor environment (such as a toxic area), bacteria tend to perform tumbling movements; in a good environment (such as an area rich in food), bacteria tend to perform swimming movements. In the life cycle of Escherichia coli, these two movements alternate, mainly aiming to find food and avoid toxic substances.
[0162] After a period of food search, some bacteria with weaker foraging abilities will be eliminated. To maintain the population size, the remaining bacteria will reproduce. This process is called the reproduction operation in the BFO algorithm. In the reproduction operation, first, the bacteria are sorted according to the quality of their positions, and the bacteria ranked behind are eliminated. The remaining bacteria then self-replicate to generate a new individual with the same position as the original bacterium, that is, a new bacterium with the same foraging ability.
[0163] The local area where bacteria live may suddenly or gradually change (such as a temperature increase or food consumption), resulting in the possible collective death or migration of the bacterial population in that area. This process is called the elimination-dispersal behavior in the BFO algorithm. The elimination-dispersal operation occurs with a certain probability. If a bacterial individual meets the elimination-dispersal probability, that bacterial individual will perish, and a new individual will be randomly generated at any position in the solution space. This new individual may have a different position from the perished individual, that is, a different foraging ability. The randomly generated new individual may be closer to the global optimal solution, which helps the chemotaxis operation to jump out of the local optimal solution and find the global optimal solution.
[0164] The BFO algorithm simulates the bacterial foraging process based on the bacterial foraging optimization algorithm, and finds the optimal hyperparameter combination through a method that combines global search and local search; during the optimization process, through multiple iterations and evaluations, the hyperparameters are continuously adjusted to make the performance of the model on the validation set reach the best.
[0165] The present invention optimizes the number of hidden layers of the model and the number of neurons in each layer to improve the prediction performance of the model on a given data set. Before applying the algorithm, the maximum number of hidden layers and the number of neurons in each layer are specified first so that the training of the model can find the optimal parameters within a certain time. In the prediction code, a function "bfoa_optimization" is created before training the model to implement the main functions of the BFO algorithm. This function is responsible for running the entire optimization process of the BFO algorithm to ensure that the optimization algorithm can be stably embedded in the model training process to achieve the search and selection of the best model.
[0166] In the initialization stage of the algorithm, the BFO algorithm generates a group of initial bacterial populations, and each bacterium represents a possible hidden layer structure of the neural network model. Each bacterium has a position, that is, its corresponding hidden layer structure, and these positions are randomly initialized. In addition, other optimization parameters such as the number of iterations, the number of bacterial populations, and the fine-tuning step size are artificially initialized. The purpose of artificially specifying these parameters is to enable the optimization to be carried out within a specified range, avoid falling into local optimal solutions, control the stability of the algorithm, and accelerate the optimization process.
[0167] In the optimization process of the BFO algorithm, each bacterium undergoes a series of fine-tuning and selection-elimination operations. The fine-tuning operation refers to the bacterium increasing or decreasing the number of neurons in its hidden layer structure in the hope of finding a better structure. This process simulates the swimming process of bacteria in the environment to find the most suitable position for survival. After fine-tuning, the algorithm calculates the fitness of each bacterium on the training set, that is, the performance index of the model. The present invention uses the mean squared error (MSE) as the fitness function. If the fitness of the bacterium after fine-tuning is better, its position is updated to the new fine-tuned structure.
[0168] At the same time, the optimization algorithm also implements selection and elimination operations to maintain the diversity of the bacterial population. After each iteration, the algorithm sorts the bacteria according to their fitness and selects the better half of the bacteria with higher fitness for replication to maintain the population size. The other half of the bacteria are re-initialized to new positions according to a certain elimination probability to increase the diversity of the population.
[0169] After all iterations are completed, the bacterium with the smallest mean squared error (MSE) among all bacteria is selected as the best model, and the hidden layer structure is determined. Then, this best model will be used for further model evaluation and prediction tasks. In this way, the BFO algorithm is successfully embedded in the neural network model optimization process, and the corresponding code is implemented to ensure the close connection between the optimization algorithm and model training, enabling the optimization algorithm to automatically optimize the neural network structure and parameters and improve the prediction performance of the model on real data.
[0170] Step 5: Call the optimized deep learning prediction model to achieve prediction by analyzing the sample data set related to drilling efficiency.
[0171] The model obtains a trained model by learning the sample data set, obtains the non-linear relationship characteristics between the target variable and the feature variable, and after saving the model and calling the model, the prediction result of the drilling efficiency is obtained according to the characteristics related to the underground cavern group to be predicted.
[0172] The above BFO algorithm, backpropagation neural network, multi-layer perceptron regression model, etc. can adopt applicable algorithms, functional modules, models and software in the prior art, or can also adopt algorithms, functional modules, models and software in the prior art and be constructed by conventional technical means.
[0173] The embodiments described above are only used to illustrate the technical ideas and features of the present invention, and the purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The patent scope of the present invention cannot be limited only by this embodiment, that is, any equivalent changes or modifications made to the spirit disclosed by the present invention still fall within the patent scope of the present invention.
Claims
1. A method for predicting the efficiency of underground cavern drilling based on deep learning, characterized in that: The steps include: Step 1, collecting relevant data on the drilling efficiency of underground cavern groups and their influencing factors, and using the collected data as original data; Step 2, preprocessing the collected raw data and compiling a sample data set; Step 3, establishing a drilling efficiency prediction model based on a back propagation neural network; using a sample data set to train the drilling efficiency prediction model; Step 4, using the BFO algorithm to optimize the hyperparameters of the hidden layer structure in the drilling efficiency prediction model; Step 5, monitoring the drilling efficiency of the underground cavern group and its influencing factors during the construction process; preprocessing the monitored data, and then predicting the drilling efficiency during the construction process through the optimized drilling efficiency prediction model.
2. The method for predicting underground cavern group rock drilling efficiency based on deep learning according to claim 1, characterized in that: In step 1, the method for collecting data related to the drilling efficiency of underground cavern groups and its influencing factors includes the following method steps: obtaining the following data related to the underground cavern group drilling operation through field investigation and historical records: drilling time, drilling depth, drill bit wear, drilling angle, drilling diameter, drilling speed, geological conditions, equipment performance and operator configuration; data sources include construction logs, equipment records, and geological survey reports; During the data collection process, basic information is first recorded, including the year, month, day, and start and end time of each drilling process. At the same time, the number and time of adding and unloading drill rods for each drilling rig, the coordinates, elevation and depth of the borehole, the type of rock, and the type of drilling rig are recorded; secondly, the meteorological monitoring station automatically records a set of meteorological data every 10 minutes, matches the drilling data and meteorological data by time, finds the meteorological data at or near the drilling time in the meteorological database and merges it into the drilling data to obtain the weather conditions corresponding to each borehole; finally, the data from different sources are integrated, and the independent variables affecting the drilling efficiency are structured, and finally a sample data set containing target variables and characteristic variables is obtained for direct input into the drilling efficiency prediction model.
3. The method for predicting underground cavern group rock drilling efficiency based on deep learning according to claim 1, characterized in that: Step 2 includes the following method steps: Step 2-1, clean the original data, delete duplicate data and process outliers and missing values; Step 2-2, quantify the non-numerical data; Step 2-3, standardize all the data obtained after processing in step 2-1 and step 2-2 so that each numerical column of the data set has the same scale; Step 2-4, extract and select features from the data processed in step 2-3; extract key features that have a greater impact on drilling efficiency by analyzing the relationship between each feature and drilling efficiency; organize these features and the corresponding drilling efficiency into a structured sample data set as data input for drilling efficiency prediction model training.
4. The method for predicting the efficiency of underground cavern drilling based on deep learning according to claim 3 is characterized in that: In step 2-2, the method for quantizing non-numeric data includes the following steps: Non-numeric data include categorical feature data, which use artificially defined dummy variables to mark the different categories of each feature.
5. The method for predicting underground cavern group rock drilling efficiency based on deep learning according to claim 1, characterized in that: In step 3, a drilling efficiency prediction model is constructed based on a multi-layer perceptron regression model using a back propagation algorithm.
6. The method for predicting underground cavern group rock drilling efficiency based on deep learning according to claim 1, characterized in that: In step 3, the method of using the sample data set to train the drilling efficiency prediction model includes the following method steps: Step 3-1, divide the sample data set into five parts for five-fold cross-validation, select four of them as training sets and one as validation set; Step 3-2, initializing the parameters of the drilling efficiency prediction model for training, wherein the hyperparameters of the drilling efficiency prediction model are set to fixed values during initialization; fitting is performed using the training set, and the parameters of the drilling efficiency prediction model are adjusted to minimize the loss function; Step 3-3, after the drilling efficiency prediction model training is completed, the validation set is used and the cross-validation method is used to evaluate the model performance.
7. The method for predicting underground cavern group rock drilling efficiency based on deep learning according to claim 6, characterized in that: Step 3-2 includes the following method steps: By calculating the error between the predicted value and the true value, the weight and bias of the model are adjusted; the gradient descent method is used to minimize the prediction error and gradually improve the prediction accuracy of the model; during the training process, the cross-validation method is used to evaluate the model performance to prevent overfitting.
8. The method for predicting underground cavern group rock drilling efficiency based on deep learning according to claim 6, characterized in that: Step 3-3 includes the following method steps: reverse standardizing the predicted values and true values, converting them back to the range of the original data, and outputting the model training and verification results.
9. The method for predicting underground cavern group rock drilling efficiency based on deep learning according to claim 1, characterized in that: In step 4, the number of hidden layers and the number of neurons in each layer of the model are optimized by the BFO algorithm to improve the prediction performance of the model on a given data set. Before applying the algorithm, the maximum number of hidden layers and the number of neurons in each layer are first specified so that the model training can find the optimal parameters within a certain time. In the initialization phase of the BFO algorithm, a group of initial bacterial populations is randomly generated, each bacterium represents a neuron in the hidden layer of the neural network model; the position of each bacterium is the position of its corresponding neuron in the hidden layer structure, and the bacterial position is randomly initialized; other optimization parameters are initialized artificially; During the optimization process of the BFO algorithm, each bacterium undergoes multiple fine-tuning and selection and elimination operations. Fine-tuning operations refer to the increase or decrease of the number of neurons in the hidden layer structure corresponding to the bacterium in the hope of finding a better structure. This process simulates the movement of bacteria in the environment to find the most suitable position for survival. After fine-tuning, the fitness of each bacterium on the training set is calculated, which is the performance index of the model. The mean square error is used as the fitness function. If the fitness of the fine-tuned bacteria is better, its position is updated to the new fine-tuned structure. After each iteration, the BFO algorithm sorts the bacteria according to their fitness and selects half of the bacteria with better fitness for replication to maintain the population size; the other half of the bacteria are reinitialized to new positions according to a certain elimination probability to increase the diversity of the population; after all iterations are completed, the bacteria with the smallest mean square error are selected among all bacteria as the best model, and the hidden hierarchy is determined.
10. A device for predicting the efficiency of underground cavern drilling based on deep learning, comprising a memory and a processor, characterized in that: The memory is used to store computer programs; the processor is used to execute the computer program and implement the steps of the method for predicting underground cavern group rock drilling efficiency based on deep learning as described in any one of claims 1 to 9 when executing the computer program.