A deep learning-based rock permeability prediction method

Abstract

This invention belongs to the field of rock permeability prediction technology and discloses a deep learning-based rock permeability prediction method. This method first conducts seepage tests on samples under different roughness and confining pressure conditions to obtain experimental data for constructing training and testing datasets. Then, a deep learning prediction model based on a fully connected neural network is built, and particle swarm optimization is used to optimize the model's parameters to avoid local optima problems that may occur with traditional gradient descent methods. This allows for better handling of complex rock properties, seepage flow processes, and experimental data, improving the efficiency and accuracy of rock permeability prediction. Furthermore, to ensure the high accuracy and practicality of the deep learning prediction model, this invention also verifies the accuracy of the rock permeability prediction results output by the deep learning prediction model using a test dataset, ensuring that the error of the prediction results is controlled within the engineering allowable range.

Description

Technical Field

[0001] This invention belongs to the field of rock permeability prediction technology, specifically relating to a rock permeability prediction method based on deep learning. Background Technology

[0002] Rock permeability is an important indicator for evaluating the ability of rocks to conduct fluids during flow, and it is widely used in geological exploration, oil extraction, hydrogeological engineering, and other fields. The measurement and prediction of permeability is a key issue in the study of rock physical properties, especially in oil and gas exploration, groundwater resource management, and underground reservoir development. Accurate assessment of rock permeability directly affects engineering design and resource development.

[0003] Traditional methods primarily predict rock permeability through laboratory permeability tests, downhole measurements, and empirical formulas based on rock type and porosity. However, these methods have limitations: firstly, laboratory permeability tests and downhole measurements typically require significant time and are costly; secondly, permeability is influenced by multiple factors, including rock porosity, fracture structure, rock type, confining pressure, and fluid properties, and the interactions between these factors are extremely complex, making accurate prediction using simple empirical formulas difficult. Therefore, designing a more efficient and accurate rock permeability prediction method is crucial.

[0004] In recent years, deep learning, as a powerful machine learning technique, has been widely applied in various fields such as natural language processing, computer vision, and speech recognition, achieving remarkable results. Through multi-layered neural network models, deep learning can automatically extract useful features from large amounts of data and make efficient predictions. Using deep learning methods to predict rock permeability can overcome many drawbacks of traditional methods, especially when dealing with large-scale, complex data, where deep learning can more accurately model and predict rock permeability.

[0005] Currently, although some research applications have been made using machine learning for rock permeability prediction, these methods still have much room for improvement in terms of large-scale data processing, feature selection, and accuracy enhancement. Summary of the Invention

[0006] The purpose of this invention is to propose a rock permeability prediction method based on deep learning. This method uses a deep learning prediction model to predict rock permeability and combines it with a particle swarm optimization algorithm to accelerate the model training process. This method can better handle complex rock properties, seepage flow processes and experimental data, thereby improving the efficiency and accuracy of rock permeability prediction.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based method for predicting rock permeability includes the following steps: Step 1. Conduct seepage tests under different roughness conditions and confining pressure conditions, and obtain real-time monitoring test data; Step 2. Perform data preprocessing on the experimental data obtained in Step 1 to generate training and test datasets; Step 3. Build a deep learning prediction model based on a fully connected neural network, which includes the following structure: The input layer is used to receive experimental data; Hidden layers are used to extract features from experimental data and perform nonlinear transformations using the ReLU activation function to obtain the input features; The output layer is used to predict rock permeability based on input features; Step 4. Based on the training dataset, train the deep learning prediction model using the particle swarm optimization algorithm, and use the deep learning prediction model obtained after this training as the current deep learning prediction model. Step 5. Input the test dataset into the current deep learning prediction model to predict rock permeability; If the prediction error of the current deep learning prediction model is lower than the preset accuracy requirement, proceed to step 4 for retraining. If the prediction error of the current deep learning prediction model is not lower than the preset accuracy requirement, then the current deep learning prediction model is used as the trained deep learning prediction model, and the process proceeds to step 6. Step 6. Use the trained deep learning prediction model to predict rock permeability.

[0008] Furthermore, based on the aforementioned deep learning-based rock permeability prediction method, this invention also proposes a computer device, which includes a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the steps of the deep learning-based rock permeability prediction method described above.

[0009] Furthermore, based on the aforementioned deep learning-based rock permeability prediction method, this invention also proposes a computer-readable storage medium storing a program thereon; when executed by a processor, this program is used to implement the steps of the aforementioned deep learning-based rock permeability prediction method.

[0010] The present invention has the following advantages: As described above, this invention relates to a rock permeability prediction method based on deep learning. This method first conducts seepage tests on samples under different roughness conditions and different confining pressure conditions, thereby more realistically simulating the permeability characteristics of rock fissures under complex geological conditions. Compared with traditional research methods that only target a single test condition, the method of this invention is more comprehensive in experimental design and has wider applicability.

[0011] Furthermore, traditional methods for predicting rock permeability mostly rely on physical experiments and empirical formulas. While these methods can provide a certain level of accuracy, their accuracy and applicability are greatly limited by the diversification of rock types and experimental conditions. The method of this invention differs from traditional theoretical formulas and empirical models. It does not rely on complex physical assumptions and prior knowledge, and can complete predictions using experimental data, reducing the difficulty of model construction. Moreover, the deep learning prediction model constructed in this invention can efficiently capture the nonlinear relationship between rock permeability and fracture roughness, confining pressure conditions, and other experimental parameters, thereby significantly improving the prediction accuracy of rock permeability.

[0012] Furthermore, the method of this invention acquires experimental data based on indoor experiments and builds a deep learning prediction model based on a fully connected neural network. Deep learning technology is used to improve prediction accuracy, and particle swarm optimization is employed to optimize the parameters of the deep learning prediction model, avoiding the local optima problem that may occur with traditional gradient descent methods. This significantly improves the model's training efficiency and predictive ability. The invention also verifies the accuracy of the rock permeability prediction results output by the deep learning prediction model using a test dataset, ensuring that the prediction error is controlled within the acceptable engineering range, thereby further guaranteeing the model's high accuracy and practicality. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the steps of a deep learning-based rock permeability prediction method in an embodiment of the present invention.

[0014] Figure 2 This is a flowchart of a deep learning-based rock permeability prediction method in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 This embodiment proposes a deep learning-based method for predicting rock permeability. The method first conducts seepage tests under different roughness and confining pressure conditions to acquire experimental data, including real-time monitoring data and permeability parameters. Next, the real-time monitoring data and permeability parameters are preprocessed to generate training and testing datasets. Then, a deep learning prediction model is trained using the training dataset, and the network parameters of the model are optimized using a particle swarm optimization algorithm. The model's accuracy is tested using the testing dataset. Finally, the deep learning prediction model that meets the accuracy requirements is used to predict rock permeability.

[0016] The method of this invention is based on indoor test data, establishes a deep learning prediction model, and uses deep learning technology to improve prediction accuracy, thus solving the problems of low accuracy, high test cost, and large difference between prediction results and actual situation in traditional rock permeability prediction methods.

[0017] like Figure 1 As shown, a deep learning-based method for predicting rock permeability includes the following steps: Step 1. Conduct seepage tests under different roughness conditions and confining pressure conditions, and obtain real-time monitoring test data.

[0018] In this embodiment, the sample opening is uniformly set to 5 mm in the rock seepage test to ensure the accuracy of the test results. The seepage test in step 1 includes two test schemes: one is to conduct seepage tests under different roughness conditions, and the other is to conduct seepage tests under different confining pressure conditions.

[0019] The specific process of conducting seepage tests under different roughness conditions is as follows: Rock blocks of the same lithology were sampled on-site according to the standard test method for engineering rock mass, and fracture specimens were prepared using the Brazilian splitting method to form specimens with different joint surface roughness coefficients. In this embodiment, the specimens are preferably cuboid specimens with a size of 200 mm × 100 mm × 100 mm, and each group of rock fractures generated by the Brazilian splitting method includes upper and lower parts.

[0020] The joint surface roughness coefficient is determined by scanning the morphology of the joint surface of the fractured sample.

[0021] Seepage tests were conducted on samples with different joint surface roughness coefficients under constant confining pressure and constant head difference. The monitored and recorded test data included: The length and width of the joint surfaces of the specimen are used to calculate the cross-sectional area of ​​the seepage.

[0022] The viscosity of the test fluid.

[0023] The fluid pressure at the inlet and outlet of the sample is used to calculate the pressure difference for seepage.

[0024] The inlet and outlet flow rates of the sample are measured. The inlet flow rate is measured by a constant speed and pressure pump, and in this embodiment, the inlet flow rate measured by the constant speed and pressure pump will be recorded in the computer. The outlet flow rate is collected by a flow meter and an electronic balance, where the flow meter is for large flow rates and the electronic balance is for small flow rates.

[0025] In this embodiment, the flow rate at the free outflow end, i.e., the flow rate q in the cracks of the sample, can also be collected based on the experiment.

[0026] Specifically, in this embodiment, the sample length along the seepage path, i.e., the length of the joint surface of the sample, is precisely recorded as 200 mm, and the joint surface width perpendicular to the flow direction is 100 mm, and the seepage cross-sectional area is calculated accordingly. The physical properties of the fluid are controlled by a high-precision isothermal system to ensure the stability and accurate acquisition of the dynamic viscosity, and a precision pressure sensor is used to directly collect the fluid pressure at the sample inlet. Fluid pressure at the outlet And calculate the pressure difference between the two in real time. .

[0027] Based on the law of cubes, the initial permeability of the sample is calculated using the recorded experimental data.

[0028] The specific process of conducting seepage tests under different confining pressures is as follows: Samples with the same size, shape, and roughness were cast using gypsum, water, and retarder in the same proportions. In this embodiment, rectangular samples with the same lithology and roughness are preferably used to apply confining pressure for seepage tests.

[0029] Under the premise of ensuring that the sample is not damaged, different confining pressures are applied to the sample according to the confining pressure conditions in the real environment to carry out seepage tests. In this embodiment, the confining pressures are set to 3 MPa, 5 MPa and 7 MPa.

[0030] The monitored and recorded test data include: The length and width of the joint surfaces of the specimen are used to calculate the cross-sectional area of ​​the seepage.

[0031] The viscosity of the test fluid.

[0032] The fluid pressure at the inlet and outlet of the sample is used to calculate the pressure difference for seepage.

[0033] The flow rate through the cracks in the sample, the inlet flow rate, and the outlet flow rate of the sample.

[0034] The inlet flow rate is measured by a constant speed and constant pressure pump, while the outlet flow rate is collected by a flow meter and an electronic balance.

[0035] Based on the law of cubes, the initial permeability of the sample is calculated using the recorded experimental data.

[0036] Step 2. Perform data preprocessing on the experimental data obtained in Step 1 to generate training and test datasets.

[0037] Step 2 will be described in detail below. Step 2 is as follows: The experimental data obtained in step 1 includes real-time monitoring data and permeability parameters. The experimental data obtained in step 1 is cleaned to obtain the selected dataset.

[0038] The data in the selected dataset are normalized to generate sample data.

[0039] The generated sample data is divided according to a preset ratio, and then used to construct training and testing datasets respectively. In this embodiment, the sample data is used to generate training and testing datasets in a 7:3 ratio.

[0040] Step 3. Build a deep learning prediction model based on the fully connected neural network FCNN structure. The deep learning prediction model includes an input layer, a hidden layer, and an output layer.

[0041] The input layer receives experimental data. The hidden layer extracts features from the experimental data and performs a nonlinear transformation using the ReLU activation function to obtain the input features. The output layer predicts rock permeability based on the input features.

[0042] The specific processing flow of signals in a deep learning prediction model based on a fully connected neural network is as follows: The input layer, acting as the receiver of experimental data, is used to receive the preprocessed experimental parameter vector X. ,in to These represent the 1st to nth preprocessed normalized eigenvalues, where n is the number of normalized eigenvalues ​​in the input parameter, i.e., the experimental parameter vector X.

[0043] Each neuron in the input layer is used to receive a normalized feature value, which includes the joint surface roughness coefficient, confining pressure, viscosity of the test fluid, pressure difference of seepage, etc., and the received normalized feature value is directly and unchanged and passed to the neuron of the first hidden layer.

[0044] Specifically, in this embodiment, the input layer consists of 5 neurons, and the number of neurons in the input layer strictly corresponds to the number of experimental parameters in the experimental data selected in this embodiment. The experimental parameters of the experimental data include the joint surface roughness coefficient JRC, confining pressure, hydrodynamic viscosity (i.e., the viscosity μ of the test fluid), and the pressure difference between the fluid at the inlet and outlet of the sample (i.e., the seepage pressure difference). And the seepage flow rate, i.e., the flow rate q in the cracks of the sample.

[0045] In this embodiment, the preprocessed experimental parameter vector X is defined as... .

[0046] There are L hidden layers, and the number of neurons in the 1st to Lth hidden layers are m1 to m2, respectively. L Number of neurons Together with the dimension of the weight matrix, it constitutes the main part of the parameters to be optimized in the particle swarm optimization algorithm.

[0047] Specifically, in this embodiment, the deep learning prediction model uses a deep structure with two hidden layers, i.e., L=2. The first hidden layer contains 64 neurons, i.e. =64. The second hidden layer contains 32 neurons, that is... =32.

[0048] make The input to the j-th hidden layer is When j=1, the input to the first hidden layer is the output of the input layer. .when At that time, the input of the j-th hidden layer is the output of the (j-1)-th hidden layer. .

[0049] Each hidden layer performs both linear and nonlinear transformations. The specific processing flow of the signal in the j-th hidden layer is as follows: The process of linear transformation is as follows: The net input is calculated by linearly combining the input of the j-th hidden layer with the weight matrix and the bias vector: .

[0050] in, This represents the net input of the j-th hidden layer. This represents the weight matrix of the j-th hidden layer. This represents the bias vector of the j-th hidden layer.

[0051] Specifically, in this embodiment, for the first hidden layer, That is, the preprocessed experimental parameter vector X input to the deep learning prediction model. It is a 64×5 matrix. It is a 64-dimensional vector. It is a 32×64 matrix. It is a 32-dimensional vector.

[0052] The process of nonlinear transformation is as follows: The ReLU activation function is used for nonlinear mapping to obtain the output of the j-th hidden layer. for: .

[0053] The process of performing linear and nonlinear transformations enables deep learning prediction models based on fully connected neural networks to learn and represent the complex nonlinear relationship between input features and the output target, i.e., rock permeability.

[0054] The input to the output layer is the output of the Lth hidden layer. In this embodiment, the output layer receives the output of the second hidden layer. The output layer performs a linear weighted calculation to obtain the predicted value of rock permeability. for: .

[0055] in, For 1×m L The weight vector, This is a scalar bias. In this embodiment, It is a 1×32 weight vector. Since rock permeability prediction is a regression problem, the output layer does not use any activation function to preserve the output value as the predicted rock permeability. The continuous range.

[0056] Step 4. Based on the training dataset, train the deep learning prediction model using the particle swarm optimization algorithm, and use the deep learning prediction model obtained after this training as the current deep learning prediction model.

[0057] This invention inputs the training dataset into a fully connected neural network of a deep learning prediction model for forward propagation. Particle Swarm Optimization (PSO) is used to optimize the parameters of the deep learning prediction model, and mean squared error (MSE) is used to measure the difference between the predicted and actual values.

[0058] Particle swarm optimization (PSO), as an intelligent optimization algorithm, can find the global optimum in high-dimensional space, avoiding the problem of traditional gradient descent methods getting trapped in local optima. Combining deep learning and PSO can not only improve prediction accuracy but also accelerate the model training process and reduce manual intervention and computational costs.

[0059] The goal of Particle Swarm Optimization (PSO) is to optimize the weights and biases of a neural network in a deep learning prediction model to minimize prediction error. PSO works collaboratively through multiple particles, each representing a solution, and the fitness of a particle is evaluated using an objective function such as mean squared error.

[0060] Step 4 will be described in detail below. Step 4 is as follows: First, the particle swarm is initialized. The parameters of the fully connected neural network in the deep learning prediction model include network weights and biases. Each solution of the parameters of the fully connected neural network is represented by the position of a particle and the particle swarm is formed.

[0061] Initialize the positions of the particles and set the velocity of each particle.

[0062] Next, velocity and position updates are performed iteratively. The particle's velocity and position are updated according to the following formula in each iteration: The speed update formula is: .

[0063] The position update formula is: .

[0064] in, , This indicates the number of samples in the training dataset. This represents the velocity of the particle in the i-th sample after the t-th iteration. Let represent the velocity of the particle in the i-th sample after the (t+1)-th iteration. This represents the position of the particle in the i-th sample after the t-th iteration. This represents the position of the particle in the i-th sample after the (t+1)-th iteration. Indicates inertia weight, Inertia is used to control the movement of particles. and Represents the learning factor. and The attraction used to control particles toward individual and global optimum. and Represents a random number between [0, 1]. This indicates the optimal position of an individual particle. This represents the optimal position of the entire swarm, i.e., the global optimal position of the particle swarm.

[0065] The fitness of the particles is then evaluated. After each iteration of updating the particle's velocity and position, the particle's fitness value is calculated. The fitness of the particles is evaluated based on the loss function value; the smaller the loss function value, the higher the particle's fitness. This embodiment performs particle swarm optimization by minimizing the mean squared error.

[0066] Specifically, the particle's fitness value is obtained by using the current particle's position as a parameter of the fully connected neural network and calculating the loss of the deep learning prediction model using the training dataset. Mean squared error is used to measure the difference between the predicted and true values. The loss function is as follows: .

[0067] in, This represents the actual permeability of the i-th sample, i.e., the true value of the rock permeability. This represents the predicted permeability of the i-th sample, i.e., the rock permeability, as predicted by the deep learning prediction model. MSE represents the mean square error.

[0068] The particle swarm continuously adjusts its position and velocity through multiple iterations, searching for the global optimum. In each iteration, the individual particle's position and the global optimum are continuously updated. When the particle's fitness value (MSE) reaches a preset threshold or the number of iterations exceeds a preset maximum number of iterations, the parameter optimization of the fully connected neural network stops, completing the training of the deep learning prediction model for this iteration. The deep learning prediction model obtained after this training is then used as the current deep learning prediction model.

[0069] In this embodiment, the particle velocity update parameter, i.e., the learning factor, that controls the learning rate in PSO is: and They are initialized to 2.0, and the learning factor is dynamically adjusted according to the error curve during training. The training epochs, i.e., the preset maximum number of iterations, are set to 100, and an early stopping strategy is used to avoid overfitting. The particle swarm is updated in each iteration until the fitness value of the particles meets the accuracy requirements or the number of iterations exceeds the preset maximum number of iterations.

[0070] This invention provides a solution for efficient and accurate prediction of rock permeability by combining a deep learning-based rock permeability prediction method with a particle swarm optimization algorithm for model training. This method can better handle complex rock properties, permeation flow processes, and experimental data.

[0071] Step 5. Input the test dataset into the current deep learning prediction model to predict rock permeability.

[0072] If the prediction error of the current deep learning prediction model is lower than the preset accuracy requirement, proceed to step 4 for retraining.

[0073] If the prediction error of the current deep learning prediction model is not lower than the preset accuracy requirement, then the current deep learning prediction model is used as the trained deep learning prediction model, and the process proceeds to step 6.

[0074] Step 5 will be described in detail below. Step 5 is as follows: The test dataset is fed into the current deep learning prediction model to predict rock permeability.

[0075] The prediction error of the current deep learning prediction model is calculated based on the true value of rock permeability in the test dataset and the predicted value of rock permeability output by the current deep learning prediction model.

[0076] Determine whether the prediction error of the current deep learning prediction model is lower than the preset accuracy requirement.

[0077] In this embodiment, the preset accuracy requirement refers to the MSE between the predicted value and the true value of the test dataset being less than a preset error value, which is 0.0001, to ensure that the deep learning prediction model has high prediction accuracy. The predicted value of the test dataset is the predicted rock permeability output by the current deep learning prediction model, and the true value of the test dataset is the actual rock permeability value in the test dataset.

[0078] If the prediction error of the current deep learning prediction model is lower than the preset accuracy requirement, proceed to step 4 to retrain the current deep learning prediction model until the accuracy of the tested model meets the requirements.

[0079] If the prediction error of the current deep learning prediction model is not lower than the preset accuracy requirement, then the current deep learning prediction model is used as the trained deep learning prediction model, and the process proceeds to step 6.

[0080] The accuracy of the rock permeability prediction results output by the deep learning prediction model is verified by testing the test dataset. This enables the model constructed by the deep learning-based rock permeability prediction method of this invention to better meet the needs of practical applications, thereby providing strong technical support for geotechnical engineering, oil exploration and other fields.

[0081] Step 6. Use the trained deep learning prediction model to predict rock permeability.

[0082] The trained deep learning prediction model is the one that meets the accuracy requirements. Step 6 uses it to predict rock permeability.

[0083] Step 6 will be described in detail below. Step 6 is as follows: Input new test data, such as new seepage pressure difference, flow rate of the fracture, viscosity of the test fluid, etc., and preprocess the test data.

[0084] The preprocessed experimental data is used as input data for the trained deep learning prediction model.

[0085] The input layer receives input data, which is then fed into the hidden layer for feature extraction. The ReLU activation function is used for nonlinear transformation, and the data is then fed into the output layer to predict rock permeability.

[0086] The rock permeability obtained from the output layer is used as the prediction result output by the trained deep learning prediction model.

[0087] Example 2 This embodiment 2 describes a computer device that includes a memory and one or more processors.

[0088] The memory stores executable code, which, when executed by the processor, implements the steps of the deep learning-based rock permeability prediction method in Embodiment 1 above.

[0089] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0090] Example 3 This embodiment 3 describes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of a deep learning-based rock permeability prediction method.

[0091] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0092] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A deep learning-based rock permeability prediction method, characterized in that, Includes the following steps: Step 1. Conduct seepage tests under different roughness conditions and confining pressure conditions, and obtain real-time monitoring test data; In step 1, the monitored and recorded test data include the length and width of the joint surface of the specimen, the viscosity of the test fluid, the fluid pressure at the inlet and outlet of the specimen, and the flow rate through the cracks in the specimen. Step 2. Perform data preprocessing on the experimental data obtained in Step 1 to generate training and test datasets; Step 3. Build a deep learning prediction model based on a fully connected neural network, which includes the following structure: The input layer is used to receive experimental data; Hidden layers are used to extract features from experimental data and perform nonlinear transformations using the ReLU activation function to obtain the input features; The output layer is used to predict rock permeability based on input features; Step 4. Based on the training dataset, train the deep learning prediction model using the particle swarm optimization algorithm, and use the deep learning prediction model obtained after this training as the current deep learning prediction model. In step 4, the parameters of the fully connected neural network in the deep learning prediction model include network weights and biases. Each solution of the parameters of the fully connected neural network is represented by the position of a particle and forms a particle swarm. Step 5. Input the test dataset into the current deep learning prediction model to predict rock permeability; If the prediction error of the current deep learning prediction model is lower than the preset accuracy requirement, proceed to step 4 for retraining. If the prediction error of the current deep learning prediction model is not lower than the preset accuracy requirement, then the current deep learning prediction model is used as the trained deep learning prediction model, and the process proceeds to step 6. Step 6. Use the trained deep learning prediction model to predict rock permeability.

2. The rock permeability prediction method based on deep learning according to claim 1, characterized in that, In step 1, the process of conducting seepage tests under different roughness conditions is specifically as follows: According to the standard test method for engineering rock mass, rock blocks of the same lithology were sampled on site, and fracture samples were prepared by Brazilian splitting method to form samples with different joint surface roughness coefficients. The joint surface roughness coefficient is determined by scanning the morphology of the joint surface of the fractured sample. Seepage tests were conducted on samples with different joint surface roughness coefficients under constant confining pressure and constant head difference. The monitored and recorded test data included: The length and width of the joint surfaces of the specimen are used to calculate the cross-sectional area of ​​the seepage flow; The viscosity of the test fluid; The fluid pressure at the inlet and outlet of the sample is used to calculate the pressure difference for seepage. The flow rate through the cracks in the sample, the inlet flow rate and the outlet flow rate of the sample; The inlet flow rate is measured by a constant speed and constant pressure pump, and the outlet flow rate is collected by a flow meter and an electronic balance. Based on the law of cubes, the initial permeability of the sample is calculated using the recorded experimental data.

3. The rock permeability prediction method based on deep learning according to claim 1, characterized in that, In step 1, the process of conducting seepage tests under different confining pressures is specifically as follows: Samples with the same size, shape and roughness were cast using the same ratio of gypsum, water and retarder. Under the premise of ensuring that the specimen is not damaged, seepage tests are conducted by applying different confining pressures to the specimen, and the test data monitored and recorded include: The length and width of the joint surfaces of the specimen are used to calculate the cross-sectional area of ​​the seepage flow; The viscosity of the test fluid; The fluid pressure at the inlet and outlet of the sample is used to calculate the pressure difference for seepage. The flow rate through the cracks in the sample, the inlet flow rate and the outlet flow rate of the sample; The inlet flow rate is measured by a constant speed and constant pressure pump, and the outlet flow rate is collected by a flow meter and an electronic balance. Based on the law of cubes, the initial permeability of the sample is calculated using the recorded experimental data.

4. The rock permeability prediction method based on deep learning according to claim 1, characterized in that, Step 2 specifically involves: The experimental data obtained in step 1 is cleaned to obtain the selected dataset; The data in the selected dataset is normalized to generate sample data; The generated sample data is divided according to a preset ratio and used to construct training and testing datasets respectively.

5. The rock permeability prediction method based on deep learning according to claim 1, characterized in that, In step 3, the signal processing flow in the deep learning prediction model based on a fully connected neural network is as follows: The input layer is the receiving end of the test data, which is used to receive the pre-processed test parameter vector , wherein to represent the pre-processed normalized feature values from the th to the th, is the number of normalized feature values in the input parameter, i.e., the test parameter vector . Each neuron in the input layer is used to receive a normalized feature value and pass the received normalized feature value to the neuron in the first hidden layer; The hidden layers have one, the first to the hidden layers contain neurons in the number of to ; make , No. The input to each hidden layer is ;when At that time, the input of the first hidden layer is the output of the input layer. ;when At that time, the first The input to the nth hidden layer is the nth... The output of each hidden layer ; The signal is processed in the first hidden layer in the following order: The processing flow in the first hidden layer is as follows: The input to the first hidden layer is linearly combined with a weight matrix and a bias vector to compute a net input of: ​ ； wherein, represents the net input of the th hidden layer, represents the weight matrix of the th hidden layer, represents the bias vector of the th hidden layer; Using a ReLU activation function for the non-linear mapping, the output of the first hidden layer is ​ ； The input of the output layer is the first The output of each hidden layer The output layer performs a linear weighted calculation to obtain the predicted value of rock permeability. for: ； in, for The weight vector, This is a scalar bias.

6. The rock permeability prediction method based on deep learning according to claim 1, characterized in that, Step 4 specifically involves: Initialize the positions of the particles and set the velocity of each particle; The velocity and position of the particles are iteratively updated using the following formula: ； ； in, , Indicates the number of samples in the training dataset; Indicates the first The sample at the th The velocity of the particle after the next iteration. Indicates the first The sample at the th The velocity of the particle after the next iteration; Indicates the first The sample at the th The position of the particle after the next iteration. Indicates the first The sample at the th The position of the particle after the next iteration; Indicates inertia weight, Inertia is used to control the movement of particles; and Represents the learning factor. and Used to control the attractive force of particles toward individual and global optimum; and Represents a random number between [0, 1]. This indicates the optimal position of an individual particle; This represents the optimal position of the entire swarm, i.e., the global optimal position of the particle swarm. After each iteration of updating the particle's velocity and position, the particle's fitness value is calculated. The particle's fitness value is obtained by using the current particle's position as a parameter of the fully connected neural network and calculating the loss of the deep learning prediction model using the training dataset. The specific loss function is as follows: ； in, Indicates the first The actual permeability of each sample is the true value of the rock permeability. This represents the prediction of the deep learning prediction model. The permeability of each sample is the predicted value of rock permeability, and MSE represents the mean square error. When the fitness value (MSE) of a particle reaches a preset threshold or the number of iterations exceeds the preset maximum number of iterations, the parameter optimization of the fully connected neural network is stopped, the training of the deep learning prediction model is completed, and the deep learning prediction model obtained after this training is used as the current deep learning prediction model.

7. The rock permeability prediction method based on deep learning according to claim 1, characterized in that, Step 5 specifically involves: The test dataset is input into the current deep learning prediction model to predict rock permeability. Based on the true values ​​of rock permeability in the test dataset and the predicted values ​​of rock permeability output by the current deep learning prediction model, the prediction error of the current deep learning prediction model is calculated. Determine whether the prediction error of the current deep learning prediction model is lower than the preset accuracy requirement; If the prediction error of the current deep learning prediction model is lower than the preset accuracy requirement, proceed to step 4 to retrain the current deep learning prediction model. If the prediction error of the current deep learning prediction model is not lower than the preset accuracy requirement, then the current deep learning prediction model is used as the trained deep learning prediction model, and the process proceeds to step 6.

8. The rock permeability prediction method based on deep learning according to claim 1, characterized in that, Step 6 specifically involves: The preprocessed experimental data is used as input data for the trained deep learning prediction model; The input layer receives input data, which is then fed into the hidden layer for feature extraction. The ReLU activation function is used for nonlinear transformation, and the data is then fed into the output layer to predict rock permeability. The rock permeability obtained from the output layer is used as the prediction result output by the trained deep learning prediction model.

9. A computer device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements the steps of the deep learning-based rock permeability prediction method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the deep learning-based rock permeability prediction method as described in any one of claims 1 to 8.

Images