A rapid identification method for rocks during drilling tests based on dynamic numerical simulation

Abstract

The present invention relates to a rapid identification method for rocks during drilling based on dynamic numerical simulation, belonging to the technical field of geotechnical engineering. The method includes drilling different types of standard rock specimens through a measurement-while-drilling (MWD) device to obtain MWD parameters; constructing a dynamic numerical simulation model for MWD; conducting dynamic numerical simulation and simulation tests to obtain dynamic numerical simulation data for MWD; constructing a rapid rock identification neural network model; obtaining small-sample training data for transfer learning, and then performing transfer learning training on the neural network model to construct a rapid identification model for rocks during drilling based on dynamic numerical simulation. The present invention can use the rapid identification model for rocks during drilling based on dynamic numerical simulation to rapidly identify rocks during drilling, with high identification accuracy, a small amount of training data required, and being easy to popularize and apply.

Description

Technical Field

[0001] The present invention belongs to the technical field of geotechnical engineering, and particularly relates to a rapid rock identification method for measurement while drilling based on dynamic numerical simulation. Background Art

[0002] In engineering geological surveys, drilling is the most commonly used survey method. However, the drilling results are only used for stratigraphic logging and providing core samples for laboratory tests. The workload of indoor physical and mechanical property tests is large and the cycle is long. During the drilling process, a large amount of information related to the properties of rock and soil is not utilized. If data is collected through measurement while drilling means and a quantitative relationship between the measurement while drilling data and the properties of rock and soil is established, more test sample data of rock and soil bodies will be obtained, which will greatly improve the efficiency of advanced geological surveys. The test data such as the drilling speed, torque, thrust force, and rotational speed during the drilling process of the drill can be used for rock identification. However, the geological conditions at the engineering site are complex, there are many influencing factors for the test data, and there are too many assumptions in the theoretical model. Currently, it can only be verified under ideal test conditions in the laboratory.

[0003] Machine learning algorithms usually can achieve good prediction effects in the analysis of measurement while drilling data. However, the amount of data required for training the model affects its practicability. In actual situations, it takes a large amount of human and time costs to obtain enough measurement while drilling data and the true calibration data of the rock for corresponding model training. It is difficult to accumulate a large amount of measurement while drilling training data that covers a wide range of rock types and considers various drilling situations.

[0004] Therefore, how to overcome the deficiencies of the existing technology is an urgent problem to be solved in the current technical field of geotechnical engineering. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems that the amount of data required for constructing the existing automatic rock identification algorithm for measurement while drilling is large, but it is difficult to accumulate a large amount of measurement while drilling training data that covers a wide range of rock types and considers various drilling situations through actual drilling tests, and the prediction accuracy is limited. A rapid rock identification method for measurement while drilling based on dynamic numerical simulation is provided. This method accumulates a large amount of dynamic numerical simulation data for measurement while drilling through the dynamic numerical simulation method, and uses the transfer learning method to enable the model established by the dynamic numerical simulation method to perform rapid rock identification for measurement while drilling.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A rapid rock identification method for measurement while drilling based on dynamic numerical simulation, comprising the following steps:

[0008] S01. Drill different types of standard rock specimens with a measurement-while-drilling (MWD) device to obtain MWD parameters. At the same time, conduct uniaxial compression tests on the standard rock specimens to obtain the specific gravity, uniaxial compressive strength, and elastic modulus of the standard rock specimens.

[0009] S02. Construct a dynamic numerical simulation model for MWD with reference to the MWD parameters obtained in S01.

[0010] S03. Use the dynamic numerical simulation model for MWD constructed in S02 to conduct dynamic numerical simulation tests to obtain dynamic numerical simulation data for MWD.

[0011] S04. Construct a neural network model for rapid rock identification based on the dynamic numerical simulation data for MWD obtained in S03.

[0012] S05. Drill known-type rocks with an MWD device of the same bit type as in S01 and obtain MWD parameters as small-sample training data for transfer learning.

[0013] S06. Based on the MWD parameters obtained in S05, conduct transfer learning training on the neural network model constructed in S04 to construct a rapid rock identification model for MWD based on dynamic numerical simulation.

[0014] S07. Conduct MWD on rocks with the same MWD device as in S05 and obtain MWD parameters. Identify the rock type based on the rapid rock identification model for MWD based on dynamic numerical simulation constructed in S06.

[0015] Furthermore, preferably, in S01, during drilling, a constant drilling speed and a constant bit rotation speed mode are adopted; the obtained MWD parameters include the torque, actual drilling speed, and actual rotation speed during drilling of the drill rig.

[0016] Furthermore, preferably, the specific method of S02 is as follows:

[0017] Construct a simulation model of the bit used in the drilling in S01; import the bit simulation model into ANASYS Workbench and use the LS-PrePost platform in LS-Dyna for mesh generation.

[0018] Cut the drilling part of the drill bit; set the drilling material as the MAT_RIGIDG rigid body material in LS-Dyna; use the PRESCRIBED_MOTION_RIGID in LS-Dyna to define the drilling speed of the drill bit in the vertical direction, and use the PRESCRIBED_MOTION_RIGID in the axial direction to define the rotation speed of the drill bit, so as to obtain the drill bit of the dynamic numerical simulation model of the while drilling test; wherein, the setting range of the drilling speed and the drill bit rotation speed needs to be set within the equipment allowable range of the while drilling test equipment used in S01;

[0019] LS-Dyna was used to construct a rock block simulation model for dynamic numerical simulation of the drilling test, and meshing was performed. The HJC constitutive model was used as the constitutive model for rock block construction. At the same time, the parameters of the rock constitutive model were determined by comparing the specific gravity, uniaxial compressive strength and elastic modulus of the standard rock sample obtained by S01. The rock block was given a failure criterion ADD_EROSION, so that the cut rock block was automatically deleted after reaching the failure condition.

[0020] The parameters of the test while drilling dynamic numerical simulation model are compared with the test while drilling parameters obtained in step S01 to ensure that the torque error obtained does not exceed 1.5 (N·m) under the same rock type, uniaxial compressive strength, drill bit drilling speed and drilling speed.

[0021] Further, preferably, the specific method of S03 is:

[0022] The dynamic numerical simulation model of the drilling test constructed in step S02 is used to export the K file of LS-Dyna, and the K files are copied and modified in batches to randomly generate a large number of dynamic numerical simulation models of different drilling speeds, drill bit rotation speeds, and different uniaxial compressive strengths of rock blocks, and a large number of model K files with random parameters are obtained, and then these model K files are batch-run and calculated to obtain a large number of dynamic numerical simulation data of the drilling test; the dynamic numerical simulation data of the drilling test includes torque, drill bit rotation speed and drilling speed;

[0023] Among them, the setting ranges of the drilling speed and the bit rotation speed need to be randomly generated within the equipment allowable range of the measurement-while-drilling test equipment used in step S01, and the uniaxial compressive strength needs to be randomly generated within the uniaxial compressive strength range of the simulated standard rock specimen; the minimum amount of K-file data generated for each rock type should be 500, and the minimum drilling time of the K-file is set to the time required to drill one layer of the rock simulation specimen, and the acquisition interval is set to 0.0002 seconds; the randomly generated dynamic numerical simulation model K-files are batch-calculated to obtain the torque, bit rotation speed, and drilling speed during the drilling of the drill rig in the calculation results of each K-file; among them, the bit rotation speed and the drilling speed are preset values each time the dynamic numerical simulation model of the measurement-while-drilling test is constructed, the torque is obtained by acquiring the Z_moment data of the contact surface between the bit and the rock during the drilling of the bit in RCFORC, and then the sum of the torques is obtained by using the sum_curves function in LS-Dyna PLOT to add the Z_moment of the rock contact surface, and then the sae filter is used to reduce the noise and interference of the data, and the filter cut-off frequency C / S is selected as 60 Hz; then the unstable data in the starting stage and the ending stage during the drilling process are removed, and only the torque data in the interval of 20% to 60% of the drilling time in the dynamic numerical simulation data of each measurement-while-drilling test is intercepted.

[0024] Further, preferably, the specific method of S04 is:

[0025] The rock rapid identification neural network model includes an input layer, a hidden layer, and an output layer;

[0026] The data input into the input layer is the normalized measurement-while-drilling test dynamic numerical simulation data obtained in S03, that is, the normalized bit torque, the normalized bit rotation speed, and the normalized bit drilling speed;

[0027] The hidden layer has 7 layers, which are 4 Batch Normalization layers and 3 RELU activation functions; a RELU activation function is set between two adjacent Batch Normalization layers;

[0028] The activation function of the output layer is softmax; the output of the output layer is the classification probability of each type of rock.

[0029] The binary_crossentropy loss function is used to calculate the loss value; the optimizer of the rock rapid identification neural network model is set to Adma.

[0030] Further, preferably, the specific method adopted for normalization is:

[0031] First, the data x is centered by the mean μ of this type of data using Z-score normalization, and then scaled by the standard deviation σ to obtain the normalized data x. * The normalized data follows a normal distribution with a mean of 0 and a variance of 1. The normalization formula is as follows:

[0032]

[0033] The data mentioned above is the bit torque, bit rotation speed, or penetration rate.

[0034] Furthermore, preferably, in S05, the measurement-while-drilling test parameters include the torque, bit rotation speed, and penetration rate during the drilling process.

[0035] Furthermore, preferably, the specific method of S06 is as follows:

[0036] Load the rock rapid identification neural network model obtained in S04, then freeze the first two RELU activation functions, and use the measurement-while-drilling test parameters obtained in S05 for retraining. After the training is completed, unfreeze it, and use the measurement-while-drilling test parameters obtained in S05 again for the final training. The model obtained after the final training is the rapid rock identification model for measurement-while-drilling based on dynamic numerical simulation.

[0037] Furthermore, preferably, the specific method of S07 is as follows:

[0038] Use the same measurement-while-drilling test equipment as in S05 to conduct measurement-while-drilling tests on the rock, obtain the torque, bit rotation speed, and penetration rate parameters during the drilling process, normalize the measurement-while-drilling test parameters according to the method of S04, and use them as the input data of the rapid rock identification model for measurement-while-drilling based on dynamic numerical simulation in S06 to identify the rock, obtain the classification probability of each type of rock, and take the rock type with the largest classification probability as the final identification result. In S02 of the present invention, if the torque error exceeds 1.5 (N·m), the rock constitutive model parameters need to be modified and corrected until the torque error does not exceed 1.5 (N·m).

[0039] In S02 of the present invention, the measurement-while-drilling dynamic numerical simulation model includes the bit and rock block simulation models of the measurement-while-drilling dynamic numerical simulation model.

[0040] In the present invention, the rock types are classified according to different rock characteristics, such as limestone, granite, and limestone, etc. Different characteristic rocks have different rock mechanical parameters such as specific gravity, compressive strength, and elastic modulus.

[0041] The present invention contemplates separately establishing simulation models for the drill bit and the rock, and using the ANSYS / LS-DYNA software to conduct dynamic numerical simulation tests on the drilling process under different confining pressures, monitoring the drilling response data during the drilling process, and calibrating the dynamic numerical simulation model for measurement-while-drilling by comparing with the measurement-while-drilling parameters obtained from the measurement-while-drilling tests. Based on the calibrated dynamic numerical simulation model, conduct dynamic numerical simulation of the typical rock drilling process to obtain a large amount of measurement-while-drilling data. Through the method of transfer learning, store the knowledge obtained when solving the rock identification task in the dynamic numerical simulation rock identification neural network model, and then perform transfer training on the model based on the small sample of measurement-while-drilling parameters obtained during drilling, so that the model is applicable to the rapid rock identification task in the real drilling process.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The establishment and collection process of the existing real drilling data set is extremely time-consuming and labor-intensive. A large number of drilling experiments need to be carried out on different types of rocks, and the drilled rock samples need to be marked and indoor tests need to be carried out. Taking the indoor drilling platform used as an example, each drilling requires a series of processes such as debugging, drilling, cleaning, and machine cooling. It takes more than 3 hours to complete a complete drilling test. At the same time, if the drilling data is collected at the real engineering site, a series of debugging and drilling work also needs to be carried out, and the rock samples obtained during drilling need to be calibrated to form data available for model training, which will take more time. Therefore, it is often difficult to obtain a large number of data samples for model training. The dynamic numerical simulation established by the present invention based on the large sample data set of the measurement-while-drilling parameters stores the knowledge obtained when solving the rock identification task in the dynamic numerical simulation rock identification neural network model through the method of transfer learning, and then performs transfer training on the model based on the small sample of drilling data obtained during drilling, so that the model is applicable to the rock classification prediction in the real drilling process.

[0044] A neural network model was constructed for a dataset with only 18,411 drilling data entries. The prediction accuracy of the neural network for granite was 91.92%, for sandstone was 80.53%, and for limestone was 76.42%. The prediction accuracies of all three types of rocks were relatively low, and the classification accuracy of rocks with less data was lower. After transfer learning using the neural network model constructed from dynamically simulated numerical data, for the same dataset, the prediction accuracy for granite was 99.82%, for sandstone was 99.53%, and for limestone was 99.36%. Compared with the prediction model without transfer learning, the prediction accuracy of granite increased by 7.9%, that of sandstone increased by 19%, and that of limestone increased by 22.94%. The classification accuracies of all three types of rocks were significantly improved, indicating that the transfer learning method based on dynamically simulated numerical data can greatly improve the problem of low prediction accuracy caused by insufficient training data of real-time drilling test parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 FIG. [X] is a flowchart of the real-time drilling test rock rapid identification method based on dynamic numerical simulation of the present invention;

[0046] Figure 2 FIG. [X] is a rock block model in the real-time drilling test dynamic numerical simulation model;

[0047] Figure 3 FIG. [X] is a structural diagram of the rock rapid identification neural network model;

[0048] Figure 4 FIG. [X] is a transfer learning model;

[0049] Figure 5 FIG. [X] is a schematic diagram of a diamond drill bit;

[0050] Figure 6 FIG. [X] is a diagram of three rock samples; from left to right are limestone, sandstone, and granite in sequence;

[0051] Figure 7 FIG. [X] is a model diagram of a diamond coring drill bit;

[0052] Figure 8 FIG. [X] is a model diagram of a dynamically simulated drill bit;

[0053] Figure 9 FIG. [X] is a model diagram of a dynamically simulated rock block;

[0054] Figure 10 FIG. [X] is a diagram of the initial state of dynamic numerical simulation;

[0055] Figure 11 FIG. [X] is a diagram of S2 and S3 torque data in dynamic numerical simulation;

[0056] Figure 12It is a specific position diagram of surfaces S2 and S3 for dynamic numerical simulation;

[0057] Figure 13 It is a diagram of the combined torque data;

[0058] Figure 14 It is a diagram of the filtered torque data;

[0059] Figure 15 It is a Mises stress nephogram of granite during dynamic numerical simulation drilling;

[0060] Figure 16 It is a diagram of the comparison results of the torque during the indoor test and numerical simulation of constant drilling speed - rotational speed; among them, (a) is granite, (b) is limestone, and (c) is sandstone;

[0061] Figure 17 It is a learning curve diagram of the rock classification results of sandstone, granite, and limestone in dynamic numerical simulation;

[0062] Figure 18 It is a confusion matrix diagram of the test set of the rock classification model in dynamic numerical simulation; among them, (a) is granite, (b) is limestone, and (c) is sandstone;

[0063] Figure 19 It is a diagram of the measurement - while - drilling test results in step S05 of the application example;

[0064] Figure 20 It is a learning curve diagram of the rock classification results based on transfer learning;

[0065] Figure 21 It is a confusion matrix diagram of the predicted data of the rock classification model based on transfer learning.

[0066] Figure 22 It is a confusion matrix diagram of the predicted data of the neural network model for rapid rock identification in example S04. Specific implementation manner

[0067] The present invention will be further described in detail below in conjunction with the embodiments.

[0068] Those skilled in the art will understand that the following embodiments are only used to illustrate the present invention and should not be regarded as limiting the scope of the present invention. For those not specified in the embodiments regarding specific technologies or conditions, they shall be carried out according to the technologies or conditions described in the literature in this field or according to the product specifications. For those materials or equipment without indicating the manufacturer, they are all conventional products that can be obtained by purchase.

[0069] As Figure 1 shown, a method for rapid rock identification during measurement - while - drilling based on dynamic numerical simulation includes the following steps:

[0070] S01. Drill different types of standard rock specimens using a measurement-while-drilling (MWD) device to obtain MWD parameters. Meanwhile, conduct uniaxial compression tests on the standard rock specimens to obtain the specific gravity, uniaxial compressive strength, and elastic modulus of the standard rock specimens.

[0071] S02. Construct a dynamic numerical simulation model for MWD with reference to the MWD parameters obtained in S01.

[0072] S03. Use the dynamic numerical simulation model for MWD constructed in S02 to conduct dynamic numerical simulation tests to obtain dynamic numerical simulation data for MWD.

[0073] S04. Construct a neural network model for rapid rock identification based on the dynamic numerical simulation data for MWD obtained in S03.

[0074] S05. Drill known-type rocks using an MWD device with the same bit type as in S01 and obtain MWD parameters as small-sample training data for transfer learning.

[0075] S06. Based on the MWD parameters obtained in S05, conduct transfer learning training on the neural network model constructed in S04 to construct a rapid rock identification model for MWD based on dynamic numerical simulation.

[0076] S07. Conduct MWD on rocks using the same MWD device as in S05 and obtain MWD parameters. Identify the rock type based on the rapid rock identification model for MWD based on dynamic numerical simulation constructed in S06.

[0077] Among them, the specific method of S01 is as follows:

[0078] Complete the drilling of rock specimens of the type to be predicted using an MWD device that can control the drilling speed and bit rotation speed of the drill rig. Real-time collect the torque, bit rotation speed, and drilling speed parameters during drilling through sensors. Use a constant drilling speed and constant bit rotation speed mode during drilling. The drilling data collection interval is 1 second, and the collected parameters are the torque during drilling of the drill rig, the actual drilling speed, and the actual rotation speed. Conduct uniaxial compression tests on the standard rock specimens, with at least 3 repetitions for each group, to obtain the specific gravity (g / cm 3 ), uniaxial compressive strength (MPa), and elastic modulus (GPa) of the standard rock specimens. The rock types of the standard rock specimens used in the drilling tests are the same as those in the uniaxial compression tests, and they are all rock samples from the same origin and the same batch.

[0079] The specific method of S02 is as follows:

[0080] First, construct the drill bit simulation model according to the real drill bit drawing of the drill bit used in the S01 drilling test. Import the drill bit simulation model into ANASYS Workbench and use the LS-PrePost platform in LS-Dyna for mesh generation.

[0081] Considering the CPU operation time of the dynamic numerical simulation, only the drilling part of the drill bit is intercepted during model construction, and the drill bit material is set as the MAT_RIGIDG rigid material in LS-Dyna. Define the drilling speed of the drill bit in the vertical direction using PRESCRIBED_MOTION_RIGID in LS-Dyna, and define the rotational speed of the drill bit in the axial direction using PRESCRIBED_MOTION_RIGID. The setting ranges of the drilling speed and the drill bit rotational speed need to be set within the equipment allowable range of the measurement-while-drilling equipment selected in step S01.

[0082] Use LS-Dyna to construct a rock block simulation model for the dynamic numerical simulation of measurement-while-drilling and perform mesh generation. To improve the CPU operation time of the dynamic numerical model and reduce the mesh generation data, the rock block is constructed as a cylinder, and the mesh generation is performed separately in the drill bit contact area, as Figure 2 shown by the brown area in the figure.

[0083] Adopt the HJC constitutive model (JOHNSON_HOLMQUIST_CONCRETE) as the constitutive model for constructing the rock block. At the same time, compare with the specific gravity, uniaxial compressive strength, and elastic modulus of the standard rock specimen obtained in S01 to determine the rock constitutive model parameters; the required input constitutive model parameters are shown in Table 1.

[0084] Table 1 HJC constitutive model parameters

[0085]

[0086]

[0087] And set the rock block to assign the failure criterion ADD_EROSION so that the cut rock block is automatically deleted after reaching the failure condition; set the maximum effective strain (Maximum effective strain at failure) EFFEPS at failure in ADD_EROSION, and other parameters are set to default.

[0088] Verify the model parameters with the real measurement-while-drilling parameters obtained in step S01 to ensure that the torque error does not exceed 1.5 (N·m) under the same rock type, uniaxial compressive strength, drill bit rotation speed, and drilling speed.

[0089] The specific method of S03 is as follows:

[0090] Export the K file of LS - Dyna using the dynamic numerical simulation model for measurement - while - drilling constructed in step S02, batch - copy and modify the K file, randomly generate a large number of dynamic numerical simulation models with different drilling speeds, bit rotation speeds, and uniaxial compressive strengths of different rock blocks, and obtain a large number of model K files with randomly generated parameters. Then, batch - run and calculate these model K files to obtain a large amount of dynamic numerical simulation data for measurement - while - drilling. The dynamic numerical simulation data for measurement - while - drilling includes torque, bit rotation speed, and drilling speed.

[0091] Among them, the setting ranges of the drilling speed and the bit rotation speed need to be randomly generated within the equipment - allowed range of the measurement - while - drilling equipment used in step S01, and the uniaxial compressive strength needs to be randomly generated within the uniaxial compressive strength range of the standard rock specimens to be simulated. The minimum amount of data in the K files generated for each rock type should be 500, and the minimum drilling time set for the K file should be the time required to drill 1 layer of rock simulation specimens. The acquisition interval is set to 0.0002 seconds. Batch - calculate the randomly generated dynamic numerical simulation model K files to obtain the torque, bit rotation speed, and drilling speed during the drilling of the drill rig in the calculation results of each K file. Among them, the bit rotation speed and the drilling speed are the preset values during the construction of each dynamic numerical simulation model for measurement - while - drilling. The torque is obtained by acquiring the Z_moment data of the contact surface between the bit and the rock during drilling in RCFORC, and then the sum of the torques is obtained by using the sum_curves function in LS - Dyna PLOT to add the Z_moment of the rock contact surface. Then, the sae filter is used to reduce noise and suppress interference, and the cut - off frequency C / S of the filter is selected as 60 Hz. Then, the unstable data in the starting and ending stages of the drilling process are removed, and only the torque data in the interval of 20% to 60% of the drilling time in each dynamic numerical simulation data for measurement - while - drilling is intercepted.

[0092] The specific method of S04 is as follows:

[0093] The rock rapid - identification neural network model used is as Figure 3As shown, the data input into the model's input layer are the bit torque, bit rotation speed, and bit penetration rate, which are obtained by the method described in S03. The model has 7 hidden layers, namely 4 Batch Normalization layers and 3 RELU activation functions, with each RELU activation function set to 500 nodes. The activation function of the model's output layer is softmax. The rock classification labels are converted into binary vectors using one-hot encoding, such as granite [1,0,0], sandstone [0,1,0], limestone [0,0,1], and the classification probabilities of each type of rock are output using Softmax. Finally, the binary_crossentropy loss function is used to calculate the loss value, and the weight and threshold values of each node are returned and adjusted. The optimizer of the model is set to Adma.

[0094] And before training, the data x is centralized by the mean μ of this type of data using Z-score normalization and then scaled by the standard deviation σ to obtain the normalized data x * ; The normalized data follows a normal distribution with a mean of 0 and a variance of 1. The normalization formula is as follows:

[0095]

[0096] The data mentioned above are the bit torque, bit rotation speed, or penetration rate.

[0097] The specific method of S05 is:

[0098] Drill known type rock blocks using the same type of measurement-while-drilling equipment as in S01 and obtain the measurement-while-drilling parameters as the small sample training data for transfer learning; the measurement-while-drilling parameters include the torque, bit rotation speed, and penetration rate parameters during drilling; the data acquisition interval for drilling is 1 second. These data are used as the small sample training data for transfer learning in step S06.

[0099] The specific method of S06 is:

[0100] The process of transfer learning includes the following steps: (1) Use the rock rapid identification neural network model based on dynamic numerical simulation data obtained in step S04 as the base model for transfer learning, and copy the weights and architecture of the model to the measurement-while-drilling rock rapid identification task. (2) Freeze a part of the model in the copied model to retain the learned features. (3) Make minor adjustments to the model, unfreeze some layers and further train on the target dataset.

[0101] The present invention first loads the rock rapid identification neural network model obtained in S04, and retains the neural network model structure of the input layer, hidden layer, and output layer in the model. Then, the first two RELU activation functions ( Figure 4 ) are frozen, and the in-the-hole test parameters obtained during the actual drilling process (obtained in S05) are used for retraining. After the training is completed, the freezing is cancelled, and the in-the-hole test parameters obtained in S05 are used again for the final training. The trained model is the final in-the-hole test rock rapid identification model based on dynamic numerical simulation. The optimizer and loss function used in the model are the same as those established in step S04 for the simulation.

[0102] The specific method of S07 is as follows:

[0103] The in-the-hole test of the rock is carried out using the same in-the-hole test equipment as in S05 to obtain the torque, bit rotation speed, and drilling speed parameters during the drilling process. The in-the-hole test parameters are normalized according to the method of S04, and used as the input data of the in-the-hole test rock rapid identification model based on dynamic numerical simulation in S06 to identify the rock, obtaining the classification probability of each type of rock, and taking the rock type with the largest classification probability as the final identification result.

[0104] Application Example

[0105] A method for rapid identification of in-the-hole test rocks based on dynamic numerical simulation, characterized by including the following steps:

[0106] S01. Drill different types of standard rock specimens through the in-the-hole test equipment to obtain in-the-hole test parameters; at the same time, conduct uniaxial compression tests on the standard rock specimens to obtain the specific gravity, uniaxial compressive strength, and elastic modulus of the standard rock specimens;

[0107] S02. Construct an in-the-hole test dynamic numerical simulation model with reference to the in-the-hole test parameters obtained in S01;

[0108] S03. Use the constructed in-the-hole test dynamic numerical simulation model in S02 to conduct dynamic numerical simulation and simulation tests to obtain in-the-hole test dynamic numerical simulation data;

[0109] S04. Construct a rock rapid identification neural network model based on the in-the-hole test dynamic numerical simulation data obtained in S03;

[0110] S05. Drill known type rocks through the in-the-hole test equipment with the same bit type as in S01, and obtain in-the-hole test parameters as the small sample training data for transfer learning;

[0111] S06. Conduct transfer learning training on the neural network model constructed in S04 based on the in-the-hole test parameters obtained in S05 to construct an in-the-hole test rock rapid identification model based on dynamic numerical simulation;

[0112] S07. Performing a while-drilling test on the rock using the same while-drilling test equipment as in S05 and obtaining the while-drilling test parameters, the rock type is identified based on the while-drilling test rock rapid identification model based on dynamic numerical simulation constructed in S06.

[0113] In this example S01, a small indoor drilling parameter rapid acquisition device (CN201920897925.9) is used to complete the rock drilling test. The system controls the drilling of the drill bit by rotation and thrust, and collects the thrust, torque, speed, drilling speed and other parameters in the drilling process in real time through sensors. When the system drills the rock and soil specimen, the drilling speed, drill bit speed, drill rod torque and thrust can be controlled, and any of the following four combined control modes can be realized: constant drilling rate-constant speed mode; constant drilling rate-constant torque mode; constant thrust-constant speed mode; constant thrust-constant torque mode. The indoor drilling test system mainly includes a servo-controlled drilling system, a servo-controlled confining pressure system, a drilling test parameter acquisition system and a control system. The main technical parameters of the drilling test system are shown in Table 2. The drilling test system collects drilling test parameters in real time by installing torque and speed sensors, drilling pressure sensors, confining pressure sensors and drilling speed sensors. The sensor parameters are shown in Table 3.

[0114] Table 2 Main technical parameters of indoor drilling test system

[0115]

[0116]

[0117] Table 3 Sensor parameters

[0118]

[0119] The drilling test used a three-edge three-sprue flat-tooth impregnated diamond core drill bit with an outer diameter of 32 mm, an inner diameter of 24 mm, and a nozzle spacing of 6 mm. Figure 5 shown.

[0120] This example uses three types of rock samples: granite, sandstone and limestone ( Figure 6 ) for indoor drilling test. The sample size is 100mm×100mm×100mm, the surface is smooth and complete, and there are no obvious cracks. Cylindrical standard rock samples made from the same batch of rocks were used for uniaxial compression test on the RMT testing machine. The sample was placed at the center of the pressure plate of the press, and the pressure plate with a spherical seat was adjusted to make the sample evenly stressed. The control mode was displacement loading, and the loading rate was 0.0020mm / s until the sample was destroyed. Each group of tests was repeated at least 3 groups, and the three rock mechanical parameters were obtained as shown in Table 4.

[0121] Table 4 Rock Mechanics Parameters

[0122]

[0123]

[0124] Multiple drilling experiments were carried out on three types of rock samples in the constant drilling rate - constant rotation speed mode. The basic parameters of the drilling tests are shown in Table 5. The types of rocks used in the drilling tests are the same as those in the compression tests, and they are all rock samples from the same origin and the same batch. In each drilling test, the drilling depth D is 100 mm, and the constant bit rotation speeds N are set to 400, 600, and 800 RPM respectively. Among them, the constant drilling speed v of the granite drilling experiment is set to vary from 5 to 40 mm / min, and the constant drilling speed v of the limestone and sandstone is set to vary from 20 to 40 mm / min. During drilling, the data acquisition interval for the in - situ measurement is 1 second, and the acquisition parameters are the torque, the true drilling speed, and the true rotation speed of the drill rig during drilling.

[0125] Table 5 Basic Parameters of Drilling Experiments

[0126]

[0127]

[0128] In this example S02, first, according to the drawing of the real diamond coring bit, a bit model with the same size as the in - situ measurement during drilling is constructed using SolidWorks software ( Figure 7 ).

[0129] The model is imported into ANASYS Workbench, the bit model is simplified using SpaceClaim, and meshing is carried out using the LS - PrePost platform in LS - Dyna ( Figure 8 ), and the bit material is set as the MAT_RIGIDG rigid material.

[0130] The drilling speed of the bit is defined in the vertical direction using the keyword PRESCRIBED_MOTION_RIGID, and the specific parameters are shown in Table 6.

[0131] Table 6 PRESCRIBED_MOTION_RIGID Parameters of Drilling Speed

[0132] DOF VAD SF DEATH BIRTH 3 2 1 0 0

[0133] The rotation speed of the bit is defined in the axial direction using the keyword PRESCRIBED_MOTION_RIGID, and the specific parameters are shown in Table 7.

[0134] Table 7 Rotational Speed PRESCRIBED_MOTION_RIGID Parameters

[0135] DOF VAD SF DEATH BIRTH 7 0 1 0 0

[0136] Using LS-Dyna to construct the rock block for the dynamic numerical simulation of measurement-while-drilling and perform mesh generation ( Figure 9 ), the height of the rock block is 9 cm and the radius is 23 cm.

[0137] By comparing with the specific gravity, uniaxial compressive strength and elastic modulus of the standard rock specimens obtained by S01, the constitutive model parameters of limestone, granite and sandstone were determined (Tables 8 - 10).

[0138] Table 8 HJC Model Parameters of Limestone

[0139]

[0140]

[0141] Table 9 HJC Model Parameters of Granite

[0142] <![CDATA[RO(kg / m 3 )]]> G (Pa) A B C N FC (MPa) 2680 7.610e+10 0.75 2.36 0.049 0.78 241.41 T (Pa) EPS0 εfmin Sfmax Pc (pa) μc PL (Pa) 1.610e+07 2.800e-05 0.015 5.4 6.300e+06 9.000e-04 1.040e+09 μL D1 D2 K1 (Pa) K2 (Pa) K3 (Pa) FS 0.1 0.046 1.02 8.600e+10 -1.730e+11 2.100e+11 1000

[0143] Table 10 HJC Model Parameters of Sandstone

[0144] <![CDATA[RO(kg / m 3 )]]> G (Pa) A B C N FC (MPa) 2416 5.670e+09 0.32 1.76 0.0127 0.79 136.00 T (Pa) EPS0 εfmin Sfmax Pc (pa) μc PL (Pa) 4.600e+06 1 0.01 7 1.833e+07 0.034 8.000e+08 μL D1 D2 K1 (Pa) K2 (Pa) K3 (Pa) FS 0.08 0.045 1 8.100e+10 -9.100e+10 8.900e+10 1.34

[0145] In the failure criterion ADD_EROSION, the maximum effective strain failure condition EFFEPS of granite is set to 0.022, the maximum effective strain failure condition EFFEPS of sandstone is set to 0.038, and the maximum effective strain failure condition of limestone is set to 0.025. At the initial state, the drill bit starts drilling close to the rock block ( Figure 10 ).

[0146] In this example S03, the K file of LS-Dyna is exported using the dynamic numerical simulation model constructed in step S02, and the K file is batch copied and modified. A large number of dynamic numerical simulation models with different drilling speeds, drill bit rotational speeds, and different uniaxial compressive strengths of rock blocks are randomly generated to obtain a large number of model K files with random parameters. Then, these model files are batch run and calculated to obtain a large amount of dynamic numerical simulation data for measurement-while-drilling.

[0147] In this example, the random generation range of the drilling speed V is 5 to 50 mm / min, the random generation range of the drill bit rotation speed N is 400 to 800 RPM, the random generation range of the uniaxial compressive strength FC of granite is 100 to 300 MPa, the random generation range of the uniaxial compressive strength of sandstone is 30 to 100 MPa, the random generation range of the uniaxial compressive strength of limestone is 60 to 170 Mpa. Other HJC model parameters of the rock are consistent with Table 8 - 10, only changing the uniaxial compressive strength FC parameter.

[0148] In this example, the set drilling time is 0.025 seconds, the time is greater than the time required to drill 1 layer of the model, and the acquisition interval is set to 0.0002 seconds. Obtain the torque, drill bit rotation speed, and drilling speed during the drilling of the drill rig. The drill bit rotation speed and drilling speed are preset constant values constructed for each simulation. The torque is obtained by acquiring the Z_moment data of the bottom surfaces S2 and S3 during the drilling of the drill bit in RCFORC ( Figure 11 ), and the specific parts represented by S2 and S3 are shown by Figure 12 . Then, use the sum_curves function in LS-Dyna PLOT to add the Z_moment of the two surfaces to obtain the torque sum ( Figure 13 ). Then, use the sae filter with a cut-off frequency C / S of 60 Hz to achieve noise reduction and interference suppression of the data. The finally obtained torque is as shown in Figure 14 . In this example, the data in the starting and ending stages of the drilling process are excluded, and only the torque data between 0.005 and 0.015 for each simulated drilling are intercepted.

[0149] In this example, 500 dynamic numerical simulation models of granite, sandstone, and limestone with different drilling speeds, drill rates, and rock masses strengths are randomly generated. The random generation range of the drilling speed V is 5 to 50 mm / min, the random generation range of the drill bit rotation speed N is 400 to 800 RPM, the random generation range of the uniaxial compressive strength of granite is 100 to 300 MPa, the random generation range of the uniaxial compressive strength of sandstone is 30 to 100 MPa, and the random generation range of the uniaxial compressive strength of limestone is 60 to 170 Mpa. Observe the Mises stress at different moments when the drill bit starts to penetrate the rock mass as shown in Figure 15 . At the beginning of the drilling process, in the rock-breaking penetration process, first, the drill teeth approach and press into the rock mass surface under the vertical pressure. At the same time, the rotation of the drill bit driven by the torque drives the drill teeth to cut the soil. In the subsequent process, under the action of the above two main effects and other frictional resistances, the drill bit as a whole continuously penetrates into the rock mass; at the same time, the interaction area and the acting force between the drill bit and the rock mass also change at all times.

[0150] Figure 16This is the comparison result of the indoor drilling experiment and numerical simulation torque for granite, limestone, and sandstone when the drilling speed V is 20 mm / min and the rotational speed N is 400 RPM. It can be seen that the comparison error between the indoor test and the numerical simulation torque does not exceed 1.5 (N·m).

[0151] In this example S04, before training, each data x is first centered by the mean μ using Z-score normalization and then scaled by the standard deviation σ to obtain the normalized data x * , and the data follows a normal distribution with a mean of 0 and a variance of 1.

[0152] The dynamic numerical simulation rock rapid identification neural network model used in this example is as Figure 3 shown. The data input into the model's inputlayer is the bit torque, bit rotational speed, and bit drilling speed, and these data are obtained by the method described in S03. The model has 7 hidden layers, including 4 Batch Normalization layers and 3 RELU activation functions, with each RELU activation function set to 500 nodes. The activation function of the model's output layer is softmax. The rock block classification labels are converted into binary vectors using one-hot encoding. In this example, granite is [1, 0, 0], sandstone is [0, 1, 0], and limestone is [0, 0, 1]. The classification probability of each type of rock is output using Softmax, and finally, the binary_crossentropy loss function is used to calculate the loss value, and the weight and threshold values of each node are returned and adjusted. The optimizer of the model is set to Adma. The training and prediction of the neural network prediction model are implemented using the keras framework of Python.

[0153] In this example, 10% is used as the test set data, 10% is used as the validation set data, and 80% is used as the training set data. The training set is used for model training, the validation set is used to verify the effect of the model during the training process, and the test set data is used to verify the final prediction accuracy of the model. The learning curve of the prediction model is as Figure 17As shown in the figure, there are 104940 training data, including numerical simulation data of three types of rocks. The number of model training rounds is 1000 epochs, and the batch size is 5000. From the learning curve, it can be seen that the prediction accuracy has been decreasing with the progress of training, indicating that the model is established correctly and is in a state where it can be trained correctly. The training set accuracy curve and the validation set accuracy curve of the three rock classification prediction models are relatively close, indicating that there is no overfitting or underfitting in the training. The prediction accuracy of the training set is 99.25%, the prediction accuracy of the validation set is 99.42, and the prediction accuracy of the test set is 99.41%. The prediction accuracy of each type of rock is relatively ideal.

[0154] Figure 18 is the confusion matrix of the test set, which indicates the number of correct and incorrect predictions for different rock types. The neural network has a prediction accuracy of 99.82% for granite with label 0, 99.53% for sandstone with label 1, and 99.36% for limestone with label 3. In summary, it can be shown that the neural network model based on dynamic numerical simulation results has high classification accuracy for the three types of rocks.

[0155] In this example S05, three kinds of rock samples, granite, limestone and sandstone, were subjected to multiple constant drilling rate-constant rotation speed while drilling test drilling experiments, and the while drilling test parameters were obtained, including the real torque, drilling speed and drill bit rotation speed of the drilling rig. The drilling results of 14 granite, 6 sandstone and 6 limestone were collected at intervals of 1 second. Figure 19 As shown. Figure 19 It can be seen that the torque is mainly distributed between 1.4 and 6.4 (N·m) when drilling granite, between 1 and 5.9 (N·m) when drilling limestone, and between 1.4 and 2.6 (N·m) when drilling sandstone. Figure 19 It can be seen from the results that the drilling speed is proportional to the torque, and the drilling rig speed is inversely proportional to the torque in the three types of rock drilling results. This result is consistent with reality.

[0156] In this example S06, the rock rapid identification neural network model obtained in S04 is first loaded, and the neural network model structure of the input layer, hidden layer and output layer in the model is retained. Then the RE LU activation function of the first two layers is frozen ( Figure 4 ), and use the drilling test parameters obtained in this example S05 for retraining, unfreeze after the training, and use S05 to obtain the drilling test parameters for final training. The trained model is the final drilling test rock rapid identification model based on dynamic numerical simulation, and the optimizer and loss function used in the model are consistent with the simulation established in step S04.

[0157] In this example, 10% of the drilling data is used as the test set data, 10% as the validation set data, and 80% as the training set data. The training set is used for model training, the validation set is used to verify the effect of the model during the training process, and the test set data is used to verify the final prediction accuracy of the model. The learning curves of the three rock prediction models are shown in Figure 1. Figure 20 The model training round number is 1000 epochs, and the batch size is 5000. From the learning curve, it can be seen that the prediction accuracy has been decreasing with the progress of training, indicating that the model is established correctly and is in a state where it can be trained correctly. The training set accuracy curve and the validation set accuracy curve of the three rock classification prediction models are relatively close, and the prediction accuracy of the training set is 98.40%.

[0158] In this example S07, the same test while drilling equipment as in example S05 is used to drill granite, sandstone and limestone, and the real torque, drilling speed and drill bit speed of the drilling rig are obtained. The test while drilling parameters are normalized and the data are input as input parameters into the test while drilling rock rapid identification model based on dynamic numerical simulation constructed in example S06 to obtain the results of various rock types. The final prediction accuracy is 97.50%. Figure 21 is the confusion matrix of the prediction results. The prediction accuracy of the neural network in predicting label 0 granite is 99.82%, the prediction accuracy of label 1 sandstone is 99.53%, and the prediction accuracy of label 3 limestone is 99.36%. If only the neural network prediction model established by S04 is used, the prediction accuracy is only 77%. Figure 22 The confusion matrix of the prediction results of the neural network prediction model established for S04 shows that the prediction accuracy of the present invention is improved by 20.5% without adding additional real drilling test parameters. Compared with the prediction model without transfer learning, the prediction accuracy of granite is improved by 7.9%, the prediction accuracy of sandstone is improved by 19%, and the prediction accuracy of limestone is improved by 22.94%. The classification accuracy of the three rocks has been greatly improved, indicating that the use of transfer learning methods can greatly improve the problem of low prediction accuracy caused by insufficient training data of real drilling test parameters.

[0159] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A rapid identification method for rocks during drilling tests based on dynamic numerical simulation, characterized in that It includes the following steps: S01. Drill different types of standard rock specimens through a measurement-while-drilling (MWD) device to obtain MWD parameters. Meanwhile, conduct uniaxial compression tests on the standard rock specimens to obtain the specific gravity, uniaxial compressive strength, and elastic modulus of the standard rock specimens; S02. Construct a dynamic numerical simulation model for MWD with reference to the MWD parameters obtained in S01; S03. Use the dynamic numerical simulation model for MWD constructed in S02 to conduct dynamic numerical simulation tests to obtain dynamic numerical simulation data for MWD; S04. Construct a neural network model for rapid rock identification based on the dynamic numerical simulation data for MWD obtained in S03; S05. Drill known-type rocks through an MWD device with the same bit type as in S01 and obtain MWD parameters as small-sample training data for transfer learning; S06. Based on the MWD parameters obtained in S05, conduct transfer learning training on the neural network model constructed in S04 to construct a rapid rock identification model for MWD based on dynamic numerical simulation; S07. Conduct MWD on rocks through the same MWD device as in S05 and obtain MWD parameters. Identify the rock type based on the rapid rock identification model for MWD based on dynamic numerical simulation constructed in S06; During drilling, adopt a constant drilling speed and constant bit rotation speed mode. The MWD parameters obtained include the torque, actual drilling speed, and actual rotation speed during the drilling of the drill rig; The specific method of S02 is as follows: Construct a simulation model of the bit used in the drilling in S01. Import the bit simulation model into ANASYS Workbench and use the LS-PrePost platform in LS-Dyna for mesh generation; Intercept the drilling part of the bit. Set the drilling material as the MAT_RIGIDG rigid material in LS-Dyna. Define the drilling speed of the bit in the vertical direction using PRESCRIBED_MOTION_RIGID in LS-Dyna and define the rotation speed of the bit in the axial direction using PRESCRIBED_MOTION_RIGID, so as to obtain the bit of the dynamic numerical simulation model for MWD. Among them, the setting ranges of the drilling speed and the bit rotation speed need to be set within the allowable range of the MWD device used in S01; Use LS-Dyna to construct a simulation model of the rock block for dynamic numerical simulation of MWD and conduct mesh generation. Adopt the HJC constitutive model as the constitutive model for constructing the rock block. Meanwhile, compare with the specific gravity, uniaxial compressive strength, and elastic modulus of the standard rock specimens obtained in S01 to determine the rock constitutive model parameters. And set the rock block to be given the failure criterion ADD_EROSION so that the cut rock block is automatically deleted after reaching the failure condition; Compare the parameters of the dynamic numerical simulation model for MWD with the MWD parameters obtained in step S01 to ensure that the torque error does not exceed 1.5 (N·m) under the same rock type, uniaxial compressive strength, bit drilling speed, and drilling speed; 2. The rapid identification method of in - situ tested rock based on dynamic numerical simulation according to claim 1, wherein, The specific method of S03 is as follows: Export the K file of LS-Dyna using the dynamic numerical simulation model for measurement-while-drilling (MWD) constructed in step S02, batch copy and modify the K file, randomly generate a large number of dynamic numerical simulation models with different drilling speeds, bit rotation speeds, and uniaxial compressive strengths of different rock blocks, and obtain a large number of model K files with randomly generated parameters. Then, batch run and calculate these model K files to obtain a large amount of dynamic numerical simulation data for MWD. The dynamic numerical simulation data for MWD includes torque, bit rotation speed, and drilling speed. Among them, the setting ranges of the drilling speed and the bit rotation speed need to be randomly generated within the equipment allowable range of the MWD equipment used in step S01, and the uniaxial compressive strength needs to be randomly generated within the uniaxial compressive strength range of the standard rock specimen being simulated. The minimum amount of data for each K file generated for each rock type should be 500, and the minimum drilling time for the K file is set to the time required to drill one layer of the rock simulation specimen, and the acquisition interval is set to 0.0002 seconds. Batch calculate the randomly generated dynamic numerical simulation model K files to obtain the torque, bit rotation speed, and drilling speed during the drilling of the drill rig in the calculation results of each K file. Among them, the bit rotation speed and the drilling speed are the preset values during the construction of each dynamic numerical simulation model for MWD. The torque is obtained by acquiring the Z_moment data of the contact surface between the bit and the rock during drilling in RCFORC, and then the sum_curves function in LS-DynaPLOT is used to add the Z_moment of the rock contact surface to obtain the torque sum. Then, the sae filter is used to reduce noise and suppress interference, and the filter cut-off frequency C / S is selected as 60 Hz. Then, eliminate the unstable data in the starting stage and the ending stage during the drilling process, and only intercept the torque data in the interval of 20% to 60% of the drilling time in each dynamic numerical simulation data for MWD.

3. The rapid rock identification method while drilling based on dynamic numerical simulation according to claim 1, characterized in that, The specific method of S04 is as follows: The rock rapid identification neural network model includes an input layer, a hidden layer, and an output layer. The data input into the input layer is the normalized dynamic numerical simulation data for MWD obtained in S03, that is, the normalized bit torque, the normalized bit rotation speed, and the normalized drilling speed of the bit. The hidden layer has 7 layers, which are 4 BatchNormalization layers and 3 RELU activation functions respectively; a RELU activation function is set between two adjacent BatchNormalization layers. The activation function of the output layer is softmax; the output of the output layer is the classification probability of each type of rock. The binary_crossentropy loss function is used to calculate the loss value; the optimizer of the rock rapid identification neural network model is set to Adma.

4. The rapid identification method of formation rocks while drilling based on dynamic numerical simulation according to claim 3, characterized in that, The specific method adopted for normalization is as follows: First, use Z-score normalization to center the data x by the mean μ of this type of data, and then scale it by the standard deviation σ to obtain the normalized data x * ; The normalized data follows a normal distribution with a mean of 0 and a variance of 1. The normalization formula is as follows: The data is the bit torque, the bit rotation speed, or the drilling speed.

5. The rapid identification method of formation rocks while drilling based on dynamic numerical simulation according to claim 1, wherein In S05, the MWD parameters include the torque, the bit rotation speed, and the drilling speed during the drilling process.

6. The rapid rock identification method during drilling based on dynamic numerical simulation according to claim 1, wherein The specific method of S06 is as follows: Load the rock rapid identification neural network model obtained in S04, then freeze the first two RELU activation functions, and use the measurement-while-drilling parameters obtained in S05 for retraining. After the training is completed, unfreeze it, and use the measurement-while-drilling parameters obtained in S05 again for final training. The model obtained after the final training is the rock rapid identification model based on dynamic numerical simulation for measurement-while-drilling.

7. The rapid identification method of in - situ testing rock based on dynamic numerical simulation according to claim 1, characterized in that, The specific method of S07 is as follows: Use the same measurement-while-drilling equipment in S05 to conduct measurement-while-drilling on the rock, obtain the torque, bit rotation speed, and penetration rate parameters during the drilling process, normalize the measurement-while-drilling parameters according to the method in S04, and use them as the input data of the rock rapid identification model based on dynamic numerical simulation in S06 to identify the rock, obtain the classification probability of each type of rock, and take the rock type with the largest classification probability as the final identification result.

Images