Rock porosity prediction method based on inversion of few-shot while-drilling data

Abstract

Disclosed in the present invention is a rock porosity prediction method based on the inversion of few-shot while-drilling data. While-drilling parameters of a drilling rig during a laboratory or on-site drilling process are collected, wherein the while-drilling parameters comprise parameters such as torque M, thrust force F, rotation speed N, penetration rate V, drill pipe amplitude A and vibration acceleration a; different types of while-drilling parameter data which have been subjected to multiple instances of denoising processing are inputted into a BP-GA model for outlier removal, data augmentation and iterative calculation, so as to form a new while-drilling data set; and while-drilling parameter time-frequency domain feature maps of rocks having different porosities are inputted into a VG-CNN prediction model, an inversion model between while-drilling parameter time-frequency domain features and the porosities is obtained by means of training and learning, and finally inversion prediction is performed on the rock porosities on the basis of on-site real-time while-drilling parameters. The present method enables fast, accurate and quantitative inversion and prediction of rock porosities on the basis of a small amount of borehole while-drilling data.

Description

A method for predicting rock porosity based on inversion of small sample drilling data Technical Field

[0001] This invention relates to a method for predicting rock porosity in underground engineering, specifically a method for predicting rock porosity based on small sample drilling data. Background Technology

[0002] Porosity is an important indicator characterizing the physical and mechanical parameters of rocks, such as strength, elastic modulus, and permeability. Rapidly determining rock porosity in underground tunnel engineering is crucial for selecting support methods and determining support parameters. Traditional methods for measuring rock porosity require processing cored rock samples into test samples and conducting laboratory testing and analysis. In fields such as oil drilling, analyzing well logging information allows for detailed characterization and quantitative inversion of oil and gas reservoir porosity. However, for shallow tunnel engineering (such as subways, tunnels, and underground tunnels), conducting well logging and well logging work is costly and cumbersome. Therefore, a quantitative, rapid, and economical in-situ method for measuring rock porosity is needed.

[0003] Currently, testing while drilling (WSDB) technology has gradually become a hot topic in the field of intelligent exploration for geotechnical engineering. WSDB technology can reflect the mechanical properties of surrounding rock by utilizing the changing characteristics of drilling parameters during the drilling process, and can identify lithology and structural features of strata. It can effectively compensate for the time lag of traditional measurement methods and does not affect on-site engineering construction, making it a convenient and rapid in-situ measurement method. Preliminary research results in the field of WSDB have verified the feasibility of inverting rock strength and identifying rock mass structural planes based on WSDB parameters. However, due to the complexity of the drill bit cutting process with the rock, and the small amount of effective data and susceptibility to external environmental interference during the on-site acquisition of WSDB parameters, it is difficult to predict rock porosity using WSDB parameters. Existing literature rarely includes methods for inverting and predicting the porosity of roadway surrounding rock, and there are no reports on how to accurately invert and predict rock porosity using small sample WSDB data. Therefore, it is necessary to establish a scientific and reasonable method for inverting and predicting rock porosity during drilling. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a method for predicting rock porosity based on small sample drilling data inversion. This method can rapidly and quantitatively predict rock porosity using small sample drilling data from laboratory or field drilling processes.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] This invention provides a method for predicting rock porosity based on small sample drilling data inversion, comprising the following steps:

[0007] Step S1: Collect drilling information during the drilling process of the drilling rig in the laboratory or underground tunnel, and obtain key drilling parameters such as torque M, thrust F, rotational speed N, thrust speed V, drill pipe amplitude A, and vibration acceleration a.

[0008] Step S2: Perform multiple noise reduction processing on the collected different types of drilling parameters, and input the multiple noise-reduced drilling parameter values ​​into the "BP-GA" model to obtain a new drilling parameter derived dataset;

[0009] Step S3: Analyze the variation curves of different types of drilling parameters with drilling time t in the new drilling parameter dataset to obtain the time-frequency domain map of drilling parameters. Use the data augmentation function in the neural network to obtain the time-frequency domain image library of different types of drilling parameters.

[0010] Step S4: Construct a convolutional neural network prediction model for rock porosity, learn and train on a time-frequency domain feature image library of drilling parameters for rocks with different porosities, determine the initial learning rate and decay coefficient of the model, and obtain the rock porosity prediction results.

[0011] Preferably, the implementation process of the multiple noise reduction methods for different types of drilling parameters in step S2 includes the following steps:

[0012] Step 1.1: Based on the raw data of different types of drilling parameters acquired continuously, calculate the extreme point m of the raw data for each type of drilling parameter. i Average value n i and local amplitude a i ;

[0013] The original drilling data x(t) is separated using a local mean function a(t), and then demodulated using an envelope estimation function m(t): h(t)=x(t)-a(t) (3)

[0014] Repeat the above operations until the original drilling data x(t) is reduced to a monotonic function or has ≤3 extreme points, then stop the iterative calculation. At this point, the original drilling data x(t) is decomposed into k decision components PF and a decomposition margin.

[0015] Step 1.2: Using kurtosis H1, correlation coefficient H2, and root mean square value H3 as selection criteria for PF components, the evaluation index V of the PF components is then determined. ij Represented as:

[0016] In the formula, j takes values ​​of 1, 2, 3, ..., k;

[0017] V is defined here ijMore than 90% of the PF components are valid components;

[0018] Step 1.3: The selected PF components are further denoised using wavelet denoising. This includes wavelet denoising of the effective PF components of torque M, thrust F, rotational speed N, thrust velocity V, drill pipe amplitude A, and vibration acceleration a. The wavelet denoising function used for torque and rotational speed is the db3 wavelet denoising function, the wavelet denoising function used for thrust and thrust velocity is the coif3 wavelet denoising function, and the wavelet denoising function used for drill pipe amplitude and vibration acceleration is the bior3.3 wavelet denoising function.

[0019] Step 1.4: Input the denoised drilling parameter data into the "BP-GA" model for iteration, cross-validation, and mutation to achieve data augmentation and data derivation.

[0020] Preferably, in step S2, a "BP-GA" model is established to address the problem of insufficient time-frequency domain plot sample data due to difficulties in on-site data acquisition for the above six types of drilling parameters. The specific method is as follows:

[0021] Step 2.1: Use a backpropagation neural network (BP) to determine the collected drilling information samples and expected output values. Use an adaptive algorithm to determine the initial weights and thresholds of the BP neural network corresponding to each type of drilling parameter, thereby eliminating unreasonable data.

[0022] Step 2.2: For the six types of drilling parameters after removing unreasonable data, a genetic algorithm is used for iteration, crossover and data mutation. The crossover probability is 0.75 and the coefficient of mutation is 0.031. The new drilling parameter data groups generated after encoding are M1, M2, ..., M6.

[0023] Step 2.3: Establish relationship curves between the newly generated drilling parameter data and the drilling time t, and then obtain the time-frequency domain feature maps of different types of drilling parameters;

[0024] Step 2.4: Utilize the data augmentation function in the neural network to randomly flip, rotate clockwise or counterclockwise, change brightness, and increase pixel value of the time-frequency domain feature map of each type of drilling parameter to expand the time-frequency domain image library of drilling parameters until there are ≥1200 time-frequency domain feature images corresponding to each type of drilling parameter, and then stop the iterative calculation of the neural network.

[0025] Preferably, in step S3, the method for obtaining time-frequency domain feature maps of different types of drilling parameters includes the following steps:

[0026] Step 3.1: Based on the drilling information processed by multiple noise reduction and "BP-GA" model, establish the variation curves of different types of drilling parameters with drilling time t, that is, obtain the time domain feature map of the drilling parameters.

[0027] Step 3.2: The time-domain characteristic map of the drilling parameters is transformed into a frequency-domain characteristic map using discrete Fourier transform. The transformation equation is as follows:

[0028] In the formula, Y(n) represents a finite-length discrete frequency sequence of length L, n = 0, 1, ..., L-1, Hz, and y(t) represents a discrete time sequence of length T.

[0029] Step 3.3: The time-domain and frequency-domain images of the drilling parameters are fused by short-time Fourier transform; that is, time is used as the horizontal axis and the acquisition frequency and instantaneous energy change of the drilling parameters are used as the vertical axis to obtain the time-frequency domain feature images of different types of drilling parameters. The expanded time-frequency domain image library of drilling parameters is then input into the rock porosity convolutional neural network prediction model, namely the VG-CNN prediction model.

[0030] Preferably, in step S4, the specific prediction process of the rock porosity convolutional neural network prediction model is as follows:

[0031] Step 4.1 First, the time-frequency domain feature map of drilling information of rocks with different porosities is input into the first convolutional layer of the VG-CNN prediction model. This convolutional layer consists of 42 convolutional kernels of size 3×3.

[0032] Step 4.2 Next, the feature map after the first convolutional layer is input into the first pooling layer of the VG-CNN prediction model. This pooling layer consists of a 2×2 pooling window.

[0033] Step 4.3 Repeat steps 4.1 and 4.2 until the learning rate of the VG-CNN prediction model is ≤3×10⁻⁶. -5 Attenuation coefficient ≤ 1×10 -6 Stop model training; at this point, substitute the rock porosity prediction confusion matrix data into the following formula to verify the accuracy of the VG-CNN prediction model in predicting rock porosity.

[0034] In the formula: P T N represents the true instance. T P represents a true counterexample. F N represents a false positive example. F This indicates a false counterexample.

[0035] The beneficial effects of this invention are as follows:

[0036] 1. This method acquires limited real-time drilling parameter data during the drilling process, performs multiple noise reduction processes, and then iteratively derives the data using a "neural network + genetic algorithm". The newly formed drilling dataset has the characteristics of large data volume, high reliability, and strong robustness, effectively solving the technical problems of large external interference, small effective data volume, and strong data dispersion in the data acquisition process of the drilling experiment.

[0037] 2. The drilling parameters selected in this method are reasonable and scientific, and can truly reflect the porosity of the rock. By optimizing the structural parameters of the convolutional neural network, the inversion accuracy of rock porosity is improved. This solves the problem of the lag caused by sending the rock porosity quantitative evaluation to the laboratory for relevant tests in the past. It can be applied to the prediction and evaluation of the porosity of the surrounding rock in underground engineering projects with variable geological conditions and limited space, such as coal mine roadways, subways, and tunnels. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 is a flowchart illustrating a method for predicting rock porosity based on small sample drilling data inversion according to an embodiment of the present invention.

[0040] Figure 2 is a structural diagram of the BP-GA intelligent algorithm-derived model. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] As shown in Figure 1, a method for predicting rock porosity based on small sample drilling data inversion includes the following steps:

[0043] Step S1: Collect drilling information during the drilling process of the drilling rig in the laboratory or on-site roadway by using sensors, and obtain key drilling parameters such as torque M, thrust F, rotational speed N, thrust speed V, drill pipe amplitude A, and vibration acceleration a.

[0044] Step S2: Perform multiple noise reduction on the collected different types of drilling parameters, and input the multiple noise-reduced drilling parameter values ​​into the "BP-GA" intelligent algorithm derived model to obtain a new drilling parameter dataset;

[0045] Step S3: Analyze the variation curves of different types of drilling parameters with drilling time t in the drilling parameter dataset to obtain the time-frequency domain map of the drilling parameters. Use the data augmentation function in the neural network to obtain the time-frequency domain image library of different types of drilling parameters.

[0046] Step S4: Construct a convolutional neural network (VG-CNN) prediction model for rock porosity. Learn and train the model on a time-frequency domain feature image library of drilling parameters during drilling of rocks with different porosities. Determine the initial learning rate and decay coefficient of the model to obtain the rock porosity prediction results.

[0047] This section uses a laboratory experiment as an example to demonstrate rock porosity prediction based on drilling parameter inversion. First, rocks with different porosities, such as sandstone, mudstone, limestone, granite, and marble, were uniformly processed into 15cm×15cm×15cm cubic specimens. These were divided into 5 groups of 9 samples each, totaling 45 specimens. The drilling test employed a controlled variable method, maintaining a constant drilling speed (V) and drill bit speed (N) during the drilling process, while monitoring real-time data of other drilling parameters. The drilling speed was set to 60mm / min, the drill bit speed to 200r / min, and the drilling depth was uniformly preset to 120mm. Once the predetermined drilling depth was reached, the drilling experimental setup automatically stopped.

[0048] First, the multi-stage noise reduction method proposed in this patent is used to denoise the drilling parameters collected by sensors, including torque M, thrust F, rotational speed N, feed rate V, drill pipe amplitude A, and vibration acceleration a. Since the drilling rate V and rotational speed N are kept constant during the experiment, the focus is on denoising torque M, thrust F, drill pipe amplitude A, and vibration acceleration a. Specifically, the wavelet denoising function used for torque M is the db3 wavelet denoising function, the wavelet denoising function used for thrust F is the coif3 wavelet denoising function, and the wavelet denoising function used for drill pipe amplitude A and vibration acceleration a is the bior3.3 wavelet denoising function. Noise reduction is stopped when the original drilling data x(t) of each type is reduced to a monotonic function or has ≤3 extreme points. After noise reduction, the datasets corresponding to each type of drilling parameter are N1, N2, ..., N6.

[0049] Figure 2 shows the structure of the "BP-GA" intelligent algorithm-derived model. Based on the results of the drilling experiment, the initial weights and screening thresholds of the sample data are determined. For the same type of drilling parameter, data points with strong dispersion are identified by calculating the data range and mean. After removing unreasonable data, the six types of drilling parameters are all iterated, crossovered, and mutated using a genetic algorithm. The crossover probability is 0.75, and the coefficient of variation is 0.031. The resulting new drilling parameter data groups are M1, M2, ..., M6.

[0050] Then, relationship curves are established between the newly generated drilling parameter data and the drilling time t to obtain time-frequency domain feature maps of different types of drilling parameters. Discrete Fourier transform is used to transform the time-domain feature maps of the drilling parameters into frequency-domain feature maps. Then, short-time Fourier transform is used to fuse the time-domain and frequency-domain maps of the drilling parameters to obtain time-frequency domain feature images of different types of drilling parameters. Using the data augmentation function in the neural network, operations such as random flipping, clockwise or counterclockwise rotation, brightness changes, and pixel value enhancement are performed on the time-frequency domain feature maps of each type of drilling parameter to expand the time-frequency domain image library of drilling parameters. The iterative calculation of the neural network is stopped when there are ≥1200 time-frequency domain feature images corresponding to each type of drilling parameter.

[0051] Finally, the time-frequency domain feature maps of drilling information for rocks with different porosities are input into the first convolutional layer of the rock porosity VG-CNN prediction model. This convolutional layer consists of 42 convolutional kernels of size 3×3. The feature map after the first convolutional layer is then input into the first pooling layer of the VG-CNN prediction model. This pooling layer consists of 2×2 pooling windows. Steps 4.1 and 4.2 are repeated until the learning rate of the VG-CNN prediction model is ≤3×10⁻⁶. -5 Attenuation coefficient ≤ 1×10 -6 Stop model training; at this point, the optimal hyperparameters of the model are shown in Table 1:

[0052] Table 1 Optimal hyperparameters of the VG-CNN prediction model for rock porosity

[0053] After the VG-CNN model was trained, a new batch of coarse-grained sandstone was selected and processed into 15cm×15cm×15cm cubic specimens. Laboratory drilling tests were conducted with a drilling speed of 60mm / min, a drill bit speed of 200r / min, and a pre-set drilling depth of 120mm. The real-time drilling parameters and time-frequency domain feature maps collected in the laboratory were input into the VG-CNN prediction model. The porosity of the coarse-grained sandstone was found to be 23.15% through inversion. Simultaneously, a laboratory mercury intrusion porosimetry test was used to test the actual porosity of the coarse-grained sandstone, and the test results showed a porosity of 25.37%. It can be seen that the rock porosity obtained by inversion using this invention is quite close to the actual porosity measured in the laboratory. This invention enables accurate prediction of rock porosity based on drilling parameter inversion.

[0054] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting rock porosity based on inversion of small sample drilling data, characterized in that, Includes the following steps: Step S1: Collect drilling information during the drilling process of the drilling rig in the laboratory or underground tunnel, and obtain key drilling parameters such as torque M, thrust F, rotational speed N, thrust speed V, drill pipe amplitude A, and vibration acceleration a. Step S2: Perform multiple noise reduction processing on the collected different types of drilling parameters, and input the multiple noise-reduced drilling parameter values ​​into the "BP-GA" model to obtain a new drilling parameter derived dataset; Step S3: Analyze the variation curves of different types of drilling parameters with drilling time t in the new drilling parameter dataset to obtain the time-frequency domain map of drilling parameters. Use the data augmentation function in the neural network to obtain the time-frequency domain image library of different types of drilling parameters. Step S4: Construct a convolutional neural network prediction model for rock porosity, learn and train on a time-frequency domain feature image library of drilling parameters for rocks with different porosities, determine the initial learning rate and decay coefficient of the model, and obtain the rock porosity prediction results.

2. The rock porosity prediction method based on small sample drilling data inversion as described in claim 1, characterized in that, The implementation process of different types of drilling parameter multi-denoising methods in step S2 includes the following steps: Step 1.1: Based on the raw data of different types of drilling parameters acquired continuously, calculate the extreme point m of the raw data for each type of drilling parameter. i Average value n i and local amplitude a i ; The raw drilling data x(t) is separated using a local mean function a(t), and then demodulated using an envelope estimation function m(t). h(t)=x(t)-a(t) (3) Repeat the above operations until the original drilling data x(t) is reduced to a monotonic function or has ≤3 extreme points, then stop the iterative calculation. At this point, the original drilling data x(t) is decomposed into k decision components PF and a decomposition margin. Step 1.2: Using kurtosis H1, correlation coefficient H2, and root mean square value H3 as selection criteria for PF components, the evaluation index V of the PF components is then determined. ij Represented as: In the formula, j takes values ​​of 1, 2, 3, ..., k; V is defined here ij >90% or more of the PF components are valid components; Step 1.3: The selected PF components are further denoised using wavelet denoising. This includes wavelet denoising of the effective PF components of torque M, thrust F, rotational speed N, thrust velocity V, drill pipe amplitude A, and vibration acceleration a. The wavelet denoising function used for torque and rotational speed is the db3 wavelet denoising function, the wavelet denoising function used for thrust and thrust velocity is the coif3 wavelet denoising function, and the wavelet denoising function used for drill pipe amplitude and vibration acceleration is the bior3.3 wavelet denoising function. Step 1.4: Input the denoised drilling parameter data into the "BP-GA" model for iteration, cross-validation, and mutation to achieve data augmentation and data derivation.

3. The rock porosity prediction method based on small sample drilling data inversion as described in claim 2, characterized in that, In step S2, a "BP-GA" model is established to address the problem of insufficient time-frequency domain plot sample data due to difficulties in on-site data acquisition for the above six types of drilling parameters. The specific method is as follows: Step 2.1: Use a backpropagation neural network (BP) to determine the collected drilling information samples and expected output values. Use an adaptive algorithm to determine the initial weights and thresholds of the BP neural network corresponding to each type of drilling parameter, thereby eliminating unreasonable data. Step 2.2: For the six types of drilling parameters after removing unreasonable data, a genetic algorithm is used for iteration, crossover and data mutation. The crossover probability is 0.75 and the coefficient of mutation is 0.

031. The new drilling parameter data groups generated after encoding are M1, M2, ..., M6. Step 2.3: Establish relationship curves between the newly generated drilling parameter data and the drilling time t, and then obtain the time-frequency domain feature maps of different types of drilling parameters; Step 2.4: Utilize the data augmentation function in the neural network to randomly flip, rotate clockwise or counterclockwise, change brightness, and increase pixel value of the time-frequency domain feature map of each type of drilling parameter to expand the time-frequency domain image library of drilling parameters until there are ≥1200 time-frequency domain feature images corresponding to each type of drilling parameter, and then stop the iterative calculation of the neural network.

4. The rock porosity prediction method based on small sample drilling data inversion as described in claim 1, characterized in that, Step S3, the method for obtaining time-frequency domain feature maps of different types of drilling parameters, includes the following steps: Step 3.1: Based on the drilling information processed by multiple noise reduction and "BP-GA" model, establish the variation curves of different types of drilling parameters with drilling time t, that is, obtain the time domain feature map of the drilling parameters. Step 3.2: The time-domain characteristic map of the drilling parameters is transformed into a frequency-domain characteristic map using discrete Fourier transform. The transformation equation is as follows: In the formula, Y(n) represents a finite-length discrete frequency sequence of length L, n = 0, 1, ..., L-1, Hz, and y(t) represents a discrete time sequence of length T. Step 3.3: The time-domain and frequency-domain images of the drilling parameters are fused by short-time Fourier transform; that is, time is used as the horizontal axis and the acquisition frequency and instantaneous energy change of the drilling parameters are used as the vertical axis to obtain the time-frequency domain feature images of different types of drilling parameters. The expanded time-frequency domain image library of drilling parameters is then input into the rock porosity convolutional neural network prediction model, namely the VG-CNN prediction model.

5. The method for predicting rock porosity based on small sample drilling data inversion as described in claim 1, characterized in that, In step S4, the specific prediction process of the rock porosity convolutional neural network prediction model is as follows: Step 4.1 First, the time-frequency domain feature map of drilling information of rocks with different porosities is input into the first convolutional layer of the VG-CNN prediction model. This convolutional layer consists of 42 convolutional kernels of size 3×3. Step 4.2 Next, the feature map after the first convolutional layer is input into the first pooling layer of the VG-CNN prediction model. This pooling layer consists of a 2×2 pooling window. Step 4.3 Repeat steps 4.1 and 4.2 until the learning rate of the VG-CNN prediction model is ≤3×10⁻⁶. -5 Attenuation coefficient ≤ 1×10 -6 Stop model training; at this point, substitute the rock porosity prediction confusion matrix data into the following formula to verify the accuracy of the VG-CNN prediction model in predicting rock porosity. In the formula: P T N represents the true instance. T P represents a true counterexample. F N represents a false positive example. F This indicates a false counterexample.

Images