Rock-soil strength parameter prediction method based on while-drilling test data

Abstract

The invention discloses a rock-soil strength parameter prediction method based on while-drilling test data, particularly relates to the technical field of data mining, and is used for solving the problem of inaccurate rock-soil strength prediction caused by complex coupling among drilling parameters in an existing method. The method comprises the following steps: firstly, acquiring bit pressure, torque and rotating speed data, separating an intrinsic signal component representing stratum mechanical properties by identifying dynamic coupling characteristics among parameters, and then comparing and screening intrinsic signals and geological background data to obtain a target signal component; and finally, a machine learning regression model is established based on the target signal to realize real-time prediction of the rock-soil strength parameters, operation interference is effectively stripped through multi-stage feature extraction and verification, and the prediction accuracy is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, and more specifically, to a method for predicting geotechnical strength parameters based on drilling test data. Background Technology

[0002] In oil drilling and geological exploration engineering, testing while drilling (WSDB) technology can acquire a series of engineering parameters reflecting the interaction between downhole conditions and the formation in real time, such as drilling pressure, torque, and rotational speed. Using this data to evaluate the rock and soil strength parameters of the drilled formation in real time is crucial for optimizing drilling techniques, assessing wellbore stability, and improving exploration and development efficiency. Current technologies typically estimate rock and soil strength based on these engineering parameters using mathematical modeling or empirical formulas.

[0003] However, these engineering parameters are inherently and complexly coupled with nonlinear relationships. For example, operational adjustments to drilling pressure and rotation speed directly cause changes in response parameters such as torque. These signal changes caused by operational behavior are highly similar to and intertwined with the response changes caused by differences in the lithology of the formation itself. This makes it extremely difficult to effectively isolate the operational influence from the mixed data and accurately identify the intrinsic information that purely characterizes the formation's mechanical properties. Consequently, the prediction results of soil and rock strength parameters have significant uncertainties and biases. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for predicting geotechnical strength parameters based on drilling test data to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for predicting soil and rock strength parameters based on drilling test data includes the following steps:

[0007] S1. Acquire real-time measured data of drilling pressure, torque, and rotational speed during the drilling process;

[0008] S2. Based on drilling pressure data, torque data, and rotational speed data, the dynamic coupling characteristics between drilling parameters are identified.

[0009] S3. Based on dynamic coupling characteristics, the intrinsic signal components characterizing formation mechanical properties are separated by comparing the differences in signal characteristics between normal drilling sections and complex working condition sections.

[0010] S4. Perform a consistency comparison analysis between the intrinsic signal components and the geological background data, and select target signal components that are consistent with the geological background.

[0011] S5. Establish a machine learning regression model for soil and rock strength parameters based on the target signal components;

[0012] S6. Use machine learning regression models to predict the rock and soil strength parameters of the currently encountered strata in real time.

[0013] Furthermore, real-time measured data on drilling pressure, torque, and rotational speed during the drilling process are acquired, including:

[0014] Real-time data collection of drill pressure, torque, and rotational speed is achieved using a drilling measurement-while-drilling instrument.

[0015] The collected drilling pressure data, torque data, and rotational speed data are synchronized and aligned in time.

[0016] The drill pressure data, torque data, and rotational speed data, after time synchronization and alignment, are stored as real-time measurement data in a standardized format.

[0017] Furthermore, based on drilling pressure data, torque data, and rotational speed data, the dynamic coupling characteristics between drilling parameters are identified, including:

[0018] Calculate the correlation trend between drill pressure data and torque data within the sliding time window;

[0019] Analyze the nonlinear interaction characteristics between speed data and torque data;

[0020] By fusing correlation trends with nonlinear interaction characteristics, dynamic coupling characteristics between drilling parameters are generated.

[0021] Furthermore, the fusion of correlation trend and nonlinear interaction characteristics includes: normalizing the correlation trend and nonlinear interaction characteristics respectively; weighting the normalized correlation trend and nonlinear interaction characteristics; and performing dimensionality reduction on the weighted combination characteristics through principal component analysis to generate dynamic coupling characteristics between drilling parameters.

[0022] Furthermore, based on dynamic coupling characteristics, intrinsic signal components characterizing formation mechanical properties are separated by comparing the differences in signal characteristics between normal drilling sections and complex operating condition sections. These components include:

[0023] Based on dynamic coupling features, the first set of signal features corresponding to the normal drilling section and the second set of signal features corresponding to the complex working condition section are identified;

[0024] Calculate the difference in energy distribution between the first signal feature set and the second signal feature set in the time-frequency domain;

[0025] Based on the difference characteristics, the intrinsic signal components characterizing the formation mechanical properties are separated from the dynamic coupling characteristics.

[0026] Furthermore, the calculation of the difference features in the time-frequency domain energy distribution includes: performing wavelet transforms on the first signal feature set and the second signal feature set respectively to extract the energy distribution of each frequency band; calculating the relative entropy of the corresponding frequency band energy distribution as a difference measure; and combining the difference measures of each frequency band to form the difference features in the time-frequency domain energy distribution.

[0027] Furthermore, a consistency comparison analysis was performed between the intrinsic signal components and the geological background data to screen out target signal components consistent with the geological background, including:

[0028] Extract key stratigraphic parameter sequences from geological background data;

[0029] The similarity between the intrinsic signal components and the key stratigraphic parameter sequence is calculated to generate a similarity index corresponding to each intrinsic signal component.

[0030] The similarity index is filtered based on a preset consistency threshold, and the intrinsic signal components with similarity indices greater than the consistency threshold are retained as target signal components consistent with the geological background.

[0031] Furthermore, geological background data is obtained through the following methods: collecting stratigraphic lithology profile data from well logging interpretation results of adjacent wells; obtaining stratigraphic structural feature data obtained from the interpretation of regional seismic data; and integrating stratigraphic age and sedimentary facies data from regional geological survey reports to form geological background data for consistency comparison analysis.

[0032] Furthermore, a machine learning regression model for soil and rock strength parameters is established based on the target signal components, including:

[0033] The target signal components are constructed into feature vectors for machine learning training;

[0034] The gradient boosting regression algorithm is used to train the feature vector and the corresponding soil and rock strength parameters to generate a trained machine learning regression model.

[0035] The gradient boosting regression algorithm establishes a nonlinear mapping relationship between feature vectors and soil strength parameters by iteratively optimizing the combination of decision trees;

[0036] The rock and soil strength parameters were obtained by: taking core samples from the drilled section and conducting indoor rock mechanics tests to obtain the rock compressive strength parameters; calculating the formation shear strength parameters using empirical formulas based on sonic logging data; and establishing a sample set of rock and soil strength parameters for training machine learning regression models by combining the results of indoor tests and logging calculations.

[0037] Furthermore, machine learning regression models are used to predict the soil and rock strength parameters of the currently encountered strata in real time, including:

[0038] Acquire real-time drilling pressure data, real-time torque data, and real-time rotational speed data during the current drilling process;

[0039] The current real-time feature vector is generated based on real-time drilling pressure data, real-time torque data, and real-time rotational speed data.

[0040] Input the current real-time feature vector into the machine learning regression model, and output the predicted value of the rock and soil strength parameters of the current drilling strata.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. By constructing a multi-level extraction and verification process from raw drilling parameters to intrinsic formation signal components, the interference of drilling operation behavior on parameter signals is effectively eliminated, significantly improving the accuracy of geotechnical strength parameter prediction. First, the dynamic coupling characteristics between drilling parameters are identified, capturing the nonlinear interaction modes between drilling pressure, torque, and rotational speed as they change during the drilling process, thus more accurately characterizing the dynamic properties of downhole conditions. Then, by comparing the differences in signal characteristics between normal drilling sections and complex operating conditions, the intrinsic signal components that purely characterize formation mechanical properties are separated from the mixed signals. This process effectively filters out signal fluctuations caused by operational adjustments, preserving characteristic information reflecting the essential lithology of the formation.

[0043] 2. By comparing the intrinsic signal components with the geological background data, the consistency between the extracted signal components and known geological patterns is ensured, enhancing the geological interpretability of the features. The machine learning regression model established based on the screened target signal components can fully explore the complex patterns hidden in the drilling big data. Through algorithms such as gradient boosting, the nonlinear mapping relationship between signal features and soil strength parameters is automatically learned. By integrating big data mining and machine learning processing methods, the limitations of traditional empirical formulas are not only overcome, but also the real-time accurate prediction of soil strength parameters is achieved, providing more reliable technical support for drilling process optimization and wellbore stability assessment. Attached Figure Description

[0044] Figure 1 This is a flowchart of a method for predicting soil and rock strength parameters based on drilling test data according to the present invention. Detailed Implementation

[0045] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0046] Example: Figure 1 This invention presents a method for predicting geotechnical strength parameters based on drilling test data, which includes the following steps:

[0047] S1. Acquire real-time measured data of drilling pressure, torque, and rotational speed during the drilling process;

[0048] S2. Based on drilling pressure data, torque data, and rotational speed data, the dynamic coupling characteristics between drilling parameters are identified.

[0049] S3. Based on dynamic coupling characteristics, the intrinsic signal components characterizing formation mechanical properties are separated by comparing the differences in signal characteristics between normal drilling sections and complex working condition sections.

[0050] S4. Perform a consistency comparison analysis between the intrinsic signal components and the geological background data, and select target signal components that are consistent with the geological background.

[0051] S5. Establish a machine learning regression model for soil and rock strength parameters based on the target signal components;

[0052] S6. Use machine learning regression models to predict the rock and soil strength parameters of the currently encountered strata in real time.

[0053] S1. Obtain real-time measured data on drilling pressure, torque, and rotational speed during the drilling process. Specifically, this is implemented as follows:

[0054] Drilling measurement-while-drilling (DWW) instruments acquire drill pressure (DPS), torque, and rotational speed (ROM) data in real time. These instruments include DPS sensors, torque sensors, and ROM sensors mounted on the drill string. During drilling, these sensors continuously monitor downhole conditions and simultaneously acquire DPS, torque, and ROM data at preset sampling frequencies. The DPS sensor generates DPS data by measuring the axial force on the drill string; the torque sensor generates torque data by measuring the torsional torque of the drill string; and the ROM sensor generates ROM data by measuring the rotational speed of the drill string. During acquisition, data is transmitted in real time to the surface data processing unit via mud pulses or electromagnetic waves to ensure data continuity and real-time performance. The sampling frequency is set according to drilling process requirements, for example, acquiring data 10 times per second to ensure the temporal resolution of the data meets the needs of real-time analysis. Each data point is accompanied by a timestamp sourced from a high-precision clock, such as a GPS clock, to ensure the accuracy of the time recording.

[0055] The collected drill pressure, torque, and rotational speed data undergo time synchronization and alignment processing. This process eliminates time offsets between data from different sensors, ensuring that the drill pressure, torque, and rotational speed data are aligned on the time axis. The process begins by extracting the timestamp for each data point. Then, using the drill pressure data's timestamp as a baseline, linear interpolation is applied to adjust the torque and rotational speed data to the same time point. The linear interpolation method calculates estimates of the torque and rotational speed data at missing time points based on the timestamp sequence, ensuring all data have a unified time series. For example, if drill pressure data is recorded at time t1, while torque data is recorded at time t2, and t1 and t2 do not coincide, the estimated torque value at time t1 is calculated by weighted averaging the torque data values ​​before and after t2. Time synchronization and alignment also involves data integrity checks, such as removing data points with abnormal or missing timestamps, to ensure the reliability of subsequent processing.

[0056] After time synchronization and alignment, drill pressure, torque, and rotational speed data are stored as real-time measurement data in a standardized format. This standardized format includes defining uniform data fields, units, and storage structures. For example, drill pressure data is stored in kilonewtons, torque data in newton-meters, and rotational speed data in revolutions per minute. Data is stored in a tabular format, with each row representing a record at a specific time point. Column fields include timestamps, drill pressure data values, torque data values, and rotational speed data values. The standardized format also involves data encoding, such as using comma-separated value formats or binary formats, to improve storage efficiency and retrieval speed. The storage process involves writing the data to a real-time database or file, along with metadata information such as well number, acquisition date, and data quality identifiers. Metadata facilitates data retrieval and verification in subsequent steps, ensuring data consistency and traceability throughout the entire process.

[0057] S2. Based on drilling pressure data, torque data, and rotational speed data, the dynamic coupling characteristics between drilling parameters are identified. Specifically, this is implemented as follows:

[0058] The dynamic coupling characteristics between drilling parameters are identified based on drill pressure (DP) data, torque data, and rotational speed data. First, the correlation trend between DP and torque data within a sliding time window is calculated. The size of the sliding time window is set according to the dynamic characteristics of the drilling process; for example, a window length of 30 seconds is set to cover sufficient operating cycles. The correlation trend is obtained by calculating the Pearson correlation coefficient between DP and torque data within continuous sliding time windows. The Pearson correlation coefficient measures the degree of linear correlation between two variables, with a value between -1 and +1. During calculation, for each sliding time window, the DP and torque data sequences within that window are extracted, and then the covariance of the two sequences is calculated divided by the product of their standard deviations. The correlation trend is formed by tracking the changes of these correlation coefficients over time, for example, by plotting a trend curve with time on the horizontal axis and the correlation coefficient on the vertical axis. The sliding time window's step size is set according to the data sampling frequency, for example, moving the window once per second to ensure the continuity and real-time nature of the trend. During the calculation, if there are insufficient data points within a window, the window is skipped or interpolation methods are used to supplement the data, ensuring the reliability of the calculation.

[0059] This study analyzes the nonlinear interaction characteristics between speed and torque data. Nonlinear interaction characteristics are used to capture the nonlinear dependencies between speed and torque data, for example, by quantifying this relationship through mutual information calculation. Mutual information measures the statistical dependence between two variables, including nonlinear relationships. The calculation first discretizes the speed and torque data, dividing continuous values ​​into several intervals, such as based on quantiles of the data distribution. Then, the joint probability distribution and marginal probability distribution are calculated based on the discretized data. Finally, the mutual information value is obtained by summing the logarithmic ratio of the products of the joint and marginal probabilities. A larger mutual information value indicates a stronger nonlinear interaction. Another approach is to use a kernel-based similarity measure, such as calculating the similarity between data points using a Gaussian kernel, and then assessing the correlation between speed and torque data in the feature space. Nonlinear interaction characteristics can also be obtained by calculating the phase synchronization index of speed and torque data, for example, by using Hilbert transform to extract the instantaneous phase of the signal and then calculating the phase lock value. During the analysis, all calculations are based on speed and torque data from the same time period to ensure feature consistency.

[0060] The correlation trend and nonlinear interaction characteristics are fused to generate dynamic coupling features between drilling parameters. The fusion process first normalizes both the correlation trend and the nonlinear interaction characteristics. Normalization uses a min-max scaling method, scaling the value of each feature to the range of 0 to 1. Specifically, for the correlation trend, the minimum and maximum correlation coefficients in the entire sequence are found, and then each correlation coefficient value is subtracted from the minimum value and divided by the difference between the maximum and minimum values. For nonlinear interaction characteristics, min-max scaling is also applied, scaling based on the minimum and maximum values ​​of their feature values. Normalization ensures the comparability of different features and avoids the impact of dimensional differences on the fusion effect.

[0061] The normalized correlation trend and nonlinear interaction features are weighted and combined. During weighting, the weights are set according to the importance of each feature in characterizing the dynamic coupling of drilling parameters. Importance can be determined by the historical correlation strength between the feature and the target variable; for example, the correlation coefficient between each feature and the soil strength parameter can be calculated in the training data, and the absolute value of the correlation coefficient can be normalized as the weight. The weights are set to a total of 1; for example, the weight of the correlation trend is 0.6, and the weight of the nonlinear interaction features is 0.4. The weighted combination is achieved by multiplying the normalized feature values ​​by their corresponding weights and then summing them to generate a combined feature vector. The combined feature vector retains the main information of the original features while highlighting the role of important features.

[0062] Principal component analysis (PCA) is used to reduce the dimensionality of the weighted combined features, generating dynamic coupling features between drilling parameters. PCA is a statistical method used to transform high-dimensional data into a low-dimensional representation while retaining most of the variance. During dimensionality reduction, a weighted combined feature vector matrix is ​​first constructed, where each row represents a time point and each column represents a feature. Then, the covariance matrix of the feature vectors is calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved. The eigenvalues ​​represent the variance contribution of each principal component, and the top few principal components with larger eigenvalues ​​are selected as the features after dimensionality reduction. For example, the first two principal components are selected because their cumulative variance contribution exceeds 85%. The dimensionality-reduced principal components are the dynamic coupling features, which comprehensively reflect the interaction patterns between drilling pressure, torque, and rotational speed data, and are used for subsequent analysis. The entire fusion process ensures feature compression and redundancy removal, improving the efficiency of subsequent machine learning.

[0063] S3. Based on dynamic coupling characteristics, intrinsic signal components characterizing formation mechanical properties are separated by comparing the differences in signal characteristics between normal drilling sections and complex working condition sections. Specifically, this is implemented as follows:

[0064] Based on dynamic coupling characteristics, intrinsic signal components characterizing formation mechanical properties are separated by comparing the differences in signal characteristics between normal drilling sections and complex operating condition sections. First, a first set of signal features corresponding to the normal drilling section and a second set of signal features corresponding to the complex operating condition section are identified based on dynamic coupling characteristics. The identification of normal and complex operating condition sections is based on operating parameters and formation response data during the drilling process. For example, historical drilling data defines a normal drilling section as one with stable drilling pressure and small torque fluctuations, and a complex operating condition section as one with sudden changes in drilling pressure and abnormally high torque. During identification, time series data of dynamic coupling characteristics are used, combined with drilling logs or real-time monitoring data, to divide the intervals corresponding to normal and complex operating condition sections. For each interval, the dynamic coupling feature values ​​within that interval are extracted to form a first set of signal features and a second set of signal features. The first set of signal features includes the dynamic coupling feature values ​​at all time points in the normal drilling section, and the second set of signal features includes the dynamic coupling feature values ​​at all time points in the complex operating condition section. The extraction of feature sets ensures data integrity, such as removing missing or outliers and maintaining chronological consistency. During the identification process, the threshold for distinguishing between normal drilling sections and complex operating conditions is set based on experience. For example, a drilling pressure change rate exceeding 5,000 kN / s is considered a complex operating condition. However, the actual threshold can be adjusted according to specific drilling conditions to ensure the accuracy of the distinction.

[0065] The difference in energy distribution between the first and second signal feature sets in the time-frequency domain is calculated. This calculation begins by performing wavelet transforms on both sets to extract the energy distribution of each frequency band. The wavelet transform uses a mother wavelet function, such as the Daubechies wavelet, to perform multi-resolution analysis on each signal feature set. The wavelet transform decomposes the signal into sub-signals in different frequency bands, for example, into four bands covering the low-to-high frequency range. Each band corresponds to a specific frequency range; for example, the first band is 0 to 10 Hz, the second band is 10 to 20 Hz, and so on. When extracting the energy distribution of each band, the energy value of each sub-signal is calculated by summing the squares of the sub-signal amplitudes. The energy distribution is represented as the energy percentage or absolute energy value of each band, forming a sequence of frequency band energy distributions for the first and second signal feature sets.

[0066] The relative entropy of the corresponding frequency band energy distribution is calculated as a measure of difference. Relative entropy measures the difference between two probability distributions; here, the frequency band energy distribution is considered as a probability distribution. The calculation first normalizes the energy distribution of each frequency band in the first and second signal feature sets, making the sum of each distribution equal to 1, thus converting it to probabilistic form. Then, for each corresponding frequency band, the relative entropy of the energy distribution of the first signal feature set relative to the energy distribution of the second signal feature set is calculated. The relative entropy is calculated based on the probability value of each frequency band, obtained by multiplying the logarithm of the sum of the probability values ​​of the first and second distributions by the probability value of the first distribution. A larger relative entropy value indicates a greater difference between the two distributions in the corresponding frequency band. This difference measure quantifies the inconsistency in the time-frequency domain energy distribution between normal drilling sections and complex operating conditions.

[0067] The difference measures of each frequency band are combined to form a difference feature in the energy distribution of the time-frequency domain. During the combination, the relative entropy values ​​of each frequency band are arranged in band order to form a multi-dimensional vector. For example, if there are four frequency bands, the difference feature is a 4-dimensional vector, with each dimension corresponding to the relative entropy value of one frequency band. The combination process ensures that the difference information of all frequency bands is preserved, while the vector form facilitates subsequent processing. The difference feature reflects the overall difference in the energy distribution of the time-frequency domain between normal drilling sections and complex operating conditions, and is used to guide the separation of intrinsic signal components.

[0068] The intrinsic signal components characterizing formation mechanical properties are separated from the dynamic coupling features based on differential characteristics. The separation process identifies the portions of the dynamic coupling features relevant to formation mechanical properties based on the values ​​of the differential characteristics. For example, a differential threshold is set, based on historical data or experience, such as the average relative entropy plus a standard deviation. For each time point of the dynamic coupling feature, the corresponding differential characteristic value is checked to see if it exceeds the differential threshold. If the differential characteristic value exceeds the threshold, the dynamic coupling feature value at that time point is retained as an intrinsic signal component; otherwise, the value is discarded or ignored. The intrinsic signal component characterizes formation mechanical properties because it highlights the significantly different portions between normal drilling sections and complex operating conditions, which are typically related to changes in formation lithology or mechanical response. After separation, the intrinsic signal components form a time series for subsequent analysis. The entire separation process ensures that the intrinsic signal components contain information representing formation mechanical properties while removing noise or irrelevant components.

[0069] S4. Perform a consistency comparison analysis between the intrinsic signal components and the geological background data, and screen out the target signal components that are consistent with the geological background. The specific implementation is as follows:

[0070] The intrinsic signal components are compared with the geological background data for consistency, and target signal components consistent with the geological background are selected. Geological background data is acquired through the following methods: collecting stratigraphic lithology profile data from adjacent well logging interpretation results, including lithology type and thickness information extracted from adjacent well logging curves; acquiring stratigraphic structural feature data from regional seismic data interpretation, including stratigraphic dip and fault information identified through seismic reflection profiles; integrating stratigraphic age and sedimentary facies data from regional geological survey reports, including stratigraphic age divisions and sedimentary environment descriptions extracted from geological maps and reports; and processing the above data to form geological background data in a unified format for subsequent consistency comparison analysis. During data acquisition, it is ensured that all data sources are aligned in spatial location and time scale, for example, through coordinate transformation and temporal stratigraphic calibration, to guarantee data consistency and comparability.

[0071] Key stratigraphic parameter sequences are extracted from geological background data. These sequences include representative stratigraphic characteristics selected from the background data, such as lithological hardness index, formation density, and acoustic velocity. The extraction process begins with preprocessing the geological background data, including data cleaning and normalization, to remove noise and outliers. Then, based on the order of the stratigraphic sequence, key parameter values ​​for each stratigraphic unit are extracted to form a parameter sequence. The construction of the parameter sequence ensures that parameters correspond to drilling depth or time, for example, by arranging parameter values ​​by depth index. The extraction of key stratigraphic parameter sequences is based on geological knowledge, such as selecting parameters that significantly affect soil and rock strength, and using interpolation methods to handle missing data and ensure sequence continuity. During extraction, the units and dimensions of the parameters are standardized; for example, lithological hardness index is dimensionless, formation density is in grams per cubic centimeter, and acoustic velocity is in meters per second.

[0072] The intrinsic signal components and key stratigraphic parameter sequences are similar to each other, generating a similarity index for each intrinsic signal component. The cosine similarity method is used to measure the directional consistency between the intrinsic signal components and the key stratigraphic parameter sequences in vector space. During calculation, the intrinsic signal components and key stratigraphic parameter sequences are first converted into vectors of the same dimension, for example, through time alignment or depth matching. Since the intrinsic signal components are time-series data and the key stratigraphic parameter sequences are depth-series data, conversion is necessary, such as mapping the intrinsic signal components to the depth domain using a depth-time relationship. Then, the dot product of the two vectors is divided by their modulus product to obtain the cosine similarity value, which is between -1 and +1; a higher value indicates higher similarity. After the similarity index is generated, a similarity value is assigned to each intrinsic signal component for subsequent filtering. During the calculation, if the vector dimensions are inconsistent, zero-padding or truncation methods are used to adjust them to ensure computational feasibility.

[0073] Similarity indices are filtered based on a preset consistency threshold, retaining intrinsic signal components with similarity indices greater than the threshold as target signal components consistent with the geological background. The consistency threshold is set based on historical data or experience, for example, by analyzing samples matching known geological backgrounds and calculating the average of the similarity indices plus a standard deviation as the threshold. The threshold setting process considers the variability of the geological background; for example, a higher threshold is set for homogeneous strata and a lower threshold for heterogeneous strata. During filtering, the similarity indices of all intrinsic signal components are iterated, and the index values ​​are compared with the consistency threshold. If the similarity index is greater than the consistency threshold, the intrinsic signal component is retained; otherwise, it is discarded. The retained intrinsic signal components form the target signal components, which have a high degree of consistency with the geological background data and represent reliable information on the mechanical properties of the strata. After filtering, the target signal components are used for subsequent machine learning model training to ensure the geological rationality of the model input data. The entire filtering process, controlled by thresholds, improves the quality of the signal components and reduces noise interference.

[0074] S5. Establish a machine learning regression model for soil and rock strength parameters based on the target signal components, specifically implemented as follows:

[0075] A machine learning regression model for soil and rock strength parameters is established based on the target signal components. Soil and rock strength parameters are obtained through the following methods: Core samples are taken from drilled sections and laboratory rock mechanics tests are conducted to obtain rock compressive strength parameters. Specifically, this involves measuring the failure strength of the rock under axial pressure using a uniaxial compressive strength tester on core samples in the laboratory. The testing process follows standard rock mechanics experimental procedures, such as controlling the loading rate within the range of 0.5 MPa to 1 MPa per second to ensure data accuracy. Formation shear strength parameters are calculated using empirical formulas based on sonic logging data. Specifically, this involves extracting P-wave velocity and S-wave velocity from sonic logging curves and then applying the Mohr-Coulomb criterion or similar empirical relationships to calculate shear strength. For example, dynamic Poisson's ratio and elastic modulus are calculated by combining P-wave velocity with density data, and then converted into static shear strength parameters. A soil and rock strength parameter sample set for training machine learning regression models is established by combining laboratory test results with logging calculation results. The integration process includes data alignment and validation, such as matching laboratory test data and logging data at the same depth points, and handling inconsistent data using statistical methods such as averaging or weighted averaging to ensure the representativeness and reliability of the sample set.

[0076] The target signal components are constructed into feature vectors for machine learning training. The construction of these feature vectors is based on the time-series data of the target signal components, extracting statistical and temporal features such as mean, variance, peak factor, and energy value. The construction process first segments the target signal components, with each segment corresponding to a fixed time window, for example, each window is 10 seconds long, to cover sufficient signal variation. Then, the aforementioned feature values ​​are calculated from each window to form a multi-dimensional feature vector; for example, if four features are extracted, the feature vector will be four-dimensional. The feature vector maintains a consistent dimension, and all feature values ​​are normalized, for example, using min-max scaling to scale the values ​​to the range of 0 to 1, to avoid dimensional differences affecting model training. After construction, the feature vectors are paired with corresponding soil and rock strength parameter sample sets to form a training dataset for subsequent machine learning model training.

[0077] A gradient boosting regression algorithm is used to train the feature vectors and their corresponding soil and rock strength parameters, generating a trained machine learning regression model. Gradient boosting regression is an ensemble learning method that iteratively adds decision trees to gradually reduce prediction errors. The training process first initializes the model, for example, using the mean of the soil and rock strength parameters as the initial prediction value. Then, multiple decision trees are trained iteratively, with each tree fitting the residual from the previous iteration. The residual is the difference between the actual soil and rock strength parameter value and the current model prediction value. In each iteration, the decision tree minimizes the loss function by splitting features in the feature vector, for example, using mean squared error as the loss function. The split point is selected based on information gain or Gini impurity. The number of iterations is determined by cross-validation, for example, setting a maximum of 100 iterations, and an early stopping mechanism to prevent overfitting. After training, the model generates a combined sequence of decision trees to predict the soil and rock strength parameters corresponding to new feature vectors.

[0078] The gradient boosting regression algorithm establishes a nonlinear mapping relationship between feature vectors and soil strength parameters by iteratively optimizing a combination of decision trees. The iterative optimization process involves calculating the gradient direction (the derivative of the loss function with respect to the predicted value) in each iteration, and then updating the model based on the gradient direction. The decision trees are integrated using a weighted summation method, with the weight of each tree set according to its ability to fit the residuals; for example, the weight is proportional to the tree's contribution to reducing the residuals. The nonlinear mapping relationship is reflected in the decision trees' ability to capture the complex interactions between feature vectors and soil strength parameters, such as handling nonlinear patterns through multi-level splitting. During optimization, hyperparameters such as learning rate, tree depth, and subsampling ratio are adjusted through grid search or random search; for example, the learning rate is set to 0.1 and the tree depth to 5 to balance model complexity and generalization ability. The resulting machine learning regression model can accurately predict soil strength parameters and provides a foundation for subsequent real-time predictions.

[0079] S6. Utilize machine learning regression models to predict the soil and rock strength parameters of the currently encountered drilling strata in real time. The specific implementation is as follows:

[0080] Machine learning regression models are used to predict the geotechnical strength parameters of the currently encountered drilling formation in real time. First, real-time drill pressure (DPP) data, real-time torque data, and real-time rotational speed data are acquired during the drilling process. DPP data is collected by a drill pressure sensor in the measurement-while-drilling (MWD) instrument. This sensor operates based on the strain gauge principle, recording the axial force value of the drill string at a fixed sampling frequency, e.g., 10 times per second. The data is transmitted to the surface system in real time via mud pulse transmission. Real-time torque data is collected by a torque sensor, which measures the torsional torque of the drill string using the magnetoelastic effect. The sampling frequency is consistent with the DPP data, e.g., 10 times per second, to ensure data synchronization. Real-time rotational speed data is collected by an encoder installed on the top drive. The encoder measures the rotational speed based on the photoelectric conversion principle, also with a sampling frequency of 10 times per second. All real-time data are timestamped during transmission using a high-precision clock source, such as a GPS clock, to ensure data consistency on the timeline. After data acquisition, a preliminary verification is performed immediately. For example, obvious outliers are removed through range checks. The thresholds for range checks are set based on historical data. For example, the normal range for drilling pressure data is set to 0 to 500 kNm, the normal range for torque data is set to 0 to 50 kNm, and the normal range for rotational speed data is set to 0 to 200 rpm. Any data point exceeding these thresholds is considered abnormal and removed to ensure data quality.

[0081] A real-time feature vector is generated based on real-time drill pressure, torque, and rotational speed data. The generation process first applies a sliding window to these data, with the window length consistent with the training phase (e.g., 30 seconds) to cover sufficient operation cycles. Within each window, features identical to those used during the machine learning regression model training are extracted, including the mean of the real-time drill pressure data, the variance of the real-time torque data, and the energy value of the real-time rotational speed data. Feature extraction is based on statistical calculations; for example, the mean is obtained by summing all values ​​within the window and dividing by the number of data points, the variance is obtained by calculating the sum of squares of the deviations of each data point from the mean and then dividing by the number of data points, and the energy value is obtained by calculating the sum of squares of the amplitudes of the data points. After feature extraction, feature normalization is performed using the minimum and maximum values ​​calculated during the training phase. A minimum-maximum scaling method is applied to scale each feature value to the range of 0 to 1. For example, if the minimum mean of the drill pressure data during the training phase is 10 kN and the maximum is 300 kN, the mean of the real-time drill pressure data is normalized by subtracting 10 and dividing by 290. The normalized features are combined in a fixed order to form the current real-time feature vector. The vector dimension matches the model input; for example, if three features are extracted, the vector will be three-dimensional. During the generation process, the feature vector construction is ensured to be consistent with the training data format to avoid model prediction bias.

[0082] The current real-time feature vector is input into a machine learning regression model, which outputs predicted values ​​of the soil and rock strength parameters of the currently encountered strata. Before input, the completeness and range of the current real-time feature vector are verified; for example, it is checked whether the feature values ​​are between 0 and 1. If any feature value exceeds this range, it is truncated, with truncation thresholds set to 0 and 1. After successful verification, the current real-time feature vector is fed into the trained machine learning regression model, which is built based on the gradient boosting regression algorithm and includes an ensemble of multiple decision trees. Internally, the model performs forward computation. Each decision tree splits nodes based on the feature values ​​of the current real-time feature vector. The node splitting threshold is determined during training by optimizing the loss function, for example, using mean squared error as the loss function. The splitting point is selected based on maximizing information gain. The outputs of each decision tree are weighted and summed to obtain the final predicted value. The weights are set during training based on each decision tree's contribution to residual reduction; for example, trees with larger contributions have higher weights. The weight values ​​are optimized through cross-validation. The model output is the predicted value of the soil and rock strength parameters, such as rock compressive strength in megapascals (MPa) and shear strength in megapascals (MPa). During the output process, the model also calculates the prediction confidence level, which is based on the variance of the decision tree output. For example, the larger the variance, the lower the confidence level. A confidence threshold, such as 0.8, is set, and an alert is triggered when the confidence level falls below this threshold. The prediction results are displayed in real time on the drilling monitoring interface and used to guide the adjustment of drilling parameters. The entire prediction process, from data input to output, has a delay of less than 2 seconds, meeting real-time requirements.

[0083] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0084] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0085] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0086] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0087] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0088] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0089] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting soil and rock strength parameters based on drilling test data, characterized in that, Includes the following steps: S1. Acquire real-time measured data of drilling pressure, torque, and rotational speed during the drilling process; S2. Based on drilling pressure data, torque data, and rotational speed data, the dynamic coupling characteristics between drilling parameters are identified. S3. Based on dynamic coupling characteristics, the intrinsic signal components characterizing formation mechanical properties are separated by comparing the differences in signal characteristics between normal drilling sections and complex working condition sections. S4. Perform a consistency comparison analysis between the intrinsic signal components and the geological background data, and select target signal components that are consistent with the geological background. S5. Establish a machine learning regression model for soil and rock strength parameters based on the target signal components; S6. Use machine learning regression models to predict the rock and soil strength parameters of the currently encountered strata in real time.

2. The method for predicting soil and rock strength parameters based on drilling test data according to claim 1, characterized in that, Acquire real-time measured data on drilling pressure, torque, and rotational speed during the drilling process, including: Real-time data collection of drill pressure, torque, and rotational speed is achieved using a drilling measurement-while-drilling instrument. The collected drilling pressure data, torque data, and rotational speed data are synchronized and aligned in time. The drill pressure data, torque data, and rotational speed data, after time synchronization and alignment, are stored as real-time measurement data in a standardized format.

3. The method for predicting soil and rock strength parameters based on drilling test data according to claim 1, characterized in that, Based on drilling pressure data, torque data, and rotational speed data, the dynamic coupling characteristics between drilling parameters are identified, including: Calculate the correlation trend between drill pressure data and torque data within the sliding time window; Analyze the nonlinear interaction characteristics between speed data and torque data; By fusing correlation trends with nonlinear interaction characteristics, dynamic coupling characteristics between drilling parameters are generated.

4. The method for predicting soil and rock strength parameters based on drilling test data according to claim 3, characterized in that, The fusion of correlation trend and nonlinear interaction characteristics includes: normalizing the correlation trend and nonlinear interaction characteristics respectively; weighting the normalized correlation trend and nonlinear interaction characteristics; and performing dimensionality reduction on the weighted combination characteristics through principal component analysis to generate dynamic coupling characteristics between drilling parameters.

5. The method for predicting soil and rock strength parameters based on drilling test data according to claim 1, characterized in that, Based on dynamic coupling characteristics, intrinsic signal components characterizing formation mechanical properties are separated by comparing the differences in signal characteristics between normal drilling sections and complex operating conditions. These components include: Based on dynamic coupling features, the first set of signal features corresponding to the normal drilling section and the second set of signal features corresponding to the complex working condition section are identified; Calculate the difference in energy distribution between the first signal feature set and the second signal feature set in the time-frequency domain; Based on the difference characteristics, the intrinsic signal components characterizing the formation mechanical properties are separated from the dynamic coupling characteristics.

6. The method for predicting soil and rock strength parameters based on drilling test data according to claim 5, characterized in that, The calculation of the difference features in the time-frequency domain energy distribution includes: performing wavelet transform on the first signal feature set and the second signal feature set respectively to extract the energy distribution of each frequency band; calculating the relative entropy of the energy distribution of the corresponding frequency band as a difference measure; and combining the difference measures of each frequency band to form the difference features in the time-frequency domain energy distribution.

7. The method for predicting soil and rock strength parameters based on drilling test data according to claim 1, characterized in that, Consistency analysis was performed between the intrinsic signal components and geological background data to identify target signal components consistent with the geological background, including: Extract key stratigraphic parameter sequences from geological background data; The similarity between the intrinsic signal components and the key stratigraphic parameter sequence is calculated to generate a similarity index corresponding to each intrinsic signal component. The similarity index is filtered based on a preset consistency threshold, and the intrinsic signal components with similarity indices greater than the consistency threshold are retained as target signal components consistent with the geological background.

8. The method for predicting soil and rock strength parameters based on drilling test data according to claim 7, characterized in that, Geological background data is obtained through the following methods: collecting stratigraphic lithology profile data from well logging interpretation results of adjacent wells; obtaining stratigraphic structural feature data obtained from the interpretation of regional seismic data; and integrating stratigraphic age and sedimentary facies data from regional geological survey reports to form geological background data for consistency comparison analysis.

9. The method for predicting soil and rock strength parameters based on drilling test data according to claim 1, characterized in that, Machine learning regression models for soil and rock strength parameters are established based on target signal components, including: The target signal components are constructed into feature vectors for machine learning training; The gradient boosting regression algorithm is used to train the feature vector and the corresponding soil and rock strength parameters to generate a trained machine learning regression model. The gradient boosting regression algorithm establishes a nonlinear mapping relationship between feature vectors and soil strength parameters by iteratively optimizing the combination of decision trees; The rock and soil strength parameters were obtained by: taking core samples from the drilled section and conducting indoor rock mechanics tests to obtain the rock compressive strength parameters; calculating the formation shear strength parameters using empirical formulas based on sonic logging data; and establishing a sample set of rock and soil strength parameters for training machine learning regression models by combining the results of indoor tests and logging calculations.

10. The method for predicting soil and rock strength parameters based on drilling test data according to claim 1, characterized in that, The machine learning regression model is used to predict the soil and rock strength parameters of the currently encountered strata in real time, including: Acquire real-time drilling pressure data, real-time torque data, and real-time rotational speed data during the current drilling process; The current real-time feature vector is generated based on real-time drilling pressure data, real-time torque data, and real-time rotational speed data. Input the current real-time feature vector into the machine learning regression model, and output the predicted value of the rock and soil strength parameters of the current drilling strata.

Images