While-drilling sparse representation classification method for tool face state of horizontal well drill string torsional pendulum system

Abstract

The invention discloses a while-drilling sparse representation classification method for tool face states of a horizontal well drill string torsional pendulum system, and relates to the technical field of intelligent prediction. Comprising the following steps: establishing a basis function library of sparse representation; tool face angle time sequence signals under different working conditions are obtained; based on the basis function library, sparse decomposition is carried out in combination with the tool face angle time sequence signal to obtain a preliminary sparse coefficient; selecting a primary function with variance greater than a threshold as a sensitive primary function, and forming a discrimination subspace of the current working condition by all the sensitive primary functions of the current working condition; taking a mean value of sparse coefficients of the tool face angle time sequence signals of the current working condition in the discrimination subspace as a working condition gravity center; repeating the operation to obtain working condition gravity centers of multiple working conditions and form a typical dictionary; the typical dictionary can be used for working condition classification. According to the method, real-time identification and state prompting of the sliding guide drilling working condition can be achieved, and reliable data support is provided for well track control, sliding efficiency optimization and abnormal vibration early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent classification, and particularly relates to a horizontal well drilling string whirling system tool face state while drilling sparse characterization classification method. BACKGROUND

[0002] In the development of unconventional energy such as shale gas and tight gas, long horizontal section wells are widely used, and the torsional vibration and whirling of the drilling string in the horizontal section are significant. The tool face angle is a key control variable in sliding guide drilling, and its dynamic change process reflects the coupling behavior between the torsional response of the drilling string, the friction state and the guide operation. However, under the action of whirling and vibration, the actual tool face at the downhole will deviate from the set tool face on the ground. Therefore, the dynamic behavior of the tool face needs to be described to provide a basis for working condition identification, vibration suppression and trajectory control.

[0003] In the process of drilling while steering, the tool face angle as a key parameter to control the direction of the well trajectory, its measurement accuracy, filtering processing and control strategy have always been the focus of related research. However, existing researches focus more on the numerical accuracy and dynamic control performance of the tool face angle, and pay less attention to the structural working condition information contained in the tool face angle time domain waveform itself.

[0004] In engineering practice, various near-bit vibration and tool face monitoring systems have been developed to collect torsional vibration, tool face angle and related dynamic data in real time, and are used for ground alarm and parameter optimization. At the same time, a series of analysis methods based on deep learning have been proposed in the field of drilling state identification at home and abroad, such as convolutional neural network (CNN), long short-term memory network (LSTM), which can realize certain fault identification and working condition classification. However, such methods generally rely on a large amount of labeled data, have complex model structure and weak interpretability, so their applicability is still limited in the environment of limited downhole transmission and limited small sample experimental data.

[0005] Although some progress has been made in the statistical feature extraction of tool face signals, frequency domain analysis and automatic classification based on deep learning, the modeling and explanation mechanism for the structural features of tool face signal waveforms are still obviously insufficient. Existing methods mostly rely on macroscopic indicators such as energy and frequency spectrum, or use black box models that are difficult to explain to distinguish categories, and cannot establish a clear correspondence between tool face waveforms and real downhole working conditions. In addition, the current technology has not formed a reusable and extensible "typical waveform library" system, which cannot summarize the inherent waveform patterns of different working conditions from the structural level, and it is also difficult to maintain stable performance under strong noise, small sample and even real-time identification conditions.

[0006] In real-world drilling environments characterized by complex well conditions, fluctuating operating conditions, and high noise levels, relying solely on traditional statistical or black-box methods is insufficient to meet the combined demands for real-time performance, interpretability, and cross-scenario adaptability. Therefore, establishing a structured, typical waveform modeling and classification method for toolface signals, mapping signal structural features to operating condition semantics, can significantly improve the efficiency and accuracy of operating condition identification while effectively reducing reliance on large amounts of labeled samples. By constructing a "typical waveform library" and covering a wider range of operating conditions with a small amount of sample data, a potentially valuable and efficient solution can be provided for downhole fault identification, operating condition classification, and traditional toolface monitoring systems. Summary of the Invention

[0007] To address at least one of the aforementioned problems, this invention provides a method for classifying the sparsity of the tool face state during drilling in a horizontal well drill string torsion system.

[0008] The technical solution of this invention to solve the above problems is as follows: A method for classifying the sparsity of the tool face state while drilling in a horizontal well drill string torsional yaw system, comprising the following steps: S1. Establish a basis function library for sparse representation, wherein the basis function library contains basis functions for handling trend terms, periodic oscillation terms, decay terms, abrupt change terms, and non-smooth edge structure terms; S2. Collect multiple tool face angle test data under different working conditions and preprocess them to obtain multiple tool face angle time series signals under different working conditions; S3. Based on the basis function library and combined with the characteristics of the tool face angle time series signal, the corresponding basis function is used to perform sparse decomposition on the tool face angle time series signal to obtain the preliminary sparse coefficients. S4. Based on the changes of the basis functions under the current working conditions, the basis functions with variance greater than the threshold are selected as sensitive basis functions. All sensitive basis functions under the current working conditions constitute the discrimination subspace of the current working conditions. S5. Take the average of the sparse coefficients of the multiple tool face angle time series signals of the current working condition in the discrimination subspace, and use this average as the working condition centroid of the current working condition. S6. Repeat S3-S5 to obtain the centroids of multiple working conditions and construct a typical dictionary; S7. Take the tool surface signal to be identified, and obtain the target sparse coefficient of the tool surface signal according to the operations of S1~S3. Calculate the Euclidean distance between the target sparse coefficient in the discrimination subspace and the centroid of the working condition in the typical dictionary. Based on the minimum distance principle, obtain the classification of the tool surface signal to be identified.

[0009] The beneficial effects of this invention are as follows: By employing the tool face signal state classification method proposed in this invention, the accuracy and stability of downhole working condition identification can be significantly improved. The constructed typical waveform dictionary exhibits good structural discrimination ability under different working conditions, enabling the Euclidean distance discrimination of sparse coefficients to truly reflect the inherent differences between various working conditions. In field environments with small samples, strong noise interference, and unstable signal acquisition conditions, this invention can still maintain reliable identification performance, overcoming the limitations of traditional methods that are susceptible to noise and local disturbances leading to misjudgment.

[0010] The discriminant subspace verification mechanism introduced in this invention further improves the reliability of the final identification result. This mechanism effectively suppresses random deviations caused by factors such as noise, sampling differences, and transient disturbances by projecting and verifying the sparse coefficients, making the working condition identification process robust and reliable.

[0011] In summary, this invention enables real-time identification and status alerts for sliding steerable drilling conditions in practical applications, providing reliable data support for wellbore trajectory control, slip efficiency optimization, and abnormal vibration early warning. It has significant engineering application value and promising prospects for widespread application. Attached Figure Description

[0012] Figure 1 This is a flowchart of the method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the tool face angle testing equipment. Figure 3 To stabilize the tool surface, the first set of original signals and the resampled signals are used; Figure 4 To stabilize the second set of original signals and resampled signals on the tool surface; Figure 5 The third set of original signals and resampled signals for stabilizing the tool surface; Figure 6 Sensitive basis function selection plot for stable tool surfaces; Figure 7 These are the three sets of resampled signals from the impact tool surface; Figure 8 A comparison of typical and original waveforms for stable tool face conditions and impact tool face conditions. Detailed Implementation

[0013] The specific embodiments of the present invention will be clearly and completely described below with reference to examples. Obviously, the described examples are only some embodiments of the present invention, and not all embodiments.

[0014] like Figure 1 As shown, the method for classifying the sparse characterization of the tool face state in a horizontal well drill string torsional yaw system includes the following steps: S1, a base function library of sparse representation is established, and the base function library includes base functions for processing a trend item, a periodic oscillation item, a decay item, a sudden change item, and a non-smooth edge structure item; Through a large number of practical experiences of the inventor, it is found that the existing tool face angle time sequence signal contains trend change, steady rotation, decay vibration, local mutation and non-smooth edge and other structural characteristics, and if only a single type of base function is used, it is difficult to effectively represent the tool face angle time sequence signal. Therefore, based on the tool face dynamics mechanism, the inventor constructs a base function library containing multiple types of base functions for multi-structural and interpretable expression of the tool face angle time sequence signal.

[0015] For the above five structural characteristics, the inventor prepares multiple base functions meeting the requirements in the base function library.

[0016] For the trend change item in the tool face angle time sequence signal, a polynomial base function can be used for processing. The polynomial base function is mainly used for describing the overall trend and slowly varying smooth component of the signal, including constant term, first-order and high-order polynomial, and can simulate the slowly varying trend caused by uneven stress of the drill string, friction accumulation and system bias.

[0017] For the steady rotation item in the tool face angle time sequence signal, a Fourier sine / cosine base function can be used for processing. The Fourier sine / cosine base function is used to represent the steady rotation, torsional vibration and harmonic components, and is the main source of periodic oscillation structure in the tool face signal, which can reflect the dynamics behavior of downhole rotation speed change and periodic torsional vibration.

[0018] For the decay vibration item in the tool face angle time sequence signal, an exponential decay base function can be used for processing. The exponential decay base function is used to capture the energy decay effect of the drill string after being impacted, rubbed or collided, and the decay rate is obtained by discretization of multiple parameters, which can simulate different physical processes from rapid decay to slow decay.

[0019] For the local mutation item in the tool face angle time sequence signal, a ReLU type segmented activation base function can be used for processing. The ReLU type segmented activation base function is used to represent local mutation, jump and breakpoint structure, and can reflect the transient jump characteristics of the tool face angle when encountering sticking, instability or short-term disturbance.

[0020] For the non-smooth edge in the tool face angle time sequence signal, one of the sawtooth wave base function, square wave base function and triangular wave base function can be used for processing. The three base functions are used to describe asymmetric mutation, non-smooth edge and non-sine periodic structure, and can supplement the nonlinear periodic behavior that the traditional Fourier base function is difficult to describe.

[0021] Unlike traditional sparse representation methods that focus on using mathematical waveform atoms, the enhanced basis function library constructed in this invention is derived from the real physical components of the tool face signal. It is built upon five structural categories—trend, periodicity, decay, abrupt change, and non-smooth edge—as its basic construction principle, ensuring a one-to-one correspondence between basis functions and downhole dynamic events. This physical mechanism-driven dictionary design significantly improves the interpretability of the representation, allowing the sparse coefficients obtained after sparse representation to not only reflect the mathematical decomposition of the signal but also directly characterize the strength of structural components under different operating conditions. This facilitates the establishment of a stable and highly generalizable typical waveform library and operating condition feature space. The construction of the basis function library enables accurate, stable, and physically meaningful sparse representations of tool face signals with a finite number of structural basis functions, providing a solid foundation for subsequent operating condition classification, waveform reconstruction, and downhole condition identification.

[0022] S2. Collect multiple tool face angle test data under different working conditions and preprocess them to obtain multiple tool face angle time series signals under different working conditions; During the testing of the tool face angle, an attitude sensor is typically installed on the drill string. This attitude sensor can acquire the acceleration of the drill string along the Y and Z axes. Subsequently, the tool face angle is calculated using the following formula: In the formula, The tool face angle is represented by Accy, the y-axis acceleration is represented by Accz, and the z-axis acceleration is represented by Accz.

[0023] However, during the data collection process for the tool face angle, anomalies were observed: in horizontal wells, the drill string is affected by factors such as friction, torsional vibration, and periodic oscillation, causing the actual tool face angle to change continuously. However, when the angle crosses the sensor's measurement range boundary, it reverses, resulting in an unrealistic jump. Therefore, considering the actual situation, corrections are usually necessary.

[0024] In this step, to recover the true angle change trajectory, an unwinding algorithm (e.g., the unwrap function in signal processing) is used to unwind the original angle sequence. By detecting the angle difference between adjacent sampling points and setting a threshold, the turning point is automatically identified. An integer period is added to or subtracted from the position where the turning point occurs, thus obtaining a continuous, non-turning true angle sequence. The unwound signal accurately reflects the drill string dynamics and exhibits good retention of structures such as oscillation, stability holding, and torsional impact, providing reliable input for subsequent typical waveform construction.

[0025] Meanwhile, under different experimental conditions and different data collection conditions, the sequence length of the obtained original data is inconsistent, and cannot be directly sparse decomposed and feature extracted in a uniformly defined domain. To solve this problem, in this step, the angle sequence is resampled by using linear interpolation, and all signals are uniformly mapped to a fixed length of standard sampling points. This processing step ensures the consistency of the time domain structure, only adjusts the sampling density, and does not change the inherent structure of the signal, so that the difference between different experiments is mainly derived from the working condition itself, rather than the sampling difference.

[0026] After unwinding and resampling, the tool face angle sequence suitable for subsequent sparse decomposition is finally obtained, and the tool face angle time sequence signal is constructed in combination with the time sequence.

[0027] In order to facilitate the establishment of the working condition gravity center, after a large number of experiments by the inventor, it is found that in each working condition, the sample collection amount of the tool face angle time sequence signal is at least 3, which can meet the actual use requirement under such condition, indicating that the method of the embodiment of the application can operate under the condition of a small amount of samples. Although more samples can make the finally obtained working condition gravity center more representative, it needs higher cost.

[0028] The device structure for collecting tool face angle test data under different working condition conditions is as shown in Figure 2 The device is composed of 12 parts: 1 central control console; 2 electrical control cabinet; 3 motor support; 4 servo motor; 5 coupling; 6 simulated drill string; 7 simulated drilling tool; 8 attitude and displacement sensor; 9 universal joint; 10 magnetic powder brake; 11 simulated wellbore; 12 experimental rack.

[0029] The servo motor 4 is installed on the motor support 3 with counterweight as the main power source to reduce vibration during operation. The servo motor 4 is connected with the simulated drill string 6 through the coupling 5 to realize stable power transmission, and drives the simulated drilling tool 7 to rotate through the simulated drill string 6. The electrical control cabinet 2 can adjust the parameters such as rotation speed and torque of the servo motor 4, so as to simulate different operating conditions.

[0030] The simulated wellbore 11 is supported and fixed by a steel support, and its inner diameter, material and other parameters are matched with the actual drilling wellbore to provide a realistic downhole environment. The simulated drill string 6 is placed inside the simulated wellbore 11 and applies static friction and dynamic friction through line contact. The magnetic powder brake 10 connected to the end of the simulated drill string 6 can accurately control the resistance torque to reproduce complex conditions such as drill string-well wall friction. The universal joint 9 connects the simulated drill string 6 and the magnetic powder brake 10 to compensate for the angular deviation during movement, ensuring continuous and stable power transmission.

[0031] The posture and displacement sensor 8 is installed on the analog drilling tool 7, can measure dynamic parameters such as acceleration, rotating speed and angular displacement in real time, and transmits data to the central control console 1 through a wireless mode, so that real-time monitoring and subsequent data analysis of the experiment process are realized.

[0032] S3, based on the base function library, combining the characteristics of the tool face angle time sequence signal, using the corresponding base function to perform sparse decomposition on the tool face angle time sequence signal to obtain preliminary sparse coefficients; In this step, a sparse solving strategy of least squares is adopted, the participation of redundant base functions is suppressed while minimizing the signal reconstruction error, so that the solved coefficients present obvious sparse characteristics.

[0033] Among them, for multiple different components in the tool face angle time sequence signal, the corresponding base function is selected for sparse decomposition, so that the trend component, periodic oscillation component, decay component and local mutation component of the tool face angle time sequence signal are distinguished in the coefficient space, realizing the structured expression of the sparse coefficient and obtaining the parameterized expression result with physical correspondence.

[0034] In specific implementation, based on the sparse solving of reconstruction error, the sparse coefficient satisfies the following optimization strategy: , wherein, represents the tool face time domain signal after preprocessing, which is a column vector with a length of , and its physical meaning is the original dynamic behavior of the tool face angle changing with time in the drilling process; is an enhanced base function dictionary matrix, which is composed of group structure base functions by columns, and the dimension is ; is the sparse coefficient vector to be solved, and the dimension is , which is used to reflect which structural components the signal is composed of and the weight size of each component; is a sparse regular weight coefficient, which is used to balance the "reconstruction accuracy" and "sparsity degree".

[0035] S4, based on the change of the base function under the current working condition, the base function with a variance greater than a threshold value is selected as a sensitive base function, and all sensitive base functions of the current working condition constitute a discriminant subspace of the current working condition; In this step, first, the preliminary sparse coefficients obtained by a plurality of base functions are calculated by variance in rows, and the calculation results are counted: , wherein, represents a variance statistical set, represents a set of the k-th row element vectors of matrix A; matrix A represents a matrix composed of all the preliminary sparse coefficients under the current working condition; subsequently, the basis functions with variances greater than a threshold value are selected as sensitive basis functions, and these sensitive basis functions are taken as a discriminant subspace of the current working condition: , wherein S represents a discriminant subspace, represents a threshold value, for which, according to the experience of the inventors, it is generally set to 0.04-0.06 times, such as 0.05 times, the maximum variance, and of course, a suitable threshold value can be set according to the actual situation by those skilled in the art.

[0036] At the same time, considering that there may be noise or local disturbance in the actual signal, in order to avoid excessive contraction of the screening result, a minimum retention number is also set in this step, which is generally set to 10-20% of the total number of basis functions, such as 5, 6, 7 or 8, and when the number of basis functions meeting the variance condition is too small, a small number of basis functions with high variance ranking will be automatically supplemented to maintain the expression ability of the discriminant subspace.

[0037] The sensitive basis function subset obtained through the above screening has the characteristics of "large inter-class difference and small intra-class change", and the corresponding sparse coefficient can form more obvious working condition distribution difference in the feature space after dimension reduction, thereby significantly improving the accuracy and interpretability of working condition recognition.

[0038] Such a strategy can balance between computational efficiency and sparsity. For typical working condition recognition or scenes with relatively simple signal structure, the least square solution method can obtain stable and obviously structured sparse coefficients. The solving framework can automatically enhance periodic characteristics, suppress noise disturbance and strengthen mutation edges, so that the tool face signal forms a significant clustering with working condition difference in the coefficient space.

[0039] S5, taking the mean value of the sparse coefficients of the plurality of tool face angle time sequence signals of the current working condition in the discriminant subspace as the working condition gravity center of the current working condition; After obtaining the discriminant subspace of the current working condition, the plurality of tool face angle time sequence signals of the current working condition are re-substituted into the discriminant subspace for sparse decomposition, and finally a plurality of sparse coefficient sequences are obtained; In particular, for tool face angle time sequence signals, as described above, they are generally divided into five different items, i.e., trend item, periodic oscillation item, decay item, mutation item and non-smooth edge structure item, and different basis functions need to be used for decomposition. However, the inventors have found that in the discriminant subspace, sensitive basis functions capable of processing the five items of trend item, periodic oscillation item, decay item, mutation item and non-smooth edge structure item may not be included at the same time.

[0040] Therefore, considering the actual situation, the inventors make the following treatment: if the basis function corresponding to an item of a certain class is not contained in the discriminant subspace, the item does not participate in the coefficient update in the discriminant process; if the basis function corresponding to the item is contained in the discriminant subspace, only the sparse coefficient is projected or reconstructed on the basis function subset, so that the sparse coefficient component consistent in the discriminant subspace is obtained; the sparse coefficient component obtained in the discriminant subspace is combined with the coefficient component not participating in the discrimination, to form a sparse coefficient vector consistent with the structure of the discriminant subspace.

[0041] Subsequently, the mean of the plurality of sparse coefficient sequences is calculated, that is, the operating condition gravity center is obtained: , wherein c represents the operating condition gravity center, n represents the sample amount, , and the sparse coefficient sequence of the nth sample is represented by sn.

[0042] The operating condition gravity center described above integrates the structural components common to each experiment, including trend changes, steady-state oscillations, decay processes, and local disturbances, and suppresses the noise, disturbances, or local abnormalities of individual test times through multi-sample averaging, so that the operating condition gravity center can more accurately represent the inherent dynamic characteristics of the operating condition.

[0043] Since the operating condition gravity center is relatively abstract, in some embodiments, a typical waveform is also provided for more intuitive and rapid preliminary judgment of the current operating condition.

[0044] The typical waveform is obtained based on the operating condition gravity center. After obtaining the operating condition gravity center, the operating condition gravity center is projected into the basis function library, and a typical waveform having complete structural expression capability and capable of reflecting the characteristics of the current operating condition is obtained through time-domain inversion. This operation is a routine operation in the field, and therefore the specific construction steps thereof will not be described herein.

[0045] The typical waveform directly presents the trend item, main periodic structure, swing amplitude, local disturbance, and transient anomaly of the operating condition in an intuitive and visual manner, so that the structural differences between operating conditions are "visualized" and "understood". Such visual expression not only facilitates rapid identification of the operating condition state by field engineers, but also provides an important reference for manual review, anomaly diagnosis, trajectory control, and parameter adjustment suggestions. In addition, as a structured template, the typical waveform actually constitutes the basic unit of the operating condition knowledge base, and can be used as a key input for future operating condition expansion, algorithm migration, model checking, and downhole anomaly identification. Therefore, even though the typical waveform is not directly used in the calculation of the discriminant formula, it is still an indispensable part of the application in engineering applications, and is a key bridge connecting the sparse structural expression and the actual drilling behavior.

[0046] S6, repeating S3-S5 to obtain the operating condition gravity centers of multiple operating conditions and form a typical dictionary; The typical waveforms mentioned above can also be included in the typical dictionary for easy reference by staff. In particular, for the sparse coefficients used to retrieve typical waveforms, considering the potential amplitude scale differences introduced by equipment gain, installation location, and measurement links, their amplitudes need to be normalized. Furthermore, the sampling time length, sampling frequency, or signal start point of different experiments are not entirely consistent. To eliminate these time-domain scale differences, this step maps all typical waveforms to a unified time axis and performs phase consistency correction on highly periodic operating conditions, ensuring that the typical structures of different operating conditions have comparability and a consistent time reference under the reference coordinates. Through this series of standardization steps, all operating condition templates in the typical dictionary are within a unified scale, a unified time reference, and a unified structural expression system, thus providing a reliable foundation for subsequent similarity measurement and projection analysis.

[0047] S7. Take the tool surface signal to be identified. Following the operations of S1 to S3, obtain the target sparse coefficient of the tool surface signal to be identified. Calculate the Euclidean distance between the target sparse coefficient and the centroid of the typical working condition in the dictionary. Based on the principle of minimum distance, obtain the classification of the tool surface signal to be identified.

[0048] To further illustrate the superiority of the method in the embodiments of the present invention, specific test examples are given below.

[0049] Using the device disclosed in this embodiment, relevant tests on the tool face working condition are performed. The attitude sensor collects the biaxial acceleration and calculates the initial tool face angle of the tool face working condition.

[0050] After unwinding and resampling the tool face angle, a tool face angle time series signal with 8 periods and 512 sampling points is obtained, as follows: Figures 3~5 As shown, where, Figure 3 To stabilize the tool surface, the first set of original signals and the resampled signals are used; Figure 4 To stabilize the second set of original signals and resampled signals on the tool surface; Figure 5 The third set of original signals and resampled signals for stabilizing the tool surface; The tool facet angle time series signal is sparsely decomposed using a basis function library. The variance of each basis function is obtained based on the initial sparse coefficients. Sensitive basis functions are then selected based on a threshold, such as... Figure 6 As shown.

[0051] This case study will utilize two types of toolface sample signals to invert typical waveforms. After preprocessing and basis function decomposition, the experimental samples are mean-valued along their dimensions using their complete sparse coefficients to obtain the centroid vector of the stable toolface condition. This centroid vector is then reprojected onto the complete basis function dictionary. Simultaneously, using the same method, the parameters of three sets of impact toolface conditions (see...) are... Figure 7) are tested, and finally, time-domain typical waveforms of two tool face working conditions are obtained as shown in Figure 8 Figure 8 (a) in FIG. 1 shows a comparison chart of typical waveforms and original waveforms in a stable tool face working condition, Figure 8 (b) in FIG. 1 shows a comparison chart of typical waveforms and original waveforms in a collision tool face working condition.

[0052] The application has been disclosed above with preferred embodiments, but those skilled in the art should understand that these embodiments are only used to describe the application, and should not be understood as limiting the scope of the application. Further improvements can also be made without departing from the principles of the application, and these improvements should also be considered as the protection of the application.​

Claims

1. A method for sparse characterization and classification of toolface state while drilling for a horizontal well drilling string whirling system, characterized in that, The method comprises the following steps: S1, a base function library of sparse representation is established, and the base function library comprises base functions for processing trend items, periodic oscillation items, decay items, mutation items and non-smooth edge structure items; S2, a plurality of tool face angle test data under different working conditions are collected and preprocessed to obtain tool face angle time sequence signals under different working conditions; S3, based on the base function library and the characteristics of the tool face angle time sequence signals, the tool face angle time sequence signals are sparsely decomposed by using corresponding base functions to obtain preliminary sparse coefficients; S4, based on the change of the base function under the current working condition, a base function with a variance greater than a threshold value is selected as a sensitive base function, and all sensitive base functions of the current working condition form a discriminant subspace of the current working condition; S5, the average of the sparse coefficients of the plurality of tool face angle time sequence signals in the discriminant subspace of the current working condition is taken, and the average is taken as the working condition gravity center of the current working condition; S6, S3-S5 are repeated to obtain the working condition gravity centers of the multiple working conditions and form a typical dictionary; S7, the tool face signal to be identified is taken, the target sparse coefficient of the tool face signal to be identified is obtained according to the operations of S1-S3, the Euclidean distance of the target sparse coefficient in the discriminant subspace and the working condition gravity center in the typical dictionary is calculated, and the classification of the tool face signal to be identified is obtained based on the minimum distance principle.

2. The method of claim 1, wherein, In S1, the base function for processing the trend item is a polynomial base function, the base function for processing the periodic oscillation item is a Fourier sine / cosine base function, the base function for processing the decay item is an exponential decay base function, the base function for processing the mutation item is a ReLU type segmented activation base function, and the base functions for processing the non-smooth edge structure item are sawtooth wave base functions, square wave base functions and triangular wave base functions.

3. The method of claim 1, wherein, In S2, the device structure for collecting tool face angle test data under different working conditions is as follows: The device is composed of a servo motor, an analog drill string, a universal joint and a displacement attitude sensor, wherein the servo motor is used for controlling the drilling of the analog drill string, the universal joint is used for turning the drill string, and the displacement attitude sensor is used for collecting Y-axis acceleration, Z-axis acceleration and angular velocity.

4. The method of claim 3, wherein, In S2, after the original data are collected, the following steps are further included: S21, calculating the tool face angle: wherein denotes the tool face angle, Accy denotes the y-axis acceleration, and Accz denotes the z-axis acceleration. S22, the obtained initial tool face angle data are processed by using an unwrapping algorithm to obtain real tool face angle time sequence data; S23, the angle sequence is resampled by using a linear interpolation method, all signals are uniformly mapped to a fixed length of standard sampling point number, and a tool face angle time sequence signal of the tool face angle changing with time is established.

5. The method of claim 1, wherein, In S3, the trend item, the periodic oscillation item, the decay item, the mutation item and the non-smooth edge structure item of the tool face angle time sequence signal are sparsely decomposed by using corresponding base functions.

6. The method of claim 1, wherein, In S4, the threshold value is 0.04-0.06 times the maximum variance of the current working condition.

7. The method of claim 1, wherein, In S5, the following steps are further included: after the working condition gravity center of the current working condition is obtained, it is projected into the base function library for inversion to obtain a typical waveform of the current working condition.

8. The method of claim 1, wherein, In S5, the solving method of the sparse coefficient of the tool face angle time series signal in the discriminant subspace is as follows: the tool face angle time series signal includes a trend item, a periodic oscillation item, a decay item, a mutation item and a non-smooth edge structure item, if the discriminant subspace does not contain the base function corresponding to a certain type of item, the item does not participate in the coefficient update in the discriminant process; if the discriminant subspace contains the base function corresponding to the item, the sparse coefficient is projected or reconstructed on the base function subset only, so as to obtain the consistent sparse coefficient component in the discriminant subspace; The sparse coefficient component obtained in the discriminant subspace is combined with the coefficient component not participating in the discrimination to form a sparse coefficient vector consistent with the structure of the discriminant subspace.

Images