Control method and system of rotary steering system

Abstract

The invention relates to the technical field of control adjustment, in particular to a control method and system for a rotary steering system, and the method comprises the following steps: obtaining the original attitude data of the rotary steering system; obtaining an attitude estimation value; obtaining an attitude angle; acquiring a control instruction; obtaining a filtering signal and a derivative, and outputting a control signal of the rotary steering system; a closed-loop optimized path is formed through links such as data preprocessing, attitude resolving, fault compensation, signal smoothing and time lag offset, and the control precision and stability of the rotary steering system are effectively improved; the influence of the fault on the control performance is reduced; smooth processing of a control instruction and time delay compensation of the extreme learning machine effectively suppress oscillation of a control signal, and the dynamic response characteristic of the system is improved. The multi-layer control architecture can adapt to changes of complex underground environments, it is guaranteed that the rotary steering system accurately drills according to a preset track, and the drilling efficiency and the well hole quality are improved.

Description

Technical Field

[0001] This invention relates to the field of control and regulation technology, and in particular to a control method and system for a rotary guide system. Background Technology

[0002] Rotary steerable drilling (RSS) is a key technology in the field of oil and gas exploration and development. It allows the drill string to rotate continuously while the drill bit direction is precisely adjusted to achieve complex wellbore trajectory control.

[0003] Compared to traditional sliding directional drilling, RSS can significantly improve drilling efficiency, wellbore trajectory smoothness, and wellbore cleanliness, making it particularly suitable for the development of unconventional oil and gas resources such as deep-sea and shale gas.

[0004] However, existing rotary steering system control methods still face many challenges. First, the downhole geological environment is extremely complex, with a combination of factors such as formation hardness, in-situ stress, friction between the drill string and the wellbore, downhole vibration, and high temperature, which cause the system to exhibit strong nonlinearity, time-varying characteristics, hysteresis, and uncertainty.

[0005] In particular, the strong vibration and noise near the drill bit can severely overwhelm the original signal of the sensor, resulting in a significant decrease in attitude measurement accuracy. Although traditional PID and robust control methods can solve some problems to a certain extent, they are highly dependent on accurate models and prior knowledge, making it difficult to cope with these complex unknown disturbances and model uncertainties. In addition, in the design of the control system, it is also difficult to efficiently handle the multimodal data processing problems caused by the coupling of multiple variables and multiple links. Summary of the Invention

[0006] The main objective of this invention is to provide a control method for a rotary guide system, which aims to solve the problem of control time delay in the face of vibration and noise in existing control methods.

[0007] To achieve the above objectives, the present invention provides a control method for a rotary guide system, the control method comprising the following steps: The raw attitude data of the rotary steering system is acquired, the filtering parameters are adjusted using the Kalman filter algorithm, and the attitude estimate is output. Based on the attitude estimation value, the attitude angle of the rotary guide system is output through speed compensation; The radial basis function neural network outputs control commands for attitude angles after fault compensation. The control commands are smoothed, and the filtered signal and derivative of the control commands are obtained. Based on the filtered signal and derivative, the execution time delay is compensated by the extreme learning machine, and the control signal of the rotary guidance system is output.

[0008] Optionally, the step of acquiring the original attitude data of the rotary guidance system, adjusting the filtering parameters using a Kalman filter algorithm, and outputting the attitude estimate includes the following steps: The analog output signal of the acquisition module is acquired, digitized, and then transmitted to the control module through the communication interface. Obtain the dynamic model of the rotary steering system, and use the dynamic model to transfer the optimal covariance of the previous moment to the current moment. After adding the process noise covariance, output the updated mean square error matrix. Based on the mean square error matrix, the attitude estimate is output after correcting the predicted state with Kalman gain.

[0009] Optionally, outputting the attitude angle of the rotary guide system through rotational speed compensation based on the attitude estimate includes the following steps: After obtaining the attitude estimate and updating the optimal covariance matrix, the attitude is calculated to output the attitude angles of the rotary steering system, which include the well inclination angle and the tool face angle.

[0010] Optionally, the step of outputting fault-compensated control commands for attitude angles via a radial basis function neural network includes the following steps: The fault dynamic function is obtained, the attitude estimate is mapped to a high-dimensional space through a nonlinear activation function, and then multiplied with the weight vector to obtain the estimated value of the fault signal. Obtain the observation error, and add the observation error to the integral of the observation error to obtain an integral sliding surface; Obtain the fault-tolerant control law, and then use the compensation force to integrate the convergence value of the attitude estimate of the rotary guide system onto the sliding surface to generate control commands.

[0011] Optionally, the step of obtaining the fault-tolerant control law, and then integrating the convergence value of the attitude estimate of the rotary guide system onto the sliding mode surface to generate control commands, includes the following steps: Obtain the nominal control law; After detecting and estimating the fault signal using a radial basis function neural network, a fault-tolerant control law is generated. This fault-tolerant control law is used to generate a compensating force to counteract the impact of the fault signal on the rotary steering system. The nominal control law and the fault-tolerant control law are superimposed to output the control command.

[0012] Optionally, the step of compensating for execution time delay using an extreme learning machine based on the filtered signal and derivative, and outputting the control signal for the rotary guidance system, includes the following steps: Obtain the approximation function of the filtered signal and its derivative under extreme learning, and model the execution delay; Obtain the adaptive law of the extreme learning machine to update the network weights, and obtain compensation signals to offset the execution delay; The control signal is obtained by combining the filtered signal, derivative, and compensation signal, and the control signal is used to drive the rotary guide system.

[0013] To achieve the above objectives, the present invention also provides a control system, the control system comprising: The acquisition module is used to acquire the original attitude data of the rotary guide system, adjust the filtering parameters through the Kalman filter algorithm, and output the attitude estimate. The compensation module is used to output the attitude angle of the rotary guide system based on the attitude estimation value and through rotational speed compensation. The control module is used to output fault-compensated control commands for attitude angles through a radial basis function neural network; The control module is also used to smooth the control commands and obtain the filtered signal and derivative of the control commands; The control module is also used to compensate for execution delays using an extreme learning machine based on the filtered signal and derivative, and to output control signals for the rotary guidance system.

[0014] Optionally, the control module includes an observer embedded with a radial basis function neural network, which is used to detect and estimate fault signals.

[0015] Optionally, the control module further includes a command filter, which is used to acquire the filtered signal and the derivative.

[0016] Optionally, the control module further includes an extreme learning machine, which is used to obtain the approximation function of the filtered signal and its derivative under extreme learning, and to model the execution delay; The extreme learning machine is also used to obtain an adaptive law to update the network weights and to obtain a compensation signal to offset the execution delay. The extreme learning machine is also used to combine the filtered signal, derivative and compensation signal to obtain a control signal, which is used to drive the rotary guide system.

[0017] The beneficial effects that this invention can achieve are as follows: This invention forms a closed-loop optimization path through data preprocessing, attitude calculation, fault compensation, signal smoothing, and time delay cancellation, effectively improving the control accuracy and stability of the rotary steering system. This, in turn, significantly enhances the control accuracy and stability of the rotary steering system in complex downhole environments. The Kalman filter algorithm with dynamic parameter adjustment significantly improves the noise immunity of attitude estimation, reducing attitude angle measurement errors. The radial basis function neural network enables adaptive compensation for system faults, minimizing the impact of faults on control performance. Smoothing of control commands and time delay compensation using the extreme learning machine effectively suppress control signal oscillations, improving the system's dynamic response characteristics. This multi-level control architecture can adapt to changes in complex downhole environments, ensuring the rotary steering system drills precisely along a predetermined trajectory, improving drilling efficiency and wellbore quality. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0019] Figure 1 This is a flowchart illustrating the control method in Embodiment 1 of the present invention.

[0020] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] 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 a part of the embodiments of the present invention, and not all of the 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.

[0022] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a specific posture. If the specific posture changes, the directional indication will also change accordingly.

[0023] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0025] Example 1 Reference Figure 1 This embodiment provides a control method for a rotary guide system, the control method including the following steps: The raw attitude data of the rotary steering system is acquired, the filtering parameters are adjusted using the Kalman filter algorithm, and the attitude estimate is output. Based on the attitude estimation value, the attitude angle of the rotary guide system is output through speed compensation; The radial basis function neural network outputs control commands for attitude angles after fault compensation. The control commands are smoothed, and the filtered signal and derivative of the control commands are obtained. Based on the filtered signal and derivative, the execution time delay is compensated by the extreme learning machine, and the control signal of the rotary guidance system is output.

[0026] It should be noted that in traditional rotary steerable drilling technology, the complex downhole environment causes severe interference from vibration and noise to the raw attitude data of sensors. The time-varying characteristics of formation parameters lead to mismatch in the dynamic model, and the multimodal data generated by the coupling of multiple components is difficult to process efficiently by traditional control architectures. Insufficient noise suppression during the attitude calculation process leads to the accumulation of angle estimation errors, and the failure to dynamically compensate for fault signals and execution delays causes oscillations in control commands, ultimately resulting in a decrease in the trajectory tracking accuracy and a delay in dynamic response of the rotary steerable system.

[0027] For example, in deep-sea drilling operations, when the drill string rotates continuously at 150 revolutions per minute, the signal-to-noise ratio of the raw signal acquired by the triaxial accelerometer near the drill bit is less than 20 dB. The fixed-parameter Kalman filter fails to adaptively adjust the process noise covariance, causing the attitude estimate to deviate from the true value by more than 0.5 degrees. Sudden changes in formation hardness trigger nonlinear changes in the drill string-wellbore contact force. The fixed weight vector of the radial basis function neural network cannot match the fault dynamic function in real time, resulting in overshoot on the integral sliding surface of the observation error. Furthermore, the second-order derivative signal output by the command filter is not correlated with the time-delay model of the actuator, causing a phase lag of 50 milliseconds in the control signal, leading to a steady-state deviation between the guide wing rib deployment angle and the set value.

[0028] Based on the above problems, this embodiment provides a control method for a rotary steering system. By combining Kalman filtering, radial basis function neural network and extreme learning machine, a multi-stage compensation mechanism is constructed. Furthermore, through the synergistic effect of noise suppression, fault tolerance and time delay compensation, the attitude control problem of the rotary steering system under strong nonlinearity and time-varying disturbances is solved.

[0029] Specifically, by acquiring the original attitude data of the rotary steering system, the Kalman filter algorithm is used to dynamically adjust the filter parameters and output the attitude estimate. The Kalman filter algorithm optimizes the state estimate in real time through two stages: prediction and update. In the prediction stage, the system model is used to predict the current state, and in the update stage, the prediction results are corrected by combining the measurement data. Dynamically adjusting the filter parameters can adapt to changes in the system noise characteristics and improve the accuracy of attitude estimation.

[0030] After obtaining the attitude estimate, the attitude angle of the rotary guide system is output through speed compensation. Speed ​​compensation takes into account the influence of drill rotation on attitude measurement, and then the measurement error caused by rotation is eliminated through compensation algorithm to obtain a more accurate attitude angle.

[0031] Next, the attitude angle is obtained, and the fault-compensated control command is output through the radial basis function neural network. The radial basis function neural network has good nonlinear mapping ability, can adaptively learn the fault characteristics of the system, and through online training, the neural network continuously optimizes its weight parameters to achieve real-time fault compensation and output more reliable control commands.

[0032] Then, the control commands are smoothed to obtain the filtered signals and derivatives of the control commands. Smoothing can reduce the high-frequency oscillations of the control commands and improve system stability. The filtered signals and derivative information provide a basis for subsequent time delay compensation.

[0033] Finally, the control signal of the rotary guide system is output by compensating for the execution time delay through an extreme learning machine. The extreme learning machine has a fast learning capability and can model and compensate for the time delay characteristics of the actuator in real time. By combining the filtered signal, derivative information and time delay compensation, the final control signal is generated to drive the rotary guide system to perform precise control.

[0034] The aforementioned multi-level control architecture, through data preprocessing, attitude calculation, fault compensation, signal smoothing, and time delay cancellation, forms a closed-loop optimization path, effectively improving the control accuracy and stability of the rotary steering system. This, in turn, significantly enhances the control accuracy and stability of the rotary steering system in complex downhole environments. The Kalman filter algorithm with dynamic parameter adjustment significantly improves the noise immunity of attitude estimation, reducing attitude angle measurement errors. The radial basis function neural network enables adaptive compensation for system faults, minimizing the impact of faults on control performance. Smoothing of control commands and time delay compensation using the extreme learning machine effectively suppress control signal oscillations, improving the system's dynamic response characteristics. This multi-level control architecture can adapt to changes in complex downhole environments, ensuring the rotary steering system drills precisely along a predetermined trajectory, improving drilling efficiency and wellbore quality.

[0035] In some embodiments, the Kalman filter algorithm refers to an algorithm that uses the state equation of a linear system and observation data to make an optimal estimate of the system state. Specifically, it can recursively adjust the filter parameters in two steps: prediction and update, and achieve noise suppression by minimizing the mean square error matrix, thereby improving the attitude estimation accuracy in complex noise environments.

[0036] In some embodiments, rotational speed compensation refers to the dynamic correction of the attitude angle of the rotary guide system. Specifically, it can be achieved by using a dynamic model combined with the process noise covariance to update the optimal covariance matrix, thereby eliminating measurement lag and errors caused by rotational motion and ensuring the accuracy of attitude calculation.

[0037] In some embodiments, a radial basis function neural network refers to a neural network with a single hidden layer feedforward structure. Specifically, it can use a nonlinear activation function to map the input to a high-dimensional space, approximate the fault signal through a linear combination of weight vectors, realize online estimation and compensation of the fault dynamic function, and enhance the fault tolerance of the system.

[0038] In some embodiments, smoothing refers to filtering and differentiating control commands. Specifically, a command filter can be used to generate a continuously differentiable filtered signal and its derivative to avoid actuator oscillations caused by sudden command changes and improve control stability.

[0039] In some embodiments, the Extreme Learning Machine refers to a training method for a single hidden layer feedforward neural network. Specifically, it can randomly initialize input weights and adaptively adjust output weights to approximate a nonlinear function of execution time delay, generate compensation signals to offset the effects of time delay, and ensure the real-time performance of control signals.

[0040] In this embodiment, the process of acquiring the original attitude data of the rotary guidance system, adjusting the filtering parameters using a Kalman filter algorithm, and outputting the attitude estimate includes the following steps: The analog output signal of the acquisition module is acquired, digitized, and then transmitted to the control module through the communication interface. Obtain the dynamic model of the rotary steering system, and use the dynamic model to transfer the optimal covariance of the previous moment to the current moment. After adding the process noise covariance, output the updated mean square error matrix. Obtain the mean squared error matrix, and output the attitude estimate after correcting the predicted state with Kalman gain.

[0041] It should be noted that the analog signals from the acquisition module undergo digital processing to reduce distortion during signal transmission. The communication interface ensures efficient data transmission to the control module. The dynamic model uses the optimal covariance from the previous moment for state prediction, and updates the mean square error matrix by superimposing the process noise covariance. This process quantifies the impact of model error on prediction. The Kalman gain is dynamically adjusted based on the updated mean square error matrix, performing a linear weighted correction on the predicted state. A larger gain value indicates a smaller impact of measurement noise on the estimation results. For example, the state transition matrix can be constructed based on a quaternion attitude update model, and the process noise covariance is set as a diagonal matrix within the range of 0.01-0.1 based on the downhole vibration intensity. By recursively updating the mean square error matrix and dynamically adjusting the Kalman gain, the accumulation of model error can be effectively suppressed in strong vibration and noise environments, improving the convergence speed and accuracy of attitude estimates.

[0042] It should also be noted that, based on the above process, the strong noise interference in the complex downhole environment is effectively handled, and the accuracy and reliability of attitude measurement are improved. As a result, the rotary steering system can more accurately sense its own attitude, providing reliable basic data for subsequent trajectory control, thereby improving overall drilling efficiency and wellbore quality.

[0043] In this embodiment, the process of outputting the attitude angle of the rotary guide system through rotational speed compensation based on the attitude estimation value includes the following steps: After obtaining the attitude estimate and updating the optimal covariance matrix, the attitude is calculated to output the attitude angles of the rotary steering system, which include the well inclination angle and the tool face angle.

[0044] Understandably, based on the attitude estimate output by the Kalman filter, the optimal covariance matrix is ​​first updated based on the current measurement noise covariance. For example, the inverse matrix weighting of the covariance matrix can be used to eliminate gyroscope drift errors caused by high-speed rotation. Then, the attitude estimate is transformed from the drill string's rotating coordinate system to a fixed geographic coordinate system using quaternion differential equations or direction cosine matrices. For example, the gravity vector direction can be calculated by fusing accelerometer and three-axis fluxgate data. Finally, the wellbore inclination angle is calculated by the angle between the drill string axis and the gravity vector, and the tool face angle is calculated by the rotation angle projected onto the horizontal plane. For example, the arctangent function can be used to calculate the angle between the magnetic north direction and the drill string projection direction. By dynamically updating the covariance matrix, noise interference in the attitude calculation process is effectively suppressed, and the accuracy of the wellbore inclination angle and tool face angle calculations is improved, providing reliable input for subsequent fault compensation by the radial basis function neural network.

[0045] In this embodiment, the step of outputting fault-compensated control commands for attitude angles via a radial basis function neural network includes the following steps: The fault dynamic function is obtained, the attitude estimate is mapped to a high-dimensional space through a nonlinear activation function, and then multiplied with the weight vector to obtain the estimated value of the fault signal. Obtain the observation error, and add the observation error to the integral of the observation error to obtain an integral sliding surface; Obtain the fault-tolerant control law, and then use the compensation force to integrate the convergence value of the attitude estimate of the rotary guide system onto the sliding surface to generate control commands.

[0046] Understandably, after the attitude estimate is input into the radial basis function neural network, it undergoes nonlinear spatial mapping through a Gaussian kernel function. The kernel center point is generated based on clustering of historical fault data, and the kernel width parameter is optimized using gradient descent. The mapped high-dimensional feature vector is multiplied by a dynamically adjusted weight vector to output the amplitude and phase estimates of the fault signal. The observation error is calculated as the difference between the current attitude angle and the reference trajectory, and its integral term is successively accumulated using a discrete-time integrator. The integral sliding surface superimposes the error and the integral term at a preset ratio to form a convergence condition with anti-interference capability. The fault-tolerant control law generates a compensation force based on the sliding surface state. This compensation force is superimposed on the nominal control command in the time domain, and the compensation force gain coefficient is determined through Lyapunov function stability analysis. By adjusting the weight vector and integral sliding surface parameters in real time, it is ensured that the control command can still accurately track the preset trajectory even when sensor noise and actuator faults coexist.

[0047] It is also understandable that, based on the above process, faults in the rotary steering system can be effectively detected and compensated, improving the system's robustness and control accuracy. Thus, the system can maintain stable operation in complex geological environments, reducing downtime and economic losses caused by faults. In summary, this method estimates fault signals in real time through a radial basis function neural network, combined with an integral sliding mode control strategy, enabling rapid response and compensation for various unknown faults and disturbances, thereby ensuring precise control of the drilling trajectory.

[0048] In this embodiment, the step of obtaining the fault-tolerant control law, and then integrating the convergence value of the attitude estimate of the rotary guide system onto the sliding mode surface to generate control commands, includes the following steps: Obtain the nominal control law; After detecting and estimating the fault signal using a radial basis function neural network, a fault-tolerant control law is generated. This fault-tolerant control law is used to generate a compensating force to counteract the impact of the fault signal on the rotary steering system. The nominal control law and the fault-tolerant control law are superimposed to output the control command.

[0049] It should be noted that the nominal control law is designed as an inverse model based on the dynamic equations of the rotary steering system, and its output includes the basic components of the pushing force and the bias force. When the radial basis function neural network detects a decrease in actuator efficiency or a sensor offset fault, the network extracts fault features through nonlinear mapping of the hidden layers, and the output layer weight matrix is ​​updated online with a learning rate of 0.01-0.1 to generate a compensation force matching the fault amplitude. The integral term of the fault-tolerant control law is used to eliminate steady-state error, and its integral action time window is limited to 10-50 milliseconds to avoid overshoot. The superposition process of the nominal control law and the fault-tolerant control law is completed synchronously within the control cycle, and their output components are linearly superimposed in the force / torque dimension to form a composite control command containing the basic control quantity and the fault compensation quantity. This composite command is applied to the guide pushing block through the actuator, so that the tracking error between the actual attitude angle and the desired trajectory converges to within ±0.5°.

[0050] It should also be noted that, through the above technical solution, the present invention realizes real-time detection and compensation of faults in the rotary steering system. The radial basis function neural network can effectively estimate unknown fault signals, and the fault-tolerant control law can generate precise compensation force to offset the impact of faults. The superposition of nominal control and fault-tolerant control ensures the control performance of the system under normal and fault conditions, thereby improving the reliability and robustness of the rotary steering system and ensuring the trajectory control accuracy under complex geological conditions.

[0051] In this embodiment, the step of compensating for execution delay using an extreme learning machine based on the filtered signal and derivative, and outputting the control signal for the rotary guidance system, includes the following steps: Obtain the approximation function of the filtered signal and its derivative under extreme learning, and model the execution delay; Obtain the adaptive law of the extreme learning machine to update the network weights, and obtain compensation signals to offset the execution delay; The control signal is obtained by combining the filtered signal, derivative, and compensation signal, and the control signal is used to drive the rotary guide system.

[0052] It should be noted that after the filtered signal and its derivative are input into the Extreme Learning Machine (ELM), they are mapped to a high-dimensional space through randomly generated hidden layer parameters. The output layer weights are dynamically adjusted according to the real-time error, and the approximation function approximates the actual time-delay model with an exponential convergence rate. The adaptive law updates the weight matrix through gradient descent of the error signal, and the derivatives of the compensation signal and the filtered signal are linearly superimposed to offset the phase lag caused by the time delay. During the control signal generation process, the single hidden layer structure of the ELM avoids traditional neural networks from getting trapped in local optima. Its millisecond-level training speed meets the real-time control requirements of downhole drilling, and the time delay of the compensated control signal is reduced to less than 5 milliseconds, effectively improving the dynamic response performance of the rotary steering system.

[0053] It should also be noted that the above process can effectively compensate for execution delay, improve the control accuracy and response speed of the rotary steering system, and the introduction of the extreme learning machine enables the system to have adaptive capabilities and cope with complex and ever-changing downhole environments. At the same time, by smoothing the control commands and utilizing derivative information, the robustness and stability of the system are further enhanced. This data-driven intelligent control method overcomes the dependence of traditional control methods on precise models and provides a new solution for high-precision control of rotary steering systems.

[0054] Example 2: This embodiment provides a control system, which includes: The acquisition module is used to acquire the original attitude data of the rotary guide system, adjust the filtering parameters through the Kalman filter algorithm, and output the attitude estimate. The compensation module is used to output the attitude angle of the rotary guide system based on the attitude estimation value and through rotational speed compensation. The control module is used to output fault-compensated control commands for attitude angles through a radial basis function neural network; The control module is also used to smooth the control commands and obtain the filtered signal and derivative of the control commands; The control module is also used to compensate for execution delays using an extreme learning machine based on the filtered signal and derivative, and to output control signals for the rotary guidance system.

[0055] It is understandable that when traditional rotary steering systems operate in complex downhole environments, there is an execution time lag in the transmission of control signals to the actuators, which causes a deviation between the control signals and the actual state of the system, affecting the steering accuracy.

[0056] This embodiment further provides a rotary guidance system, including a data acquisition module, a compensation module, and a control module. The data acquisition module acquires raw attitude data and outputs an attitude estimate using a Kalman filter algorithm; the compensation module performs rotational speed compensation based on the attitude estimate and outputs the attitude angle; the control module generates fault-compensated control commands using a radial basis function neural network, smooths the control commands to obtain a filtered signal and its derivative, and then compensates for the execution time delay using an extreme learning machine before outputting the final control signal.

[0057] The control module includes a command filter, a radial basis function neural network, and an extreme learning machine. The command filter performs low-pass filtering on the control commands to suppress high-frequency noise and extract the signal derivatives. The extreme learning machine adopts a single hidden layer feedforward neural network structure, with Gaussian functions selected as the activation functions for the hidden layer nodes. The input layer receives the filtered signal and its derivatives, and the output layer generates a compensation signal. The network weights are updated online through an adaptive law, and the compensation signal is superimposed with the filtered signal to generate the control signal.

[0058] More specifically, after the control command is processed by the command filter, the filtered signal retains the low-frequency components of the command, and the derivative reflects the signal change trend. The Extreme Learning Machine (ELM) establishes an approximation model of the execution delay based on the filtered signal and the derivative, adjusts the network weights through an adaptive law, estimates the delay in real time, and generates a compensation signal. The compensation signal is superimposed on the filtered signal to offset the phase deviation caused by the actuator's response delay. For example, the number of hidden layer nodes is set to 10-20, and the adaptive law learning rate is selected in the range of 0.01-0.1. This ensures convergence speed while avoiding weight oscillations. Finally, when the control signal drives the guiding mechanism, the compensated signal is synchronized with the actual system state, improving the dynamic response accuracy of the rotary guiding system.

[0059] As a preferred embodiment, the preferred implementation process of this embodiment is as follows: The acquisition module obtains the raw attitude data of the rotary guide system, adjusts the filtering parameters through the Kalman filter algorithm, and outputs the attitude estimate. The compensation module obtains the attitude estimate and outputs the attitude angle of the rotary guide system through speed compensation. The control module obtains the attitude angle and outputs the fault-compensated control command through the radial basis function neural network. The control module also obtains the control command, smooths the control command, and obtains the filtered signal and derivative of the control command. The control module further obtains the filtered signal and derivative, compensates for the execution time delay through limit learning, and outputs the control signal of the rotary guide system.

[0060] The acquisition module may include an accelerometer and a three-axis fluxgate magnetometer for measuring the acceleration and magnetic field strength of the rotary steering system. The compensation module may include a rotational speed sensor and a compensation algorithm for measuring the drill string rotational speed and compensating for the attitude estimate. The control module may include an observer, a command filter, and an extreme learning machine. The observer is embedded with a radial basis function neural network for detecting and estimating fault signals. The command filter is used to acquire the filtered signal and its derivative. The extreme learning machine is used to acquire the approximation function of the filtered signal and its derivative under extreme learning, model the execution delay, acquire an adaptive law to update the network weights, acquire a compensation signal to offset the execution delay, and combine the filtered signal, derivative, and compensation signal to obtain the control signal.

[0061] The above technical solutions effectively address the problems of strong nonlinearity, time-varying, lag, and uncertainty in complex downhole environments. The acquisition module improves the accuracy of attitude estimation through the Kalman filter algorithm, the compensation module improves the accuracy of attitude angle through rotational speed compensation, the control module achieves fault compensation through radial basis function neural network, and compensates for execution time delay through extreme learning machine.

[0062] In this embodiment, the control module includes an observer embedded with a radial basis function neural network, which is used to detect and estimate fault signals.

[0063] Understandably, during operation, the observer first inputs the attitude estimate into the input layer of the radial basis function neural network. The Gaussian function in the hidden layer performs a nonlinear transformation on the input, and the output layer generates the fault estimate by multiplying the weight vector with the output of the hidden layer. The center vector and width parameters of the hidden layer are pre-determined through sample training, while the weight vector is adjusted online based on real-time observation errors. When downhole vibration causes abnormal sensor signals, the neural network separates the fault component through high-dimensional space mapping and superimposes this component as a compensation signal into the control command. By dynamically adjusting the weight vector, the neural network can adaptively track changes in the fault signal, thereby maintaining robustness in fault detection under strong noise environments. This structure effectively solves the problem of insufficient fault separation capability of traditional observers under nonlinear disturbances and improves the fault tolerance performance of the control system.

[0064] In this embodiment, the control module further includes a command filter, which is used to acquire the filtered signal and the derivative.

[0065] In some preferred embodiments, the command filter is implemented using a second-order Butterworth low-pass filter, whose transfer function is determined by the system control bandwidth. When a control command is input to the command filter, the filter's cutoff frequency is adaptively adjusted according to the current attitude angle error, specifically by dynamically configuring the filter parameters after online identification of the system response frequency. During the filtering process, high-frequency noise components of the control command are filtered out, generating a smooth filtered signal; simultaneously, the first derivative of the filtered signal is calculated using a numerical differentiation algorithm to obtain the signal rate of change. The filtered signal and the derivative signal are transmitted to the Extreme Learning Machine module through a parallel transmission channel to construct a compensation model for execution delay.

[0066] In this embodiment, the control module further includes an extreme learning machine, which is used to obtain the approximation function of the filtered signal and its derivative under extreme learning, and to model the execution delay; The extreme learning machine is also used to obtain an adaptive law to update the network weights and to obtain a compensation signal to offset the execution delay. The extreme learning machine is also used to combine the filtered signal, derivative and compensation signal to obtain a control signal, which is used to drive the rotary guide system.

[0067] It should be noted that the Extreme Learning Machine (ELM) first constructs a time-delay dynamic model using the filtered signal and its derivative from the input layer. It then maps the input signal to a high-dimensional space using a Gaussian kernel function and fits the nonlinear characteristics of the time delay through hidden layer nodes. An adaptive law adjusts the network weights in real time based on the error between the control signal and the desired response, ensuring that the compensation signal can offset the phase lag caused by the time delay. Finally, the compensation signal, the filtered signal, and the derivative are superimposed in a preset ratio to generate a control signal that eliminates the effects of the time delay. For example, when the actuator has a 0.5-second time delay, the ELM predicts the system state change within the next 0.5 seconds using an approximation function and generates a corresponding compensation signal, causing the control signal to act on the actuator 0.5 seconds earlier, thereby eliminating the tracking error caused by the time delay.

[0068] Example 3: As a preferred embodiment, the preferred implementation process of this embodiment is as follows: In the stage of acquiring the original attitude data of the rotary guidance system, an inertial measurement unit composed of a high-precision three-axis accelerometer and a three-axis gyroscope is used to collect the original attitude data at a sampling frequency of 200Hz. After the original data is converted from analog to digital by 16 bits, it is transmitted to the control processor via the CAN bus.

[0069] The Kalman filter algorithm employs a 9-dimensional state vector, including three axes of angle, angular velocity, and angular acceleration. In the prediction phase, the fourth-order Runge-Kutta method is used to solve the system state equations. In the update phase, an adaptively adjusted measurement noise covariance matrix is ​​used. The dynamic adjustment of the filter parameters is achieved by analyzing the statistical characteristics of the innovation sequence, and the process noise covariance matrix is ​​updated every 100ms.

[0070] The rotational speed compensation adopts a quaternion-based attitude calculation algorithm, which combines the drill string rotational speed information to compensate the gyroscope measurement value. The attitude angle calculation frequency is 10Hz, and the output well inclination angle and tool face angle are output.

[0071] The radial basis function neural network uses Gaussian radial basis functions, with 50 hidden layer nodes. The network input is the attitude angle and its rate of change, and the output is the fault compensation amount. The network weights are updated online using the recursive least squares method, with a learning rate of 0.01.

[0072] The control command smoothing process uses a second-order Butterworth low-pass filter with a cutoff frequency of 5Hz. The filtered signal is then used to calculate the first and second derivatives using the center difference method.

[0073] The Extreme Learning Machine uses the sigmoid activation function, has 100 hidden layer nodes, and its input includes filtered control instructions and their derivatives. The output is a time delay compensation quantity. The network parameters are updated through an online sequence extreme value learning algorithm, with an initial learning rate of 0.1.

[0074] The final control signal is generated by superimposing the original control command, fault compensation amount and time delay compensation amount, and output to the servo driver of the actuator at a frequency of 50Hz.

[0075] Example 4: To make the technical solution of the present invention clearer, the process of obtaining the original attitude data and the attitude estimation value is described in detail here; specifically: The acquisition module obtains analog output signals from the seven-axis sensors (three-axis accelerometer, three-axis fluxgate, and in some embodiments, angular rate gyroscope). These signals are digitized and transmitted to the main control chip of the control module through a communication interface (such as SPI).

[0076] In the main control chip, the system no longer uses a fixed noise covariance matrix, but adaptively adjusts the process noise covariance matrix based on the real-time acquired drill string data. and measurement noise covariance matrix The value is used to more accurately reflect the noise level of the current environment. Under severe vibration, the system will increase... and The value of makes the filtering process rely more on model prediction, thus avoiding interference from abnormal noise. Subsequently, the system estimates the optimal attitude from the previous time step. The system dynamics model predicts the state at the current moment.

[0077] The specific expression satisfies: ; in, This represents the predicted state at the current moment, i.e., the optimal attitude estimate. For the system at time k The state transition matrix from 1 to k; For the previous time step k The attitude estimate is 1. The input transition matrix of the system; Let k be the control input vector of the system at time k; This refers to system process noise.

[0078] It is understandable that the above expression is the state prediction part of the time update equation of the Kalman filter, which is to predict the state by taking the optimal state from the previous time step. Multiply by the state transition matrix In addition to the influence of external inputs To predict the state at the current moment. The purpose is to provide a prior estimate based on the system dynamic model before new measurement data arrives.

[0079] Subsequently, after the prediction status is established, the system updates the prediction mean squared error matrix. This matrix incorporates adaptive adjustment. Its expression satisfies: ; in, for The state transition matrix, and the transpose matrix are obtained by interchanging the rows and columns of the original matrix; Let be the prior covariance matrix for time step k; For time step k The posterior covariance matrix of 1 represents the best estimate of the system state at the previous time step. The uncertainty is that this value is obtained in the previous filtering cycle by fusing the predicted value and the actual measured value.

[0080] The process noise covariance matrix represents the uncertainty of the system model itself.

[0081] For measurement updates and output: The system acquires sensor measurements and calculates the Kalman gain. Then, it uses the Kalman gain to correct the predicted state and outputs an attitude estimate, which satisfies the expression: ; in, Indicates Kalman gain; The measurement matrix describes how the system state vector is mapped to the measurement vector. It can be understood as converting the internal state (such as well inclination angle and angular velocity) into physical quantities that the sensor can measure. This matrix is ​​preset according to the relationship between the sensor and the system model. This is the transpose of the measurement matrix.

[0082] The expression for obtaining the attitude estimate satisfies: ; The posterior state estimate, i.e. the final output of the Kalman filter, represents the optimal estimate of the system state at time step k. It combines information from both prediction and measurement, and is therefore more accurate than any single source of information. This is the prior state estimate, which is the output of the previous time update and represents the system state based on the model prediction. For measurement vectors; To measure the residual, the difference between the actual measured value and the measured value derived from the prediction model is measured, and this difference is the basis for the filter to make corrections.

[0083] Regarding the attitude angle, the specific steps include: Obtain the pose estimate output from the previous step. ; Update the optimal covariance matrix The expression satisfies: ;in, The identity matrix is ​​used in the expression above. The purpose is to perform matrix subtraction operations, ensuring the mathematical integrity of the formula, by using... minus This yields a correction factor, which will act on the prior covariance matrix. This allows us to obtain a more accurate posterior covariance matrix. This correction process essentially reduces uncertainty in order to obtain more accurate information through new measurement data.

[0084] Using the rotational speed data measured by the angular rate gyroscope and based on the compensation formula, the errors caused by the centrifugal force and tangential force generated by the high-speed rotation of the drill bit are compensated. The compensated signal is calculated as well inclination angle θ and tool face angle. Its expression satisfies: ; The compensation formula satisfies the following expression: ; It also satisfies the expression: ; Understandable These are the raw output signals from the triaxial accelerometer, which include the additional acceleration generated by rotation; The signal value corresponding to the gravitational acceleration measured by the accelerometer in a static state is used as a reference to convert the acceleration value into actual physical units; R is the distance from the accelerometer to the center of the drill bit's rotation. The sensor's installation position determines the magnitude of the centrifugal force and tangential force it senses. ω is the angular velocity of the drill bit measured by the angular rate gyroscope; This is the angular acceleration of the drill bit, i.e., the rate of change of angular velocity. When rotating at a constant speed, this value is zero. These are the compensated signals.

[0085] In this embodiment, for the specific compensation process, during drilling, the angular rate gyroscope continuously and in real time measures the angular velocity ω and angular acceleration of the drill string. Simultaneously, the triaxial accelerometer measures and outputs the raw signal, which includes rotational error. The system will input the real-time acquired angular velocity, angular acceleration, and preset sensor installation parameters (such as distance R) into the compensation formula. For X-axis (tangential) acceleration, the compensation formula will subtract an error term proportional to the angular acceleration; for Y-axis (centrifugal) acceleration, the compensation formula will subtract an error term proportional to the square of the angular velocity; thus obtaining the compensated... After receiving the signal, the system will combine these corrected values ​​with the filtered values. Once the signal is received, the final attitude calculation is performed to determine the precise wellbore inclination angle θ and tool face angle. .

[0086] Example 5: Regarding the process of acquiring control commands: Understandably, not all system states can be directly measured during actual drilling operations. Therefore, a full-dimensional state observer with a built-in control module is designed. This observer uses measurable output data, combined with the system's dynamic model, to estimate the states of all systems in real time and output state estimates. This ensures that the controller can obtain complete state information even when some states are missing.

[0087] Inside the state observer, a radial basis function neural network is embedded. Its main function is to approximate and reconstruct unknown fault signals online. This radial basis function neural network uses the system state estimate output by the state observer. As input, it leverages its powerful nonlinear fitting capability to output real-time reconstructed values ​​of fault dynamics. The expression satisfies: ; in, Let be the estimated value of the fault signal at time k; The weight vectors for the online learning of the radial basis function neural network are updated in real time through an adaptive law; For radial basis function neural networks, use Gaussian activation function vectors. This is the system state estimate output by the state observer.

[0088] After reconstruction using a state observer and a radial basis function neural network, an estimate containing fault dynamics information is obtained. .

[0089] Understandably, in some embodiments, the control module further includes a fault-tolerant controller. This controller, after acquiring an estimated value of the fault signal, generates a control command capable of proactively compensating for the impact of the fault, ensuring continuous and stable tracking of the wellbore trajectory. Specifically, the state estimate value is input to the fault-tolerant controller. and fault signal estimation .

[0090] The core task of a fault-tolerant controller is to adjust control commands in real time based on fault estimation to counteract the effects of faults. The fault-tolerant control law of a fault-tolerant controller... It consists of two parts: a nominal control law and a fault compensation control law .in, The design directly relies on the fault signals reconstructed in the aforementioned steps. This ensures that the system can compensate immediately when a fault occurs, and the process satisfies the expression: ; in, This represents the final control command at time k; This indicates the nominal control law, used for trajectory tracking in the absence of faults; This represents the fault compensation control law, the magnitude and direction of which depend on the fault signal estimated by the radial basis function neural network.

[0091] Example 6: In other embodiments, a fault-tolerant control law based on integral sliding mode can also be used; specifically, an integral sliding surface is calculated. To reduce the observation error of the system The observation error input to the fault-tolerant controller in this embodiment is weighted and summed, and its cumulative value (i.e., integral) varies with the dimensionless distance ξ. With observation error The essence is the state estimate. and fault signal estimation The only difference lies in the symbolic meaning, and their expressions satisfy: ; It is understandable that the above expression defines the integral sliding surface. The current observation error of the system and the accumulation of historical observation errors By combining the gain matrices c1 and c2 and weighting them, the aim is to make the system converge to a preset sliding surface in the state space. Once the system state reaches this surface, it can maintain stable operation on the surface even if there are disturbances, thereby achieving robust control of faults.

[0092] In the formula, Let be the integral sliding surface, which is a vector representing the distance between the system state and the sliding surface; All of these are positive sliding mode gain matrices, which are pre-designed parameters used to adjust the characteristics and convergence speed of the sliding surface. They are derived from constants set by the controller based on system performance indicators and stability requirements. The observation error of the system state can be represented by the aforementioned fault signal estimation. In other embodiments, it can also be the difference between the output of the state observer and the actual system state. The integral of the observation error represents the accumulation of the observation error over time (or dimensionless distance), reflecting the average deviation of the system over a period of time. τ is the integration variable used to traverse the distance from 0 to ξ, derived from the real-time integration of the observation error during operation.

[0093] It is also understandable that, after defining the integral sliding surface, this step will generate a fault-tolerant control law, the purpose of which is to generate a compensating force that forces the system state to converge toward the sliding surface and remain there.

[0094] Fault-tolerant control law of design The purpose is to force the system state to converge toward the sliding surface and remain there. This control law consists of multiple parts, each of which serves a specific control objective.

[0095] The expression for this process satisfies: ; In the formula, This is a fault-tolerant control law, derived from the compensation force calculated by the controller based on the system state, observation error, and fault estimation. G is the control gain matrix, a preset constant matrix used to map the output of the control law to the system input. It is derived from the constants set by the controller according to system requirements.

[0096] B is the system input matrix, which describes how the control input affects the system state. It is derived from the dynamic model of the rotary steering system. The control parameter for the convergence law is a positive constant used to adjust the convergence speed and chattering magnitude. It is a constant set by the controller according to the system performance requirements.

[0097] It is the hyperbolic tangent function, a nonlinear function used to smooth the approach process and reduce chattering; All are system matrices, describing the evolution of the system state under different time delays, and are derived from the dynamic model of the rotary steering system; This refers to the observation error, which includes the error at the current time and the time delay. F is the fault input matrix, which describes how faults affect the system state; The fault signal is reconstructed by the radial basis function neural network, and its source is the output of the aforementioned radial basis function neural network.

[0098] Finally, the system superimposes the fault-tolerant control law with the normal nominal control law to form the final control command.

[0099] Example 7: For command filters, in the control method of rotary steering system, the control commands output from the radial basis function neural network are usually used to drive the actuator. However, in practical applications, since the output of the neural network may have high-frequency chattering, direct use will wear down the actuator and may cause system instability. Therefore, it is necessary to smooth the control commands.

[0100] In the traditional backstepping control design of high-order nonlinear systems, it is necessary to perform stepwise differentiation of the virtual control signal. This process may lead to multimodal data processing problems. That is, as the system order increases, the expression of the control law becomes extremely complex and verbose, making it difficult to implement the controller. In this embodiment, the virtual control signal and its derivative are approximated by a command filter through a dynamic system, thereby avoiding the tedious analytical differentiation process and greatly simplifying the controller design.

[0101] Its expression satisfies: ; In the formula, u This represents the control commands from the radial basis function neural network, i.e., the input signal of the filter; u c1 The state variable represents the filter, which is the control signal after smoothing, i.e., the filtered signal. u c2 This represents another state variable of the filter, which is the derivative of the filtered signal; ω n This represents the natural frequency and is used to adjust the bandwidth of the filter. It is usually designed according to the dynamic response requirements of the system. The larger the frequency, the faster the filter response, but the worse the smoothing effect may be; conversely, the smaller the frequency, the slower the response and the better the smoothing effect.

[0102] ζ represents the damping ratio, which is used to adjust the response characteristics of the filter and prevent overshoot and oscillation.

[0103] Example 8: For extreme learning machines, in rotary guidance systems, actuator response, signal transmission and computation processing may all introduce execution delay. Delay is a nonlinear characteristic of system dynamics that causes the controller's output signal to lag behind the actual needs of the system, resulting in a decrease in control accuracy and, in severe cases, even system oscillation and instability.

[0104] Understandably, the Extreme Learning Machine (ELM) is a type of feedforward neural network. Its significant advantage lies in the fact that its input layer weights and hidden layer biases are randomly generated, eliminating the need for iterative training. The output weights can be obtained in a single step, making its training speed extremely fast and making it very suitable for online control applications.

[0105] Suppose the system has an unknown function. Caused by execution delay, the Extreme Learning Machine can be used to approximate this unknown function, whose expression satisfies: ; In the formula, Let be the unknown function to be approximated, where This represents the system state at time t-τ, reflecting the time delay characteristic; Indicates the extreme learning machine pair The estimated value; L represents the number of hidden layer nodes; This represents the activation function of the hidden layer, such as the Sigmoid function; This represents the weight vector from the input layer to the hidden layer, which is randomly generated and remains constant. The bias of the hidden layer neurons is randomly generated and remains constant. The weights from the hidden layer to the output layer need to be determined through learning. This represents the hidden layer output matrix; This represents the estimated output weights of the Extreme Learning Machine, and is a vector that needs to be adjusted online.

[0106] In other embodiments, in order to ensure that the extreme learning machine can effectively compensate for time delays and guarantee the stability of the entire closed-loop system, an adaptive law needs to be designed to update the weight estimates in real time. The derivation of this adaptive law preferably relies on the Lyapunov-Kravsky functional.

[0107] The final control signal will be a superposition of multiple factors, including the control command smoothed by the command filter and the time delay compensation term generated by the extreme learning machine. This control signal is transmitted to the actuator to drive the rotary guide system, thereby achieving accurate tracking of the desired trajectory while ensuring the stability and robustness of the system.

[0108] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0109] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0110] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a multimedia terminal device (which may be a mobile phone, computer, television receiver, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0111] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A control method for a rotary guide system, characterized in that, The control method includes the following steps: The raw attitude data of the rotary steering system is acquired, the filtering parameters are adjusted using the Kalman filter algorithm, and the attitude estimate is output. Based on the attitude estimation value, the attitude angle of the rotary guide system is output through speed compensation; The radial basis function neural network outputs control commands for attitude angles after fault compensation. The control commands are smoothed, and the filtered signal and derivative of the control commands are obtained. Based on the filtered signal and derivative, the execution time delay is compensated by the extreme learning machine, and the control signal of the rotary guidance system is output.

2. The control method for a rotary guide system as described in claim 1, characterized in that, The process of acquiring the raw attitude data of the rotary steering system, adjusting the filtering parameters using a Kalman filter algorithm, and outputting attitude estimates includes the following steps: The analog output signal of the acquisition module is acquired, digitized, and then transmitted to the control module through the communication interface. Obtain the dynamic model of the rotary steering system, and use the dynamic model to transfer the optimal covariance of the previous moment to the current moment. After adding the process noise covariance, output the updated mean square error matrix. Based on the mean square error matrix, the attitude estimate is output after correcting the predicted state with Kalman gain.

3. The control method for a rotary guide system as described in claim 1, characterized in that, Based on the attitude estimation value, the attitude angle of the rotary guide system is output through speed compensation, including the following steps: After obtaining the attitude estimate and updating the optimal covariance matrix, the attitude is calculated to output the attitude angles of the rotary steering system, which include the well inclination angle and the tool face angle.

4. The control method for a rotary guide system as described in claim 3, characterized in that, The control command for attitude angle after fault compensation, which is output through a radial basis function neural network, includes the following steps: The fault dynamic function is obtained, the attitude estimate is mapped to a high-dimensional space through a nonlinear activation function, and then multiplied with the weight vector to obtain the estimated value of the fault signal. Obtain the observation error, and add the observation error to the integral of the observation error to obtain an integral sliding surface; Obtain the fault-tolerant control law, and then use the compensation force to integrate the convergence value of the attitude estimate of the rotary guide system onto the sliding surface to generate control commands.

5. The control method for a rotary guide system as described in claim 4, characterized in that, The process of obtaining the fault-tolerant control law, and then integrating the convergent value of the attitude estimate of the rotary guide system onto the sliding mode surface to generate control commands by obtaining the compensation force, includes the following steps: Obtain the nominal control law; After detecting and estimating the fault signal using a radial basis function neural network, a fault-tolerant control law is generated. This fault-tolerant control law is used to generate a compensating force to counteract the impact of the fault signal on the rotary steering system. The nominal control law and the fault-tolerant control law are superimposed to output the control command.

6. The control method for a rotary guide system as described in claim 1, characterized in that, The process of compensating for execution time delay using an extreme learning machine based on the filtered signal and derivative, and outputting control signals for the rotary steering system, includes the following steps: Obtain the approximation function of the filtered signal and its derivative under extreme learning, and model the execution delay; Obtain the adaptive law of the extreme learning machine to update the network weights, and obtain compensation signals to offset the execution delay; The control signal is obtained by combining the filtered signal, derivative, and compensation signal, and the control signal is used to drive the rotary guide system.

7. A control system, characterized in that, The control system is used to execute the control method of the rotary guide system according to claim 6, and the control system includes: The acquisition module is used to acquire the original attitude data of the rotary guide system, adjust the filtering parameters through the Kalman filter algorithm, and output the attitude estimate. The compensation module is used to output the attitude angle of the rotary guide system based on the attitude estimation value and through rotational speed compensation. The control module is used to output fault-compensated control commands for attitude angles through a radial basis function neural network; The control module is also used to smooth the control commands and obtain the filtered signal and derivative of the control commands; The control module is also used to compensate for execution delays using an extreme learning machine based on the filtered signal and derivative, and to output control signals for the rotary guidance system.

8. A control system as described in claim 7, characterized in that, The control module includes an observer embedded with a radial basis function neural network, which is used to detect and estimate fault signals.

9. A control system as described in claim 7, characterized in that, The control module also includes a command filter, which is used to acquire the filtered signal and derivative.

10. A control system as described in claim 7 or 9, characterized in that, The control module also includes an extreme learning machine, which is used to obtain the approximation function of the filtered signal and its derivative under extreme learning, and to model the execution delay. The extreme learning machine is also used to obtain an adaptive law to update the network weights and to obtain a compensation signal to offset the execution delay. The extreme learning machine is also used to combine the filtered signal, derivative and compensation signal to obtain a control signal, which is used to drive the rotary guide system.

Images