A deep adaptive drilling rig and adaptive method thereof

Abstract

The present application belongs to the technical field of drilling machines, and in particular to a depth adaptive drilling machine and an adaptive method thereof, comprising a drill string channel physical modeling module, a multi-modal vibration signal acquisition and preprocessing module, a channel-signal collaborative decoupling module, a control sensitive feature directional extraction module, a depth adaptive decision module and an actuator collaborative control module; a drill string channel model is established, downhole and ground vibration signals are synchronously acquired and preprocessed, a depth decoupling network is used to inversely decouple the aliasing signals, and control sensitive features are extracted. By constructing a drill string distributed parameter channel model and embedding it as a differentiable constraint into a depth decoupling network, the depth fusion of physical mechanism and data-driven algorithm is realized, and independent vibration modes close to the real state of the well bottom can be stably and explainably reconstructed from the ground signals with severe modal aliasing, fundamentally solving the signal transmission distortion problem.

Description

Technical Field

[0001] This invention belongs to the field of drilling machine technology, specifically a depth-adaptive drilling rig and its adaptive method. Background Technology

[0002] In the field of intelligent drilling and automated control, the adaptive capability of drilling rigs highly depends on accurate perception of downhole conditions. Drill pipe vibration signals, as a key physical information carrier reflecting formation lithology, drill bit condition, and drilling efficiency, are widely used in closed-loop feedback control. However, in deep or ultra-deep well operations, the long-distance mechanical transmission channel formed by the drill pipe exhibits significant internal material damping, non-uniform cross-sectional characteristics, and complex wellbore boundary constraints. This leads to severe dispersion, attenuation, and modal coupling phenomena—i.e., modal aliasing—as the original vibration signal propagates upwards along the wellbore. This effect causes vibration modes of different frequencies and propagation paths to superimpose and interfere with each other, resulting in signal waveform distortion, characteristic frequency shifts, and even the annihilation of key transient components. Consequently, the vibration data acquired by the surface receiver deviates significantly from the actual excitation state at the bottom of the well.

[0003] The effectiveness of formation identification and adaptive drilling parameter control technology based on vibration signals directly depends on signal fidelity. Existing technologies generally employ post-processing methods such as conventional filtering, time-frequency analysis, or empirical mode decomposition to attempt to extract useful information from distorted signals. However, these methods are essentially passive compensation strategies, unable to decouple the already aliased multimodal components, and struggling to reconstruct the subtle geological response characteristics lost due to high-frequency attenuation. More importantly, traditional control architectures separate signal acquisition from control decision-making, failing to establish a collaborative mechanism between the physical model of vibration wave propagation in the drill string medium and the adaptive algorithm. This results in adjustment commands made by the drilling rig based on distorted information not only being ineffective but potentially exacerbating drilling rig instability.

[0004] Existing technologies lack dynamic modeling of drill pipe as a distributed parameter drilling rig and fail to incorporate material properties, geometric configurations, and wellbore constraints into the prior knowledge system of the signal evolution process; Furthermore, the signal recovery algorithm is not aligned with the control objective, but only pursues local optimization in the time or frequency domain, neglecting the targeted enhancement of control-sensitive features (such as impact events and resonant frequency drift). Secondly, adaptive control logic is still based on static mapping or low-order models, and cannot make forward-looking interventions using potential physically interpretable features.

[0005] Therefore, the present invention provides a depth-adaptive drilling rig and its adaptive method to solve the technical problems mentioned above. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a depth-adaptive drilling rig and its adaptive method, which can effectively solve the problems in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a depth-adaptive drilling rig, comprising the following components: The drill string channel physical modeling module constructs a distributed parameter drilling rig model characterizing the propagation characteristics of vibration waves in the drill string based on the material properties, geometric configuration, and wellbore constraint parameters of the drill pipe. This model describes the signal transmission function from the bottom of the well to the surface. The multimodal vibration signal acquisition and preprocessing module, deployed near the drill bit end and at the surface receiver, synchronously acquires the original downhole vibration signal and the aliased vibration signal received on the surface, and performs synchronous alignment and noise reduction preprocessing on the signals. The channel-signal collaborative decoupling module is communicatively connected to the drill string channel physical modeling module and the multimodal vibration signal acquisition and preprocessing module. It incorporates a deep decoupling network guided by a physical model. This network uses the transmission function as a priori constraint to perform inverse decoupling of the aliased signals on the surface, separating and reconstructing the signals corresponding to different transmission functions. The system comprises: independent vibration modes of the drilling path and frequency components; a control-sensitive feature directional extraction module, connected to the channel-signal collaborative decoupling module, for directional extraction of physical features strongly correlated with drilling control decisions from the reconstructed independent vibration modes, including impact event energy, main resonant frequency drift, and mode energy ratio; a depth adaptive decision module, connected to the control-sensitive feature directional extraction module, which has a built-in reinforcement learning controller trained based on physical features and historical drilling performance data, for generating optimized control commands to dynamically adjust drill pressure, rotation speed, and tool face angle; and an actuator collaborative control module, connected to the depth adaptive decision module, for receiving the optimized control commands and collaboratively driving the feed rig, rotary rig, and guide tool to achieve real-time and collaborative adjustment of drilling parameters.

[0008] Preferably, the distributed parameter drilling rig model constructed by the drill string channel physical modeling module is based on a wave equation considering material damping, cross-sectional non-uniformity, and wellbore boundary effects. This model discretizes the drill string into a series of elements with mass, stiffness, and damping. The dynamic characteristics of each element are defined by material density ρ, elastic modulus E, Poisson's ratio ν, cross-sectional moment of inertia I, outer diameter D, and inner diameter d. Wellbore constraints are modeled by introducing equivalent spring-damped boundary conditions related to the drill string element positions, with stiffness coefficients... and damping coefficient The model is parameterized based on formation lithology, well diameter, and drilling fluid properties. Its output is the frequency domain transfer function H(ω) from the drill bit position to the surface sensor position, expressed as: in, and These are the frequency domain representations of the vibration signals at the ground and the bottom of the well, respectively. Let be the transfer matrix for the i-th drill string element or joint. The attenuation coefficient is frequency-dependent, L is the total well depth, and N is the total number of discrete elements. The transfer function H(ω) is... It encapsulates the dispersion, attenuation, and mode coupling effects of signals, serving as a mathematical description of channel characteristics.

[0009] Preferably, the multimodal vibration signal acquisition and preprocessing module includes a downhole near-bit measurement unit and a surface vibration monitoring unit. The downhole near-bit measurement unit is integrated within the drill collar or near-bit stabilizer and includes a triaxial accelerometer and a dynamic pressure sensor. Its sampling frequency is not less than 2kHz, used to acquire the original downhole vibration signal as close as possible to the actual excitation source. The ground vibration monitoring unit is installed in the top drive or turntable bearing housing and also includes a triaxial accelerometer for acquiring ground aliasing vibration signals transmitted through the drill string channel. The module incorporates a high-precision clock synchronization mechanism to ensure that the timestamp alignment error between downhole and surface data is less than 1 millisecond. The preprocessing process includes low-pass filtering (cutoff frequency 1kHz) of the raw signal to combat aliasing and using wavelet thresholding to suppress random noise.

[0010] Furthermore, the channel-signal collaborative decoupling module uses a physical model-guided deep decoupling network with an encoder-decoder architecture and introduces a physical constraint layer. The encoder receives the pre-processed ground aliasing signal. Using its corresponding short-time Fourier transform spectrum as input, the decoder's target output is K reconstructed independent modal signals { , , ..., The physical constraint layer will use the transfer function H(ω) calculated by the drill string channel physical modeling module. The inverse H^{-1}(ω) As a differentiable layered embedded network, during network training and inference, this constraint layer enforces that the weighted sum of all remodeled states be processed through the transfer function H(ω). After forward propagation, the signal should approximate the input ground aliasing signal as closely as possible, and its loss function L should include data reconstruction loss. and physical consistency loss : Where λ is the weighting coefficient. These represent the Fourier transform and inverse transform, respectively. This design allows the network to learn data-driven learning while strictly adhering to the physical laws of vibration wave propagation, thus enabling stable and interpretable separation of independent modes from severely aliased signals.

[0011] Preferably, the control-sensitive feature orientation extraction module is configured for each reconstructed independent modality. The system performs three types of feature extraction operations in parallel. The first type is impact event detection: calculating the envelope of the modal signal and setting a dynamic threshold. When the envelope value exceeds the threshold, it is marked as an impact event, and the time, energy, and duration of the event are calculated. The second type is main resonance frequency tracking: performing continuous spectrum estimation on the modal signal and tracking the frequency corresponding to its power spectral density peak. And calculate its drift relative to the baseline frequency. The third category is modal energy ratio calculation: calculating the energy of each modal signal within a sliding time window. And calculate its proportion of the total energy. The impact energy and frequency shift. and energy ratio The feature vector F(t) at the current time step It is fed into the deep adaptive decision-making module in real time.

[0012] Preferably, the reinforcement learning controller in the deep adaptive decision-making module has a state space The feature vector F(t) at the current time The action space is composed of a sequence of features from several historical moments and the current drilling parameters. Defined as the adjustment amounts (ΔW, ΔR, Δθ) to drilling pressure, rotational speed, and tool face angle, the reward function... The design combines drilling efficiency with equipment protection: , in Instantaneous mechanical drilling speed, For vibration intensity, to The weights are used to train the controller online or offline using a proximal policy optimization algorithm. Its policy network... Based on the current state Output optimal adjustment action The probability distribution is used to sample or select the action with the highest expected value based on this distribution, and generate specific optimized control instructions.

[0013] Preferably, an adaptive method for a depth-adaptive drilling rig includes the following specific steps: Step S110: Based on the drill pipe and wellbore parameters, establish a distributed parameter channel model that characterizes the physical laws of vibration wave propagation along the drill string, and obtain the channel transfer function; Step S120: Simultaneously acquire the original vibration signal near the drill bit end downhole and the superimposed vibration signal from the ground receiving end, and perform time alignment and preprocessing. Step S130: Using the channel transfer function as a physical constraint, a deep decoupling network is used to reverse decouple the ground aliased vibration signal and reconstruct multiple independent vibration mode signals. Step S140: From the reconstructed independent vibration mode signals, extract three types of control-sensitive features: impact event energy, main resonance frequency drift, and mode energy ratio. Step S150: Using the control-sensitive features and current drilling parameters as state inputs, a reinforcement learning controller is used to generate coordinated adjustment commands for drilling pressure, rotation speed, and tool face angle. Step S160: Execute the coordinated adjustment command to drive the feed, rotation, and guidance drilling rig, and complete the adaptive optimization of drilling parameters.

[0014] Preferably, in step S130, the training process of the deep decoupling network adopts a two-stage strategy. The first stage is pre-training: using the distributed parameter channel model, a large number of simulated bottom hole mode signal combinations and their corresponding surface aliasing signals are generated as training datasets to perform preliminary training on the network, enabling the network to learn the decoupling mapping relationship under ideal physical constraints. The second stage is online fine-tuning: during the actual drilling process, the signal segments obtained by the downhole near-bit measurement unit are used as approximations of the quasi-real modes, and compared with the network's results of decoupling the surface signals for the corresponding time period. The difference loss is calculated, and the network parameters are updated in small batches using this loss, enabling the network to adapt to small deviations between the actual channel and the model.

[0015] Preferably, in step S150, the online learning mechanism of the reinforcement learning controller includes: during each action execution... And observed a new state and rewards Then, this empirical tuple ( , , , The data is stored in an experience replay buffer. The controller periodically samples a batch of experience data from the buffer to update its policy network and value network parameters. At the same time, an independent explore-utilization scheduler is set up to reduce the exploration rate when drilling conditions are stable, focusing on utilizing learned policies. When a drastic change in characteristics is detected (such as entering a new formation), the exploration rate is temporarily increased to encourage the controller to try new actions to quickly adapt to the new environment.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By constructing a drill string distributed parameter channel model and embedding it as a differentiable constraint into a deep decoupled network, a deep integration of physical mechanisms and data-driven algorithms is achieved. This enables the stable and interpretable inverse reconstruction of independent vibration modes that closely approximate the actual state at the bottom of the well from surface signals with severe modal aliasing, fundamentally solving the problem of signal transmission distortion. Through a targeted extraction mechanism of control-sensitive features, subsequent control decisions are directly focused on the physical indicators that have the most critical impact on drilling stability and efficiency, improving the targeting of information utilization and the effectiveness of decision-making. By adopting an adaptive decision-making architecture based on reinforcement learning, combining physical features, drilling parameters, and multi-dimensional rewards, the optimal control strategy under complex formation conditions can be automatically discovered through online learning, achieving a leap from control based on static models or rules to intelligent control based on dynamic environmental interaction and long-term benefit optimization. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic diagram of the architecture of the depth-adaptive drilling rig proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the deep decoupling network guided by the physical model in this invention; Figure 3 This is a logical flowchart of the drill string channel physical modeling and multimodal vibration signal acquisition in this invention; Figure 4 This is a logical flowchart of the channel-signal collaborative decoupling and feature-oriented extraction in this invention; Figure 5 This is a schematic diagram illustrating the principle framework of the deep adaptive decision-making module in this invention for generating optimized control instructions; Figure 6 This is a schematic diagram illustrating the principle of the deep collaborative architecture of the entire link in this invention, which includes physical channel modeling, signal collaborative decoupling, feature-oriented extraction, and intelligent decision optimization. Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0020] Example 1: In deep oil and gas exploration or geothermal drilling operations, the drill string serves as the sole mechanical and information channel connecting surface equipment and downhole drilling tools, and its dynamic behavior is extremely complex. During drilling, the rich vibration information generated by the interaction between the drill bit and the rock (such as impact, torsional vibration, and longitudinal vibration) carries key physical characteristics of formation lithology, drill bit condition, and drilling efficiency; However, these vibration signals originating from the bottom of the well undergo severe distortion, attenuation, dispersion, and multimodal aliasing as they propagate upwards along the drill string, which can extend for thousands of meters, to the surface sensors. Traditional methods directly analyze the aliased vibration signals acquired from the surface, attempting to extract effective features for control decisions. However, because the channel transmission effects cannot be reversed, the obtained features are severely distorted, making the decision-making basis unreliable and hindering the achievement of truly depth-adaptive drilling.

[0021] See Figure 1 This drilling system forms a complete closed loop from physical perception to intelligent decision-making and execution. Its core lies in constructing a deep collaborative architecture across the entire link, encompassing physical channel modeling, signal collaborative decoupling, feature-oriented extraction, and intelligent decision optimization. Specifically, it includes a drill string channel physical modeling module, a multimodal vibration signal acquisition and preprocessing module, a channel-signal collaborative decoupling module, a control-sensitive feature directional extraction module, a depth adaptive decision-making module, and an actuator collaborative control module. Each module is interconnected via a high-speed real-time industrial network and data bus to ensure that the timeliness of data transmission and command execution meets the control requirements of complex dynamic processes in downhole drilling.

[0022] The drill string channel physics modeling module is the builder of the drilling rig's physical foundation. Its core task is to establish a mathematical model that accurately characterizes the propagation characteristics of vibration waves within the drill string-wellbore drilling rig—a distributed parameter drilling rig model. Model construction begins with the refined parameter input of the drill string and drilling rig. These parameters include the drill string's material properties: density ρ, elastic modulus E, Poisson's ratio ν; geometric configuration: the outer diameter D, inner diameter d, and length of each drill string; and the corresponding parameters for all drill string assemblies (such as drill collars, stabilizers, and slammers). In addition, wellbore constraint parameters are crucial, including well diameter, formation elastic modulus, and drilling fluid density and rheological properties obtained from logging-while-drilling or adjacent well data.

[0023] In this embodiment, based on these parameters, the module discretizes the entire drill string rig into N cascaded dynamic elements using the finite element method or transfer matrix method. Each element is considered as a beam element with concentrated mass, torsional / longitudinal / bending stiffness, and structural damping. The wellbore constraint on the drill string is modeled as an equivalent spring-damped boundary condition acting on each element node. The stiffness coefficient of this boundary condition is... and damping coefficient It is not a fixed value, but a dynamic calculation based on the lithology of the formation at the depth of the unit (such as mudstone, sandstone, granite) and the annular space between the well diameter and the outer diameter of the drill string at that location; For example, in well sections with regular diameters and hard formations, The value is relatively large; in easily collapsible strata or well sections with large indentations, The value is relatively small. This reflects the viscous damping effect of the drilling fluid.

[0024] As mentioned earlier, by solving the dynamic equations of this series of units under boundary conditions, the module ultimately calculates the frequency domain transfer function H(ω) from the drill bit position (excitation point) to the ground sensor installation position (receiving point). This function is a complete mathematical description of the channel characteristics, and its expression encapsulates all the physical effects of signal propagation: in, and These are the frequency domain representations of the vibration signals at the ground and the bottom of the well, respectively. Let be the transfer matrix for the i-th drill string unit or joint. It is a frequency-dependent complex attenuation coefficient that integrates energy losses caused by internal material friction (hysteretic damping), drilling fluid viscous damping, and wellbore friction. L is the total well depth. The transfer function H(ω) As prior physical knowledge, it is output and solidified to guide subsequent signal decoupling processes. During drilling, as well depth increases or drill string assemblies change, this model can be triggered for online updates to maintain the accuracy of the channel description.

[0025] The multimodal vibration signal acquisition and preprocessing module is responsible for acquiring the drilling rig's input data. This module employs a dual-end observation strategy of simultaneous downhole and surface acquisition. See Figure 3 The near-bit measurement unit (NB-IoT) is crucial for acquiring source signals and is typically integrated within the drill collar or near-bit stabilizer, located 5 to 10 meters from the drill bit. This unit includes a high dynamic range triaxial accelerometer (measuring longitudinal, lateral, and axial vibrations) and a dynamic pressure sensor (monitoring bottomhole pressure fluctuations). Its sampling frequency is configured to be no less than 2 kHz to ensure the capture of high-frequency components (typically hundreds of hertz) generated by drill bit rock breaking. This unit transmits the compressed raw vibration signal via high-speed mud pulses or electromagnetic waves from the measurement-while-drilling rig. Data can be uploaded to the surface in real time or stored in a downhole storage device for retrieval after drilling is completed.

[0026] It should be noted that the ground vibration monitoring unit is installed on key nodes in the vibration transmission path, such as the top drive or the turntable bearing housing. It is also equipped with a triaxial accelerometer to collect ground aliased vibration signals after the drill string channel has been contaminated. To ensure accurate comparison and alignment of downhole and surface signals, the entire data acquisition rig employs high-precision cesium atomic clocks or GPS synchronization technology for global time synchronization, ensuring that the timestamp synchronization error between any two acquisition points is strictly less than 1 millisecond.

[0027] The acquired raw signal contains a large amount of noise. The preprocessing procedure first performs anti-aliasing low-pass filtering on the signal, with the cutoff frequency set to 1kHz to retain the main vibration frequency band; Subsequently, a threshold denoising method based on wavelet transform is employed. Specifically, the signal is decomposed into multi-scale wavelet components. Based on the statistical characteristics of the wavelet coefficients at each scale (e.g., using BayesShrink thresholding), the small coefficients representing random noise are set to zero, and then wavelet reconstruction is performed. This effectively suppresses background noise while preserving the signal's abrupt change characteristics (such as shocks). The preprocessed downhole signal... With ground signals It is sent to the downstream module.

[0028] In this embodiment, the channel-signal cooperative decoupling module is tasked with decoupling severely distorted ground aliased signals. In the process, multiple independent vibration mode signals that closely approximate the actual state at the bottom of the well are separated and reconstructed in reverse. See Figure 2 and Figure 4 The core of the channel-signal collaborative decoupling module is a deep decoupling network guided by a physical model. This network adopts an encoder-decoder architecture. The encoder receives the pre-processed ground aliasing signal. The short-time Fourier transform spectrum of the signal is used as the dual-channel input, and high-dimensional abstract features of the signal are extracted through multi-layer convolution and attention mechanisms. The decoder aims to reconstruct K independent time-domain modal signals based on these abstract features. , , ..., The value of K can be determined based on engineering priors, and typically includes the longitudinal vibration principal mode, the torsional vibration principal mode, and several higher-order bending modes.

[0029] The key to the network lies in the introduction of a differentiable physical constraint layer. This layer uses the frequency domain transfer function calculated by the drill string channel physical modeling module. and its reverse The forward and backward propagation processes are embedded in matrix operations. During network training and inference, physical constraints mandate that all modal signals reconstructed by the decoders { } are superimposed, and then passed through a forward channel model The transmission must be performed, and the result obtained must be compared with the input ground aliasing signal. As consistent as possible. This constraint is expressed through the physical consistency loss function. To achieve: in, These represent the Fourier transform and inverse Fourier transform, respectively. The total loss function L of the network is the data reconstruction loss. (Measures of the smoothness, sparsity, etc. of the reconstructed signal itself) and physical consistency loss Weighted sum: L = + λ· λ is a trade-off hyperparameter.

[0030] This design transforms the network's learning process from a purely black-box data fitting into a process strictly guided by physical laws. The network is forced to learn how to decompose a ground signal into several components, which, after being transmitted through a known physical channel, can be precisely synthesized into the original signal. Therefore, the decoder output It possesses clear physical interpretability and can be considered as an approximate bottom-hole source signal corresponding to different vibration modes (such as drill bit longitudinal impact and drill string torsional oscillation). The network is trained using a two-stage strategy: The first stage is offline pre-training: using the drill string channel physical modeling module, a large number of simulated bottom hole modal signal combinations are randomly generated, and then... The corresponding ground aliasing signals are synthesized to form a large-scale simulation dataset, and the network is fully pre-trained. The second stage is online fine-tuning: during actual drilling, the data obtained from the downhole near-bit measurement unit is used... As a reference for the near-real mode (although it also has local channel effects), it is compared with the main mode decoupled from the ground signal in the same time period by the network, the difference loss is calculated, and the network parameters are updated in small batches online, so that the network can adapt to the small deviation between the actual channel and the theoretical model.

[0031] The control-sensitive feature directional extraction module receives the independent vibration mode signals reconstructed by the channel-signal cooperative decoupling module. This module abandons traditional full-band spectrum analysis or broad time-domain statistics, instead focusing on extracting three core physical features strongly correlated with drilling control decisions. See also... Figure 4 For each mode The module performs the following operations in parallel: As mentioned earlier, impact event detection and energy calculation are performed. First, the Teager-Kaiser energy operator envelope of the signal is calculated to enhance the transient impact component. A threshold is set, dynamically adjusted based on the root mean square value of the signal within a sliding window. When the envelope value exceeds the threshold, it is marked as an impact event. The start time, peak amplitude, and duration of the event are recorded, and the total signal energy within the event window is calculated as the impact energy. Continuous impact energy sequences reflect the severity of drill bit rock breaking and the heterogeneity of the formation. Following the above, the main resonant frequency is tracked and the drift is calculated. For the modal signal... Perform continuous complex modulation and demodulation or sliding window Welch spectral estimation to track the frequency corresponding to the peak power spectral density in real time. Set a baseline frequency for this mode. (This can be obtained statistically during the stable drilling phase). Frequency drift. Frequency drift is usually associated with changes in the equivalent stiffness of the drill string and may indicate drill bit wear, changes in wellbore contact conditions, or formation lithology transformation. Next, the modal energy ratio is calculated. Within a sliding time window of length T (e.g., 2 seconds), the energy of each modal signal is calculated. Then, the proportion of energy of each mode to the total energy (the sum of the energies of all K modes) is calculated: The modal energy ratio reflects the contribution distribution of different vibration modes in the total vibration energy. Its change can indicate the shift of the dominant vibration type. For example, a change from longitudinal vibration dominance to torsional vibration dominance may mean the occurrence of stick-slip phenomenon.

[0032] The above three types of characteristic impact energy sequences and frequency shifts of each mode. Energy ratio of each mode At each processing time, they are combined into a multidimensional feature vector F(t). The vector F(t) The physical quantities that have a direct indicative significance for control are fed into the deep adaptive decision-making module in real time.

[0033] See Figure 5 The depth adaptive decision-making module is the intelligent control center of the drilling rig. Its core is a controller trained using a reinforcement learning algorithm, designed according to a Markov decision process framework. Its state space... It is a high-dimensional vector, consisting of the control-sensitive feature vector F(t) at the current time step. The feature sequence of several past moments (to capture dynamic trends), and the current drilling operation parameters (drilling pressure). Rotation speed Tool face angle Together, they constitute the action space. Defined as the adjustment amount for these three operating parameters, i.e. = (ΔW, ΔR, Δθ Each adjustment value takes a value within a continuous range.

[0034] The goal of the controller is to learn the optimal policy. This maximizes long-term cumulative rewards. Reward function The design integrates drilling efficiency, equipment protection, and operational safety, and is key to driving learning. Its specific form is as follows: .

[0035] in, The instantaneous mechanical drilling speed is measured by the drilling footage sensor; Ground vibration intensity (e.g., root mean square value of acceleration); This is the sum of the energies of all shock events detected within the current decision-making cycle; It represents the absolute value of the frequency drift of a key mode (such as the torsional frequency). to The weighting coefficient is used to balance efficiency rewards, vibration reduction rewards, shock suppression penalties, and frequency stability penalties. For example, increasing the mechanical drilling speed may result in a positive reward, but excessive vibration or strong impact may result in a negative reward (penalty). The controller must learn to find a balance between conflicting objectives.

[0036] The controller is trained using a proximal policy optimization method for this type of robust reinforcement learning algorithm. Training can be performed on two levels: As mentioned earlier, a basic policy network is obtained by conducting extensive offline training using historical drilling data and a simulation environment. Following the above, online learning and fine-tuning are performed during actual drilling. The online learning mechanism includes experience playback and exploration using scheduling. Each time, the controller adjusts its settings based on the state. Select and execute an action Afterwards, the drilling rig entered a new phase. And the reward r_t was observed, the empirical tuple ( , , , The experience is stored in a recurring experience replay buffer. The controller periodically samples a batch of experiences randomly from the buffer to update the parameters of its policy network and value network, thereby continuously optimizing the policy. Simultaneously, an independent scheduler monitors the feature change rate. When the drilling rig is in a stable drilling phase, the exploration rate is reduced, primarily based on learned strategies (utilization); when the feature vector F(t)... When drastic changes occur, indicating potential entry into a new geological stratum or an anomaly, the exploration rate is temporarily increased to encourage the controller to try new action combinations (exploration) to quickly adapt to the new environment and discover better strategies. Ultimately, the controller determines the optimal strategy based on the current state. With the optimal strategy π, specific optimized control commands are output, such as slowly increasing the drilling pressure by 20kN and reducing the rotation speed by 5rpm over the next 5 seconds.

[0037] The actuator coordination control module is the physical execution layer of control commands. This module receives digitally optimized control commands from the depth adaptive decision module. These commands include target values ​​or change trajectories for drill pressure, rotational speed, and tool face angle, as well as execution time requirements. The module contains three parallel precision servo control loops. For drill pressure adjustment, the command is converted into the valve core position setting curve of the feed hydraulic rig servo valve. Through a high-response electro-hydraulic proportional valve or servo valve, the flow rate into the feed hydraulic cylinder is precisely controlled, achieving smooth and rapid adjustment of drill pressure. For rotational speed adjustment, the command is converted into the torque or speed setpoint of the top drive or rotary table drive motor, achieving stepless speed regulation through a high-performance frequency converter. For tool face angle adjustment (in directional drilling), the command is converted into internal control commands for the guiding tool (such as a rotary steerable drill), adjusting the direction of the push force or the amount of pad extension. The control actions of these three loops require a high degree of coordination. The module performs decoupling calculations and cross-compensation to ensure that secondary oscillations are not caused by parameter coupling during the adjustment process. The actual feedback values ​​of all actuators (such as actual drilling pressure and actual rotation speed) are collected in real time and sent back to the depth adaptive decision module to form a closed-loop control, and are also used as part of the state information for the reinforcement learning controller.

[0038] See Figure 6 The entire drilling rig operates continuously throughout the drilling process, forming a closed-loop system that deeply integrates physical and intelligent technologies. The drill string channel physical modeling module provides a physical map of vibration propagation. The multimodal vibration signal acquisition and preprocessing module acquires the raw, fuzzy observations. The channel-signal collaborative decoupling module acts as a physically guided translator, using the physical map to reverse-translate the fuzzy observations into clear bottom-hole source information. The control-sensitive feature directional extraction module extracts key control semantics from the source information. The depth-adaptive decision-making module, acting as the intelligent brain, understands these semantics and, combined with historical experience, makes optimal control decisions. The actuator collaborative control module ultimately translates the decisions into control actions. This chain enables the drilling rig to penetrate channel interference across thousands of meters of drill string, directly perceive the mechanical state at the bottom of the well, and perform forward-looking, adaptive intelligent control based on this, thereby achieving comprehensive optimization of drilling efficiency, drill string safety, and wellbore quality in complex deep wells.

[0039] Example 2 When this drilling rig is used for drilling operations in ultra-deep wells or horizontal wells with extended reach, the complexity of the channel model and the dimensionality of the control decisions need to be further enhanced. In ultra-deep wells, the drill string is longer, signal attenuation and dispersion effects are more significant, and there may be multiple resonant modes densely coupled. At this point, the drill string channel physical modeling module needs to employ a higher-order element discretization method and a more refined damping model, especially considering the impact of minute changes in drill string material properties on wave velocity under high temperature and high pressure conditions. Transfer function The calculation may require the introduction of iterative algorithms to handle the multiple reflection effects of waves at the variable diameter and variable stiffness drill bit assembly.

[0040] In this embodiment, for horizontal wells with large displacement, the wellbore trajectory is curved, and the contact friction between the drill string and the wellbore becomes one of the main factors affecting vibration propagation. In the channel model, the nonlinearity of frictional contact needs to be incorporated into the boundary conditions. This can be achieved by introducing Coulomb friction elements at the nodes where the drill string and wellbore may contact each other. The tangential stiffness of these elements is related to the normal contact force, making the transfer function... To some extent, it depends on the current axial load and torque conditions, making it a weakly nonlinear time-varying drilling rig. The channel-signal coordinating decoupling module needs to be able to adapt to this time-varying nature, and its online fine-tuning phase requires more frequent calibration using downhole reference signals.

[0041] It should be noted that, in terms of feature extraction, features related to drill string buckling and self-locking become particularly important for horizontal drilling. The control-sensitive feature directional extraction module needs to add monitoring of specific low-frequency modes (such as frequencies related to drill string sinusoidal or helical buckling) and extract their amplitude growth trends as features. Simultaneously, due to increased frictional torque, the features of torsional vibration (such as the frequency and amplitude of the stick-slip cycle) require more refined quantification.

[0042] The action space and reward function of the deep adaptive decision-making module also need to be expanded. The action space may need to include adjustments to drilling fluid flow rate or performance parameters to help reduce friction. The reward function needs to incorporate considerations of wellbore trajectory smoothness (such as dogleg severity) and penalty terms for pressure buildup and friction torque. Training the reinforcement learning controller faces a larger state-action space, which may require the use of hierarchical reinforcement learning or an option framework to decompose the complex control task into different levels (such as macro-level well section strategy and micro-level parameter adjustment) to improve learning efficiency and strategy stability.

[0043] In horizontal wells, the actuator coordination control module needs to be deeply integrated with advanced rotary steerable drilling rigs. Tool face angle control is no longer simply about tracking setpoints; it requires dynamic adjustments based on geological steering commands and real-time vibration characteristics to achieve integrated geological-engineering optimization. The overall rig coordination requirements are higher, necessitating the establishment of a unified real-time data platform to ensure seamless integration of physical models, signal data, feature vectors, and control.

[0044] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A depth-adaptive drilling rig, characterized in that, include: The drill string channel physical modeling module constructs a distributed parameter drilling rig model that characterizes the propagation characteristics of vibration waves in the drill string based on the material properties, geometric configuration, and wellbore constraint parameters of the drill pipe. This model is used to describe the transmission function of signals from the bottom of the well to the surface. The multimodal vibration signal acquisition and preprocessing module is deployed at the drill string near the drill bit end and the ground receiving end. It is used to synchronously acquire the raw downhole vibration signal and the superimposed vibration signal received on the ground, and to perform synchronous alignment and noise reduction preprocessing on the signal. The channel-signal collaborative decoupling module is communicatively connected to the drill string channel physical modeling module and the multimodal vibration signal acquisition and preprocessing module. It has a built-in deep decoupling network guided by the physical model. This network uses the transfer function as a priori constraint to perform inverse decoupling of the ground aliasing signal, and separates and reconstructs independent vibration modes corresponding to different propagation paths and frequency components. A control-sensitive feature directional extraction module, which is connected to the channel-signal cooperative decoupling module, is used to directionally extract physical features that are strongly correlated with drilling control decisions from the reconstructed independent vibration modes. The features include impact event energy, main resonance frequency drift, and mode energy ratio. The depth adaptive decision module, which is connected to the control sensitive feature orientation extraction module, has a built-in reinforcement learning controller trained based on physical features and historical drilling performance data, used to generate optimized control commands that dynamically adjust drilling pressure, rotation speed and tool face angle. The actuator collaborative control module, which is connected to the depth adaptive decision module, is used to receive the optimized control command and collaboratively drive the feed drilling rig, rotary drilling rig and guiding tool to realize real-time and collaborative adjustment of drilling parameters.

2. The depth-adaptive drilling rig according to claim 1, characterized in that, The distributed parameter drilling rig model constructed by the drill string channel physical modeling module is based on a wave equation considering material damping, cross-sectional non-uniformity, and wellbore boundary effects. This model discretizes the drill string into a series of elements with mass, stiffness, and damping. The dynamic characteristics of each element are defined by material density ρ, elastic modulus E, Poisson's ratio ν, cross-sectional moment of inertia I, outer diameter D, and inner diameter d. Wellbore constraints are modeled by introducing equivalent spring-damped boundary conditions related to the drill string element positions, with stiffness coefficients... and damping coefficient The model is parameterized based on formation lithology, well diameter, and drilling fluid properties. Its output is the frequency domain transfer function H(ω) from the drill bit position to the surface sensor position, expressed as: in, and These are the frequency domain representations of the vibration signals at the ground and the bottom of the well, respectively. Let be the transfer matrix for the i-th drill string element or joint. The attenuation coefficient is frequency-dependent, L is the total well depth, and N is the total number of discrete units.

3. The depth-adaptive drilling rig according to claim 1, characterized in that, The channel-signal collaborative decoupling module uses a physical model-guided deep decoupling network with an encoder-decoder architecture and a physical constraint layer. The encoder receives the pre-processed ground aliasing signal. Using its corresponding short-time Fourier transform spectrum as input, the decoder's target output is K reconstructed independent modal signals { , , ..., The physical constraint layer will use the transfer function H(ω) calculated by the drill string channel physical modeling module. The inverse H^{-1}(ω) As a differentiable layered embedded network, during network training and inference, this constraint layer enforces that the weighted sum of all remodeled states be processed through the transfer function H(ω). After forward propagation, the loss function L approximates the input ground aliased signal, and includes data reconstruction loss. and physical consistency loss : Where λ is the weighting coefficient. These represent the Fourier transform and the inverse transform, respectively.

4. The depth-adaptive drilling rig according to claim 1, characterized in that, The control-sensitive feature orientation extraction module targets each reconstructed independent modality. It performs three types of feature extraction operations in parallel. The first type is impact event detection: calculate the envelope of the modal signal and set a dynamic threshold. When the envelope value exceeds the threshold, it is marked as an impact event, and the time, energy and duration of the event are calculated. The second type is main resonance frequency tracking: continuous spectrum estimation of the modal signal is performed to track the frequency corresponding to the peak power spectral density. And calculate its drift relative to the baseline frequency. ; The third category is modal energy ratio calculation: calculating the energy of each modal signal within a sliding time window. And calculate its proportion of the total energy. .

5. A depth-adaptive drilling rig according to claim 1, characterized in that, The reinforcement learning controller in the deep adaptive decision-making module has a state space The feature vector F(t) at the current time The action space is composed of a sequence of features from several historical moments and the current drilling parameters. Defined as the adjustment amounts (ΔW, ΔR, Δθ) to drilling pressure, rotational speed, and tool face angle, the reward function... It combines drilling efficiency with equipment protection: , in Instantaneous mechanical drilling speed, For vibration intensity, to The weights are used to train the controller online or offline using a proximal policy optimization algorithm. Its policy network... Based on the current state Output optimal adjustment action The probability distribution.

6. A depth-adaptive drilling rig according to claim 5, characterized in that, The reinforcement learning controller employs an online learning mechanism, which includes: In each action And observed a new state and rewards Then, this empirical tuple ( , , , The data is stored in an experience replay buffer; the controller periodically samples a batch of experience data from the buffer to update its policy network and value network parameters; at the same time, an independent explorer is set up to reduce the explorer rate when drilling conditions are stable and temporarily increase the explorer rate when drastic changes in features are detected.

7. An adaptive method for a depth-adaptive drilling rig, applicable to a depth-adaptive drilling rig as described in any one of claims 1-6, characterized in that, The method includes the following: S110: Based on drill pipe and wellbore parameters, establish a distributed parameter channel model characterizing the physical laws of vibration wave propagation along the drill string, and obtain the channel transfer function; S120: Synchronously acquire the raw vibration signal near the drill bit end downhole and the superimposed vibration signal at the ground receiving end, and perform time alignment and preprocessing. S130: Using the channel transfer function as a physical constraint, a deep decoupling network is used to reverse decouple the ground aliased vibration signal and reconstruct multiple independent vibration mode signals. S140: From the reconstructed independent vibration mode signals, extract three types of control-sensitive features: impact event energy, main resonance frequency drift, and mode energy ratio. S150: Using the control-sensitive features and current drilling parameters as state inputs, a reinforcement learning controller is used to generate coordinated adjustment commands for drilling pressure, rotational speed, and tool face angle. S160: Execute the aforementioned coordinated adjustment command to drive the feed, rotation, and guidance drilling rig, thereby completing the adaptive optimization of drilling parameters.

8. The adaptive method for a depth-adaptive drilling rig according to claim 7, characterized in that, In S130, the training process of the deep decoupled network adopts a two-stage strategy, the first stage being pre-training: Using the distributed parameter channel model, a large number of simulated bottom hole mode signal combinations and their corresponding surface aliasing signals are generated as training datasets to perform preliminary training on the network. The second stage is online fine-tuning: During the actual drilling process, the signal segments obtained by the downhole near-bit measurement unit are used as approximations of the quasi-true modes. These are compared with the results obtained by the network from decoupling the ground signals for the corresponding time period, and the difference loss is calculated. The network parameters are then updated in small batches using this loss.

9. The adaptive method for a depth-adaptive drilling rig according to claim 7, characterized in that, In S140, the directional extraction of control sensitive features specifically includes: calculating the envelope of the modal signal and setting a dynamic threshold to detect impact events and calculate their energy; performing continuous spectrum estimation on the modal signal to track the main resonance frequency and calculate its drift; and calculating the energy of each modal signal and its proportion of the total energy within a sliding time window.

Images