Self-adaptive drilling method and system based on hole wall stability and deep learning

Abstract

The invention relates to a self-adaptive drilling method and system based on hole wall stability and deep learning, and belongs to the technical field of coal mine underground drilling. Aiming at the technical problems that the drilling efficiency is low and the construction safety risk is high due to the fact that an existing drilling control method is difficult to effectively predict and avoid hole wall instability, the method adopts the core technical scheme that a drilling parameter sequence is collected in real time through an orifice sensor; calculating a hole wall instability risk probability in real time by using a pre-trained lightweight hole wall instability prediction network; further, the scheme evaluation network predicts future multi-target performance based on the current state and the candidate control set; and finally, constructing a scoring function by taking the hole wall stability as a primary constraint, preferentially generating a control instruction and issuing the control instruction to a drilling machine for execution. According to the method, intelligent prediction and active control over the stability of the hole wall are achieved, the safety and reliability of drilling construction are remarkably improved, meanwhile, the drilling efficiency and the hole forming quality are both considered, the system calculation load is low, and on-site deployment and application are easy.

Description

Technical Field

[0001] This invention belongs to the field of underground drilling technology in coal mines, and relates to an adaptive drilling method and system based on borehole wall stability and deep learning. Background Technology

[0002] In coal mine drilling operations, borehole wall stability is a key factor affecting drilling efficiency and safety. Borehole instability manifests as cracking, collapse, and crack propagation. These phenomena not only significantly reduce drilling speed but can also damage equipment and even cause work stoppages. Therefore, effectively predicting and preventing borehole instability is crucial for improving drilling efficiency and ensuring wellbore safety.

[0003] Traditional borehole stability control methods typically rely on empirical settings and rule-based control of drilling parameters, such as judging whether borehole instability has occurred based on borehole vibration signals or cuttings discharge. However, these traditional methods exhibit significant limitations in drilling environments with complex and dynamically changing lithology. Their judgments lack scientific basis, are not adaptable enough, and are difficult to achieve precise control.

[0004] Currently, in the field of adaptive drilling control, existing methods mainly focus on improving drilling efficiency as the sole control objective, and generally lack adaptive control strategies with borehole wall stability as the core control objective. However, in reality, good borehole quality is just as important as drilling efficiency, and effective control of borehole wall stability is a key link in improving borehole quality and ensuring long-term drilling efficiency. Taking mechanical drilling speed as an example, although a high mechanical drilling speed means rapid rock breaking efficiency, neglecting borehole wall stability can easily induce borehole wall instability, leading to a decrease in overall efficiency or even project failure.

[0005] In summary, existing technologies have significant gaps, creating an urgent need for an intelligent adaptive drilling method that prioritizes borehole stability as the core control objective and can adapt to complex geological conditions. This method should be able to perceive working conditions in real time, scientifically predict risks, and dynamically optimize drilling parameters, thereby achieving efficient drilling while ensuring safety. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide an adaptive drilling method and system based on borehole wall stability and deep learning, which combines a deep learning model to predict borehole wall stability in real time and adjusts drilling parameters according to the prediction results, thereby realizing intelligent control of borehole wall stability.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An adaptive drilling method based on borehole wall stability and deep learning, the method comprising the following steps: S1: Data acquisition and processing steps: Acquire multi-channel sensing sequences through sensors installed at the orifice, and preprocess the multi-channel sensing sequences; S2: Borehole instability risk prediction step: Input the preprocessed multi-channel sensing sequence and static formation parameters into the pre-trained borehole instability prediction network (Rock Failure Net) to calculate the borehole instability risk probability at the current moment; S3: Control scheme evaluation step: Based on the probability of instability risk of the orifice wall and the multi-channel sensing sequence, the current state features are constructed, and combined with the candidate control set, the multi-objective performance index after executing each candidate control scheme is predicted by the pre-trained scheme evaluation network (Plan Net); S4: Optimal decision-making and safe projection steps: Construct a scoring function based on the predicted multi-objective performance index, select the optimal control scheme under the condition of satisfying the hole wall stability constraint, and perform safe projection processing on the optimal control scheme to generate the final control command; S5: Command execution and closed-loop update steps: The final control command is sent to the drilling rig's electrical control system for execution to adjust drilling parameters and continuously collect new sensor data to update the current state characteristics, thereby achieving closed-loop control.

[0008] Furthermore, in S1, the multi-channel sensing sequence includes rotational speed, drilling pressure, torque, and vibration signals, with a sampling frequency of not less than 200Hz, a sampling window length of 2 seconds, and a sliding step size of 0.2 seconds; the preprocessing includes synchronous resampling and alignment of the multi-channel time-series signals, detrending, 2-80Hz bandpass filtering, and standardization according to the physical dimensions of each sensor.

[0009] Furthermore, in S2, the static formation parameters include confining pressure intensity. Poisson's ratio Compressive strength or shear strength internal friction angle Cohesion c and elastic modulus E .

[0010] Furthermore, in S2, the construction and operation of the borehole wall instability prediction network includes: copying and expanding the static formation parameter matrix into a pseudo-sequence of the same length as the multi-channel sensing sequence, and then splicing it with the multi-channel sensing sequence in the channel dimension to form a joint input. X t The combined input X tFeatures are extracted sequentially through three layers of one-dimensional convolutional neural networks (1DCNN), then temporal dynamics are captured by bidirectional gated recurrent units (BiGRU), and weighted by a self-attention mechanism. Finally, the probability of hole wall instability is output through global average pooling and a fully connected layer. s t .

[0011] Furthermore, in S3, the candidate control set u t Including speed increment Incremental drilling pressure Torque limit increment and mechanical drilling rate increment .

[0012] Furthermore, in S3, the construction and computation of the scheme evaluation network Plan Net includes: encoding the current state features using a state encoder to obtain a state embedding. The action embedding is obtained by encoding the schemes in the candidate control set using an action encoder. ψ ( i ); splicing the embedded states and the action embedded ψ ( i After that, the fusion evaluation module outputs the predicted probability of hole wall instability. Changes in mechanical drilling speed Impact strength change and energy consumption .

[0013] Furthermore, in S4, the scoring function is expressed as:

[0014] in, This represents the predicted probability of borehole wall instability. This represents the change in mechanical drilling speed. This represents the change in impact strength. For energy consumption, , , , Let be the weight coefficient, and satisfy... ; Select the scoring function J The candidate solution that is maximized is taken as the optimal control solution.

[0015] Furthermore, in S4, if the predicted probability of hole wall instability of the candidate scheme is... Exceeding the safety threshold determined by the on-site geological and engineering safety level If the candidate solution is not found, it will be eliminated.

[0016] Furthermore, in S4, the safety projection process uses a quadratic programming solver to project the optimal control scheme under hard constraints of upper and lower bounds for the rotational speed, drilling pressure, and torque of the drilling rig actuator, generating feasible final control commands. u safe .

[0017] Furthermore, the training data for the borehole wall instability prediction network and the scheme evaluation network are generated through a simulation system based on the drill string dynamics model and the borehole wall surrounding rock constitutive model; the drill string dynamics equation is expressed as:

[0018] in, The unit mass matrix, For the element damping matrix, The element stiffness matrix, For physical constraint, To account for the contact force matrix considering formation boundary constraints, For acceleration, For the speed term, This is the displacement term.

[0019] An adaptive drilling control system for implementing the method, the system comprising: The sensor module is used to collect the rotational speed, drilling pressure, torque and vibration signals at the borehole opening, forming a multi-channel sensing sequence; A data processing module, connected to the sensor module, is used to preprocess the multi-channel sensing sequence; The borehole wall instability prediction module is connected to the data processing module and has a built-in borehole wall instability prediction network. It is used to receive the preprocessed multi-channel sensing sequence and static formation parameters, and calculate and output the probability of borehole wall instability risk at the current moment. The scheme evaluation module is connected to the data processing module and the orifice wall instability prediction module. It has a built-in scheme evaluation network, which is used to predict the multi-objective performance index after executing each candidate control scheme based on the orifice wall instability risk probability and the multi-channel sensing sequence to form the current state features and combine them with the candidate control set. The decision control module, connected to the scheme evaluation module, is used to construct a scoring function based on the multi-objective performance index to select the optimal control scheme and perform safety projection processing to generate the final control command. An actuator, connected to the decision control module, is used to receive the final control command and adjust the drilling rig's rotational speed, drilling pressure, and torque parameters.

[0020] Furthermore, the sensor module includes a vibration sensor, a torque sensor, a speed sensor, and a drill pressure sensor. The sampling frequency of the vibration sensor is preferably 500-1000Hz, and the sampling frequencies of the torque sensor, speed sensor, and drill pressure sensor are not less than 100Hz.

[0021] Furthermore, the borehole wall instability prediction module and the scheme evaluation module are deployed on the drilling rig's local controller, which is an embedded graphics processor (GPU) or a programmable logic controller (PLC) AI module.

[0022] Furthermore, the decision control module is also configured to perform online fine-tuning of the borehole instability prediction network and the scheme evaluation network based on real-time drilling data.

[0023] The beneficial effects of this invention are as follows: (1) The core advantage of this invention lies in taking borehole wall stability as the primary control objective, achieving a fundamental shift from passive response to proactive prevention. Through a borehole wall instability prediction network, the system accurately predicts borehole wall instability risks in real time and online, enabling it to take control measures before signs of instability such as collapse or rupture appear. This forward-looking intelligent control greatly reduces the probability of borehole wall instability accidents, effectively avoiding serious engineering problems such as drill bit jamming, drill bit burial, and even equipment damage caused by borehole wall issues, thus providing a solid safety guarantee for the entire drilling operation.

[0024] (2) This invention breaks through the limitations of traditional reliance on human experience and creatively integrates two types of lightweight deep learning networks into the control loop. This method can not only perceive the complex dynamic working conditions downhole in real time, but also intelligently evaluate the future effects of various control schemes and autonomously select the optimal strategy. The system has strong adaptive capabilities and can automatically adjust key parameters such as rotation speed and drilling pressure according to different formation lithology and equipment status, so that the drilling process always maintains the optimal range of high efficiency and stability, realizing truly intelligent drilling.

[0025] (3) Traditional methods struggle to balance drilling speed and borehole quality, often sacrificing one to preserve the other. This invention utilizes a scheme evaluation network for multi-objective collaborative optimization. Under the rigid constraint of ensuring borehole wall stability, it simultaneously pursues maximizing mechanical drilling speed, minimizing drill string impact vibration, and optimizing energy consumption. This enables the drilling rig to achieve high drilling efficiency while simultaneously forming regular, stable, and high-quality boreholes, solving the technical challenge of balancing efficiency and quality, and maximizing overall benefits.

[0026] (4) The core neural network models designed in this invention are all lightweight time-series models, and their computational intensity is far lower than that of the artificial intelligence modules of embedded graphics processors or high-end programmable logic controllers in existing drilling rig electrical control systems. This feature allows the entire intelligent control system to be deployed directly on the local controller of the drilling rig without relying on expensive high-performance computing equipment or complex external systems, which greatly reduces the cost and complexity of system modification and is conducive to the rapid promotion and field application of the technology.

[0027] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This describes the methods for obtaining the training and validation sets; Figure 2 This is a diagram of the Rock Failure Net network architecture. Figure 3 This is a diagram of the Plan Net network architecture. Figure 4 This is an adaptive control method for the drilling process based on borehole wall stability and deep learning. Detailed Implementation

[0029] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0030] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0031] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0032] 1. Methodology and Principles By utilizing two types of neural networks—Rock Failure Net and Plan Net—control variables such as rotational speed, drilling pressure, torque limiting, and vibration excitation are dynamically optimized while meeting safety constraints. This reduces the risk of borehole failure, increases the rate of mechanical drilling (ROP), suppresses impact vibration, and reduces drilling energy loss.

[0033] By monitoring multiple parameters at the borehole in real time, and taking into account the geological and lithological parameters of the coal and rock strata being drilled, the system can predict whether borehole instability will occur, and adaptively adjust drilling parameters accordingly to ensure the safety and efficiency of the drilling process.

[0034] 2. Methods for obtaining training and validation sets This method uses deep learning to construct the relationship between borehole parameters and borehole wall stability. Since the drill string motion inside the borehole is chaotic and highly complex, and the contact between the drill string and the borehole wall involves uncertainties in both time and space, it is currently difficult to obtain large amounts of borehole wall stability data experimentally. Furthermore, there is limited research on borehole wall stability models for rotary drilling processes. Therefore, a borehole wall stability model is established, and this model is used to obtain a large amount of neural network training and validation data through simulation.

[0035] Numerous studies have shown that drill string motion is closely related to wellhead parameters such as drill pressure, rotational speed, vibration signal, and torque, as well as the properties and parameters of the drilled formation. Borehole stability is affected by drill string motion; therefore, it can be considered that wellhead drilling parameters and lithological parameters are related to borehole stability. In addition, the author believes that borehole stability is also related to the drilling length. Based on this, a borehole stability equation can be written: (1) in: x For the three-dimensional vibration signal of the drill pipe, T For torque, n For speed, F For drilling pressure,r For formation parameters, h For drilling in coal mines, parameters such as confining pressure, Poisson's ratio, and elastic modulus of the formation have a certain influence on drill string movement and borehole stability. The formation parameter matrix can be expressed as: (2) in: For confining pressure strength, For Poisson's ratio, For compressive or shear strength, For internal friction angle, c For cohesion, E It is the elastic modulus.

[0036] Because boreholes are elongated hollow cylinders in space, and the movement of the drill string within the borehole is irregular, while borehole wall stability is closely related to the contact between the drill string and the borehole wall, a drill string dynamics model should be established to describe the motion of the drill string at different spatial positions within the borehole. The stability of rock strata parameters at any point within the borehole is reflected in two aspects: First, the drill string movement is affected by the properties of the rock strata, and drill string dynamics should fully consider these properties; therefore, a drill string dynamics model considering the surrounding rock properties should be constructed. Second, when the drill string contacts the borehole wall, the contact force is related to the rock properties, and this contact force directly affects borehole wall stability; therefore, a constitutive model of the borehole wall surrounding rock should be constructed.

[0037] According to the finite element method, the drill string dynamics equation considering the surrounding rock properties can be expressed as:

[0038] in, The unit mass matrix, For the element damping matrix, The element stiffness matrix, For physical constraint, To account for the contact force matrix considering formation boundary constraints, For acceleration, For the speed term, This is the displacement term.

[0039] Traditional drill string dynamics treats the contact between the drill string and the formation as a rigid contact, but this is not realistic. The drill string and the surrounding rock of the borehole wall exhibit elastoplastic contact, meaning that contact forces... Geological factors need to be considered, therefore a constitutive model of the borehole wall surrounding rock needs to be constructed. According to the generalized Hooke's law, the basic equations of the elastoplastic constitutive relationship can be written as follows: (4) in: Contact stress; This refers to the elastic modulus, with units of MPa. For elastic strain; This is plastic strain.

[0040] The contact force matrix between the drill string and the surrounding rock of the borehole wall can be expressed as: (5) in: For the contact area, This refers to contact stress.

[0041] By considering the properties of the surrounding rock of the borehole wall, a drill string dynamics model is used to solve the motion state of the drill string inside the borehole based on the boundary conditions at the borehole opening. This yields the motion state and contact points of the drill string within the borehole space. Contact stress is then calculated based on the contact state, and the Mohr-Coulomb criterion is used as the criterion for borehole wall instability. The Mohr-Coulomb criterion can be expressed as: (6) in: This represents the normal stress on the shear plane, expressed in MPa. This represents the maximum principal stress at which the rock sample fails, expressed in MPa.

[0042] The instability condition of the surrounding rock of the borehole wall can be expressed as: (7) in: The actual shear stress of the surrounding rock of the borehole wall is obtained through the drill string dynamics model.

[0043] Based on the above-mentioned instability model inside the borehole, the data information of the contact between the drill string and the borehole wall in the time domain can be obtained in real time by using the borehole parameters, and the stability data of each point in the borehole space domain can be obtained. This data can be used as the training set and validation set of the deep learning neural network.

[0044] The methods for obtaining the training and validation sets are as follows: Figure 1 As shown.

[0045] 3. Data Acquisition, Weak Label Construction, and Control Target Setting 3.1 Data Acquisition Sensors are installed at the orifice to collect data and obtain sensor sequences. x t ∈{ C × T}, C The dimension for acquiring signals from the orifice sensor. T Let be the number of sampling points within the window, and let the sliding step size be . δ Units are in seconds. Vibration ≥200Hz (preferably 500~1000Hz), torque / speed / drilling pressure and other parameters ≥100Hz; window L =2s, step size δ =0.2s.

[0046] In intelligent monitoring of the drilling process, the multi-channel time-series signals are first synchronously resampled and aligned at a frequency of 200 Hz or higher to eliminate sensor clock drift and phase difference. Subsequently, detrending and 2–80 Hz bandpass filtering are performed to retain the effective frequency bands related to bottom hole drill string vibration, stick-slip, and borehole wall mechanical response. The signals are then standardized according to the physical dimensions of each sensor to ensure the comparability of cross-dimensional characteristics.

[0047] 3.2 Weak Tag Construction Weak labels are constructed for supervised or semi-supervised learning, and the proximity of hole wall instability is defined based on the Mohr-Coulomb criterion. FP : (8) A weak positive label is defined as one that meets any of the following conditions: ① FP > θ f (Empirical threshold, obtained through experiments, with a reference value of 0.15–0.25 provided here); ② Abnormal rise in vibration energy in the 30–60Hz frequency band (vibration energy greater than 1.25 times the mean and lasting for more than 2 seconds); ③ Instantaneous fluctuations in drill weight (WOB) or torque exceeding 1.25 times the standard deviation of their respective 10-second sliding window.

[0048] 3.3 Adaptive Control Objective Adaptive control is achieved by controlling the rotational speed, drilling pressure, upper limit of torque, and drilling speed. The control step matrix can be written as: (9) in: , , and These correspond to incremental settings for rotational speed, drilling pressure, upper limit of torque, rotational speed, main vibration frequency, and amplitude, respectively.

[0049] 4. Construction of the Rock Failure Net for Predicting Hole Wall Instability 4.1 Input Definition Rock Failure Net was used to predict borehole instability risk. Rock Failure Net simultaneously received two heterogeneous information sources, one of which was a real-time acquired 2-second high-frequency drilling sequence. xt (The data structure is 8×400), covering parameters such as rotational speed, drilling speed, and torque sampled at 200Hz, used to characterize the instantaneous dynamics inside the hole and as input for network training; the other is a static formation description. r(The data structure is 6×400), which consists of mechanical parameters such as internal friction angle, cohesion, and pore pressure obtained from drilling measurements or geological exploration. It is used to provide a priori information about the stratigraphy and serve as input parameters for the network.

[0050] 4.2 Network Architecture Input layer: Embeds static stratigraphic branch data into time-series stratigraphic branches, matrix r Copy and expand according to time sequence length to the same batch x t Equal-length pseudo-sequences, then with x t By splicing the channels together, a joint input is formed. X t ∈ R (C+d)×T The data is then fed into a 1DCNN to extract the dynamic-stratigraphic coupling features.

[0051] Convolutional layers: The first layer of the 1D CNN uses a 32-core 1D convolutional layer. k =5, s =1, same zero-padding) While maintaining a length of 400 points, the primary local patterns are extracted and output as 32×400 feature maps after normalization and ReLU. The second layer further expands the receptive field with dilated convolutions of 64 kernels and an inflation rate of 2, capturing long-range multi-scale associations without increasing the number of parameters. It also obtains a 64×400 representation after ReLU. The third layer halves the sequence length to 200 with max pooling of k=2, while maintaining the number of channels at 64. This significantly compresses the computational load and retains significant features, providing a refined temporal input of "abstraction-dimensionality reduction-multi-scale" for BiGRU.

[0052] BiGRU layer: The 64×200 feature sequence obtained by pooling is first captured synchronously through a bidirectional gated recurrent unit (BiGRU, hidden state 128×2) to capture the dynamic evolution of the hole wall stress in the forward and backward directions, and outputs a 256×200 context tensor.

[0053] Attention mechanism: A self-attention mechanism is used to calculate the weights for each time step, and the time-channel weights are calculated synchronously in the feature space to highlight the key response dimensions at critical moments.

[0054] Output layer: After being compressed into a 256-dimensional vector by global average pooling, it is then processed through two fully connected layers (256→64→1) and activated by Sigmoid. This process reduces the risk of output aperture wall instability. s t ∈[0,1], to achieve end-to-end estimation of the probability of wellbore mechanical stability.

[0055] Rock Failure Net network architecture as follows Figure 2 As shown.

[0056] 5. Construction of the Plan Net Given the current drilling status and a set of candidate control schemes, a multi-objective evaluation model is established to predict the future performance of each candidate scheme and select the optimal one for execution. The specific process is as follows: 5.1 Input Definition Current state characteristics: Hole wall instability risk output by Rock Failure Net s t With the original sensing sequence x t Together they form the state vector z t .

[0057] Candidate control set: ,in: , , and These correspond to incremental settings for rotational speed, drilling pressure, upper limit of torque, and rotational speed, respectively.

[0058] 5.2 Network Architecture State encoder: for x t Apply a one-dimensional convolution stack to extract higher-order dynamic embeddings. .

[0059] Action encoder: converts discrete or continuous signals u ( i Mapping to same-dimensional action embedding ψ ( i ).

[0060] Fusion-Evaluation Module: Stitching [ ; ψ ( t The input is then fed into two fully connected layers, and the output is a scheme-specific prediction: (10) Execute respectively u ( i )back τ The probability of borehole wall instability within seconds, the change in mechanical drilling speed, drilling impact, and drilling energy consumption.

[0061] The PlanNet network uses an "input-encoding-fusion evaluation-decision" process as its core: the input layer receives state input consisting of the hole wall instability risk and the original sensing sequence (number of sensors C=6, time step T=400), as well as... The system takes four control parameters as input. The state encoder extracts local and temporal dependency features through 1D convolution (32 kernels, kernel size 5) and 1D dilated convolution (64 kernels, dilation rate 2), and then obtains a 64-dimensional state embedding through global temporal average pooling. The action encoder maps the four-dimensional control parameters to a 16-dimensional action embedding through two fully connected layers (including LayerNorm normalization). The fusion-evaluation module first concatenates the two types of embeddings to obtain 80-dimensional joint features, which are then enhanced by a shared fully connected layer with an output dimension of 128 (Dropout probability 0.1). The multi-task head then predicts the hole wall instability probability (Sigmoid activation), mechanical drilling rate change, impact, and energy consumption (the latter three are linearly output). The decision module uses a weighted score of -w1·Risk + w2·ΔROP + w3·ΔShock + w4·P, combined with the on-site safety threshold to filter out risk exceeding the standard, and finally selects the optimal control parameter with the highest score to achieve high drilling speed, low impact, low energy consumption, and safe and reliable drilling control. Network architecture such as Figure 3 As shown.

[0062] The detailed design of the network architecture is shown in Table 1.

[0063] Table 1 Plan Net Network Architecture Design Parameters

[0064] 5.3 Decision-making objectives Construct a weighted linear multi-objective scoring function using the prediction results, while satisfying ( Under the premise of safety constraints such as on-site geological conditions and engineering safety levels, the optimal control scheme that maximizes drilling speed and minimizes impact is selected. u t .

[0065] 6. Adaptive control scheme for drilling process An adaptive drilling control framework with borehole wall stability as the primary optimization objective solves for the optimal control sequence online using a "rolling time-constraint optimization" strategy after the prediction-evaluation phase. The complete process can be divided into the following five stages: Candidate instruction generation: based on the state at the current time t z t =[ x t ; s t ; r ](in x t It is a 2-second high-frequency sensing sequence. s t This represents the probability of hole wall instability output by the Rock Failure Net.r (Static formation mechanical parameters), within the allowable control increment space Δ U K candidate instructions are generated using Gaussian perturbation. U t Each instruction It covers the adjustment range of rotational speed, drilling pressure, upper limit of torque and vibration reduction parameters to ensure that the feasible range is fully explored and the engineering operability is preserved.

[0066] Multi-objective performance extrapolation: Utilizing a fully trained Plan Net, parallel forward extrapolation is performed on the candidate set to output future performance. τ =Probability of borehole wall instability within a 5-second time period ∈[0,1] (directly associated with rockfall, collapse, and fracturing risks), mechanical drilling rate variation used to measure rock breaking efficiency. and the change in impact intensity reflecting drill string fatigue and wellbore impact failure tendency In addition, the network provides additional command power consumption. It represents the energy consumed during the drilling process.

[0067] Multi-objective optimization: Construct a scalar evaluation function with pore wall stability as the core, as shown below: (11) The weights satisfy ,and λ 0 Real-time adaptive increase (more conservative for higher risk). For any condition satisfying... The candidate solutions are directly assigned Eliminate them to achieve implicit processing of "hard constraints on stability".

[0068] Safe projection: The initial instruction with the highest score. u Input the quadratic programming (QP) solver, project it under hard constraints (rotation speed, drilling pressure, and upper and lower bounds of torque) to generate a feasible and safe final control. u safe .

[0069] Closed-loop execution and rolling updates: u safe The data is sent to the drilling rig's electrical control box via fieldbus to control the hydraulic system response, thereby controlling the actuators such as the power head and frame, and controlling parameters such as speed, drilling pressure, and torque. At the same time, new sensor data is continuously collected to update the inputs of the Rock Failure Net and Plan Net, forming a rolling time-domain closed loop of "perception-prediction-optimization-execution".

[0070] This framework uses borehole stability as a hard constraint and overall drilling efficiency as a soft objective. Under the premise of ensuring borehole stability, it achieves an online trade-off between mechanical drilling rate and energy consumption, and ultimately completes adaptive drilling control for complex formations.

[0071] Adaptive control logic such as Figure 4 As shown.

[0072] 7. Neural Network Training and Adaptive Adjustment 7.1 Training Process Rock Failure Net: The neural network is trained using drill string motion data calculated by a theoretical model considering borehole wall stability, borehole wall instability data (i.e., simulation data), existing experimental data, and existing historical drilling data to learn the relationship between features in the data and borehole wall instability.

[0073] Plan Net: Uses simulation data or historical data to generate samples, trains the network to evaluate the effectiveness of different control schemes.

[0074] 7.2 Adaptive Adjustment As drilling progresses and factors such as changes in lithology and equipment aging occur, the control system requires online fine-tuning. The system continuously adjusts model parameters based on new data and evaluates the results after each update to ensure effective control.

[0075] Example 1: Network Deployment and Basic Control Flow Based on Simulation Model Training First, a large amount of data is needed to train the neural network. Since obtaining borehole stability data directly through downhole experiments is extremely difficult, this embodiment uses a high-fidelity drill string dynamics model and a borehole wall surrounding rock constitutive model for simulation. Specifically, a drill string dynamics equation considering the surrounding rock properties is established, which includes element mass matrix, element damping matrix, element stiffness matrix, body constraint forces, and a contact force matrix considering formation boundary constraints. Simultaneously, an elastoplastic constitutive model of the borehole wall surrounding rock is constructed based on the generalized Hooke's law to accurately calculate the contact stress when the drill string contacts the borehole wall. Finally, the Mohr-Coulomb criterion is used as the borehole wall instability criterion. By inputting different borehole parameters and formation parameters, the simulation system can output the drill string's motion state within the borehole and detailed data on its contact with the borehole wall, thereby generating a training dataset labeled with borehole wall stability. The methods for obtaining the training and validation sets are as follows... Figure 1 As shown.

[0076] Next, the above simulation data is used to train a hole wall instability prediction network, namely Rock Failure Net. The architecture of this network is as follows: Figure 2As shown, its input includes a real-time 2-second high-frequency drilling sequence and static formation description parameters. The network first expands the static formation parameters into a pseudo-time series and concatenates it with the sensing sequence along the channel dimension. The concatenated joint input is then processed by a three-layer one-dimensional convolutional neural network to extract local and multi-scale features, with the second layer employing dilated convolution to increase the receptive field. The convolutionally processed features are then fed into a bidirectional gated recurrent unit to capture the forward and backward temporal dynamics of borehole wall stress evolution. Subsequently, a self-attention mechanism is used to weight the time steps, highlighting key information. Finally, after global average pooling and a fully connected layer, a borehole wall instability risk probability value between 0 and 1 is output.

[0077] Then, the training scheme evaluates the network, namely Plan Net. The architecture of this network is as follows: Figure 3 As shown, its input consists of two parts: first, the current state features, composed of the risk probability output by Rock Failure Net and the original sensor sequence; second, the candidate control set, including the rotational speed increment, drilling pressure increment, torque limit increment, and mechanical drilling rate increment. The network encodes the current state through a state encoder and the candidate control schemes through an action encoder. The embedded vectors encoded by both are concatenated and fed into a fusion evaluation module, which outputs the predicted values ​​for a future period after executing each candidate scheme, including the probability of borehole instability, the change in mechanical drilling rate, the change in impact intensity, and energy consumption.

[0078] Finally, the trained lightweight network model is deployed on the drilling rig's local controller, such as an embedded GPU or the AI ​​module of a high-end PLC. During on-site construction, the system... Figure 4 The workflow shown is as follows: real-time acquisition of borehole sensor data; calculation of the current borehole wall instability risk using Rock Failure Net; generation of a batch of candidate control commands based on the current state, and prediction of the future effects of each command using Plan Net; selection of the optimal solution based on a scoring function with borehole wall stability as the primary objective; and issuance of control commands to the actuators after safety projection to adjust parameters such as drilling rig speed and drilling pressure, thereby achieving closed-loop control.

[0079] Example 2: Enhanced control process for complex and fractured formations Based on Example 1, this paper further details how the control system ensures safety by strengthening risk assessment and constraint handling when drilling encounters complex and fractured formations.

[0080] When the drilling rig enters fractured formations with low internal friction angles, low cohesion, and well-developed joints, the risk of borehole instability increases significantly. In this situation, the system adaptively adjusts its decision-making strategy. First, Rock Failure Net monitors the continuous increase in the probability of borehole instability risk. During the scheme evaluation phase, the decision module dynamically increases the weighting coefficient of the borehole instability risk term in the scoring function, making the system's decision more conservative. Specifically, the screening of candidate control schemes is more stringent; any scheme that might cause the risk probability to exceed the safety threshold is directly rejected.

[0081] Furthermore, when generating candidate control sets, the system narrows the search range of control parameters, particularly reducing the incremental upper limits of drill pressure and torque, to avoid disturbing the borehole wall due to overly aggressive operations. During the safe projection phase, the quadratic programming solver strictly enforces more stringent torque and drill pressure constraints. Simultaneously, the system may prioritize schemes incorporating active vibration reduction parameters to suppress the impact of the drill string on the fractured borehole wall. Throughout the process, the system achieves safe, slow drilling in highly unstable formations through higher control frequencies and more cautious optimization strategies, with the primary goal of maintaining borehole stability rather than pursuing drilling speed.

[0082] Example 3: System's Online Self-Learning and Parameter Fine-Tuning Process As drilling progresses, the system continuously accumulates new wellhead sensor data and corresponding feedback on actual drilling results. This field data is stored and used for incremental learning of the model. For example, the system periodically compares and analyzes sensor data sequences over a period of time with subsequent actual wellbore conditions. If a systematic deviation is found between the Rock Failure Net's predictions and the actual situation, the network weights are fine-tuned using new data to make the prediction model more closely reflect the actual formation response characteristics of the work area.

[0083] Similarly, for Plan Net, the system records the executed control schemes and their resulting effects on mechanical drilling speed, equipment vibration, and other actual conditions. This data serves as new samples to optimize the accuracy of scheme evaluation. This online self-learning capability allows the control system to gradually break free from complete dependence on initial simulation data, continuously evolve, and ultimately form a highly customized intelligent control expert system for the current drilling rig and specific geological environment, thereby continuously improving the prediction accuracy and control effectiveness of drilling performance.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive drilling method based on borehole wall stability and deep learning, characterized in that: The method comprises the following steps: S1: data acquisition and processing step: acquiring a multi-channel sensing sequence through a sensor arranged at the orifice, and pre-processing the multi-channel sensing sequence; S2: borehole wall instability risk prediction step: inputting the pre-processed multi-channel sensing sequence and static formation parameters into a pre-trained borehole wall instability prediction network to calculate the borehole wall instability risk probability at the current time; S3: control scheme evaluation step: based on the borehole wall instability risk probability and the multi-channel sensing sequence, a current state feature is formed, and combined with a candidate control set, a multi-objective performance index after executing each candidate control scheme is predicted through a pre-trained scheme evaluation network; S4: optimal decision and safety projection step: a scoring function is constructed according to the predicted multi-objective performance index, an optimal control scheme is selected under the condition of meeting the borehole wall stability constraint, and safety projection processing is performed on the optimal control scheme to generate a final control instruction; S5: instruction execution and closed-loop update step: the final control instruction is issued to the electric control system of the drilling rig for execution to adjust the drilling parameters, and new sensing data is continuously acquired to update the current state feature, thereby realizing closed-loop control.

2. The adaptive drilling method based on borehole wall stability and deep learning of claim 1, wherein: In S1, the multi-channel sensing sequence includes rotation speed, drilling pressure, torque and vibration signals, the sampling frequency is not less than 200 Hz, the sampling window length is 2 seconds, and the sliding step is 0.2 seconds; the pre-processing includes synchronous resampling and alignment of multi-channel time series signals, detrending, 2-80 Hz band-pass filtering, and standardization according to the physical dimension of each sensor.

3. The adaptive drilling method based on borehole wall stability and deep learning of claim 1, wherein: In S2, the static formation parameters include confining stress intensity , Poisson's ratio , compressive or shear strength , internal friction angle , cohesion c , and elastic modulus E .

4. The adaptive drilling method based on borehole wall stability and deep learning of claim 1, wherein: In S2, the construction and operation of the hole wall instability prediction network comprises: copying and extending the static formation parameter matrix into a pseudo sequence with the same length as the multi-channel sensing sequence, and then splicing the pseudo sequence with the multi-channel sensing sequence in the channel dimension to form a joint input X t ; the joint input X t The features are extracted through three one-dimensional convolutional neural networks (1DCNN) in sequence, the time sequence dynamics are captured through a bidirectional gated recurrent unit (BiGRU), and the hole wall instability risk probability is output through global average pooling and a fully connected layer after weighting by a self-attention mechanism s t .

5. The adaptive drilling method based on borehole wall stability and deep learning of claim 1, wherein: In S3, the candidate control set u t including a rotational speed increment , a weight on bit increment , a torque upper limit increment and a rate of penetration increment .

6. The adaptive drilling method based on borehole wall stability and deep learning of claim 1, wherein: In S3, the construction and operation of the scheme evaluation network Plan Net includes: encoding the current state feature by a state encoder to obtain a state embedding , encoding the schemes in the candidate control set by an action encoder to obtain an action embedding Ψ ( i );splicing the state embedding and the action embedding Ψ ( i ), and outputting, by a fusion evaluation module, a predicted borehole wall instability probability , a mechanical drilling speed change amount , an impact strength change amount and energy consumption .

7. The adaptive drilling method based on borehole wall stability and deep learning of claim 6, wherein: In S4, the scoring function is represented as: wherein, is a predicted hole wall instability probability, is a mechanical drilling rate change, is an impact strength change, is an energy consumption, , , , is a weight coefficient, and satisfies ; selecting a candidate solution that maximizes the score function J as the optimal control solution.

8. The adaptive drilling method based on borehole wall stability and deep learning of claim 7, wherein: In S4, if the predicted probability of hole wall instability of the candidate scheme is... Exceeding the safety threshold determined by the on-site geological and engineering safety level If the candidate solution is not found, it will be eliminated.

9. The adaptive drilling method based on borehole wall stability and deep learning of claim 1, wherein: In S4, the safety projection processing is performed by a quadratic programming solver to project the optimal control scheme under the upper and lower boundary hard constraints of the rotation speed of the rig actuator, the drilling pressure, and the torque to generate the final control instruction u safe .

10. The adaptive drilling method based on borehole wall stability and deep learning of claim 1, wherein: The training data of the borehole wall instability prediction network and the scheme evaluation network is generated by a simulation system based on a drill string dynamics model and a borehole wall surrounding rock constitutive model; the drill string dynamics equation is represented as: wherein, is a unit mass matrix, is a unit damping matrix, is a unit stiffness matrix, is a body force constraint force, is a contact force matrix considering the stratum boundary constraint, is an acceleration term, is a velocity term, is a displacement term.

11. An adaptive drilling control system for implementing the method of any one of claims 1 to 10, characterized by: The system comprises: A sensor module for acquiring rotation speed, drilling pressure, torque and vibration signals at the orifice to form a multi-channel sensing sequence; A data processing module connected to the sensor module for pre-processing the multi-channel sensing sequence; A borehole wall instability prediction module connected to the data processing module and internally provided with a borehole wall instability prediction network for receiving the pre-processed multi-channel sensing sequence and static formation parameters and calculating and outputting the borehole wall instability risk probability at the current time; A scheme evaluation module connected to the data processing module and the borehole wall instability prediction module and internally provided with a scheme evaluation network for forming a current state feature based on the borehole wall instability risk probability and the multi-channel sensing sequence, and combining a candidate control set to predict a multi-objective performance index after executing each candidate control scheme; A decision control module connected to the scheme evaluation module for constructing a scoring function according to the multi-objective performance index to select an optimal control scheme and performing safety projection processing to generate a final control instruction; An execution mechanism connected to the decision control module for receiving the final control instruction and adjusting the rotation speed, drilling pressure and torque parameters of the drilling rig.

12. The adaptive drilling control system of claim 11, wherein: The sensor module comprises a vibration sensor, a torque sensor, a rotation speed sensor and a drilling pressure sensor, the sampling frequency of the vibration sensor is preferably 500-1000 Hz, and the sampling frequency of the torque sensor, the rotation speed sensor and the drilling pressure sensor is not less than 100 Hz.

13. The adaptive drilling control system of claim 11, wherein: The borehole wall instability prediction module and the scheme evaluation module are deployed on a local controller of a drilling rig, and the local controller is an AI module of an embedded graphic processing unit (GPU) or a programmable logic controller (PLC).

14. The adaptive drilling control system of claim 11, wherein: The decision control module is further configured to perform online fine-tuning on the borehole wall instability prediction network and the scheme evaluation network according to real-time drilling data.

Images