An adaptive electric-oil intelligent digital driving method and system for oil drilling conditions

Abstract

The application discloses an electric-oil intelligent digital driving method and system self-adapting to oil drilling conditions, and belongs to the field of oil drilling engineering.The method comprises the following steps: collecting oil drilling data through a multi-source sensor group, sending the collected data to an edge gateway controller for preprocessing, sending multi-source data streams obtained through the preprocessing to a cloud end, and constructing a multi-source database; deploying an intelligent torque prediction model on the cloud end, realizing real-time estimation of a current drill bit load torque according to data in the multi-source database, and obtaining a drill bit predicted torque value; the cloud end sends the calculated predicted torque value to the edge gateway controller, and the edge gateway controller generates actual control instructions according to the received predicted torque value.The application can effectively guarantee the safe and stable operation of drilling operations.

Description

Technical Field

[0001] This invention relates to the field of oil drilling engineering, and in particular to an intelligent digital drive method and system for electric oil replacement that is adaptive to oil drilling conditions. Background Technology

[0002] With the increasing intensity and expansion of oil and gas field development, the drilling environment is becoming increasingly complex, and formation conditions are highly variable. During drilling, the pore fluid pressure is often dynamically changing as the drill bit passes through different geological layers. With the development of monitoring while drilling (WSMD) technology, formation pressure signals can now be directly acquired by integrating sensors onto the drill string. However, existing drilling prediction models still have some problems. First, traditional drilling methods rely on experience and manual judgment, leading to frequent deviations and guidance errors in dynamic environments, which can easily reduce drilling efficiency and increase safety hazards. Although there are theoretical methods for applying neural networks to predict wellbore trajectories, existing neural network prediction methods often rely on single or multiple indicators and their thresholds, resulting in limited information. Furthermore, the thresholds for various parameters differ across different mines, and there is still no good solution for comprehensively judging when multiple indicators approach the thresholds at different levels. In addition, as drilling depth increases, downhole conditions become increasingly complex, and the working environment of the drill string becomes harsh, making it impossible to accurately know the working status of the drill bit in real time. Therefore, there is an urgent need for a method that can accurately evaluate the working status of the drill bit in real time in order to improve the efficiency and safety of drilling operations. Summary of the Invention

[0003] One of the objectives of this invention is to provide an intelligent digital drive method for electric oil drilling that is adaptive to oil drilling conditions, in order to solve the problem in the prior art that the working status of the drill bit cannot be known in real time and accurately due to the complexity of the downhole working conditions.

[0004] This invention is achieved through the following technical solution: an intelligent digital drive method for adaptive oil drilling conditions, comprising the following steps: S100, collecting oil drilling data through a multi-source sensor group, sending the collected data to an edge gateway controller for preprocessing, and sending the preprocessed multi-source data stream to the cloud to construct a multi-source database; S200, deploying an intelligent torque prediction model in the cloud, based on the data in the multi-source database, to achieve real-time estimation of the current drill bit load torque and obtain the predicted drill bit torque value. The intelligent torque prediction model includes: a feature space for reflecting the essential laws of torque signals during drilling, and a feature space for encoding logging-while-drilling data into a form with a clear physical... The system comprises a physical encoding sub-model for meaningful parameters and a prediction engine that makes predictions based on the feature space and physical meaning parameters. The prediction engine adopts a dual-stream architecture that integrates the physical sub-model and the data-driven sub-model. The physical sub-model is a set of differential equations constrained by physical laws, constructed based on the Rossler model, used to analyze the essential physical characteristics and obtain high-dimensional features of drill bit torque. The data-driven sub-model, through a deep learning network, autonomously learns from the high-dimensional features and captures complex patterns that the physical model fails to fully describe. The S300 and the cloud send the calculated predicted torque values ​​to the edge gateway controller, which generates actual control commands based on the received predicted torque values.

[0005] Furthermore, the multi-source data includes: electrical condition data, mechanical condition data, and geological condition data; the electrical condition data includes: stator three-phase current and bus voltage of the motor and rotor position data; the mechanical condition data includes: drill string vibration acceleration data, real-time drill pressure, rotational speed and mechanical drilling speed data; and the geological condition data includes: real-time logging data obtained through the logging-while-drilling system interface.

[0006] Furthermore, the intelligent torque prediction model also includes a dual-timescale learning framework designed to address issues caused by abrupt changes in geological formations. This dual-timescale learning framework includes a fast recurrent kernel based on extended Kalman filtering, used for real-time fine-tuning of the physical sub-model to enable it to quickly track short-term operating condition changes; and a slow recurrent kernel based on meta-learning, enabling rapid reshaping of the deep learning network of the data-driven sub-model.

[0007] Furthermore, the intelligent torque prediction model is constructed according to the following steps: S210, multi-source heterogeneous measurement data is transformed into a standardized feature space that can reveal the internal dynamic state of the drilling system. At the same time, by introducing geological prior knowledge, the original logging curves are transformed into physical coefficients that directly affect the mechanical process. By constructing a formation hardness mapping function, the natural gamma is transformed into a dimensionless formation hardness coefficient, and the conversion relationship from acoustic transit time to pore pressure gradient is established; S220, in order to achieve accurate prediction of torque signals, a prediction engine is constructed through a dual structure of physics and data. The physical sub-model of the prediction engine is constructed based on the Rossler model. By coupling the control parameters in the original Rossler model into physical coefficients transformed from the logging curves, and using mechanical operation parameters as external driving terms of the model, and using physical coefficients as intrinsic parameters of the state evolution equation, the endogenous variables of the internal evolution law of the Rossler model are adjusted; The data-driven sub-model predicts future torque through a lightweight adaptive spatiotemporal convolutional network based on the high-dimensional feature tensor calculated by the physical sub-model.

[0008] Furthermore, the standardized feature space is constructed through the following sub-steps: S211, the phase space of the core dynamic torque time-series signal of the drill bit is reconstructed to obtain the reconstructed dynamic state vector; S212, the logging-while-drilling data curve is encoded into parameters with clear physical meaning according to the physical parameters of the formation to obtain the encoded formation physical parameters; S213, the reconstructed dynamic state vector is fused with the encoded formation physical parameters to form a standardized feature space that includes the current motion state of the system and the physical constraints of the environment.

[0009] Furthermore, the reconstructed dynamic state vector is obtained by phase space reconstruction using time-delayed coordinate embedding. By combining the torque value at the current moment with the torque values ​​at multiple historical moments into a high-dimensional vector, the topological structure of the high-dimensional system is recovered from the one-dimensional time series, thereby providing a complete state description.

[0010] Furthermore, the reconstructed dynamic state vector can be represented by the following equation:

[0011] ,

[0012] in, For the reconstructed dynamic state vector, For time, Indicates time The reconstructed dynamic state vector; This is the torque value. This represents the torque value at the current moment. The torque value is the result of a past delay time. The torque value at a specific point in the past. For embedded dimensions, For delay time; A collection representing the embedded dimensions; This is the vector transpose symbol.

[0013] Furthermore, the formation hardness mapping function is constructed based on the Sigmoid function to encode the natural gamma value, thereby converting the natural gamma value into a continuous, normalized hardness parameter; the conversion relationship between acoustic transit time and pore pressure gradient is based on a rock physics model using acoustic data for a first-order approximation, thereby realizing the conversion between acoustic transit time and pore pressure gradient.

[0014] Furthermore, the formation hardness mapping function can be expressed as follows:

[0015] ,

[0016] in, This is the formation hardness coefficient, a dimensionless normalized value. It is a natural exponential function. A coefficient that controls the steepness of the function curve. This is the natural gamma value measured while drilling. The gamma baseline value is used as a reference stratum.

[0017] Furthermore, the conversion relationship between acoustic transit time and pore pressure gradient can be expressed as follows:

[0018] ,

[0019] in, For the Nabla operator, Pore ​​pressure, The pore pressure gradient; The acoustic transit time conversion factor is determined based on regional geological data. For sound wave time difference, These are the linear transformation coefficients determined based on regional geological data.

[0020] Furthermore, the physical sub-model and the data-driven sub-model complement each other through a fusion mechanism, which includes dynamic fusion weights and a pseudo-Hamilton function. The dynamic fusion weights, through a Sigmoid function, transform the uncertainty gradient of the physical sub-model, external disturbances, and the unreliability of the data source into a smooth weight, thereby achieving a dynamic trade-off between physical mechanisms and data-driven approaches. The pseudo-Hamilton function follows the principles of energy conservation and system uncertainty. By constructing a pseudo-Hamilton function that characterizes the total energy of the system, the gradient of the pseudo-Hamilton function in phase space is used to measure the degree to which the system deviates from steady state. When the system fluctuates violently or encounters high uncertainty in stratigraphic information, the fitting ability of the data sub-model is trusted; when the system operates smoothly and the stratigraphic characteristics are clear, the physical sub-model with clear physical meaning is trusted.

[0021] Furthermore, the physical sub-model based on the Rossler model can be represented by the following equation:

[0022]

[0023]

[0024]

[0025] in, For strain energy, It is inertial kinetic energy. For energy dissipation, This is the coupling strength parameter for the formation hardness coefficient. The coupling strength parameter is the pore pressure. These are the mechanical operating parameters. and Standard Rossler parameters for controlling the basic form of chaotic attractors; This indicates the derivative with respect to time.

[0026] Furthermore, the data-driven sub-model based on the adaptive spatiotemporal convolutional network can be represented by the following equation:

[0027] ,

[0028] in, For the predicted torque, For time parameters, This indicates the final predicted future torque; For dynamic fusion weights; This represents a lightweight adaptive spatiotemporal convolutional network. The fusion tensor processed by ASTCN; It is a pseudo-Hamiltonian function.

[0029] Furthermore, the dynamic fusion weights and pseudo-Hamilton function can be calculated using the following formula:

[0030] ,

[0031] ,

[0032] in, The gradient of the pseudo-Hamiltonian function. Let Frobenius norm be the matrix. The norm of the Hamiltonian function gradient reflects the rate of change of the system's energy. The standard deviation of the motor input power over a time window. The confidence factor for lithological data is high when the lithology is homogeneous and low when it is not. For high-dimensional feature tensors, is the energy transfer coefficient.

[0033] Furthermore, the intelligent torque prediction model also includes: S230, a dynamic adaptive mechanism based on a dual-timescale learning framework. The dynamic adaptive mechanism first uses a mutation detector based on a chaotic index to determine whether a new formation has been encountered. When the drill bit encounters a formation mutation, the system dynamics change drastically, causing the formation mutation detector to generate a detectable sharp pulse, thereby triggering a slow loop update. When a mutation is detected, the meta-learning optimizer is activated, and rapid model reconstruction is achieved through inner-outer two-layer optimization models.

[0034] Furthermore, a mutation detector based on chaos indicators can be represented by the following equation:

[0035]

[0036] in, This represents the estimated value of the local Lyapunov exponent at time t; For a specific moment, The symbol is for exponent. The symbol is for partial differentials. The length of a short time window; Let k represent the phase space state vector at discrete time step k. The Jacobian matrix represents the state transition of the system from time k to time k+1.

[0037] Furthermore, the meta-learning optimizer can be represented as follows:

[0038]

[0039] in, The expected value of the task. In the mission The parameters were updated after n steps of gradient descent were performed. These are the parameters of the neural network model, specifically the ASTCN neural network parameters of the data-driven model. This represents a specific training task sampled from all possible abrupt geological changes, such as a drilling process from shale to sandstone. For the overall distribution of tasks; For loss function, Indicates the task of updating the outer layer. The loss function is usually the error between the predicted torque and the actual torque, and in this embodiment it can be the Huber loss function. Task for updating the inner layer loss function, This is the single-step learning rate for the inner loop.

[0040] Furthermore, step S300 may also include: using the local Lyapunov exponent calculated by the intelligent torque prediction model to achieve vibration monitoring and early warning. This exponent is a magnifying glass for measuring the dynamic stability of the system and is extremely sensitive to impending instability. The sharp increase of this exponent usually occurs much earlier than the appearance of a significant peak on the FFT spectrum. This exponent can be used as an early warning signal to achieve rapid early warning.

[0041] Another aspect of the present invention provides an intelligent digital drive system for adapting to oil drilling conditions, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the intelligent digital drive method for adapting to oil drilling conditions as described above.

[0042] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0043] 1. This invention uses a multi-source sensor array to collect real-time data on the electrical and mechanical status of oil drilling and the dynamic changes in the downhole geological environment, and constructs a multi-source database to achieve comprehensive and real-time monitoring of drilling conditions. This overcomes the limitations of traditional methods that rely on experience and manual judgment, effectively reduces the deviation and guidance error rate during drilling, and improves the safety and efficiency of drilling operations.

[0044] 2. The intelligent torque prediction model of this invention adopts a nonlinear torque dynamic millisecond-level online prediction mechanism. Combining data from a multi-source database, it transforms multi-source heterogeneous measurement data into a standardized feature space. In conjunction with prior geological knowledge, it encodes well logging data into parameters with clear physical meaning, achieving effective data fusion and transformation. This provides comprehensive information support for subsequent predictions. At the same time, combined with a dual-stream architecture prediction engine, it achieves dynamic adaptation to complex formation conditions, overcomes the limitations of a single judgment basis, improves the accuracy and reliability of predictions, and realizes real-time estimation and high-precision prediction of current load torque. It overcomes the limitations of existing static models and qualitative analysis, and significantly improves the matching degree between drill bit torque prediction values ​​and actual conditions.

[0045] 3. By introducing a dual-timescale learning framework, this invention achieves millisecond-level real-time adaptation and second-level slow loop optimization, significantly improving the model's adaptability and learning efficiency, and effectively solving the real-time problem of downhole engineering parameter evaluation. Attached Figure Description

[0046] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0047] Figure 1 This is a flowchart of the method provided in Embodiment 1 of the present invention.

[0048] Figure 2 This is a timing diagram of the method provided in Embodiment 1 of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0050] Example 1

[0051] Figure 1 The flowchart of the intelligent digital drive method for adaptive oil drilling conditions in this embodiment is shown. Figure 2 This is the timing diagram in this embodiment, from Figure 1 As can be seen from this embodiment, it includes the following steps:

[0052] Step 1: The edge gateway controller of the drilling platform collects relevant data on oil drilling in real time through a multi-source sensor group. The edge gateway controller preprocesses the collected data to obtain a multi-source data stream, and sends the preprocessed multi-source data stream to the cloud to build a multi-source database.

[0053] Specifically, the relevant data in this embodiment includes: data that can comprehensively and in real-time reflect the electrical and mechanical status of oil drilling operations, and even the dynamic changes in the downhole geological environment. For example, electrical condition data, mechanical condition data, and geological condition data.

[0054] Electrical operating condition data can be collected by current transformers and voltage sensors installed on the incoming side of the power electronic controller or at the motor junction box, which collect the three-phase stator current and bus voltage of the motor. Rotor position data is obtained by a rotary encoder installed on the motor shaft.

[0055] Mechanical operating condition data can be collected by MEMS accelerometers installed near the top drive unit or rotary table to obtain vibration acceleration data of the drill string. This data directly reflects the stability of the drilling process. Real-time data on drilling pressure, rotational speed, and mechanical drilling speed are obtained through the drill rig instrument system interface or dedicated pressure sensors. These data are key parameters reflecting drilling load and efficiency.

[0056] Geological condition data can be obtained in real time through the MWD (Log While Drilling) system interface of the edge gateway, such as sonic transit time and natural gamma, which are used to sense the formation lithology, hardness and pore pressure in real time.

[0057] Step 2: Build an intelligent torque prediction model in the cloud to achieve millisecond-level online prediction of nonlinear torque dynamics. This model can combine data from multiple source databases to achieve real-time estimation of the current drill bit load torque and obtain a high-precision drill bit predicted torque value.

[0058] Specifically, in this embodiment, the intelligent torque prediction model can be constructed through the following steps:

[0059] 1) First, the received multi-source heterogeneous measurement data is transformed into a standardized feature space that can uniformly reveal the internal dynamic state of the drilling system.

[0060] This standardized feature space can be constructed through the following steps:

[0061] First, the core dynamic torque time-series signal of the drill bit is reconstructed in phase space to obtain the reconstructed dynamic state vector. Then, the logging-while-drilling data curve is encoded into parameters with clear physical meaning according to the formation physical parameters to obtain the encoded formation physical parameters. Finally, the reconstructed dynamic state vector and the encoded formation physical parameters are fused to form a cooperative feature tensor (i.e., a standardized feature space) that includes the current motion state of the system and the physical constraints of the environment, providing comprehensive information input for the subsequent prediction engine.

[0062] Generally, torque signals during drilling typically exhibit strong nonlinear and chaotic characteristics. Existing technologies usually capture and predict torque signals within a future time window by analyzing their time-series values, but this method struggles to capture the essential patterns of torque signals during drilling. Therefore, in this implementation, the hidden dynamic structure is revealed by unfolding the one-dimensional time-series signal of multi-source heterogeneous measurement data into a high-dimensional phase space.

[0063] Specifically, for the core dynamic torque time-series signal, a time-delay coordinate embedding technique can be used for phase space reconstruction. This involves combining the current torque value with torque values ​​from several historical time points into a high-dimensional vector. This method allows the topological structure of a high-dimensional system to be recovered from a one-dimensional time series, providing a complete state description for subsequent dynamic modeling. In this embodiment, the embedding dimension and delay time are not arbitrarily chosen but adaptively determined using mathematical tools such as the Cao method and minimizing mutual information to ensure that the reconstructed phase space most accurately reflects the attractor morphology of the original dynamic system.

[0064] For example, in this embodiment, the phase space reconstruction of the torque signal can be expressed by the following formula:

[0065]

[0066] in, This is the phase space state vector, which is also the reconstructed dynamic state vector. For time, Indicates time The phase space state vector reconstructed in time; This is the torque value. This represents the torque value at the current moment, which is a single projection of the dynamic behavior of the drilling process. The torque value is the result of a past delay time. For the torque value at a specific point in the past, each such value represents a snapshot of the torque value at a historical moment; This is the embedding dimension, which can be calculated using pseudo-nearest neighbor methods such as the Cao method; For delay time; A collection representing the embedded dimensions; This is the symbol for the transpose of a vector. The entire vector combines a series of discrete observations from the present and the past to form a point in an m-dimensional space. This point is no longer an isolated numerical value, but rather contains the state of the system's recent evolutionary history.

[0067] It should be noted that this state vector is constructed based on the Tukens embedding theorem. For a deterministic dynamic system, even if only one variable of the system can be observed, a phase space that is topologically equivalent to the original system's dynamic behavior can still be reconstructed by embedding the delayed coordinates of that variable. In other words, although the real drill string-formation system is determined by countless variables (temperature, pressure, vibration modes, etc.), the shape of its overall behavior (e.g., its chaotic attractor) can be recovered from just one time series of torque. The phase space reconstruction formula shown allows for the reconstruction of a one-dimensional time series... Unfold into a high-dimensional geometric object This allows the chaotic dynamics hidden behind the torque signal to be explicitly expressed and analyzed mathematically, providing an initial premise for subsequent chaos analysis.

[0068] Meanwhile, to enable the model to better understand the impact of formation characteristics on drill bit dynamics, it is necessary to encode the logging-while-drilling (MWD) data into parameters with explicit physical meaning. In this implementation, by introducing prior geological knowledge, the original logging curves are transformed into physical coefficients that directly influence the drilling process. For example, formation hardness is a key factor affecting drill bit rock-breaking efficiency, and the natural gamma value is usually positively correlated with clay content, thus indirectly reflecting rock hardness. Therefore, a mapping function can be constructed to transform the natural gamma into a dimensionless formation hardness coefficient; similarly, sonic transit time (Δtc) is closely related to rock porosity, and porosity directly affects pore pressure, so a conversion relationship from sonic transit time to pore pressure gradient can be established.

[0069] For example, in this embodiment, the physical coding formula can be calculated using the following formula:

[0070]

[0071]

[0072] in, This is the formation hardness coefficient, a dimensionless normalized value. It is a natural exponential function. A coefficient that controls the steepness of the function curve. This is the natural gamma value measured while drilling. The gamma baseline value is used as a reference stratum. For the Nabla operator, Pore ​​pressure, The pore pressure gradient; The acoustic transit time conversion factor is determined based on regional geological data. For sound wave time difference, These are the linear transformation coefficients determined based on regional geological data.

[0073] It should be noted that the goal of the two formulas mentioned above is to transform raw, abstract well logging data into parameters with clear physical meaning that can be directly used by physical models. For formation hardness coefficients, the natural gamma value is usually related to clay content; high gamma values ​​typically correspond to mudstone (softer lithology), while low gamma values ​​correspond to sandstone or carbonate rocks (harder lithology). Encoding the gamma value using the sigmoid function transforms this qualitative geological knowledge into a continuous, normalized hardness parameter, allowing it to be directly embedded as a multiplier term in subsequent kinetic equations. For pore pressure gradients, an empirical rock physics model is used—a common engineering method that utilizes readily available sonic data for a first-order approximation when pore pressure cannot be directly measured. In this embodiment, the two formulas shown encode well logging curves into values ​​that can be directly calculated by physical equations.

[0074] 2) After completing the standardized feature space construction, in order to achieve accurate prediction of torque signals, it is necessary to build a core prediction engine that can simulate the complex interaction between the formation and the drill string. This prediction engine can combine the physical mechanism of the feature space with data-driven approach to achieve high-precision prediction through deep collaboration.

[0075] In this embodiment, the prediction engine employs a dual-stream architecture that integrates a physical model and a data-driven model. The physical model, based on mechanical principles, constructs a dynamic model capable of generating chaotic torque signals to analyze the essential physical characteristics; while the data-driven model, through a deep learning network, autonomously learns from the high-dimensional features output by the physical model and captures complex patterns that the physical model fails to fully describe.

[0076] Specifically, in this embodiment, the physical model can be constructed based on the Rossler model, a common chaotic system. The control parameters in the original Rossler model are replaced or coupled with lithological physical parameters derived from drilling data. This allows mechanical operating parameters such as drilling rate and drilling pressure to be used as external driving terms of the model. The formation hardness coefficient and pore pressure gradient obtained in the previous step are used as intrinsic parameters of the state evolution equation to adjust the endogenous variables of the model's internal evolution law. This results in a set of differential equations constrained by physical laws. The evolution trajectory of the solution to this set of equations in phase space simulates the physical evolution process of the torque state.

[0077] For example, in this embodiment, the differential phase space physical model based on the Rossler model can be expressed by the following equation:

[0078]

[0079]

[0080]

[0081] in, For strain energy, It is inertial kinetic energy. For energy dissipation, This is the coupling strength parameter for the formation hardness coefficient. The coupling strength parameter is the pore pressure. These are the mechanical operating parameters. and Standard Rossler parameters for controlling the basic form of chaotic attractors; This represents the derivative with respect to time. The final solution obtained from this expression is: This is the high-dimensional feature tensor of the reconstructed phase space state vector.

[0082] It should be noted that in the above formula, This is the part responsible for rotation and stretching in the standard Rossler model; it forms the basic framework for chaotic behavior, while the rest... This is a newly added physical driving term, and also the core modification to the Rossler model in this embodiment. It introduces externally applied mechanical energy as the driving force of the system, and the efficiency of this energy injection is modulated by the formation hardness coefficient; a larger hard formation leads to a stronger driving effect, thus simulating the scenario of a drill bit generating severe vibrations on hard rock. And... middle This is the key nonlinear term in the standard Rossler model that produces the folding effect, and it is the root cause of chaos. This is the second core modification to the Rossler model in this embodiment: the pore pressure gradient is introduced into the nonlinear term to characterize how the pore pressure of the rock affects its fracturing mode and the rate of energy release. High pore pressure may make the rock more fracturing, thereby changing the way the system dissipates energy and thus affecting the morphology of the chaotic attractor.

[0083] Specifically, in this embodiment, the data-driven model can directly learn to predict future torque from the fused high-dimensional feature tensor through a lightweight adaptive spatiotemporal convolutional network (ASTCN). To enable these two models to complement each other, an organic fusion mechanism needs to be designed. In this embodiment, the fusion mechanism follows the principles of energy conservation and system uncertainty. It constructs a pseudo-Hamiltonian function representing the total energy of the system, and its gradient in phase space measures the degree to which the system deviates from its steady state. When the system fluctuates drastically (i.e., the gradient is large) or encounters high uncertainty in geological information, the fitting ability of the data-driven model should be relied upon more; conversely, when the system operates smoothly and the geological characteristics are clear, the physical model with clear physical meaning should be trusted more.

[0084] For example, in this embodiment, the data-driven model can be represented by the following formula:

[0085] ,

[0086] in, For the predicted torque, For time parameters, This indicates the final predicted future torque; For dynamic fusion weights; This represents a lightweight adaptive spatiotemporal convolutional network. The fusion tensor processed by ASTCN; It is a pseudo-Hamiltonian function.

[0087] For example, in the data-driven model of this embodiment, the dynamically fused weights and pseudo-Hamilton function can be calculated using the following formula:

[0088] ,

[0089] ,

[0090] in, The gradient of the pseudo-Hamiltonian function. Let Frobenius norm be the matrix. The norm of the Hamiltonian function gradient reflects the rate of change of the system's energy. The standard deviation of the motor input power over a time window. The confidence factor for lithological data is high when the lithology is homogeneous and low when it is not. The energy transfer coefficient is denoted as . The dynamic fusion weight formula uses a Sigmoid function to transform the uncertain gradient of the physical model, external perturbations, and the unreliability of the data source into a smooth weight, thereby achieving an intelligent and dynamic trade-off between physical mechanisms and data-driven approaches.

[0091] 3) The core prediction engine is still essentially a static prediction model, which cannot well cope with the complex formation conditions in actual drilling. After the prediction engine is established, in order to achieve millisecond-level online adaptation to cope with problems such as drill bit wear caused by sudden formation changes, this embodiment introduces an online, fast self-evolving adaptive mechanism to enable the model to achieve adaptive learning.

[0092] In this embodiment, considering that geological changes require a response time within seconds (slow loop) while torque fluctuations require real-time compensation (fast loop), a dynamic adaptive mechanism is constructed by introducing a dual-timescale learning framework to achieve millisecond-level online adaptation. In the millisecond-level fast loop, this learning framework uses an extended Kalman filter as the fast loop kernel to fine-tune the mechanical coupling coefficients in the physical model in real time, enabling it to quickly track short-term operating condition changes. In the second-level slow loop, a meta-learning framework is used as the slow loop kernel to rapidly reshape the entire neural network.

[0093] Specifically, in this embodiment, the mechanism first uses a mutation detector based on chaos indicators to determine whether a new formation has been encountered. When the drill bit encounters a formation mutation, the dynamic behavior of the system will change drastically, causing the formation mutation detector to generate a sharp pulse that can be detected, thereby triggering a slow cycle update.

[0094] Once a mutation is detected, the meta-learning optimizer is activated. Unlike traditional online learning, this model is trained on various geological change tasks during the pre-training phase. Through inner-outer two-layer optimization, the initial parameters learned by the model are forced to a specific position: from this position, regardless of the new geological formation encountered next, only minor adjustments (one or two steps of gradient descent) are needed to quickly achieve good performance. This enables rapid model reconstruction for geological mutations within seconds or even less, greatly improving the model's adaptability.

[0095] For example, in this embodiment, the formation abrupt change detector can be represented by the following formula:

[0096]

[0097] in, This represents the estimated value of the local Lyapunov exponent at time t; For a specific moment, The symbol is for exponent. The symbol is for partial differentials. The length of a short time window; Let k represent the phase space state vector at discrete time step k. The Jacobian matrix represents the state transition of the system from time k to time k+1.

[0098] It should be noted that the formation abrupt change detector in this embodiment is based on the Lyapunov exponent. The interaction between the drill bit and the rock is essentially a complex dynamic process. When the drill bit is drilling through homogeneous formations (such as large sections of mudstone), the entire dynamic system is in a relatively stable state, which can be imagined as a system operating on a specific "attractor," and its dynamic behavior (such as sensitivity and degree of chaos) has statistical consistency. However, when the drill bit suddenly drills from one type of lithology to another completely different type (for example, from soft mudstone to hard and dense limestone), the physical laws governing this dynamic system will fundamentally change. This will inevitably lead to a change in the structure of the attractor of the dynamic system, manifested as a dramatic change in the local divergence / convergence characteristics of the phase space orbit. Therefore, the value of the Lyapunov exponent becomes a perfect formation abrupt change detector. When it increases sharply, it indicates that the underlying dynamic laws of the system have changed, and the existing model parameters may no longer be applicable. This signal can be used as a trigger condition to activate deeper model adaptation mechanisms, such as meta-learning in slow loops.

[0099] For example, in this embodiment, the slow-loop updating meta-learning optimizer can be calculated using the following formula:

[0100]

[0101] in, The expected value of the task. In the mission The parameters were updated after n steps of gradient descent were performed. These are the parameters of the neural network model, specifically the ASTCN neural network parameters of the data-driven model. This represents a specific training task sampled from all possible abrupt geological changes, such as a drilling process from shale to sandstone. For the overall distribution of tasks; For loss function, Indicates the task of updating the outer layer. The loss function is usually the error between the predicted torque and the actual torque, and in this embodiment it can be the Huber loss function. Task for updating the inner layer loss function, This is the single-step learning rate for the inner loop.

[0102] It should be noted that in the above formula, The outer layer optimization is the core of the entire optimization process. Its goal is not to minimize the model's loss on the current task, but rather to minimize the model's loss on the new task after performing the inner layer update. In other words, the object of the outer layer optimization is a parameter obtained through the inner layer update. Its optimization objective is not to minimize the model's direct performance on any single task, but rather to minimize its expected performance after rapid adaptation. The inner layer update section describes a simulated, rapid fine-tuning process for any given formation mutation task. The inner update calculates the loss using a small amount of data from the task and adjusts the current model parameters along the gradient direction. Perform one or more updates to obtain a temporary set of new parameters adapted to the task. In other words, the essence of inner layer updates is to simulate how the model will quickly self-adjust when it actually encounters new formations online.

[0103] Step 3: The cloud sends the predicted torque value calculated by the intelligent torque prediction model to the edge gateway controller. The edge gateway controller then generates the actual control commands based on the received predicted torque value.

[0104] Vibration monitoring and early warning are achieved based on the local Lyapunov exponent calculated by the intelligent torque prediction model. This exponent is a magnifying glass for measuring the dynamic stability of the system and is extremely sensitive to impending instability (such as stick-slip vibration). The sharp increase of this exponent usually occurs much earlier than the obvious peak on the FFT spectrum. This exponent can be used as an early warning signal to achieve rapid early warning.

[0105] Furthermore, by setting a vibration suppression threshold, when the monitored index exceeds the threshold, the edge gateway controller automatically superimposes a compensation torque command on the generated control command. This compensation torque command is a sine wave or a pulse width modulation (PWM) signal that is in the same frequency as the main vibration frequency but out of phase. Its amplitude is positively correlated with the monitored vibration amplitude or the degree of exceeding the Lyapunov exponent.

[0106] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent digital drive method for electric oil drilling that adapts to different oil drilling conditions, characterized in that, The intelligent digital driving method includes, S100: Collects oil drilling data through a multi-source sensor group, sends the collected data to the edge gateway controller for preprocessing, and sends the preprocessed multi-source data stream to the cloud to build a multi-source database; The S200, with its cloud-based intelligent torque prediction model, uses data from multiple databases to estimate the current drill bit load torque in real time and obtain the predicted drill bit torque value. The intelligent torque prediction model includes: The feature space used to reflect the essential laws of torque signals during drilling. Used to encode logging-while-drilling data into physical sub-models with parameters of definite physical meaning. And prediction engines that make predictions based on feature space and physical meaning parameters, The prediction engine employs a dual-stream architecture that integrates a physical sub-model and a data-driven sub-model. The physical sub-model is a set of differential equations constrained by physical laws, constructed based on the Rossler model. It is used to analyze the essential physical characteristics and obtain the high-dimensional features of drill bit torque. The data-driven sub-model, through deep learning networks, autonomously learns and captures complex patterns that the physical sub-model fails to fully describe from high-dimensional features. The S300 and cloud-based systems will send the calculated predicted torque values ​​to the edge gateway controller. The edge gateway controller generates actual control commands based on the received predicted torque value; The intelligent torque prediction model also includes: A dual-timescale learning framework was designed to address the problems caused by abrupt changes in geological formations. The dual-timescale learning framework includes: A fast loop kernel based on extended Kalman filtering is used to fine-tune the physical sub-model in real time, enabling it to quickly track short-term operating condition changes. The slow-loop kernel, with meta-learning at its core, enables the rapid reshaping of deep learning networks that drive data-driven sub-models. The intelligent torque prediction model is constructed according to the following steps: S210. Transform multi-source heterogeneous measurement data into a standardized feature space that can reveal the intrinsic dynamic state of the drilling system. Meanwhile, by introducing prior geological knowledge, the original logging curves are transformed into physical coefficients that directly influence the geological process. By constructing a formation hardness mapping function, the natural gamma is transformed into a dimensionless formation hardness coefficient, and the conversion relationship from sonic transit time to pore pressure gradient is established. S220. To achieve accurate prediction of torque signals, a prediction engine is constructed through a dual structure of physical and data components. The physical sub-model of the prediction engine is constructed based on the Rossler model. By coupling the control parameters in the original Rossler model into physical coefficients derived from well logging curves, and by using mechanical operating parameters as external driving terms of the model, Furthermore, physical coefficients are used as intrinsic parameters of the state evolution equation to adjust the endogenous variables of the internal evolution law of the Rossler model; The data-driven sub-model predicts future torque based on the high-dimensional feature tensor calculated by the physical sub-model using a lightweight adaptive spatiotemporal convolutional network. The intelligent torque prediction model also includes: S230, A dynamic adaptive mechanism based on a dual-timescale learning framework. The aforementioned dynamic adaptive mechanism first uses a mutation detector based on chaotic indices to determine whether new formations have been encountered during drilling. When the drill bit encounters a sudden change in formation, the dynamic behavior of the system changes drastically, causing the formation change detector to generate a sharp pulse that can be detected, thereby triggering a slow cycle update. When a mutation is detected, the meta-learning optimizer is activated, which optimizes the model through inner and outer layers to achieve rapid model reconstruction.

2. The intelligent digital drive method for adaptive oil drilling conditions according to claim 1, characterized in that, The multi-source data includes: electrical condition data, mechanical condition data, and geological condition data; The electrical operating data includes: three-phase stator current and bus voltage of the motor, as well as rotor position data. Mechanical operating data includes: drill string vibration acceleration data, real-time drill pressure, rotational speed, and mechanical drilling speed data. Geological condition data includes real-time logging data obtained through the logging-while-drilling system interface.

3. The intelligent digital drive method for adaptive oil drilling conditions according to claim 1, characterized in that, The standardized feature space is constructed through the following sub-steps: S211. The core dynamic torque timing signal of the drill bit is reconstructed in phase space to obtain the reconstructed dynamic state vector. S212. Based on the physical parameters of the formation, the logging-while-drilling data curves are encoded into parameters with clear physical meanings to obtain the encoded formation physical parameters. S213. The reconstructed dynamic state vector is fused with the encoded formation physical parameters to form a standardized feature space that includes the current motion state of the system and the physical constraints of the environment.

4. The adaptive oil drilling condition-based intelligent digital drive method for oil replacement, as described in claim 1, is characterized in that: The formation hardness mapping function is constructed based on the Sigmoid function, which encodes the natural gamma value and transforms it into a continuous, normalized hardness parameter. The conversion relationship between acoustic transit time and pore pressure gradient is based on a rock physics sub-model using acoustic data for a first-order approximation, thereby realizing the conversion between acoustic transit time and pore pressure gradient.

5. The adaptive oil drilling condition-based intelligent digital drive method for oil replacement, as described in claim 1, is characterized in that, The physical sub-model and the data-driven sub-model complement each other through a fusion mechanism. The fusion mechanism includes dynamic fusion weights and a pseudo-Hamilton function. The dynamic fusion weights use a sigmoid function to transform the uncertain gradients of the physical sub-model, external perturbations, and the unreliability of the data source into a smooth weight, thereby achieving a dynamic trade-off between physical mechanisms and data-driven approaches. The pseudo-Hamiltonian function follows the principles of energy conservation and system uncertainty. By constructing a pseudo-Hamiltonian function that characterizes the total energy of the system, the magnitude of the gradient of the pseudo-Hamiltonian function in phase space is used to measure the degree to which the system deviates from steady state. When the system experiences severe fluctuations or encounters high uncertainty in stratigraphic information, the fitting ability of the data sub-model is trusted. When the system operates smoothly and the formation characteristics are clear, then a physical sub-model with clear physical meaning is trusted.

6. The intelligent digital drive method for adaptive oil drilling conditions according to claim 3, characterized in that, The reconstructed dynamic state vector is obtained by phase space reconstruction using time-delayed coordinate embedding. By combining the torque value at the current moment with the torque values ​​at multiple historical moments into a high-dimensional vector, the topology of the high-dimensional system can be recovered from the one-dimensional time series, thus providing a complete state description.

7. An intelligent digital drive system for electric oil drilling that adapts to different oil drilling conditions, characterized in that: The intelligent digital drive system includes: processor; The memory stores a computer program that, when executed by a processor, implements the intelligent digital drive method for adaptive oil drilling conditions as described in any one of claims 1 to 6.

Images