Data-mechanism fusion method for inversion of bottom hole pressure and torque in rotary drilling

Abstract

This invention relates to the field of coal mine underground drilling engineering technology, specifically to a data-mechanism fusion method for inverting bottom hole pressure and torque in rotary drilling. First, a mechanistic model is established to calculate and output predicted values ​​of bottom hole pressure and bottom hole torque based on current drilling parameters. Then, the following data are organized into the input feature vector of the data model: sampling time; real-time drilling speed; current drilling depth; feed pressure; power head rotation speed; predicted bottom hole pressure; actual bottom hole pressure; predicted bottom hole torque; actual bottom hole torque. After receiving the above inputs, the data model extracts temporal features through an LSTM layer and performs nonlinear mapping through a BP layer, outputting the predicted residuals of bottom hole pressure and bottom hole torque at the current moment. Finally, the predicted residuals are superimposed with the mechanistic prediction values ​​to obtain the inverted results of bottom hole pressure and bottom hole torque. This method can be applied to coal mine underground drilling rig monitoring systems.

Description

Technical Field

[0001] This invention relates to the field of coal mine drilling engineering technology, specifically to a data-mechanism fusion method for inverting bottom drilling pressure and torque in rotary drilling. Background Technology

[0002] In coal mine underground gas extraction, water exploration and drainage, and high-level directional drilling of the roof, horizontal rotary drilling is the mainstream construction method. Weight on bit (WOB) and torque on bit (TOB), as core dynamic parameters for rock breaking by the drill bit at the bottom of the hole, are crucial for optimizing drilling parameters, preventing stuck drill bits, and improving drilling efficiency. However, due to the length of the drill string (hundreds of meters or even thousands of meters), the feed pressure and rotational torque applied by the surface drilling rig experience severe attenuation, lag, and nonlinear distortion during transmission to the bottom of the hole. This results in the parameters displayed on the borehole instruments failing to accurately reflect the actual stress state at the bottom of the hole. Therefore, achieving high-precision real-time inversion of bottom-hole weight on bit and torque has become a pressing technical challenge in the field of underground coal mine drilling.

[0003] Currently, scholars both domestically and internationally have conducted extensive research on downhole parameter inversion, mainly resulting in three technical approaches: The first type is the indirect estimation method based on orifice parameters. This method relies on the classic drill string mechanics model and inversely estimates the bottom hole values ​​by deducting the friction between the drill string and the borehole wall. For example, Chinese patent CN116362143A discloses a drill string friction analysis method, which uses a gated cyclic unit neural network to establish a bottom hole drilling pressure torque prediction model and combines it with a friction calculation theoretical model to inversely calculate the friction coefficient. However, this type of method is highly dependent on the accuracy of the friction coefficient. In horizontal boreholes in coal mines, the lithology of the rock strata changes drastically, the borehole wall has high roughness and is prone to collapse, and rock cuttings in the horizontal section easily accumulate to form rock cuttings beds, resulting in the friction coefficient not only being extremely unevenly distributed along the borehole depth but also exhibiting strong time-varying characteristics. Existing estimation models are mostly based on empirical friction coefficients or simple Coulomb friction assumptions, which are difficult to accurately characterize the above complex working conditions. The estimation error is usually above 30%, which is insufficient to meet the requirements of precise construction.

[0004] The second category is direct measurement based on measurement while drilling. This method directly acquires drill pressure and torque by installing a measuring sub near the drill bit and transmitting them to the wellhead via mud pulses or through-the-hole drill pipe. However, underground coal mine drilling largely uses water or pneumatic slag removal, lacking a stable mud pulse channel; while through-the-hole drill pipe can transmit signals, the near-the-bit measuring sub is expensive, costing tens to hundreds of thousands of yuan per set, and its lifespan is extremely short in harsh environments with strong vibrations, high water pressure, and coal dust intrusion, requiring frequent maintenance and replacement, making it difficult to widely promote in underground coal mines.

[0005] The third category is purely data-driven methods. In recent years, some researchers have attempted to use neural network models to directly predict bottom hole parameters from wellhead logging parameters. For example, Chinese patent CN119333123A discloses a downhole parameter prediction model based on multi-source data, which uses LSTM, BP, and other networks to learn the mapping relationship between surface data and bottom hole parameters; while CN121524935A further proposes a mechanism-deep learning hybrid mechanical drilling rate prediction method, which uses a CNN-GRU-Attention network to compensate for the residuals of the mechanism model. However, purely data-driven models rely entirely on statistical regularities in historical data and lack constraints on the physical mechanisms of drill string dynamics. When faced with new drilling sites, new formations, or sudden changes in operating conditions, their generalization ability drops sharply, and the prediction results may violate basic physical laws, such as predicting negative torque or sudden spikes in drilling pressure. Although the latter introduces a mechanistic model as a benchmark, its mechanistic model mainly considers drill bit wear and bottom hole temperature, and the prediction target is mechanical drilling speed. It does not involve the inversion of drilling pressure and torque, and it does not conduct special dynamic modeling for the key influencing factor of time-varying friction caused by cuttings bed in horizontal holes.

[0006] In summary, there is currently no existing technology that can comprehensively utilize drill string dynamics and drilling time-series data to achieve high-precision real-time inversion of borehole pressure and torque in horizontal rotary drilling in coal mines, relying solely on existing sensors on the drilling rig. There is an urgent need in this field for a borehole parameter inversion scheme that balances physical constraints with data-driven adaptive capabilities, specifically models cuttings beds and time-varying friction in horizontal holes, and does not rely on expensive downhole measurement tools. Summary of the Invention

[0007] This invention aims to address the shortcomings of existing technologies in horizontal rotary drilling in coal mines, including insufficient accuracy in bottom hole pressure and torque inversion, poor generalization ability of purely data-driven models due to lack of physical constraints, limited ability of mechanistic models to characterize cuttings beds and time-varying friction in horizontal sections, and the difficulty in large-scale application of expensive downhole measurement-while-drilling (MWD) tools. It provides a data- and mechanism-driven method for bottom hole pressure and torque inversion. By establishing a residual-compensated fusion architecture between a three-degree-of-freedom lumped mass method mechanistic model considering the characteristics of cuttings removal in horizontal holes and an LSTM-BP neural network data model, it achieves high-precision real-time inversion of bottom hole pressure and torque using only existing sensors on the drilling rig.

[0008] To achieve the above objectives, the present invention provides a method for bottom hole drilling pressure and torque inversion driven by data and mechanism fusion for horizontal rotary drilling in coal mines. This method can be executed by the on-site monitoring system of the coal mine drilling rig. The monitoring system includes a data acquisition module, a mechanism model calculation module, a data model calculation module, and a fusion output module.

[0009] The method includes the following steps: S1: Construct a three-degree-of-freedom lumped mass method mechanism model for horizontal rotary drilling in coal mines.

[0010] Based on the lumped mass method, the drill string system is discretized along the borehole trajectory curve into a lumped mass model with three degrees of freedom: axial, torsional, and lateral. Drill string dynamics equations considering drill bit cutting characteristics, drill pipe friction mechanisms, and horizontal hole cuttings removal mechanisms are established to calculate predicted values ​​for bottom hole pressure and bottom hole torque mechanisms. Furthermore, the mechanism model incorporates a horizontal hole cuttings removal mechanism term that considers the additional friction caused by cuttings bed deposition.

[0011] Specifically, the drill bit cutting characteristics decompose the drill bit-rock interaction into cutting components and frictional components, among which... Drilling pressure cutting component , Torque cutting component , Drilling pressure friction component , Torque friction component , Where Wc is the cutting component of the bottom hole drilling pressure, Tc is the cutting component of the bottom hole torque, Wf is the friction component of the bottom hole drilling pressure, and Tf is the friction component of the bottom hole torque. ε is the inherent strength of the rock stratum; a is the drill bit radius; ζ is the angular parameter of the drill bit; σ is the contact strength; l is the geometric parameter of the drill bit; μ is the friction coefficient between the drill bit and the rock stratum; γ is the geometric parameter of the drill bit. R(·), H(·), and S(·) represent the Ramp function, Heaviside function, and Sign function, respectively. d(t) represents the instantaneous cutting thickness of the drill bit; Indicates the angular velocity of the drill bit rotation; This indicates the drill bit feed rate.

[0012] Specifically, to address the additional friction caused by the deposition of rock cuttings in horizontal boreholes, a slag discharge mechanism correction term based on a two-layer flow model is introduced into the aforementioned mechanism model. A slag discharge friction calculation model incorporating the self-weight and circumferential distribution of the coal slag is established, and its relevant calculation formulas are shown below:

[0013]

[0014] Where Tfw represents the frictional torque generated by the drill pipe during the rotation of the slag; Ffw represents the additional frictional force of the slag bed on the drill string; Fcw represents the contact force between the slag bed and the wellbore; fws represents the coefficient of friction between the slag bed and the drill string; θ represents the contact angle between the slag bed and the wellbore; and R represents the outer radius of the drill pipe.

[0015] The dynamic differential equations are numerically solved using the fourth-order Runge-Kutta method, and the predicted values ​​of the bottom hole drilling pressure mechanism and the bottom hole torque mechanism are output in real time.

[0016] S2: Construct an LSTM-BP neural network data model.

[0017] The first stage uses an LSTM neural network for temporal dynamic feature extraction, and the hidden state output of the last time step of the LSTM is passed as a feature vector to the BP neural network. The BP neural network is used as the second stage for nonlinear mapping and feature fusion, and outputs the bottom hole pressure prediction residual and the bottom hole torque prediction residual.

[0018] The inputs to the LSTM network include: sampling time, real-time drilling speed, drilling depth, feed pressure, power head rotation speed, and the predicted values ​​of bottom hole drilling pressure and torque mechanism output by the aforementioned mechanistic model, along with their corresponding actual values. The above input data is sampled using a fixed time window to construct a time-series sample sequence.

[0019] The output mode of the LSTM network is set to last, meaning that only the hidden state of the last time step is retained. The BP neural network contains at least one fully connected layer, and each fully connected layer is followed by a ReLU activation function and a Dropout layer.

[0020] Optionally, an attention mechanism layer can be connected after the LSTM network to dynamically weight the features at different time steps and pass the weighted feature vector to the BP neural network.

[0021] Optionally, the LSTM network can be replaced with a gated recurrent unit network, and / or the BP network can be replaced with a convolutional neural network.

[0022] S3: Establish a residual compensation fusion architecture.

[0023] The prediction residuals output by the data model are superimposed onto the predicted values ​​of the bottom hole drilling pressure mechanism and the bottom hole torque mechanism output by the mechanism model in an additive compensation manner to obtain the final predicted values ​​of the bottom hole drilling pressure and the bottom hole torque.

[0024] S4: Model Training and Online Applications.

[0025] During the model training phase, a training set was constructed using complete logging data from drilled boreholes, and time-series samples were generated using a sliding time window approach. The training process consisted of two stages: first, the LSTM-BP data model was trained separately, using the residuals output by the mechanistic model as the learning objective; then, the mechanistic model and the data model were jointly fine-tuned to optimize the overall performance of the fused model. Mean squared error was used as the loss function, and the Adam optimizer was used for parameter updates. The initial learning rate was set to 0.001 with an exponential decay strategy, and early stopping was employed to prevent overfitting.

[0026] Optionally, a hybrid model update strategy is also adopted: when a single or vertical shaft is drilled, a real-time incremental update is performed, the fully connected layer is locked, and the hidden layer is updated only using real-time drilling sequence data, and one training cycle is performed; when tripping in and out of the drill string, a periodic full update is performed, collecting all drilling data for the current tripping in and out of the drill string and updating the fully connected layer and hidden layer after random shuffling.

[0027] During the online inversion phase, the trained fusion model is deployed on the drilling rig's on-site monitoring system. The system collects parameters such as feed pressure, power head speed, and drill rod displacement from the drilling rig's sensors in real time. These parameters are then processed sequentially: the mechanistic model calculates the baseline predicted value, the data model calculates the residual correction value, and finally, the system outputs the real-time inversion results of the bottom hole drilling pressure and torque, which are then sent to the drilling rig's monitoring and display terminal.

[0028] The present invention also provides a coal mine downhole drilling pressure and torque inversion system applying the above method, comprising: The data acquisition module is used to acquire in real time the feed pressure, power head speed, drill rod displacement and slag discharge medium parameters of the drilling rig sensors during the drilling process; The mechanism model calculation module is used to run a three-degree-of-freedom lumped mass method model that considers the slag removal mechanism of horizontal holes, and output the mechanism prediction values ​​of bottom hole drilling pressure and torque. The data model calculation module is used to run the LSTM-BP neural network data model and output the predicted residuals of bottom hole drilling pressure and torque. The fusion output module is used to superimpose the mechanism prediction value and the prediction residual to output the final predicted values ​​of bottom hole drilling pressure and torque, and send them to the drilling rig monitoring and display terminal.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Significantly improves the accuracy and robustness of borehole bottom parameter inversion. This invention organically combines the physical constraints of the mechanistic model with the adaptive capabilities of data-driven approaches through a residual compensation fusion architecture. The three-degree-of-freedom lumped mass method mechanistic model provides benchmark predictions that conform to the dynamic laws of the drill string; the LSTM-BP neural network data model is specifically designed to perform residual learning and compensation for complex factors such as time-varying friction and rock heterogeneity that are difficult for the mechanistic model to accurately characterize. The two complement each other, maintaining high-precision prediction even in scenarios with drastic changes in rock strata and sudden changes in working conditions.

[0030] (2) Accurately characterize the time-varying friction of cuttings bed in horizontal holes to solve the problem of inaccurate prediction in long horizontal sections. Most existing mechanistic models ignore or simply handle the cuttings discharge effect in horizontal holes, resulting in a sharp increase in the prediction error of drilling torque in long horizontal sections. This invention introduces a cuttings discharge mechanism correction term based on a bilaminar flow model into the lumped mass method mechanistic model, and establishes a cuttings discharge friction calculation model that includes the self-weight of coal slag and its circumferential distribution. It can dynamically capture the changes in additional friction torque and additional friction force caused by the accumulation and migration of cuttings bed, and make up for the shortcomings of traditional models in characterizing friction in horizontal sections from a mechanistic perspective.

[0031] (3) Fully utilize the dynamic features of the time series to achieve accurate learning of the prediction residuals. An LSTM-BP cascaded network structure is adopted. In the first stage, LSTM is used to extract the long-term temporal dependence between drilling parameters and the deviation of the mechanism model, effectively capturing the lag and cumulative effect of the changes in the bottom hole state. In the second stage, the BP network is used to perform nonlinear feature fusion and mapping to accurately output the prediction residuals at the current moment. Compared with a single static mapping model, this structure can more fully explore the dynamic evolution law of the drilling process and improve the accuracy and temporal continuity of residual prediction.

[0032] (4) It relies solely on existing sensors on the drilling rig, resulting in low cost and easy promotion. This invention only requires the use of existing standard sensor data such as feed pressure, power head speed, and drill pipe displacement on the drilling rig. It does not require the installation of expensive near-bit downhole measurement subs, nor does it rely on special communication facilities such as cable-connected drill pipes or mud pulse channels. It completely solves the pain points of existing near-bit direct measurement methods, such as high cost, easy instrument damage, and difficulty in large-scale promotion. It provides an economical and feasible technical solution for the large-scale popularization of real-time monitoring of bottom hole parameters in coal mines.

[0033] (5) It has online self-evolution capability and can adapt to complex and ever-changing working conditions. The model hybrid update strategy adopted in this invention combines real-time incremental updates after a single drill bit with periodic full updates during tripping in and out of the drill bit. This enables the fused model to continuously adapt to working condition drift caused by drilling site migration, formation changes, equipment wear, etc., avoids model performance decay over time, and ensures continuous high-precision inversion capability during long-term construction. Attached Figure Description

[0034] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will now be described in detail with reference to the accompanying drawings, wherein... Figure 1 The overall flowchart of the bottom hole drilling pressure and torque inversion method driven by data and mechanism fusion for horizontal rotary drilling in coal mines provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the physical structure of the three-degree-of-freedom lumped mass method mechanism model considering the horizontal hole slag discharge mechanism in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the LSTM-BP neural network data model in an embodiment of the present invention; Figure 4 This is a comparison curve of the predicted drilling pressure at the bottom of the hole and the measured value in an embodiment of the present invention; Figure 5 This is a comparison curve of the predicted torque at the bottom of the hole and the measured value in an embodiment of the present invention. Detailed Implementation

[0035] 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.

[0036] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures, and 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.

[0037] Example 1 This embodiment provides a method for bottom hole drilling pressure and torque inversion driven by data and mechanism fusion for horizontal rotary drilling in coal mines.

[0038] This method can be executed by the on-site monitoring system of an underground coal mine drilling rig, the monitoring system comprising: Data acquisition module: A PLC or explosion-proof data acquisition card that connects to the feed pressure sensor, power head speed encoder, and drill pipe displacement sensor; Mechanism model calculation module: Numerical solution program deployed in explosion-proof industrial control computer or embedded controller; Data model calculation module: LSTM-BP neural network inference engine deployed in industrial control computer; Fusion output module: A communication unit connected to the drilling rig monitoring display screen.

[0039] Reference Figure 1 The method includes the following steps: S1: Construct a three-degree-of-freedom lumped mass method mechanism model for horizontal rotary drilling in coal mines.

[0040] In this step, the monitoring system discretizes the drill string system along the borehole trajectory curve into a lumped mass model with three degrees of freedom: axial, torsional, and lateral, based on the lumped mass method, and establishes the drill string dynamic equation considering the drill bit cutting characteristics, drill pipe friction mechanism, and horizontal hole drill pipe slag removal mechanism.

[0041] Specifically, refer to Figure 2 The physical structure diagram shown simplifies the drill string system into three lumped mass units. The interaction between the drill bit and the rock is decomposed into cutting and friction components: (1) Modeling of drill bit cutting characteristics During drilling operations, the bottom hole pressure and bottom hole torque exerted by the downhole rock formations on the drill bit can be divided into two linear superpositions. One component, the cutting force and cutting torque, is proportional to the thickness dn of the rock formation being cut, and this component is called the cutting component. The other component, the cutting force and cutting torque, is related to the frictional force it experiences, and is called the frictional component.

[0042] The monitoring system calculates the bottom hole pressure and bottom hole torque at the drill bit according to the following formula:

[0043]

[0044] Where Wb is the bottom hole pressure at the drill bit, Wc represents the cutting component of the drill bit pressure, and Wf represents the friction component of the drill bit pressure; Tb is the bottom hole torque at the drill bit, Tc represents the cutting component of the drill bit torque, and Tf represents the friction component of the drill bit torque.

[0045] The specific calculation formulas for the cutting component and the friction component are as follows:

[0046]

[0047]

[0048]

[0049] Where Wc is the cutting component of the bottom hole drilling pressure, Tc is the cutting component of the bottom hole torque, Wf is the friction component of the bottom hole drilling pressure, and Tf is the friction component of the bottom hole torque. ε is the inherent strength of the rock stratum; a is the drill bit radius; ζ is the angular parameter of the drill bit; σ is the contact strength; l is the geometric parameter of the drill bit; μ is the friction coefficient between the drill bit and the rock stratum; γ is the geometric parameter of the drill bit. R(·), H(·), and S(·) represent the Ramp function, Heaviside function, and Sign function, respectively, and are defined as follows:

[0050]

[0051]

[0052] d(t) represents the instantaneous cutting thickness of the drill bit, and its mathematical expression is: Where N is the number of drill bit teeth; τ is the time required for the drill bit to rotate 2π / N angles in each tooth cycle, i.e., the time delay variable during the drilling process, and its mathematical expression is: .

[0053] Indicates: angular velocity of the drill bit rotation; Indicates: Drill bit feed depth; Indicates: Drill bit feed rate.

[0054] (2) Modeling of friction mechanism of horizontal hole drill rod The frictional torque generated between the drill pipe and the borehole wall rock strata during drilling includes axial sliding friction and circumferential rolling friction, which is mainly affected by the material and structure of the drill pipe itself, as well as the power input. To simulate the increased frictional resistance caused by the increased contact area between the drill pipe and the rock strata, and the increased contact area between the drill pipe and the borehole wall with increasing drilling depth, the monitoring system calculates the frictional torque between the drill pipe and the borehole wall using the following formula:

[0055] Where R is the outer radius of the drill pipe; V0 is the feed rate (drilling speed); V1 is the tangential velocity; μ is the friction coefficient between the drill bit and the rock formation; r is the inner radius of the drill pipe; and ρ is the density of the drill pipe material. It is the acceleration due to gravity; This refers to the drilling time.

[0056] The mathematical expression for the tangential velocity V1 is: ;in, This indicates the angular velocity of the drill bit's rotation.

[0057] (3) Modeling of the slag removal mechanism of horizontal hole drill rod Specifically, to address the additional frictional resistance caused by the deposition of rock cuttings in horizontal boreholes, the monitoring system introduces a slag discharge mechanism correction term based on a two-layer flow model into the mechanism model, establishing a slag discharge frictional resistance calculation model that considers the self-weight and circumferential distribution of coal slag. The relevant calculation formulas are as follows:

[0058]

[0059] Where Tfw represents the frictional torque generated by the drill pipe during the rotation of slag; Ffw represents the additional frictional force of the slag bed on the drill string; Fcw represents the contact force between the slag bed and the wellbore; fws represents the coefficient of friction between the slag bed and the drill string; and θ represents the contact angle between the slag bed and the wellbore.

[0060] When coal slag and rock cuttings accumulate at various points on the drill pipe, and the contact area with the drill pipe covers the lower semicircle of the drill pipe, the contact angle θ = The radial contact force of the coal slag on the drill pipe is sinusoidally distributed along the circumference, reaching its maximum at the center of the contact area (directly below the borehole) and smoothly transitioning to zero towards the two edges (θ=Π and θ=0); the self-weight of the coal slag per unit length is shared by the drill pipe and the borehole wall, with each bearing half; when the coefficient of friction between the coal slag and the drill pipe is constant; the additional frictional force of the coal slag bed on the drill string is expressed as: ;in, , , Ac is the cross-sectional area of ​​coal slag per unit length. Where A is the bulk density of coal slag, Ahole is the semi-circular area of ​​the borehole, and Apipe is the sector area of ​​the buried portion of the drill rod.

[0061] (4) Three-degree-of-freedom dynamic equations The monitoring system establishes a set of dynamic differential equations for each degree of freedom based on Newton's second law, and numerically solves the set of differential equations using the fourth-order Runge-Kutta method, outputting real-time predicted values ​​of the bottom hole drilling pressure mechanism. Predicted values ​​of borehole bottom torque mechanism .

[0062] Based on Newton's second law, a system of dynamic differential equations is established for each degree of freedom:

[0063] Where J1 and J2 represent the moments of inertia of the drill pipe and drill bit, respectively; c1 and c2, and k1 and k2 represent the torsional stiffness and damping of the drill string system, respectively; k0 and c0 represent the longitudinal stiffness and damping, respectively. The angular velocity of the drilling rig; , , The rotation angle of the drill pipe at the middle and its first and second derivatives are respectively calculated. , , Here, represents the rotation angle at the drill bit and its first and second derivatives, respectively; Tb is the drill bit torque; and M is the concentrated mass. Let be the drilling depth of the drill bit and its first and second derivatives.

[0064] Using the lumped mass method, the mass and moment of inertia of the drill pipe are uniformly distributed at both ends. The equivalent mass is the sum of half the mass of the bottom drill string assembly and half the mass of the drill pipe. The equivalent moment of inertia is... , Je is the differential factor, and Jp and Jb are the moments of inertia of the drill pipe and drill bit, respectively.

[0065] As an optional implementation, the three-degree-of-freedom model can be extended to more degrees of freedom, such as five or seven degrees of freedom, to further improve the accuracy of the mechanism model. In this case, the drill string is discretized into more elements, which are connected by stiffness and damping, and the form of its dynamic equations is correspondingly extended. In addition, the contact force distribution assumption in the slag discharge mechanism model can be adjusted to a uniform distribution, a parabolic distribution, or other functional forms according to the actual working conditions.

[0066] S2: Construct an LSTM-BP neural network data model.

[0067] The monitoring system constructs an LSTM-BP concatenated neural network as its data model, where the LSTM network is the first stage and the BP neural network is the second stage. (Refer to...) Figure 3 The input layer of the LSTM network receives time-series feature data, and the output layer of the BP network outputs the bottom hole pressure prediction residual and the bottom hole torque prediction residual.

[0068] (1) LSTM temporal feature extraction layer The input data for an LSTM neural network includes nine dimensions of features: sampling time. Real-time drilling speed Current drilling depth Pressure input Power head speed Bottom hole drilling speed Predicted values ​​of borehole bottom drilling pressure mechanism Actual value of drilling pressure at the bottom of the hole Predicted value of borehole bottom torque mechanism True value of borehole bottom torque The input data above is constructed as a time-series sample sequence in the form of a fixed time window, with a window length of... Pick One sampling point.

[0069] The core computation process of the LSTM network is as follows: · represents the outer product of vectors: ① Forget Gate. The output vector of the forget gate at time t is:

[0070] Where ft represents time t; σ is the Sigmoid activation function, which compresses the input value to the (0, 1) interval; Wf represents the weight matrix corresponding to the forget gate; bf is the bias vector corresponding to the forget gate; xt is the feature vector input to the LSTM unit at time t; ht-1 is the hidden state (output) of the LSTM unit at time t-1; [ht-1, xt] means concatenating the hidden state ht-1 of the previous time step and the input xt of the current time step into a longer vector.

[0071] ② Input gate. Determines which new information is stored in the cell state. The output vector it of the input gate at time t is:

[0072] Where Wi represents the weight matrix corresponding to the input gate; bi is the bias vector corresponding to the input gate.

[0073] ③ Candidate cell states. The vector of candidate cell states at time t. for:

[0074] Where tanh is the hyperbolic tangent activation function, which compresses the input value to the (-1,1) interval; WC represents the weight matrix corresponding to the candidate cell state; and bC represents the bias vector corresponding to the candidate cell state.

[0075] ④ Cell state update. Combine the old state Ct-1 with the outputs of the forget gate and the input gate to update it to the new cell state Ct:

[0076] Where Ct-1 represents the old cell state vector at time t-1; * indicates element-wise multiplication between vectors.

[0077] ⑤ Output gate. The output at time t is determined based on the cell state. At time t, the output vector ot of the output gate is:

[0078] Where Wo represents the weight matrix corresponding to the output gate; bo is the bias vector corresponding to the output gate. ⑥ Hidden State. The hidden state ht of the LSTM cell at time t is:

[0079] The monitoring system will hide the state of the last time step of the LSTM network. As a condensed temporal feature vector, it is passed to the subsequent BP neural network. The output mode of the LSTM layer is set to `last`, that is, only the output of the last time step is retained.

[0080] (2) Nonlinear mapping layer of BP neural network The backpropagation (BP) neural network serves as the second stage of this model, used for nonlinear mapping and feature fusion. When the hidden state is output by the LSTM... After being fed into the input layer of the BP neural network, the BP layer is constructed according to the preset number of hidden layers. The initial number of neurons in each layer is set to 64.

[0081] BP neural network reception As input, a nonlinear mapping is performed through at least one hidden layer, and a two-dimensional vector is output, corresponding to the bottom hole pressure prediction residuals. and hole bottom torque prediction residual :

[0082] Where fBP is the global mapping of the BP neural network, WBP is the weight of the BP neural network, and bBP is the bias of the BP neural network. .

[0083] Each fully connected layer is connected to a ReLU activation function and a Dropout layer in sequence. The expression for the ReLU activation function is: This can effectively alleviate the vanishing gradient problem; the Dropout layer, during training, uses probability... Randomly discard a portion of the neuron outputs, with the Dropout ratio set to 0.2, to prevent the model from overfitting to the training data.

[0084] As an optional implementation, the LSTM network can be replaced with a Gated Recurrent Unit (GRU) network to simplify the computational structure and reduce the number of parameters. The number of hidden layers, the number of neurons, and the dropout ratio of the BP network can be optimized using orthogonal experimentation; for example, three hidden layers can be set, with 128, 64, and 32 neurons respectively. The attention mechanism can be further added after the LSTM layers to dynamically weight features at different time steps.

[0085] S3: Establish a residual compensation-based fusion architecture.

[0086] The monitoring system uses the predicted values ​​of bottom hole drilling pressure and bottom hole torque from the S1 mechanistic model, along with the input features defined in S2, as input to the data model. The prediction residuals output by the data model are added to the baseline predicted values ​​output by the mechanistic model using additive compensation to obtain the final predicted value of bottom hole drilling pressure. Final predicted value of borehole bottom torque :

[0087]

[0088] The monitoring system sends the final predicted value to the drilling rig monitoring and display terminal for operators to refer to and adjust drilling parameters.

[0089] As an optional implementation, the residual compensation method is not limited to addition; weighted addition or multiplicative compensation can also be used. Furthermore, the fusion architecture can be further extended to multi-level fusion, where the intermediate states of the mechanistic model are also used as input to the data model to learn finer-grained residuals.

[0090] S4: Model Training and Online Applications.

[0091] The monitoring system performs the following operations during the model training phase: A training set is constructed using complete logging data from the drilled boreholes. Training set samples are generated using a sliding time window, with each sample containing continuous data. The sequence of input parameters for each time step. For underground drilling speeds in coal mines, this is typically... Sampling interval Second, Pick This is quite suitable. Measured values ​​of borehole bottom drilling pressure and torque are used as training labels.

[0092] The training process is divided into two phases: Phase 1: Train the LSTM-BP data model separately, using the residuals output by the mechanistic model (i.e., and The learning objective is to enable the data model to predict the patterns of residuals. Mean squared error (MSE) is used as the loss function. .

[0093] The second stage involves jointly fine-tuning the mechanistic model and the trained data model to optimize the overall performance of the fused model. The Adam optimizer is used for parameter updates, with the initial learning rate set to... It employs an exponential decay strategy with a decay rate of... Early stopping is used during training. Training is stopped when the validation set loss no longer decreases for 10 consecutive epochs to prevent overfitting.

[0094] During the online inversion phase, the monitoring system collects parameters such as feed pressure, power head speed, and drill rod displacement from the drilling rig sensors in real time. The parameters are then calculated sequentially using a mechanistic model to determine the baseline prediction value and a data model to determine the residual correction value. Finally, the system outputs the real-time inversion results of the bottom hole drilling pressure and torque.

[0095] As an optional implementation, the model update strategy can combine periodic full updates with real-time incremental updates: when a single or vertical shaft is drilled, the monitoring system collects the current drilling data, locks the fully connected layer, and updates the hidden layer only using real-time drilling sequence data, training for one round to achieve real-time incremental updates; when tripping in or out of the drill string, all drilling data for the current tripping in or out of the drill string are collected and randomly shuffled, and periodic full updates are initiated. This hybrid update strategy can simultaneously take into account both real-time changes in operating conditions and the stability of global patterns.

[0096] In this embodiment, the various modules of the monitoring system can be implemented through software, hardware, or a combination thereof. For example, the mechanism model calculation module and the data model calculation module can be integrated into the same industrial control computer and implemented using programming languages ​​such as C++ and Python; the data acquisition module can be connected to the sensor via RS485 bus, CAN bus, or Ethernet; and the fusion output module can be connected to the explosion-proof display screen via HDMI or VGA interfaces.

[0097] Example 2 In this embodiment, the BP neural network part of the data model is optimized for hyperparameters using orthogonal experimentation to adapt to the characteristics of coal mine drilling data.

[0098] 1. Orthogonal experimental design Four key hyperparameters were selected as experimental factors: Factor A: Number of hidden layers in BP, with horizontal layers set to 1, 2, or 3; Factor B: The number of neurons in each layer, using a decreasing structure, with levels set as (32), (64,32), and (128,64,32); Factor C: Dropout ratio, with levels set at 0.1, 0.2, and 0.3; Factor D: Activation function, set to Sigmoid, Tanh, or ReLU.

[0099] Select An orthogonal array was used for the experimental design, resulting in a total of 9 hyperparameter combinations. Each combination was repeated 3 times on the same training and validation sets, and the average value was used as the evaluation metric.

[0100] 3. Evaluation Indicators The root mean square error of the bottom hole pressure prediction on the validation set ( ) and root mean square error of borehole bottom torque prediction ( The weighted sum of the values ​​is used as a comprehensive evaluation index:

[0101] The baseline value is the result of a benchmark model with 32 neurons in a single hidden layer.

[0102] 3. Experimental Results and Analysis Range analysis of the orthogonal experimental results showed that the order of influence of each factor on the model performance was: number of hidden layers (A) > number of neurons (B) > activation function (D) > dropout ratio (C).

[0103] Further analysis of variance showed that the performance was optimal when the number of hidden layers was 3; the decreasing number of neurons (128, 64, 32) achieved the best balance between expressive power and generalization performance; the Dropout ratio of 0.2 effectively suppressed overfitting, while further increasing the ratio would lead to underfitting; the ReLU activation function combined with the Adam optimizer resulted in the most stable convergence.

[0104] Based on the results of the orthogonal experiments, the optimal hyperparameter combination was determined to be: 3 hidden layers, 128, 64, and 32 neurons respectively, a Dropout ratio of 0.2, and ReLU activation function. Figure 4 , Figure 5 As shown, under this configuration, the root mean square error of the data model in predicting bottom hole drilling pressure on the test set is 1395.30 N, with a relative error of 1.34% and a coefficient of determination R² of 0.9776; the root mean square error in predicting bottom hole torque is 30.29 N·m, with a relative error of 0.15% and an R² of 0.9956.

[0105] Example 3 The application of this method is verified in the field by taking the high-level directional long borehole in the roof of a gas extraction roadway in a coal mine as an example.

[0106] 1. Project Overview Drilling design hole depth Among them, the opening section For the upward-facing tunnel section, For the construction of the inclined section, This is a near-horizontal target stratum drilling section. The target coal seam is Coal Seam No. 2, with a thickness of... Protodextrin coefficient The drilling rig contains localized rock inclusions. The drilling rig model is ZDY12000LD, a fully hydraulic directional drilling rig. The drill rod specifications are... External flat drill pipe, slag removal method is clean water slag removal, slag removal volume During drilling, parameters such as feed pressure, power head rotation speed, drilling speed, and hole depth are collected in real time at a specific frequency. .

[0107] To obtain verification data, a self-recording near-bit measurement sub was installed at the third drill pipe post (used only in this test) to record the bottom hole pressure and torque data throughout the entire process as a reference for the actual values.

[0108] 2. Model Training and Deployment Using data from three similar boreholes previously drilled at the same drilling site, the cumulative footage was approximately The fusion model of this invention is trained. After training, the model is deployed on the drilling rig monitoring industrial control computer to calculate and display the predicted values ​​of bottom hole pressure and torque in real time during drilling.

[0109] 3. Result Comparison and Analysis Throughout the drilling process, the prediction results of the method of this invention are compared point by point with the measured values ​​of the short section measured near the drill bit, and at the same time, they are compared horizontally with the pure mechanism model and the pure LSTM data-driven model.

[0110] Hole bottom drilling pressure prediction results: (1) Pure mechanistic model: In The prediction results for the inclined section are acceptable, but... In the subsequent long horizontal section, due to the gradual formation of the cuttings bed and the increase in back pressure from slag discharge, the mechanistic model failed to accurately capture the nonlinear growth of friction, resulting in systematically overestimating predicted values ​​and an average relative error. The maximum error reached .

[0111] (2) Pure LSTM model: The overall trend matches the measured values ​​well, but when drilling encounters interbedded rock layers (sudden drop in drilling speed and violent fluctuations in feed pressure), the prediction shows obvious oscillations and local outliers, with an average relative error of 0. .

[0112] (3) The fusion model of this invention: The mechanistic benchmark provides a stable baseline prediction, the data model outputs a negative residual to correct the increased friction effect during the rock cuttings accumulation stage, and outputs a positive residual to compensate for the sudden change in cutting force at the interlayer. The average relative error across the entire segment... The maximum error did not exceed .

[0113] Hole bottom torque prediction results: (1) Pure mechanistic model: average relative error The error increases significantly at the end of the horizontal section, mainly due to the fact that the additional rotational friction caused by the rock cuttings bed was not adequately modeled.

[0114] (2) Pure LSTM model: average relative error However, there is a response lag when the drilling speed changes rapidly.

[0115] (3) The fusion model of this invention: average relative error The torque variation trend is highly consistent with the measured value.

[0116] 4. Engineering significance After applying the method of this invention, the driller can monitor the actual drilling pressure at the bottom of the hole in real time, avoiding the false feed phenomenon caused by normal feed pressure display at the holehead but insufficient actual drilling pressure at the bottom of the hole, thus increasing the mechanical drilling speed by approximately [missing information]. Furthermore, by providing early warning of abnormal torque at the bottom of the hole, two potential stuck drill accidents were successfully avoided. These results fully verify the effectiveness and engineering practical value of the method of this invention in horizontal rotary drilling operations in coal mines.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to specific embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention, such as replacing LSTM with a GRU network, adjusting the discretization accuracy of the lumped mass method, and correcting the slag discharge mechanism parameters for different slag discharge media, without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A data-mechanism fusion method for inversion of bottom hole drilling pressure and torque in rotary drilling, characterized in that, Includes the following steps: S1: Construct a three-degree-of-freedom lumped mass method mechanism model, simplify the drill string system into a lumped mass model with three degrees of freedom: axial, torsional and transverse. Establish a drill string dynamics equation that considers the cutting characteristics of the drill bit, the friction mechanism of the drill pipe and the slag removal mechanism of the horizontal hole drill pipe, and use it to calculate the predicted values ​​of the bottom hole drilling pressure mechanism and the bottom hole torque mechanism. The mechanistic model incorporates a horizontal hole slag removal mechanism that takes into account the additional friction caused by cuttings bed deposition. S2: Construct an LSTM-BP neural network data model, use the LSTM neural network as the first stage for temporal dynamic feature extraction, and pass the hidden state output of the last time step of the LSTM as a feature vector to the BP neural network. The BP neural network is used as the second stage to perform nonlinear mapping and feature fusion, and outputs the bottom hole drilling pressure prediction residual and the bottom hole torque prediction residual. The inputs to the data model include: sampling time, real-time drilling speed, drilling depth, feed pressure, power head rotation speed, and the predicted values ​​of bottom hole drilling pressure and torque mechanism output by the mechanism model and their corresponding actual values. S3: Establish a residual compensation fusion architecture, and superimpose the predicted residuals output by the data model onto the predicted values ​​of the hole bottom drilling pressure mechanism and the hole bottom torque mechanism output by the mechanism model in an additive compensation manner to obtain the final predicted values ​​of the hole bottom drilling pressure and the hole bottom torque.

2. The method of claim 1, wherein, In S1, the drill bit cutting characteristics decompose the drill bit-rock interaction into cutting and friction components, wherein, Weight on bit cut component , Torque cutting component , Weight on bit friction component , Torque friction component , Where Wc is the cutting component of the bottom hole drilling pressure, Tc is the cutting component of the bottom hole torque, Wf is the friction component of the bottom hole drilling pressure, and Tf is the friction component of the bottom hole torque. ε is the inherent strength of the rock stratum; a is the drill bit radius; ζ is the angular parameter of the drill bit; σ is the contact strength; l is the geometric parameter of the drill bit; μ is the friction coefficient between the drill bit and the rock stratum; γ is the geometric parameter of the drill bit. R(·), H(·), and S(·) represent the Ramp function, Heaviside function, and Sign function, respectively. d(t) represents the instantaneous thickness of cut of the drill bit; represents the angular velocity of rotation of the drill bit; represents the feed speed of the drill bit.

3. The method of claim 1, wherein, In S1, the mechanism model introduces a horizontal hole slag discharge mechanism term that considers the additional friction caused by the deposition of rock cuttings beds. Specifically, for the additional friction caused by the deposition of rock cuttings beds in the horizontal holes, a slag discharge mechanism correction term based on a two-layer flow model is introduced into the mechanism model to establish a slag discharge friction calculation model that includes the self-weight of the coal slag and its circumferential distribution. The relevant calculation formulas are as follows: Where Tfw represents the frictional torque generated by the drill pipe during the rotation of the slag; Ffw represents the additional frictional force of the slag bed on the drill string; Fcw represents the contact force between the slag bed and the wellbore; fws represents the coefficient of friction between the slag bed and the drill string; θ represents the contact angle between the slag bed and the wellbore; and R represents the outer radius of the drill pipe.

4. The method according to claim 1, characterized in that, In S2, the input of the LSTM network specifically includes: sampling time , real-time drilling speed , current drilling depth , feed pressure , power head rotating speed , hole bottom drilling speed , hole bottom drilling pressure mechanism predicted value , hole bottom drilling pressure real value , hole bottom torque mechanism predicted value , hole bottom torque real value ; the input features are constructed into time sequence sample sequences in the form of fixed time windows.

5. The method according to claim 1, characterized in that, In S2, the output mode of the LSTM network is set to last, i.e. only the hidden state of the last time step is retained as a time series feature vector fed into the BP neural network; the BP network comprises at least one fully connected layer, and a ReLU activation function and a Dropout layer are connected after each fully connected layer.

6. The method of claim 1, wherein, In S2, an attention mechanism layer is connected after the LSTM network. The attention mechanism layer is used to dynamically weight the features at different time steps and pass the weighted feature vector to the BP neural network.

7. The method of claim 1, wherein, It also includes model training steps: constructing a training set using logging data from drilled boreholes, using the residuals output by the mechanism model as the learning objective, employing mean squared error as the loss function, using the Adam optimizer to update parameters, and using early stopping to prevent overfitting.

8. The method according to claim 1, characterized in that, The LSTM network is replaced with a gated recurrent unit network, and / or the BP network is replaced with a convolutional neural network.

9. The method according to claim 1, characterized in that, It also includes a model hybrid update strategy: when a single or vertical shaft is drilled, a real-time incremental update is performed, the fully connected layer is locked, and the hidden layer is updated only using real-time drilling sequence data, and one training cycle is performed; when tripping in and out of the drill string, a periodic full update is performed, collecting all drilling data for the current tripping in and out of the drill string and updating the fully connected layer and hidden layer after random shuffling.

10. A coal mine underground borehole bottom drilling pressure and torque inversion system applying the method of any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire in real time the feed pressure, power head speed, drill rod displacement and slag discharge medium parameters of the drilling rig sensors during the drilling process; The mechanism model calculation module is used to run a three-degree-of-freedom lumped mass method model that considers the slag removal mechanism of horizontal holes, and output the mechanism prediction values ​​of bottom hole drilling pressure and torque. The data model calculation module is used to run the LSTM-BP neural network data model and output the predicted residuals of bottom hole drilling pressure and torque. The fusion output module is used to superimpose the mechanism prediction value and the prediction residual to output the final predicted values ​​of bottom hole drilling pressure and torque, and send them to the drilling rig monitoring and display terminal.

Images