Multi-objective real-time optimization method and device for borehole washing process

Abstract

The application discloses a drilling and hole washing process multi-target real-time optimization method and device. The method comprises the following steps: obtaining a plurality of sample points in a parameter space; simulating each sample point according to at least one pre-constructed digital twin to obtain at least one output index corresponding to the sample point; constructing a simulation sample of the sample point and the output index and training a lightweight proxy model; obtaining the working condition parameters of the current hole section and loading the lightweight proxy model, obtaining search recommended process parameters based on the lightweight proxy model and the working condition parameters; and issuing the recommended process parameters to the control unit of the hole washing equipment. The application solves the problem of insufficient real-time performance of online simulation optimization of the digital twin, that is, the simulation sample generated offline in batches by the digital twin is used as a training data source, and the lightweight proxy model can replace online simulation through training.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to a multi-objective real-time optimization method and apparatus for borehole washing process. Background Technology

[0002] Borehole washing is a crucial step in removing mud cake and suspended matter from the borehole wall before spectral observation or borehole television layer determination. The washing effect is determined by a combination of multiple process parameters, such as pump pressure, flow rate, jet direction angle, cleaning agent concentration, and washing time. Moreover, the requirements for parameter combinations vary significantly for different borehole sections due to differences in formation conditions.

[0003] There are three main ways to determine the hole cleaning process parameters in the existing technology.

[0004] Among them, the experience-based fixed value method sets a fixed pump pressure and flow rate for uniform flushing of the entire hole based on the historical experience of this work area. The method is simple but difficult to adapt to the different conditions of different hole sections, and is prone to problems of over-flushing or insufficient flushing.

[0005] Furthermore, the manual trial-and-error adjustment method involves first washing a section of the drill string and observing the effect. If the result is unsatisfactory, the parameters are adjusted and the process is repeated. Although this method is intuitive, the trial-and-error cost is high, the progress is slow, and the parameter selection cannot be reproduced.

[0006] Furthermore, the online simulation optimization method of digital twins, under the condition of having constructed a high-fidelity digital twin, successively changes the parameter combination and runs physical simulation to compare the output indicators. However, the simulation time of a single high-precision digital twin is usually on the order of minutes. To search for a better combination in the multi-parameter space, hundreds or thousands of simulations are required. The total calculation time far exceeds the short decision window allowed for in-hole operations.

[0007] In existing solutions, empirical setting and manual trial and error cannot achieve quantitative multi-objective optimization based on the current working conditions of the borehole section; although online simulation optimization using digital twins can theoretically achieve the optimal result, its real-time performance is insufficient to support real-time decision-making on-site within the borehole. Summary of the Invention

[0008] In view of this, the first aspect of the present invention discloses a multi-objective real-time optimization method for a drilling and hole washing process, the method comprising, Obtain multiple sample points within a parameter space; Simulate each of the sample points using at least one pre-constructed digital twin to obtain at least one output index corresponding to each of the sample points; Multiple simulation samples, including the sample points and the output indicators, are constructed. A lightweight proxy model is trained from the simulation samples. The lightweight proxy model takes the hole washing process parameters and operating condition parameters as inputs and the output vector containing the observation signal-to-noise ratio, water consumption and time consumption as the prediction target to approximate the input-output relationship of the digital twin. Obtain the operating parameters of the current hole segment and load the lightweight proxy model; search for recommended process parameters based on the lightweight proxy model and the operating parameters. The recommended process parameters are sent to the control unit of the hole cleaning equipment.

[0009] In a preferred embodiment of the present invention, the lightweight proxy model is any one of a quadratic polynomial response surface model, a parametric gridded lookup table and multidimensional interpolation model, or a shallow fully connected neural network.

[0010] In a preferred embodiment of the present invention, searching for the recommended process parameters includes, Using the hole washing process parameters as decision variables, a multi-objective optimization problem is constructed with the optimization objectives of maximizing the observation signal-to-noise ratio, minimizing water consumption, and minimizing time consumption. The observed signal-to-noise ratio, water consumption, and time consumption are normalized and then weighted and unified into a scalar evaluation function. The weights of each objective are given by the user according to a preset mode or a custom ratio. The optimal value of the scalar evaluation function is obtained by solving the lightweight proxy model to obtain the recommended process parameters.

[0011] In a preferred embodiment of the present invention, when the lightweight surrogate model is differentiable, the optimal parameter combination of the multi-objective evaluation function is solved on the lightweight surrogate model using the quasi-Newton method; When the lightweight surrogate model is non-differentiable, the solution is obtained on the lightweight surrogate model using the simplex search method.

[0012] In a preferred embodiment of the present invention, the search and recommendation process parameters further include, A Pareto front consisting of a set of non-dominated solutions is generated on the lightweight surrogate model using a multi-objective genetic algorithm. The optimal compromise point for the overall objectives is selected from the front as the automatic recommendation.

[0013] In a preferred embodiment of the present invention, the method further includes, The actual effect indicators were obtained by the sensors inside the hole after the actual hole was cleaned. The relative error is calculated by comparing the actual performance index with the predicted value of the lightweight proxy model at the recommended process parameters. When the relative error of any performance metric exceeds a preset threshold, incremental correction is performed on the lightweight proxy model.

[0014] In a preferred embodiment of the present invention, the parameter space is composed of borehole washing process control parameters and operating condition parameters. The borehole washing process control parameters include at least the borehole washing pump outlet pressure, borehole washing flow rate, jet axial direction angle, jet circumferential deflection angle, cleaning agent mass concentration, and target borehole washing time per unit borehole section. The operating condition parameters include at least the borehole diameter, mud dynamic viscosity, mud density, formation porosity, and median particle size of rock powder.

[0015] In a preferred embodiment of the present invention, the acquisition of multiple sample points within a parameter space is achieved using Latin hypercube sampling or uniform grid sampling. When performing Latin hypercube sampling, each parameter dimension is divided into equally probable intervals, and random sampling is performed within each interval followed by random pairing across dimensions. During uniform grid sampling, each parameter dimension is separated into multiple horizontal values ​​within its value range according to a preset step size, and all cross combinations of the horizontal values ​​of each dimension are taken to form the sample points.

[0016] In a preferred embodiment of the present invention, when constructing the simulation sample, the simulation sample is further filtered for physical rationality and engineering feasibility. At least samples with an observation signal-to-noise ratio lower than the minimum acceptable threshold, samples with a flow rate exceeding the pump's rated range, and samples with water consumption exceeding the upper limit of construction economy are removed. After filtering, the effective samples are divided into training set and test set according to a preset ratio.

[0017] Furthermore, a second aspect of the present invention discloses a multi-objective real-time optimization device for a drilling and hole cleaning process, which applies the aforementioned multi-objective real-time optimization method for the drilling and hole cleaning process. The device includes, The lightweight proxy model construction and training unit is used to acquire sample points in the parameter space, obtain output indicators from digital twin simulation and construct simulation samples, and train the lightweight proxy model from the simulation samples. The online multi-objective optimization unit is used to receive the current hole section working condition parameters and load the lightweight proxy model, search for recommended process parameters on the lightweight proxy model, and send the recommended process parameters to the hole cleaning equipment control unit. The closed-loop correction unit is used to obtain the relative error between the measured effect index after the physical hole cleaning and the predicted value of the lightweight proxy model. When the error exceeds the threshold, incremental correction is performed on the lightweight proxy model.

[0018] Compared with existing technologies, this invention solves the problem of insufficient real-time performance in online simulation optimization of digital twins. Specifically, it uses simulation samples generated offline in batches as training data sources, and replaces repeated online simulations with a lightweight proxy model that can complete inference in much less than a second. This allows multi-objective optimization to be completed on-site in the hole in real time. Furthermore, the lightweight proxy model achieves continuous closed-loop adaptation across hole sections and work areas through Bayesian incremental correction. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the multi-objective real-time optimization method for the drilling and cleaning process described in this embodiment; Figure 2 This is a schematic diagram of the online weighted and multi-objective optimization process in this embodiment; Figure 3 This is a schematic diagram of the closed-loop correction Bayesian incremental update process described in this embodiment; Figure 4 This is a schematic diagram of the multi-objective real-time optimization device for the drilling and washing process described in this embodiment. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] This embodiment provides a multi-objective real-time optimization method for drilling and hole cleaning processes. The method can be applied to drilling and hole cleaning systems. The drilling and hole cleaning system includes a pre-constructed digital twin simulation platform, an in-hole control terminal, a hole cleaning pump, a flow valve, a hole cleaning agent addition pump, and in-situ sensors within the hole.

[0023] The digital twin pre-constructed in the digital twin simulation platform is formed based on the radiative transfer theory and historical measured hole washing data: the optical simulation kernel of the hole wall is established based on the physical mechanism of light attenuation and scattering in the downhole drilling fluid, and the optical property parameters in the simulation kernel are calibrated by historical measured spectrum or borehole television image data to form a twin that can predict the signal-to-noise ratio, water consumption and time consumption of each hole section under given hole washing parameters.

[0024] It should be noted that the innovation of this invention does not lie in constructing the physical model of the digital twin itself, but rather in positioning it as an offline training data source, driving a lightweight proxy model to achieve online real-time optimization through its system architecture and execution method. In this embodiment, the radiative transfer simulation kernel of the digital twin is constructed based on the well-known radiative transfer equation (RTE) and discrete coordinate method (DOM). The particle settling model is based on the Stokes settling formula, and its specific mathematical form and numerical solution are existing technologies, referencing core content from drilling fluid optical monitoring literature and Chandrasekhar's *Radiative Transfer*.

[0025] Therefore, the present invention provides an exemplary description of the construction of the physical model of the digital twin and its connection with this embodiment.

[0026] The digital twin's inputs are borehole washing process control parameters (pump outlet pressure, washing flow rate, jet direction angle, cleaning agent concentration, target washing duration) and operating parameters (borehole diameter, mud dynamic viscosity and density, formation porosity, median rock powder particle size). The outputs are three performance indicators: observed signal-to-noise ratio (SNR), water consumption per unit borehole section (W), and time per unit borehole section (T). The observed SNR is calculated from the ratio of rock and mineral characteristic signals to suspended noise in the target band using radiative transfer simulation. The water consumption per unit borehole section (W) is obtained by directly multiplying the washing flow rate by the target washing duration. The time per unit borehole section (T) is obtained by adding the post-washing settling time to the target washing duration, where the settling time is calculated backwards from the particle settling model based on rock powder particle size, mud viscosity, and a preset turbidity threshold.

[0027] Furthermore, the construction of a digital twin may include establishing a simulation kernel and calibrating parameters using historical test data.

[0028] Establishing a simulation kernel involves modeling the downhole drilling fluid as an absorption-scattering turbid medium containing suspended rock powder particles. The attenuation and scattering process of light in the medium is described using the radiative transfer equation, and the settling process of suspended particles after well flushing is described using the Stokes equation. A discrete coordinate method is used to perform layered numerical solutions along the optical path from the well wall to the spectral probe. The physical models and numerical methods used in the above simulation kernel steps are all existing technologies.

[0029] Parameter calibration based on historical measured data refers to selecting measured spectral (or borehole television image) data from no fewer than five historical borehole segments in the current or adjacent work areas as the calibration set. The mean square error between the effective spectrum (or image contrast) of the simulation output and the measured value is used as the objective function, and nonlinear least squares fitting is employed to calibrate the undetermined optical parameters (the proportionality constants between the absorption and scattering coefficients) in the simulation kernel. The above parameter calibration steps are also standard techniques in this field.

[0030] Based on this, after the digital twin is completed, it acts as an offline data generator to generate simulation samples in batches according to the sampling strategy.

[0031] Furthermore, this invention uses a lightweight proxy model to replace online simulation to achieve second-level optimization, and does not depend on the specific physical equations or numerical solutions used by the digital twin. Those skilled in the art can independently construct a usable digital twin based on the aforementioned well-known physical models, or directly use commercial downhole optical simulation software (such as ZemaxOpticStudio's non-sequential mode, TracePro, etc.) to achieve equivalent simulation functions.

[0032] Figure 1 This is a flowchart illustrating the multi-objective real-time optimization method for the drilling and cleaning process described in this embodiment. Figure 1 The execution of the multi-objective real-time optimization method for the drilling and washing process is shown to include the following steps.

[0033] 10. Select sample points in a parameter space according to the sampling strategy.

[0034] The parameter space is composed of the hole cleaning process control parameters and the operating condition parameters.

[0035] The control parameters for the hole cleaning process include the outlet pressure of the hole cleaning pump, the hole cleaning flow rate, the axial direction angle of the jet, the circumferential deflection angle of the jet, the mass concentration of the cleaning agent, and the target hole cleaning time per unit hole segment.

[0036] The operating parameters include borehole diameter, mud dynamic viscosity, mud density, formation porosity, and median particle size of rock powder.

[0037] The sampling strategy can be uniform grid sampling, which means that each parameter dimension is separated into multiple horizontal values ​​within its value range according to a preset step size, and all cross combinations of the horizontal values ​​of each dimension are taken to form a sample point set, so that the sample points are distributed in a regular grid in the parameter space.

[0038] In some preferred embodiments, the sampling strategy is preferably Latin hypercube sampling, which divides each parameter dimension into equally probable intervals, randomly selects a point within each probability interval, and then randomly pairs it across dimensions to ensure a uniform marginal distribution and no overlapping intervals in each dimension. Alternatively, when the parameter dimension does not exceed three dimensions, uniform grid sampling can also be used to ensure higher conditional space coverage.

[0039] 20. Simulate each sample point in a pre-built digital twin to obtain output metrics.

[0040] In this context, simulating sample points in a digital twin means simulating the entire process of hole washing for each sample point in the digital twin.

[0041] The output indicators include at least the observation signal-to-noise ratio, water consumption per unit well section, and time consumption per unit well section.

[0042] The observed signal-to-noise ratio (SNR) refers to the ratio of the amplitude of the characteristic signal to the root mean square of the noise within the target band, and its calculation formula is expressed as follows: ,in The average amplitude of the characteristic reflection peaks of rocks and minerals within the target band is obtained from the normalized reading of the spectrometer, with a value range of 0 to 1. The root mean square of the baseline noise caused by residual suspended matter and stray light from the orifice cleaning process is output from the radiative transfer simulation in the digital twin. Therefore, a higher observation signal-to-noise ratio (SNR) indicates clearer spectral features.

[0043] In some preferred embodiments, in drilling scenarios equipped only with a drilling camera and without a spectral probe, the local contrast ratio C of the borehole wall image can be used instead of the observed signal-to-noise ratio (SNR), and its calculation formula is expressed as follows: Imax and Imin are the maximum and minimum brightness values ​​of a local area in the borehole wall image acquired by the borehole camera, respectively.

[0044] The water consumption W per unit borehole section refers to the water consumption per meter of borehole section during the washing operation, and its calculation formula is expressed as follows: Where Q is the orifice flow rate (L / min). The target cleaning time per unit hole segment (s / m).

[0045] The time T per unit hole segment refers to the time spent washing one meter of hole segment, and its calculation formula is expressed as follows: ,in The settling time (min / m) for suspended solids to settle after rinsing is stopped is determined by the digital twin based on the particle settling rate and the turbidity threshold within the observation depth.

[0046] 30. Create multiple simulation samples and divide them into training and test sets.

[0047] The simulation samples include the sample coordinates representing each sample in the parameter space in step 10 and the output index obtained from the simulation of each sample in step 20.

[0048] In some preferred embodiments, the generated samples are filtered for physical rationality and engineering feasibility before the dataset is divided, in order to at least remove samples with an observation signal-to-noise ratio below the minimum acceptable threshold, samples with a flow rate exceeding the pump's rated range, and samples with water consumption exceeding the upper limit of construction economics.

[0049] Specifically, simulation samples with an observation signal-to-noise ratio below the minimum acceptable threshold are removed because the hole washing effect corresponding to such parameter combinations cannot meet the minimum data quality requirements for layer discrimination and are considered invalid samples. Retaining them would mislead the lightweight surrogate model into wasting fitting degrees of freedom in infeasible regions. Meanwhile, simulation samples with orifice flow rates exceeding the pump's rated range were removed because some flow values ​​generated during the sampling phase may exceed the rated upper limit of the nameplate of the orifice pump on site. Such parameter combinations can run in simulation but cannot be executed physically, and are therefore invalid samples. The optimal solution that may be recommended by the lightweight proxy model but cannot be achieved on site was retained. Samples with water consumption exceeding the upper limit of construction economics were excluded because, although the unit water consumption corresponding to certain parameter combinations was physically feasible, it far exceeded the on-site water supply capacity and cost budget, and was therefore considered invalid samples.

[0050] After this filtering, the lightweight surrogate model fits only to the meaningful parameter subspace, avoiding the dilution of model accuracy by invalid samples and ensuring that the recommended solution output by subsequent optimization is practically feasible for on-site construction. The filtered valid samples are divided into training and test sets according to a preset ratio.

[0051] 40. Select training and test sets to train a lightweight agent model.

[0052] This involves training an approximate mapping relationship based on the data from the training and test sets. The lightweight proxy model makes the lightweight proxy model approximate the actual input-output relationship of the digital twin.

[0053] The lightweight proxy model in this embodiment can be any one of the following: a quadratic polynomial response surface model, a parametric gridded lookup table and multidimensional interpolation model, or a shallow fully connected neural network. The quadratic polynomial response surface model, parametric gridded lookup table and multidimensional interpolation model, or shallow fully connected neural network are all existing technologies.

[0054] Preferably, the lightweight proxy model described in this embodiment can be selected according to the following principles.

[0055] When the parameter space has sufficient smoothness and gradient information is needed to support efficient optimization, the quadratic polynomial response surface model is preferred. Sufficient smoothness means that the digital twin output is continuous and differentiable in the parameter space.

[0056] When the input dimension does not exceed three dimensions and the reasoning process needs to be completely transparent and auditable, a parametric gridded lookup table is used. When the parameter space exhibits strong nonlinearity and the polynomial fitting residual is large, a shallow fully connected neural network is employed.

[0057] Furthermore, to provide a clear and complete explanation of this embodiment, this embodiment provides a detailed training process for the quadratic polynomial response surface model, the mathematical form of which is: Where m=1,2,3 corresponds to three output channels (SNR, W, T), and each output channel is fitted independently. The input is linearly scaled to the [-1,1] interval before modeling to eliminate dimensional differences. Model coefficients are estimated using the least squares method, and L2 regularized regression is used to address multicollinearity. The regularization coefficient λ is selected by k-fold cross-validation to minimize the mean square error. When the input dimension is large, LASSO regularization is used to remove insignificant terms.

[0058] Furthermore, when the parameter space exhibits strong nonlinearity (e.g., complex formation conditions such as interbedded mudstone and carbonate rock fissures and cavities), and the cross-validation RMSE of the quadratic polynomial response surface exceeds a preset percentage of the full scale, a shallow fully connected neural network is used as a surrogate model.

[0059] Furthermore, to provide a clear and complete explanation of this embodiment, this embodiment provides a network construction for a parameter gridded lookup table and a multidimensional interpolation model. For each parameter dimension to be optimized, the grid is uniformly divided within its value range: if the parameter range is [a, b], the number of grids is... Then the parameter value of the i-th grid node This embodiment recommends... (When the parameter dimension is 1) (When the parameter dimension is 2) (When the parameter dimension is 3), the total number of grid nodes is kept below approximately 10,000, ensuring that batch simulations of digital twins can be completed within hours. Furthermore, during online queries, for parameter combinations not on grid nodes, multidimensional linear interpolation (when the parameter space is a regular grid) or multidimensional cubic spline interpolation (when the parameter space has strong local curvature variations) is used. Multidimensional linear interpolation: For a query point x_query in d-dimensional space, its target value is obtained by weighted averaging of the values ​​of the 2^d nearest grid nodes surrounding that point, with the weights being the product of the relative distances of the query point to the adjacent nodes in each dimension. Multidimensional cubic spline interpolation: After independently performing one-dimensional cubic spline interpolation in each dimension, the results are combined according to dimension priority to obtain the multidimensional interpolation result. The total time for table lookup and interpolation is short, meeting the real-time requirements of the borehole field.

[0060] Furthermore, to provide a clear and complete explanation of this embodiment, this embodiment provides an architecture for a shallow fully connected neural network, which includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer includes 11 nodes, corresponding to all input parameters (6 process control parameters and 5 operating condition parameters), each input being linearly scaled to the [-1, 1] interval before entering the network. The first hidden layer includes 64 nodes, with the activation function being LeakyReLU (negative slope 0.1), followed by a dropout layer (dropout rate 0.1) to prevent overfitting. The second hidden layer includes 32 nodes, with the activation function being LeakyReLU, followed by a dropout layer (dropout rate 0.1); the output layer includes 3 nodes, corresponding to the three prediction targets SNR, W, and T, with a linear activation function. Therefore, the total number of parameters in the shallow fully connected neural network is approximately 3,107, far less than that of a typical deep network, and the number of floating-point operations (FLOPs) per inference iteration is approximately 6,200.

[0061] 50 Online multi-objective optimization.

[0062] Figure 2 This is a schematic diagram of the online weighted and multi-objective optimization process described in this embodiment. Figure 2 The execution of the online multi-objective optimization is shown to include the following steps.

[0063] 51. Obtain the current working condition of the hole section and load the corresponding lightweight proxy model.

[0064] In step 51, the operating parameters for the current borehole section are obtained from caliper logging, mud logging, and data from adjacent wells. If multiple lightweight proxy models have been trained according to operating conditions, the version that best matches the current operating conditions is loaded.

[0065] 52. Construct and solve a multi-objective optimization problem on a lightweight surrogate model.

[0066] In step 52, a multi-objective optimization problem is constructed with the hole cleaning process parameters as decision variables x: minimizing the negative signal-to-noise ratio (i.e. maximizing the signal-to-noise ratio), minimizing water consumption, minimizing time consumption, while satisfying constraints such as pump pressure range, flow rate range, minimum acceptable signal-to-noise ratio, and upper limit of hole cleaning agent concentration.

[0067] This embodiment uses the normalized weighted sum method to unify the three objectives into a scalar evaluation function. ,in , , The normalized reference span for each objective is W1+W2+W3=1, and the preference weights are given by the user through the interactive interface by selecting a preset mode or a custom ratio.

[0068] When the lightweight surrogate model is a quadratic polynomial response surface The gradient of x has a closed analytical expression. The quasi-Newton method (L-BFGS-B) is used to iteratively optimize the gradient norm until it is less than the threshold or the upper limit of iteration is reached.

[0069] When the lightweight proxy model is in a non-differentiable form such as a lookup table, the simplex search method is used for iterative optimization.

[0070] In a preferred embodiment, this example employs a multi-objective genetic algorithm to run on a lightweight surrogate model at extremely low cost (population size 50, evolution 100 generations), outputting the Pareto front for users to view the complete trade-off surface before making a selection. Since the time for a single inference iteration of the lightweight surrogate model is much shorter than the time for a single simulation iteration of the digital twin, the entire optimization process can be completed within seconds, meeting the timeliness requirements for real-time decision-making within the well.

[0071] 60. The recommended parameters are sent to the hole cleaning equipment control unit.

[0072] In step 60, the optimal combination of process parameters obtained from the search is encoded into control commands that can be read by the hole cleaning pump, flow valve, and hole cleaning agent addition pump and then executed.

[0073] Therefore, this invention first replaces repeated online digital twin simulations with an offline trained lightweight agent model to compress the time of a single inference, enabling multi-objective optimization to be completed quickly, thus solving the timeliness bottleneck of real-time decision-making in the borehole.

[0074] Furthermore, this invention utilizes various surrogate model selection strategies (such as quadratic polynomial response surface, lookup table interpolation, and shallow neural networks) to flexibly adapt to the requirements of parameter space smoothness and interpretability, thus balancing accuracy and transparency.

[0075] Meanwhile, this invention uses a normalized weighted sum method to unify the signal-to-noise ratio, water consumption, and time consumption into a scalar evaluation function, and presets three weight modes: quality assurance, balance, and fast forward, allowing operators to quickly select according to working conditions, thus reducing the human threshold for multi-objective optimization.

[0076] 70 Closed-loop correction.

[0077] Figure 3 This is a schematic diagram of the closed-loop correction Bayesian incremental update process described in this embodiment. Figure 3 The execution of the closed-loop correction is shown to include the following steps.

[0078] 71. Obtain the actual effect indicators after the solid hole cleaning and calculate the error.

[0079] In step 71, after the physical hole washing system is executed according to the recommended parameters, the actual observed signal-to-noise ratio, water consumption and time are obtained by the in-situ sensor inside the hole, and the relative error is calculated by comparing it with the predicted value of the lightweight surrogate model channel by channel.

[0080] 72. Determine if a correction is triggered and execute it.

[0081] In step 72, a periodic correction is triggered when the absolute value of the relative error of any output channel exceeds a preset trigger threshold (e.g., 15%) or after a preset number of aperture segments. When the lightweight proxy model is a quadratic polynomial response surface, the correction uses Bayesian linear regression incremental updates, and its calculation formula is as follows: ; ; in, Let Z* be the Kalman gain, Z* be the row vector of the design matrix corresponding to the current recommended parameters, σ² be the empirical estimate of the observation noise variance, and μ and Σ be the posterior mean and covariance matrix of the coefficients, respectively. Only one matrix-vector update is needed for each new observation, resulting in minimal computation. When the lightweight surrogate model is a lookup table model, the correction uses an exponential moving average, with weighted merging of the nearest grid points: v_new = α·v_old + (1 α)·y_real. In the initial stage of the new work area, α=0.5 is used for rapid adaptation, and then gradually increased to α=0.9 to maintain stability.

[0082] Therefore, this invention achieves closed-loop incremental correction through Bayesian linear regression or exponential moving average, enabling the lightweight surrogate model to continuously adapt in cross-hole and cross-work area scenarios without requiring full retraining, thus significantly reducing model maintenance costs.

[0083] In summary, the method described in this invention is driven by offline simulation data and does not rely on on-site trial and error, which significantly shortens the debugging cycle of borehole washing parameters. It is applicable to various drilling operation scenarios such as geological exploration, well observation, and hydrological drilling, and has significant economic and practical advantages.

[0084] Furthermore, Figure 4 This embodiment illustrates a multi-objective real-time optimization device for a drilling and cleaning process. The device includes a surrogate model construction and training unit, an online multi-objective optimization unit, and a closed-loop correction unit.

[0085] The lightweight proxy model construction and training unit is used to generate a set of mapping samples offline in batches according to a sampling strategy in the digital twin, and to train the lightweight proxy model using the set of samples. The online multi-objective optimization unit is used to receive the current hole section operating parameters, search for recommended process parameters on the lightweight surrogate model using a multi-objective evaluation function, and send the recommended parameters to the hole cleaning equipment control unit. The closed-loop correction unit is used to obtain the relative error between the measured effect index after the physical hole washing and the predicted value of the lightweight proxy model. When the error exceeds the threshold, incremental correction is performed on the lightweight proxy model.

[0086] The above description is merely a preferred embodiment of the present invention and does not limit the scope of the patent. Any equivalent structural modifications made based on the inventive concept of the present invention and the description and drawings, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of the present invention.

Claims

1. A multi-objective real-time optimization method for drilling and hole washing processes. Its features are, The method includes, Obtain multiple sample points within a parameter space; Simulate each of the sample points using at least one pre-constructed digital twin to obtain at least one output index corresponding to each of the sample points; Multiple simulation samples, including the sample points and the output indicators, are constructed. A lightweight proxy model is trained from the simulation samples. The lightweight proxy model takes the hole washing process parameters and operating condition parameters as inputs and the output vector containing the observation signal-to-noise ratio, water consumption and time consumption as the prediction target to approximate the input-output relationship of the digital twin. Obtain the operating parameters of the current hole segment and load the lightweight proxy model; search for recommended process parameters based on the lightweight proxy model and the operating parameters. The recommended process parameters are sent to the control unit of the hole cleaning equipment.

2. The multi-objective real-time optimization method for drilling and hole washing process according to claim 1, characterized in that, The lightweight proxy model can be any one of the following: a quadratic polynomial response surface model, a parametric gridded lookup table and multidimensional interpolation model, or a shallow fully connected neural network.

3. The multi-objective real-time optimization method for drilling and hole washing process according to claim 1, characterized in that, The search for the recommended process parameters includes, Using the hole washing process parameters as decision variables, a multi-objective optimization problem is constructed with the optimization objectives of maximizing the observation signal-to-noise ratio, minimizing water consumption, and minimizing time consumption. The observed signal-to-noise ratio, water consumption, and time consumption are normalized and then weighted and unified into a scalar evaluation function. The weights of each objective are given by the user according to a preset mode or a custom ratio. The optimal value of the scalar evaluation function is obtained by solving the lightweight proxy model to obtain the recommended process parameters.

4. The multi-objective real-time optimization method for drilling and hole washing process according to claim 3, characterized in that, When the lightweight surrogate model is differentiable, the optimal parameter combination of the multi-objective evaluation function is solved on the lightweight surrogate model using the quasi-Newton method. When the lightweight surrogate model is non-differentiable, the solution is obtained on the lightweight surrogate model using the simplex search method.

5. The multi-objective real-time optimization method for drilling and hole washing process according to claim 3, characterized in that, The search and recommendation process parameters also include, A Pareto front consisting of a set of non-dominated solutions is generated on the lightweight surrogate model using a multi-objective genetic algorithm. The optimal compromise point for the overall objectives is selected from the front as the automatic recommendation.

6. The multi-objective real-time optimization method for drilling and hole washing process according to claim 1, characterized in that, The method also includes, The actual effect indicators were obtained by the sensors inside the hole after the actual hole was cleaned. The relative error is calculated by comparing the actual performance index with the predicted value of the lightweight proxy model at the recommended process parameters. When the relative error of any performance metric exceeds a preset threshold, incremental correction is performed on the lightweight proxy model.

7. The multi-objective real-time optimization method for drilling and hole washing process according to claim 1, characterized in that, The parameter space is composed of borehole washing process control parameters and operating condition parameters. The borehole washing process control parameters include at least the borehole washing pump outlet pressure, borehole washing flow rate, jet axial direction angle, jet circumferential deflection angle, cleaning agent mass concentration, and target borehole washing time per unit borehole section. The operating condition parameters include at least the borehole diameter, mud dynamic viscosity, mud density, formation porosity, and median particle size of rock powder.

8. The multi-objective real-time optimization method for drilling and hole washing process according to claim 1, characterized in that, The acquisition of multiple sample points within a parameter space is achieved using Latin hypercube sampling or uniform grid sampling. When performing Latin hypercube sampling, each parameter dimension is divided into equally probable intervals, and random sampling is performed within each interval followed by random pairing across dimensions. During uniform grid sampling, each parameter dimension is separated into multiple horizontal values ​​within its value range according to a preset step size, and all cross combinations of the horizontal values ​​of each dimension are taken to form the sample points.

9. The multi-objective real-time optimization method for drilling and hole washing process according to claim 1, characterized in that, When constructing the simulation samples, the process also includes filtering the simulation samples based on physical rationality and engineering feasibility. At least samples with an observation signal-to-noise ratio lower than the minimum acceptable threshold, samples with a flow rate exceeding the pump's rated range, and samples with water consumption exceeding the upper limit of construction economics are removed. After filtering, the effective samples are divided into training set and test set according to a preset ratio.

10. A multi-objective real-time optimization device for drilling and hole cleaning processes, employing the multi-objective real-time optimization method for drilling and hole cleaning processes as described in any one of claims 1 to 9. Its features are, The device includes, The lightweight proxy model construction and training unit is used to acquire sample points in the parameter space, obtain output indicators from digital twin simulation and construct simulation samples, and train the lightweight proxy model from the simulation samples. The online multi-objective optimization unit is used to receive the current hole section working condition parameters and load the lightweight proxy model, search for recommended process parameters on the lightweight proxy model, and send the recommended process parameters to the hole cleaning equipment control unit. The closed-loop correction unit is used to obtain the relative error between the measured effect index after the physical hole cleaning and the predicted value of the lightweight proxy model. When the error exceeds the threshold, incremental correction is performed on the lightweight proxy model.

Images