Drill rod thread ultimate bearing capacity prediction model training method and system

Abstract

The invention provides a drill rod thread ultimate bearing capacity prediction model training method and system, and belongs to the field of oil drilling equipment, and the method comprises the steps: collecting sample points of a thread failure related parameter space; constructing a finite element model according to the sample points, and calculating limit torque values of the sample points based on the finite element model; constructing a training data set according to the failure related parameters and the limit torque values of the sample points; and inputting the training data set into the neural network model, and adjusting the neural network model by taking the physical information loss function as an optimization target to obtain a thread ultimate bearing prediction model. According to the method, a physical information loss function is used as a core optimization target, so that physical rule constraint is applied to a neural network predicted value, the learning process is forced to fit data and follow the basic principle of structural mechanics, and the generalization ability and prediction reliability of the model under the extrapolation working condition are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of oil drilling equipment, specifically relating to a training method and system for predicting the ultimate bearing capacity of drill pipe threads. Background Technology

[0002] Aluminum alloy drill pipes are widely used in oil drilling due to their light weight and corrosion resistance. However, the ultimate load-bearing capacity, such as the ultimate torque of the drill pipe threads, is a key indicator for their safe service. Traditional methods for calculating ultimate load-bearing capacity mainly rely on experiments or finite element simulations: experiments require the construction of large loading devices, which are costly and time-consuming; finite element simulations require high-precision modeling, which involves a huge amount of computation.

[0003] With the development of deep learning, models such as XGBoost have been increasingly applied to the field of mechanical analysis due to their ability to reduce computational load and improve prediction efficiency. For example, the XGBoost model uses data fitting as its core objective, optimizing parameters by minimizing the error between predicted and true values, ultimately achieving the prediction of ultimate torque under unknown working conditions. However, experiments or simulations of aluminum alloy drill pipe thread failure are extremely costly and typically only a small amount of sample data can be obtained. Existing deep learning models, relying on training datasets built from limited sample data, are prone to overfitting, resulting in weak generalization ability. Consequently, the prediction results of the trained models under extreme working conditions contradict physical laws, leading to poor reliability in the prediction of ultimate bearing capacity. Summary of the Invention

[0004] To address the issue of low reliability in existing prediction models, this invention provides a training method and system for predicting the ultimate bearing capacity of drill pipe threads.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Training methods for predicting the ultimate bearing capacity of drill pipe threads include: Sample points are collected from the domain space of failure-related parameters of drill pipe threads; a finite element model is constructed based on the sample points, and the ultimate torque value of the sample points is calculated based on the finite element model; a training dataset is constructed based on the failure-related parameters and ultimate torque values ​​of the sample points. The training dataset is input into the Transformer model, and the parameters of the neural network model are iteratively updated with the physical information loss function as the optimization objective to obtain the thread ultimate bearing capacity prediction model. The physical information loss function includes a data-driven loss function and a physical prior loss function. The data-driven loss function is used to calculate the predicted ultimate torque value and the average prediction error of the ultimate torque value output by the Transformer model through the Smooth L1 function. The physical prior loss function is used to calculate the theoretical ultimate torque value through a pre-built quadratic physical model, and to calculate the average squared difference between the predicted ultimate torque value and the theoretical ultimate torque value. The quadratic physical model includes the squared terms of multiple failure-related parameters and their mutual interaction terms. The coefficients of the quadratic terms in the quadratic physical model are constrained to be non-negative.

[0006] Optionally, the drill pipe thread ultimate bearing capacity prediction model training method provided by the present invention further includes: Sample points were selected in the failure-related parameter space of drill pipe threads using the Latin hypercube sampling method.

[0007] Optionally, the drill pipe thread ultimate bearing capacity prediction model training method provided by the present invention further includes: A three-dimensional nonlinear finite element model was constructed using hexahedral elements based on the failure-related parameters of the sample points. Based on geometric nonlinear constraints, contact nonlinear constraints, and material nonlinear constraints, the corresponding ultimate torque value is determined by the failure-related parameters of the sample points.

[0008] Optionally, the drill pipe thread ultimate bearing capacity prediction model training method provided by the present invention further includes: Failure-related parameters in the training dataset that exceed the preset upper bound of the interquartile range are truncated to the values ​​corresponding to the upper bound of the interquartile range, and failure-related parameters in the training dataset that exceed the preset lower bound of the interquartile range are truncated to the values ​​corresponding to the lower bound of the interquartile range, thus obtaining the training dataset after interquartile range judgment. Additional features are constructed based on failure-related parameters, and these additional features are added to the training dataset after interquartile range determination to obtain the preprocessed training dataset.

[0009] Optionally, the Transformer model is a PI-TransCNN model. The PI-TransCNN model includes a shallow feature extraction module, a deep feature extraction module, and a prediction output module connected sequentially. The deep feature extraction module consists of multiple stacked residual modules based on Transformer modulation. The residual modules include parallel Transformer branches and convolutional branches, as well as feature modulation units connected to the Transformer branches and convolutional branches respectively. The training dataset is input into the neural network model, and the parameters of the neural network model are iteratively updated with the physical information loss function as the optimization objective, including: The shallow feature vector is obtained by extracting failure-related parameters from the training dataset through the shallow feature extraction module; Based on the self-attention mechanism, the shallow feature vectors are globally modeled through the Transformer branch to obtain the global feature vector; the local feature vectors are extracted from the shallow feature vectors through the convolution branch; and the global feature vectors and local feature vectors are fused through the feature modulation unit to obtain the deep feature vector. The predicted value of the ultimate torque is determined by the deep feature vector through the prediction output module; The loss is calculated for the predicted and ultimate torque values ​​based on the physical information loss function, and the PI-TransCNN model is adjusted based on the loss calculation results.

[0010] Optionally, the physical information loss function also includes a monotonic loss function, which is used to calculate the gradient of the predicted ultimate torque value relative to the failure-related parameters, and to penalize the mean of the negative gradient using the ReLU function.

[0011] Optionally, failure-related parameters include temperature parameters and tensile load parameters. The drill pipe thread ultimate bearing capacity prediction model training method provided by this invention also includes: Obtain the temperature parameters and tensile load parameters at the thread of the rod to be predicted; By inputting the temperature parameters and tensile load parameters into the thread ultimate bearing prediction model, the ultimate bearing prediction result at the thread of the rod to be predicted is obtained.

[0012] This invention also provides a training system for a prediction model of the ultimate bearing capacity of drill pipe threads, comprising: The dataset construction module is used to collect sample points from the domain space of failure-related parameters of drill pipe threads; construct a finite element model based on the sample points; calculate the ultimate torque value of the sample points based on the finite element model; and construct a training dataset based on the failure-related parameters and ultimate torque values ​​of the sample points. The model training module is used to input the training dataset into the Transformer model and iteratively update the parameters of the neural network model with the physical information loss function as the optimization objective, thereby obtaining the thread ultimate bearing capacity prediction model. The physical information loss function includes a data-driven loss function and a physical prior loss function. The data-driven loss function is used to calculate the predicted ultimate torque value and the average prediction error of the ultimate torque value output by the Transformer model through the Smooth L1 function. The physical prior loss function is used to calculate the theoretical ultimate torque value through a pre-built quadratic physical model and to calculate the average squared difference between the predicted ultimate torque value and the theoretical ultimate torque value. The quadratic physical model includes the squared terms of multiple failure-related parameters and their mutual interaction terms. The coefficients of the quadratic terms in the quadratic physical model are constrained to be non-negative.

[0013] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in the drill pipe thread ultimate bearing capacity prediction model training method.

[0014] The present invention also provides a computer-readable storage medium storing a computer program, which, when loaded by a processor, is capable of executing any step of the drill pipe thread ultimate bearing capacity prediction model training method.

[0015] The drill pipe thread ultimate bearing capacity prediction model training method provided by this invention has the following beneficial effects: Because the drill pipe thread ultimate bearing capacity prediction model training method provided by this invention uses the physical information loss function as the core optimization objective, it imposes physical constraints on the neural network prediction values, forcing its learning process not only to fit the data but also to follow the basic principles of structural mechanics. This significantly improves the model's generalization ability and prediction reliability under extrapolation conditions. Specifically, the physical information loss function includes not only the calculation of the average error between the predicted and actual values ​​but also the calculation of the average squared difference between the theoretical and predicted values ​​calculated by the quadratic physical model. The quadratic polynomial architecture of the quadratic physical model is widely used in structural mechanics. For example, the squared term is analogous to formulas related to energy accumulation, such as strain energy, while the cross term describes various failure correlation coefficients, such as the coupling effect between temperature and load, covering the influence of temperature on multiple mechanical properties. This ensures that the quadratic physical model macroscopically follows mechanical laws. Furthermore, considering that some physical quantities in structural mechanics are scalars, their squared term coefficients are non-negative. The ReLU activation function forces them to be non-negative, ensuring that the model learns that as the absolute value of temperature or load increases, its effect on the structure always tends towards failure, thus avoiding predictions that contradict physical principles, such as higher temperatures leading to stronger structures. The final trained thread ultimate bearing capacity prediction model can still give reasonable predictions that conform to physical common sense when faced with extreme parameters not fully covered by the training data. This significantly improves the reliability of the prediction results of the drill pipe thread ultimate bearing capacity and provides a reliable basis for drilling operation safety judgment. Attached Figure Description

[0016] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. 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.

[0017] Figure 1 This is a schematic diagram of the training method for the prediction model of the ultimate bearing capacity of drill pipe threads provided in an embodiment of the present invention; Figure 2 This is an example of a model training process provided in an embodiment of the present invention; Figure 3 This is an example of a sample scatter plot for Latin hypercube sampling provided in an embodiment of the present invention; Figure 4 This is an example of a finite element model provided in an embodiment of the present invention; Figure 5 This is an example of a PI-TransCNN model framework diagram provided in an embodiment of the present invention; Figure 6 This is an example of the prediction performance graph of the PI-TransCNN model provided in an embodiment of the present invention; Figure 7 This is a comparison example of the root mean square error of multiple models provided in the embodiments of the present invention; Figure 8 This is an example of the ultimate load distribution and safety line diagram provided in an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0019] With the development of deep learning, models such as MLP and XGBoost have been gradually used for ultimate load prediction. However, these models lack physical constraints and have low reliability. Moreover, due to the extremely high cost of experimental or simulation experiments on aluminum alloy drill pipe thread failure, only small sample data can usually be obtained. Traditional deep learning models are prone to overfitting and have poor generalization ability. Furthermore, traditional normalization methods such as Z-scores are easily affected by outliers such as experimental errors in features, resulting in poor robustness.

[0020] To address the aforementioned shortcomings, the drill pipe thread ultimate bearing capacity prediction model training method provided by this invention uses Latin hypercube sampling to obtain aluminum alloy drill pipe thread failure samples, and robustly processes the sample features based on IQR truncation and additional feature construction, thereby inputting them into a PI-TransCNN model that integrates physical priors for model training. This achieves rapid prediction of ultimate bearing capacity under small sample conditions, improves the computational efficiency and prediction accuracy of ultimate bearing capacity, and ensures the reliability of ultimate bearing capacity prediction under extreme working conditions. It can be used in engineering to quickly determine the safety status of drilling operations.

[0021] Example 1 This invention provides a training method for a predictive model of the ultimate bearing capacity of drill pipe threads, specifically as follows: Figure 1 As shown, it includes the following steps: Step 11: Collect sample points from the domain space of the failure-related parameters of the drill pipe thread; construct a finite element model based on the sample points, calculate the ultimate torque value of the sample points based on the finite element model; construct a training dataset based on the failure-related parameters and ultimate torque value of the sample points.

[0022] Specifically, step 11 includes: Step 111: Select sample points in the thread failure related parameter space using the Latin hypercube sampling method.

[0023] Step 112: Using hexahedral elements, construct a three-dimensional nonlinear finite element model based on the failure-related parameters of the sample points.

[0024] Step 113: Based on geometric nonlinear constraints, contact nonlinear constraints, and material nonlinear constraints, determine the corresponding ultimate torque value from the failure-related parameters of the sample points.

[0025] Specifically, such as Figure 2 As shown, taking the aluminum alloy drill pipe body of oil drilling equipment as an example, sample points are first selected in the failure-related parameter space, such as the temperature parameter T space and the tensile load parameter F space, in the thread fracture failure area of ​​the aluminum alloy drill pipe body. For example... Figure 3 As shown, for example, samples with a failure value of 0 were eliminated within a temperature range of 25-200℃ and a tensile load range of 0-2000kN, resulting in a total of 48 sample points.

[0026] Subsequently, to reduce computational costs, hexahedral elements can be used to construct, for example... Figure 4 The diagram shows a three-dimensional nonlinear finite element model of the thread. Furthermore, to maximize the representation of the actual thread stress, calculations are performed based on failure-related parameters according to conditions such as geometric nonlinearity, contact nonlinearity, and material nonlinearity to obtain the ultimate torque value. A dataset is constructed from the failure-related parameters and the corresponding ultimate torque value for each sample point.

[0027] Step 12: Truncate the failure-related parameters in the training dataset that exceed the preset upper bound of the interquartile range to the values ​​corresponding to the upper bound of the interquartile range, and truncate the failure-related parameters in the training dataset that exceed the preset lower bound of the interquartile range to the values ​​corresponding to the lower bound of the interquartile range, to obtain the training dataset after interquartile range judgment.

[0028] Step 13: Construct additional features based on failure-related parameters, and add the additional features to the training dataset after interquartile range judgment to obtain the preprocessed training dataset.

[0029] Specifically, for key failure characteristics such as temperature parameter T and tensile load parameter F, an outlier truncation strategy based on interquartile range (IQR) is adopted, as shown in formula (1): (1) in, As the first quartile, it can be set to 25%; As the third quartile, it can be set to 75%. , This refers to limiting or truncating a given value within a range, that is, the failure-related parameter value of the i-th sample. Limit to the lower limit and upper limit Between. Furthermore, to enhance the model's ability to capture nonlinear interactions, additional features can be constructed based on key failure characteristics such as temperature parameter T and tensile load parameter F, for example, " Additional features such as "" are added to the training dataset.

[0030] Step 14: Input the training dataset into the Transformer model, and iteratively update the parameters of the neural network model with the physical information loss function as the optimization objective to obtain the thread ultimate bearing capacity prediction model. The physical information loss function includes a data-driven loss function and a physical prior loss function. The data-driven loss function is used to calculate the average prediction error of the ultimate torque value output by the Transformer model and the ultimate torque value through the Smooth L1 function. The physical prior loss function is used to calculate the theoretical value of the ultimate torque through the pre-built quadratic physical model, and calculate the average squared difference between the ultimate torque prediction value and the ultimate torque theoretical value. The quadratic physical model includes the squared terms of multiple failure-related parameters and their mutual interaction terms. The coefficients of the quadratic terms in the quadratic physical model are constrained to be non-negative.

[0031] The neural network model is a PI-TransCNN model, which includes a shallow feature extraction module, a deep feature extraction module, and a prediction output module connected in sequence. The deep feature extraction module consists of multiple stacked residual modules based on Transformer modulation. The residual module includes parallel Transformer branches and convolutional branches, as well as feature modulation units connected to the Transformer branches and convolutional branches respectively. Step 14 includes: Step 141: Extract shallow feature vectors from failure-related parameters in the training dataset using the shallow feature extraction module.

[0032] Step 142: Based on the self-attention mechanism, the shallow feature vectors are globally modeled through the Transformer branch to obtain the global feature vector; the local feature vectors are extracted from the shallow feature vectors through the convolution branch; and the global feature vectors and local feature vectors are fused through the feature modulation unit to obtain the deep feature vector.

[0033] Step 143: Determine the predicted value of the ultimate torque from the deep feature vector through the prediction output module.

[0034] Step 144: Calculate the loss of the predicted and ultimate torque values ​​based on the physical information loss function, and adjust the PI-TransCNN model based on the loss calculation results.

[0035] The physical information loss function also includes a monotonic loss function, which is used to calculate the gradient of the predicted ultimate torque value relative to the failure-related parameters, and to penalize the mean of the negative gradient using the ReLU function.

[0036] Specifically, such as Figure 5 As shown, the drill pipe thread ultimate bearing capacity prediction model training method provided by this invention solves the bottleneck of feature association utilization in ultimate torque prediction by implicitly modeling failure-related parameters and derived features such as temperature and tensile load. Among them, the PI-TransCNN model includes a shallow feature extraction module, a deep feature extraction module and a prediction output module. The shallow feature extraction module is responsible for extracting initial features from the input physical features, including basic features and interaction information; the deep feature extraction module enhances and fuses the initial features through the residual module modulated by Transformer to extract implicit feature patterns of different dimensions; the prediction output module uses the enhanced features to output the ultimate torque prediction value. The processing is shown in formula (2):

[0037] (2) in, This is the mapping function for the shallow feature extraction module. This is the mapping function for the deep feature extraction module. This is the mapping function for predicting the output module.

[0038] Specifically, the shallow feature extraction module is responsible for extracting features from the input feature vector. Extract preliminary local feature information to generate shallow feature vectors. For example, using 1D convolutional layers to capture input feature vectors The temperature and tensile load are directly correlated, and the low-dimensional input features are transformed into high-dimensional features to enhance expressive power, as shown in formula (3): (3) in, and These are the parameters of the convolutional layer. For batch normalization processing, As the activation function, the above single convolutional layer avoids the computational overhead and overfitting risk of multiple stacked layers, balancing processing efficiency and generalization performance, and laying the foundation for subsequent global feature modeling.

[0039] Then, the deep feature extraction module enhances the expressive power of shallow features by using a Transformer-modulated residual module. By stacking these residual modules, it adaptively models long-distance dependencies between features while extracting local features, thus improving the depth and breadth of feature representation. Furthermore, to balance performance and complexity, the shallow feature vectors are first processed before being input into the residual module. Dimensionality reduction is performed to obtain the shallow feature vectors after dimensionality reduction. .

[0040] Each residual module in the deep feature extraction module contains a Transformer branch, a convolutional branch, and a feature modulation unit. The Transformer branch uses a self-attention mechanism to globally model the input features, capturing long-range dependencies, such as the inverse correlation between F / T and T / F, and generating a global feature representation. Specifically, it first processes the dimensionality-reduced shallow feature vectors... Perform linear projection to obtain the query matrix. Key matrix The sum matrix V is shown in formulas (4), (5) and (6):

[0041] (4) (5) V (6) in, , and All are learnable projection matrices. Then, the query matrix is... Key matrix The sum matrix V is used for feature extraction to obtain the global feature vector, as shown in formulas (7) and (8):

[0042] (7) (8) Where A represents the output of the self-attention or attention layer. For feature dimension, This represents the fused or intermediate features after residual connections, Dropout, and layer normalization. The softmax activation function is used. This is the layer normalization function.

[0043] In the convolutional branch, local features, such as the direct interaction strength of T×F, are extracted through 1D convolution, thereby focusing on the short-range dependencies of neighboring features. For example, local features are captured by combining two layers of 1D convolution, namely depthwise convolution, with pointwise convolution, as shown in Equations (9) and (10):

[0044] (9) (10) in, and All are 1D convolutional kernels. For batch normalization processing, and For bias parameters, The features are obtained from the first layer of depthwise convolution. The features are obtained by the second layer of depthwise convolution.

[0045] Next, the outputs of the Transformer branch and the convolutional branch are adaptively fused through a feature modulation unit. This fusion ensures that the deep features output not only contain local details but also reflect global correlations by adjusting the feature weights of the convolutional branch. For example, this applies to prior texture representations. Modulation parameters are calculated using two sets of 1×1 convolutions and ReLU layers. and As shown in formulas (11) and (12):

[0046] (11) (12) Where H represents the characteristic of the Transformer branch output, and the modulation parameter. and Convolutional layers originating from the same shared parameters are independently updated during training through backpropagation and optimization algorithms, differentiated into modulation parameters with different functions, such as... As a scaling factor, As a bias term. Then, the modulation parameters are... and When applied to convolutional features, the modulated convolutional features are determined by omitting the whitening-like batch normalization operation, as shown in formula (13):

[0047] (13) in, The modulated convolutional features have dimensions similar to H. Consistent with element-wise multiplication, the modulated convolutional features enrich the prediction information, contributing to the final prediction. In summary, the deep feature extraction process is shown in formula (14):

[0048] (14) in, For deep features, and These are the mapping functions for the Transformer branch and the convolution branch, respectively. By stacking two of these residual modules, the expressive power of complex feature patterns is gradually enhanced, providing accurate feature support for extreme torque prediction.

[0049] Finally, the deep features are converted into the ultimate torque prediction value through the prediction output module composed of fully connected layers, as shown in Equation (15): (15) in, This is a single-value prediction result, specifically the predicted limit torque value. and These are the parameters for the fully connected layer. Global average pooling is used. In summary, the prediction output module performs nonlinear mapping on deep features through multiple fully connected layers and combines global average pooling to compress feature dimensions, ultimately outputting a single-value prediction result. Specifically, the prediction output module enhances the mapping ability from features to predicted values ​​through nonlinear transformations, ensuring that the output single-value prediction result reflects the complex correlations of the input features.

[0050] Then, the loss function is calculated based on the predicted limit torque value and the pre-measured limit torque value, as shown in formulas (16), (17), (18), and (19): (16) (17) (18) (19) in, For physical information loss, and Set the weights corresponding to the loss, for example, to 0.5 and 0.1 respectively. For data-driven loss, outlier identification is handled based on Smooth L1 loss. This is the predicted limit torque value from the model. The true value is the pre-measured limit torque value, and N is the number of samples. For physical prior loss, To integrate the coupling effects of temperature and load, a secondary physical model is constructed using an adaptive coefficient prediction network. Among them, “d,e,f,g,h,i,a,b,c,j” are dynamically predicted and determined by an independent MLP based on input features, and the coefficients of quadratic terms such as “d,e,f” are constrained to be non-negative by the ReLU activation function, thus conforming to the principle of energy accumulation. The monotonicity loss is used to apply a physical constraint that the limiting torque increases non-decreasingly with increasing load parameter F and temperature parameter T, and the gradient is calculated by automatic differentiation and negative gradients are penalized.

[0051] The training set can be 80% of the samples and the validation set can be 20%. The training iterations are 1000 times, the learning rate is 0.001, and the Adam optimizer is used to minimize the physical information loss of the PI-TransCNN model. The model parameters are iteratively updated to finally obtain the prediction model for the ultimate bearing capacity of the thread.

[0052] The final prediction performance of the PI-TransCNN model is shown in the figure below. Figure 6 As shown, by comparing the model's predicted values ​​with the actual values ​​in terms of residual distribution and relative error distribution, it can be seen that the PI-TransCNN model's prediction results have a certain degree of accuracy.

[0053] like Figure 7 As shown, the RMSE of the PI-TransCNN model provided by this invention is compared with that of the conventional RandomForest model, XGBoost model, KNN model and MLP model. It can be seen that the PI-TransCNN model provided by this invention has a lower overall RMSE than the conventional models, indicating that the PI-TransCNN model provided by this invention has better performance in predicting thread ultimate bearing capacity than existing models.

[0054] Step 15: Obtain the temperature parameters and tensile load parameters at the thread of the rod to be predicted.

[0055] Step 16: Input the temperature parameters and tensile load parameters into the thread ultimate bearing prediction model to obtain the ultimate bearing prediction result at the thread of the rod to be predicted.

[0056] Specifically, after the model training is complete, the temperature parameters and tensile load parameters of the aluminum alloy drill pipe thread to be predicted are input into the model to obtain the predicted ultimate torque value. This allows the ultimate bearing capacity to be determined by combining the material fracture failure criteria. For example, if the input temperature parameter is 160℃ and the tensile load parameter is 50kN, the model will give a predicted value of 82kN·m, which, combined with the material fracture failure criteria, will yield the following result: Figure 6 The safe range shown is 160°C, which is used to determine whether the drilling operation is in a safe condition.

[0057] In summary, the training method for the drill pipe thread ultimate bearing capacity prediction model provided by this invention employs Latin hypercube sampling, ensuring coverage while reducing the number of samples. Combined with the physical constraints of the PI-TransCNN model, it solves the overfitting problem of deep learning models under small sample conditions, achieving high generalization ability of the model under small sample conditions. Compared to the finite element method's single calculation exceeding 4 hours, the neural network prediction calculation time is less than 5 seconds, enabling real-time prediction on-site and achieving rapid prediction of ultimate bearing capacity. Robust feature engineering preprocessing removes outliers and enhances nonlinear features. The MLP regression module of the PI-TransCNN model captures complex patterns, and the PINN loss function, which incorporates physical priors, makes the model's prediction accuracy significantly higher than existing models. Monotonicity loss constrains the non-decreasing trend of ultimate torque with increasing load / temperature, and the physical prior loss incorporates the energy accumulation principle, making the model's prediction results consistent with physical intuition and significantly improving reliability under extreme conditions. Finally, the model can quickly predict the ultimate bearing capacity of aluminum alloy drill pipe threads, providing a safety judgment basis for drilling operations and reducing experimental and simulation costs.

[0058] Specifically, this invention provides a training method for a predictive model of the ultimate bearing capacity of drill pipe threads. Sample points are selected using Latin hypercube sampling, a method that achieves efficient and uniform sampling in the thread failure-related parameter space, ensuring comprehensive sample point coverage and avoiding localized over- or under-sparse problems that may occur with traditional random sampling. This sampling method effectively represents the entire parameter space even with a limited number of samples, thereby reducing data bias and the risk of overfitting. This directly improves the representativeness and quality of the training data, making the model more generalizable under small sample conditions, and ultimately enhancing the reliability of the predictive model by increasing the credibility of the data foundation.

[0059] Furthermore, a three-dimensional nonlinear finite element model is constructed using hexahedral elements, and the ultimate torque value is calculated based on geometric, contact, and material nonlinear constraints, thereby accurately reproducing the stress state of a real thread. By considering various nonlinear factors, the calculation of the ultimate torque value is ensured to be closer to physical reality. Compared to simplified models, this finite element model provides a more reliable source of training data, reduces simulation errors, and lays an accurate foundation for model training. By improving the physical realism of the input data, the predictive model's adaptability to complex working conditions is indirectly enhanced, thereby improving the reliability of prediction results under extreme conditions. After constructing the training set, outlier truncation based on interquartile range (IQR) and additional feature construction steps are added. IQR truncation can effectively identify and process outliers in the data, preventing these noises from affecting model training; while additional features, by introducing nonlinear interaction terms, enhance the model's ability to capture complex relationships between features. This robust feature preprocessing method improves the quality and consistency of the data, reduces instability during model training, and makes the model more adaptable to small sample sizes and noisy environments.

[0060] The physical information loss function used in model training includes data-driven loss, physical prior loss, and monotonicity loss. Data-driven loss ensures the model fits the training data; physical prior loss constrains predicted values ​​to conform to physical laws through a quadratic physical model; and monotonicity loss forces the ultimate torque to increase with load or temperature rather than decrease, aligning with structural mechanics intuition. This combination of multiple losses creates an optimization objective that allows the model to not only rely on data but also incorporate physical common sense, avoiding unreasonable predictions under extreme conditions caused by purely data-driven models.

[0061] Furthermore, after model training is complete, a step is added to obtain the temperature and tensile load parameters of the threaded rod to be predicted, and directly input them into the model to obtain the ultimate bearing capacity prediction result. This achieves a closed loop from training to application, enabling the model to quickly respond to field requirements. By combining material fracture criteria, this claim ensures that the prediction results can be directly used for drilling operation safety judgment, improving the model's practicality and real-time performance.

[0062] Example 2 This invention also provides a training system for a prediction model of the ultimate bearing capacity of drill pipe threads, comprising: The dataset construction module is used to collect sample points from the domain space of failure-related parameters of drill pipe threads; construct a finite element model based on the sample points; calculate the ultimate torque value of the sample points based on the finite element model; and construct a training dataset based on the failure-related parameters and ultimate torque values ​​of the sample points. The model training module is used to input the training dataset into the Transformer model and iteratively update the parameters of the neural network model with the physical information loss function as the optimization objective, thereby obtaining the thread ultimate bearing capacity prediction model. The physical information loss function includes a data-driven loss function and a physical prior loss function. The data-driven loss function is used to calculate the predicted ultimate torque value and the average prediction error of the ultimate torque value output by the Transformer model through the Smooth L1 function. The physical prior loss function is used to calculate the theoretical ultimate torque value through a pre-built quadratic physical model and to calculate the average squared difference between the predicted ultimate torque value and the theoretical ultimate torque value. The quadratic physical model includes the squared terms of multiple failure-related parameters and their mutual interaction terms. The coefficients of the quadratic terms in the quadratic physical model are constrained to be non-negative.

[0063] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in the embodiment of the drill pipe thread ultimate bearing capacity prediction model training method. Specific implementation methods can be found in the method embodiment, and will not be repeated here.

[0064] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions, on which a computer program is stored. For example, a memory containing instructions, which can be executed by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in the embodiment of the drill pipe thread ultimate bearing capacity prediction model training method. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0065] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A training method for a predictive model of the ultimate bearing capacity of drill pipe threads, characterized in that, include: Sample points were collected from the domain space of failure-related parameters of drill pipe threads; A finite element model is constructed based on the sample points, and the ultimate torque value of the sample points is calculated based on the finite element model. A training dataset is constructed based on the failure-related parameters and ultimate torque values ​​of the sample points; The training dataset is input into the Transformer model, and the parameters of the neural network model are iteratively updated with the physical information loss function as the optimization objective to obtain the thread ultimate bearing capacity prediction model. The physical information loss function includes a data-driven loss function and a physical prior loss function. The data-driven loss function is used to calculate the predicted ultimate torque value output by the Transformer model and the average prediction error of the ultimate torque value through the Smooth L1 function. The physical prior loss function is used to calculate the theoretical ultimate torque value from the failure-related parameters through a pre-constructed quadratic physical model, and to calculate the average squared difference between the predicted ultimate torque value and the theoretical ultimate torque value. The quadratic physical model includes the squared terms of each of the failure-related parameters and the interaction terms between them. The coefficients of the quadratic terms in the quadratic physical model are constrained to be non-negative.

2. The training method for the prediction model of the ultimate bearing capacity of drill pipe threads according to claim 1, characterized in that, Sample points were collected from the domain space of failure-related parameters of drill pipe threads: The sample points are selected in the failure-related parameter space of the drill pipe thread using the Latin hypercube sampling method.

3. The training method for the prediction model of the ultimate bearing capacity of drill pipe threads according to claim 1, characterized in that, Constructing a finite element model based on the sample points, and calculating the ultimate torque value of the sample points based on the finite element model, includes: A three-dimensional nonlinear finite element model is constructed using hexahedral elements based on the failure-related parameters of the sample points. Based on geometric nonlinear constraints, contact nonlinear constraints, and material nonlinear constraints, the corresponding ultimate torque value is determined by the failure-related parameters of the sample points.

4. The training method for the prediction model of the ultimate bearing capacity of drill pipe threads according to claim 1, characterized in that, After constructing the training dataset based on the failure-related parameters and ultimate torque values ​​of the sample points, the following is also included: The failure-related parameters in the training dataset that exceed the preset upper bound of the interquartile range are truncated to the values ​​corresponding to the upper bound of the interquartile range, and the failure-related parameters in the training dataset that exceed the preset lower bound of the interquartile range are truncated to the values ​​corresponding to the lower bound of the interquartile range, thus obtaining the training dataset after interquartile range judgment. Additional features are constructed based on the failure-related parameters, and these additional features are added to the training dataset after the interquartile range determination to obtain the preprocessed training dataset.

5. The training method for predicting the ultimate bearing capacity of drill pipe threads according to claim 4, characterized in that, The Transformer model is a PI-TransCNN model, which includes a shallow feature extraction module, a deep feature extraction module, and a prediction output module connected in sequence. The deep feature extraction module is a series of stacked residual modules based on Transformer modulation. The residual module includes parallel Transformer branches and convolutional branches, as well as feature modulation units connected to the Transformer branches and convolutional branches respectively. Inputting the training dataset into the neural network model and iteratively updating the parameters of the neural network model with the physical information loss function as the optimization objective includes: The shallow feature vector is obtained by extracting the failure-related parameters from the training dataset through the shallow feature extraction module. Based on the self-attention mechanism, the shallow feature vector is globally modeled through the Transformer branch to obtain the global feature vector; local feature vectors are extracted from the shallow feature vector through the convolution branch; and the global feature vector and local feature vector are fused through the feature modulation unit to obtain the deep feature vector. The prediction output module determines the predicted value of the ultimate torque from the deep feature vector. The loss is calculated on the predicted value of the ultimate torque and the ultimate torque value based on the physical information loss function, and the PI-TransCNN model is adjusted based on the loss calculation results.

6. The training method for the prediction model of the ultimate bearing capacity of drill pipe threads according to claim 5, characterized in that, The physical information loss function also includes a monotonic loss function, which is used to calculate the gradient of the predicted ultimate torque value relative to the failure-related parameters, and to penalize the mean of the negative gradient in the gradient using the ReLU function.

7. The training method for the prediction model of the ultimate bearing capacity of drill pipe threads according to claim 6, characterized in that, The failure-related parameters include temperature parameters and tensile load parameters. After obtaining the thread ultimate bearing capacity prediction model, it also includes: Obtain the temperature parameters and tensile load parameters at the thread of the rod to be predicted; The temperature parameters and tensile load parameters are input into the thread ultimate bearing prediction model to obtain the ultimate bearing prediction result at the thread of the rod to be predicted.

8. A training system for predicting the ultimate bearing capacity of drill pipe threads, characterized in that, include: The dataset construction module is used to collect sample points from the domain space of failure-related parameters of drill pipe threads; A finite element model is constructed based on the sample points, and the ultimate torque value of the sample points is calculated based on the finite element model. A training dataset is constructed based on the failure-related parameters and ultimate torque values ​​of the sample points; The model training module is used to input the training dataset into the Transformer model and iteratively update the parameters of the neural network model with the physical information loss function as the optimization objective to obtain a thread ultimate bearing capacity prediction model. The physical information loss function includes a data-driven loss function and a physical prior loss function. The data-driven loss function is used to calculate the predicted ultimate torque value output by the Transformer model and the average prediction error of the ultimate torque value using the Smooth L1 function. The physical prior loss function is used to calculate the theoretical ultimate torque value from the failure-related parameters using a pre-constructed quadratic physical model, and to calculate the average squared difference between the predicted ultimate torque value and the theoretical ultimate torque value. The quadratic physical model includes the squared terms of each of the failure-related parameters and their mutual interaction terms. The coefficients of the quadratic terms in the quadratic physical model are constrained to be non-negative.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the drill pipe thread ultimate bearing capacity prediction model training method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to execute the steps of the drill pipe thread ultimate bearing capacity prediction model training method according to any one of claims 1 to 7.

Images