A drilling multi-parameter collaborative prediction method and device based on multi-task learning

Abstract

The application discloses a kind of based on multi-task learning's drilling multi-parameter collaborative prediction method and device, the method includes: selecting prediction target and input parameter combination from the multidimensional drilling parameter of target well, according to prediction target, task set is constructed, task set includes multiple prediction tasks for prediction target, each prediction task is associated;The input parameter combination of each prediction task is input to multi-task prediction model, and the prediction result of each prediction task is obtained;Wherein, multi-task prediction model includes shared layer, expert layer, gate unit and tower layer, shared layer is used to extract the common characteristics of multiple prediction tasks, expert layer is used to extract the unique characteristics of each prediction task, gate unit is used to weightedly fuse common characteristics and unique characteristics, obtain the fusion characteristics of each prediction task, tower layer is used to output the prediction result of each prediction task according to the fusion characteristics of each prediction task.The application can accurately and efficiently realize the collaborative prediction of multiple drilling parameters at a lower cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas exploration, and particularly relates to a drilling multi-parameter collaborative prediction method and device based on multi-task learning. BACKGROUND

[0002] Drilling parameters not only reflect the running status of drilling equipment and the stability of the well wall, but also reveal the complexity of formation characteristics and drilling environment. Accurate prediction of drilling parameters can realize real-time monitoring and early warning of the drilling process, timely discovery of abnormal states, and guarantee of the continuity and safety of drilling operations. In addition, it can also provide reliable data support for the optimization and scheme adjustment of drilling technology, effectively reduce drilling cost, and improve operation efficiency.

[0003] However, the existing drilling parameter prediction method usually only trains for a single target to solve a certain specific task. For example, a drilling speed prediction model can only predict drilling speed, and a torque prediction model can only predict torque. However, in the drilling process, there is often a demand for multi-parameter prediction. The existing drilling parameter prediction method ignores the internal relationship and interaction between various drilling parameters, is difficult to capture the synergistic effect between parameters under complex geological conditions, and needs to build an independent model for each drilling parameter, which increases the complexity and training cost of the model. That is, the existing drilling parameter prediction method cannot accurately and efficiently realize the collaborative prediction of multiple drilling parameters (drilling multi-parameters) at a low cost.

[0004] At present, no effective solution has been proposed for the above problems. SUMMARY

[0005] The embodiments of the present application provide a drilling multi-parameter method and device based on multi-task learning to solve the problem that the prior art cannot accurately and efficiently realize the collaborative prediction of multiple drilling parameters at a low cost.

[0006] In a first aspect, the embodiments of the present application provide a drilling multi-parameter collaborative prediction method based on multi-task learning, comprising:

[0007] selecting a prediction target and an input parameter combination corresponding to the prediction target from multi-dimensional drilling parameters of a target well, constructing a task set according to the prediction target, the task set comprising a plurality of prediction tasks for the prediction target, and each prediction task having an associated relationship;

[0008] The input parameters of each prediction task are combined and input to a multi-task prediction model to obtain a prediction result of each prediction task; wherein the multi-task prediction model comprises a shared layer, an expert layer, a gating unit and a tower layer, the shared layer is used to extract common features of multiple prediction tasks, the expert layer is used to extract unique features of each prediction task, the gating unit is used to weight and fuse the common features and the unique features to obtain fusion features of each prediction task, and the tower layer is used to output the prediction result of each prediction task according to the fusion features of each prediction task.

[0009] In some embodiments, the multi-dimensional drilling parameters include logging parameters, logging parameters and engineering parameters, the logging parameters include at least one of the following: drilling pressure, torque, rotating speed, hook load, the logging parameters include at least one of the following: gamma, resistivity, and the engineering parameters include at least one of the following: drilling speed, specific energy, bottom hole pressure, formation pressure, bottom hole temperature, equivalent density.

[0010] In some embodiments, the multi-task prediction model is trained in the following manner:

[0011] The input parameter samples of each prediction task sample are combined and input to an initial multi-task prediction model to output the prediction result of each prediction task sample, and the input parameter samples are selected from the preprocessed multi-dimensional drilling parameter samples;

[0012] The prediction error of each prediction task sample is calculated, and when the prediction error of at least one prediction task sample is greater than a preset error threshold, a weight adjustment strategy is invoked to adjust the weight of each prediction task sample;

[0013] According to the adjusted weight of each prediction task sample, a total loss value is calculated, and the model parameters in the initial multi-task prediction model are updated according to the total loss value until the prediction error of each prediction task sample is less than the preset error threshold, and the multi-task prediction model is obtained.

[0014] In some embodiments, the shared layer comprises a plurality of sub-expert networks with attention mechanisms, each sub-expert network is used to process different input parameter combinations and extract corresponding common features, and the attention mechanism is used to calculate shared weights corresponding to the common features.

[0015] Correspondingly, the shared layer is used to extract common features of multiple prediction tasks, including:

[0016] The shared layer is used to weight and sum the corresponding common features based on the shared weights to obtain the common features of the multiple prediction tasks.

[0017] In some embodiments, the expert layer comprises a plurality of independent expert networks for the prediction target, each independent expert network being configured to process a different input parameter combination and extract unique features for each prediction task.

[0018] In some embodiments, the gating unit comprises a fully connected layer, a ReLU activation layer and a Softmax activation layer.

[0019] Correspondingly, the gating unit is configured to weight and fuse the general features and the unique features, comprising:

[0020] The gating unit is configured to concatenate the general features and the unique features, activate the concatenated features through the ReLU activation layer after inputting the concatenated features into the fully connected layer, and output the gating weights for each prediction task based on the Softmax activation layer.

[0021] The general features and the unique features for each prediction task are weighted and fused according to the gating weights for each prediction task to obtain the fused features for each prediction task.

[0022] In some embodiments, the weighting and fusing of the general features and the unique features for each prediction task according to the gating weights for each prediction task comprises:

[0023] The general features and the unique features for each prediction task are weighted and fused according to the following formula:

[0024] h final =g. h shared +(1-g). h task

[0025] wherein h final is the fused features for the prediction task, g is the gating weight for the prediction task, h shared is the general features for the plurality of prediction tasks, and h task is the unique features for the prediction task.

[0026] In a second aspect, the embodiments of the present specification also provide a drilling multi-parameter collaborative prediction device based on multi-task learning, comprising:

[0027] A construction module is configured to select a prediction target and an input parameter combination corresponding to the prediction target from multi-dimensional drilling parameters of a target well, construct a task set according to the prediction target, and the task set comprises a plurality of prediction tasks for the prediction target, and each prediction task has a correlation relationship.

[0028] The cooperative prediction module is configured to combine input parameters of each prediction task into the multi-task prediction model to obtain a prediction result of each prediction task; wherein the multi-task prediction model comprises a shared layer, an expert layer, a gating unit and a tower layer, the shared layer is configured to extract common features of multiple prediction tasks, the expert layer is configured to extract unique features of each prediction task, the gating unit is configured to weight and fuse the common features and the unique features to obtain fused features of each prediction task, and the tower layer is configured to output the prediction result of each prediction task according to the fused features of each prediction task.

[0029] In a third aspect, the embodiments of the present specification also provide a computer device, comprising a memory, a processor and a computer program / instruction stored in the memory, wherein the processor executes the computer program / instruction to implement the steps of the multi-task learning based drilling multi-parameter cooperative prediction method described above.

[0030] In a fourth aspect, the embodiments of the present specification also provide a computer readable storage medium, which stores a computer program / instruction, wherein the computer program / instruction is executed by a processor to implement the steps of the multi-task learning based drilling multi-parameter cooperative prediction method described above.

[0031] The embodiment of the present specification provides a drilling multi-parameter collaborative prediction method and device based on multi-task learning. First, a prediction target and an input parameter combination corresponding to the prediction target are selected from multi-dimensional drilling parameters of a target well. A task set is constructed according to the prediction target. The task set includes multiple prediction tasks for the prediction target, and each prediction task has an associated relationship. Then, the input parameter combination of each prediction task is input into a multi-task prediction model to obtain a prediction result of each prediction task. The multi-task prediction model includes a shared layer, an expert layer, a gating unit and a tower layer. The shared layer is used to extract common features of multiple prediction tasks. The expert layer is used to extract unique features of each prediction task. The gating unit is used to weight and fuse the common features and the unique features to obtain fusion features of each prediction task. The tower layer is used to output the prediction result of each prediction task according to the fusion features of each prediction task. In the embodiment of the present specification, by selecting the prediction target and the input parameter combination corresponding to the prediction target from the multi-dimensional drilling parameters of the target well and constructing the task set according to the prediction target, it can be ensured that the input data set of the prediction target belongs to the same well, that is, the input data set has an associated relationship. In addition, each prediction task also has an associated relationship. Therefore, on the premise of considering the mutual influence and internal relationship between different drilling parameters, the multi-task prediction model can accurately and efficiently realize the collaborative prediction of multiple drilling parameters, reduce the model redundant parameters and training cost. The multi-task prediction model includes the shared layer, the expert layer, the gating unit and the tower layer. The shared layer can extract common features of multiple prediction tasks, and the expert layer can extract unique features of each prediction task, so as to ensure that the multi-task prediction model can capture the commonality between the prediction tasks and retain the characteristics of each prediction task. By weighting and fusing the common features and the unique features through the gating unit, negative transfer can be avoided. Finally, based on the tower layer, the prediction result of each prediction task can be output according to the fusion features of each prediction task, so as to accurately and efficiently realize the collaborative prediction of multiple key drilling parameters in the drilling process, greatly reduce the training cost, improve the model generalization ability, and provide technical support for efficient drilling under complex geological conditions. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief introductions will be given to the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort. In the drawings:

[0033] Figure 1 is a relationship diagram among multiple models involved in the drilling process provided by the embodiment of the present specification;

[0034] Figure 2is a flow diagram of a drilling multi-parameter collaborative prediction method based on multi-task learning provided by an embodiment of the present specification;

[0035] Figure 3 is an architecture comparison diagram of a single-task learning method and a multi-task learning method provided by an embodiment of the present specification;

[0036] Figure 4 is an architecture diagram of a drilling multi-parameter collaborative prediction method based on multi-task learning provided by an embodiment of the present specification;

[0037] Figure 5 is a whole flow diagram of a drilling multi-parameter collaborative prediction system based on multi-task learning provided by an embodiment of the present specification;

[0038] Figure 6 is a prediction effect diagram of a multi-task prediction model on three important drilling parameters provided by an embodiment of the present specification;

[0039] Figure 7 is a prediction effect diagram of a multi-task prediction model on three logging-while-drilling parameters provided by an embodiment of the present specification;

[0040] Figure 8 is a prediction effect diagram of a multi-task prediction model on another three logging-while-drilling parameters provided by an embodiment of the present specification;

[0041] Figure 9 is a structural composition diagram of a drilling multi-parameter collaborative prediction device based on multi-task learning provided by an embodiment of the present specification;

[0042] Figure 10 is a structural composition diagram of an electronic device provided by an embodiment of the present specification. DETAILED DESCRIPTION

[0043] In order to enable persons skilled in the art to better understand the technical solutions in the present specification, the technical solutions in the embodiments of the present specification will be described clearly and completely in conjunction with the drawings in the embodiments of the present specification. Obviously, the described embodiments are only some of the embodiments of the present specification, not all the embodiments. Based on the embodiments in the present specification, all other embodiments obtained by persons skilled in the art without creative labor should belong to the scope of protection of the present specification.

[0044] Drilling parameter prediction, as one of the key steps of drilling state characterization, is of great significance to ensure the efficiency, safety and economy of drilling operations. Various drilling parameters are involved in the drilling process, such as drilling parameters in mud logging parameters, such as drilling pressure, torque, rotation speed, hook load, etc., logging parameters such as gamma, resistivity, etc., engineering parameters such as drilling speed, specific energy, bottom hole pressure, etc. These parameters not only reflect the running status of drilling equipment and the stability of well wall, but also reveal the complexity of formation characteristics and drilling environment. Accurate prediction of drilling parameters can realize real-time monitoring and early warning of the drilling process, and timely detection of abnormal conditions, so as to avoid accidents and other potential risks that may occur in the drilling process, and ensure the continuity and safety of drilling operations. In addition, accurate prediction of drilling parameters can also provide reliable data support for the optimization and scheme adjustment of drilling technology, effectively reduce drilling cost, and improve operation efficiency. Especially in complex geological conditions, accurate prediction of drilling parameters has far-reaching significance for the intelligent and automated management of drilling operations.

[0045] However, the drilling parameter prediction model used by the existing drilling parameter prediction method is usually trained only for a single target to solve a certain specific task, which is called single task learning (STL). For example, a drilling speed prediction model can only predict drilling speed, and a torque prediction model can only predict torque. However, in the drilling process, there is often a need for multi-parameter prediction, and the traditional drilling parameter prediction model ignores the internal relationship and interaction between parameters, making it difficult to capture the synergistic effect between parameters under complex geological conditions. Figure 1 As shown in FIG. 1, drilling process may involve drilling speed model, pipe string model, flow model and formation pressure profile. The drilling speed model can be constructed according to the surface control variables such as drilling pressure, rotation speed, displacement, etc., the pipe string model can be constructed according to other logging data such as torque, natural gamma, etc., and the pipe string model can analyze the mechanical behavior of the drill string. There is a coupling relationship or mutual influence between the above models, such as the pipe string model can pass the downhole parameters (such as drill string vibration) to the drilling speed model as input parameters, which may cause drilling pressure fluctuations and affect the actual drilling speed prediction. The drilling speed model can pass the drilling speed change to the pipe string model as a dynamic boundary condition, which affects the torque, axial load and vibration characteristics of the drill string. Drilling speed also determines the rate of cuttings generation, which directly affects the cuttings concentration and drilling fluid rheological property in the flow model, and the cuttings removal efficiency of the drilling fluid affects the drilling bit cleaning degree, and the cuttings accumulation will reduce the drilling speed. The formation pressure profile provides pressure boundary conditions (pressure constraints) for drilling fluid flow to ensure that the wellbore pressure is within the safety window, and the formation pressure profile also determines the drilling bit rock breaking efficiency, which directly affects the drilling speed. Drill string rotation also affects flow, and drilling fluid flow also affects the drill string.

[0046] Secondly, since each parameter is modeled separately, this method is less robust in the face of data noise and outliers, and is easily affected by local data bias, affecting the overall prediction accuracy. In addition, the single-task learning mode often needs to build an independent model for each parameter, which not only increases the cost of training and maintenance, but also causes a significant delay in time consumption, low data utilization, and a large number of potential shared features and information that cannot be effectively extracted due to model architecture isolation, further weakening the prediction accuracy of the model.

[0047] To solve the above problems, the embodiments of the present specification provide a drilling multi-parameter collaborative prediction method and device based on multi-task learning, first selecting a prediction target and an input parameter combination corresponding to the prediction target from the multi-dimensional drilling parameters of the target well, constructing a task set according to the prediction target, the task set includes a plurality of prediction tasks for the prediction target, and each prediction task has an associated relationship. Then input the input parameter combination of each prediction task into the multi-task prediction model to obtain the prediction result of each prediction task; wherein the multi-task prediction model includes a shared layer, an expert layer, a gating unit and a tower layer, the shared layer is used to extract general features of multiple prediction tasks, the expert layer is used to extract unique features of each prediction task, the gating unit is used to weightedly fuse the general features and the unique features to obtain the fusion features of each prediction task, and the tower layer is used to output the prediction result of each prediction task according to the fusion features of each prediction task.

[0048] Through the above scheme, the problems of low prediction accuracy, insufficient generalization ability and information island between each drilling parameter in the prior art can be solved. By constructing a multi-task prediction model including a shared layer, an expert layer, a gating unit and a tower layer, the shared layer can fully exploit the internal correlation of multi-dimensional drilling parameters, the expert layer can extract unique features of each prediction task, the gating unit can weightedly fuse the general features and the unique features, and then based on the tower layer, the prediction result of each prediction task can be output according to the fusion features of each prediction task. The collaborative prediction of multiple key drilling parameters in the drilling process can be accurately and efficiently realized, the training cost is greatly reduced, the model generalization ability is improved, and effective data support is provided for the optimization of drilling technology, real-time early warning and risk prevention and control of abnormal state.

[0049] It can be understood that the above method provided by the embodiments of the present specification can be applied to an electronic device, and the electronic device can refer to an electronic device with data calculation, processing and storage capabilities. The electronic device can be a terminal such as a PC (Personal Computer), a tablet computer, a smart phone, a wearable device, a smart robot, etc.; or a server. The server can be a standalone physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0050] ReferenceFigure 2 As shown, the embodiments of the present specification provide a drilling multi-parameter collaborative prediction method based on multi-task learning. In specific implementation, the method can include the following contents.

[0051] S201: selecting a prediction target and an input parameter combination corresponding to the prediction target from multi-dimensional drilling parameters of a target well, constructing a task set according to the prediction target, the task set including multiple prediction tasks for the prediction target, and each prediction task having a correlation relationship;

[0052] S202: inputting the input parameter combination of each prediction task into a multi-task prediction model to obtain a prediction result of each prediction task; wherein the multi-task prediction model includes a shared layer, an expert layer, a gating unit and a tower layer, the shared layer is used to extract common features of multiple prediction tasks, the expert layer is used to extract unique features of each prediction task, the gating unit is used to weight and fuse the common features and the unique features to obtain fusion features of each prediction task, and the tower layer is used to output the prediction result of each prediction task according to the fusion features of each prediction task.

[0053] Based on the above embodiments, by selecting a prediction target and an input parameter combination corresponding to the prediction target from multi-dimensional drilling parameters of a target well, and constructing a task set according to the prediction target, it can be ensured that the input data set of the prediction target belongs to the same well, i.e. the input data set has a correlation. In addition, there is also a correlation relationship between each prediction task, so that the collaborative prediction of multiple drilling parameters can be accurately and efficiently realized based on the multi-task prediction model under the premise of considering the mutual influence and internal relationship between different drilling parameters, and the model redundancy parameters and training cost are reduced. Specifically, the multi-task prediction model includes a shared layer, an expert layer, a gating unit and a tower layer, the shared layer can extract common features of multiple prediction tasks, and the expert layer can extract unique features of each prediction task, which can ensure that the multi-task prediction model can capture the commonality between the prediction tasks and retain the characteristics of each prediction task. By weighting and fusing the common features and the unique features through the gating unit, negative transfer can be avoided. The above multi-task prediction model can not only utilize the correlation between multiple prediction tasks to improve the generalization ability, but also maintain the prediction accuracy of each prediction task. Based on the multi-task prediction model, intelligent collaborative prediction of drilling parameters can be realized, which provides technical support for efficient drilling under complex geological conditions.

[0054] In some embodiments, the above multi-dimensional drilling parameters can include logging parameters, logging parameters and engineering parameters, the logging parameters can include at least one of the following: drilling pressure, torque, rotation speed, hook load, the logging parameters can include at least one of the following: gamma, resistivity, and the engineering parameters can include at least one of the following: drilling speed, specific energy, bottom hole pressure, formation pressure, bottom hole temperature, equivalent density.

[0055] Specifically, the logging parameters, the well logging parameters and the engineering parameters of a well (i.e., a target well) can be collected to ensure that the multi-dimensional drilling parameters share the same data set, thereby providing one of the necessary conditions for the collaborative prediction of the multiple drilling parameters. The logging data can include, but is not limited to, the drilling pressure, the torque, the rotation speed, and the hook load. The well logging parameters can include, but are not limited to, the gamma and the resistivity. The engineering parameters can include, but are not limited to, the drilling speed, the specific energy, the bottom hole pressure, the formation pressure, the bottom hole temperature, and the equivalent density.

[0056] According to the actual prediction requirements, the prediction target and the input parameter combination corresponding to the prediction target can be selected from the multi-dimensional drilling parameters of the target well. For example, the prediction target is selected to be the drilling speed (the mechanical drilling speed), the formation pressure and the bottom hole pressure, and the drilling speed prediction task A, the formation pressure prediction task B and the bottom hole pressure prediction task C can be constructed accordingly. When the prediction target is the drilling speed, the input parameter combination corresponding thereto can be the parameters related to the mechanical drilling speed, such as the drilling pressure, the rotation speed and the torque. When the prediction target is the formation pressure, the input parameter combination corresponding thereto can be the parameters related to the formation pressure, such as the gamma, the resistivity and the acoustic travel time. When the prediction target is the bottom hole pressure, the input parameter combination corresponding thereto can be the parameters related to the bottom hole pressure, such as the bottom hole temperature, the equivalent circulating density and the mud performance.

[0057] Before the collaborative prediction of the multiple drilling parameters (e.g., the collaborative prediction of the drilling speed, the formation pressure and the bottom hole pressure), in addition to the multi-dimensional drilling parameters collected from the same well to ensure that the multi-dimensional drilling parameters share the same data set, it is also necessary to ensure that the prediction tasks in the task set have a correlation relationship. For example, the increase of the mechanical drilling speed will increase the amount of cuttings, resulting in the fluctuation of the equivalent circulating density, and further affecting the bottom hole pressure. In addition, it is also necessary to ensure that the different prediction tasks belong to different aspects of the same system. For example, the key drilling parameters such as the mechanical drilling speed, the torque and the equivalent circulating density have the physical coupling relationship such as the mechanical transmission and the energy conservation in the drilling process.

[0058] After meeting the above necessary conditions, the learning process of each prediction task can be optimized by sharing common features based on the multi-task prediction model, improving the prediction accuracy and the generalization ability of the model. Among them, the multi-task prediction model is implemented based on the multi-task learning (MTL) framework, which is a machine learning method that can handle multiple related tasks simultaneously during model training to achieve feature sharing and knowledge transfer between tasks. In the problem of drilling parameter prediction, multiple key drilling parameters can be considered as related tasks and jointly trained in the same model through the multi-task learning framework. Compared with single-task learning, multi-task learning has obvious advantages in the following aspects: (1) reduces model complexity and parameter quantity, improves computational efficiency and model deployment efficiency; (2) can fully exploit the internal correlation between drilling parameters, thereby improving the overall accuracy of prediction; (3) can significantly improve data utilization and reduce overfitting risk by sharing underlying information (common features); (4) the collaborative training mode helps to enhance the generalization ability and robustness of the model, making it perform more stably in complex drilling environments.

[0059] As shown in Figure 3 Single-task learning STL, task 1, task 2, task n need to train corresponding model 1, model 2, model n, based on model 1 output target 1, based on model 2 output target 2, based on model n output target n, this way training cost is high, model generalization ability is insufficient. And multi-task learning MTL, can train a model based on task 1, task 2, task n, output target 1, target 2, target n, can realize real-time, efficient collaborative prediction of multiple related tasks, greatly reduce the training cost, improve the model generalization ability.

[0060] In some embodiments, the above multi-task prediction model can be trained in the following way:

[0061] The input parameter samples of each prediction task sample are combined and input into the initial multi-task prediction model, and the prediction results of each prediction task sample are output, and the input parameter sample combination is selected from the preprocessed multi-dimensional drilling parameter samples;

[0062] The prediction error of each prediction task sample is calculated, and when the prediction error of at least one prediction task sample is greater than a preset error threshold, the weight adjustment strategy is called to adjust the weight of each prediction task sample;

[0063] According to the adjusted weight of each prediction task sample, the total loss value is calculated and the model parameters in the initial multi-task prediction model are updated according to the total loss value, until the prediction error of each prediction task sample is less than the preset error threshold, and the multi-task prediction model is obtained.

[0064] Specifically, due to the complexity of the downhole environment, the logging parameter samples and the recording parameter samples in the multi-dimensional drilling parameter samples will be affected by the sensor accuracy, the environment, the collection and transmission equipment, and the personnel operation in the process of collection and transmission, and there are missing values, abnormal values, and noise, which leads to a decrease in data quality and affects the prediction effect. Therefore, in order to adapt to the high requirements of the subsequent multi-task prediction model on data quality, the multi-dimensional drilling parameter samples can be preprocessed before modeling. The preprocessing can include missing value processing, abnormal value processing, data filtering and noise reduction, data normalization / standardization and the like. The missing values can be filled by averaging the data before and after to ensure the continuity of the data sequence. The abnormal values can be processed by the box plot method to identify and process abnormal points far from other observation values. The Savitzky-Golay filter can be used to filter and denoise the data to improve the quality of the data samples. The min-max normalization method and the z-score standardization method can be used to compress the data values to 0~1 and scale them to a state with a mean of 0 and a variance of 1.

[0065] It should be noted that the above input parameter sample combination and the corresponding prediction target sample can be selected from the preprocessed multi-dimensional drilling parameter samples, and a task sample set can be constructed according to the prediction target sample. The task sample set can include a plurality of prediction task samples for the prediction target sample. The data type of the above multi-dimensional drilling parameter sample can be the same as the data type of the above multi-dimensional drilling parameter, which will not be described herein. In addition, the above multi-dimensional drilling parameter can also be preprocessed and then a task set is constructed to improve the prediction accuracy of the multi-task prediction model.

[0066] After outputting the prediction results of each prediction task sample, the prediction error (the prediction error can also be referred to as mean square error MSE) of each prediction task sample can be calculated according to the following formula:

[0067]

[0068] Wherein, MSE is the mean square error, which represents the average of the square error between the model prediction value and the true value; n is the number of prediction task samples; y i is the true value of the i-th prediction task sample; is the prediction result or prediction value of the i-th prediction task sample.

[0069] Afterwards, the weight adjustment strategy can be invoked to adjust the weight of each prediction task sample when the prediction error of at least one prediction task sample is greater than a preset error threshold. The preset error threshold can be set according to actual needs, and the present specification does not make specific limitations thereon. The weight of each prediction task sample can be adjusted according to actual needs from one of the following weight adjustment strategies:

[0070] (1) Fixed weight method: based on experience or experimental test, a fixed weight is assigned to each task.

[0071] (2) Uncertainty weighting method: automatically adjust the loss weight based on the uncertainty of the task. The weight of each task is a function of its uncertainty, that is, tasks with high uncertainty will be given lower weights.

[0072] (3) Gradient normalization method: adjust the gradient of each task loss function to a similar order of magnitude, so that the model learns different tasks at similar speeds as much as possible.

[0073] (4) Dynamic weighted average method: use the speed of each task loss to measure the learning speed of each task, and then adaptively calculate the weight of different tasks.

[0074] (5) Multi-objective optimization method: consider the model training process containing loss functions of different tasks as a multi-objective optimization problem, and solve its Pareto front.

[0075] Afterwards, the prediction error of each prediction task sample can be taken as the loss value of each prediction task sample, and the total loss value can be calculated according to the adjusted weight of each prediction task sample and the loss value of each prediction task sample according to the following formula:

[0076]

[0077] wherein, L MTL is the total loss value or the overall loss value; w i is the weight of the i-th prediction task sample; L i is the loss value of the i-th prediction task sample.

[0078] Afterwards, the model parameters in the initial multi-task prediction model can be updated according to the total loss value according to the following formula:

[0079]

[0080] wherein, is the updated model parameter; is the model parameter before updating; L r is the learning rate; w i is the weight of the i-th prediction task sample; L iLoss value of the i-th prediction task sample.

[0081] Finally, the input parameter sample of the prediction task sample can be combined and input to the initial multi-task prediction model after updating the model parameters, the prediction error of the prediction result output by the initial multi-task prediction model after updating the model parameters and the true value is calculated, if the prediction error is less than the preset error threshold, the model training is stopped, and the multi-task prediction model with the minimum MSE is obtained. Otherwise, the process of updating the model parameters is repeated until the prediction error is less than the preset error threshold.

[0082] In some embodiments, the above-mentioned multi-task prediction model can include a parameter input layer, which can receive the pre-processed multi-dimensional drilling parameters. The parameter input layer can serve as a data input portal to pass the received parameters or data to the shared layer and the expert layer.

[0083] In some embodiments, the above-mentioned shared layer can include a plurality of sub-expert networks with attention mechanisms, each sub-expert network can be used to process different input parameter combinations and extract corresponding common features, and the above-mentioned attention mechanism can be used to calculate shared weights corresponding to the common features.

[0084] Correspondingly, the above-mentioned shared layer is used to extract general features of multiple prediction tasks, and in specific implementation, can include:

[0085] The shared layer is used to weight and sum the corresponding common features based on the shared weights to obtain general features of multiple prediction tasks.

[0086] Specifically, the shared layer can include a plurality of sub-expert networks with attention mechanisms, and the sub-expert network can include a GRU network or an LSTM network or a Transformer+CNN network, which is not limited in the present specification, that is, the shared layer can use different networks according to actual needs. The shared layer can learn the common rules of multiple prediction tasks and extract general features.

[0087] Each sub-expert network can be used to process different input parameter combinations and extract corresponding common features. For example, assuming that mechanical drilling speed (ROP), bottom hole pressure (BHP), and formation pressure (FP) are to be predicted cooperatively, the input parameters can include: well depth, mechanical drilling speed, bottom hole pressure, formation pressure, displacement, torque, acoustic travel time, natural gamma, etc. The shared layer sets three GRU networks (GRU_1, GRU_2, GRU_3) with attention mechanisms, each GRU network processes different input parameter combinations and extracts corresponding common features, for example, GRU_1 extracts time series features, GRU_2 extracts local features, and GRU_3 extracts global features.

[0088] The attention mechanism can be used to calculate the shared weight corresponding to the common feature output by each sub-expert network. Finally, the shared layer can perform weighted summation on the corresponding common feature based on the shared weight to obtain the general feature of multiple prediction tasks. For example, the shared weight corresponding to the common feature can be calculated according to the following formula:

[0089]

[0090] wherein h i is the common feature output by each sub-expert network; 、 is a learnable parameter; is the shared weight corresponding to the common feature (i.e., the attention weight of each sub-expert network).

[0091] The general feature of multiple prediction tasks can be obtained by performing weighted summation according to the following formula:

[0092]

[0093] wherein h shared is the general feature of multiple prediction tasks; is the shared weight corresponding to the common feature; and h i is the common feature output by each sub-expert network.

[0094] It should be noted that the common features in the shared layer can be fully shared or partially shared, that is, the shared layer can dynamically adjust which common features output by each sub-expert network are transmitted to the expert layer through the self-attention mechanism. In this way, the shared layer can balance the learning needs of commonality and difference between prediction tasks and avoid the negative transfer risk caused by the traditional fixed sharing mode.

[0095] In some embodiments, the above-mentioned expert layer can include multiple independent expert networks for prediction targets, and each independent expert network is used to process different input parameter combinations and extract unique features of each prediction task.

[0096] Specifically, each prediction task or prediction target has an independent expert network, and the network type (such as MLP, lightweight CNN, etc.) can be selected according to actual needs. Each independent expert network can receive the general features output by the shared layer and the input parameter combination input by each prediction task, refine the general features output by the shared layer, capture the unique nonlinear mapping relationship of the input parameter combination, and extract the unique features of the corresponding prediction task. For example, assuming that the mechanical drilling speed (ROP), well bottom pressure (BHP), and formation pressure (FP) are to be predicted cooperatively, the input parameters can include well depth, mechanical drilling speed, well bottom pressure, formation pressure, displacement, torque, acoustic travel time, natural gamma, etc. Correspondingly, the expert layer can include a drilling speed prediction expert network, a well bottom pressure expert network, and a formation pressure expert network. The drilling speed prediction expert network can extract unique features or exclusive features that only affect the drilling speed based on the general features output by the shared layer. The well bottom pressure expert network can extract unique features that only affect the well bottom pressure based on the general features output by the shared layer. The formation pressure expert network can extract unique features that only affect the formation pressure based on the general features output by the shared layer.

[0097] Among them, the drilling speed prediction expert network can include a HydroNet sub-network (which can be referred to as a first expert network) responsible for capturing the influence of hydraulic factors on drilling speed, a GeoNet sub-network (which can be referred to as a second expert network) responsible for representing the influence of formation properties on drilling speed, a WearNet sub-network (which can be referred to as a third expert network) responsible for representing the influence of drill bit wear on drilling speed, and an ROPNet sub-network (which can be referred to as a fourth expert network) for representing the mapping relationship between drilling speed and hydraulic factors, formation properties, and drill bit wear. The information extracted by the HydroNet sub-network, the GeoNet sub-network, and the WearNet sub-network can be integrated with the general features output by the shared layer first, and then transmitted to the ROPNet sub-network, and finally output the unique features that only affect the drilling speed (which can be referred to as drilling speed unique features).

[0098] The well bottom pressure expert network can include a Net 1 (FCNN) network (which can be referred to as a fifth expert network) responsible for processing depth and time, a Net 2 (GRU) network (which can be referred to as a sixth expert network) responsible for processing time series data, and a Net 3 network (which can be referred to as a seventh expert network) responsible for information integration. The information extracted by the Net 1 (FCNN) network responsible for processing depth and time and the Net 2 (GRU) network responsible for processing time series data can be integrated with the general features output by the shared layer first, and then transmitted to the Net 3 network responsible for information integration, and finally output the unique features that only affect the well bottom pressure (which can be referred to as well bottom pressure unique features).

[0099] The formation pressure expert network can include a Net 1 (FCNN) network (which can be referred to as an eighth expert network) responsible for processing parameters such as DT, a Net 2 (GRU) network (a sixth expert network) responsible for processing time series data, and a Net 3 network (a seventh expert network) responsible for information integration. The information extracted by the Net 1 (FCNN) network responsible for processing parameters such as DT and the Net 2 (GRU) network responsible for processing time series data can be integrated with the general features output by the shared layer first, and then transmitted to the Net 3 network responsible for information integration, and finally output the unique features (which can be referred to as formation pressure unique features) that only affect the formation pressure.

[0100] By setting each prediction task to have an independent expert network, each independent expert network can extract unique features for the corresponding prediction task, and the task-specific features can be retained, that is, the differences can be retained.

[0101] In some embodiments, the above gating unit can include a fully connected layer, a ReLU activation layer, and a Softmax activation layer.

[0102] Correspondingly, the above gating unit is used to weight and integrate the general features and the unique features, and in specific implementation, can include:

[0103] The gating unit is used to concatenate the general features and the unique features, and after inputting the concatenated features to the fully connected layer, the features are activated by the ReLU activation layer, and the gating weights of each prediction task are output based on the Softmax activation layer.

[0104] The general features and the unique features of each prediction task are weighted and integrated according to the gating weights of each prediction task to obtain the integrated features of each prediction task.

[0105] In some embodiments, the above weighting and integration of the general features and the unique features of each prediction task according to the gating weights of each prediction task can include:

[0106] The general features and the unique features of each prediction task are weighted and integrated according to the following formula:

[0107] h final = g. h shared + (1-g). h task

[0108] where h final is the integrated features of the prediction task; g is the gating weight; h shared is the general features of the multiple prediction tasks; and h task is the unique features of the prediction task.

[0109] Specifically, the gating unit can quantify the dependency strength between prediction tasks, dynamically select and combine the output results of different expert layers according to the needs of the prediction task, and adjust the information flow from the shared layer to the task branch. Each prediction task can contain an information-sharing gating unit, which is composed of a fully connected layer with 32 neurons, a ReLU activation layer, and a final Softmax activation layer connected in series. The gating unit of each prediction task can receive the general features output by the shared layer and the specific features of each prediction task output by the expert layer, and output the weighted fusion to the corresponding tower layer of each prediction task.

[0110] Specifically, the gating unit for each prediction task can calculate the gating weight for each prediction task according to the following formula:

[0111] g=Softmax(W g [h shared ; h task ]+b g )

[0112] Where g is the gating weight for the prediction task; W g b g h is the gating parameter. shared This is a common feature for multiple prediction tasks; h task For features specific to the prediction task; [h shared ; h task [] represents the spliced ​​features.

[0113] After determining the gating weights for each prediction task, the weighted sum of the specific and general features of each prediction task can be performed according to the following formula to obtain the fused features of each prediction task:

[0114] h final =g. h shared +(1-g) . h task

[0115] Among them, h final The fusion features for the prediction task; g represents the gating weights for the prediction task; h represents the fusion features for the prediction task. shared This is a common feature for multiple prediction tasks; h task These are features specific to the prediction task. The gating weights determine the fusion ratio between general features and prediction task-specific features.

[0116] By setting a corresponding gating unit for each prediction task, the gating unit of each prediction task can determine the gating weight of each prediction task, thereby determining the fusion ratio of general features and features specific to each prediction task, and thus accurately determining the fusion features of each prediction task, improving the data foundation for the subsequent output of prediction results of each prediction task.

[0117] In some embodiments, the common features output by the sub-expert networks in the shared layer are allocated by attention weights, which determine which features are shared. The weights can be dynamically calculated by task relevance. For example:

[0118] Suppose task A (drilling speed) and task B (bottom hole pressure) are highly related in a certain section of the well (e.g., an increase in drilling speed causes pressure fluctuations), the shared layer will automatically increase the proportion of shared features for both tasks (e.g., 70%). If task C (formation pressure) is weakly related to other tasks, the shared layer will reduce its sharing proportion (e.g., 30%), and rely more on task-specific expert networks.

[0119] The above gating unit can dynamically adjust the weights of shared features and task-specific features to balance commonality and difference. The information fusion ratio of the shared layer and the expert layer can be dynamically adjusted according to task requirements. For example:

[0120] Suppose there are task A (drilling speed) and task B (bottom hole pressure), task A is highly related to shared features (e.g., drilling speed is significantly affected by general drilling parameters), then the gating unit will emphasize the general features output by the shared layer for task A. If task A needs to rely on expert layer-specific features, the gating unit will strengthen the role of expert layer features for task A. In addition, during the training process, if the prediction error of task A is high, the multi-task prediction model will also automatically adjust the weight distribution of the gating unit to explore a better sharing and expert feature ratio. If the correlation between tasks changes, the gating unit will also adaptively adjust the shared layer weight to avoid negative transfer.

[0121] In some embodiments, the above tower layer can be used to generate the final prediction value based on the fused features. For example, the fused features output by the gating unit are input into a fully connected network to map them to the prediction result of a specific parameter, such as inputting the fused features output by the gating unit into a fully connected layer to regress the predicted mechanical drilling speed. Each prediction task can have a corresponding tower layer, and each prediction task's tower layer can cooperatively output the prediction result of each prediction task based on the fused features of each prediction task.

[0122] Based on the above embodiments, the following technical effects can be achieved: 1. Improved multi-parameter collaborative prediction accuracy: the drilling parameter prediction accuracy is significantly improved compared to single-task learning methods. 2. Enhanced model generalization ability and robustness. 3. Intensive use of computing resources: the shared layer architecture significantly reduces redundant parameters, training time is significantly shortened compared to the sum of multiple single-task models, and GPU memory usage is significantly reduced. The present invention can effectively solve the problems of complex, low-precision, and poor generalization of existing drilling parameter prediction processes, providing a new approach for drilling parameter prediction, and having high practical value and wide application prospects.

[0123] In some embodiments, after S201, in specific implementation, the following steps can also be included:

[0124] According to the prediction results of each prediction task, the stratum parameter fluctuation law is captured; and according to the stratum parameter fluctuation law, the abnormal working condition in the drilling operation is identified and a risk early warning is performed on the abnormal working condition.

[0125] Specifically, after obtaining the prediction results (cooperative prediction results) of each prediction task, real-time risk early warning (cooperative early warning) can be performed to ensure the continuity and safety of the drilling operation, and to provide technical support for efficient drilling under complex geological conditions.

[0126] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. Specifically, the description of the related processing of the foregoing embodiments can be referred to, and will not be described here.

[0127] The above describes the present method, however, it is worth noting that the specific embodiments are only for better illustrating the present application, and the description of the specific embodiments in the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve the desired result. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

[0128] In a specific implementation scenario, refer to Figure 4As shown, task A can be a drilling rate prediction task, task B can be a formation pressure prediction task, and task C can be a bottom hole pressure prediction task. The input parameters of task A can be weight on bit, rotary speed, and torque, the input parameters of task B can be gamma, resistivity, and acoustic travel time, and the input parameters of task C can be bottom hole temperature, equivalent circulating density, and mud performance. All drilling parameters can be weight on bit, rotary speed, torque, gamma, resistivity, acoustic travel time, bottom hole temperature, equivalent circulating density, and mud performance. Task A can be input into Expert A (drilling rate prediction expert network) of task A to extract drilling rate specific features, task B can be input into Expert B (bottom hole pressure expert network) of task B to extract bottom hole pressure specific features, and task C can be input into Expert C (formation pressure expert network) of task C to extract formation pressure specific features. Task A, task B, and task C can be input into the shared Expert (i.e., the shared layer) to extract common features of task A, task B, and task C. Expert A can further input the drilling rate specific features to the gating unit of task A, and the gating unit of task A can fuse the drilling rate specific features and the common features according to the gating weights of task A. Expert B can further input the bottom hole pressure specific features to the gating unit of task B, and the gating unit of task B can fuse the bottom hole pressure specific features and the common features according to the gating weights of task B. Expert C can further input the formation pressure specific features to the gating unit of task C, and the gating unit of task C can fuse the formation pressure specific features and the common features according to the gating weights of task C. Finally, the gating unit of task A can transmit the fused features of task A to Tower A of task A to output the prediction result of task A, the gating unit of task B can transmit the fused features of task B to Tower B of task B to output the prediction result of task B, and the gating unit of task C can transmit the fused features of task C to Tower C of task C to output the prediction result of task C. That is, the collaborative prediction results of task A, task B, and task C are finally output.

[0129] Referring to Figure 5 As shown, the drilling multi-parameter collaborative prediction system based on multi-task learning can include the following modules:

[0130] The data management module can include data acquisition, data preprocessing, and feature optimization. Specifically, historical logging-while-drilling data and drilling real-time data can be acquired, and then missing values, outliers, data filtering, normalization / standardization processing, and finally feature optimization can be performed to select key features with high data quality.

[0131] The multi-task learning network architecture module (the multi-task prediction model described above in this invention is implemented through this network structure) may include a parameter input layer, a shared feature extraction layer (i.e., the shared layer mentioned above), an expert layer, a gating unit, and a tower layer. Specifically, the parameter input layer can receive multi-dimensional drilling parameters, the shared feature extraction layer can extract general features, the expert layer can refine the feature combination of input parameters for each prediction task based on the general features, the gating unit can adjust the information flow and dynamically adjust the fusion ratio of general features and features specific to each prediction task, and the tower layer can output the prediction results for each prediction task based on the fused features.

[0132] The model optimization module can employ loss function balancing methods (i.e., the five weight adjustment strategies mentioned above) to optimize hyperparameters, such as training epochs, learning rate, optimizer, and weight decay coefficient, ultimately resulting in a well-trained multi-task prediction model. For example, to simplify the model size, all FCNNs are set to 3 hidden layers and 16 neurons per layer, with the number of neurons in each hidden layer adjusted to 8. Other hyperparameters used in model training are optimized using Optuna, such as training epochs, learning rate, optimizer, and weight decay coefficient.

[0133] Based on the above modules, a prediction system that takes into account both parameter coupling relationships and specific feature expressions is established, which significantly reduces the mean square error of the prediction task and greatly improves the accuracy and reliability of drilling parameter prediction.

[0134] See Figure 6 As shown in (a), (b), and (c), where, Figure 6 (a) in the diagram shows the prediction effect of the multi-task prediction model on drilling speed. Figure 6 (b) in the diagram illustrates the prediction effect of the multi-task prediction model on bottom hole pressure. Figure 6 Figure (c) illustrates the prediction performance of the multi-task prediction model on formation pressure. Overall, the predicted values ​​for drilling speed (m / h), bottom hole pressure (MPa), and formation pressure (MPa) closely match the actual values, demonstrating the model's ability to accurately capture data trends. For instance, despite significant fluctuations in drilling speed at certain well sections (around 5360m), the multi-task prediction model still accurately predicts these trends, indicating that it has successfully learned the relationship between drilling speed and input features, avoiding overfitting and exhibiting good generalization ability. Similarly, the multi-task prediction model accurately predicts dramatic fluctuations in formation pressure (around 5240m and 5480m). In summary, the multi-task prediction model constructed in this invention demonstrates extremely high accuracy across all three prediction targets.

[0135] See Figure 7 (a), (b), and (c) Figure 8As shown in (a), (b), and (c), where, Figure 7 (a) in the figure is a schematic diagram of the prediction effect of the multi-task prediction model on torque. Figure 7 (b) in the diagram illustrates the prediction effect of the multi-task prediction model on the opposing pressure. Figure 7 (c) in the diagram illustrates the prediction effect of the multi-task prediction model on the pool volume. Figure 8 (a) in the figure is a schematic diagram of the prediction effect of the multi-task prediction model on natural gamma. Figure 8 (b) in the diagram illustrates the prediction effect of the multi-task prediction model on acoustic time difference. Figure 8 (c) in the diagram illustrates the prediction effect of the multi-task prediction model on spontaneous potential, including torque (kN.m), vertical pressure (MPa), and pool volume (m³). 3 The model coordinates and predicts six logging parameters (natural gamma, sonic transit time, spontaneous potential, etc.), and the predicted values ​​are in high agreement with the actual values, demonstrating the model's ability to accurately capture data trends. This showcases the versatility of the multi-task prediction model in predicting different engineering parameters, logging parameters, and well logging parameters, while also demonstrating high accuracy.

[0136] Although this specification provides the following examples or appendices Figure 9 The method or apparatus structure shown may include more or fewer combined operational steps or module units based on conventional or non-inventive methods. In steps or structures where there is no logically necessary causal relationship, the execution order of these steps or the module structure of the apparatus is not limited to the execution order or module structure shown in the embodiments or drawings of this specification. When the method or module structure is applied in actual devices, servers, or terminal products, it can be executed sequentially or in parallel according to the method or module structure shown in the embodiments or drawings (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed processing or server cluster implementation environment). Based on the above-described drilling multi-parameter collaborative prediction method based on multi-task learning, this specification also proposes an embodiment of a drilling multi-parameter collaborative prediction apparatus based on multi-task learning. Figure 9 As shown, the device may specifically include the following modules:

[0137] The construction module 901 can be used to select the prediction target and the corresponding input parameter combination from the multi-dimensional drilling parameters of the target well, and construct a task set according to the prediction target. The task set includes multiple prediction tasks for the prediction target, and the prediction tasks are related to each other.

[0138] The cooperative prediction module 902 can be configured to combine input parameters of each prediction task into the multi-task prediction model to obtain a prediction result of each prediction task. The multi-task prediction model includes a shared layer, an expert layer, a gating unit, and a tower layer. The shared layer is configured to extract common features of multiple prediction tasks. The expert layer is configured to extract unique features of each prediction task. The gating unit is configured to weight and fuse the common features and the unique features to obtain fused features of each prediction task. The tower layer is configured to output a prediction result of each prediction task according to the fused features of each prediction task.

[0139] In some embodiments, the multi-dimensional drilling parameters in the construction module 901 can include logging parameters, logging parameters, and engineering parameters. The logging parameters can include at least one of the following: drilling pressure, torque, rotation speed, and hook load. The logging parameters can include at least one of the following: gamma and resistivity. The engineering parameters can include at least one of the following: drilling speed, specific energy, bottom hole pressure, formation pressure, bottom hole temperature, and equivalent density.

[0140] In some embodiments, the multi-task prediction model in the cooperative prediction module 902 is trained in the following manner: inputting input parameter samples of each prediction task sample into an initial multi-task prediction model to output a prediction result of each prediction task sample, wherein the input parameter samples are selected from preprocessed multi-dimensional drilling parameter samples; calculating a prediction error of the prediction result of each prediction task sample and a true value, and adjusting the weight of each prediction task sample when the prediction error of at least one prediction task sample is greater than a preset error threshold; calculating a total loss value according to the adjusted weight of each prediction task sample and updating the model parameters in the initial multi-task prediction model according to the total loss value until the prediction error of each prediction task sample is less than the preset error threshold, thereby obtaining the multi-task prediction model.

[0141] In some embodiments, the shared layer in the cooperative prediction module 902 can include a plurality of sub-expert networks with attention mechanisms. Each sub-expert network can be configured to process different input parameter combinations and extract corresponding common features. The attention mechanism can be used to calculate shared weights corresponding to the common features. Accordingly, the cooperative prediction module 902 can be configured to perform weighted summation on the corresponding common features based on the shared weights to obtain common features of multiple prediction tasks.

[0142] In some embodiments, the expert layer in the cooperative prediction module 902 can include a plurality of independent expert networks for prediction targets. Each independent expert network can be configured to process different input parameter combinations and extract unique features of each prediction task.

[0143] In some embodiments, the gating unit in the collaborative prediction module 902 can include a fully connected layer, a ReLU activation layer, and a Softmax activation layer. Accordingly, the collaborative prediction module 902 can also be specifically configured to use the gating unit to concatenate the general features and the specific features of each prediction task, input the concatenated features into the fully connected layer, activate the features through the ReLU activation layer, and output the gating weights of each prediction task based on the Softmax activation layer. The general features and the specific features of each prediction task are fused according to the gating weights of each prediction task to obtain the fused features of each prediction task.

[0144] In some embodiments, the collaborative prediction module 902 can also be specifically configured to fuse the general features and the specific features of each prediction task according to the following formula:

[0145] h final =g. h shared +(1-g). h task

[0146] wherein h final is the fused features of each prediction task, g is the gating weight of each prediction task, h shared is the general features of the multiple prediction tasks, and h task is the specific features of each prediction task.

[0147] As can be seen from the above, the drilling multi-parameter collaborative prediction device based on multi-task learning provided by the embodiments of the present specification effectively solves the pain points of low prediction accuracy, poor generalization ability, and insufficient resource utilization of drilling parameters under the traditional single-task learning mode. The present invention can efficiently predict multiple drilling parameters with associated drilling parameters while reducing model redundancy parameters and training costs, capture the fluctuation law of formation parameters, improve the real-time early warning ability of abnormal working conditions, provide intelligent technical support for efficient drilling, and has broad engineering application prospects.

[0148] The embodiment of the present specification also provides an electronic device based on the above-mentioned multi-task learning-based drilling multi-parameter collaborative prediction method, comprising a processor and a memory for storing programs / instructions executable by the processor. The processor, when implemented, can execute the following steps according to the programs / instructions: selecting a prediction target and an input parameter combination corresponding to the prediction target from multi-dimensional drilling parameters of a target well, constructing a task set according to the prediction target, the task set comprising a plurality of prediction tasks for the prediction target, and each prediction task having an associated relationship; inputting the input parameter combination of each prediction task into a multi-task prediction model to obtain a prediction result of each prediction task; wherein the multi-task prediction model comprises a shared layer, an expert layer, a gating unit and a tower layer, the shared layer is used to extract common features of a plurality of prediction tasks, the expert layer is used to extract unique features of each prediction task, the gating unit is used to weight and fuse the common features and the unique features to obtain fusion features of each prediction task, and the tower layer is used to output the prediction result of each prediction task according to the fusion features of each prediction task.

[0149] In order to be able to more accurately complete the above-mentioned instructions, referring to Figure 10 The embodiment of the present specification also provides another specific electronic device, wherein the electronic device comprises a network communication port 1001, a processor 1002 and a memory 1003, and the above-mentioned structures are connected through internal cables so that each structure can perform specific data interaction.

[0150] The processor 1002 can be specifically used to select a prediction target and an input parameter combination corresponding to the prediction target from multi-dimensional drilling parameters of a target well, construct a task set according to the prediction target, the task set comprising a plurality of prediction tasks for the prediction target, and each prediction task having an associated relationship; input the input parameter combination of each prediction task into a multi-task prediction model to obtain a prediction result of each prediction task; wherein the multi-task prediction model comprises a shared layer, an expert layer, a gating unit and a tower layer, the shared layer is used to extract common features of a plurality of prediction tasks, the expert layer is used to extract unique features of each prediction task, the gating unit is used to weight and fuse the common features and the unique features to obtain fusion features of each prediction task, and the tower layer is used to output the prediction result of each prediction task according to the fusion features of each prediction task.

[0151] The memory 1003 can be specifically used to store corresponding instruction programs.

[0152] In the embodiment, the network communication port 1001 can be a virtual port bound with different communication protocols, so as to send or receive different data. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for mail data communication. In addition, the network communication port can also be an entity communication interface or a communication chip. For example, it can be a wireless mobile network communication chip such as GSM, CDMA, etc.; it can also be a Wifi chip; it can also be a Bluetooth chip.

[0153] In the embodiment, the processor 1002 can be implemented in any appropriate manner. For example, the processor can take the form of a microprocessor or processor and a computer readable medium storing computer readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, Application Specific Integrated Circuit (ASIC), programmable logic controller, and embedded microcontroller, etc. The present specification is not limited in this regard.

[0154] In the embodiment, the memory 1003 can include multiple levels, and in a digital system, as long as it can save binary data, it can be a memory; in an integrated circuit, a circuit without a physical form and with a storage function is also called a memory, such as RAM, FIFO, etc.; in a system, a storage device with a physical form is also called a memory, such as a memory stick, a TF card, etc.

[0155] The embodiment of the present specification also provides a computer storage medium based on the above-mentioned multi-task learning-based drilling multi-parameter collaborative prediction method, and the computer storage medium stores computer programs / instructions, and when the computer programs / instructions are executed, the following is implemented: selecting a prediction target and an input parameter combination corresponding to the prediction target from multi-dimensional drilling parameters of a target well, constructing a task set according to the prediction target, the task set including multiple prediction tasks for the prediction target, and each prediction task having an associated relationship; inputting the input parameter combination of each prediction task into a multi-task prediction model to obtain a prediction result of each prediction task; wherein the multi-task prediction model includes a shared layer, an expert layer, a gating unit, and a tower layer, the shared layer is used to extract common features of multiple prediction tasks, the expert layer is used to extract unique features of each prediction task, the gating unit is used to weightedly fuse the common features and the unique features to obtain fusion features of each prediction task, and the tower layer is used to output the prediction result of each prediction task according to the fusion features of each prediction task.

[0156] In the present embodiment, the storage medium includes, but is not limited to, a Random Access Memory (RAM), a Read-Only Memory (ROM), a Cache, a Hard Disk Drive (HDD), or a Memory Card. The storage medium can be used to store computer program instructions. The network communication unit can be an interface configured according to a standard specified by a communication protocol, and used for network connection communication.

[0157] In the present embodiment, the functions and effects realized by the program instructions stored in the computer storage medium can be explained in comparison with other embodiments, and will not be described here again.

[0158] Although the present specification provides method operation steps as described in the embodiments or flowcharts, more or fewer operation steps can be included based on conventional or non-inventive means. The order of steps listed in the embodiments is only one of the many execution orders of the steps, and does not represent the only execution order. In actual device or client product execution, the method order shown in the embodiments or the drawings can be executed in sequence or in parallel (for example, in a parallel processor or multi-thread processing environment, or even in a distributed data processing environment). The terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, product or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, product or equipment. Without further limitation, it does not exclude the presence of other same or equivalent elements in the process, method, product or equipment including the elements. The terms "first", "second" and the like are used to represent names, and do not represent any particular order.

[0159] Those skilled in the art also know that, in addition to implementing the controller in the form of pure computer readable program code, the controller can also be implemented in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps to realize the same functions. Therefore, such a controller can be considered as a hardware component, and the means included therein for realizing various functions can also be considered as structures within the hardware component. Alternatively, the means for realizing various functions can be considered as both software modules for implementing the method and structures within the hardware component.

[0160] The specification can be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform particular tasks or implement particular abstract data types. The specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules can be located in both local and remote computer storage media including memory storage devices.

[0161] From the above description of the embodiments, those skilled in the art can clearly understand that the specification can be implemented by means of software plus necessary universal hardware platforms. Based on such an understanding, the technical solutions of the specification can essentially be embodied in a form of software product. The computer software product can be stored in a storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to cause a computer device (which can be a personal computer, a mobile terminal, a server, or a network device, etc.) to execute the methods described in each of the embodiments or some parts of the embodiments.

[0162] The embodiments in the specification are described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment mainly describes the difference from other embodiments. The specification can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0163] Although the specification is described through the embodiments, those skilled in the art know that the specification has many modifications without departing from the spirit of the specification, and it is expected that the appended claims include these modifications without departing from the spirit of the specification.

Claims

1. A multi-task learning-based drilling multi-parameter collaborative prediction method, characterized in that, The method comprises the following steps: selecting a prediction target and an input parameter combination corresponding to the prediction target from multidimensional drilling parameters of a target well, constructing a task set according to the prediction target, the task set comprising a plurality of prediction tasks for the prediction target, and each prediction task having an association relationship; the prediction target comprising a rate of penetration, a bottom hole pressure and a formation pressure; inputting the input parameter combination of each prediction task into a multi-task prediction model to obtain a prediction result of each prediction task; wherein the multi-task prediction model comprises a shared layer, an expert layer, a gating unit and a tower layer, the shared layer is used to extract common features of a plurality of prediction tasks, the expert layer is used to extract unique features of each prediction task, the gating unit is used to weight and fuse the common features and the unique features to obtain fusion features of each prediction task, and the tower layer is used to output the prediction result of each prediction task according to the fusion features of each prediction task; wherein the shared layer comprises a plurality of sub-expert networks with attention mechanisms, each sub-expert network is used to process different input parameter combinations and extract corresponding common features, and the attention mechanism is used to calculate shared weights corresponding to the common features; accordingly, the shared layer is used to extract common features of a plurality of prediction tasks, comprising: the shared layer is used to weight and sum the corresponding common features based on the shared weights to obtain the common features of the plurality of prediction tasks; the expert layer comprises a plurality of independent expert networks for the prediction target, each independent expert network is used to receive the common features output by the shared layer and different input parameter combinations input by each prediction task, perform feature refinement based on the common features, capture nonlinear mapping relationships unique to the input parameter combinations, and extract unique features of the corresponding prediction task; the plurality of independent expert networks for the prediction target comprises a rate of penetration prediction expert network, a bottom hole pressure expert network and a formation pressure expert network, the rate of penetration prediction expert network comprises a first expert network responsible for capturing the influence of hydraulic factors on the rate of penetration, a second expert network responsible for representing the influence of formation properties on the rate of penetration, a third expert network responsible for representing the influence of bit wear on the rate of penetration, and a fourth expert network for representing the mapping relationship between the rate of penetration and the hydraulic factors, the formation properties and the bit wear, the features output by the first expert network, the second expert network and the third expert network are first integrated with the common features output by the shared layer, and then transmitted to the fourth expert network to output the unique features of the rate of penetration; the gating unit comprises a fully connected layer, a ReLU activation layer and a Softmax activation layer; accordingly, the gating unit is used to weight and fuse the common features and the unique features, comprising: the gating unit is used to splice the common features and the unique features, input the spliced features into the fully connected layer, activate the features through the ReLU activation layer, and output gating weights of each prediction task based on the Softmax activation layer; the common features and the unique features of each prediction task are weighted and fused according to the gating weights of each prediction task to obtain the fusion features of each prediction task; the gating unit is also used to adaptively adjust the shared weights when the association between the prediction tasks changes.

2. The method of claim 1, wherein, The multi-dimensional drilling parameters include logging parameters, logging parameters, and engineering parameters, the logging parameters include at least one of the following: drilling pressure, torque, rotating speed, hook load, the logging parameters include at least one of the following: gamma, resistivity, the engineering parameters include at least one of the following: drilling speed, specific energy, bottom hole pressure, formation pressure, bottom hole temperature, equivalent density.

3. The method of claim 1, wherein, The multi-task prediction model is trained in the following manner: The input parameter sample combination of each prediction task sample is input into an initial multi-task prediction model, and the prediction result of each prediction task sample is output, and the input parameter sample combination is selected from the preprocessed multi-dimensional drilling parameter sample; The prediction error of each prediction task sample is calculated, and when the prediction error of at least one prediction task sample is greater than a preset error threshold, the weight adjustment strategy is called to adjust the weight of each prediction task sample; According to the adjusted weight of each prediction task sample, the total loss value is calculated, and the model parameters in the initial multi-task prediction model are updated according to the total loss value until the prediction error of each prediction task sample is less than the preset error threshold, and the multi-task prediction model is obtained.

4. The method of claim 1, wherein, The general features and the specific features of each prediction task are weighted and fused according to the gating weight of each prediction task, including: The general features and the specific features of each prediction task are weighted and fused according to the following formula: h final = g. h shared + (1 - g). h task wherein h final is a fusion feature for the prediction task; g is a gating weight for the prediction task; h shared is a general feature for multiple prediction tasks; h task is a unique feature for the prediction task.

5. A drilling multi-parameter collaborative prediction device based on multi-task learning, characterized in that, Including: The construction module is used to select the prediction target and the input parameter combination corresponding to the prediction target from the multi-dimensional drilling parameters of the target well, construct a task set according to the prediction target, the task set includes a plurality of prediction tasks for the prediction target, and each prediction task has an associated relationship; the prediction target includes mechanical drilling speed, bottom hole pressure and formation pressure; The collaborative prediction module is used for inputting the input parameter combination of each prediction task into a multi-task prediction model to obtain the prediction result of each prediction task; wherein the multi-task prediction model includes a shared layer, an expert layer, a gating unit and a tower layer, the shared layer is used to extract general features of a plurality of prediction tasks, the expert layer is used to extract specific features of each prediction task, the gating unit is used to weightedly fuse the general features and the specific features to obtain the fusion features of each prediction task, and the tower layer is used to output the prediction result of each prediction task according to the fusion features of each prediction task; The shared layer includes a plurality of sub-expert networks with attention mechanisms, each sub-expert network is used to process different input parameter combinations and extract corresponding common features, and the attention mechanism is used to calculate shared weights corresponding to the common features; accordingly, the shared layer is used to extract general features of a plurality of prediction tasks, including: the shared layer is used to weightedly sum the corresponding common features based on the shared weights to obtain the general features of the plurality of prediction tasks; The expert layer comprises a plurality of independent expert networks for the prediction target, each independent expert network being configured to receive the shared layer output common features and different input parameter combinations of each prediction task input, perform feature refinement based on the common features, capture nonlinear mapping relationships specific to the input parameter combinations, and extract specific features for the corresponding prediction task; the plurality of independent expert networks for the prediction target comprises a drilling speed prediction expert network, a bottom hole pressure expert network, and a formation pressure expert network; the drilling speed prediction expert network comprises a first expert network responsible for capturing the influence of hydraulic factors on drilling speed, a second expert network responsible for representing the influence of formation properties on drilling speed, a third expert network responsible for representing the influence of drill bit wear on drilling speed, and a fourth expert network for representing the mapping relationships between drilling speed and hydraulic factors, formation properties, and drill bit wear; the features output by the first, second, and third expert networks are first integrated with the common features output by the shared layer and then transmitted to the fourth expert network to output drilling speed specific features. The gating unit comprises a fully connected layer, a ReLU activation layer, and a Softmax activation layer; accordingly, the gating unit is configured to weight and fuse the common features and the specific features, including: the gating unit is configured to concatenate the common features and the specific features, input the concatenated features to the fully connected layer, activate the features through the ReLU activation layer, and output gating weights for each prediction task based on the Softmax activation layer; weight and fuse the common features and the specific features of each prediction task according to the gating weights of each prediction task to obtain the fused features of each prediction task; the gating unit is also configured to adaptively adjust the shared weights when the correlation between the prediction tasks changes.

6. A computer device comprising a memory, a processor, and a computer program / instructions stored on the memory, wherein, The processor executes the computer program / instructions to implement the steps of the method of any one of claims 1 to 4.

7. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions, when executed by the processor, implement the steps of the method of any one of claims 1 to 4.

Images