WO2023009027A1 - Method and system for warning of upcoming anomalies in a drilling process - Google Patents

Abstract

A computerized method for warning of a future anomaly in a drilling process includes: generating a first training set of data for a first machine learning algorithm (MLA) on the basis of historic borehole data; training a first MLA using the first training set of data to generate a first normal response model; determining a potential drilling anomaly on the basis of the first normal response model and real-time borehole data; generating a second training set of data for a second MLA on the basis of historic borehole data and sample data generated by the first normal response model; training the second MLA using the second training set of data to generate a second normal response model; predicting the main parameters of a drilling process on the basis of the second normal response model and real-time borehole data; determining an actual drilling anomaly on the basis of the predicted main parameters of the drilling process and real-time borehole data; generating a warning on the basis of the potential drilling anomaly and/or the actual drilling anomaly.

Description

Method and system for warning about upcoming anomalies during drilling

The field of technology to which the invention belongs

The invention generally relates to means for predicting and indicating potential problems during well drilling. In particular, the invention relates to a system and method for alerting operating personnel to a future anomaly in the drilling process that could lead to failure.

Prior Art

Drilling a new well is quite a challenge. The drilling process must be strictly controlled in order to prevent incidents and accidents, which requires appropriate training of operating personnel and ensuring their readiness to prevent accidents and quickly eliminate their consequences. Drilling is a technologically complex process, and therefore any malfunction during drilling operations leads to significant losses associated with the costs of situation analysis, repair and resumption of work, including the working time of engineers and other personnel necessary to solve the problem.

Currently, some companies have the ability to predict certain types of drilling complications either automatically or with the help of human experts and their knowledge. However, due to the complexity of this task and the lack of specialists, many incidents, failures and accidents remain unforeseen until they occur. Moreover, due to the fact that the means of detecting such failures and accidents are often based on the analysis of only indirect indicators, it can sometimes even be difficult to say for sure whether a failure has already occurred or not.

AU2013403353B2 describes systems and methods for real-time risk prediction during drilling operations using real-time well data, a trained thick layer model, and a trained thin layer model for each respective layer of the trained thick layer model. These systems and methods can also be used to track other in-progress wells and perform a statistical analysis of the duration of each level of risk for the monitored well.

US6722450B2 describes an apparatus and method for monitoring and detecting a drill bit failure. Sensors said device is located on an auxiliary node (which is removed from the drill bit) and transmits data to a neural network or other adaptive filter. The neural network uses previous sensor readings to predict future sensor readings. To determine the prediction error, the predicted value is subtracted from the actual value. The increase in prediction error is an indicator based on which the operator can conclude that the drill bit has failed.

US9970266B2 describes methods and systems for improving drilling efficiency by using real-time drilling data to predict bit wear, rock lithology, pore pressure, rotational friction coefficient, rock permeability, and drilling costs. operations in real time, as well as to adjust drilling parameters in real time based on these forecasts. Operational prediction of the lithological composition of the rock is carried out by processing drilling data received in real time by a multilayer neural network. Real-time bit wear prediction is performed using real-time drilling data to predict bit efficiency and determine how it changes over time. These forecasts can be used to automatically correct drilling parameters during drilling operations with the possibility of manual intervention by the operator. Such methods and systems may also include means for determining various parameters of the hydrodynamic conditions in the well and the coefficient of rotational friction. At the same time, along with real-time data, historical data can be used here to provide expert assistance to the system and identify industrial safety problems.

US20170308802A1 describes methods and systems for predicting the outcome of various situations, such as production equipment failures, in heavy industry. The predicted outcomes can be used by the owners and operators of drilling rigs, mines, factories and other industrial facilities to identify potential failures and take preventive and/or corrective actions in relation to industrial facilities and equipment. In one common implementation, historical data associated with multiple selections is fed to one or more servers at a central site from one or more data sources. Based on this historical data, the system generates data arrays, using which a set of models is trained to predict outcomes. Each model includes sub-models corresponding to the hierarchy of components of each industrial object. The set of models is combined into an ensemble model, which is transferred to production facilities.

Patent document WO2015123591 A1 contains a description of a system for providing operational assistance to operating personnel in the process of drilling wells. Said system consists of a computer and a machine medium for long-term storage of information, on which a program is written, under the control of which the computer receives the initial data coming in real time from sensors that monitor the progress of drilling operations and/or the state of the well; clears the raw data, including deleting all raw data obtained in the following modes: auxiliary operations while drilling, running the tool, raising the tool, drilling forward, drilling back and/or cyclic drilling - in order to obtain cleaned data; feeds at least part of the cleaned data to a neural network trained on the basis of information related to the drilling process, including geological information in relation to the region in which drilling is carried out; receives from the neural network an operational forecast of the probability that certain conditions will arise in the course of drilling operations in the future; and displays this forecast.

Patent document CN109829561A contains a description of a method for predicting incidents, failures and accidents based on smoothing data processing and machine learning of a network model. This method involves building an early warning system for possible failures in the process of drilling wells and marking emergency anomalies based on smoothing data processing; creation of a training network of wells based on a machine learning model in order to form a similar model based on models with several classifications corresponding to adjacent wells by minimizing the objective function; using the measured data obtained from the training well to train the optimization model in order to obtain a parametric solution for this model; search in the network for an offset well similar to the test well, followed by evaluation of the parameters of the test well model for compliance with the offset well model, and prediction of failures and accidents of the test well based on the estimated parameters.

US20020120401A1 describes a predictive drilling management system based on the use of a neural network. The downhole processor controls the operation of various devices in downhole assembly by changing drilling parameters and drilling direction to autonomously optimize drilling efficiency. The neural network iteratively updates the drilling forecasting model and provides recommendations to the operator for adjusting the drilling parameters.

Patent document WO2019216891A1 describes a method for optimizing a drilling process in real time based on machine learning using a multilayer deep neural network built on the basis of input data from a drilling process. This network allows you to extract a large number of different drilling parameters, on the basis of which a linear regression model is built. This model is used to predict one or more drilling parameters.

In an article by Qi Peng et al. [1] provides a description of a drilling deviation warning system designed to monitor the status of the drilling process in real time and provide early warning of possible failures and accidents during drilling operations. The integrated logging tool allows the alert system to collect data on drilling parameters (weight on bit, torque, hook load) that can be displayed on a computer screen so that engineers can track trends. To determine the causes of parameter changes, mathematical statistics tools (mean time, dispersion, amplitude, etc.) are used. The most important operating principle of an early warning system is to establish the relationship between failures in the drilling process and technical parameters. The choice of anomalous drilling parameters is based on the Gray correlation theory. This system is able to accurately track abnormal drilling parameters and provide accurate (over 80% accuracy) early warning signals of the most common drilling failures and accidents (blowout, well seal failure, abnormal H2S content, drilling tool failure, etc.) either directly on a drilling rig, or at a remote monitoring center.

In an article by Giulio Gola et al. [2] considers two ways to use artificial intelligence to improve the efficiency of monitoring drilling processes in terms of reducing uncertainty and increasing the reliability of forecasting. The first method is based on the use of the combined opinion of various experts in the framework of the so-called ensemble approach; the second is based on the so-called "grey box" approach, which combines a physical model and artificial intelligence. These two methods are used to solve the problem of predicting bottom hole pressure during drilling with dynamic control. pressure to improve the accuracy and reliability of forecasting, as well as to avoid misleading information about the situation in the well or receiving inconsistent reports on operating conditions.

In the article by Jan Avreytsevich et al. [3] describes a software tool for analyzing the geological causes of lost circulation. Classification of wells on the map according to the intensity of lost circulation and the construction of new nodes for each class is carried out using an artificial neural network. The probability of assigning a new well to a particular loss intensity class is calculated based on the distance between the well and the nearest node of each class. When technological factors are taken into account, a probabilistic forecast of the occurrence of losses in the drilling fluid is formed by processing field data using the theory of fuzzy sets. The software makes it possible to predict losses caused by each group of considered factors.

In an article by Shennan Wu et al. [4] considers an approach to determining the conditional probabilities of dangerous events and their consequences. This approach involves the use of models that take into account the effects of natural degradation of equipment and (if applicable) new information about changes in model parameters (for example, changes in drilling fluid density) coming in real time. This approach is based on the theory of dynamic Bayesian network (DBN) and involves the use of additional nodes to eliminate model uncertainties and parameter uncertainties. In addition, the effect of degradation is taken into account. When a hazardous event occurs, the model can be used to predict the evolution of risk and determine its root causes in the process of offshore drilling. To link possible incident scenarios to equipment and reservoir parameters, a bow tie model is created, which is then converted to DBN. The conclusion formulated by the DBN is adapted to perform prediction and diagnostics for dynamic risk assessment, and then a sensitivity analysis is performed to determine the relative weight of each root cause.

In the article by Kishuay Inya et al. [5] proposes an intelligent method based on reservoir pressure data, spatial location of offset wells and spatial continuity of the target area, which provides the formation of a pressure profile with a high degree of certainty, the determination of a safe pressure window and the corresponding drilling fluid densities, as well as the construction of a probabilistic risk profile and drilling risk interval. Thus, given the method ensures the identification and assessment of the risks of drilling in complex formations. First, the concept of a reservoir pressure matrix is introduced and an epitaxial stratigraphic pressure migration algorithm is defined to calculate pressure in target wells of a particular tectonic area based on a depth adjustment and an inverse distance weighting algorithm. Secondly, the risk probability of drilling is calculated. Thirdly, with the help of fuzzy set theory and using data on failures and accidents, a membership function is constructed with different degrees of intervals. Finally, the degree of ownership of the risk of drilling operations is determined to provide intelligent identification of the risk of drilling.

In the article by Ekaterina Gurina et al. [6] contains a description of an intelligent model trained on the basis of historical data on failures and accidents during drilling. It is capable of detecting various types of accidents using real-time signals. Using time series comparison based on aggregated statistics and gradient boost classification, it detects process anomaly and determines its type by comparing current telemetry data (MWD) with values stored in the accident database.

In the article by Ksenia Antipova et al. [7] contains a description of a machine learning algorithm that allows you to detect anomalies using drilling process data (telemetry data). The specified algorithm automatically extracts patterns of anomaly precursor symptoms and then recognizes them during the drilling process. Its training data set comes from more than 20 oilfields and consists of sections related to more than 80 accidents of various types: stuck string, lost circulation, oil and gas show, string washout, drilling tool failure, stuffing box formation, lifting problems, and descent of the tool, problems when drilling out. The machine learning model used is based on gradient boosted decision trees. It analyzes drilling parameters in real time within a sliding 4-hour window. For each measurement, the model calculates the probability of an accident and warns of an anomaly of a certain type when the probability exceeds a set threshold. This solution provides prediction of failures and accidents during drilling, optimizes work with data and reduces downtime associated with accidents by up to 20%.

In an article by Sridharan Chandrasekaran et al. [8] describes an artificial neural network (ANN) capable of predicting ROP changes by using real-time vertical well surface displacement parameters while drilling. I/O mapping here is carried out using an interconnected neural network with training using a feedforward backpropagation algorithm, which makes it possible to effectively predict ROP on the bit. To enable the model to be trained efficiently, raw data is transformed into processed data using data screening techniques and feature engineering techniques. The completed ANN model is cross-validated to generalize across a range of inputs and compared with field measurements. The model is then used to optimize the ROP using a metaheuristic algorithm, which reduces the overall cost per meter of penetration. This is achieved using a Particle Swarm Optimization (PSO) algorithm that finds the best combination of drilling parameters (weight on bit, bit RPM, mud consumption) to provide maximum performance under actual conditions.

Unlike the above patent documents and articles, the invention described in the application provides an assessment of the overall "health" of the drilling process, not limited to predicting certain types of failures and accidents. The models provided by this invention are based on a number of parameters recorded in real time, and take into account all of them. In addition, derived elements are used in addition to the raw data, which include, among other things, human expertise. Such models do not need any data markup; instead, they use unsupervised learning techniques, which allow them to work with larger training datasets and achieve more accurate generalizations. The described invention provides not only a warning about possible accidents, but also a prediction of the main parameters of the drilling process.

Brief summary of the invention

The described invention aims to detect abnormal parameters during the drilling process in order to predict potential accidents and provide operational personnel with sufficient time to make decisions and prevent real problems, thereby providing significant time and cost savings.

Instead of predicting specific failures and accidents, it is proposed to monitor a number of key parameters of the drilling process and evaluate its overall "health". This makes it possible to predict a wide range of possible drilling problems, including problems not directly related to known types of failures and accidents.

The assessment of the "health" of the drilling process is carried out using a predictive model. One of the main difficulties in training predictive models crashes is the lack of labeled data. The volume of data associated with correctly labeled accidents (it is such data that is considered labeled) is usually small, so trained models are difficult to generalize. To overcome this problem, the described invention implements the technology of unsupervised learning, in which the model itself decides which data is suitable for training and which contains anomalies. This approach avoids the subjectivity caused by individual expert opinions and does not require any labeling of data on specific failures and accidents or other anomalies in the available (previously accumulated) data. Due to this, the model can be trained on sufficiently large arrays of training data.

An element-by-element normal behavior model (herein referred to as the "Internal Model") is used to define the training dataset, which encompasses both the original drilling data and the derived elements. The normal response model is based on "normal" pieces of data that refer to a "normal" workflow that occurs without significant disruptions, crashes, malfunctions, or any other anomalies that cause problems. To select “normal” data fragments, fragments corresponding to failures and conditions preceding failures, or containing failure precursors (for example, increased traction or increased resistance to rotation before the tool is stuck under the action of a pressure drop), fragments with abnormal or erroneous written data (for example, corrupted or late data), etc. After eliminating such "abnormal" data fragments, the remaining data is considered "normal" and is grouped into clusters of "normal" data.

Derived elements include expert knowledge related to all known failures and accidents, which eliminates the need to specifically tell the model where an anomaly (or failure) occurs, but instead informs the model which signals should be received in attention beyond the raw drilling data. The selected data is then clustered using a separate model, and some clusters are defined as reference clusters. The data from the reference clusters are considered free from anomalies and are further used to train a number of other element-wise normal behavior models (referred to here as "Regression Models").

Regression models predict a number of key drilling parameters. In particular, this application provides an illustrative example of predicting the average riser pressure (SPPA), average torque (TQA) and average hook load (HKLA). Since these models are trained only on data from the reference clusters, the predicted parameters represent a kind of "healthy" drilling process. The prediction is performed in real time and the predicted parameter values along with the corresponding actual parameter values are displayed on the system display. Operations personnel can tell at any time that something is going wrong while drilling, just by looking at the display screen. To ensure timely risk mitigation, the system also automatically predicts and warns of possible accidents.

The system indicates for which parameter there is the largest difference between the predicted value and the actual value, and triggers an accident probability indicator to indicate a potential problem (for example, using signal colors - yellow or red). Anomalies are determined by the Regression Models and the Internal Model independently. The results of anomaly detection are included in the general signaling of the problem.

The system proposed by the described invention is based on a computerized method of warning about a future anomaly during drilling. This method includes:

- formation of the first training data set for the first machine learning algorithm (MLA) based on historical well data;

- training the first MLA algorithm using the first training data set to form the first normal response model;

- determination of a potential drilling anomaly based on the first normal response model and real-time well data;

- generating a second training data set for the second MLA algorithm based on historical well data and sample data generated by the first normal response model;

- training the second MLA algorithm using the second training data set to form the second normal response model;

- prediction of the main parameters of the drilling process based on the second normal response model and real-time data of wells;

- determination of the actual drilling anomaly based on the predicted main parameters of the drilling process and real-time data of wells;

- generating a warning based on a potential drilling anomaly and/or an actual drilling anomaly. A drilling anomaly warning system includes at least one processor, RAM, and computer readable instructions for executing the drilling anomaly warning method described above.

Systems for warning about upcoming anomalies during drilling may also additionally include devices for drilling wells, at least a drill bit.

The described invention provides for the possibility of generating multiple training datasets for the second MLA, after which one or more of these arrays can be selected based on operations and steps corresponding to real-time drilling data. Accordingly, real-time drilling data can be pre-processed to determine operations and steps.

Prediction of key parameters of the drilling process can be performed in real time. Within one embodiment of the described invention, average hook load (HKLA), average torque (TQA) and average riser pressure (SPPA) can be key parameters of the drilling process.

To generate an alert, potential drilling anomaly data and actual drilling anomaly data can be combined. The merging of the potential drilling anomaly data and the actual drilling anomaly data may be performed using a separate merging algorithm. For example, the join algorithm may be implemented by a separate MLA. In some embodiments of the described invention, the combining algorithm may be implemented as a neural network.

The first and/or second normal behavior model may be updated based on the combined drilling anomaly data. Updating models allows you to improve the quality of forecasts.

Warnings generated by the system provided by the described invention may be visual and/or audible. Visual warnings may be based on the use of signal colors (eg, similar to traffic lights). Visual warnings can also be based on the use of text messages describing a predicted or detected anomaly. Visual warnings may also be based on the display of images (eg, icons) indicating a predicted or detected anomaly. Sound warnings can be based on the use of special audible signals such as an alarm. Audible warnings can also be based on the use of voice messages indicating a predicted or detected anomaly.

Brief description of the drawings

On fig. 1 shows a general schematic diagram of a warning method implemented in accordance with the described invention.

On fig. 2 shows a schematic diagram of the model generation stage.

On fig. 3 shows a schematic diagram of the model training phase.

On fig. 4 shows a schematic diagram of the stage of using models to process real-time data.

On fig. 5 shows a schematic diagram of a warning system implemented in accordance with the described invention.

On fig. 6 shows an example of identifying potential anomalies.

On fig. 7 shows an example of a user interface in a system implemented in accordance with the described invention.

On fig. 8 shows an enlarged fragment of the interface shown in fig. 7, containing indications of the values of the key parameters of the "health" of the well (comparison of actual values with predicted ones) and indications of possible risks.

Detailed description of embodiments of the described invention

Implementation details of the warning method (10) and the warning system (200) in accordance with the described invention are given below with reference to Fig. 1-8.

On fig. 1 shows a general schematic diagram of a warning method (10) in accordance with the described invention. As a rule, the warning method (10) includes the following steps:

- stage (11) generation of models, which generates the Internal model and several Regression models;

- stage (12) training models, which trains the generated models; And

- step (13) using trained models to process real-time data, at which actual and potential anomalies are determined;

- step (14) generating warnings, which generates and outputs warnings, signals and/or indications; and optional - step (15) updating the used models, in which the used models are updated based on the feedback.

The schematic diagram of stage (11) of creating models is shown in fig. 2, and the schematic diagram of stage (12) of model training is shown in fig. 3. Internal Model and Regression Models are generated and trained based on historical well data and expert knowledge.

Historical well data and expertise

As shown in fig. 2, within the framework of the modeling stage, an initial array of historical well data (111) is formed. To form the initial array, it is necessary to collect well data from specific fields of interest. A field here refers to a certain place where hydrocarbons accumulate, where one or more wells are located. For example, deposits may be owned by a particular company or developed by a particular company. Since such well data is intended to train computer production models, the data to be included in the historical well data set is selected from the raw input data in such a way as to ensure maximum , using telemetry (MWD) and/or inclinometry, or obtained from any other appropriate source. Parameter values can be recorded (for example, by direct measurements) in real time or calculated later on the basis of the measurement results. The list of typical parameters obtained as a result of the GTI is presented in Table 1.

Table 1

Various parameters may be used to detect and/or predict various failures or other failures in wells. For example, the problem "sticking tool due to differential pressure" can be associated with parameters HKLA, BPO, DBTM, DMEA, TQA, WOB, RPMA, SPPA and MFIA, while the problem "gas-oil-water show" can be associated with parameters HKLA, TQA , SPPA and ROPA, and the "loss of circulation" problem can be associated with the parameters TQA, ROPA, SPM1, SPM2, TV01-TV04 and HERE.

It should be noted that the above list of parameters is not exhaustive or comprehensive. For example, some parameters such as MDOA (Drilling Fluid Outlet Density (avg.)), MTOA (Drilling Fluid Outlet Temperature (avg.)) or MFOA (Mud. in MWD data; however, they can be included in the list of parameters that should be considered in relation to some types of failures that should be predicted by the models. The more representative the dataset is with respect to key parameters, the more predictive a model trained on that dataset can be. In this regard, the list of parameters that should be taken into account when generating an array of historical well data can be adjusted depending on the most important failures and accidents that need to be prevented, which, in turn, often depend on the nature, design and age of the well. , as well as on the geological conditions of the deposit. The list of parameters can also be adjusted depending on the purpose of the model that will be trained using the generated data array.

In addition, the array of historical well data is pre-processed in order to exclude certain types of failures, taking into account the availability of data on these failures, the frequency of their occurrence, the severity of the consequences, etc. An exception can be implemented using thresholds, which can be static or dynamic. The static threshold may be determined, for example, by data availability. A dynamic threshold value can be implemented, for example, depending on the frequency of occurrence of a failure, the severity of its consequences, etc.

The number of failures or accidents covered by the array of historical well data can range from several tens to several thousand. However, in some embodiments of the described invention, the historical well data set may not include any failures (in particular, this applies to data sets of the normal behavior model). In addition, some failures that are not present in the original historical well dataset may be deliberately added to the dataset (in particular, this applies to test datasets). Exclusion and/or inclusion of data on failures and accidents can be done using human expertise (112).

Expert knowledge includes the knowledge of scientists and real experience in the construction and operation of wells. For example, expertise is used to select candidate wells, determine the importance of specific parameters to be included in the historical well data set, divide the workflow into operations and/or sections and/or steps, and determine which operations and/or sections and /or stages must be included in or excluded from the historical well data set for each specific project. Expert knowledge is also used to identify and analyze anomalous patterns in historical well data. In addition, expert knowledge can be used to expand the set of elements used to characterize models, taking into account expert experience, drilling technologies and best practices. During model testing, expert knowledge can also be used to analyze prediction results. Pre-processing of historical well data

Before using historical well data (111) to create models, this data can be pre-processed. In particular, pre-processing may include (but not be limited to) determining the operations and steps to which the data relates, filtering and/or averaging data values, collecting statistical information, etc.

Training data set for the Internal model

In step (115), a training data set for the Internal Model is created. The pre-processed historical well data is divided into fragments corresponding to specific fields, operations, sections and stages to ensure that the most relevant and representative data is included in the training sets. A field here refers to a certain place where hydrocarbons accumulate, where one or more wells are located. Typically, all or most of the wells in a particular field have the same or similar properties. A section here means a separate part of the downhole equipment, for example, a conductor, a drill string, a production string, a liner, etc. A stage here refers to a part of an operation, such as drilling, raising a tool, running a tool, etc.

In addition, the processed historical well data is divided into segments of a certain duration in time. Since there are no truly “perfect” wells for extracting reference training data, the data packets are divided into segments of several minutes each. The length of each segment may depend on the sampling rate, the number of parameters to track, the target of the model being trained, and so on. After analysis for the relevance and integrity of the data and appropriate screening, the segments of all wells are clustered. During clustering, data dimensionality reduction can be performed in order to reduce the computational load and reduce the model execution time. Any appropriate approach may be used to reduce the dimensionality. The choice of a dimensionality reduction method is a trivial task for a person skilled in the art, so its description is omitted here for brevity.

Some indicators can be used to evaluate the results of clustering. Metrics can include parameters such as cluster size, cluster health, data distribution, time coverage, and so on. Metrics may include some weights and/or features. Based on the selected indicators, the most representative pool of clusters is accepted, and the data of all The segments in these clusters are considered reference data, while the rest of the data are considered anomaly-related data.

The clustered reference data is then used to train the Internal Model.

Internal model training

The learning process (122) for the Internal Model is based on a machine learning algorithm (MLA) implementing an unsupervised learning approach. This approach is used when there are many objects (situations) and many reactions (behaviors). The set of object-response pairs is called the training set. Training is aimed at forming a kind of "perception" by the neural network, which could give a fairly good response to any object. For example, sensor readings for WPO, TQA, SPPA (and/or their derivatives such as arithmetic mean, median, variance, quantiles, etc.) are used as objects to determine HKLA values considered as responses. In particular, the gradient boosting algorithm can be used as the MLA.

The main idea of the gradient boosting method is to sequentially add some new models to the ensemble of models, where at each iteration a new weak base predictive model is trained against the error that the entire ensemble has already learned. In other words, the goal is to obtain new learning models that correlate as much as possible with the negative gradient of the ensemble-related loss function. As a general rule, any gradient boosting method tends to approximate as a weighted sum of weak predictive models. Some applications of this technology may use various kinds of gradient boosting algorithms, such as decision tree-based gradient boosting, stochastic gradient boosting, probability-based boosting, AdaBoost or Gentle Boost, and others.

Training Datasets for Regression Models

Once the Internal Model has been trained, it can also be used to generate (114) datasets for training Regression Models. In particular, the trained Internal Model allows you to create reference data grouped into reference clusters. The data that form the reference clusters are considered free from any anomalies and are further used to train the Regression Models.

Also, similar to the clustered data used to train the Inner Model, some special features can be computed additionally. These elements are typically field, well, operation, and stage specific and can usually be derived directly or indirectly from the clustered data. For example, these elements may include normalized data for a period of time or condition, statistics, velocity, flow, weight or displacement values, triggers, thresholds, etc. that are calculated from the clustered data. It should be noted that different Regression Models may require different elements. In the future, clustered data aggregated with objects is used to train Regression models.

It should also be noted that in addition to the training datasets, other datasets for testing, validation, etc. can also be additionally generated. Unlike training datasets, test datasets for the same models may include data related to wells with some deviations, for example, high non-productive time (or downtime), failures, accidents, etc.

Training Regression Models

Training Regression Models is designed to generate a stream of predicted values of certain parameters, such as HKLA, SPPA and TQA, based on real-time well data, the predicted values of which will be used as reference values indicating the "normal" workflow. The occurrence of anomalies can be determined by comparing predicted values and actual values obtained from downhole sensors and/or calculated from actual values.

The learning process (121) for Regression Models can be essentially the same as for an Internal Model. However, different training data sets are used here, and the goal of training is to identify problems based on real-time well data. For different sections and stages, own (separate) Regression models are trained. This allows in each case to select the appropriate regression models for training (123) depending on the training data sets used at this stage.

It should be noted that both the Internal model and any of the Regression models are able to detect anomalies, but they do it in different ways, which will be discussed below.

Stage (13) of using models is schematically shown in fig. 4.

Selecting Regression Models Since individual Regression Models are trained for different operations, sections, and stages, the appropriate Regression Models (131) are selected from among the trained Regression Models. As mentioned above with respect to the selection of Regression Models during the training phase, the appropriate Regression Models (131) can be selected during the use phase depending on the composition of the pre-processed well data coming in real time. In particular, when real-time pre-processed well data is related to drilling operations, Regression Models for the relevant drilling parameters (eg, SPPA, TQA, and HKLA) can be selected.

Downhole data in real time

The real-time downhole data (135) is a stream of essentially the same parameters as the historical well data discussed above. It should be noted that real-time stored well data only becomes historical well data for future periods. However, real-time well data may have a higher sampling rate, such as being updated every second, while historical well data may be averaged and stored once every ten seconds. Real-time well data is assumed to arrive without significant delays and/or significant gaps.

Pre-processing of well data in real time

Similar to historical well data pre-processing, real-time well data is also pre-processed (136) to determine the operations, sections, and steps to which they relate. In addition, for some parameters in the real-time well data, statistical information is usually calculated, such as, for example, the minimum value, maximum value, average value and standard deviation. The parameters to be statistically analyzed may be target parameters to be tracked and predicted using the method provided by the described invention. Depending on the conditions of the project, other information can be obtained from the real-time well data using pre-processing (for example, the definition of some characteristics, as described below). Statistical information can be generated in relation to given time intervals (for example, each segment lasting 10 minutes). However, different statistics for the same parameter may be generated for different time intervals and/or the same statistics may be generated for different parameters.

Typically, the pre-processing of real-time well data destined for the Internal Model may differ from the pre-processing of the same real-time well data destined for Regression Models. Internal Model and Regression Models are Normal Behavior Models (NRMs). This means that they are designed to represent a normal workflow without significant failures, accidents, malfunctions or any other anomalies that cause problems in the drilling process. Thus, both the Internal Model and each of the Regression Models are defined by the data set used to train the respective model.

Prediction of key parameters of the drilling process

The key parameters of the drilling process are the target parameters that must be predicted by the Regression Models. Target parameters are predicted (132) each time new real-time well data is entered into the respective trained Regression Model. For example, new reference values for the HKLA, SPPA and TQA parameters can be generated. Reference values are the values that would be observed in an "ideal" well when other monitored parameter values are equal to the entered values.

Definition of actual anomalies

If the difference between the generated reference values and the actual real-time well data exceeds a threshold, the Regression Model determines the presence of an anomaly (133). Determination of actual anomalies may include zooming in a sliding window. The size of the sliding window is determined automatically, depending on the achievement of the maximum values of the metrics. For example, the sliding window size can range from 10 minutes to 6 hours. Zooming in the sliding window minimizes false positives. Identifying Potential Anomalies

The trained Internal Model can be used to identify potential anomalies. Real-time well data is entered into the Internal Model. New data is combined with previously known data and clustered. The newly defined clusters are analyzed and the clusters associated with anomalies are separated from the reference clusters. The separated cluster data are compared with real-time well data and potential anomalies are identified based on the results of the comparison (137).

Typically, both the Regression Models and the Internal Model work for each target parameter determined from the results of the GTI as follows:

- mud logging data is extracted for the entire history of well drilling;

- fragments of data that are not subject to consideration by the model are filtered out;

- a sliding time window is determined;

- the corresponding parameter values are normalized based on the average (or median) value of the parameter in the sliding window;

- additional (derivative) elements are calculated (for example, the rate of change of the parameter value);

- derived element values determined from the data contained in the last 10 minutes (offset value) of the target parameter change curve are discarded to avoid overtraining the model;

- the remaining values of the parameters and their derived elements for the last 10 minutes are used for forecasting.

The predicted values of the target parameters (eg HKLA, SPPA and TQA) are compared with the corresponding parameters determined in real time. The comparison is based on a metric for each time reading. The metric is accumulated in a sliding window. The comparison result is additionally compared with a threshold value. The sliding window size and threshold are determined after the model has been trained.

Combining Anomaly Data

The anomaly predictions made by the Internal Model and the Regression Models are combined (134) in such a way as to generate appropriate warnings, signals and indications. At the same time, it should be remembered that the forecasts of the Internal Model and Regression Models are of a different nature, therefore, to aggregate these forecasts and decision making regarding warnings, a separate model can be used. In some embodiments of the described invention, the data merging step may be performed by a separate ad hoc model.

As a rule, the Internal Model shows the presence of an abnormal piece of data, which indicates a possible abnormal course of the workflow. Regression models focus on specific parameters (eg, HKLA, SPPA, and TQA), where the difference between the predicted parameter value and the corresponding real-time parameter value also indicates a possible abnormal workflow. When the readings of the Internal Model and one or more Regression Models match, one or more warnings may be generated and displayed. The type of warnings may depend on the severity of the abnormal situation, which, in turn, may depend on the degree of agreement between the indications of different models.

However, the data aggregation step is not absolutely necessary anyway, so it can be omitted or replaced by other solutions. In particular, in some embodiments of the described invention, alerts may be generated separately based on Inner Model predictions and Regression Model predictions.

Alert generation

Warnings, signals and instructions are generated (14) to draw the attention of the operating personnel to possible problems detected by the warning system. Alerts may take any suitable form, including, but not limited to, various audible signals, voice messages through a loudspeaker, text messages sent to a pager, voice and/or text messages sent over a cellular communication system (for example, sending messages to mobile phones). ), light signaling (for example, similar to traffic lights), flashing beacon signals, etc. Warnings or other indications may also take any appropriate form, including, but not limited to, interface messages, icons, and signals displayed on the user interface of the warning system in accordance with the described invention.

Updating Internal and Regression Models (Optional)

As a rule, updating (15) of the Internal Model and Regression Models is carried out with the participation of three specialists: a drilling expert, a systems analyst and a deployment engineer. Drilling expert initiates request to update models in case of errors (incorrect results) in the work of models. The drilling expert has the right to approve or reject the updated model. The system analyst is responsible for handling requests from the drilling expert and for developing new models. The systems analyst can also initiate a model update if the model is found to be performing less than expected. The deployment engineer is responsible for deploying the updated model to a test or production environment.

Thus, model updates can be initiated by either the customer, the supplier, or both. In any case, the system analyst receives artifacts of incorrect model operation, for example, version number, prediction results, actual parameter values, notes and comments from staff, etc. The system analyst summarizes the information received and analyzes the incident. Some additional information may also be used, such as system logs. Based on the information collected, the system analyst reproduces the behavior of the system and finds the cause of the failure in its operation. If the failure is caused by other systems, such as data corruption due to data transmission errors from the well, the systems analyst escalates the problem to the parties responsible for such systems.

In the event that the failure is caused by internal conditions, the system analyst performs an analysis of the root causes of the failure in order to find ways to eliminate them. In most cases, failures are caused by a situation that is not covered by the current release of the warning system. To resolve this failure, the systems analyst adds observational data to the training dataset and starts retraining the corresponding model. After retraining, the system analyst obtains quality scores for the retrained model, validates the performance of the model, and performs regression testing.

After the problem is found, isolated and fixed, the system analyst prepares a report for the drilling expert, including information about the changes made, updated model principles, illustrations, updated model quality scores, and other relevant information.

The drilling expert approves or rejects the updated model. If the drilling expert rejects the updated model, he must provide good reasons to justify such a decision. In this case, the system analyst starts a new update procedure. If the updated model is approved, the system analyst passes it to the deployment engineer along with additional information (for example, a description of new parameters and features of the model, etc.) that may be useful when deploying the model.

The deployment engineer performs additional functional tests and, if they pass, updates the model in a test or production environment. If necessary, the deployment engineer also updates the values of statistical parameters, such as the thresholds used in the model.

On fig. 5 shows a generalized diagram of a warning system (200) in accordance with the described invention.

The warning system (200) typically includes a historical well data subsystem (201), an expert knowledge subsystem (202), a modeling subsystem (203), a processing subsystem (204), and a user interface (205). Downhole data in real time are sent to the processing subsystem via a communication line (206).

The historical well data subsystem implements the collection, processing, storage, maintenance and updating of historical well data used as input to the warning system in accordance with the described invention. The implementation of the historical well data subsystem does not have any particular limitations. In one embodiment of the described invention, the historical well data subsystem may comprise one or more servers and/or one or more databases. In another embodiment of the described invention, the historical well data subsystem may be implemented using a different architecture, such as a peer-to-peer network.

The expert knowledge subsystem collects, processes, stores, maintains and updates information, which is expert knowledge, usually related to drilling, completion, commissioning and operation of wells, etc. The implementation of the hardware/software components of the historical well data subsystem does not have any particular limitations. It should be noted that the human component of the historical well data subsystem is of particular importance to ensure the relevance of models built on the basis of expert knowledge; however, this aspect is outside the focus of this patent application.

The modeling subsystem implements the generation of training datasets for the Internal Model and Regression Models based on historical well data and expert knowledge, as well as training of the Internal Model and Regression Models using training datasets. The implementation of the modeling subsystem does not have any particular limitations. In one embodiment described of the invention, the modeling subsystem may comprise one or more servers and/or one or more databases. In another embodiment of the described invention, the modeling subsystem can be implemented using a different architecture, for example, in the form of a peer-to-peer network.

The processing subsystem performs pre-processing of real-time downhole data to determine the operations, sections and stages to which the data belongs; identification of potential anomalies based on the results of the Internal Model; selection of Regression Models based on the sections and steps to be modeled; prediction of key parameters of the drilling process using the results of Regression Models; determination of actual anomalies based on predicted key parameters of the drilling process; and combining the anomaly data resulting from the determination of potential and actual anomalies as described above. The implementation of the processing subsystem does not have any particular restrictions. In one embodiment of the described invention, the processing subsystem may contain one or more servers and/or one or more databases. In another embodiment of the described invention, the processing subsystem can be implemented using a different architecture, for example, in the form of a peer-to-peer network.

It should be noted that the above subsystems are described as separate entities only to facilitate understanding of the concept of the invention. One skilled in the art would understand that these functions may be combined or distributed among one or more entities without any particular limitation.

Real-time well data is delivered to the processing subsystem via a communication line. The implementation of the communication line does not have any particular limitations. In various embodiments of the described invention, the communication line may have one of the following implementation methods: Ethernet, Fieldbus, wireless data network, 1oT data network. The only limitation associated with the implementation of the communication line is the ability to transmit real-time data with the required degree of reliability.

The user interface implements the issuance of warnings, signals and instructions that are communicated to operational personnel and possibly other interested parties. Depending on the nature of the warnings, signals, and instructions, the user interface may include one or more displays, loudspeakers, sources of audible alarms, buzzers, color alarms, signal indicators, flashing beacons, etc. It may also include any means of delivering text messages using a paging system, means of delivering text and/or voice messages using a local wireless communication system, a cellular communication system, a satellite communication system, and so on. Some alerts and guidance may be issued and routed on a corporate local area network (LAN).

Illustrative variant

Below is an example of the implementation of the method of warning about anomalies during the drilling process, schematically presented in fig. 1-4. It should be understood that this example is intended only to facilitate understanding of the essence of the described invention and in no way limits its content.

In the example below, historical well data (111) is taken from 60 wells that did not experience any significant disruptions or accidents while drilling and had minimal non-productive time. The selection was conducted by a team of experts with deep scientific knowledge and extensive practical experience in the field of well construction technologies (112). The resulting array of historical well data was used to train the Internal Model and Regression Models designed to predict the current values of HKLA, SPPA and TQA coming in real time in order to detect abnormal situations during drilling.

Historical data for each well was divided into 10 minute chunks, analyzed for relevance, and then the respective chunks were clustered. With a 10-second sampling of historical data, each 10-minute fragment contains their 10x6 parameter values, where n is the number of elements to be used in the model. Depending on the approach used in each specific project, n can be in the range from 5 to 21, so the clustering problem must be solved with a dimension in the range from 300 to 1200, which will require excessive computing resources. Therefore, to reduce the dimension of the clustering problem, the UMAP algorithm (https://umap-leam.readthedocs.io/en/latest/) was used. The final dimension is determined by estimation or by testing during the execution of the UMAP algorithm.

Then, the resulting reduced-dimensional training dataset was clustered using the DBSCAN method (https://en.wikipedia.org/wiki/DBSCAN). The set of clusters was evaluated based on the size of the cluster (i.e. the number of fragments lasting 10 minutes in each cluster) and the health of the cluster (percentage of fragments containing failures in each cluster). The clustering task was performed with various combinations of task parameters until the best result was obtained, in which "normal" fragments without significant failures, artifacts or other violations were grouped into large clusters. All fragments within these clusters were subsequently considered reference fragments, while the remaining fragments were considered anomalous fragments. Thus, the training dataset for the Internal Model (115) was generated.

Similar to preparing the training data for the Internal Model, the historical data for each well was divided into 10-minute chunks to create a training dataset for each Regression Model. Historical data was then analyzed for reliability, and unreliable data (eg, negative HKLA, SPPA, and TQA values) were filtered out. For each point in time, the corresponding operation and stage were defined. In sliding time windows of various widths, the following characteristics were calculated, which were determined automatically depending on the achievement of the maximum values of the above indicators:

- rate of change of the distance between the bit and the bottom of the well;

- the rate of change in the depth of the bottomhole of the well without taking into account the fragments corresponding to the build-up of the tool;

- normalized values for sliding windows;

- standard deviation for HKLA values in a small sliding window;

- the rate of change of the position of the block (VRO);

- offset value for all derived elements associated with the target element; an offset value used to prevent overtraining of the model;

is the average frequency of the Fourier transform.

The Internal Model was then used to cluster the training data to form a training dataset for each Regression Model. The above data was aggregated to generate a training dataset for each Regression Model (114).

In the exemplary embodiment shown, the Inner Model has been trained (122) for clustering only to determine reference tiles. Regression models were trained (121) to predict anomalies in the target parameters (HKLA, SPPA and TQA) of well data coming in real-time, which are introduced into the Regression models after they have been trained. For each section and stage, separate Regression models are trained.

Downhole data (135) containing HKLA, SPPA, TQA, WOB, TUT, MFIA and ROPA parameters are received once per second in real time. Statistical parameters for HKLA, SPPA and TQA, namely minimum value, maximum value, mean value and standard deviation, are calculated for each 10-minute fragment (136).

On fig. 6 shows an illustrative example of identifying potential anomalies (137). Curve N° 1 represents the actual (i.e. measured) HKLA value; curve N° 2 shows the predicted value of HKLA; curve N° 3 shows the threshold value of HKLA; and curve o 4 shows the difference between the actual and predicted HKLA values. As can be seen from the above diagram, a possible problem "sticking tool due to differential pressure" was identified at time ti, and the actual problem occurred at time t2 after about 1.5 hours, which gives the operating personnel enough time to take preventive action.

On fig. 7 shows an example of the user interface (in Russian) of the warning system provided by the described invention. The interface contains a chart showing the progress of drilling a well over time, as well as the values of the target parameters of the "health" of the well and potential risks.

On fig. 8 shows an enlarged right side of the interface shown in fig. 7 (in English), containing indications of target parameters (actual and predicted values of HKLA, SPPA and TQA and corresponding pictograms if the value increases or falls) and indications of potential risks. In particular, red color indicates possible problems. When a risk is detected, some recent relevant events are indicated.

It should be understood that this warning method and this warning system are described herein merely as an exemplary implementation of the technology of the invention. Thus, their description is only intended to provide illustrative examples. This description is not intended to define the scope or establish any boundaries of the technology of the invention. In some cases, where it is considered appropriate, examples of modifications to the warning method and warning system provided by the described invention are given. This is done solely to facilitate understanding of the essence of the invention, and not to determine the scope applications or set the boundaries of the technology of the invention. These modifications are not intended to be exhaustive, and therefore any person skilled in the art should understand that other and/or further modifications are possible. Where examples of modifications are not given, this fact should not be interpreted to mean that no modifications are possible and/or that what has been described is the only way to implement the relevant part, component, or element of the technology of the invention. In addition, it should be understood that the warning method and warning system provided by the described invention may, in certain cases, be simplified implementations of the technology of the described invention, and that their description is thus provided as an aid to a better understanding of the technology of the invention. In this regard, any person skilled in the art should understand that other implementations of the technology of the invention may be more complex than described in this application.

The examples and conventions provided herein are primarily intended to help the reader understand the underlying technology of the invention described and not to limit its scope to such specific examples and terms. It should be understood that a person skilled in the art can develop various mechanisms that, although not explicitly described or shown in this application, nevertheless embody the essence and essence of the technology of the described invention and are consistent with its spirit, therefore, such mechanisms are also included in the scope legal protection of the described invention.

Moreover, all statements that set forth the principles, aspects, and implementations of the technology of the described invention, as well as specific examples thereof, are intended to cover both structural and functional equivalents, whether they are currently known or may be developed in the future. . Thus, for example, it should be understood that any functional diagrams contained in this application are a conceptual vision of illustrating circuits and/or algorithms embodying the principles of the technology of the described invention. Likewise, it should be understood that any block diagrams, state transition diagrams, pseudocode, and the like. are representations of various processes that may be represented on computer-readable media and, accordingly, executed by a computer, processor, or controller, whether such computer, processor, or controller is explicitly shown or referred to or not. Any functions of the various elements shown in the drawings contained in this application, including any functional unit designated or referred to as "processor" or "data processing unit", can be implemented using special equipment, as well as equipment capable of executing program code in combination with the corresponding software components. The functions of a processor may be performed by a single dedicated processor, a single shared processor, or multiple individual processors, some of which may be shared. In some non-limiting embodiments of the technology of the invention described, the functions of a processor may be performed by a general purpose processor such as a central processing unit (CPU) or a processor dedicated to a specific purpose such as a graphics processing unit (GPU). In addition, explicit use of the term "processor" or "controller" should not be construed solely as referring to hardware capable of executing program code, and such terms may implicitly include, but are not limited to, digital signal processor (DSP), network processor, application specific integrated circuit (ASIC), field-programmable logic integrated circuit (FPGA), read-only memory (ROM) for storing program code, random access memory (RAM) for storing data and/or program year, and any non-volatile data storage. Also, these terms may include any other serial equipment and / or equipment made to special order.

In the context of the present application, a "server" is a hardware/software system capable of receiving data and/or signals directly and/or over a network, processing the data, and producing some resulting data and/or signals. The hardware in such a complex may be one physical computer or one physical computer system, or several physical computers or several physical computer systems. In the context of this application, any use of the term “server” does not imply that all tasks (e.g., all received data, signals, instructions, or requests) or any one particular task must be received, executed, or processed by the same server (i.e., i.e. the same software and/or hardware); this means that any number of software elements or hardware devices may be involved in receiving/sending, executing or processing any task or request or the results of any task or request, or processing any data or signals; and the entirety of such software and hardware may be one server or multiple servers, both of which mean "at least one server".

In the context of this application, the term "component" means software (corresponding to specific equipment) that is necessary and sufficient to perform the specific function(s) in question.

In the context of this application, the term "computer storage media" is used to refer to media of any nature and type, including RAM, ROM, disks (CD-ROMs, DVDs, floppy disks, hard drives, etc.), USB drives, solid state drives, tape drives, flash memory cards, etc.

In the context of the present application, the ordinal numbers "first", "second", "third", etc. can be used solely to distinguish between the nouns they refer to, and not to describe any particular relationship between those nouns. Thus, for example, it should be understood that the use of the terms "first element" and "third element" does not imply any particular order, type, chronology, hierarchy, or ranking of (for example) elements, and their use (by itself) does not imply that any "second element" must necessarily exist in any particular situation. Also, referring to a "first" element and a "second" element does not preclude that the two elements are the same actual real world element. Thus, for example, in some cases the "first" computing device and the "second" computing device may be the same software and/or hardware, and in other cases they may be different software and/or hardware.

In the context of this application, the term "database" is any structured data set, regardless of its specific structure, data management software, or computer hardware on which the data is stored, implemented, or otherwise made available for use. The database may reside on the same hardware as the process that stores or uses the data stored in the database, or it may reside on separate hardware such as a dedicated server or multiple servers.

In the context of this application, the terms "information" and "data" include information of any nature and of any kind that may be stored in a database. Thus, information includes, but is not limited to, data (spatial data, time data, telemetry data, knowledge base data, etc.), textual information (conclusions, comments, notes, questions, messages, etc.), images, sound recordings, documents, spreadsheets, etc. .

It should be noted that the above description covers those actions that are most important for achieving the goal of the described invention. It should be obvious to a person skilled in the art that other actions must also be performed in order to ensure the operation of the system provided by the described invention, for example, connecting equipment, initializing it, starting the appropriate software, sending and receiving instructions and confirmations, exchanging auxiliary data, synchronizing etc., and their description is omitted here for brevity.

It should also be noted that the above description covers those elements and components that are most important for achieving the purpose of the invention. One skilled in the art will appreciate that these elements and components must or may include other parts necessary to enable the warning system and warning method to function, and their description is omitted here for brevity.

The devices and their components or parts, as well as the methods and steps thereof, mentioned in the above description and shown in the diagrams related to it, refer to one or more specific embodiments of the described invention when they are given with reference to a numerical designation, or refer to all applicable embodiments of the invention when given without reference to a numerical designation.

Devices and their parts mentioned in the above description, diagrams and claims may be combined hardware/software, in which the hardware of some devices may differ or partially or completely coincide with the hardware of other devices, unless expressly stated otherwise. . The hardware of some devices may be located in different parts of other devices, unless expressly stated otherwise. The software may be implemented as computer code stored in a storage device or storage medium.

The sequence of steps in the description of the warning method given in this application is only an illustration and may differ in some embodiments of the described invention, if the corresponding function is supported and the desired result is achieved. Some stages of this method may be combined together, while some of the other steps may be divided into sub-steps or separate steps. Thus, the order of the steps is not limiting for the described technology; otherwise, this order is not explicitly specified or not originally provided.

Components or parts and elements of the described invention can be combined in various embodiments of the invention, if they do not contradict each other. The embodiments of the invention discussed above are presented by way of illustration only, and they are not intended to limit the essence and content of the invention described in the claims. All and any reasonable modifications, changes and equivalent substitutions in design, configuration and modes of operation, corresponding to the essence of the invention, are included in the scope of the described invention, indicated in the attached claims.

Each implementation of the technology of the described invention has at least one of the above objectives and/or aspects, but not necessarily all of them. It should be understood that some embodiments of the technology of the described invention that result from an attempt to achieve any of the aforementioned objectives may not serve that purpose and/or may serve other purposes not specifically listed in this patent application.

Links

1. Qi Peng, Honghai Fan, Shuguo Xu, Haobo Zhou, Minbin Lai, Guangxi Ma, Suiyi Fu. A Real-Time Warning System for Identifying Drilling Accidents. SPE Annual Caspian Technical Conference and Exhibition, 12-14 November, 2014, Astana, Kazakhstan, DOI: 10.2118/172303-MS

2. Giulio Gola, Roar Nybo, Dan Sui, Davide Roverso. Improving Management and Control of Drilling Operations with Artificial Intelligence. SPE Intelligent Energy International, 27-29 March, 2012, Utrecht, The Netherlands, DOI: 10.2118/150201-MS

3. J. Awrejcewicz, Yu.B. Lind, AR Kabirova, LF Nurislamova, IM Gubaidullin, RA Mulyukov. Prediction of drilling fluids loss during oil and gas wells construction. lP ^h Conference on Dynamical Systems - Theory and Applications, December 5-8, 2011, Lodz, Poland

4. Shengnan Wu, Laibin Zhang, Wenpei Zheng, Yiliu Liu, Mary Arm Lundteigen. A DBN-based risk assessment model for prediction and diagnosis of offshore drilling incidents. Journal of Natural Gas Science and Engineering, vol. 34, August 2016, pp. 139-158, DOI: 10.1016/j.jngse.2016.06.054 5. Qishuai Yin, Jin Yang, Shujie Liu (CNOOC) | Ting Sun, Wenlong Li, Lilin Li, Nanding Hu, Gang Tong, Xiaoliang Chen, He Deng. Intelligent Method of Identifying Drilling Risk in Complex Formations Based on Drilled Wells Data. SPE Intelligent Oil and Gas Symposium, 9-10 May, 2017, Abu Dhabi, UAE, DOI: 10.2118/187472-MS

6. Ekaterina Gurina, Nikita Klyuchnikov, Alexei Zaitsev, Evgenia Romanenkova, Ksenia Antipova, Igor Simon, Viktor Makarov, Dmitry Koroteev. Detection of failures in directional drilling using real-time analogue search. Preprint presented in the Journal of Petroleum Science and Engineering, June 7, 2019.

7. Ksenia Antipova, Nikita Klyuchnikov, Alexei Zaitsev, Ekaterina Gurina, Evgenia Romanenkova, Dmitry Koroteev. Data-based model for predicting drilling accidents. SPE Annual Technical Conference and Exhibition, September 30 - October 2, 2019, Calgary, Alberta, Canada, DOI: 10.2118/195888- MS

8. Sridharan Chandrasekaran, G. Suresh Kumar. Optimization of Rate of Penetration with Real Time Measurements Using Machine Learning and Meta-Heuristic Algorithm. International Journal of Scientific & Technology Research, vol. Issue 9, September 2019

Claims

Claim

1. A computerized method of warning about a future anomaly during drilling, including:

- formation of the first training data set for the first machine learning algorithm (MLA) based on historical well data;

- training the first MLA algorithm using the first training data set to form the first normal response model;

- determination of a potential drilling anomaly based on the first normal response model and real-time well data;

- generating a second training data set for the second MLA algorithm based on historical well data and sample data generated by the first normal response model;

- training the second MLA algorithm using the second training data set to form the second normal response model;

- prediction of the main parameters of the drilling process based on the second normal response model and real-time data of wells;

- determination of the actual drilling anomaly based on the predicted main parameters of the drilling process and real-time data of wells;

- generating a warning based on a potential drilling anomaly and/or an actual drilling anomaly.

2. The method of claim. 1, which generates several second training data sets for the second MLA algorithm.

3. The method of claim 2, wherein one or more training datasets from multiple second training datasets for the second MLA algorithm are selected based on operations and phases in real-time well data.

4. The method of claim 3, wherein real-time well data is pre-processed to determine operations and phases.

5. The method of claim 1 wherein the prediction is performed in real time.

6. The method of claim. 1, in which the main parameters of the drilling process are the average weight on the hook, the average torque and the average inlet pressure.

7. The method of claim. 1, prior to generating an alert, further comprising combining the potential drilling anomaly data and the actual drilling anomaly data.

8. The method of claim 7, further comprising updating the first and/or second normal response model based on the combined anomaly data.

9. The method of claim 1, wherein the warning is visible.

10. The method of claim 9 wherein the visible warning is based on colored light.

11. The method of claim 9, wherein the visible warning is text-based.

12. The method of claim 9, wherein the visible warning is based on an image.

13. The method of claim 1 wherein the warning is audible.

14. The method of claim 13, wherein the audible warning is based on a particular sound.

15. The method of claim 13, wherein the audible warning is based on a voice message.