Intelligent Directional Survey Data Analysis to Improve Directional Data Acquisition

In the oil and gas industries, some problems might cost a lot of money and effort to solve. With Artificial intelligence and machine learning techniques, many problems can be solved by using various algorithms such as Artificial neural networks (ANN). The literature review provides a lot of problems discussed in the field of oil and gas. It also provides a perspective on what type of algorithms are used and which works better. Working on the literature review is a critical factor in solving any problem. The developed item UGI42.03 provided some solutions for some problems that are divided into two parts. The first part used is mainly working with operations, and the second part is about improving the algorithms. Binder et al. [3] used an operational algorithm and implemented some equations. After the development, the researchers have gotten their results for the geophysical surveys they achieved to create a tool that is resistant to shock and high temperatures, as well as the issues of surveying for wellbores that are vertical, were solved by using micromechanical mechanics. In conclusion, by working on the study, they were focused on improving new methods for some types of surveys.

Seibi et al. [4] discusses the complexity of drilling wellbores that has been increasing over time, and it includes some of the techniques for the existing technologies of drilling, measurements, and data transmission. In this article, there will be a comparison of the advantages and disadvantages of each technique. MWD A method used to provide wellbore path and other measurements. Engineers are looking for new methods and techniques for surveying and drilling; directional drilling and surveying were mentioned in this article, including their techniques and transmission of the data. Researchers used many techniques like MWD System and bending moment measurement, which are used mainly for surveying, Gravity MWD System, which is used for directional drilling, curvature deduction, and dogleg survey.

Akinpelu et al. [5] points out a mapping of the pollution in the environment using a system of artificial intelligence, including different algorithms and techniques. For the model, they used algorithms such as support vector regression (SVR), which can be enhanced by applying (MS), which is searching manually, and genetic algorithm (GA) which is considered a searching algorithm. To predict, they supported the following model by using correlation coefficient (CC), root means error, mean absolute percentage, and error. As shown, GA and SVR-Gauss had better performance than MS and SVR, GA and SVR Poly by the accuracy of 63% and 536%, also MS and SVR performed better than GA and SVR Poly by 288%. While ranking the algorithms, it will be as the following, GA-SVR-Gauss > MS-SVR > GA-SVR-Poly. In conclusion, this work used the algorithm SVR as a hybrid for the parameters GA and MS. After Using the Correlation Coefficient, and RMSE, MAE, and MAPD, it appeared that GA-SVR-Gaussian is better than other algorithms as listed.

Xu [6] focused mainly on learning algorithms for machine learning, such as the Extreme Learning Machine (ELM), which is used for hidden layers. In the paper, they compared two algorithms which are Extreme Learning Machine (ELM) and Least Squares Support Vector Machine (LSSVM). For a regression model, Neural Networks (NN) had many risks for the work; on the other hand, Support Vector Machine was the right algorithm to be used in these conditions. Neural Network (NN) takes a longer time to create a regression model, and nowadays, they have brought a new algorithm that is called Extreme Learning Machine (ELM). ELM is much faster in learning and easier than CNN, as well as solving different problems of engineering. As discussed, this paper points out the impact of algorithms on the localization of a plate structure, and models view the location of the issue. For a security system there must be some types of tools as detection tools to prevent unwanted outcomes.

Ahmad et al. [7], many algorithms were used, such as Support Vector Machine (SVM), to solve many issues like false alarms by decreasing it and the rate of detection by increasing it. Some techniques weren't efficient for large datasets, and for that, some of the classification techniques must be used like Support Vector Machine (SVM), Random Forest (RF), and Extreme Learning Machine (ELM), as they applied these techniques while working on this issue. After training and testing, the full data sample had 65,535 records. Support Vector Machine (SVM) had a 98.7% accuracy, Random Forest (RF) had a 97.7% accuracy, and Extreme Learning Machine (ELM) had 99.5% accuracy. These accuracies are received using full records of the Dataset, and it points out that ELM got the best result.

Škrjanc and Vulić [8] provides information about directional wells and the anti-collision because collisions can be a big issue while planning for a position for a wellbore as well as surveying and drilling. It was in Alaska on the north slope, which is a huge space between seas. A comparison was made for error models, which are Operator Wellbore Survey Group (OWSG) and Industry Steering Committee for wellbore surveying accuracy (ISCWSA). The goal of this study was to improve anti-collision methods and guides for drilling to avoid collisions, which means drilling in a safe direction due to the problems occurring while drilling in a wrong direction or place. The separation factor (SF) indicates collision avoidance as its equation is by dividing the center-to-center distance by minimum separation. As a result, it appears that the Operator Wellbore Survey Group (OWSG) has better performance, and it should be used as a default model.

Mahajan et al. [9] provides information about drilling operations for fluids that are mud. There are many objectives to achieve while drilling mud, mainly avoiding blowouts and kicks and by keeping the wellbore stable. Researchers have used the machine learning (ML) approach to achieve some of their goals. Different Machine learning methods were discussed and compared for either supervised or unsupervised methods to predict the issues that could appear while drilling. Machine learning algorithms used are artificial neural networks (ANN), random forest (RF), which is used for a regression approach, XGBoost (XGB), support vector machines (SVM), and K-nearest neighbors (KNN), and lastly, Markov decision process (MDP). As a result, they applied these algorithms in many approaches, and some of the mentioned techniques performed the best and achieved the greatest result, such as ANN and RF.

Gul [10], investigates a random forest algorithm was applied that observed the productivity of field-wide, including the data injection as a given value to estimate the timelapse oil profiles within wellbore locations. The progress is estimated using calculated field data by getting the complex structure, heterogenous, and heavy faulted offshore reservoir. This algorithm was obtained to apply the inverse modeling standard, which uses the productivity of the field, and it is an injection to estimate the time-lapse saturation. Wireline logging has limitations not only in the productivity and it is injection but includes interruption, limitation of the operation tools, difficulties in deviated wells, and pulling the tubing pump. The limitations and risks could include well intervention and how it can face an increase in the cost besides risks of its operations that could bring limitations for the frequency of data acquisition. Data acquisition could cause some limitations while applying the model since the given data is not dependable when it comes to visibility, and it is a sufficient perspective. The given input features in this model are wide-field injection and productivity, and the produced output will represent how deeply the time-lapse of oil saturation inside the location of the well by using the reservoir matching history technique. The outcome of the model shows the efficiency of applying the random forest algorithm for getting matched results between productivity and injection with the oil saturation.

Wang et al. [11], there is a concentration on the progress of novel-driven data and their models. By using deep learning with its analysis techniques of the data to get productivity progresses using good horizontal pairs and their temperature from vertical wells observation to customize the barriers. Data were calculated and gathered from many different fields which are working on the oil, such as geometrics, proportions, and spatial distributions of barriers for the building of SAGD models. The neural network of deep learning that was applied is known as CNN to catch the complexity of nonlinear relations between inputs and their outputs parameters. Because of how the thickness of bitumen oil sands, conventional recovery methods like water flooding do not work in all situations. There were two horizontal wells drilled parallelly in the ground of the reservoir, the upper well considered as the injector and the lower well considered as the producer. Usually, it is difficult to locate the barriers and customize their sizes and dimensions. The CNN model was implemented to solve this issue in two ways. Getting a bunch of geological models by the static data and applying the changed operations of models each time by the dynamic data, which in addition will be helpful to decrease the mismatch of productivity between models and their data. To analyze the inputs and outputs of the driving of the data in the models, there were several parameterization phases applied. One of the main purposes behind doing this experiment is to verify if the deep learning techniques could be applied within the dynamic integration of the data for reservoir customization.

Table 1. Summary of related work

Liang et al. [12], a Genetic algorithm (GA) and Back Propagation neural network (BPNN) are applied to analyze the relations functions between predicted outputs and the number of customized parameters in the drilling related data. The model decreased error occurred by the usage of one feature parameter, and higher the speed of BP using genetic algorithm, additionally avoiding dealing with local extreme value. By that, the overflow drilling diagnosis accomplished and decreased the percentage of wrong and missed judgments of drilling kicks. Drilling kicks are considered as big damage to the drilling speed, and it is produced quality. Also, it could cause fluid and leakage of the drilling progress, which results can be considered as a large amount of loss in the cost. These risks of drilling can be prohibited and decreased in their possibility of occurring if it is discovered at the beginning of their occurrence. We will compare the literature reviews by first comparing the key elements in every paper that we have reviewed such as the method, the accuracy rate, and the number of samples in the dataset, and discuss the findings of the studies alongside the techniques that performed well in the studies and managed to achieve the highest accuracy. This comparison will be between related works under specific criteria for the methods that were used, the accuracy rate, and the dataset used in the related work. A summary of the related work is listed in Table 1.

From the comprehensive literature review, we have noticed that problems that were solved using ANN have shown satisfiable and reliable results. However, the Dataset plays a huge role. Some datasets may perform poorly using ANN.

For the undergoing study, we believe that trying several techniques such as ANN, Recurrent neural networks (RNN), Deep neural networks (DNN), and Long Short-term memory (LSTM) will help us to get the best and most reliable model that is possible by using these techniques. Some of the papers consisted of surveying and drilling [3]; it was noted that to maintain a high accuracy during the operations, many techniques were used to achieve higher accuracy and to maximize the efficiency of drilling [4]. Methods used such as MWD System, Gravity MWD System, curvature deduction, dogleg survey [8] and the tangential method and their improved versions. Some of these methods were used for surveying, and others were used for drilling [5]. The algorithm Support Vector Regression provided a good result for predicting discrete values, as well as being enhanced by other algorithms like (MS) that are manual searching, or genetic algorithm (GA), that is used as a searching algorithm [6]. After comparing different types of algorithms, it appeared that Extreme Learning Algorithm (ELM) has better performance than algorithms like CNN because ELM is faster [7]. ELM had the best accuracy result in some of the reviewed papers, which proves how useful the algorithm [9]. Issues like collision occurrence should be solved by methods such as OWSG that had better performance and result, and ISCWSA that are considered anti-collision methods [10]. ANN was one of the best algorithms used during searching and studying these papers because it had some of the best accuracies and results [11]. The results of the model show the efficiency of applying the random forest algorithm for getting matched results between productivity and injection with the oil saturation [12]. Nonlinear relations contain a lot of complex computations. Depending only on the general algorithms, the model fit of the data will not be accomplished as it is supposed to be. Therefore, neural networks will be helpful to get the diagnosis faults and their recognition patterns.

This section describes the dataset and how it was cleaned. It also includes a detailed explanation of the experimental setup and how the models are constructed. Lastly, the optimization process and the performance measures that were applied in the proposed work.

4.1 Dataset description

The dataset comprises 34068 rows allocated for 157 wellbores and 13 columns which represent the wellbore number, and it is main measurements which are MD, Inclination, Azimuth, and the validity of the station based on those three measurements. The rest columns are representing the eight rules which clarify why the station is not considered a valid station, with one column for each rule. All the algorithms considered in the study have been trained with the same set of instances after due preprocessing was conducted on the dataset. Such as normalization, eliminating redundant instances and missing values and handling outliers.

4.2 Experimental environment

The experiment was implemented using Python, machine learning algorithms were done using Jupyter notebook, and deep learning algorithms were done using Google Colab due to the simplicity of libraries in installation on Google Colab.

4.3 Performance measures

The proposed model is evaluated using the confusion matrix, precision, and recall, accuracy, and F1-score measures. The confusion matrix will contain a summary of the model predictions, to identify the mistakes for each class in the model and calculate the correct and incorrect predictions [31, 32]. In the Precision, true positive values are required to determine the precision. In the recall, false positive values are required to determine the recall. The accuracy rate is another type of performance measure, the classification accuracy can be computed by dividing the number of true positives and false positives by the total number of samples (in the formula N indicates the total number of samples). the F1-score can be defined as the harmonic mean between precision and recall, it indicates how stable and how accurate is the classifier [33-40]. The measures are given in Eq. (3) to Eq. (6).

Precision $=\frac{\text { True } \text { Positive } (T P)}{\text { True } \text { Positive } (T P)+\text {False } \text { Positive }(F P)}$          (3)

Recall $=\frac{\text { True } \text {Positive}(T P)}{\text { True Positive} (T P)+\text {False }\text {Negative}(F N)}$          (4)

Accuracy $=\frac{\# \text { Correct } \text {Predections }}{N}$          (5)

$F 1=2 * \frac{1}{\frac{1}{\text { Precision }}+\frac{1}{\text { Recall }}}$          (6)

4.4 Results and discussion

Table 2 indicates the result of the seven considered algorithms, most of the algorithms managed to score a high accuracy rate above 80%. The parameters for the network’s structures are already explained in the previous sections. The deep learning algorithms ANN, LSTM, and GRU managed to get a 100% on all the evaluation metrics including the accuracy rate, Precision, Recall and F1 score. BI-LSTM had a slightly lower accuracy rate achieving 98.7%, BI-Simple RNN was lower than BI-LSTM achieving 82% accuracy rate, and ELM had performed poorly making it the lowest score across all the algorithms reaching an accuracy rate of 55.3%. The machine learning algorithms SVM and Naïve Bayes provided better results, SVM managing to get 100% in recall, 81.1% in accuracy rate, and 98.6% in F1 score. Naïve Bayes has provided exceptional results managing to get 100% in Accuracy, Precision, Recall, and F1 Score.

Table 2. Summary of the result

Figure 7 presents a comparison of the proposed scheme with the state-of-the-art schemes in the literature namely [7] and [11]. Apparently the proposed scheme outperforms in terms of accuracy, precision, recall and F1-score, respectively.

7.png

Figure 7. Comparison with the state-of-the-art schemes