Abstract
Drilling and workover operations represent a crucial part of a well lifecycle in terms of deliverability and
economics. Understanding the underlying phenomena that cause operational anomalies is the stepping
stone into early detection and control of undesired events, such as a kick.
The evolution of artificial intelligence and machine learning applications lend itself to well operations to
gain new efficiencies and unveil hidden insightful observations about downhole and surface operating
conditions. Incorporating the mechanisms of natural phenomena and big data, retrieved from sources
such as logging while drilling (LWD) and measurement while drilling (MWD), into machine learning
models boost capabilities for early detection of operational anomalies and mitigation of potential
negative consequences, while eliminating human-bias.
This paper highlights a novel machine learning model developed to streamline early detection for the
operational anomaly of uncontrolled hydrocarbon flow during well operations such as drilling. The
proposed technique detects and classifies the risk level of a kick much before it reaches the surface to
extend the safe response time limit. When this method is integrated with LWD data in real-time mode
by means of a software, and alarm system can be embedded to alert field hands about downhole
conditions. This does not only promote safer operations, but also significantly improve the availability
and reliability of critical information.
To further fine-tune the accuracy of the predictive model, multiple rounds of cross-validation were executed on the training data. It is evident that training machine learning models allows for more
learning through practice. The technique presented shows that big data and machine learning
algorithms are powerful tools to uncover hidden information and enable continuum leap improvement
in operational leadership.
Introductory Background
The daily flow of vast, fast and varied data represents vast opportunities for the oil and gas industry as
this data is processed and refined into meaningful information (Baaziz, 2014). Accordingly, machine
learning models are used to analyze the collected data to generate insights that would describe, predict
and prescribe operating trends. Other advantages of machine learning are elimination of human-bias
and streamlining error-prone repetitive tasks. Machine learning models are categorized into four major
categories: unsupervised learning, supervised learning, semi-supervised learning and reinforcement
learning (Noshi and Schubert, 2018).
One machine learning application that is gaining wider applicability is anomaly detection during drilling,
workover and completion operations. Many companies find it appealing to use machine learning algorithms to identify undesired outcomes, such as kicks and blowout, in real-time mode (Al-Ghazal,
2018). According to Unrau and Torrione (2017), anomaly detection machine learning algorithms offer
new capabilities, such as reduction in false alarm reports, higher anomaly detection rating and reduction
in detection time lag.
Influx of fluids into wellbore is an operational anomaly encountered during a drilling operation, even
though the flow of fluids at surface is controlled using methods, including change in mud weight and
shutting well. A severe case of fluid influx is uncontrolled flow of hydrocarbons from a high pressure
subsurface structure is termed as blowout. The Foundation for Scientific and Industrial Research at
Norwegian Institute of Technology (SINTEF) conducted a study and found that about 117 well control
incidents occurred in the period from 2000 through 2014. Given the importance of safety while drilling a
well, a relatively limited volume of literature is published in the area of kick detection and management
using data-driven models and machine learning algorithms.
Fraser et al. (2014) suggests a method to early detection of influx flow rate for offshore wells. Installing a
Coriolos Flowmeter outside of the riser at the seabed gives the ability to measure flowrate even before
the fluids reach the surface, reducing kick detection volume by a factor of 2. Unrau et al. (2017)
developed a machine learning algorithm to prevent false alarms of fluid influx. Fluctuations in flow rate
and mud levels due to operations, including changing mud pump rates, making connections and
transferring mud, were accounted for and the system was trained to accurately detect fluid influx or
mud loss.
Predictive Data Analytics Model Development
In this study, the operational anomaly of hydrocarbon kick is being detected using parameters recorded
through logging while drilling (LWD) and measurements while drilling (MWD). At the most fundamental
level, this novel approach considers the following conditions for a well in order to experience a kick:
Subsurface rock is porous enough to hold any fluid.
Subsurface rock pores contain fluid.
The actual subsurface pressure is higher than expected pressure (calculated based on pressure gradient data).
Using LWD tool data, it is possible to determine if the above three conditions are met in real-time. Kicks
are conventionally identified by a rise in fluid level at the mud pits. The proposed method will detect a
kick before it reaches the surface and alarm the crew ahead of time to be prepared for necessary well
control measures including to eliminate or abate any potential negative consequences. LWD tool
assembly parameters required for this study are as follows: Resistivity Log, Neutron Porosity Log, Bulk
Density Log and Pressure Measurement (Dowell et al., 2006).
Considering each of the three aforestated conditions separately and expanding on how it can be
achieved using LWD data is useful in developing the proposed predictive machine learning model. First,
subsurface rock should be porous enough to hold any fluid. Combination of Bulk density and Neutron
Porosity logs can determine the porosity of the subsurface formation. Figure 1 illustrates typical bulk density and neutron porosity logs. The positive separation of the two logs indicates a porous formation
with fluids. Second, subsurface rock pores should have presence of fluids. Resistivity logs can investigate the presence of fluids in the shallow, near to wellbore and deep regions of the hole. In this study, we consider any spike (in case of hydrocarbons) or drop (in case of brine) in resistivity measurement for an extended depth interval with a minimum 15 ft. This is to compensate for noise in measurements attributed to presence of fluid. This is illustrated in Figure 1.
Third, the actual subsurface pressure should be higher than expected pressure. Assuming downhole
pressure measurement tool is part of LWD and MWD assembly, the instrument measures downhole
pressure every 10 ft and the pressure gradient is calculated (Δpsi/10ft) and recorded every 10 ft. Any
unexpected rise in pressure gradient can be attributed to influx of fluids. However, noise in the data
could indicate a false kick. To mitigate such false indications, pressure gradient is calculated over 30 ft
interval (Δpsi/ 30ft), even though the tool may measure pressure every 10 ft. If the recorded pressure
gradient is higher than the expected pressure gradient by 30psi/30ft (green line on Figure 2), fluid influx
is assumed, Figure 2. Expected pressure gradient can be calculated knowing the depth of the well, hole
geometry, mud properties and volume of annular hydrostatic column.
This method can identify a kick just 30 ft after the pressure measurement tool records the pressure at a
given depth. Unlike traditional detection methods that rely on surface flow parameters and visual
inspection of mud pits volumes for kick identification, the proposed predictive method identifies a kick
much before the downhole influx reaches the surface. When this method is integrated with LWD data in
real-time by means of a software, an alarm system can be developed to notify the rig crew about
possibility of fluid influx. This ensures safer drilling operations and assists the crew to be prepared for
any well control activities. This kick identification method requires minimal cost of installation (only the
cost of alarm system) when a comprehensive LWD assembly is already installed.
A decision matrix, as shown in Table 1, is designed to indicate the risk of kick based on the data
recorded. Knowledge of the typical measurements of different logs and pressure gradients for a
particular oilfield is helpful to assign risk classifications of high, medium or low based on recorded
parameters’ value range.
Model Optimization
The novel machine learning technique, introduced in this paper, eliminates the time lag, often
encountered to detect and classify undesired events, utilizing big data retrieved in real-time mode from
logging while drilling and measurement while drilling. A pattern recognition logic, which is fit to existing
operating workflow frames, is used to analyze the LWD and MWD data and identify symptoms of
downhole formation fluid kick and its corresponding risk level.
The input parameters for the data-driven machine learning model are LWD and MWD data, including
resistivity, neutron porosity, bulk density and pressure gradient. These inputs (represented as vector
matrices) are the features and observations used to understand the fluid kick. The target variable of the
model is fluid influx classification (Yes or No influx), which is a discrete data set. The data science
classification decision tree method was selected to categorize the target variable because it is suitable
for discrete set of values, which is the case in this application. The ultimate purpose of the classification
decision tree method is to split the dataset into a class of either influx or no influx.
As far as model training to fine-tune accuracy and reliability, MWD and LWD data was used from
multiple wells to feed and calibrate the model. The trained model was further tested using multi-fold
cross validation test method. Additionally, multiple rounds of cross-validation were executed to
streamline the predictive data analytics model.
It is evident that machine learning model training allows the algorithm model to learn through repetitive
practices. This work is a testimony to the power of big data and machine learning algorithms to unveil
insights about drilling and completion operations that would enhance decision-making to safeguard
human lives and valuable assets.
Summary
Digitalization, in the form of artificial intelligence, machine learning and big data analytics, continues to
progress in the oil and gas industry, spanning upstream, midstream and downstream applications. In the
specific case of this paper, a novel machine learning model for early detection of the operational
anomaly of kick has been outlined. The machine learning model was developed to streamline early
detection of a kick to enable new control capabilities towards operational robustness, which is the need
of the hour to add non-existing efficiencies.
The model uses a decision matrix to categorize the risk of kick based on the data recorded from LWD
and MWD. The input parameters for the predictive machine learning model are LWD and MWD data,
including resistivity, neutron porosity, bulk density and pressure gradient. These inputs are the features
and observations used to describe the downhole conditions. The target variable of the model is fluid
influx identification and classification. The data science classification decision tree method was selected
to categorize the target variable because it is suitable for discrete set of values, which is the case in this
application. The ultimate purpose of the classification decision tree method is to split the dataset into a
class of either influx or no influx based on pre-determined value ranges. Finally, multiple rounds of
cross-validation were executed to verify the predictive data analytics model and its accuracy.
References
Al-Ghazal, M.A. “The Value of Digital Data Analytics,” Oil & Gas Vision 14: 10-11, 2018.
Baaziz, A. “How to Use Big Data Technologies to Optimize Operations in Upstream Petroleum Industry,” 21 st World Petroleum Congress, 15-1 June 2014, Moscow, Russia.
Fraser, D., Lindley, R., Moore, D. D., & Vander Staak, M. “Early Kick Detection Methods and Technologies,” SPE 170756-MS presented at the SPE Annual Technical Conference and Exhibition, 27-29 October 2014, Amsterdam, The Netherlands.
Dowell, I., Mills, A., Ridgway, M. and Lora, M. (2006). Petroleum Engineering Handbook. (Volume II). Society of Petroleum Engineers. 647-685.
Noshi, C. I. and Schubert, J. J. “The Role of Machine Learning in Drilling Operations; A Review,” SPE 191823-18ERM-MS presented at the SPE Eastern Regional Meeting, 7-11 October 2018, Pittsburgh, Pennsylvania, USA.
Unrau, S. and Torrione, P. “Adaptive Real-Time Machine Learning-Based Alarm System for Influx and Loss Detection,” SPE 187155-MS presented at the SPE Annual Technical Conference and Exhibition, 9-11 October 2017, San Antonio, Texas, USA.
Unrau, S., Torrione, P., Hibbard, M., Smith, R., Olesen, L., & Watson, J. “Machine Learning Algorithms Applied to Detection of Well Control Events,” SPE 188104-MS presented at the SPE Kingdom of Saudi Arabia Annual Technical Symposium and Exhibition, 24-27 April 2017, Dammam, Saudi Arabia.
Technical Paper authored by: Mohammed A. Al-Ghazal and Viranchi Vedpathak