A new methodology for kick detection during petroleum drilling using long short-term memory recurrent neural network
Introduction
Machine learning is an interesting field which involves finding relationship amongst data. There is a lot of interest in machine learning because of its success in challenging domains such as speech recognition, medical imaging, etc. (Alom et al., 2019). The level of success achieved in solving a task through the use of machine learning depends on the training data as well as the type of machine learning algorithm used. For example, while convolution neural network (CNN) is designed to take advantage of spatial relationship in two-dimensional data, recurrent neural network (RNN) is designed to explore temporal relationships in data (Alom et al., 2019).
During drilling for exploration/production of hydro-carbon, kick (i.e. influx of hydrocarbon into the well bore) can occur. This can be very dangerous especially if it is a gas influx, as this can lead to blowout when the gas influx is not properly controlled. Therefore, occurrence of kick is detected by locating sensors to monitor drilling parameters. Some parameters typically monitored for kick detection are torque, rate of penetration (ROP), weight on bit (WOB), flow differential and stand pipe pressure (SPP). Kick is typically indicated by sudden increase in torque, increase of outflow over inflow, increase in SPP, increase in ROP and decrease in WOB. Eq. 1 describes the d-exponent also known as the normalised rate of penetration. The d-exponent can be used as a means by which kick occurrence can be identified; this was demonstrated in the article by Mao et al., (Mao and Zhang, 2019) and by Tang et al., (Tang et al., 2019). refers to rotary speed and refers to bit size.
Several researchers have utilized machine learning and a combination of different sensors for kick detection. This can be seen in Table 1.
In addition to the articles on machine learning approach for kick detection presented in Table 1, a systematic approach for kick detection was presented in the article by Karimi et al., (Karimi and Van, 2015). The systematic approach shows how kick occurrence can be detected by observing flow in, flow out, pit gain, pump pressure and annular pressure profile.
Sensor data used for kick detection in several cases are time series in nature. When temporal relationship exist in sensor time series data, the event indicated at a time (kick or no kick) could be described as a function of the sensor reading both at that particular time, as well as sensor readings at earlier times in order to improve the accuracy of kick detection. For example, temporal relationship can be observed in Fig. 3 of the article by (Hargreaves et al. (2001)). This figure shows that flow out ramps up over a period of time in response to kick. We can observe that using the flow out value at an instant in time is not sufficient to indicate kick occurrence. Instead, several consecutive values of flow out will be needed in order to observe the trend or pattern of flow out over time for accurate kick detection.
Several dynamic neural network algorithms such as focused time delay neural network, classical recurrent neural network, recursive neural network can be utilized in learning temporal relationships in data. When it comes to learning long term temporal relationships in data, the long short-term memory recurrent neural network (LSTM-RNN) offers the advantage of overcoming the gradient vanishing problem which makes it difficult for other dynamic neural network algorithms to learn long term dependencies. The LSTM-RNN achieves this by having an architecture made up of memory cell, input gate, forget gate and output gate (Graves, 2012).
The objective of our work is to develop a methodology which can help achieve automatic, early and accurate detection of kick onset using d-exponent and standpipe pressure data. While effort has been made to develop equations which can be used to detect kick as a function of d-exponent, standpipe pressure and flow measurement data, a lot of uncertainties are still encountered. For example in the article by Tang et al., (Tang et al., 2019), kick indicating parameters (namely, flow-in rate, flow-out rate, SPP, WOB and ROP) were used in two equations. These equations are: (i) drilling parameter group (DPG) which is the same as d-exponent; (ii) flow parameter group (FPG) which utilises flow-in rate, flow-out rate, compressibility of the drilling mud () and SPP as shown by Eq. 2. represents reference pressure which is 14.7 psig at ground surface.
Also, in the article by Mao et al., (Mao and Zhang, 2019), DPG, FPG and pit volume gain were used for kick detection. The methodology used in (Tang et al., 2019) was also applied in (Mao and Zhang, 2019).
Although FPG, DPG and pit volume gain can be utilized together for kick detection, some challenges such as: (i) the ratio by which the probability of kick occurrence using DPG, FPG and pit volume gain are combined for early kick detection with minimal chances of false alarm, (ii) the tolerance value of DPG, FPG and pit volume gain adopted for accurate kick detection, and (iii) the threshold value of kick-risk index (KRI) to indicate kick show areas for further research (Tang et al., 2019). The benefit of machine learning is that with sufficient training data, the machine learning algorithm can capture the relationship between input parameters/equations such as the ratio by which the inputs are combined for kick detection. In this paper, data-driven approach using LSTM-RNN is implemented to utilize the relationship between SPP and d-exponent in order to achieve early detection of kick without false alarm. The industrial significance of the proposed methodology is that it can improve the chances of obtaining automatic, accurate, and early kick detection with sufficient training SPP and d-exponent data for different scenarios of kick and no-kick periods during drilling.
This paper is structured as follows. Section 2 describes the proposed kick detection methodology. Section 3 describes the data used for verification of the methodology. Section 4 reports the result obtained by using the proposed methodology on the data described in Section 3. Concluding remarks are provided in Section 5.
Section snippets
Methodology
The flowchart of the proposed methodology is presented in Fig. 1. Detailed description of the methodology is presented in subsequent subsections.
Data used for verifying methodology
The data used for this study were obtained from the article by Tang et al., (Tang et al., 2019). The smoothed DPG (d-exponent) data and SPP data of Fig. 6, Fig. 8, representing cases 1, 2, 3 and 4, of (Tang et al., 2019) were obtained. This was done by reverse engineering the figures using WebPlotDigitizer and UN-SCAN-IT software. These software have been used in some articles for data extraction and analysis (Phattanawasin et al., 2016) (Mani et al., 2018). The values obtained by reverse
Results and discussion
Considering that data corresponding to four cases are available, four categories of training and testing will be done. Category 1 involves using all data points/samples of case 2, case 3 and case 4 to train while all data points/samples of case 1 is used for testing. Category 2 involves training with all data points/samples of cases 1, 3 and 4 while all data points/samples of case 2 is used for testing. Category 3 involves training with all data points/samples of cases 1, 2 and 4 while all data
Conclusions
We have proposed a methodology for the successful use of data-driven approach for kick detection by using an ensemble of LSTM-RNNs. The aim of using LSTM-RNN is to overcome the challenges of learning how to utilize d-exponent and stand pipe pressure data for early kick detection without false alarm.
The use of an ensemble of LSTM-RNNs is able to achieve early kick detection without miss and false alarm in all the four drilling cases considered here, whereas the ensemble of simple ANNs is able to
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
The authors gratefully acknowledge the Niger Delta Development Commission for their financial support, as well as the financial support provided through Canada Research Chair Program in Offshore Safety and Risk Engineering.
References (30)
- et al.
Time series data analysis for automatic flow influx detection during drilling
J. Pet. Sci. Eng.
(2019) - et al.
Synthetic Well Logs Generation Via Recurrent Neural Networks
Pet. Explor. Dev.
(2018) - et al.
Ensemble machine learning: an untapped modeling paradigm for petroleum reservoir characterization
J. Pet. Sci. Eng.
(2017) - et al.
Canadian association of radiologists white paper on artificial intelligence in radiology
Can. Assoc. Radiol. J.
(2018) - et al.
A State-of-the-Art Survey on Deep Learning Theory and Architectures
Electronics
(2019) - et al.
An automated kick alarm system based on statistical analysis of real-time drilling data
Old Spe J.
(2019) - et al.
Early kick detection for deepwater drilling: New probabilistic methods applied in the Field
- et al.
Spotting a false alarm. Integrating experience and Real-time analysis with artificial intelligence
- et al.
Improved and robust drilling simulators using past Real-time measurements and artificial intelligence
- et al.
Early kick detection using Real time data analysis with dynamic neural network: a Case study in Iranian oil fields