DOI: 10.

2118/204462-PA Date: 23-March-21 Stage: Page: 551 Total Pages: 12

Early Stuck Pipe Sign Detection with

Depth-Domain 3D Convolutional Neural
Network Using Actual Drilling Data
Naoki Tsuchihashi, Ryota Wada*, and Masahiko Ozaki, University of Tokyo; Tomoya Inoue, Japan Agency for
Marine-Earth Science and Technology; Konda Reddy Mopuri and Hakan Bilen, University of Edinburgh;
Tazuru Nishiyama, Japan Petroleum Exploration Co., Ltd.; Kazuhiro Fujita, INPEX Corporation;
and Kazuya Kusanagi, Japan Oil, Gas and Metals National Corporation

Summary
A real-time stuck pipe prediction using the deep-learning approach is studied in this paper. Early signs of stuck pipe, hereinafter called
stuck, are assumed to show common patterns in the monitored data set, and designing a data clip that well captures these features is crit-

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

ical for efficient prediction. With the valuable input from drilling engineers, we propose a 3D-convolutional neural network (CNN)
approach with depth-domain data clip. The clip illustrates depth-domain data in 2D-histogram images with unique abstraction of the
time domain. Thirty field well data prepared in multivariate time series are used in this study—20 for training and 10 for validation.
The validation data include six stuck incidents, and the 3D-CNN model has successfully detected early signs of stuck in three cases
before the occurrence. The portion of the data clip contributing to anomaly detection is indicated by gradient-weighted class activation
map (grad-CAM), providing physical explanation of the black box model. We consider such explanation inevitable for the drilling engi-
neers to interpret the signs for rational decision-making.

Introduction
Stuck is one of the major drilling problems that accounts for nonproductive time. In the past, one-third of the wells drilled in the Gulf
of Mexico and the North Sea experience stuck pipe problems (Howard et al. 1994). Moreover, stuck may lead to wellbore abandonment
in the worst scenario. Regarding its severity, there is a strong demand for early stuck sign-detection methods to avoid stuck incidents.
The drillstring is considered to be stuck when it loses the freedom of movement; that is, it can neither rotate nor move up and down.
Also, it cannot be retrieved from the wellbore due to external force. According to the mechanisms, stuck is roughly classified into three
types, namely mechanical stuck, differential stuck, and geometrical stuck (Alshaikh et al. 2018). Each stuck type could be separated
into smaller classes, for example, mechanical stuck includes both packoff and bridging. Also, each stuck type has respective indicators.
As a result, one of the major challenges in detecting early sign of stuck is that there is not a single pattern in a single parameter indicat-
ing all stuck pipe incidents (Alshaikh et al. 2018).
Conventionally, when detecting an anomaly during drilling operation, alerts are generated when there is an anomalous deviation
between the measured value and predicted value (Salminen et al. 2017). Predicted value is often calculated from preliminary simulation
by a physics-based model, which requires geological survey in advance. However, the conventional method raises many false alarms
due to the insufficient accuracy of geological information. Unrau et al. (2017) claim that this fosters an environment in which drillers
can become desensitized to alarms. As a result, drillers often rely on their experience and intuition instead of the alarm system, which
occasionally results in incidents where meaningful alarms are ignored. It could be concluded that the current alarm systems require
higher reliability.
Recently, several researches have focused on the use of drilling data obtained in real time with the machine-learning approach,
being motivated with its success to detect complex patterns in high-dimensional data. Deep neural networks have become the state-
of-the-art in many disciplines including computer vision, natural language processing, and speech recognition. Unrau et al. (2017) pro-
posed an adaptive alarm system to reduce false alarms in kick detection. The system uses a machine-learning algorithm (not specified)
and moves the alarm threshold dynamically in real time by using the drilling data obtained during the operation. As a result, false
alarms were dramatically reduced. Although the system was designed to detect kick events or loss circulation, it has shown an effi-
ciency of the data-driven approach against general anomalous events during drilling.
There are various studies that address machine learning for stuck pipe prediction, such as Heinze and Al-Baiyat (2012), Jahanbakh-
shi et al. (2012), Lind and Kabirova (2014), Agwu et al. (2018), Magana-Mora et al. (2019), and Zhu et al. (2019). These studies intro-
duce machine learning with selected drilling parameters in the time domain to develop a prediction model. The model’s results are
successful, although the details of the model are not provided. Some suggest the possibility of including a longer duration of past data.
Alshaikh et al. (2019) constructed three data-driven models to detect the stuck incident at the time of occurrence. Each model uses dif-
ferent machine-learning algorithms, namely decision tree, artificial neural network (ANN), and support vector machine. All models
have identical input and output. The input is 13 surface-drilling parameters including modified value such as moving average, and the
output is the probability of stuck. The models were trained, validated, and tested with nine stuck cases using nested cross-validation.
All models achieved high accuracy and precision scores beyond 90%, which indicates accurate prediction with few false positives. The
result was further validated by inspecting the real-time prediction of the decision-tree model. The decision-tree model has successfully
detected stuck at the time of occurrence. Although the model was designed to detect the stuck itself instead of its early sign, the model
raised an alarm way before the stuck occurrence, and they argued that the model may have caught an early sign of stuck. Abbas et al.
(2019) applied ANN and the support vector machine to predict the best remediation after stuck based on past experience.
We assume that there are several issues in the previous study, which are listed as follows:
• From the drilling engineering point of view, the input data shall include information of long duration to detect the change of drill-
ing behavior that indicates early signs of stuck pipe.

*Corresponding author; email: [email protected]

Copyright V
C 2021 Society of Petroleum Engineers

Original SPE manuscript received for review 17 July 2020. Revised manuscript received for review 3 October 2020. Paper (SPE 204462) peer approved 12 October 2020.

April 2021 SPE Journal 551

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 552 Total Pages: 12

• Regarding various stuck mechanisms and indicators, abundant training cases are necessary to learn various early signs of respec-
tive stuck mechanisms.
• Considering the practical use, it is recommended that the produced stuck risk is interpretable to drillers to make the correct deci-
sion of following operations. However, physical interpretation methods are not presented.
The objective of this study is to construct an effective early stuck sign-detection model by a data-driven approach with implementa-
tion of drilling engineers’ domain expertise. In addition, an effective model shall be interpretable to support the decision-making of
drilling engineers.
In this paper, we propose a model based on pattern recognition, assuming that early signs of stuck will show common patterns in the
monitored data set. Though deep-learning techniques are powerful and work well on raw input, how to represent the input to a deep net-
work is also important. Thus, the design of the data clip that well captures these features is critical for efficient prediction. Domain
knowledge of drilling engineers is implemented into data clip design by working with oil and gas companies and research organizations
with abundant drilling experience. Various early sign prediction models will be trained and validated with actual drilling data including
up to 43 stuck cases. Moreover, we implement a physical interpretation method of the stuck risk to improve the reliability of the
machine-learning model.

Field Data Set

Data Collection. The actual drilling data from drilling fields are provided by several oil and gas companies. The data are prepared in
multivariate time series with 19 drilling parameters. Along with the sensor data, timestamps of the stuck incident are provided for each

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

stuck case with 15-minute precision according to the daily drilling reports. Statistics of the data is shown in Table 1.

Company A Company B Company C

Sampling interval (seconds) 4 4 1
Number of well sites 34 5 7
Number of stuck events 43 1 5
Total length (days) 3,382 305 131
Location Onshore Onshore Offshore

Table 1—Statistics of the drilling data.

In general, the drilling parameters used in this study are considered important and available during the drilling operation. We were
able to extract these data from most of the wells used in this study, although some wells did not record them despite obtaining them.
Regarding the rapid evolution of data-driven technology, we assume that the availability of these drilling data will increase even more.
There is a great variation in the data due to several factors. One is diverse measurement conditions (e.g., drilling sites, formations,
instruments, sensors), which causes variation across different wells. The data includes data from different drilling rigs. They also have
different hole sizes (maximum hole size is 36 in. and minimum hole size is 6 in.), various bottomhole assemblies (both steerable motors
and rotary steerable systems), mud types (water-based mud and synthetic-based mud), bit types (polycrystalline diamond compact
[PDC] bit, insert bit, milled tooth bit, hole opener), and well profiles (vertical wells, inclined wells, and horizontal wells).
Another is various drilling operations, which gives variation within a well. The others are sensor drifts and errors, which could not
be identified precisely.

Data Preprocessing. Sampling interval is unified to 4-second intervals, and apparent sensor errors are removed by setting a loose
threshold for each drilling parameter. Also, timestamps of the stuck incidents with 15-minute precision were considered to be insuffi-
cient to learn early signs appearing within a few minutes before stuck. Therefore, looking into the time series data carefully, we identi-
fied the time when the stuck started with 4-second precision.
Data calibration was not considered in the preprocessing. The prediction model will deal with the real-time drilling data that are not
calibrated. Thus, the prediction model will be robust to such calibration errors.
As can be seen from Table 1, field data from Company A contains most of the stuck events. This study will mainly focus on the data
from this company.

Methodology
Machine learning systematically applies algorithms to synthesize the underlying relationships among data and information (Awad and
Khanna 2015). Machine-learning methods could be roughly divided into three types: supervised learning, unsupervised learning, and
reinforced learning. In this study, we assume that early signs of stuck appear before the stuck occurrence. Therefore, supervised learning
is adopted, and the model is designed to learn the patterns of early signs. Data obtained shortly before the stuck is labeled 1 (positive),
and data obtained further before are labeled 0 (negative). The model is trained with the binary labeled data to output the probability of
the early sign, namely the stuck risk. The model is intended to output high stuck risk shortly before the stuck occurrence. The concept
explained above is shown in Fig. 1.

CNN. CNNs are a kind of ANN that are widely adopted for 2D pattern recognition (LeCun and Bengio 1995). A simple ANN has an
architecture of fully connected perceptrons, whereas the CNN has an additional architecture of locally connected perceptrons before the
simple ANN layers. This specific architecture with local connections limits the spatial information transferred to each neuron in the
next layer and allows the model to catch the local pattern of the input. As well as general ANN models, weights are updated iteratively
to minimize the loss function. Because the problem setting is a binary classification of positive and negative samples, binary cross
entropy is adopted as a loss function and Adam from Kingma and Ba (2015) is used as an optimizer.
In the CNN, weights in the convolution layer (kernels) are also updated, which represents the pattern learning. CNNs are widely
applied to 1D data in natural language processing and speech problems, 2D data in image recognition tasks, and 3D data in video recog-
nition tasks. In addition, CNNs are extensible in the dimension of pattern learning.

552 April 2021 SPE Journal

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 553 Total Pages: 12

Stuck

Stuck risk Early sign Stuck

t
1

0 t
N P

0: negative 1: positive

Fig. 1—Our concept of supervised learning.

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

Feature Engineering. Because CNN learns the patterns in the input data (clip) for each class, the model performance is greatly
dependent on the design of the clip. Although raw data are provided as a time series, time series clips are not necessarily the most effec-
tive format to learn the pattern. For example, voiceprint could be clearly identified by converting the raw time series data into a spectro-
gram. A method to improve the model performance by devising the clips is called feature engineering, and expertise in the domain is
critical for an effective feature engineering. As for early stuck sign detection, an effective feature engineering is not investigated in the
previous studies. Therefore, the feature engineering method was investigated in an exploratory manner. As a first step, clips were pre-
pared in time series with 1-hour length, and the 1D-CNN model was adopted to learn the pattern of time series clips. However, we
could not observe meaningful predictions with this setup.
Later, 1-hour time clips were converted into a power spectrum, and the 1D-CNN model was adopted to learn the patterns in the
power spectrum clips. Although the results showed few meaningful predictions, the performance was not satisfactory. Further feature
engineering from the depth-domain approach was conducted.

Advanced Feature Engineering from the Depth Domain. The depth-domain feature engineering incorporates drilling engineers’ per-
spectives. Drilling engineers often inspect the distribution of surface-measured drilling parameters in comparison with the past record
obtained at the nearby depth. For example, during run-in-hole (RIH), if the hookload is smaller than that recorded during the previous
RIH at the same depth, engineers detect an increase in friction drag. This suggests a deterioration of hole condition. Another example
can be seen in the topdrive torque—the rotational torque on drillstring measured at the topdrive—behavior during drilling operations.
Even though records are not available for the same depth, an abrupt change of torque behavior in depth suggests a change in the
hole condition.
It is worth mentioning that capturing such changes in a data clip is not straightforward with a finite time duration. The data clip shall
include not just the data indicating anomaly but also the data recorded in the nearby depth. Otherwise, the model cannot compare the
data with past records. However, it is not clear how long a period the data clip should include. Ideally, it would be to use the whole data
set for that well operation. The increase of input parameters would make the model training significantly difficult.
In this paper, we propose a data clip using a novel representation of the drilling data set. Fig. 2 is a scatter plot of data points with
depth as vertical axis. Here, the depth is measured depth, the length of wellbore along the path. The red data points indicate the data
recorded in the last 24 hours (hereafter mentioned as “signal”), and the green data points indicate all the data recorded in that well
excluding the last 24 hours (hereafter mentioned as “history”). Each of the two panels in Fig. 2 corresponds to RIH with hookload on
the horizontal axis and drilling case previously mentioned with topdrive torque (TD_trq) on the horizontal axis. These both are data
obtained shortly before the stuck incident. Here, we can visually observe the anomalous behavior of the recent data. To the best of our
knowledge, this is the first attempt to apply depth-domain representation of data clip for machine learning.
The example illustrates that there are cases in which patterns of early sign are apparent when depicted in comparison with recorded
data in the depth domain. Such representation allows us to directly compare the data. The simplicity of representation is achieved by a
bold abstraction of time-domain information to a binary categorization of present and past.
Assuming the features of early stuck signs are better captured in depth-domain representation, the scatter plot shall be proceeded to
a data clip that can be handled by CNN. Here, the scatter plots are converted into 2D histograms by defining a bin. The process to con-
verting raw time series into the proposed format is shown in Fig. 3. Although the depth data point is created for each drilling parameter,
the explanation below focuses on a single drilling parameter X for simplicity. As for the data obtained at time t, a 2D histogram of bit-
depth and X is created, for both signal [t  2ðhoursÞ t] and history [t  7ðdaysÞ t  2ðhoursÞ]. Here, 2 hours is a hyperparameter, and
7 days is sufficiently long to cover the whole data set during well operation. In stacking the signal and history along a new axis, the
depth data is converted into a 3D tensor. Next, logð1 þ xÞ is applied to every element in the tensor to emphasize the low-frequency bins
and suppress the effect of high-frequency bins.
The length of signal and history, the spatial range, and the resolution of the 2D histograms are manually determined and presented
in Table 2. The horizontal resolution is set higher than the vertical resolution to obtain precise information of the horizontal deviation.
The vertical range is set relative to the current bitdepth to focus on the nearby depth. Also, we use only four parameters, topdrive speed,
the rotational speed of drill bit (TD_spd), the rotational torque on the drillstring measured at the topdrive (TD_trq), hookload, and
standpipe pressure (SPP_pressA), the pressure of mud injection monitored at the standpipe manifold to avoid model complexity that
could mislead the optimization in the learning phase. The chosen parameters are generally investigated by the drilling engineers in the
case of suspicious activities.
The T, V, H-sized 3D tensor introduced above is created for respective drilling parameters, and are stacked along new parameter
axis, resulting in a P, T, V, H-sized 4D tensor. Then 4D tensors from different time instances are gathered and stacked to make a mini-
batch; thus, a clip is finally a 5D-tensor with the size of B, P, T, V, H.
• B: batch size (¼ 16)
• P: drilling parameters (¼ 4)

April 2021 SPE Journal 553

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 554 Total Pages: 12

• T: time size (¼ 2, signal and history)

• V: vertical size of the 2D histogram (¼ 32)
• H: horizontal size of the 2D histogram (¼ 128)

Hookload (ton)
TD_trq (kNm)
100 125 150 175 200 225 250 275 300 0 2 4 6 8
4000
Stuck depth
Red: signal 4520
Green: history
4200

4540
Bitdepth (m)

Bitdepth (m)
4400
4560

4600
4580

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

4800
4600

5000

Fig. 2—Scatter plot of observed data in depth domain. Both plots represent a stuck case, in which the solid red line indicates the
depth of stuck occurrence. Red dots indicate the data recorded in the last 24-hour “signal,” and green dots indicate all the data
recorded in that well, excluding the last 24-hour “history.” The left figure corresponds to RIH with hookload on the horizontal axis,
and the right figure corresponds to the drilling case previously mentioned with TD_trq on the horizontal axis.

Horizontal range:
Prefixed for each parameter
3D-clip
Bitdepth (t) – 100 (m) 0

5
(kN/m)

25 TD_trq
10
0
1000 15
Bitdepth Vertical Vertical range:
(m)

20
0 bins: 32
t Bitdepth (t) 25

30
t - 7 (days) t - 2 (hours) Bitdepth (t) + 10 (m) 0 20 40 60 80 100 120
Horizontal bins: 128

Fig. 3—Creating a depth-domain data clip from raw time series data.

Symbol Parameter Value

Ts Signal length 2 (hours)
Th History length 7 (days)
V Number of vertical bins 32
H Number of horizontal bins 128
Vertical range of histogram [Bitdepth (t) – 100 m, Bitdepth (t) þ10 m]
TD_spd: [0, 300] rev/min
TD_trq: [0, 100] kNm
Hookload: [0, 500] ton
Horizontal range of histogram SPP_pressA: [0, 50] MPa

Table 2—Hyperparameter settings.

The 3D-CNN model accepts the 4D tensor (single channel) or 5D tensor (multichannel) as an input, and here we select multichannel
3D-CNNs to proceed the 5D tensor. The 3D-CNN applies convolution to an input tensor and extracts local features of the last three
dimensions. Thus, there is a degree of freedom in selecting three dimensions out of five. Ignoring the batch dimension, local feature
extraction is necessary for the V and H axes, but not required for the P and T axes, thus either B, P, T, V, H or B, T, P, V, H is expected
to work. Note that local feature extraction of T-axis will be necessary in general when the size of T is larger, but in this study, we use
only two snapshots in the time axis; signal and history, thus not required. For the convenience of multiparameter grad-CAM analysis

554 April 2021 SPE Journal

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 555 Total Pages: 12

(Selvaraju et al. 2017) that will be mentioned later, the shape of the tensor is transposed to B, T, P, V, H (Fig. 4). This is because the
feature dimension (the second dimension) will be squeezed into a single feature at the first layer, when using a multichannel CNN, by
the inner product. By setting P (drilling parameter dimension) to the third dimension, the information of respective drilling parameters
are kept unsqueezed until the final convolution layer; hence, grad-CAM analysis is available to respective drilling parameters.

SPP_pressA

p
TD_trq

t
Hookload
h
V

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

Input layer First-layer kernels

First-layer outputs

Signal
t

History P

h
V

Input layer First-layer kernels

First-layer outputs

Fig. 4—Transposition of 5D input tensor [upper: (B, P, T, V, H), lower: (B, T, P, V, H)].

To summarize, the 3D-CNN model learns the patterns of the 5D tensor in the P, V, H axis by using a 3D kernel. The model architec-
ture is presented in Table 3. Because local feature extraction of the P-axis is not required, the 3D kernel is substantially a 2D kernel.
The objective of using 3D-CNN is to enable multiparameter grad-CAM analysis instead of 3D pattern recognition.

Sublayers Hyperparameters Input Shape Output Shape

Dropout Probability ¼ 0.25 (16,2,4,32,128) (16,2,4,32,128)
Convolution-3D Input_features ¼ 2, (16,2,4,32,128) (16,8,4,28,96)
Output_features ¼ 8,
Kernel size ¼ (1,5,33),
Padding ¼ 0, dilation ¼ 1
1st layer ReLU (Activation function) (16,8,4,28,96) (16,8,4,28,96)

Dropout Probability ¼ 0.0 (16,8,4,28,96) (16,8,4,28,96)

Convolution-3D Input_features ¼ 8, (16,8,4,28,96) (16,16,4,24,64)
Output_features ¼ 16,
Kernel size ¼ (1,5,33),
Padding ¼ 0, dilation ¼ 1
2nd layer ReLU (Activation function) (16,16,4,24,64) (16,16,4,24,64)

Dropout Probability ¼ 0.5 (16,16,4,24,64) (16,16,4,24,64)

Linear (fully connected) Input_features ¼ 16  4  24  64 (16,98304) (16,1) (squeezed)
Output_features ¼ 1
3rd layer Sigmoid (Activation function) (16,1) (16,1)

Table 3—Model architecture. ReLU ¼ rectified linear unit.

April 2021 SPE Journal 555

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 556 Total Pages: 12

The codes are written in the open-source programming language Python 2.7, and PyTorch 1.0.0 (PyTorch 2020) is used to imple-
ment CNN. The information presented in this section (e.g., programming tools, data preprocessing method, hyperparameters, and
model architecture) specifies the proposed method and thus supports its reproducibility.
The implemented model is first trained with train data, and then applied to unknown clips, which is prepared as validation data and
is evaluated with the metrics presented in the next section.

Model Evaluation. When evaluating a binary classification model, a confusion matrix (Skiena 2017) is often used, especially when the
target class is hugely imbalanced. A confusion matrix is a cross-tabulation of prediction class and target class (as shown in Table 4).

Predicted Class
Negative Positive
Target class Negative TN FP
Positive FN TP

Table 4—Confusion matrix

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

Referring to the confusion matrix, the rate of correct prediction (accuracy) could be calculated by Eq. 1. Accuracy is a common
metric of binary classification model when the target classes are balanced.
TN þ TP
Accuracy ¼ ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð1Þ
FN þ TN þ FP þ TP
where TN is true negative, TP is true positive, FN is false negative, and FP is false positive. When the target classes are hugely imbal-
anced (e.g., P:N ¼ 1:r; 1  r; in this case approximately 1:10,000), which is the case with anomaly detection, there should be a larger
weight on TP compared to TN. In this case, other metrics that focus on TP, such as precision (Eq. 2) and recall (Eq. 3) are often used
instead of accuracy.
TP
Precision ¼ ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð2Þ
FP þ TP
TP
Recall ¼ : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð3Þ
FN þ TP
However, a CNN model does not directly predict the class. Instead, it predicts the probability for each class. In the case of this
study, the output of the model is a continuous scalar value in [0,1] representing the probability for the positive prediction. Therefore,
the prediction class, the confusion matrix, and the metrics previously presented are all dependent on the classification threshold.
Although the model could be evaluated using these metrics by setting a fixed threshold, it shows us only a single perspective of the
model. Instead, we adopt the area under the curve (AUC) (Skiena 2017) and average precision (AP) (Kekäläinen and Järvelin 2002),
which aggregate the performance along all such classification thresholds.
The AUC is an aggregation of recall and the false positive rate (Eq. 4) along all the classification thresholds defined by Eq. 5. By
integration, the model performance could be determined uniquely. The AUC is equivalent to the probability that the classifier will rank
a randomly chosen positive instance higher than a randomly chosen negative instance.
FP
False Positive Rate ¼ ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð4Þ
TN þ FP
X 1 X C1 X C0
AUC ¼ Recallðsk Þ  dFPRðsk Þ ¼ HðPj  Ni Þ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð5Þ
k
C0 C1 j¼1 i¼1

where FPR is the false positive rate, Ni is the prediction value for a negative sample, Pj is the prediction value for a positive sample, C0
is the total number of negative samples, C1 is the total number of positive samples, and H(x) is a step function defined by Eq. 6 for the
convenience of describing the “if” statement in programming.

1 ðx  0Þ
XðxÞ ¼
0 ðx < 0Þ                                                                 ð6Þ

AP is similar to AUC but is an aggregation of recall [¼ True Positive Rate (TPR)] and precision along all classification thresholds,
defined by Eq. 7.
X 1 X C1
j
AP ¼ Precisionðsk Þ  dRecallðsk Þ ¼ ; ....................................... ð7Þ
k
C1 j¼1 X
C0
HðPj  Ni Þ þ j
i¼1

where each symbol is shared with Eq. 5. AP is sensitive to the P:N ratio in the target class, whereas AUC is completely independent of
P:N ratio. Therefore, by using AUC, model performances could be compared even when they are applied to a different data set (e.g.,
cross-validation), but at the same time, it is permissive to low precision. AP is strict to low precision, but it is unsuitable when applying
the model to a different data set.
As mentioned previously, the reliability of label used for training is not sufficient. Thus, AUC and AP results were used to evaluate
the approximate performance of the CNN model. To evaluate the model performance in detail, the time series of the stuck risk indicator
was investigated in detail for each stuck case. Moreover, further interpretation methods are required to consider the reliability of the
model. In the next section, a method to visualize the basis of model prediction is introduced.

556 April 2021 SPE Journal

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 557 Total Pages: 12

Model Interpretation. Despite the ability of machine-learning techniques to extract information from various data, the lack of trans-
parency of the internal logic of the black box is a challenge (Teso and Kersting 2019). Deep-learning-based approaches, such as neural
networks, increase the model capability with a high degree of complexity. As a consequence, there is the trade-off between performance
and explainability (Gunning et al. 2019). An opaque model is also problematic when considering possible uncertainty and biases caused
by insufficient sample data or limitations of the adopted algorithms (Guidotti et al. 2019). Considering the actual drilling operation, the
early stuck detection model will be used as decision support system for drilling engineers. Thus, it is important that the cause of the
alarm is traceable by the drillers for them to make the decisions with trust.
Recently, the field of explainable artificial intelligence has obtained a lot of attention (Arrieta et al. 2020). There are several ways to
provide explanation. A description of how the black box works is desired but only possible for simplified models that sacrifice the accu-
racy of the machine-learning approach. A compromised approach is to explain the decisions without understanding the details on how
the opaque decision systems work in general. This is called the model-agnostic interpretation method (Ribeiro et al. 2016). If the early
stuck sign detection model can indicate which feature of the data set leads to the increase of stuck risk, the drilling engineers can specu-
late the situation based on the operation conditions and experience to make rational decisions.
Here, we adopted grad-CAM for CNN model interpretation. Previously, CAM was proposed to produce a visual explanation of the
CNN models (Zhou et al. 2016). Selvaraju et al. (2017) further developed grad-CAM, which is a technique for producing visual explan-
ations for decisions from a large class of CNN-based models, making them more transparent. Grad-CAM uses the gradients of any
target concept, flowing into the final convolutional layer to produce a coarse localization map highlighting important regions in the
image for predicting the concept. Grad-CAM is overlayed on the input data to highlight the data, which had high contribution to the
output. Grad-CAM is an aggregation of feature maps with low spatial resolution, so the contribution map may not be spatially accurate.

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

Results
Scores. The 3D-CNN prediction performance is first evaluated with AUC and AP. Twenty wells are used for training, and 10 wells are
used for validation, both from Company A. Our 3D-CNN model was trained with training data and applied to validation data. Scores
are presented in Tables 5 and 6. Because the performance of a prediction model should be evaluated by the prediction against the
unknown data, the most remarkable score is shown in the “validation, trained model” cell. The “train” row is complementary informa-
tion showing that the model has successfully learned the train data, and the columns on the right of the “trained model” is the baseline
scores made by poor prediction models.

Trained Before Training [0,1] Uniform

AUC P:N Model (five runs) All 0/1 (five runs)
Train 1:2 0.9969 0.5369 0.5 0.5052
Validation 1:10.5 0.6527 0.4958 0.5 0.4949

Table 5—Stuck prediction performance in depth domain (AUC).

Trained Before Training [0,1] Uniform

AP P:N Model (five runs) All 0/1 (five runs)
Train 1:2 0.9902 0.3744 0.3333 0.3330
Validation 1:10.5 0.3941 0.0845 0.0868 0.0878

Table 6—Stuck prediction performance in depth domain (AP).

Along with the scores, the raw prediction distributions for train data and validation data are presented in Fig. 5. It can be observed
that the model can distinguish each class clearly for the training data, and for the validation data, approximately 40% of positive sam-
ples were correctly predicted as positive.

Prediction Distribution for Train Data Prediction Distribution for Validation Data
(Positive:Negative = 1:2) (Positive:Negative = 1:10.5)
350
Negative Negative
Positive 5000 Positive
4000 300
Frequency (pos)

4000 250
Frequency (neg)
Frequency

3000
3000 200

2000 150
2000
100
1000 1000 50

0 0 0
0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
Prediction Prediction

Fig. 5—Prediction distribution for train data (left) and validation data (right).

Here, we compare the performance of the 3D-CNN model with a that of a typical approach, that is, a CNN model applied to a
n-hour time series of the drilling parameters. Such an approach is similar to previous literature, such as Siruvuri et al. (2006). The
results are presented in Table 7. The AUC result shows that the 3D-CNN (AUC 0.6527) outperforms time-domain CNN (AUC 0.475).
Note that AUC is a metric independent of the P:N ratio and provides fair comparison.

April 2021 SPE Journal 557

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 558 Total Pages: 12

Trained Before Training [0,1] Uniform

AUC P:N Model (five runs) All 0/1 (five runs)
Train 1:2 0.980 0.5378 0.5 0.5015
Validation 1:11 0.475 0.5048 0.5 0.4935

Table 7—Stuck prediction performance in time domain (AUC).

Stuck Scenarios. Because it is impossible to define when the early sign of stuck pipe is evident, the normal classification metric is not
capable of evaluating the model performance. We conducted case studies to see how the model responds to each stuck event to see the
ratio of “true positive (alerts when needed)” and “false negative (no alerts when needed).” In addition, we randomly selected eight
cases of 2-hour data sets in which we assume the operation is normal to see occurrence of “false positive (alerts when not needed).”
There are various mechanisms of stuck occurrence. Table 8 depicts the stuck mechanism for the six cases used for validation. These
stuck mechanisms were extracted from daily drilling reports.

Stuck Frequency Depth

Mechanism 1D-CNN 3D-CNN

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

Case 1 Packoff x x
Case 2 Packoff x x
Case 3 Differential x o
Case 4 Unknown x o
Case 5 Unknown x x
Case 6 Unknown o o

Table 8—Case studies in validation data.

The depth-domain approach detected early signs for three cases out of six. Note that some stuck events may be sudden. In these
cases, it is difficult, if not impossible, to detect early signs. Thus, we cannot judge the performance just from the ratio of success.
The model did not give stuck risk higher than 0.5 in any of the randomly selected cases. This suggests the model was free of “false
positive (alerts when not needed)” for the tested cases. Because the actual performance of early stuck pipe sign detection can be tested
for stuck events, a further validation study is desired.

Case Study. For Case 3, real-time prediction of 1D-CNN and 3D-CNN is presented in Fig. 6. In this case, 1D-CNN could not detect
signs, whereas 3D-CNN does. In the output of 3D-CNN, the stuck risk has been high for 30 minutes before the occurrence of stuck (1).
The 3D-CNN was catching signs of stuck at the time of discharge (2), and was in the process of discharge at the time of stuck (3). From
the figure, it is indicated that early stuck signs were appearing from earlier than an hour before the occurrence.
Looking into a stuck case in the training data (Fig. 7), it could be observed that the stuck risk is increasing from 23:00, an hour
before the stuck, even the data before 23:54 was not used for training. The result indicates that the proposed model is successfully learn-
ing the common patterns of early stuck sign, instead of just memorizing.
By comparing input data and grad-CAM where the alarm was fired, we could understand that the model has an intense reaction on
hookload (Figs. 8 and 9). In this case, stuck happened during RIH, thus the alarm reacting on hookload does make sense according to
driller’s intuition, possibly indicating the increase of breakover weight.
From the result and case studies, we expect further possibilities in the depth-domain approach. It is considered that the depth
domain was effective for the differential stuck that gradually shows signs in the depth domain. However, feature engineering is required
for stuck where features appear in other areas. Also, further improvements are expected by parameter tuning such as horizontal range,
resolution of 2D histogram, and length of time history information to be included.

Performance of the Depth-Domain Approach. The improvement of depth-domain 3D-CNN performance compared to time-domain
and frequency-domain 1D-CNN is clearly due to the data clip representation. This suggests the features are “well captured” or “easier
to learn” with the depth-domain representation. The amount of information conveyed in the depth-domain data clip increases by includ-
ing all the data for nearby depth in the same well but decreases by abstraction of the time-domain information to just current “signal”
and past “history.” We assume the explicit comparison of “signal” and “history” in the depth domain has effectively highlighted the fea-
tures in the data that indicate early stuck signs. The depth-domain representation was inspired by discussion with experienced drilling
engineers suggesting the importance of feature engineering with domain expertise.
It is not trivial to evaluate the reliability of the learned model. As previously mentioned, there may be cases in which an early sign did
not exist due to the abrupt change of situation. Therefore, a flawless detection of an early sign would be impossible. However, we assume
stuck caused by differential pressure to be unique from others. For the drilling engineers, stuck caused by differential pressure occurs as a
sudden event. However, the process of drillpipe being embedded in the mudcake does not occur at once, that is, a certain lead time exists.
The increase in difference between the hydrostatic and pore pressures, thick mudcake development, and the pressure drop in the mudcake
occur gradually, and finally results in the stuck (Rehm et al. 2013). Breakover weight or breakover torque are known to be early signs of
differential stuck. Unlike the other stuck mechanisms, we assume that there is sufficient lead time for differential stuck in which the
observed drilling parameters may show features of stuck risk increasing. In this sense, we acknowledge the fact that the depth-domain
3D-CNN model being successful in detecting the early signs for this case as an indication of good performance.
From the result and case studies, we expect further possibilities in the depth-domain approach. It is considered that the depth
domain was effective for the differential stuck that gradually shows signs in the depth domain. However, feature engineering is required
for stuck in which features appear in other areas. Also, further improvements are expected by parameter tuning such as horizontal
range, resolution of 2D histogram, and length of time history information to be included.

558 April 2021 SPE Journal

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 559 Total Pages: 12

(kN/m)(rev/min)
200
TD_spd
0
30
20 TD_trq
10

1050 3
(m)
Bitdepth
1025
100
(ton)
Hookload
50
(mPa)

10 SPP_pressA
0
Totdepth
(m)

4500
4250
1
Pred_1D
()

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

0
1
Pred_3D 1
()

0
0 0 0 0 0 0 0
1:0 :3 :0 :3 :0 :3 :0
1 11 12 12 13 13 14

0.8
Bitdepth
1065 2 0.7
Pred_3D
0.6
1060
Bitdepth (m)

0.5

Pred_3D
1055 0.4
0.3
1050
0.2
1045 0.1

1040 0
30 35 40 45 50 55 00
1: 1: 1: 1: 1: 1: 2:
9:1 9:1 9:1 9:1 9:1 9:1 9:1
2 2 2 2 2 2 2

Fig. 6—Real-time stuck risk prediction with validation data (Case 3).
(rev/
(kN/m) min)

50 TD_spd

0
5 TD_trq

0
4000
(m)

Bitdepth
2000
150
(mPa) (ton)

Hookload
100
50

10 SPP_pressA

0
4750 Totdepth
(m)

4500
1
Pred
3D-CNN output
()

0
0 0 0 0 0 0 0 0 0
:0 :0 :0 :0 :0 :0 :0 :0 :0 e
16 17 18 19 20 21 22 23 00 Jun
-
06

Fig. 7—Real-time stuck risk prediction with training data.

As Decision Support Tool for Engineers. The developed model is considered to be applicable to various drilling operations because
the parameters used in this study are general. The model only neglects measurement outliers so that the model is robust to calibration
errors. In the training data and validation data, data from various drilling operations are used. Here, we have excluded parameters that
specify the aforementioned drilling operation. Because we were able to predict some of the stuck events, we understand that our goal is
somewhat achieved. However, the specification of this background information has the potential to increase prediction accuracy,
although that would require a much larger data set for training the model.

April 2021 SPE Journal 559

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 560 Total Pages: 12

0 0

5 5

Pixelized Depth (unitless)

Pixelized Depth (unitless)

10 10
0.0004
15 15

20 20

25 25
0.0003

30 30
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Pixelized TD_spd (unitless) Pixelized TD_trq (unitless)

0 0

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

0.0002

5 5
Pixelized Depth (unitless)

Pixelized Depth (unitless)

10 10

15 15 0.0001

20 20

25 25

30 30 0

0 20 40 60 80 100 120 0 20 40 60 80 100 120

Pixelized Hookload (unitless) Pixelized SPP_pressA (unitless)

Fig. 8—Grad-CAM showing reaction to the input data. The color code scale is nondimensional grad-CAM magnitude, which repre-
sents the degree of contribution to the positive class activation.

0 0 0 0
Pixelized Depth (unitless)

Pixelized Depth (unitless)

Pixelized Depth (unitless)

Pixelized Depth (unitless)

5 5 5 5

10 10 10 10

15 15 15 15

20 20 20 20

25 25 25 25

30 30 30 30
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120
Pixelized TD_spd (unitless) Pixelized TD_trq (unitless) Pixelized Hookload (unitless) Pixelized SPP_pressA (unitless)
0 0 0 0
Pixelized Depth (unitless)

Pixelized Depth (unitless)

Pixelized Depth (unitless)

Pixelized Depth (unitless)

5 5 5 5

10 10 10 10

15 15 15 15

20 20 20 20

25 25 25 25

30 30 30 30
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120
Pixelized TD_spd (unitless) Pixelized TD_trq (unitless) Pixelized Hookload (unitless) Pixelized SPP_pressA (unitless)

Fig. 9—Grad-CAM compared with input depth clip. In the upper row, green data represents history and red data represents signal.

The model can be applied to every timestep the drilling data are obtained. An average workstation with an IntelV R XeonV R E5-2603

v4 @ 1.70 GHz (Intel Corporation, Santa Clara, California, USA) central processing unit and a GeForceV R GTX 1080 Ti (NVIDIA Cor-

poration, Santa Clara, California, USA) graphics card could process the data in 0.303 seconds on average. This includes the process of
producing depth-domain clips and running the model. Because the sampling rate of drilling data is usually less than 1 Hz, we consider
the method to be applicable in real-time operation.
It is the drilling engineer who makes the final decision on how to operate. The purpose of an early warning alarm is to remind the drill-
ing engineer (Zhu et al. 2019). The alarm generated by our model is valuable because it provides a new point of view to detect early signs

560 April 2021 SPE Journal

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 561 Total Pages: 12

that drilling engineers may not be aware of. Thus, the drilling engineers can have an adequate time for operational decisions in advance.
The reliability of the machine-learning model is not and will never be perfect, even though the model is continuously updated. Therefore,
we focus to make our model explainable or understandable for the drilling engineers to decide whether to trust the model output or not.
In the design of our detection model, we applied a model-agnostic interpretation method. The portion of data, that is, when and
which parameter, contributing to the model response is highlighted with grad-CAM methodology. Case studies show that the alarm
reacting on a specific portion of the data provides intuitive understanding of possible phenomena occurring in the well. Such informa-
tion supports the engineers to make rational decisions.

Unsupervised Learning. We also see limitations in supervised approach. The model learns to detect the early signs of stuck from
training data, where the data set of the last 6 minutes before stuck occurrence is labeled to be positive. The duration of 6 minutes is an
arbitrary choice because the drilling engineers do not know where the early signal actually starts. Six minutes is a relatively short dura-
tion to avoid contamination of negative data. As shown in Case 3, stuck signs may appear a few hours before the stuck occurrence, and
it is difficult to identify and label all stuck signs.
It is worth noting that the model was successful with such rough estimation. However, as previously mentioned, it is not apparent
whether an early sign existed or not in general. This means the accuracy of labeling early sign is always uncertain. In addition, the data
set for learning positive cases are limited to the available stuck cases.
One way to avoid the shortcomings of supervised learning is to just focus on the negative data set. Anomaly detection with unsuper-
vised learning identifies the data that are different from the majority of the data set, that is, the negative data set. This can be imple-

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

mented by various methods, such as autoencoders. This can avoid the ambiguity of data labeling and also build the model on abundant
negative data sets.

Conclusions
In this study, we propose an early stuck sign-detection model based on CNN with unique depth-domain data clips. Field well data are
provided by oil and gas companies and 30 well data from a single company is used in this work. Data are split into 20 training wells
and 10 validation wells. By applying the depth-domain 3D-CNN model to the validation well data in real time, the model has success-
fully detected early signs of stuck in three cases out of six stuck incidents, before the occurrence. In addition, we did not observe “false
positive” alerts in the randomly selected validation well where stuck pipe incidents did not occur. The result indicates the possibility of
the machine-learning model to detect early stuck signs in real time.
Field drilling data suggests that depth-domain representation of data clips improves performance of early stuck detection by includ-
ing a longer duration of data with adequate abstraction. The model identifies the portion of data that contributed to the anomaly detec-
tion, providing insight to the black-box model behavior for the drilling engineers.
The basic concepts of these models were inspired by drilling experts, suggesting the need for cocreation of machine learning using
domain expertise. Further improvement of the model also needs to incorporate experience of engineers. The first step will be a demon-
stration for the drilling engineers to get familiar with the detection model and grad-CAM illustration. Such interpretation can be further
extended by observing the time variation of the risk indicator or the grad-CAM results. Based on the findings of this research, we are
now considering further improvement by fusing unsupervised learning and the depth-domain approach.

Acknowledgments
This research project was financially supported by JOGMEC (Japan Oil, Gas and Metals National Corporation). We would like to thank
project members (i.e., oil/gas companies and research institutions) for providing data and valuable advice.

References
Abbas, A. K., Flori, R., Almubarak, H. et al. 2019. Intelligent Prediction of Stuck Pipe Remediation Using Machine Learning Algorithms. Paper pre-
sented at the SPE Annual Technical Conference and Exhibition, Calgary, Alberta, Canada, 30 September–2 October. SPE-196229-MS. https://
doi.org/10.2118/196229-MS.
Agwu, O. E., Akpabio, J. U., Alabi, S. B. et al. 2018. Artificial Intelligence Techniques and Their Applications in Drilling Fluid Engineering: A Review.
J Pet Sci Eng 167: 300–315. https://doi.org/10.1016/j.petrol.2018.04.019.
Alshaikh, A., Magana-Mora, A., Al Gharbi, S. et al. 2019. Machine Learning for Detecting Stuck Pipe Incidents: Data Analytics and Models Evaluation.
Paper presented at the International Petroleum Technology Conference, Beijing, China, 26–28 March. IPTC-19394-MS. https://doi.org/10.2523/
IPTC-19394-MS.
Alshaikh, A. A., Albassam, M. K., Al Gharbi, S. H. et al. 2018. Detection of Stuck Pipe Early Signs and the Way Toward Automation. Paper presented
at the Abu Dhabi International Petroleum Exhibition and Conference, Abu Dhabi, UAE, 12–15 November. SPE-192975-MS. https://doi.org/10.2118/
192975-MS.
Arrieta, A. B., Dı́az-Rodrı́guez, N., Del Ser, J. et al. 2020. Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Chal-
lenges Toward Responsible AI. Inf Fus 58: 82–115. https://doi.org/10.1016/j.inffus.2019.12.012.
Awad, M. and Khanna, R. 2015. Efficient Learning Machines: Theories, Concepts, and Applications for Engineers and System Designers. Berkeley,
California, USA: Apress.
Guidotti, R., Monreale, A., Ruggieri, S. et al. 2019. A Survey of Methods for Explaining Black Box Models. ACM Comput Surv 51 (5): 1–42. https://
doi.org/10.1145/3236009.
Gunning, D., Stefik, M., Choi, J. et al. 2019. XAI—Explainable Artificial Intelligence. Sci Rob 4 (37): eaay7120. https://doi.org/10.1126/
scirobotics.aay7120.
Heinze, L. and Al-Baiyat, I. A. 2012. Implementing Artificial Neural Networks and Support Vector Machines in Stuck Pipe Prediction. Paper presented
at the SPE Kuwait International Petroleum Conference and Exhibition, Kuwait City, Kuwait, 10–12 December. SPE-163370-MS. https://doi.org/
10.2118/163370-MS.
Howard, J. A. and Glover, S. B. et al. 1994. Tracking Stuck Pipe Probability while Drilling. Paper presented at the IADC/SPE Drilling Conference,
Dallas, Texas, USA, 15–18 February. SPE-27528-MS. https://doi.org/10.2118/27528-MS.
Jahanbakhshi, R., Keshavarzi, R., Aliyari Shoorehdeli, M. et al. 2012. Intelligent Prediction of Differential Pipe Sticking by Support Vector Machine
Compared with Conventional Artificial Neural Networks: An Example of Iranian Offshore Oil Fields. SPE Drill & Compl 27 (4): 586–595. SPE-
163062-PA. https://doi.org/10.2118/163062-PA.

April 2021 SPE Journal 561

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178

 DOI: 10.2118/204462-PA Date: 23-March-21 Stage: Page: 562 Total Pages: 12

Kekäläinen, J. and Järvelin, K. 2002. Using Graded Relevance Assessments in IR Evaluation. J Am Soc Inf Sci Technol 53 (13): 1120–1129. https://
doi.org/10.1002/asi.10137.
Kingma, D. and Ba, J. 2015. Adam: A Method for Stochastic Optimization. In Proc., 3rd International Conference for Learning Representations
(ICLR’15). San Diego, California, USA, 7–9 May.
LeCun, Y. and Bengio, Y. 1995. Convolutional Networks for Images, Speech, and Time Series. In The Handbook of Brain Theory and Neural Networks,
ed M. A. Arbib. Cambridge, Massachusetts, USA: MIT Press.
Lind, Y. B. and Kabirova, A. R. 2014. Artificial Neural Networks in Drilling Troubles Prediction (Russian). Paper presented at the SPE Russian Oil and
Gas Exploration and Production Technical Conference and Exhibition, Moscow, Russia, 14–16 October. SPE-171274-RU. https://doi.org/10.2118/
171274-RU.
Magana-Mora, A., Gharbi, S., Alshaikh, A. et al. 2019. AccuPipePred: A Framework for the Accurate and Early Detection of Stuck Pipe for Real-Time
Drilling Operations. Paper presented at the SPE Middle East Oil and Gas Show and Conference, Manama, Bahrain, 18–21 March. SPE-194980-MS.
https://doi.org/10.2118/194980-MS.
PyTorch, Version 1.0.0. 2020. https://pytorch.org/ (accessed 2 September 2020).
Rehm, B., Schubert, J., Haghshenas, A. et al. 2013. Managed Pressure Drilling, first edition. Houston, Texas, USA: Gulf Publishing Company.
Ribeiro, M. T., Singh, S., and Guestrin, C. 2016. Model-Agnostic Interpretability of Machine Learning. In Proc., International Conference on Machine
Learning Workshop on Human Interpretability in Machine Learning, New York, New York, USA, 23 June.
Salminen, K., Cheatham, C., Smith, M. et al. 2017. Stuck-Pipe Prediction by Use of Automated Real-Time Modeling and Data Analysis. SPE Drill &
Compl 32 (3): 184–193. SPE-178888-PA. https://doi.org/10.2118/178888-PA.
Selvaraju, R. R., Cogswell, M., Das, A. et al. 2017. Grad-Cam: Visual Explanations from Deep Networks via Gradient-Based Localization. Paper pre-

Downloaded from http://onepetro.org/SJ/article-pdf/26/02/551/2431588/spe-204462-pa.pdf by U. of Alberta Library user on 15 September 2023

sented at the IEEE International Conference on Computer Vision (ICCV), Venice, Italy, 22–29 October. https://doi.org/10.1109/ICCV.2017.74.
Siruvuri, C., Nagarakanti, S., and Samuel, R. 2006. Stuck Pipe Prediction and Avoidance: A Convolutional Neural Network Approach. Paper presented
at the IADC/SPE Drilling Conference, Miami, Florida, USA, 21–23 February. SPE-98378-MS. https://doi.org/10.2118/98378-MS.
Skiena, S. S. 2017. The Data Science Design Manual. New York, New York, USA: Springer.
Teso, S. and Kersting, K. 2019. Explanatory Interactive Machine Learning. Paper presented at the AIES AAAI/ACM Conference on AI, Ethics, and
Society, Honolulu, Hawaii, USA, 27–28 January. https://doi.org/10.1145/3306618.3314293.
Unrau, S., Torrione, P., Hibbard, M. et al. 2017. Machine Learning Algorithms Applied to Detection of Well Control Events. Paper presented at the SPE
Kingdom of Saudi Arabia Annual Technical Symposium and Exhibition, Dammam, Saudi Arabia, 24–27 April. SPE-188104-MS. https://doi.org/
10.2118/188104-MS.
Zhou, B., Khosla, A., Lapedriza, A. et al. 2016. Learning Deep Features for Discriminative Localization. Paper presented at the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), Las Vegas, Nevada, USA, 26 June–1 July. https://doi.org/10.1109/CVPR.2016.319.
Zhu, Q., Wang, Z., and Huang, J. 2019. Stuck Pipe Incidents Prediction Based on Data Analysis. Paper presented at the SPE Gas and Oil Technology
Showcase and Conference, Dubai, UAE, 21–23 October. SPE-198672-MS. https://doi.org/10.2118/198672-MS.

562 April 2021 SPE Journal

ID: jaganm Time: 20:33 I Path: S:/J###/Vol00000/200178/Comp/APPFile/SA-SPE-J###200178