SPE-205677-MS

An Unsupervised Learning Model for Pipe Stuck Predictions Using a Long

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/1-21APOG/D011S009R006/2496805/spe-205677-ms.pdf by U. of Alberta Library user on 15 September 2023

Short-Term Memory Autoencoder Architecture

Yujin Nakagawa and Tomoya Inoue, Japan Agency for Marine-Earth Science and Technology; Hakan Bilen,
University of Edinburgh; Konda R. Mopuri, Indian Institute of Technology; Keisuke Miyoshi and Shungo Abe3,
Japan Oil, Gas and Metals National Corporation; Ryota Wada, University of Tokyo; Kouhei Kuroda, Japan
Petroleum Exploration Co., Ltd; Hitoshi Tamamura, INPEX corporation

Copyright 2021, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE/IATMI Asia Pacific Oil & Gas Conference and Exhibition held virtually on 12 - 14 October, 2021. The official
proceedings were published online on 4 October, 2021.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Pipe-sticking during drilling operations causes severe difficulties, including economic losses and safety
issues. Therefore, stuck-pipe predictions are an important tool to preempt this problem and avoid the
aforementioned troubles. In this study, we have developed a prediction technique based on artificial
intelligence, in collaboration with industry, the government, and academia. This technique was an
unsupervised learning model built using an encoder-decoder, long short-term memory architecture. The
model was trained with the time series data of normal drilling operations and based on an important
hypothesis: reconstruction errors between observed and predicted values are higher around the time of pipe
sticking than during normal drilling operations. The trained model was then applied to 34 actual stuck-
pipe events, where it was found that reconstruction errors increased prior to the pipe sticking in some cases
(thereby partly confirming our hypothesis) and were sensitive to large variations in the drilling parameters.

Introduction
The occurrence of pipe-sticking during drilling operations leads to extreme delays in the completion of
the operation, well abandonment and, in the worst case, an uncontrolled spout of crude oil or gas into the
surroundings. These troubles are unexpected and cause economic losses and safety issues. An important
reason behind these troubles is the difficulty in the antecedent collection of information about the lithological
characteristics of the wells. Of the techniques used to reduce and avoid these troubles, two approaches are
eminent: one is the detection of stuck pipes as soon as possible after the pipe-sticking has started (hereafter
referred to as stuck-pipe detections), while the other is the prediction of pipe-sticking before it has occurred
(hereafter referred to as stuck-pipe predictions). The latter approach is more appropriate because it could
give crews sufficient time to take preventative or remedial actions. Therefore, stuck-pipe predictions that
use mud logging data that reflect the conditions inside the wells are an important step taken to avoid or
reduce the troubles associated with pipe-sticking.
2 SPE-205677-MS

Several ideas have been proposed to make stuck-pipe predictions using mud logging data. One involves
the use of statistical and/or analytical models (Hess 2016; Magana-Mora et al. 2019; Meor Hashim et
al. 2021b; Salminen et al. 2017; Shoraka et al. 2011), while another employs a combination of physics-
based and data-driven models (Zhang et al. 2019). Ideas based on supervised machine learning have been
proposed using the architecture of neural networks and/or support vector machines (Abbas et al. 2019;
Ahmed et al. 2019; Al-Baiyat et al. 2012; Alshaikh et al. 2019; Brankovic et al. 2020; Chamkalani et al.
2013; Meor Hashim et al. 2021a; Siruvuri et al. 2006; Zhu et al. 2019). In addition, prediction methods based

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/1-21APOG/D011S009R006/2496805/spe-205677-ms.pdf by U. of Alberta Library user on 15 September 2023

on unsupervised machine learning have also been put forth, including a three-dimensional convolutional
neural network with depth-domain data (Tsuchihashi et al. 2021) and a mixture probability model with fully-
connected neural networks (Inoue et al. 2021).
To make stuck-pipe predictions using artificial intelligence, we embarked on a project in collaboration
with industry, the government, and academia (Inoue et al. 2020). Most existing works of research into
stuck-pipe predictions are conducted using models based on supervised machine learning. The time-tagged
mud logging data (hereafter referred to as the time series data) of the actual drilling operations in various
wells, which were obtained by the collaborating companies, were used to make stuck-pipe predictions
based on a binary classification model involving supervised machine learning. This model requires negative
and positive labels, which indicate the time periods during normal operation and before the start of pipe-
sticking, respectively; these labels were mapped to the time series data. The performance of this model is,
however, inadequate in the following ways. The first flaw is the insufficient accuracy of the labels given to
the time series data; this is caused by a difficulty in defining the time periods of the positive labels since
the time point corresponding to the exact start of pipe-sticking is unknown. The second deficiency is the
significant imbalance in the numbers of negative and positive labels; this arises from the fact that the pipe-
sticking scenario required for positive labels is a rare event relative to the time periods of normal drilling
operations. Models based on supervised machine learning, which requires accurate labeling, are therefore
unsuitable for stuck-pipe predictions. In contrast, models based on unsupervised machine learning can be
trained solely with the time series data of normal drilling operations and are hence more suited to make
stuck-pipe predictions. In this study, we developed an unsupervised learning model as mentioned above,
to make stuck-pipe predictions. The novelty of our research is the development of an LSTM-AE model
based on unsupervised machine learning to make stuck-pipe predictions which can be applied to various
types of wells. We believe that our model has the possibility of making stuck-pipe predictions that cannot
be achieved by supervised learning approaches.

Methods
The fundamental design of our unsupervised learning model was based on two concepts. The first concept
is that an unsupervised learning model is trained only with the time series data of normal drilling operations.
The second concept is the hypothesis that the differences between the actual and predicted values of the
time series data increase around the starting of pipe sticking, when compared to the time of normal drilling
operations. If this hypothesis could be proven to be correct, stuck-pipe predictions can be certainly made
before the start of pipe-sticking.

Data
This study was performed using the time series data of actual drilling operations in 34 distinct wells,
obtained by the collaborating companies. The time series data spans 3,382 days and has a resolution of 4 s.
To remove anomalous, possibly inaccurate values, the data were filtered so that all 13 drilling parameters
simultaneously satisfied the minimum and maximum criteria presented in Table 1. The data were then
normalized by a min-max scaler whose variables were calculated for each drilling parameter. As a result, we
SPE-205677-MS 3

found 76,278,319 datasets, where a single dataset consisted of 150 continuous time bins (i.e., 600 seconds)
and 13 drilling parameters.

Table 1—Summary of Drilling Parameters Used in This Study.

Parameter Name Description Unit Minimum Value Maximum Value

TD_spd Top drive rotation speed rpm −300 300

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/1-21APOG/D011S009R006/2496805/spe-205677-ms.pdf by U. of Alberta Library user on 15 September 2023

TD_trq Top drive torque kNm −100 100
ROP_ave Average rate of penetration m/h −9000 9000
Bitdepth Bit depth m −50 9000
Totdepth Total hole depth m −50 9000
Hookheight Hook height m −10 100
Hookload Hook load ton −20 500
Weight on bit (calculated from
WOB kN – –
hook load)
MRetFlow Return flow rate of mud % −30 100
SPP_pressA Pressure of stand-pipe manifold A MPa −5 50
MPP_SPM1 Stroke of mud pump #1 spm −1 160
MPP_SPM2 Stroke of mud pump #2 spm −1 160
FlowIn Flow from the mud pumps L/min −100 6000

As described in the "Introduction" section, the long short-term memory autoencoder (LSTM-AE) model
was designed to be trained with the time series data of normal drilling operations. These data points were
obtained by filtering the complete dataset using two boundaries that indicated normal drilling operations:
360 s before the start of pipe-sticking and 3,600 s after its end. As a result, we obtained 47,667,472 datasets
corresponding to the normal drilling operations of all 34 wells, among which we extracted 38,394,875
datasets pertaining to 24 of the wells for training the model.
To evaluate the performance of the trained LSTM-AE model, we defined the time periods corresponding
to the normal operation of the drill and to the detection and prediction of pipe-sticking by using the daily
drilling reports obtained from the collaborating companies as follows.
Labels of detector time periods indicate the duration between the start and end of pipe-sticking (hereafter
referred to as detector labels). We identified the detector time periods of 34 stuck-pipe events using the daily
drilling report. Labels of predictor time periods, on the other hand, indicate the duration from 360 to 60 s
before the start of the pipe-sticking event (hereafter referred to as predictor labels). In addition, labels of
the normal operating time periods were categorized into two types of durations — one before the predictor
time period and one after the detector time period. In this study, the first type was defined as 360 s before
the start of pipe-sticking, while the second type was defined as 3,600 s after the end of pipe-sticking.
Based on the above three aspects, detector, predictor, and normal operation labels were assigned to each
time bin of the datasets and used to evaluate the performances of stuck-pipe detections, predictions, and both
detections and predictions, respectively. It should be noted here that the performances are highly dependent
on the definitions of the time periods of normal operation, predictor, and detection.
Test datasets were then extracted from all the datasets for all 34 stuck-pipe events. The definition
of negative and positive labels was necessary at this point to evaluate the performances of the stuck-
pipe predictions and detections. To evaluate the performance of stuck-pipe detections, the time periods
corresponding to normal operation and detection were taken as the negative and positive labels. Similarly,
the durations corresponding to normal operation and prediction were considered as the negative and positive
labels to evaluate the performance of stuck-pipe predictions.
4 SPE-205677-MS

Model
We developed an autoencoder model using an encoder-decoder LSTM architecture — a feedback neural
network that is more suitable for time series predictions than standard feedforward neutral networks. The
LSTM-AE model was intended to reconstruct the input data consisting of the 13 drilling parameters by first
encoding them and then decoding them back to their original values to make the output data consistent with
input data. A major detail of this model is that the dimensions of input data were narrowed down into a
latent layer during the encoding stage of the process. Therefore, the appropriate extraction of features from

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/1-21APOG/D011S009R006/2496805/spe-205677-ms.pdf by U. of Alberta Library user on 15 September 2023

the latent layer is required for accurate reconstructions of input and output data. If the LSTM-AE model
were trained solely with the time series data from normal drilling operations, differences between the input
and output data would be expected to be small. Prior to and during pipe-sticking, however, these differences
were expected to increase. Thus, reconstruction errors were used as indicators of the stuck-pipe predictions
and detections.
The input and output shapes of the LSTM-AE model are represented by (150, 13), consistent with the
shape of a single dataset consisting of 150 time bins and the 13 drilling parameters defined in the "Data"
section. Here, we used 150 time bins (i.e., 600 seconds) by assuming the values of the drilling parameters to
be affected by the time-based variations during the 600 seconds under consideration. Following discussions
with the collaborating companies, we also chose the 13 drilling parameters (Table 1) that were likely to
have the most influence on pipe-sticking. For simplicity, we adopted only a single latent layer. The choice
of suitable latent dimensions is highly dependent on the training datasets used. Therefore, considering the
(150, 13) shapes of the input and output data, a latent dimension of 128 was adopted. Finally, an Adam
optimizer (Diederik et al. 2015; Sashank et al. 2018) was employed with a learning rate of 0.01. The LSTM-
AE model was then, overall, trained by minimizing the reconstruction errors calculated with a mean squared
error function, which is defined as:

(1)

where n is the number of time bins and xi and denote vectors of true and predicted values, respectively.
Since both these vectors consist of the 13 drilling parameters; therefore, L is a vector of their averaged
reconstruction errors.

Model Learning
Using the training datasets, parameters of the LSTM-AE model were updated through 30,000 iterations. A
mini-batch of one iteration consisted of 50 randomly chosen datasets from the larger superset of training
data. To avoid overfitting, we utilized a function to halt the training process if the reconstruction errors
did not improve even after a certain number of iterations. Upon implementation, this function stopped
the training after 16,000 iterations. Fig. 1 shows the averaged reconstruction error calculated by the mean
squared error function over each 500 iterations; these errors decreased with progressive iterations.
SPE-205677-MS 5

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/1-21APOG/D011S009R006/2496805/spe-205677-ms.pdf by U. of Alberta Library user on 15 September 2023

Figure 1—The variation of averaged reconstruction error with the number of iterations.

Case Studies
Statistics of Reconstruction Errors and Area Under Cover
Using the test datasets for 6 stuck-pipe events from 10 wells which were not used in the training process,
reconstruction errors were calculated for the detector, predictor, and normal operating time periods defined
in the "Data" section. In Fig. 2, the left panel shows distributions of the reconstruction errors for the stuck-
pipe predictions, where histograms labeled as negative and positive indicate the normal operation and
predictor time periods, respectively. On the same note, the right panel in Fig. 2 shows distributions of the
reconstruction errors for stuck-pipe detections, where histograms labeled as negative and positive indicate
the normal operation and detector time periods, respectively.

Figure 2—Left: Distributions of the reconstruction error for stuck-pipe predictions. Histograms labeled
as negative and positive indicate the normal operation and predictor time periods, respectively.
Right: Distributions of the reconstruction error for stuck-pipe detections. Histograms labeled
as negative and positive indicate the normal operation and detector time periods, respectively.
6 SPE-205677-MS

To evaluate the trained LSTM-AE model, area under cover (AUC), a performance indicator, was
calculated for each stuck-pipe event using the receiver operating characteristic (ROC) curve of the above
reconstruction errors, which quantifies the performance of a classification model at all classification
thresholds. The ROC curve depicts the relation between the true positive rate (RTPR) and the false positive
rate (RFPR), which are defined as:

(2)

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/1-21APOG/D011S009R006/2496805/spe-205677-ms.pdf by U. of Alberta Library user on 15 September 2023

and

(3)

where NTP, NFN, NFP, and NTNdenote the numbers of results where the model correctly predicted positive
labels, incorrectly predicted negative labels, incorrectly predicted positive labels, and correctly predicted
negative labels, respectively. The AUC is then defined as the complete two-dimensional area underneath
the entire ROC curve. Using the above reconstruction errors, AUC values were calculated for the 6 stuck-
pipe events: 0.72 and 0.78 for the predictor and detector time periods, respectively.

Time Variations of Reconstruction Errors

To systematically evaluate the LSTM-AE model, reconstruction errors of each drilling parameter were
calculated for the test datasets representative of all 34 stuck-pipe events. Predictions were made by the
trained LSTM-AE model for each of the 150 continuous time bins by individually shifting through them.
This was followed by the calculation of reconstruction errors for each time bin and averaging them out over
all 150 time bins for each drilling parameter. As a result, we obtained 13 reconstruction errors for every
dataset. Fig. 3 shows an example of the time variations of the reconstruction errors averaged over 13 drilling
parameters (top) and the contour map of the reconstruction errors for each drilling parameter (bottom). The
gray-hatched areas in Fig. 3 indicate the time periods corresponding to pipe-sticking.
SPE-205677-MS 7

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/1-21APOG/D011S009R006/2496805/spe-205677-ms.pdf by U. of Alberta Library user on 15 September 2023

Figure 3—Time variations of reconstruction errors averaged over 13 drilling parameters
(top) and a contour map of the reconstruction errors for each drilling parameter (bottom).
The gray-hatched areas indicate the time periods corresponding to pipe-sticking.

Discussion
Performance Evaluation Using AUC
There seems to be a little difference in the distributions of the reconstruction errors between the negative
labels and the positive labels for both the predictor and detector time periods (Fig. 2). The reconstruction
errors for the positive labels seems to be greater than those for the negative labels for both the predictor and
detector time periods. The reason of the unclear difference is presumably because the distributions of the
reconstruction errors (i.e., Fig. 2) were summarized for all 6 stuck-pipe events with different performances.
On the other hand, marginally good performances are given by the AUC values of 0.72 and 0.78 for the
predictor and detector time periods, respectively. These imply the capability of the trained LSTM-AE model
to detect and predict certain types of stuck-pipe events. Further investigations into specific types of stuck-
pipe events and their AUC values are required for a more detailed evaluation of the performances of the
trained LSTM-AE model.

Performance Evaluation Using Time Variation of Reconstruction Errors

In some cases, reconstruction errors increased prior to the pipe-sticking event (Fig. 3). Among all 13 drilling
parameters, the reconstruction errors of MPP_SPM1 and MPP_SPM2 (depicted by the bottom panel in Fig.
3) were particularly high. In addition, the reconstruction errors of SPP_pressA and FlowIn affected the
reconstruction errors averaged over 13 drilling parameters (indicated by the top panel in Fig. 3). This might
8 SPE-205677-MS

be caused by lost circulation which occurred during the time period around this pipe-sticking event: the
lost circulation reduced SPP_pressA, and increased MPP_SPM1 and MPP_SPM2. Lost circulation with a
high rate generally reduces the cuttings transport efficiency and increases a probability of pipe-sticking due
to pack-off. Therefore, it is not clear whether the increasing of the reconstruction errors indicates the lost
circulation during the time period around the pipe-sticking event or the predictors of the pipe-sticking event.
However, the above reconstruction errors are sensitive to large variations in the drilling parameters; this
was not unique to stuck-pipe predictions, but was common to all time series data. Therefore, the use of this

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/1-21APOG/D011S009R006/2496805/spe-205677-ms.pdf by U. of Alberta Library user on 15 September 2023

LSTM-AE model made it difficult to distinguish reliable, true predictors of pipe-sticking from misleading,
false ones caused by these large variations. The optimization of hyperparameters such as the number of
latent dimensions and the learning rate during model training is expected to lead to a decrease in the number
of false predictors and, thereby, more accurate predictions of pipe-sticking.

Conclusions
To make stuck-pipe predictions using artificial intelligence, we undertook a project in collaboration with
industry, the government, and academia (Inoue et al. 2020). This led to the development of an LSTM-
AE model based on unsupervised machine learning. The model was trained with the time-tagged mud
logging data of normal drilling operations based on the hypothesis that reconstruction errors between true
and predicted values increase around the time of occurrence of the stuck-pipe event, when compared to
the time of normal drilling operations. This model was applied to all 34 stuck-pipe events, and results
showed high AUC values for some of these events. We also found that the reconstruction errors did increase
prior to pipe-sticking in some cases (thereby partially confirming our hypothesis), while also sensitively
responding to large variations in the drilling parameters. Therefore, we conclude that while the LSTM-
AE model appears promising for the making of stuck-pipe predictions which could give crews sufficient
time to take preventative or remedial actions, hyperparameter optimization is required for more accurate
predictions.

Acknowledgements
This work was supported by the Japan Oil, Gas and Metals National Corporation.

References
Abbas, A. K., Flori, R., Almubarak, H. et al. 2019. Intelligent Prediction of Stuck Pipe Remediation Using Machine
Learning Algorithms. Paper Presented at the SPE Annual Technical Conference and Exhibition, Calgary, Alberta,
Canada, September. SPE-196229-MS. https://doi.org/10.2118/196229-MS
Ahmed, O. S., Aman, B. M., Zahrani, M. A. et al. 2019. Stuck Pipe Early Warning System Utilizing Moving Window
Machine Learning Approach. Paper Presented at the Abu Dhabi International Petroleum Exhibition & Conference,
Abu Dhabi, UAE, November. SPE-197674-MS. https://doi.org/10.2118/197674-MS
Al-Baiyat, I., and Lloyd H. 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, December. SPE-163370-MS. https://doi.org/10.2118/163370-MS
Alshaikh, A., Magana-Mora, A., Gharbi, S. A. 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, March. IPTC-19394-MS. https://doi.org/10.2523/IPTC-19394-MS
Brankovic, A., Matteucci, M., Restelli, M. et al. 2020. A Data-Based Approach for the Prediction of Stuck-Pipe Events
in Oil Drilling Operations. Paper Presented at the Abu Dhabi International Petroleum Exhibition & Conference, Abu
Dhabi, UAE, November. SPE-202625-MS. https://doi.org/10.2118/202625-MS
Chamkalani, A., Pordel S. M., and Saeed P. 2013. Support Vector Machine Model: A New Methodology for Stuck Pipe
Prediction. Paper Presented at the SPE Unconventional Gas Conference and Exhibition, Muscat, Oman, January.
SPE-164003-MS. https://doi.org/10.2118/164003-MS
Diederik, K., Jimmy, B. 2015. Adam: A Method for Stochastic Optimization. Published as a conference paper at the 3rd
International Conference for Learning Representations, San Diego. https://arxiv.org/abs/1412.6980v8
SPE-205677-MS 9

Hess, J. 2016. Pipe Sticking Prediction Using LWD Real-Time Measurements. Paper Presented at the IADC/SPE Drilling
Conference and Exhibition, Fort Worth, Texas, USA, March. SPE-178828-MS. https://doi.org/10.2118/178828-MS
Inoue, T., Wada, R., Miyoshi, K. et al. 2020. Research Project on Safety Improvement by Applying Digital Technology
for Drilling Operations (in Japanese), Proceedings of the Japan Society of Naval Architects and Ocean Engineers.
ISSN: 2185–1840.
Inoue, T., Nakagawa, Y., Wada, R. et al. 2021. Attempt of Early Stuck Detection Using Unsupervised Deep Learning with
Probability Mixture Model. Proceedings of the ASME 2021 40th International Conference on Ocean, Offshore and
Arctic Engineering OMAE 2021. Virtual, Online.

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/1-21APOG/D011S009R006/2496805/spe-205677-ms.pdf by U. of Alberta Library user on 15 September 2023

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, March. SPE-194980-MS. https://doi.org/10.2118/194980-MS
Meor Hashim, M. M., Yusoff, H. M., Arriffin, F. M. et al. 2021a. Utilizing Artificial Neural Network for Real-Time
Prediction of Differential Sticking Symptoms. Paper Presented at the International Petroleum Technology Conference,
Virtual, March. IPTC-21221-MS. https://doi.org/10.2523/IPTC-21221-MS
Meor Hashim, M. M., Yusoff, H. M., Arriffin, F. M. et al. 2021b. Performance Improvement of Wells Augmented Stuck
Pipe Indicator via Model Evaluations. Paper Presented at the International Petroleum Technology Conference, Virtual,
March. IPTC-21455-MS. https://doi.org/10.2523/IPTC-21455-MS
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 (2017): 184–193. SPE-178888-PA. https://doi.org/10.2118/178888-PA
Sashank, J. R., Satyen, K., and Sanjiv, K. 2018. On the Convergence of Adam and Beyond. International Conference on
Learning Representations. https://openreview.net/forum?id=ryQu7f-RZ
Shoraka, S. A., Shadizadeh, S. R., and Pordel Shahri, M. 2011. Prediction of Stuck Pipe in Iranian South Oil Fields Using
Multivariate Statistical Analysis. Paper Presented at the Nigeria Annual International Conference and Exhibition,
Abuja, Nigeria, July. SPE-151076-MS. https://doi.org/10.2118/151076-MS
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, February.
SPE-98378-MS. https://doi.org/10.2118/98378-MS
Tsuchihashi, N., Wada, R., Ozaki, M. et al. 2021. Early Stuck Pipe Sign Detection with Depth-Domain 3D
Convolutional Neural Network Using Actual Drilling Data. SPE J. 26 (2021): 551–562. SPE-204462-PA. https://
doi.org/10.2118/204462-PA
Zhang, F., Islam, A., Zeng, H. et al. 2019. Real Time Stuck Pipe Prediction by Using a Combination of Physics-
Based Model and Data Analytics Approach. Paper Presented at the Abu Dhabi International Petroleum Exhibition &
Conference, Abu Dhabi, UAE, November. SPE-197167-MS. https://doi.org/10.2118/197167-MS
Zhu, Q., Wang, Z., and Jian H. 2019. Stuck Pipe Incidents Prediction Based On Data Analysis. Paper Presented
at the SPE Gas & Oil Technology Showcase and Conference, Dubai, UAE, Octobe. SPE-198672-MS. https://
doi.org/10.2118/198672-MS