ARMA 23–0326

Drilling Stuck Probability Intelligent Prediction Based on

LSTM Considering Local Interpretability
Jianhua Wang. And Zhen Guan.
Kunlun Digital Technology Co., Ltd., Beijing, China

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

Muchen Liu. And Xianzhi Song.
China University of Petroleum(Beijing), Beijing, China
Zheng Wang.
China University of Petroleum(Beijing), Beijing, China

Copyright 2023 ARMA, American Rock Mechanics Association

This paper was prepared for presentation at the 57th US Rock Mechanics/Geomechanics Symposium held in Atlanta, Georgia, USA, 25-28 June
2023. This paper was selected for presentation at the symposium by an ARMA Technical Program Committee based on a technical and critical
review of the paper by a minimum of two technical reviewers. The material, as presented, does not necessarily reflect any position of ARMA, its
officers, or members. Electronic reproduction, distribution, or storage of any part of this paper for commercial purposes without the written consent
of ARMA is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 200 words; illustrations may not be copied. The
abstract must contain conspicuous acknowledgement of where and by whom the paper was presented.

ABSTRACT: Stuck pipe in the drilling can increase the non-productive time and lead to drilling risks such as buried tool and drilling
tool fracture. Due to complex sequential feature analysis and the lag of downhole data transmission, traditional methods which are
mostly used for post-drilling analyses have poor timeliness and low accuracy so that it cannot meet the requirements of field
application. This study provides a data-driven prediction method for stuck pipe probability. First, the dataset is established combining
drilling records and logging data from 108 wells, and 15 input features are selected. Then, the intelligent prediction model based on
the Long Short Term Memory (LSTM) network is developed considering the timing characteristics of drilling parameters. Finally,
Shapley Additive Explanations (SHAP) is used to calculate the quantitative contribution of input parameters to solve the problem of
the black box. The results show that the accuracy of the LSTM model is 97%. Through local interpretation and analysis, the model
accords with the natural physical law. The proposed model has great timeliness, high accuracy, high reliability and quantitative risk
prediction. This study provides useful insights into predicting stuck pipe risk.

and a new side drilling scheme need to be redesigned,

1. INTRODUCTION which seriously restricts the safety and efficiency of
drilling. Therefore, we focus on the real-time prediction
Drilling is the only way to get oil and gas which is an of stuck pipe in the process of drilling.
uncertain undertaking. Due to poor underground
conditions, complex conditions (lost circulation, well Many approaches have been developed to predict the risk
collapse, stuck pipe, etc.) often occur, among which of stuck pipe. For the knowledge-driven model, cuttings,
complicated conditions of stuck pipe often occur in the wellbore tortuosity, and other factors will eventually lead
drilling construction process. According to Chunbo Zhao to an increase in torque and drag which can lead to a stuck
(1998), stuck drilling accidents account for 40%~50% of pipe accident. Therefore, researchers use the torque and
the total drilling complications in western oilfields, which drag model to invert the friction coefficient to realize the
seriously restricts safe, economical, and efficient drilling. qualitative analysis of the stuck pipe trend in the view of
Therefore, the timely prediction of stuck pipe events is mechanics. Lesage, M (1988) shows how to estimate two
considered a primary necessity to assist the drilling team friction coefficients on a foot-by-foot basis at the well site
in the decision-making process, so that appropriate with both measurement-while-drilling (MWD) and
countermeasures can be put into effect before the situation surface values of weight on bit (WOB) and torque. A log
slips out of hand. of the coefficients with depth can be used to diagnose
drilling problems in directional wells. Field examples are
Mudlog data shows that stuck pipe accidents occur in the given that show how the technique detects incipient stuck
process of tripping and drilling with varying degrees of pipe. Annis, M.R. (1962) describes an investigation of the
severity. For example, in some highly deviated wells and friction between the pipe and the mud cake. It was found
horizontal wells, the stuck pipe accident is often that the apparent coefficient of friction, or the “stuck pipe”
accompanied by the process of tripping, but it is usually coefficient, was not a constant; instead, it increased with
not serious. On the contrary, more attention is paid to the increased time of contact between the plate and mud cake.
stuck pipe during drilling. Once the drill string is They think that the friction factor and mud cake quality
completely stuck, it can only be lifted by the explosion are factors that cannot be ignored. Due to various factors,
difficult sequential feature analysis, and the lag of As shown in Fig. 1, it has a Recurrent Neural Network
downhole data transmission, the traditional stuck (RNN) that enables information from a hidden layer of
prediction methods have poor timeliness and low neurons to be transmitted to the next hidden layer of
accuracy, which are mostly used for post-drilling analyses neurons through the use of self-feedback neurons, has a
but not for real-time quantitative analysis · . Therefore certain memory capacity, and has obvious advantages in
methods cannot meet the requirements of field application. processing sequential problems. However, because the
RNN structure shares weight and bias, the gradient is
For the data-driven models, The use of artificial continuously multiplied in the process of backpropagation,
intelligence methods can help to solve the complex and

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

which makes the gradient get bigger or smaller, and then
strong nonlinear problems which are difficult to solve in produces the problem of gradient explosion or gradient
the traditional engineering field. Many scholars have disappearance. Hochreiter (1997) gave birth to the LSTM
applied artificial intelligence algorithms to the prediction network by improving the traditional RNN network
of stuck pipe. Murillol (2009) introduced mathematical structure.
methods such as fuzzy comprehensive discrimination and
neural networks to predict stuck pipe and proposed new
methods for early warning and risk assessment of stuck
pipe. Albaiyat (2019) used ANN along with the support
vector machine to predict stuck pipe events. Naraghi, M.E.
(2013) used Fuzzy logic and active learning for stuck pipe
prediction. Runqiao Yu (1996) selected nine parameters
and used the method of system dynamic cluster analysis
to divide the stuck pipe accidents into three categories:
mechanical stuck pipe, differential pressure stuck pipe, Fig. 1. Structure of long-short-term memory Neural network.
and circulation stuck pipe. However, due to the limitation
of data quality and algorithm principle, the model could The neurons of LSTM are designed with input gate i,
not consider the timing nature of the stuck pipe and could forgetting gate f, output gate o and memory unit C. LSTM
not quantify the results, resulting in low accuracy. Most relies on the structure of these "gates" to allow
machine learning methods belong to the black box model information to selectively influence the state of the neural
and are uninterpretable, resulting in insufficient reliability network at each moment. A fully connected neural
of the model. network that uses sigmoid as an activation function
outputs a value between 0 and 1 that describes how much
Based on previous research, we establish the Long Short
information can pass through the structure in the current
Term Memory (LSTM) model for stuck pipe probability
input, thus avoiding gradient explosion or gradient
prediction, where the LSTM is utilized for the sequential
disappearance.
input features. which is expected to realize an accurate
prediction of stuck pipe probability and ensure the forgetting gate：
timeliness of the model. Moreover, the SHAP method is
applied to quantitatively calculate the contribution of the ft =  (W f   ht −1 , xt  + b f ) (1)
input features to stuck pipe and analyze the local
interpretability of the model. Input gate：

it =  (Wi   ht −1 , xt  + bi ) (2)
2. METHODOLOGY
2.1. LSTM structure and adaptability analysis output gate：

Ot =  (WO   ht −1 , xt  + bO )
The LSTM is designed to address long-term dependencies
(3)
and retain information over long periods. It is suitable for
dealing with dynamic changes during drilling. In this Where 𝑥𝑡 is the current input; ℎ𝑡−1 is the output of the
section, the nonlinear fitting of neural network and the
previous time; 𝑤𝑓 is the weight matrix of forgetting gate
ability of long-term and short-term memory are used to
establish a function between timing data and stuck pipe neurons in LSTM; 𝑏𝑓 is the bias vector of amnestic gate
risk. Due to the network structure characteristics of neurons in LSTM. 𝑓𝑡 is the output of amnestic gate
LSTM, the established model not only considers the neuron in LSTM. 𝑤𝑖 is the weight matrix of input gate
complex relationship between drilling engineering neurons in LSTM; 𝑏𝑖 is the bias vector of the input gate
parameters and influencing factors but also considers the neuron in LSTM; 𝑖𝑡 is the output of the input gate neuron
changing relationship of parameters themselves in time in LSTM; 𝑤𝑜 is the weight matrix of output gate neurons
dimension. in LSTM; 𝑏𝑜 is the bias vector of output gate neuron in
LSTM. 𝑜𝑡 is the output of output gate neuron in LSTM; 𝜎 analyze the interpretability of the established intelligent
is sigmoid activation function; model to enhance its reliability.
In the process of drilling, there are frequent incidents of According to Chunyan Zeng (2021), Machine-learning
stuck pipe, such as turning plate or top drive. Although interpretability methods are mainly divided into global
the occurrence of stuck pipe accidents is often sudden and interpretation methods and local interpretation methods.
the response time of field drilling engineers is very short, The global interpretation method refers to a
there may be some anomalies in the surface data comprehensive understanding of the complex operational
monitored by the comprehensive logging instrument logic and causes inside the black-box model (such as how

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

before stuck pipe occurs, but these anomalies are difficult to understand the decision-making causes of the model
to be detected, and the anomalies are different for from a human perspective). Limited by application
different types of stuck pipe. However, no matter which scenarios and specific algorithms, it is difficult to achieve
type of stuck pipe, it is different from normal drilling. global interpretation at the current stage. Local
Therefore, the prediction of stuck pipe in the drilling interpretation methods are more focused on analyzing the
process can be treated as a classification problem (stuck logical relationship between model inputs and outputs,
pipe and not stuck pipe). Due to the irregularity and strong explaining how inputs affect the output changes, rather
randomness of the anomalies generated before the stuck than explaining the decision function itself. Shouling Ji
pipe is about to occur, it is difficult for traditional methods (2019) concluded that the local interpretation method is
to identify the early signs. Therefore, machine learning relatively mature and widely used, so this paper uses this
can be used to build an intelligent prediction model of method for interpretability analysis.
stuck pipe, which is good at dealing with complex
SHAP is a method inspired by the game theory that
nonlinear problems, and identifying the anomalies in
analyses interpretability by calculating the importance of
advance.
each feature to a single predicted outcome. It is a model-
For SVM, random forest, and other machine learning independent approach, meaning that it works for any
classification algorithms, they can only give the final single output model. Linardatos P (2020) concludes that
coon of stuck and non-stuck. However, the early signs of In the interpretation of the arbitrary black box model, the
stuck are characterized by strong randomness and poor SHAP method is by far the most comprehensive and
regularity and are greatly affected by field engineering dominant in terms of visual interaction and feature
equipment. Therefore, it is necessary to be cautious when importance. In essence, the input features are arranged
identifying early signs of stuck through a data-driven and combined, the feature subset is used as input, and the
method, and it is safe to give the probability of stuck in other feature inputs are set to 0. According to the changes
the future. The neural network model can deal with in model output, the influence degree of each input feature
regression problems as well as classification problems, on the output is quantitatively evaluated. The method is
for which the output of the neural network can give the shown in Fig. 2.
probability of the target category. Although the accident
happened suddenly at the site, it was still a gradual
process using a short time, before the obvious gradual
process implied early signs of stuck pipe. Therefore, the
prediction of stuck accidents should consider the data of
the current moment and a period before, and analyze the
complex relationship between the change characteristics
of the data in this period and the early signs of stuck pipe.
LSTM can analyze the time series data and is suitable for
dealing with the prediction and classification problems
that change with time. Fig. 2. The SHAP method.
2.2. Model interpretation method based on SHAP
Casalicchio G (2018) quantified influence degree by
Most intelligent models are black box models. Although
Shapley value, as shown in Eq. (4).
the performance of the models can reach a good level, it
is impossible to know the operational mechanism and the ( S − 1) !( n − S )！

influence of input on output, which reduces the reliability ij =  SN 
n!  ( )
  f ( S ) − f S \  xij   (4)

of the models and limits the wide application of machine
learning methods in the engineering field. To improve the In Eq. (1), 𝜑𝑖𝑗 represents the i-th feature Shapley value of
rationality and reliability of the proposed stuck the j-th data sample. 𝑥𝑖𝑗 represents the i-th feature of the
probability intelligent prediction model and enhance the j-th data sample; 𝑁 represents the set of all input features;
application potential of this study, it is necessary to 𝑆 represents the contained feature set, and there is a
relationship 𝑆 ∈ 𝑁 ; |𝑆| represents the number of the In Eq. (5), 𝑆𝑆𝐻𝐴𝑃𝑖 represents the SHAP coefficient of the
contained feature set; 𝑛 represents the number of all input i-th feature, and its value is [0,1]. 𝜑𝑖𝑗 represents the i-th
features; 𝑓(𝑆) represents the model output value when feature Shapley value based on the j-th data sample; 𝑛
the feature set 𝑆 serves as the model input. Features not represents the number of all input features; 𝑀 represents
included are set to 0 during model calculation. 𝑓(𝑆\ the number of data samples.
{𝑥𝑖𝑗 }) represents the model output value when the
feature set 𝑆 is used as model input after the feature 𝑥𝑖𝑗 is 3. EXPERIMENTS AND ANALYSIS

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

excluded. 3.1. Data set construction and feature selection
Through the calculation of Eq. (4), the Shapley value of In this study, data from 108 Wells were collected. Drilling
the i-th feature in the j-th data sample is obtained. To logs were first analyzed to obtain stuck pipe samples. For
analyze local interpretability from a macro perspective one well, it can be found from the stuck pipe logs that the
(considering all data), the SHAP coefficient was date of the log is 2018/05/03. Due to the recording lag or
introduced based on the Shapley value. The SHAP dislocation in some logs, the main control parameters of
coefficient of the i-th feature was defined as Eq. (5). stuck pipe for 05/02, 05/03, and 05/04 days are drawn to
accurately judge the stuck bit date, as shown in Fig. 3.
M


j =1
ij

S SHAPi = n M
100% (5)
 
i =1 j =1
ij

Fig. 3. The bit position, hook load, and torque change diagram of 05 / 02, 05 / 03, and 05 / 04 days during drilling are retained.

According to the log, the stuck pipe date was 05/03, but torque has a significant change from 14.2kN·m to
as can be seen from Fig. 3, the actual stuck pipe date was 19.6kN·m. Therefore, based on the above analysis, the
05/02. The analysis process was as follows. First, the log actual date of the stuck pipe can be finally determined as
recorded that the PowerV drilled from 8:00 to 20:50, so 05/02.
the bit depth should have increased gradually during this
After determining the actual date of stuck pipe, it is
period, while the bit position did not increase gradually
necessary further to determine the specific time point of
throughout the day of 05/03. Therefore, it can be
stuck pipe to facilitate further labeling of normal/stuck
preliminarily determined that the actual date of stuck drill
working condition samples. Fig. 4 shows the changes in
should be 05/02, and at around 20:50:00 of 05/02, the
hook load and torque on the whole day of 05/02. It can be
seen from the Fig. 4 that the stuck pipe occurs around suddenly increased from 14.2 kN·m to 19.6 kN·m, as
20:48:08, and the interception period is from 20:11:28 to recorded in the log, so the exact stuck time was
21:21:28. During 20:50:28-20:51:28, The torque determined to be 05/02 20:51:28.

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

Fig. 4. Variation diagram of hook load and torque in the whole day on 05 / 02.

Because the stuck pipe accident tends to be a mutation pipe data, it is necessary to mark the stuck pipe working
accident, from the beginning of the sign of stuck pipe to condition based on the stuck time point of the drill tool in
the drilling tool stuck in a short time, insufficient time can advance, to achieve the purpose of predicting in advance.
not be set aside to adjust drilling parameters in time to
avoid stuck pipe. Therefore, when calibrating the stuck

Table 1. Prediction effect of the model under different division ratio

Partition ratio stuck samples normal samples Accuracy Precision Recall F1
1:1 1/2 1/2 0.97 0.97 0.98 0.97
1:2 1/3 2/3 0.96 1.0 0.90 0.94
1:3 1/4 3/4 0.82 1.0 0.31 0.47

As shown in Table 1, a reasonable partition ratio is Stuck pipe prediction is a complicated problem, the
selected by comparing the model prediction effect under essence of which is that every link in the drilling process
different sample partition ratios. The division ratio is not solved well. For example, the dogleg Angle of the
between the stuck pipe and normal samples varies from well track is too large, the structure design of the drill
1:1, 1:2 to 1:3, and the corresponding model prediction string is not reasonable and large rocks fall off the wall,
effect (F1) also gradually decreases. There is little easy to cause stuck pipe; The unreasonable design of
difference between the prediction effect of the model with hydraulic parameters will lead to the unreasonable
an equilibrium ratio of 1:1 and a non-equilibrium ratio of cleaning of boreholes and lead to stuck pipe. No matter
1:2. Considering that the proportion of stuck pipe samples which link causes stuck pipe, it will eventually reflect the
in the total samples in field practice is very small, a stress of drill string. The parameters on the ground that
disequilibrium ratio (1:2) was selected to divide the stuck can reflect the force of drill string include hook load,
pipe sample set. data points from 20:28:28 to 20:50:28 are rotary torque, weight on bit, and rotary speed. In this
marked as stuck working conditions, and 20:11:28 to paper, riser pressure, pump stroke, inlet flow, and outlet
20:27:28 are marked as normal working conditions. flow are also selected as input parameters, which can
characterize the degree of cuttings in the hole to a certain
extent. In addition, some basic parameters of drilling fluid
(density, temperature, etc.) and bit position are also
included.

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

Fig. 5. Division diagram of normal drilling and sticking samples.

The characteristic parameters selected in this paper are used to set up the test scheme, as shown in Table 2 and
mainly based on an analysis of factors affecting stuck pipe Table 3.
and field data acquisition. 15 input parameters are adopted,
2) Evaluation metrics selection
which are: well depth, hook load, riser pressure, rotary
torque, weight on bit, rotary speed, pump stroke, inlet As shown in Table 4, there are four common model
temperature, outlet temperature, inlet conductance, outlet evaluation indexes for classification problems in machine
conductance, inlet density, outlet density, inlet flow rate, learning: Accuracy, Precision, Recall, and F1-score.
and outlet flow rate. For the factors such as inclination
Angle and azimuth Angle, although they are closely Table 4 Evaluation index of the classification problem
related to stuck pipe, cannot be measured in real-time in Prediction result
the drilling process, so this paper does not consider such fact
Stuck pipe normal
factors. After processing, 15 stuck events were obtained Stuck pipe TP FN
from 11 Wells, with more than 105000 samples, and the normal FP TN
ratio of normal samples to stuck was 1:2.
3.2. Training and optimization of LSTM model In a binary classification problem, the calculation formula
1) Experiment design is as follows:
TP + TN
In the training process of the LSTM model, the Accuracy = (6)
optimization of hyperparameters is a work based on TP + TN + FP + FN
experiment and experience. Generally, the optimal TP + TN
combination of hyperparameters is obtained through Precison = (7)
repeated trials. In this paper, the time step, learning rate, TP + FP
number of LSTM network layers, number of LSTM FAR ( False _ alarm _ rate ) = 1 − Pr ecison (8)
neurons, and activation function of input data are
optimized. Each factor involves three levels. If all tests TP
are conducted, 3^5=234 tests are needed. To reduce the Recall = (9)
number of tests, the orthogonal test method (L18.3.5) was TP + FN
MAR ( Missed _ alarm _ rate ) = 1 − Re call (10)
Table 2 Factor level table of neural network model test
factors
level
Timesteps learning rate number of layers number of neurons activation function
1 5 0.1 1 16 tanh
2 10 0.01 2 32 relu
3 20 0.001 3 64 sigmoid

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

Table 3 Orthogonal test table of the neural network model
number Timesteps learning rate number of layers number of neurons activation function
1 5 0.1 1 16 tanh
2 5 0.1 2 32 sigmoid
3 5 0.01 1 64 sigmoid
4 5 0.01 3 16 relu
5 5 0.001 2 64 relu
6 5 0.001 3 32 tanh
7 10 0.1 1 64 relu
8 10 0.1 3 16 sigmoid
9 10 0.01 2 32 relu
10 10 0.01 3 64 tanh
11 10 0.001 1 32 sigmoid
12 10 0.001 2 16 tanh
13 20 0.1 2 64 tanh
14 20 0.1 3 32 relu
15 20 0.01 1 32 tanh
16 20 0.01 2 16 sigmoid
17 20 0.001 1 16 relu
18 20 0.001 3 64 sigmoid

2  Pr ecison  Re call indicators are auxiliary indicators. The MAR refers to the
F1_ score = (11) data sample that is a stuck pipe is predicted as normal,
Pr ecison + Re call
while the FAR refers to the normal sample is predicted as
Among them, TP (True Positive): correctly predicted the stuck pipe.
stuck pipe sample as a stuck pipe;
3) Analysis of experimental results
TN (True Negative): The normal sample is correctly
Six indexes, including FAR, MAR, Precision, Recall, F1-
predicted as normal;
score, and model complexity, are integrated to optimize
FP (False Positive): The normal sample is incorrectly the LSTM stuck pipe prediction model. The evaluation
predicted as stuck; indexes of each model are shown in Table 5 and Fig. 6 to
FN (False Negative): indicates that the stuck pipe sample Fig. 11.
is incorrectly predicted as normal. As shown in Fig. 9, there are models with high F1-score
In the field of engineering, for the prediction of complex indexes: 1, 2, 4, 7, 9, 10, 13, 15, and 17, and their values
working conditions, especially when the cost of missing are very similar. As there are not many stuck pipe training
the prediction of risk is large, the engineering would samples, to prevent model overfitting and improve model
rather misreport than miss the report. According to the generalization ability, models with low complexity are
field data statistics, if the stuck occurs during drilling and selected, as shown in Fig. 11. Models with low
fails to release the stuck successfully, it is necessary to complexity include the first and 17th models. In
blast release the stuck. It often takes several months from engineering, the risk prediction would rather be a false
stuck drilling tools to re-striking, increasing non- alarm than a missed alarm. Therefore, more attention
productive time and severely affecting drilling efficiency. should be paid to the MAR when the FAR and MAR are
Therefore, the FAR and MAR are very important to used in the optimization model. As shown in Fig. 7, the
forecast stuck pipe. In this paper, F1-score, FAR, MAR, first model has a small MAR, so the optimal model is
and model complexity are the main indicators, and other finally selected.
Table 5 108 results of the Neural network model
Model
Number Precision FAR Recall MAR F1-score
complexity
1 0.9796 0.0204 1.0 0 0.9897 1681
2 0.9919 0.0081 0.9966 0.0034 0.9942 13713
3 0.9132 0.0868 0.9635 0.0365 0.9377 18961
4 0.9911 0.0089 0.9961 0.0039 0.9936 5905
5 0.9883 0.0117 0.9707 0.0293 0.9794 51985

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

6 0.9151 0.0849 0.9407 0.0593 0.9277 22033
7 0.9874 0.0126 1.0 0 0.9936 18961
8 0.9670 0.033 0.9817 0.0183 0.9743 5905
9 0.9949 0.0051 0.9987 0.0013 0.9968 13713
10 0.9936 0.0064 0.9978 0.0022 0.9957 85009
11 0.5284 0.4716 1.0 0 0.6914 5393
12 0.9288 0.0712 0.9567 0.0433 0.9426 3793
13 0.9970 0.003 0.9995 0.0005 0.9983 51985
14 0.9527 0.0473 0.9995 0.0005 0.9756 22033
15 0.9949 0.0051 0.9987 0.0013 0.9968 5393
16 0.9312 0.0688 0.9809 0.0191 0.9554 3793
17 0.9867 0.0133 0.9771 0.0229 0.9819 1681
18 0.9319 0.0681 0.9635 0.0365 0.9475 85009

Fig. 6. Recall of different network models. Fig. 8. Precision of different network models.

Fig. 7. MAR of different network models. Fig. 9. F1-score of different network models.
 no missing report occurs. For a certain well, the prediction
results are shown in Fig. 12. The probability of sticking
predicted by the LSTM model, represented by blue dots,
is gradually increasing, and it is finally judged to be stuck
at 6108.8m. The actual situation was that the stuck pipe
did occur at 6109.4m. The prediction was realized 14min
in advance.

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

Fig. 10. FAR of different network models.

Fig. 13. SHAP coefficients of parameters.

There are many kinds of stuck pipe, such as differential

pressure, collapse, sand bridge, diameter reduction,
keyway, falling object, mud bag, etc., which can be
properly divided into several categories of stuck
conditions during drilling. Runqiao Yu (1996) used the
systematic dynamic cluster analysis method to analyze
the actual data of 63 Wells and divided more than a dozen
Fig. 11. Complexity of different network models.
types of stuck pipe into three categories: mechanical stuck,
differential pressure stuck, and circulation stuck. The
3.3. Results and interpretability analysis
mechanical stuck included falling object stuck drilling
and mud bag stuck drilling, and the circulation stuck
drilling included sand settling, well collapse, and sand
bridge stuck pipe.
First of all, no matter what reason the pipe is stuck, it is
reflected by the friction of the drill string. The friction will
be increased before or after the drill string is stuck. Under
drilling conditions, simple statics force analysis can be
expressed by Eq. (12):
f + hookload + wob = G (12)
Where, 𝑓: friction of drill string, 𝑁; 𝐺: floating weight of
drill string, 𝑁; ℎ𝑜𝑜𝑘𝑙𝑜𝑎𝑑: hook load, 𝑁; 𝑤𝑜𝑏: weight on
bit, 𝑁;
Fig. 12. Prediction results of the optimal model. In normal drilling, the axial movement rate of the drill
string is slow, so the force is balanced, thus satisfying the
In three Wells, all four stuck events were successfully above equation. The floating weight of the drill string
predicted. The prediction accuracy rate of stuck pipe does not change significantly. When a stuck pipe occurs,
samples is 97.96%, and the FAR is 2.04%. Although some the friction 𝑓 increases, and the hook load and bit weight
normal samples are wrongly predicted as stuck pipe, the will decrease. At the same time, in the circumferential
model can accurately predict all stuck pipe samples, and direction, the drill string will encounter resistance, and the
torque of the surface torque will increase sharply until the 2. Casalicchio G , Molnar C , Bischl B . Visualizing the
turntable or top drive will be halted. Therefore, in the Feature Importance for Black Box Models[J]. 2018.
process of drilling, the obvious characteristics of 3. Chunyan Zeng and Kang Yan, (2021). A review of
mechanical stuck are determined to be the sharp decline research on the interpretability of deep learning models (in
Chinese). J. Computer engineering and applications.57(8):
of hook load and the sharp increase of rotary torque, and
9.
the obvious characteristics of circulating stuck are 4. Chunbo Zhao and Yuewen Kang, (1998). Research review
determined to be the sharp increase of pump pressure or of stuck drill prediction technology (in Chinese). J.
standpipe pressure. Xinjiang Petroleum Science and Technology.8, 9.

Downloaded from http://onepetro.org/ARMAUSRMS/proceedings-pdf/ARMA23/All-ARMA23/ARMA-2023-0326/3251831/arma-2023-0326.pdf by U. of Alberta Library user on 15 September 2023

According to the well-stuck log, the incident occurred as 5. Heinze, L.and Al-Baiyat.(2012) Implementing Artificial
Neural Networks and Support Vector Machines in Stuck
follows: "At 20:50, the torque increased from 14.2 kN·m
Pipe Prediction. In Proceedings of the SPE Kuwait
to 19.6 kN·m when the well was drilled to 6110.58 m with International Petroleum Conference and Exhibition,
PowerV and the top drive was halted. 20:57 Pull up and Kuwait City, Kuwait, 10–12 December 2012.
back reaming until the pipe is stuck at a depth of 6109 m." 6. Hochreiter, Sepp, Schmidhuber and Jurgen. (1997). Long
It can be concluded that mechanical stuck occurred in the short-term memory. Neural Computation Heinze.
well. Fig. 13 shows the quantified contribution of features 7. Lesage, M.and Falconer, I.G.(1988). Evaluating Drilling
to the stuck probability. The SHAP coefficients of hook Practice in Deviated Wells with Torque and Weight Data.
load, weight on bit, and rotary torque, which are used to SPE Drill. Eng.3:248–252. [CrossRef].
judge the mechanical stuck, rank first, second, and fourth. 8. Linardatos P, Papastefanopoulos V and Kotsiantis S.(2020).
Standpipe pressure and pump stroke, which are used to Explainable AI: A Review of Machine Learning
Interpretability Methods[J]. Entropy (Basel, Switzerland),
judge circulation stuck, ranked 12th and 14th. It is shown
23(1): 18.
that the model can correctly predict the stuck pipe 9. Murillo, A, Neuman, J and Samuel, R.(2009) Pipe stuck
according to the main controlling factors of mechanical pipe prediction and avoidance using adaptive fuzzy logic
stuck, but not according to the main controlling factors of and neural network modeling. In Proceedings of the SPE
circulating stuck. This is following the real physical Production and Operations Symposium, Oklahoma City,
situation. Therefore, the developed LSTM model for OK, USA, 244–258.
intelligent prediction of stuck pipe probability is locally 10. Naraghi, M.E, Ezzatyar, P and Jamshidi, S. (2013).
interpretable and conforms to the real physical law in Prediction of drilling pipe stuck pipe by active learning
terms of input and output based on SHAP analysis, which method (ALM). J. Pet. Gas Eng. 4, 173–183. [CrossRef]
further enhances the reliability of the model. 11. Runqiao Yu. (1996). Research on Prediction Technology
of stuck drill Accident(in Chinese)). Oil drilling
technology. 024, 15–17.
4. CONCLUSION 12. Shouling Ji, Jinfeng Li and Tianyu Ji. (2019). A review of
machine learning model interpretability methods,
In this study, the LSTM network is selected and the model applications, and security(in Chinese). J. Computer
training is conducted based on the stuck pipe sample data, research and development, 56(10): 2071-2096.
and the intelligent prediction model of the stuck 13. Yu, R.Q. (1996). Research on prediction technology of
probability is established. The intelligent model was stuck pipe accidents. Pet. Drill. Tech, 24, 15–17.
optimized by model complexity, false alarm rate, false
alarm rate, and other indicators. The accuracy rate of the
optimal model was 97%, the false alarm rate was 2%, the
false alarm rate was 0, and the model complexity was
1681. Based on the SHAP method, the local
interpretability of the model is verified from the
perspective of input and output, and the reliability of the
model is enhanced.
For future work, we plan to collect more stuck pipe
samples, so that we can increase the stability and
generalization ability of the intelligent prediction model,
and determine a reasonable threshold of stuck pipe instead
of setting thresholds artificially.

REFERENCES
1. Annis, M.R.; Monaghan, P.H. Differential Pressure Stuck
pipe-Laboratory Studies of Friction Between Steel and
Mud Filter Cake. J. Pet. Technol. 1962, 14, 537–543.
[CrossRef].