SPE-205627-MS

Bi-Directional Long Short-Term Memory Variational Autoencoder for

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/3-21APOG/D031S025R004/2495836/spe-205627-ms.pdf/1 by Shanghai Jiao Tong University user on 29 March 2024
Real-Time Bit-Wear Estimation

Trieu Phat Luu and John A.R. Bomidi, Baker Hughes; Arturo Magana-Mora, Alawi Alalsayednassir, and Guodong
David Zhan, Saudi Aramco

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
Drilling operations rely on learned expertise in monitoring the drilling performance data and the rock data to
assess the dull condition of the drill bit. While human learning can subjectively pick up the indicators based
on rig surface data streams, this information is highly convoluted with changes in rock and drilling data.
Recent approaches for bit wear estimation also include model-based and traditional supervised machine
learning methods, which are usually costly and time-consuming. In this study, we developed a bi-directional
long short-term memory-based variational autoencoder (biLSTM-VAE) to project raw drilling data into a
latent space in which the real-time bit-wear can be estimated. The proposed deep neural network was trained
in an unsupervised manner, and the bit-wear estimation is demonstrated as an end-to-end process.

Introduction
One of the critical decisions in real-time drilling operations is to trip out to replace the drilling bit or
continue drilling with a suboptimal rate of penetration. The decision requires a combination of knowledge,
such as an apriori estimate of formation ahead of where the bit is currently drilling or refined projection
by recently learned knowledge of the formation, and importantly data-driven assessment of the bit dull
condition. Drilling operations rely on learned expertise in monitoring the drilling performance data and the
rock data to assess the dull condition of the drill bit. While human learning can subjectively pick up the
indicators based on rig surface data streams, this information is highly convoluted with changes in rock
and drilling data.
Recently, data-driven approaches using machine learning (ML) and deep learning (DL) algorithms have
achieved remarkable results across different domains and industries where the data are highly convoluted,
scarce, sparse, and noisy. For instance, ML and DL models have been applied to genomics (Albaradei et
al., 2020; Albalawi et al., 2019; Ashoor et al., 2012), chemistry (Rahman et al., 2021; Soufan et al., 2018),
manufacturing (Ademujimi et al., 2017), and oil and gas (Aljubran et al., 2021; Magana-Mora et al., 2021;
Zhan et al., 2021; Chinthaka et al., 2020), among others. Similarly, different ML/DL approaches for bit
wear estimation have been proposed (Liu et al., 2018; Agostini and Sampaio, 2020; Wilson, 2018).
2 SPE-205627-MS

Although not applied directly to the estimation of bit wear, manufacturing applications can relate to the
problem. For instance, Qiao and colleagues (Qiao et al., 2020) proposed a multiscale convolutional long
short-term memory model (MCLSTM) to complete the monitoring task and a bi-directional LSTM model
(biLSTM) to complete the prediction task to successfully monitor general tool wear for manufacturing
processes. Additionally, studies with a direct focus on bit wear estimation have shown that DL models
can be applied to address this problem. Recently, Agostini and Sampaio (Agostini and Sampaio, 2020)

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/3-21APOG/D031S025R004/2495836/spe-205627-ms.pdf/1 by Shanghai Jiao Tong University user on 29 March 2024
proposed an ML model to aid drill bit wear assessment for real-time operations, using a probabilistic neural
network and Bayes theorem, combined with power spectral density (PSD) of surface torque image and
feature extraction by deep convolutional autoencoder (DCAE). Their results and analysis obtained from a
case study showed that a feature extraction based on DCAE is more efficient to discriminate the bit wear
state when compared with PSD torque image and raw torque data. However, most of the proposed methods
for bit wear estimation rely on traditional supervised ML/DL models that require hyperparameter tuning
and training processes using labeled data (i.e., bit wear grade), which are usually costly and time-consuming
to extract. Additionally, the bit wear grade data are sparse and only available at the end of each run.
In this study, we propose an unsupervised representation learning method to learn the mapping from high
dimensional observations (e.g., logging while drilling - LWD subsurface and surface drilling data) to lower-
dimensional representation (latent) space such that the original signals can be reconstructed from the latent
space and the downstream task (e.g., bit wear estimation) can be facilitated. Specifically, we developed a bi-
directional long short-term memory-based variational autoencoder (biLSTM-VAE) to project raw drilling
data into a latent space in which the real-time bit-wear can be estimated. We trained the biLSTM-VAE model
on subsurface/surface data from multiple runs. The original multivariate time series data were successfully
projected into a two-dimensional (2D) latent space. The results obtained showed a significant separation of
bit-wear states in the lower dimensional latent space, suggesting the feasibility of the real-time monitoring
and tracking of bit wear states in the representation space. The estimation is evaluated on both runs/wells
that the learning was conducted and those that were unseen during the learning phase. The proposed deep
neural network was trained using unsupervised learning techniques that do not require labels.
To the best of our knowledge, the proposed DL-assisted model for the real-time drill bit wear estimation is
one of the first to be trained in an unsupervised manner and to provide an end-to-end process. Moreover, the
commonly available data is selected to ensure their availability during real-time operations. Our approach
also introduces a new way of estimating bit wear based on the tracking of its trajectory in the latent space.
This helps improve the efficiency in drilling operations and can significantly affect the economics of well
engineering.

Methods and Approach

Background
LSTM is a variant of an artificial recurrent neural network (RNN) architecture that belongs to the category
of the DL algorithms. As opposed to RNN, LSTM intends to address the vanishing gradient problem
(Kalkatawi et al., 2019) and has feedback connections that can process both single data points (e.g., images)
and sequences of data (e.g., videos, speech, time series data). This network architecture is well-suited to
classifying and making predictions based on time series data.
An autoencoder is a type of artificial neural network that is designed for the task of representation learning
in an unsupervised manner. Autoencoders consist of two main components: 1) An encoder that maps the
input into the code in a latent space and 2) a decoder that maps the code to a reconstruction of the input (Fig
1). Autoencoders were designed to compressed knowledge representation of the original input and filtering
out the "noise". Variational autoencoder are generative models. As compared to discriminative autoencoders
that directly output values of latent codes in the representation space, variational autoencoder (VAE) will
SPE-205627-MS 3

instead formulate each dimension in the latent space into distribution function. VAEs aim to search for the
parameters of the distribution functions that describe the latent codes rather than directly predicting them.

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/3-21APOG/D031S025R004/2495836/spe-205627-ms.pdf/1 by Shanghai Jiao Tong University user on 29 March 2024
Figure 1—Schematic structure of a variational autoencoder

Approach
BiLSTM-VAE model architecture. In this study, we implemented a biLSTM-VAE to train a sequence of
5 ft data (e.g., formation, drilling, drilling distance) and map the sequential data into a representation space
or latent space. A downstream task for bit wear classification is later applied in the trained latent space.
The biLSTM-VAE model was designed using Tensorflow (Abadi et al., 2016) and Keras (Chollet, 2018)
frameworks. Fig 2. shows the derived model architecture. The input data are matrices with the size of (10
× 11). We used 11 feature vectors that represent rock formation, drilling parameters, and drilling distance
as inputs. We also selected 2D (z1, z2) for latent space, and the number of LSTM units is 128.
Model training. The model was trained for 300 epochs with a batch size of 32. The total loss function
to train the biLSTM-VAE model consists of Kullback-Leibler (KL) divergence (Hershey and Olsen, 2007)
and reconstruction loss. KL divergence is the statistical measure of how one distribution is different from
another. In VAE, let X and P(X) be the data and its probability that we want to model, z and P(z) be the
latent variable and its probability distribution, and P(X | z) be the distribution of the input data given latent
variable. The VAE was designed to infer P(z) from P(z | X) which is the probability distribution that projects
the input data into the latent space. Since P(z | X) is unknown, it was estimated by using a simpler parametric
distribution Q(z | X) (i.e., Gaussian distribution). The VAE was trained to capture the Q(z | X) such that it
is as close as possible to the actual distribution P(z | X). We used KL divergence to measure this distance.
We used Huber loss to measure the reconstruction loss (i.e., difference between reconstruction outputs
from the decoder and the original input data). It combines the best properties of L2 squared loss and L1
absolute loss by being strongly convex when close to the target and less sensitive to extreme values.
4 SPE-205627-MS

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/3-21APOG/D031S025R004/2495836/spe-205627-ms.pdf/1 by Shanghai Jiao Tong University user on 29 March 2024
Figure 2—Schematic architecture of our proposed bi-directional
LSTM Variational Autoencoder latent space learning from drilling data

Data Collection and Processing

We processed data from three different sources: formation data (wireline or lwd combined for training
purposes), drilling data (in this analysis depth data was used), and bit run record data (bit depth-in/out, dull
grade, etc. generally available from drilling reports).
All the input data were pre-processed before feeding to train the DNN model. The signal processing
criteria are described in Table 1. Fig 3 shows an example of the processed data.
SPE-205627-MS 5

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/3-21APOG/D031S025R004/2495836/spe-205627-ms.pdf/1 by Shanghai Jiao Tong University user on 29 March 2024
Figure 3—Overview of input data (i.e., formation data, drilling data, drilling distance) for seven runs. The 5th run was
used as test-set. The first and second value in the Dull notation represents the inner and outer wear, respectively.

Table 1—Data processing criteria.

Criteria Description

Data alignment All the input data were formatted into pandas dataframe in Python and aligned by the
common depth.
Interpolation of missing data Any missing data in each dataframe was filled by using linear interpolatin.
Outliers We used simple thresholding method to exclude extreme values in each signal.
Data scaling After pre-processing steps, formation data, drilling data, and bit dull data from seven
runs were combined into one big dataframe. We applied the Z-score normalization
(i.e., removing column mean and normalizing by standard deviation) on the combined
dataframe before training the neural network.
Rolling window Rolling windows (5 ft window size, 0.5 ft step size) were applied on the combined data
set. We hypothesized that the bit wear within 5 ft depth does not change substantially.
6 SPE-205627-MS

Results and Discussion

We trained the biLSTM-VAE model on data from seven runs. The original multivariate time series data were
successfully projected into a 2D latent space. Our results show a significant separation of bit-wear states
in the lower dimensional latent space, suggesting the feasibility of the real-time monitoring and tracking of
bit wear states in the representation space. The estimation is evaluated on both wells that the learning was
conducted and wells that were unseen during the training phase.

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/3-21APOG/D031S025R004/2495836/spe-205627-ms.pdf/1 by Shanghai Jiao Tong University user on 29 March 2024
Fig 4 A) shows the training loss against training epochs. The loss value starts at 45.3 at the beginning of
the training process and quickly decreases to approximately 28 at 100th epoch. At the end of the training
process (300th epoch), the loss value is stable at about 24.1. Fig 4 B) to E) illustrate the organization of
training data in the latent space at epochs 5, 100, 200, and 300, respectively. The bit-wear levels were
color coded from dark-green (new bit) to red (worn-bit). The distance between the new and the worn bits
in the latent space steadily increases during the training process as shown from Fig 4 B) to E). A clear
separation among clusters of bit-wear values is desired because it will enhance the accuracy of the bit-wear
classification algorithm in the downstream process.

Figure 4—A) Training loss across 300 training epochs. B) to E) The organization of trained latent space at different epochs

Fig 5 shows the mapping of both training and testing data in the latent space by our trained LSTM-VAE
model. The circle and triangle markers represent training and testing data, respectively. Bitwear values are
colored coded from dark-green for new bit to red for worn bit. We applied our model on the testing set for
SPE-205627-MS 7

an entire run. The gradient colors from white to black on the triangle markers represents the mapping of
testing data from the beginning to the end of the run. The mapping results suggest that the dril bit started
as new and gradually worn down at the end of the test run. At the end of the test run, the black triangle
markers concentrate around the 2-2 and 1-3 wear level clusters. The result is very close to the actual label
of the bit-wear at the end of the test run (2-2).

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/3-21APOG/D031S025R004/2495836/spe-205627-ms.pdf/1 by Shanghai Jiao Tong University user on 29 March 2024
Figure 5—Bit-wear latent space generated by trained biLSTM-VAE model for both train and test-set. The triangle markers
represent testing data. The gradient colors from white to black encode time data from the beginning to the end of the test run.

Conclusions
A biLSTM-VAE was developed to project raw drilling data into a latent space in which the real-time bit-
wear can be estimated. The proposed deep neural network was trained in an unsupervised manner, and the
bit-wear estimation is an end-to-end process. The original multivariate time-series data from multiple runs
were successfully projected into 2D latent space. Our results show a significant separation of bit-wear states
in the lower dimensional latent space. We evaluated the bit-wear estimation using the data from a bit run
that was not used during the learning phase, showing a very clear separation in the reduced latent space
with accurate estimations relative to the actual label of the bit-wear at the end of the test run. From these
results, we demonstrated the feasibility of the real-time monitoring and tracking of bit wear states in the
representation space.
8 SPE-205627-MS

Acknowledgements
The authors would like to thank Saudi Aramco and Baker Hughes for supporting the endeavor of research
and technology development and for granting permission to publish this paper.

References
Abadi, Martín, Barham, Paul, Chen, Jianmin et al. 2016. Tensorflow: A system for large-scale machine learning. Proc.,

Downloaded from http://onepetro.org/SPEAPOG/proceedings-pdf/21APOG/3-21APOG/D031S025R004/2495836/spe-205627-ms.pdf/1 by Shanghai Jiao Tong University user on 29 March 2024
12th USENIX symposium on operating systems design and implementation (OSDI 16)265–283.
Ademujimi, Toyosi Toriola, Brundage, Michael P, and Prabhu, Vittaldas V. 2017. A review of current machine
learning techniques used in manufacturing diagnosis. Proc., IFIP International Conference on Advances in Production
Management Systems407–415.
Agostini, Cristiano Eduardo and Sampaio, Marcio Augusto. 2020. Probabilistic Neural Network with Bayesian-based,
spectral torque imaging and Deep Convolutional Autoencoder for PDC bit wear monitoring. Journal of Petroleum
Science and Engineering 193: 107434.
Albalawi, Fahad, Chahid, Abderrazak, Guo, Xingang et al. 2019. Hybrid model for efficient prediction of poly (A) signals
in human genomic DNA. Methods 166: 31–39.
Albaradei, Somayah, Magana-Mora, Arturo, Thafar, Maha et al. 2020. Splice2Deep: An ensemble of deep convolutional
neural networks for improved splice site prediction in genomic DNA. Gene: X 5: 100035.
Aljubran, Mohammad, Ramasamy, Jothibasu, Bassam, Mohammed et al. 2021. Deep Learning and Time-Series Analysis
for the Early Detection of Lost Circulation Incidents During Drilling Operations. IEEE Access.
Ashoor, Haitham, Mora, Arturo M, Awara, Karim et al. 2012. Recognition of translation initiation sites in Arabidopsis
thaliana. In Systemic approaches in bioinformatics and computational systems biology: recent advances, 105–116.
IGI Global.
Chinthaka, Gooneratne, Arturo, Magana-Mora, and Otalvora, Contreras William. 2020. Drilling in the Fourth Industrial
Revolution-Vision and Challenges. IEEE Engineering Management Review.
Chollet, François. 2018. Keras: The python deep learning library. Astrophysics Source Code Library: ascl: 1806.022.
Hershey, John R and Olsen, Peder A. 2007. Approximating the Kullback Leibler divergence between Gaussian mixture
models. Proc., 2007 IEEE International Conference on Acoustics, Speech and Signal Processing-ICASSP'074, IV-317-
IV-320.
Kalkatawi, Manal, Magana-Mora, Arturo, Jankovic, Boris et al. 2019. DeepGSR: an optimized deep-learning structure
for the recognition of genomic signals and regions. Bioinformatics 35 (7): 1125–1132.
Liu, Yu, Kibbey, Justin, Bai, Yanbin et al. 2018. Real-time bit wear monitoring and prediction using surface mechanics data
analytics: a step toward digitization through agile development. Proc., IADC/SPE Drilling Conference and Exhibition.
Magana-Mora, Arturo, Affleck, Michael, Ibrahim, Mohamad et al. 2021. Well Control Space Out: A Deep-Learning
Approach for the Optimization of Drilling Safety Operations. IEEE Access.
Qiao, Huihui, Wang, Taiyong, and Wang, Peng. 2020. A tool wear monitoring and prediction system based on multiscale
deep learning models and fog computing. The International Journal of Advanced Manufacturing Technology 108:
2367–2384.
Rahman, MM, Vadrev, SM, Magana-Mora, A et al. 2021. FDMine: a graph mining approach to predict and evaluate
food-drug interactions.
Soufan, Othman, Ba-Alawi, Wail, Magana-Mora, Arturo et al. 2018. DPubChem: a web tool for QSAR modeling and
high-throughput virtual screening. Scientific reports 8 (1): 1–10.
Wilson, Adam. 2018. Real-Time Data Analytics Allows for Bit-Wear Monitoring and Prediction. Journal of Petroleum
Technology 70 (12): 67–68.
Zhan, Guodong David, Magana-Mora, Arturo, Moellendick, Eric et al. 2021. Hybrid Physics-Field Data Approach
Improves Prediction of ROP/Drilling Performance of Sharp and Worn PDC Bits. Proc., International Petroleum
Technology Conference.