You are on page 1of 9

SPE-206254-MS

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/21ATCE/1-21ATCE/D011S015R005/2489891/spe-206254-ms.pdf/1 by Rajiv Gandhi Institute of Petroleum Technology user on 04 May 2023
Dull Bit Grading Using Video Intelligence

Merit P. Ekeregbe, Nigerian Petroleum Development Company/University of Houston; Mina S. Khalaf, University of
Houston; Robello Samuel, Halliburton

Copyright 2021, Society of Petroleum Engineers

This paper was prepared for presentation at the 2021 SPE Annual Technical Conference and Exhibition held in Dubai, UAE, 21 - 23 September 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
Although visual data analytics using image processing is one of the most growing research areas today and
is largely applied in many fields, it is not fully utilized in the petroleum industry. This study is inspired
by medical image segmentation in detecting tumor cells. This paper uses a supervised Machine Learning
technique through video analytics to identify bit dullness that can be used in the drilling industry in place
of the subjective screening approach. The evaluation of bit performance can be affected by subjective
evaluation of the degree of dullness. The present approach of using video analytics is able to grade bit
dullness to avoid user subjectivity.
The approach involves the use of datasets in good quantity and quality by separating them into training
datasets, testing datasets, and validation datasets. Due to the large datasets, Google Collaboratory was used
as it provides access to its Graphic Processing Unit (GPU) online for the processing of the bit datasets.
The processing time and resource consumption are minimized using Google GPU. Using the Google GPU
resources, the procedure is automated without any installation. After the bit is pulled out and cleaned, a video
is taken around and up and down in 360°. Further, it is compared against the green bit. By this approach,
multiple video datasets are not required.
The algorithm was validated with new sets of bit videos and the results were satisfactory. The
identification of the dullness or otherwise of each screened bit is done with the aid of a bounding box with
a stamp of a level of confidence (range 0.5–1) and the algorithm assigns for its decision on the identified
or screened object. This method is also able to screen multiple bits stored in a single place. In an event
where several drill bits are to be screened, manual grading will be a huge task and will require a lot of
resources. This model and algorithm will take a few minutes to screen and provide grading for several bits
while videos are passed through the algorithm. It has also been found that the grading with the video was
much better than the single image as the contextual information extracted are much higher at the level of
the entire video, per segment, per shot, and per frame. Also, methodology is made robust so that the video
model test starts successfully without error. The time penalty for the processing is fast and it took less time
for a single video screening.
The work developed here is probably the first to handle the dull bit grading using video analytics. With
more of these datasets available, the future automation of the IADC bit characterization will soon evolve
into an automated process.
2 SPE-206254-MS

Introduction
Machine Learning (ML) has become the new direction for doing things more easily while also improving the
accuracy of results. The uses of machine learning span from health to sport and to Oil and Gas. In Oil and Gas
seismic processing, machine learning has also improved well log data interpretation as a way of extending

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/21ATCE/1-21ATCE/D011S015R005/2489891/spe-206254-ms.pdf/1 by Rajiv Gandhi Institute of Petroleum Technology user on 04 May 2023
the interpretation of a section of well log by compensating for incomplete values and providing full potential
of the well logs; and as a result, bringing the time needed for a well log interpretation down from 7 days of
human effort to 7 min of automated computation with minimum human input using ML (Nordloh & Roub
2020). Machine learning has also provided security coverage of Oil and Gas assets. A novel ML algorithm—
the deviant learning algorithm—has been tested in performing predictive classification of onshore pipeline
incidence/threat datasets and has proven to be successful (Osegi et al., n.d.). Some other operators utilize
magnetic flux leakage sensors to probe Oil and Gas pipelines to localize and size different defect types
(Mohamed & Hamdi, 2015). In addition, machine learning techniques have many applications in production
data analysis (Li et al., 2021) and well testing (Carpenter, 2016).
One major area of machine learning is in the area of big data analysis. Supervised and unsupervised
machine learning has the potential to greatly help with the analysis of large data volumes (Odi et al., 2018).
One strategy in big data analysis is the training of dataset using linear regression to predict and forecast
future possibilities when the algorithm is trained. One of the applications of machine learning is in reservoir
engineering (Chaki et al., 2020) and this has also broadened its application in the Oil and Gas industry. Using
machine learning algorithms, we can build a reservoir model by learning the historical pattern of bottom
hole pressure and flow rate (or training set) without the physics of the flow being programmed explicitly
(Ristanto, 2018). This is possible by using historical flow rate and bottom-hole pressure and the model will
be able to detect any deviation and track any multiphase flow problems.
The drilling industry is not without some touches of this shift in the application of ML in problem-solving.
The evaluation of bit performance under down-hole conditions was investigated by many researchers
(Macini et al., 2005–2007). Deep learning techniques have been used considerably in drilling engineering
to perform different data analytics (Ding & Samuel, 2016; Lakhanpal & Samuel, 2017; Mirani &
Samuel, 2016). Machine learning was also used in determining wellbore stability. The prediction of
mechanical characteristics of the reservoir formations, such as static Young’s modulus, is very important
for the evaluation of the wellbore stability and development of the geomechanical earth model. Elasticity
considerably varies with the change in the lithology (Mahmoud et al., 2020). The ML has become an easier
way of doing things and its application in many areas of drilling can be explored. One of such thing is in
the dull bit grading for which this paper is developed.

Bit Dullness
Dull bit grading is defined as the evaluation/description of dull conditions (of teeth, bearing, and gage) of
the used drill bits. Dull grading of rock bits is essential for well planners as well as bit manufacturers. It helps
field personnel, responsible for preparing well programs, select proper bits to drill efficient and economic
wells. The technology developer uses dull grading to design better tools for drilling.
SPE-206254-MS 3

In the past, evaluating dull bits (to decide what type of bit to run next) was a skill that a driller would
have through experience. Moreover, this was a part of art that distinguishes a driller from others. The dull
grading system standardization emerged to make it easier for drilling personnel to learn how to identify bit
dullness and have same universal interpretation once the conditions are equally understood. The evaluation
of bit performance can be affected by subjective evaluation of the degree of dullness but automation using

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/21ATCE/1-21ATCE/D011S015R005/2489891/spe-206254-ms.pdf/1 by Rajiv Gandhi Institute of Petroleum Technology user on 04 May 2023
machine learning makes the most objective of all decisions. The motivation of this project is to present an
automated technique able to detect bit dullness to defeat the unavoidable subjective and relative Engineers’
judgment.
Over the last decades, a considerable amount of bit dullness data was recorded. The data can be
incorporated with the advances in computers to present an efficient technique able to identify bit dullness
and advance further in classifying them according to IADC grading system. This may be accomplished
by developing a programming code that utilizes visual analytics of videos to detect bit dullness. There are
several robust libraries that makes ML approach to bit dullness identification an easy task and the model
should be sufficiently trained with datasets.

Methodology
The You Only Look Once (YOLO) algorithm is an upgraded version of the object detection model with a
higher detection speed. It is modeled after the human–object detection process-humans look at an object
once and identify the same one as the name implies. YOLO captures the overall image once by using
convoluted neural network. The model uses an S × S grid to locate the image as the first step as depicted in
Fig. 2. The YOLO system (1) resizes the input image to 448 × 448, (2) runs a single convolutional network
on the image, and (3) thresholds the resulting detections by the model’s confidence (After, Redmon et al.,
n.d.) The object detection is pivoted at the center of the object in the grid block and if the center of the object
is in the center of grid cell, that cell enables the object detection.

Fig. 1—YOLO detection and processing system


4 SPE-206254-MS

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/21ATCE/1-21ATCE/D011S015R005/2489891/spe-206254-ms.pdf/1 by Rajiv Gandhi Institute of Petroleum Technology user on 04 May 2023
Fig. 2—The YOLO model

The YOLO algorithm uses the neural network which is trained using the loss function. The YOLO
algorithm uses the loss function to minimize the error and scale its performance. The loss function is shown
in Equation 1 (Redmon et al., n.d.).
The loss function evaluates the performance of the algorithm detecting the object using the grid size, S
with the spatial coordinates x, y on the grid block (number of cells); B number of bounding boxes, confidence
which is the intersection of the ground truth and the model bounding boxes on space with object and space
without object within the grid and class probability. The loss function also includes the width and the height
of the object as stated in Equation (1). Confidence prediction represents the Intersection over Union (IOU)
between the predicted box and any ground truth box (Redmon et al., n.d.). The uniform sizing of the object
to be detected is important for the algorithm to do one-off full image detection. If the object fed into the
YOLO model is greater than the model required size or the average size that the algorithm was trained with
(as was discovered using the algorithm, without it being resized) the algorithm will only be considering the
cutting structure rather than the whole bit structure (see Fig. 3). Because of the bit bigger than the required
image size, the bounding and confidence are placed on the structure instead of the whole bit.

(1)
SPE-206254-MS 5

where
C = confidence in the bounding box
xi, yi = variables to define at the center of the bounding box
wi, hi = dimensions of the bounding box
λcoord, = manipulation to increase or decrease the emphasis of the bounding box

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/21ATCE/1-21ATCE/D011S015R005/2489891/spe-206254-ms.pdf/1 by Rajiv Gandhi Institute of Petroleum Technology user on 04 May 2023
λnoobj

Fig. 3—Oversized input image

Model Implementation
All machine learning and AI use different methods based on the expectation from the processed data.
Supervised and unsupervised learning deal with prior classified and unclassified datasets, respectively. The
classified dataset provides training of the algorithm to learn intelligence by the pattern and classification
of the structure and nature of the dataset. The unsupervised deals with the algorithm learning unstructured
datasets. All supervised learning problems are classification and/or regression problems. The algorithm
performs a regression calculation on the data from the training and on the strength of the model parameter
able to predict accurately.

Fig. 4—Ungraded bits (1) and graded bit (2–4) with bounding boxes (model input and output
dataset) Courtesy: YouTube (Ungraded-without bounding boxes-drill bits 1–3) and Field (4)
6 SPE-206254-MS

The model development is implemented by splitting the object videos or images into the formative set
for model development and the summative set for model testing. The formative dataset is further split into
training data to compute the model parameters and the validation dataset to evaluate the model capacity.
During the training stage, the images are labeled manually as "dull bit" or "fully graded."
The accuracy of the model is dependent on the number of the dataset used to train the algorithm. For this

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/21ATCE/1-21ATCE/D011S015R005/2489891/spe-206254-ms.pdf/1 by Rajiv Gandhi Institute of Petroleum Technology user on 04 May 2023
implementation of drill bit grading algorithm, above 1000 images were used for the training of the model
and a further 1000 images were used for validation of the model. The larger the datasets, the better the
accuracy. It means that the model can be more robust as the dataset increases.
The algorithm performance indicators are as depicted in Fig. 5. Of most important to the reader is the
bounding box and the precision values. The algorithm has a confidence level of 50–95% next to the bounding
box in its prediction.

Fig. 5—Algorithm performance indicators

One beautiful thing about the supervised learning approach in bit grading is that grading will not be
restricted to any certain type of bit as it is with the IADC grading technique. The different type of bit datasets
used to train and make the algorithm demonstrate unlimited scope in bit grading irrespective of the type
of bit.

Algorithm Performance
The model performance is depicted in Fig. 7. The precision of the model is determined by the percentage
of model prediction that is correct during the testing phase of the model. It is the ratio of total number of
true dull bits detected by the model to the total number of dull bits that the algorithm recognized or detected
throughout the screening. Recall is the ratio of the total true dull bits detected by the algorithm to the total
number of dull bits (detected and undetected).
Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/21ATCE/1-21ATCE/D011S015R005/2489891/spe-206254-ms.pdf/1 by Rajiv Gandhi Institute of Petroleum Technology user on 04 May 2023
7

Fig. 6—Graphical definition of recall and precision (Source: HPE Intro 2 ML)

Fig. 7—Model performance


SPE-206254-MS
8 SPE-206254-MS

Machine Learning approach is a reliable technology to automate and evaluate bit performance to avoid
being affected by human subjectiveness. In addition, ML techniques can evaluate a lot of bits most
effectively without the rigor of manual determination and that saves time and energy. With advancement
in technology, incorporation of real-time and down-hole continuous dullness monitoring will provide an
effective drilling practice and ensures cost saving. The birth of such a process starts with this model.

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/21ATCE/1-21ATCE/D011S015R005/2489891/spe-206254-ms.pdf/1 by Rajiv Gandhi Institute of Petroleum Technology user on 04 May 2023
A supervised learning project requires a lot of datasets for the training of the algorithm which is sometimes
not easily available for use. There is no single repository of bit records for the drilling industry where such
data can be available for the research. The getting of datasets for this purpose is a rigorous exercise. Due
to the limited dataset and time, the study was restricted to bit recognition, and bit dullness and full-grade
identification.

Mode of Algorithm Implementation


Jupiter Notebook and Google Collaboratory are major modes that easily provide access to implement the
algorithm by using online resources. The Google Collaboratory easily provides syntax that is easy to read,
not machine language and easy to implement. Here, we take advantage of Google Collaboratory to process
large datasets (video of drill bits) in a fast and efficient way. Google Collaboratory provides access to
its Graphic Processing Unit (GPU) online for the processing of the bit datasets. The processing time and
resource consumption are minimized using Google GPU. Google Collaboratory helps to utilize online
resources by using the available GPU of Google

Conclusion
Used drill bit grading is a process that requires skilled personnel to evaluate the bit dullness. Sometimes,
the dull feature detection can be impacted by personal subjectivity. In addition, in an event where several
drill bits are to be screened, manual grading will be a huge task and will require a lot of resources. The
development of a machine learning approach to drill bit dullness identification is a better approach to bit
grading in a most non-subjective and time and energy-saving manner. In this work, we develop an automated
technique to recognize drill bits (roller cones) in video and identify the dullness using the YOLO algorithm
which works like the human–object detection. The developed approach takes a few minutes to screen and
provide grading for several bits in a moment that videos are passed through the algorithm. The developed
model is a supervised ML technique written in Python but implementation is done using online resources
like the Google Collaboratory Notebook. The developed model was validated with drill bit videos and the
results were very good with a confidence range of 0.5–0.95.

References
Carpenter, C. (2016). Machine Learning Applied to Multiwell-Test Analysis and Flow-Rate Reconstruction. Journal of
Petroleum Technology, 68(05), 76–78. https://doi.org/10.2118/0516-0076-jpt
Chaki, S., Zagayevskiy, Y., Shi, X., Wong, T., & Noor, Z. (2020, January). Machine Learning for Proxy Modeling of
Dynamic Reservoir Systems: Deep Neural Network DNN and Recurrent Neural Network RNN Applications. Day 2
Tue, January 14, 2020. https://doi.org/10.2523/IPTC-20118-MS
Ding, B., & Samuel, R. (2016). Visual Data Analytics : Morphological Image Processing Approach on the Evaluation of
Foamed Cement Quality. https://doi.org/10.2118/181476-MS
Lakhanpal, V., & Samuel, R. (2017). Real-Time Bit Wear Prediction Using Adaptive Data Analytics. https://
doi.org/10.2118/187445-MS
Macini, P., Magagni, M., Da Dalt, G., & Valente, P. (2006). How minimum interval concept can improve bit performance
evaluation.
Macini, P., Magagni, M., Da Dalt, G., & Valente, P. (2007). Bit performance evaluation revisited by means of bit index
and formation drillability catalogue. Proceedings of the SPE/IADC Middle East Drilling Technology Conference and
Exhibition, 150–156. https://doi.org/10.2118/107536-ms
SPE-206254-MS 9

Macini, P., Magagni, M., & Valente, P. (2005). Drill-bit catalog and bit index: A new method for bit performance
evaluation. https://doi.org/10.2118/94798-ms
Mahmoud, A. A., Elkatatny, S., & Shehri, D. Al. (2020). Application of Machine Learning in Evaluation of the Static
Young’s Modulus for Sandstone Formations.
Mirani, A., & Samuel, R. (2016). Mitigating Vibration Induced Drillstring Failures Using Data Analytics : Workflow and

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/21ATCE/1-21ATCE/D011S015R005/2489891/spe-206254-ms.pdf/1 by Rajiv Gandhi Institute of Petroleum Technology user on 04 May 2023
Case Study. https://doi.org/10.2118/178849-MS
Mohamed, A., & Hamdi, M. S. (2015). 2015 3rd International Conference on Future Internet of Things and Cloud. https://
doi.org/10.1109/FiCloud.2015.54
Nordloh, V. A., & Roub, A. (n.d.). Machine Learning for Gas and Oil Exploration.
Odi, U., Nguyen, T., & Energy, D. (2018). Geological Facies Prediction Using Computed Tomography in a Machine
Learning and Deep Learning Environment. 1–11. https://doi.org/10.15530/urtec-2018-2901881z
Osegi, E. N., Technology, C., State, L., & State, R. (n.d.). p-DLA: A Predictive System Model for Onshore Oil and Gas
Pipeline Dataset Classification and Monitoring - Part 1.
Ristanto, T. (2018). Machine Learning Applied to Multiphase Production Problems. June. Redmon, J., Divvala, S.,
Girshick, R., & Farhadi, A. (n.d.). You Only Look Once : Unified, Real-Time ObjectDetection.

You might also like