Вы находитесь на странице: 1из 14

URTeC: 253

Development and Application of a Real-Time Drilling State


Classification Algorithm with Machine Learning
Yuxing Ben*, Chris James, Dingzhou Cao; Anadarko Petroleum Corporation.
Copyright 2019, Unconventional Resources Technology Conference (URTeC) DOI 10.15530/urtec-2019-253

This paper was prepared for presentation at the Unconventional Resources Technology Conference held in Denver, Colorado, USA,
22-24 July 2019.

The URTeC Technical Program Committee accepted this presentation on the basis of information contained in an abstract submitted
by the author(s). The contents of this paper have not been reviewed by URTeC and URTeC does not warrant the accuracy, reliability,
or timeliness of any information herein. All information is the responsibility of, and, is subject to corrections by the author(s). Any
person or entity that relies on any information obtained from this paper does so at their own risk. The information herein does not
necessarily reflect any position of URTeC. Any reproduction, distribution, or storage of any part of this paper by anyone other than the
author without the written consent of URTeC is prohibited.

Abstract
A fundamental component of a real-time drilling analytics system is automatic rig state detection. High
frequency time series data (typically one data point per second) from multiple sensors on a drilling rig is
processed and labeled with drilling states including: slide drilling, rotate drilling, pick up, in slips, and
others. With labeled time-series data, the real-time system can derive operational KPIs (key performance
indicators) with extremely high resolution, e.g., a statistical summary of rotary versus slide drilling time
for the rig supervisor and drilling engineer to analyze efficiency. Later, such information can be leveraged
to develop algorithms to detect abnormal drilling events and drive closed loop control.

A workflow was developed to clean and fill in any missing data. A rules-based model was then applied to
classify the data into seventeen rig states. For the state “drilling”, a sub-classification was made to label
rotate drilling and slide drilling. However, it is difficult to categorize “slide drilling” solely based on
surface RPM due to top drive oscillation. In order to achieve acceptable accuracies, three machine
learning models to classify “rotate drilling” and “slide drilling” were evaluated: Random Forest,
Convolutional Neural Network (CNN), and a hybrid Convolutional Neural Network / Recurrent Neural
Network (CNN/RNN).

Machine learning models were built for two basins, one model each, to accommodate different drilling
styles. For the Delaware Basin, 10 wells with 9 million rows of data were chosen, and for the DJ Basin,
12 wells with 2 million rows of data were chosen. A legacy, rules-based algorithm was applied to label
each row as rotate or slide drilling, and the misclassified records were manually corrected. The machine
learning models were found to be far superior to rules-based models. For the wells in the training set, the
accuracies of our rules-based models were 70% and 90% respectively, while the accuracies of our
machine learning models were over 99%. The CNN model was proven to be the best model, excelling
with high accuracy, short computation time, and scalability for big data applications.

The data cleaning, preprocessing, and machine learning algorithm has been deployed in Anadarko’s Real-
Time Drilling (RTD) ecosystem (Cao et al., 2018, 2019), which consists of four layers: a data source,
analytics, data storage, and UI layer. KPIs, directional statistics, and engineering models are calculated in
real-time and visualized through a web-based UI. This system can be accessed by any member of the
drilling operations team. The system is regularly used to evaluate, compare, and optimize well
performance. Future plans include pushing analytical models to the rig site with edge computing to
URTeC 253 2

facilitate drilling guidance and levels of automation. To our knowledge, this is the first time that a deep
learning model has been used to analyze drilling time series data in a production real-time system.

Introduction
There are few published papers on rig state classification. Dunlop et al. (2006) filed a patent on rig state
detection that makes use of probabilistic approach and in particular, is based on particle filtering
techniques. Veres and Sabeur (2015) investigated several machine learning algorithms to classify 10
defined rig states, and the accuracy of the best classifier was greater than 90% for states such as: Drilling
Rotate, Drilling Slide, and Make Connection. Arnaout et al. (2012, 2014) used discrete polynomial
moments and a distributed recognition system to classify drilling operations and five drilling states,
achieving an average accuracy of 90%. Recently, Coley (2019) built a rig state classifier using supervised
machine learning to support invisible lost time analysis. Machine learning in drilling has predominantly
found applications in anomaly detection. For example, Pournazari et al. (2015) applied SAX to enhance
kick detection; Ambrus et al. (2017) used a Bayesian Network to detect drilling dysfunction.

There has been a considerable amount of work on time series classification using machine learning/deep
learning outside of the drilling domain, especially in the past several years. Bagnall et al. (2017) reviewed
and evaluated time series classification algorithms published from 2011 to 2016, but none leveraged deep
learning. More recently, Fawaz et al. (2018) reviewed deep learning applications in time series
classification. Rajpurkar et al. (2017) developed an algorithm which exceeded the performance of board
certified cardiologists in detecting a wide range of heart arrhythmias from electrocardiograms with a CNN
model. However, the previously mentioned time series classification problems leveraged either the entire
time series, intervals within the series, the pattern represented by the whole time series, or the whole
interval corresponding to a single class.

The purpose of this work is to analyze the data stream coming from various rig sensors each second, and
develop a model that can detect the drilling rig state automatically, in real-time. Hence, the model has to
be able to satisfy the following conditions:

1) The model can only use data up to the current time step;
2) The inference of the model should be fast enough to not cause congestion within the real-time
system. In other words, the model execution time should be less than the real-time data arrival
interval. For example, in the case of 1Hz data for one well in the system, the execution time must
be less than one second;
3) If a window of previous data is used in the model, the window cannot be excessively large due to
limitations of buffering in the real-time system.
In the following sections, a detailed workflow will be presented, describing how machine learning was
applied to increase the detection accuracy of two of the most important drilling states for onshore
unconventional wells (rotate drilling and slide drilling), enabling workflows that allow drilling teams to
capture significant business value.

Data Preprocessing
In this work, the authors focused on two onshore basins where Anadarko (APC) has drilled hundreds of
wells. Offshore wells typically use rotary steerable systems, and, as such, would not include intervals of
slide drilling. The drilling time-series data preprocessing workflow included several steps:

1) Data quality control to ensure that there are minimal calculation errors. For example, if there is a
missing value, interpolation is used to fill the gap. Figure 1 shows the time series channels used
to classify rig states.
URTeC 253 3

2) Rig state components calculation. These components identify the behavior of major rig systems
and high-level operational states. Figure 1 shows each of the rig state components. Each
component relies on thresholds determined by engineering judgement, but generally apply to
most wells without the need for active modification. One exception is the rig-specific hook load
threshold required for Slip Status.
3) The rig state is classified using additional logic applied to the rig state components for every
second of drilling data. The 17 individual states are shown in Figure 1.
4) Classify “drilling” as rotate or slide drilling. Regarding the “drilling” state, determination of
rotate/slide drilling becomes very difficult with modern drilling systems and techniques. In Figure
2, the time series data for RPM for a typical well is shown: the blue dots represent “slide
drilling”, while the grey dots represent “rotary drilling”. In a classical sense, for RPM = 0 the
drilling state is slide, while for RPM>=0 the drilling state is rotate. However, due to surface RPM
reductions while rotary drilling caused by auto driller torque limits, the use of top drive oscillators
to reduce static friction while sliding, and intermittent RPM applied to re-orient the motor while
sliding, a static RPM threshold to determine the drilling type has been observed to be unreliable.
In fact, the best rules-based model the authors developed could only achieve 70% accuracy for
the Delaware basin and 90% accuracy for the DJ basin when left unsupervised. Hence, several
machine learning models to classify rotate and slide drilling were evaluated.
URTeC 253 4

Figure 1. Workflow for rig state identification.

Figure 2. RPM vs time plot for a typical well


URTeC 253 5

To prepare training data for the machine learning algorithms, the following tasks were performed:

1) Data labeling: A rules-based algorithm was applied to label each drilling row as rotate or slide,
then the misclassified rows were corrected manually. As with any machine learning problem,
quality of the data is critical to model success. The importance of carefully labeling data and
allocating adequate time for this exercise cannot be understated.
2) Feature selection:
a. Using lessons learned from the manual labeling exercise, surface RPM and torque were
chosen as the main two features.
b. Given that the use of top drive oscillation and commonly observed RPM reductions due
to surface torque limits were observed to occur in specific hole sections (i.e. vertical,
curve, or lateral), wellbore trajectory data was used to label the hole section and used as a
feature; although, as covered in the proceeding sections, this feature was eventually
removed.
c. As shown in Figure 3, the preceding 20 data points (seconds, in this case) of a particular
drilling window (i.e. a continuous interval of on-bottom drilling time) were used to
determine the drilling state. If 20 seconds of drilling in a given window had not elapsed,
data was repeated until 20 seconds had passed.

Figure 3. Using a moving window to select data and construct the ML models.

10 wells from the Delaware Basin and 12 wells from the DJ basin were selected across multiple rigs for
the training data set, totaling 11 million rows of data. A dataset 𝐷𝐷 = {(𝑋𝑋1 , 𝑌𝑌1 ), (𝑋𝑋2 , 𝑌𝑌2 ), … , (𝑋𝑋𝑁𝑁 , 𝑌𝑌𝑁𝑁 ) was
generated with a collection of pairs (𝑋𝑋𝑖𝑖 , 𝑌𝑌𝑖𝑖 ) , where the input 𝑋𝑋𝑖𝑖 included the last 20 points of rotary
torque, the last 20 points of RPM, and the hole section, and the output 𝑌𝑌𝑖𝑖 was the labeled drilling state at
time 𝑡𝑡. The task of the machine learning model was to train a classifier on this dataset to a probability
distribution over the labeled drilling state (rotate or slide).
URTeC 253 6

Methods
Random Forest (RF), Convolutional Neural Network (CNN), and hybrid Convolutional Neural Network /
Recurrent Neural Network (CNN/RNN) models were trained and evaluated. The ML architectures are
shown in Figure 4. Scikit-learn, a machine learning library in Python, was used to implement the RF
model. TensorFlow, specifically tensorflow.keras, a library by Google, was used to implement the CNN
and CNN/RNN models. All three approaches generated a distribution of probability of 0 to 1, and the
threshold was set at 0.5 for one class.

Figure 4. ML/DL Architectures.

For each basin, one well was randomly selected as the test well, while the rest of the wells were used for
training. The training data set was split into training and validation sets in order to facilitate a reliable,
accurate model without overfitting. A 32 core CPU was used to train the RF models, while a single
NVIDIA GPU was used to train the CNN and hybrid CNN/RNN models.

Results
The machine learning models were far superior to previously used rules-based methods. Figure 5 shows
the results for each basin across each of the models. The leftmost table shows the results for the wells in
the Delaware Basin. The prediction accuracy of the RF model was promising, but two outliers testing
below 90% accuracy raised concerns about the model’s potential accuracy in the production system. Both
the CNN and CNN/RNN models had consistent results, at roughly 99% accuracy. The rightmost table in
Figure 5 shows the results for wells in the DJ basin. In this case, the RF model produced more consistent
results, with only one well falling below 99% accuracy. The resulting differences suggest that deep
learning models are capable of handling more data and complex drilling patterns.
In the first production release, RF models were used for both basins to facilitate ease of deployment.
However, as the production results were analyzed over time, challenges arose in both input data and
model accuracy. Eventually, a universal CNN model was developed to classify drilling states across all
basins. Figure 6 shows the accuracy of the universal CNN model for both basins’ test wells.
URTeC 253 7

Figure 5. The left shows the results for wells in the Delaware Basin, while the right shows the results for wells in the DJ Basin.

Figure 6. The results of universal CNN model for both Basins’ wells.

Model Deployment, Evolution, and Lessons Learned


Inference speed of a model in real-time is a key consideration when choosing which model to deploy. The
inference of the hybrid CNN/RNN model was slower than the CNN model. Thus, only the RF and CNN
models were chosen as deployment candidates in the Real-Time Drilling (RTD) analytics system. Figure
7 shows an overview of the RTD system, which consists of four layers: data source, analytics, data
storage, and UI. The data source layer includes streaming time series data and static data sources such as
an Enterprise Data Management (EDM) database and digitalization transformation applications that are
built internally. The analytics layer includes the rig state classification logic, KPI aggregations, directional
calculations, and other advanced analytics, such as physics-based models and drilling dysfunction
URTeC 253 8

detection algorithms. For a more detailed description of the RTD system, please refer to Cao et.al (2018,
2019).

Figure 7. Architecture of Real-Time Drilling Analytics System.

As previously stated, when the models were deployed into production, many challenges arose. Five
versions of the drilling state models have been developed to date, as highlighted in Figure 8.
1) Rules-Based (Delaware and DJ)
Originally, a basic, rules-based methodology was deployed in StreamBase to classify rotate and slide
drilling. Torque and RPM variance were analyzed over the preceding 20 seconds, and if/then thresholds
were assigned by basin. However, poor performance (~70% accuracy) and the need to continuously
monitor and manage thresholds necessitated improved methods.
2) Random Forest (Delaware) / Random Forest (DJ)
Illustrated in Figure 5, the accuracy of the RF model for the DJ basin was very high, but some concerns
arose over the outliers in the Delaware Basin test set. However, for ease of deployment, the RF models
were deployed to production. To deploy, the two RF models were packaged as pickle files within
StreamBase and inferred by the Python adapter. Inference occurred every second, using time-series
drilling data, the preceding 20 seconds of buffered data, and the wellbore section label. As predicted,
accuracy was much better than the rules-based model, but a more robust classifier was required by end
users.
3) CNN (Delaware) / Random Forest (DJ)
As the accuracy of the model for Delaware Basin was not ideal using traditional machine learning
methods (the RF model), a deep learning CNN model was chosen to enhance accuracy. When considering
system complexity, deploying a combination of RF models and CNN models into the RTD system
(without CI/CD pipelines) did not proceed. Model accuracy results in the Delaware Basin inspired
confidence that a CNN model would also work well for the DJ Basin, so production deployment of both
RF and CNN models was skipped in favor of training a CNN model for the DJ Basin.
4) CNN (Delaware) / CNN (DJ)
URTeC 253 9

As predicted, the CNN model test results for the DJ Basin were more accurate than the RF model. Thus
both CNN models were selected for production deployment. During this time period, another issue arose
related to model features. Wellbore section (vertical, curve, and lateral) was previously being used as a
model input. This led to a dependency on the planned wellbore trajectory being entered into the RTD
system before real-time data began streaming. When the wellbore trajectory was not passed into the RTD
system, there would not be a rotate or slide drilling classification. To overcome this obstacle, the models
were retrained without that feature, and the accuracies remained largely unchanged from the original
CNN test results. Therefore, both CNN models, without wellbore section as a feature, were deployed to
production. Over time, continued misclassification was observed, and the introduction of operations in a
new basin, Powder River, inspired further model evolution.
5) Universal CNN (DJ, Delaware, and Powder River)
To overcome the latest round of challenges and further decrease the complexity of the RTD system, a
combined CNN model was trained. However, in an effort to further improve accuracy, the size of the
labeled, training data set was doubled to include at least one well from every rig in APC’s fleet, as well as
including data from the wells drilled in the Powder River Basin. The test accuracies were largely the same
as those from the previous, basin-specific CNN models. With a newly developed CI/CD pipeline, the
model was deployed to the Google Kubernetes Engine (GKE) environment to ensure scalability. The
StreamBase RTD system makes a REST call to infer the model in real-time. Average inference time for
the model is ~60 milliseconds, which is acceptable and does not cause any detrimental data traffic within
the StreamBase system. The results of this model have been very accurate, and are currently being used in
the downstream, web-based application by the asset drilling teams.

Figure 8. Evolution of drilling state classification models.


URTeC 253 10

Application
The development of the rig state and subsequent ML-based drilling state models has been driven by
requirements from engineers, managers, and field personnel in asset drilling teams. During drilling
operations, team members ensure safety as their number one priority, while they also deal with other
conflicting priorities, including, but not limited to: minimizing wellbore tortuosity, drilling lateral sections
in-zone, achieving efficient and repeatable performance, solving unplanned operational issues, managing
logistics, and planning future operations, among others. As a matter of practicality, optimization and
performance improvement often takes a lower priority to urgent issues in the critical path of operations. In
addition, optimization projects aimed at identifying and re-engineering performance limitations can be
time consuming workflows that have limited repeatability.
The RTD system uses low-level time-series data labeled with a rig state for multiple higher-level models,
which provide data for user requested workflows on the web-based user interface. Three examples of
workflows include:
1) Performance KPI comparisons
2) Invisible lost time detection
3) Directional performance analysis
A common use-case for real-time drilling data is KPI comparisons. As a member of the drilling team, it
can be extremely useful to compare performance and statistics within or between hole sections, wells,
rigs, basins, and custom groupings of the previously mentioned categories. Historically, these
comparisons were assembled on-demand or as part of a quarterly reporting cycle, both of which can be
time consuming and have limited repeatability for day-to-day needs. To solve this problem and deliver an
efficient, repeatable method to facilitate data-driven decision making, three workflows were constructed.
They are: single-well KPI analysis, multiple-well (or group) KPI comparisons, and a custom grouping
tool to enable advanced KPI comparisons. At the core of all three workflows is labeled time-series data.
Using patterns in the rig state, the system aggregates KPIs such as rotary and slide ROP, footage, and
time, allowing users to rapidly visualize statistics and trends. In a recent example, an engineer was able to
quickly compare the lateral slide percent between two wells on a pad that used different bit and motor
combinations. After quantifying the lagging performance, an excessively high slide percent, and cross
referencing the actual drilling parameters and toolface control trends to rule out other causes of a high
slide percent between the runs, the decision was made to re-design the BHA. The next well drilled by that
rig realized a nearly 20 hour reduction in lateral drilling time. See Figure 9 for the end-user view of a KPI
comparison.
URTeC 253 11

Figure 9. Rotate/slide footage and ROP comparisons across a pad.

Another powerful use-case for real-time drilling analytics is the identification and quantification of
invisible lost time; that is, time spent inefficiently without being labeled truly non-productive, from a rig
downtime perspective. In this case, the time-series data labeled with a rig state is processed to extract
drilling and tripping connection metrics. The connection statistics are then displayed on the UI for
engineers and field personnel to compare various KPIs between crews, wells, and rigs, enabling
identification of top performers and quick insight into performance improvement opportunities. Since the
introduction of the drilling and tripping connection metrics, countless examples of quick discussions
between engineers and rig supervisors as well as rig supervisors and drillers, fueled by actual performance
metrics, have yielded up to 70% reductions in connection times. To add perspective, this can amount to
six or seven hours saved per well, multiplying rapidly when realized across a fleet of rigs. See Figure 10
for an example of the end-user view of connection KPIs.

Figure 10. Drilling connection time KPIs for the lateral section of a Delaware Basin well.
URTeC 253 12

Perhaps one of the most useful applications of the rig state detection algorithm has been in directional
performance analysis. Using the labeled time series data, a high quality depth log is generated by re-
indexing time-series data into 0.5 foot intervals during periods of on-bottom drilling. The depth log is
labeled with the rig state (rotate or slide drilling), and the data is leveraged in a multi-dimensional user
interface to analyze directional drilling performance. In Figure 11, the main section of the plot area is
shaded to indicate rotate and slide intervals, displaying toolface values during periods of slide drilling.
That data is combined with wellbore trajectory and depth-based drilling parameters to enable members of
the drilling team to quickly understand trends during operations. This has led to a variety of instances
where teams were able to provide data-driven explanations to phenomena observed during operations. For
example, a common theme that arises in unconventional wells is bottom-hole assemblies failing to yield
expected build rates. Frequently, the subsequent action involves running a motor with a higher bend
angle. However, armed with quick access to data, teams frequently observe the true root cause to not be in
an inadequate motor, but rather, in excessive burn footage attempting to control toolface. Data-driven
decision making enables teams to treat the root cause of problems opposed to the symptoms, and as a
result, reduce operational risk while improving overall performance.

Figure 11. Multi-dimensional analysis of directional drilling performance. The bottom plot is shaded with the lighter intervals indicating rotary
drilling and the darker intervals indicating slide drilling. The toolface trend is shown in pink during slide drilling.

Conclusions
Applying data science to the oil and gas industry’s operations can be challenging; problems have ever
increasing complexity and uncertainty levels exist surrounding downhole physics. However, the
application of machine learning in drilling time-series rig state classification is an example of a practical
use-case to solve a well formulated problem, leading to significant business value. The authors have
found that deep learning can be leveraged by combining domain knowledge and data science to solve
problems that are otherwise impractical or that have resulting inaccuracies by other means.

In contrast to rules-based methods or traditional machine learning, deep learning has been found to have
extremely high levels of accuracy when trained with appropriate data sets. Aside from domain knowledge
and the patience required to assemble training data sets, a scalable, real-time system was required to
provide a venue to deploy and validate the models against real-time data. To accomplish this, a significant
amount of data engineering effort was required to control data quality, ensure minimal latency in the
streaming pipeline, and manage the storage of analytical outputs for use by drilling teams in day to day
URTeC 253 13

workflows. In addition, a thorough understanding of workflow requirements from asset drilling teams was
crucial. Knowledge of the workflows engineers and field personnel need to enable allowed the
development of specific models to accomplish that goal.

By assembling a multi-disciplinary team containing application development, data science, data


engineering, drilling engineering, and project management skills, the authors were able to successfully
develop and deploy a robust, deep-learning-based solution to solve multiple workflow-based drilling
optimization problems.

Acknowledgements
The authors are thankful to the following colleagues for their help with labeling data as well as training,
testing, and deploying the models: Weilu Han, Sam Ariabod, Ryan Gard, and Don Hender. The authors
are also grateful to the management of Anadarko Petroleum Corporation for their insights, support, and
permission to publish.

Nomenclature

APC - Anadarko Petroleum Corporation


API - Application Programming Interface
CI/CD Pipeline - Continuous Integration / Continuous Deployment Pipeline
CNN - Convolutional Neural Network
CPU - Central Processing Unit
DJ Basin - Denver-Julesburg Basin
EDM - Enterprise Data Management
GKE - Google Kubernetes Engine
GPU - Graphics Processing Unit
KPI - Key Performance Indicator
RNN - Recurrent Neural Network
ROP - Rate of Penetration
RPM - Revolutions per Minute
RTD - Real-Time Drilling
UI - User Interface

References
Ambrus, A., Ashok, P. K., Chintapalli, A., Ramos, D., Behounek, M., Thetford, T. S., and Nelson, B.
2017. A novel probabilistic rig based drilling optimization index to improve drilling performance. SPE
Offshore Europe Conference & Exhibition. SPE 186166-MS. https://doi.org/10.2118/186166-MS.
Arnaout, A., O’Leary, P., Esmael, B., and Thonhauser, G. 2014. Distributed recognition system for
drilling events detection and classification. International Journal of Hybrid Intelligent Systems 11 25-39.
Arnaout, A., Fruhwirth, R., Esmael, B., and Thonhauser G. 2012. Intelligent real-time drilling operations
classification using trend analysis of drilling rig sensors data. Presented at the SPE Kuwait International
Petroleum Conference and Exhibition held in Kuwait City, Kuwait, 10-12 December. SPE 163302-MS.
https://doi.org/10.2118/163302-MS.
URTeC 253 14

Bagnall, A., Lines, J., Bostrom, A., Large, J., and Keogh, E. 2017. The great time series classification
bake off: a review and experimental evaluation of recent algorithmic advances. Data Min Kowl Disc 31:
606-660.
Cao, D., Ben, Y., James, C., and Ruddy, K. 2019. Developing an Integrated Real-Time Drilling
Ecosystem to Provide a One-Stop Solution for Drilling Monitoring and Optimization. SPE Annual
Technical Conference and Exhibition to be held 30 September – 2 October. SPE 196228-MS.
Cao, D., Loesel, C., and Paranji S. 2018. Rapid Development of Real-Time Drilling Analytics System.
IADC/SPE Drilling Conference and Exhibition, 6-8 March, Fort Worth, TX, USA. SPE-189595-MS.
https://doi.org/10.2118/189595-MS
Coley, C. 2019. Building a Rig State Classifier Using Supervised Machine Learning to Support Invisible
Lost Time Analysis. Presented at the SPE.UADC Driling International Conference and Exhibition,
Hague, Netherlands, 5-7 March. SPE/IADC 194136-MS. https://doi.org/10.2118/194136-MS.
Dunlop, J., Lesso, W., Aldred, W., Meehan, A., Orton, M.R., and Fitzgerald, W.J. 2006. System and
method for rig state detection. US Patent US 7128167B2.
Fawaz, H.I., Forestier, G., Weber, J., Idoumghar, L., and Muller, P. 2018. Deep Learning for time series
classification: a review https://arxiv.org/abs/1809.04356, Sep 12.
Pournazari, P., Ashok, P. K., Van Oort E., Unrau, S., and Lai, S. 2015. Enhanced kick detection with low-
cost rig sensors through automated pattern recognition and real-time sensor calibration. SPE Middle East
Intelligent Oil & Gas Conference & Exhibition. SPE 176790-MS. https://doi.org/10.2118/176790-MS.
Rajpurkar, P., Hannun, A. Y., Haghpanahi, M., Bourn, C., and Ng, A. 2017. Cardiologist-level
arrhythmia detection with convolutional neural networks. https://arxiv.org/abs/1707.01836.
Veres, G. and Sabeur Z. 2015. Date analytics for drilling operational states classifications. European
Symposium on artificial neural networks, Computational intelligence and machine learning. Bruges
(Belgium), 22-24 April.

Вам также может понравиться