IPTC-19714-MS

Deep Learning for Monitoring Well Testing Burner Efficiency

Darine Mansour, Hakim Arabi, Yingwei Yu, and Hugues Trifol, Schlumberger

Copyright 2020, International Petroleum Technology Conference

This paper was prepared for presentation at the International Petroleum Technology Conference held in Dhahran, Saudi Arabia, 13 – 15 January 2020.

This paper was selected for presentation by an IPTC Programme Committee following review of information contained in an abstract submitted by the author(s).
Contents of the paper, as presented, have not been reviewed by the International Petroleum Technology Conference and are subject to correction by the author(s). The
material, as presented, does not necessarily reflect any position of the International Petroleum Technology Conference, its officers, or members. Papers presented at
IPTC are subject to publication review by Sponsor Society Committees of IPTC. Electronic reproduction, distribution, or storage of any part of this paper for commercial
purposes without the written consent of the International Petroleum Technology Conference 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 where and by whom the paper was presented.
Write Librarian, IPTC, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A., fax +1-972-952-9435.

Abstract
During well testing operations, an incorrect burner combustion may pose a risk of adverse impact on the
environment or people. The combustion efficiency is assessed by personnel who observe the flame. This
practice lacks consistency and may pose challenges, from environmental and safety considerations to data
quantity and quality. An automatic alarm triggered by suboptimal combustion is needed. In this paper, we
propose a solution that uses a deep neural network that learns from flame videos to define the quality of
the combustion.
Flame features that help determine the combustion quality are the flame color, the color of the smoke
around the flame and the flame shape and pattern, among many others. To extract the features relevant to
the problem and qualify the combustion, we performed supervised learning using a convolutional neural
network. We trained the network using videos of flames from burner combustion; the network extracts
relevant features and classifies the flame in video images as acceptable or undesirable.
We labeled a set of flame videos obtained from well testing operations. The videos include various flame
conditions, in day and night scenarios and acceptable and undesirable combustion contexts. We trained the
convolutional neural network using the labeled flame videos.
The network prediction accuracy is 100% on the training set and 96.8% on the test set. The misclassified
images are in the transition periods between acceptable and undesirable combustion events.
We demonstrated the potential of a deep learning solution using a convolutional neural network for
classifying the combustion quality of burner images as acceptable or undesirable. The results are promising,
and they show that this solution is a good candidate for real-time burner efficiency monitoring and automatic
alarm triggering and optimization.

Introduction
Burners used during well testing operations need special monitoring as poor oil and/or gas combustion
may cause damage to both people and the environment. When the proper conditions are not fulfilled for an
efficient combustion of effluent oil/gas with a burner or flaring system, small drops of oil or soot can drop
into the sea, land, or other surrounding environment. Also, the smoke and heat generated by the burners and
2 IPTC-19714-MS

radiated to any nearby facilities can be affected by changes in the environment, such as wind speed and/or
direction, potentially resulting in unsafe conditions for operators (such as toxic black smoke) and equipment.
To account for large potential fluctuations in burning efficiency, the current practice is for the operators
to look at the burner flame. The operators adjust the flow rates or other variables within their control that
affect burning efficiency, based on what they see and their experience and ability to analyze in real-time.
There is no consistency in the evaluation of the combustion efficiency because the people in charge of
watching the burners have their own interpretations of acceptable combustion. There is often a late detection
of combustion issues. The automation of the evaluation of combustion efficiency (Arabi 2015; Arabi
and Piasseski 2016) would promote consistent combustion efficiency and timely reaction to undesirable
combustion.
Today, in field operations, cameras are positioned so they have a direct view of the burner flames. We have
a dataset of video images captured by the cameras and labelled by domain experts as images of acceptable
or undesirable combustion. We want to automate the human monitoring of the combustion by processing
these images and automatically classifying them as images of acceptable or undesirable combustion. We
applied a supervised learning approach to solve our problem. It consists of extracting the flame features
from each image and feeding the features to a classifier that classifies the combustion.
The classifier is trained using the extracted features and images combustion labels as input.
The traditional approach to feature extraction is the manual feature engineering. It involves extracting
hand crafted features using various image processing algorithms and domain knowledge. The process is
tedious and time consuming. A lot of effort is required to extract all relevant features in various contextual/
environmental conditions. The extracted features are problem dependent, and the process may require
revisiting the algorithms to adapt to new data that hold new features or new representations of existing
features.
Considering these limitations, we chose a deep learning approach to automatically extract the features
relevant to our specific combustion classification problem and to train a classifier.
Deep learning is showing state of the art results in the fields of computer vision, speech recognition
and others. Convolutional Neural Networks (CNN) are the most popular deep learning models used in the
field of image classification and object recognition (Ciresan et al. 2011; Krizhevsky et al. 2012; Szegedy
et al. 2015). CNN learn the filters that are hand engineered in traditional image classification algorithms;
this reduces the tedious and time-consuming process of feature engineering and with CNN, the features
are automatically learned from data. The cost of training CNN is less than the traditional approach and the
network can be improved perpetually by training it on new richer datasets. The learnt deep features can
generalize well to new scenarios.
In this paper, we propose using a deep learning approach to classify the combustion of well testing burners
flame images as acceptable or undesirable. We tested the approach on a set of labelled burner flame videos
taken during well testing field operations.
The rest of the paper is organized as follows. Next section describes the deep learning approach used.
It is followed by a full analysis of the results of the combustion classification tasks. Last section presents
the conclusion with way forward.

Approach
This work aims to classify flame images of well testing burners as images of acceptable or undesirable
combustion. To achieve the classification, we train a deep CNN using a dataset of images labelled by domain
experts. In this section, we describe the CNN architecture and the workflow for our approach.
There are numerous CNN architectures that were trained on large image classification and detection
datasets such as ImageNet (Russakovsky et al. 2014) and that achieved very good classification accuracy
results such as AlexNet and VGGNet. Using our flame natural image dataset captured by well testing
IPTC-19714-MS 3

cameras, we trained and fine-tuned a CNN with an architecture similar to VGG-16 (Simonyan and
Zisserman 2014) which was the runner-up at the ImageNet ILSVRC 2014 competition, and which learned
to extract powerful features from natural images.
The network architecture is illustrated in Figure 1 and detailed in Table 1. We changed the number of
neurons in the last fully connected layer of VGG-16 to 2 because we have two target classes: acceptable
combustion and undesirable combustion. The network input is a fixed size 224×224×3 image, and it has
two outputs which represent the probabilities that the image belongs to each of the two target classes. The
network consists of 16 weight layers: 13 convolutional layers for feature extraction and 3 fully connected
layers for classification. The input image is normalized, and it passes through the stack of 13 convolutional
layers where filters with a very small receptive field (3 × 3) are used. The resulting 512 feature maps
go through the classifier 3 fully connected layers. All hidden layers have a Rectified Linear Unit (ReLU)
activation function. The final layer has a softmax activation function.

Figure 1—CNN architecture: The dimensions in the feature extraction section correspond to
the size x number of feature maps at the output of each stack of convolutional layers. The
dimensions in the classification section correspond to the size of the fully connected layers

Table 1—CNN layers and configuration

Layer Type Parameters

Input image Size: 224×224×3

Convolutional 64 filters

Convolutional 64 filters

Max Pooling Size: 2×2, Strides: 2

Convolutional 128 filters

Convolutional 128 filters

Max Pooling Size: 2×2, Strides: 2

Convolutional 256 filters

Convolutional 256 filters

Convolutional 256 filters

Max Pooling Size: 2×2, Strides: 2

Convolutional 512 filters

Convolutional 512 filters

Convolutional 512 filters

4 IPTC-19714-MS

Layer Type Parameters

Max Pooling Size: 2×2, Strides: 2

Convolutional 512 filters

Convolutional 512 filters

Convolutional 512 filters

Max Pooling Size: 2×2, Strides: 2

Fully Connected 4096 units, activation: ReLU

Fully Connected 4096 units, activation: ReLU

Fully Connected 2 units, activation: softmax

All convolutional layers have the following parameters: size: 3×3, activation: ReLU, strides: 1, padding: "same"

To increase the size of our dataset and introduce variants of the representation of the flame features, we
augmented the original dataset images by mirroring and rotating them. We then resized the images to fit
the input size 224×224 of the chosen CNN and we normalized the images. We then randomly shuffled the
dataset and split it between training and test sets. The results of the CNN training and testing are presented
in the next section.

Results
We have a dataset of video images captured by the cameras during well testing field operations. The dataset
is labeled by domain experts as images of acceptable or undesirable combustion. The dataset consists of
three videos of 16 min each, 1 frame per second, and a resolution of 640×480. They include different lighting
conditions (light and dark) and both cases of acceptable and undesirable combustion (black or white smoke,
with or without activated water screens).
Overall, we have 5984 images. We ran the CNN training on 70% of the dataset (original and augmented).
We tested the trained CNN on the remaining 30% of the dataset. We have two target classes: acceptable
combustion and undesirable combustion.
The network was trained for 60 epochs using Adam optimizer, with a batch size of 64. The learning rate
is 1×10−4. The training ran on a NVIDIA Tesla K80 GPU for 103 min.
The training accuracy is 100%, and the accuracy of the testing set is 96.8%. The confusion matrix of the
testing set in shown in Figure 2.

Figure 2—Confusion matrix of the testing set

A closer look at the mis-classified images within the video sequence shows that undesirable combustion
events are never missed and most of the mis-classified images are located in a transition phase when the
IPTC-19714-MS 5

combustion goes from acceptable to undesirable state or from undesirable to acceptable state. Few images
within some transition phases are mis-classified and therefore there is a discontinuity in the classification
within this critical phase; this may be due to the fact that the shuffled images are fed into the network as
independently and identically distributed samples. The result may be improved by allowing the network to
capture the temporal relationship between consecutive image frames and this will be the scope for future
work.
The images fed into the CNN are transformed into a representation at the last hidden layer in which the
two target classes are separable by a linear classifier. The space dimension at the last hidden layer is 4096.
Images that have representations close to each other or clustered together, are considered "similar" by the
CNN; the similarity is class/semantic based. To get an idea about the topology of the high dimensional
space, we can embed it into two dimensions where the pairwise distances between the high dimensional
representation of the images is preserved. This would also allow analyzing the features extracted by the
network. To perform this embedding, we use the t-Distributed Stochastic Neighbor Embedding (t-SNE)
method (Van der Maaten and Hinton 2008). We input the dataset images into our trained CNN and extract
their high dimensional representation at the last hidden layer. We then feed this representation into the t-
SNE algorithm and we get the two-dimensional t-SNE components space (see Figure 3) which we illustrate
in Figure 4.

Figure 3—Method to embed the CNN representation at CNN last hidden layer into t-SNE representation

Figure 4—t-SNE representation of the dataset. a and b show the two-dimensional t-SNE components of the
dataset representation at the last hidden layer of the trained CNN. In a, the red color refers to the images
representation that are classified by the trained CNN as acceptable combustion and the blue color refers to
the ones classified as undesirable combustion. In b, the red color refers to the images representation that are
labeled as acceptable combustion and the blue color refers to the ones labeled as undesirable combustion.
6 IPTC-19714-MS

In Figure 4, every dot corresponds to an image in the dataset; more specifically each dot is the t-SNE
components of the representation at the last CNN hidden layer of each image in the dataset. The red dots are
the ones labeled as undesirable combustion in Figure 4-b and the ones classified by the CNN as undesirable
combustion in Figure 4-a. The blue dots are the ones labeled as acceptable combustion in Figure 4-b and
the ones classified by the CNN as acceptable combustion in Figure 4-a.
We can clearly see that there are two major clusters that correspond each to one target class: acceptable
combustion (blue cluster) and undesirable combustion (red cluster). We can also notice a transition region
between the two clusters; in Figure 4-a, the CNN classification has a strict spatial separation between the two
classes within the transition region while in Figure 4-b, the images labels are overlapping in the transition
region. Finally, we can see that we have clusters within each class. We describe these different regions in
Figure 5; it shows how the trained CNN clusters together the flame and smoke patterns that have common
features like undesirable combustion flames with predominant white smoke or predominant black smoke
with their respective characteristic flame patterns.

Figure 5—Regions/clusters in t-SNE representation. The blue cluster (Regions A and B) corresponds to the acceptable
combustion class; Region A includes non augmented images and Region B includes the augmented ones. The red
cluster (Regions C, D and E) corresponds to the undesirable combustion class; Regions C and D include images of
the flames with predominant black smoke and a flower flame pattern; Region C covers the augmented images and
Region D covers the non-augmented ones; Region E includes the images of the flames with predominant white smoke
and a different flame pattern. The labels and CNN classification mostly agree about the acceptable combustion class
within the blue cluster and the undesirable combustion class within the red cluster (see Figure 4). Region F is the
region of transition between the blue (acceptable combustion) cluster and the red (undesirable combustion) cluster.

Most of the mis-classified images belong to the region of transition (Region F) between the blue cluster
(acceptable combustion class) and the red cluster (undesirable combustion class) in Figure 5. By looking at
the images within this region, we do not distinguish visually the difference between the images classified by
the trained CNN as undesirable combustion and labeled by humans as acceptable or undesirable combustion.
The same applies for the images classified by the CNN as acceptable combustion. We could have some cases
of mis-labeling in the transition region. The mis-classification could also be due to the fact that the images
are isolated from their temporal context when they are classified independently by the CNN; the results
could be improved through temporal averaging applied on the classification of consecutive video frames.
We could also improve our approach by making the network capture the temporal and spatial relationship
IPTC-19714-MS 7

between consecutive video frames, which would allow capturing the flame or smoke volume and changes
in volume and the flame temporal pattern.

Conclusion
The aim of this paper is to automate the human monitoring of the burners combustion during well testing
field operations; the monitoring is done today by simply looking at the flame. We propose using a supervised
deep learning approach to classify the combustion of well testing burners flame images as acceptable
or undesirable. We trained and fine-tuned a convolutional neural network with an architecture similar to
VGG16. We trained the CNN and tested the approach on a set of labeled burner flame images taken by
an optical camera during well testing field operations. The classification results are very promising. They
demonstrate the potential of the proposed method in classifying the combustion performance and they open
the door to consider deep learning techniques for burner monitoring and automatic alarm triggering.
In a future work, it is important to acquire richer datasets in various environmental contexts and apply
consistent labeling of combustion performance for a better training of the network. We could also consider
real-time labelling by operators to mature the network perpetually.
The proposed algorithm can be improved by capturing the temporal relationship between consecutive
video frames for a better flame assessment.

References
1. Arabi, H. 2015. Optical monitoring system for a burner. European Patent 15290246.6
2. Arabi, H., Piasseski, D. 2016. Burner monitoring and control systems. International patent
application PCT/US2016/054041
3. Ciresan, D. C., Meier, U., Masci, J., Gambardella, L. M., Schmidhuber, J. 2011. Flexible, High
Performance Convolutional Neural Networks for Image Classification. International Joint
Conference on Artificial Intelligence (IJCAI), pp. 1237–1242.
4. Krizhevsky, A., Sutskever, I., Hinton, G. 2012. Imagenet Classification with Deep Convolutional
Neural Networks. Neural Information Processing Systems (NIPS).
5. Russakovsky, O., et al. 2014. ImageNet Large Scale Visual Recognition Challenge.
ArXiv14090575 Cs.
6. Simonyan, K., Zisserman, A. 2014. Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556.
7. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and
Rabinovich, A. 2015. Going Deeper with Convolutions. IEEE Conference on Computer Vision
and Pattern Recognition (CVPR).
8. Van der Maaten, L. J. P., Hinton, G. E. 2008. Visualizing High-Dimensional Data Using t-SNE.
Journal of Machine Learning Research. 9, 2579–2605.