^{*}

This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).

This paper presents novel research on the development of a generic intelligent oil fraction sensor based on Electrical capacitance Tomography (ECT) data. An artificial Neural Network (ANN) has been employed as the intelligent system to sense and estimate oil fractions from the cross-sections of two-component flows comprising oil and gas in a pipeline. Previous works only focused on estimating the oil fraction in the pipeline based on fixed ECT sensor parameters. With fixed ECT design sensors, an oil fraction neural sensor can be trained to deal with ECT data based on the particular sensor parameters, hence the neural sensor is not generic. This work focuses on development of a generic neural oil fraction sensor based on training a Multi-Layer Perceptron (MLP) ANN with various ECT sensor parameters. On average, the proposed oil fraction neural sensor has shown to be able to give a mean absolute error of 3.05% for various ECT sensor sizes.

In the oil industry, it is vital to measure fluids in oil pipelines for optimizing exploitation, production and transportation (

The first ECT sensor system was developed in 1970s by the US Department of Energy in Morgantown, WV to visualize the solid-particle distribution in fluidized beds [

The first image reconstruction algorithm for ECT, known as the Linear Back Projection (LBP), was developed by Xie

The current work focuses on utilizing a direct method. Using this method, a more accurate calculation of oil concentration is expected, along with shorter processing times because no pixels are involved, thus making it possible to cut costs and increase efficiency. Previous work by Xie

ANN is renowned as a robust artificial learning tool for solving numerous ECT-based problems; with applications ranging from pattern recognition of multiphase flows in pipelines to direct process interpretation without recourse to image reconstruction. Previous works by Duggan and York [

Despite the success, all the ECT works have been based on fixed ECT sensor parameters, which in turn have produced non-generic neural oil fraction sensor systems. In order to build a generic neural oil fraction system, various ECT parameters can be considered. Several works [

Basically, an ECT system works based on Poisson's equation given by:

By applying Gauss law, the induced charge at electrode _{j}_{ij}_{ij}

The basic principle of ECT is to employ a desired number of adjacent rectangular electrodes (separated by a small gap) as sensors. The electrodes are made of a conductive material placed around some process equipment containing the materials to be visualised. Electrode capacitance sensors detect materials which are non-conductive or have different dielectric properties. If the two flowing materials have different permittivities, the electrical capacitance between all possible electrode pairs will change. The values of capacitance measurement depend on the dielectric constant value of each component in the mixture and their distribution inside the process equipment. The changes are measured. The next step is to obtain an image of such material distribution using the capacitance measurements with an appropriate image reconstruction algorithm.

The ECT system also consists of a data acquisition system (DAS) and computer system as illustrated in

Basically, an ECT system consists of a sensor, a data acquisition system (DAS) and a computer system consisting of various process interpretation components as illustrated in

The choice in the number of electrodes is a tradeoff between resolution, sensitivity and image capture rate. More sensing electrodes will produce better resolution images, but the measurement sensitivity will be lower because the size of electrodes becomes smaller and at the same time the capture rate is reduced. The sensitivity of the sensor and the image capture rate can be increased by using a bigger electrode size, though this will decrease the image resolution [

With 12-electrodes, there are 66 possible capacitance values that can be measured. This is based on:

As higher capacitance sensitivity is desirable and can be achieved by using larger angles of primary electrodes, the angles under investigation are chosen to be within the range of 20° until 26° with 0.5° interval. The minimum size of 20° is chosen because any smaller size of primary electrode leads to lower sensitivity distribution of the capacitance [

The ECT sensor data can be collected in two ways, by means of the actual plant data or generated from the ECT sensor simulator [

Various sets of ECT sensor data have been simulated based on different flow patterns of each flow regime. The numbers of simulated flow patterns for each flow regime are as depicted in

This work proposes that each set of the measured capacitance is normalized based on the average capacitance measurement of empty and full flows of various θ ranging from 20° to 26°. This is because a generic oil fraction neural sensor system deals with ECT sensor data generated from primary electrode sizes, θ of 20° to 26°. Hence, every set of ECT sensor data corresponding to each flow pattern is normalized using:

The term component fraction, _{mat} refers to the ratio of the cross-sectional area covered by the material of interest _{mat}, to the area of the entire cross-section of sensing area _{pipe}. In this work, the material to be estimated is oil and hence, the equation is given by:

The values of component fractions range from zero (an empty pipe) to one (a pipe full either with oil or water). In this work, training and testing sets are chosen to be as large as possible and equal in numbers but of different sets. This approach has a potential to produce a more intelligent sensing system as larger number of training set is utilized and also to produce more accurate performance assessment as a larger number of testing datapoints is used [

The robustness demonstrated by ANN in previous works has motivated the use in this work of a similar approach for developing a generic intelligent oil fraction sensor system. There are two types of learning with ANN; either supervised or supervised. The works by Mohamad Saleh

The MLFF ANN used in this work consists of multiple-input neurons as shown in _{1}_{2}_{3}_{R}_{1,1}_{1,2}_{1,R}

In matrix form the equation can be written as:

These multiple-input neurons stack together to produce multiple layers that operate in parallel. Finally, these layers are cascaded together to form a fully-connected MLP network as depicted in

An MLP ANN with 66 input neurons (for 66 capacitance measurements) and one output neuron is constructed for developing a neural oil fraction sensor. The optimum number of hidden neurons is determined using a network-growing approach [

The training process of an MLP is run based on a training algorithm. The Back Propagation algorithm [

The BR training algorithm is an extended or modified version of the LM training algorithm. Although, this training algorithm updates weight and bias values according to LM optimization, it has a different performance cost function than LM training algorithm in that it also minimizes ANN error, along with weights as well as biases, to produce a network that generalizes well [

A sigmoidal activation function is a nonlinear function that transforms the weighted sum of inputs to output values. It is applied to either hidden or output neurons. The choice of activation function for hidden neurons is important as the selection is usually problem-dependent and may affect an ANN's performance.

As shown in

Another most common activation function used in Back Propagation learning is the hyperbolic tangent sigmoid activation function. This activation function is similar to the logarithmic sigmoid in that the input values range within [−∞, +∞], but it transforms into an output ranging from −1 to 1 as depicted in

In this work, the output values for oil fraction can range from 0.0001 to 0.9999; therefore, the most suitable activation function to be applied at output neuron is logarithmic sigmoid (Logsig). Meanwhile, hyperbolic tangent sigmoid (Tansig) and logarithmic sigmoid (Logsig) functions are investigated to select the most suitable activation function for hidden neurons.

An experimental method has been devised for our investigation towards developing a generic neural oil fraction sensor using various θ ranging from 20° to 26° with 1° intervals, as illustrated in

To develop a generic neural network for oil fraction estimation, 2800 ECT data have been used and divided into 40% for training, 20% for validation and 40% for test sets. The raw data have been normalised using

The performance of a MLP oil fraction sensor is evaluated based on mean absolute error (MAE). This error measure is used because it gives good indicator of how much MLP estimations of oil fraction differ from their actual values. MAE is given by:
_{i}^{th}_{i}^{th}

Non-generic neural sensors have been developed by separately training seven MLPs with ECT data with a single θ. A total of 160 training data, 80 validation data and 160 test data are used for each θ of the ECT design. The seven MLPs have also been trained using the best training algorithm and the best hidden activation function used to develop the generic neural sensor. Then,the evaluation of these non-generic neural sensors is carried out based on

Upon determining the best procedure for developing a generic intelligent oil fraction sensor, the work proceeds with utilization of PCA technique on ECT sensor datasets. This step is an attempt to improve the capability of the generic neural oil fraction sensor. Basically, the PCA technique is applied to reduce correlated information in the ECT sensor data. This correlation in ECT sensor data creates confusion over MLPs during learning process, thus degrading their generalization capability. PCA is a procedure which converts a set of data of possibly correlated variables into a set of values of uncorrelated variables. The uncorrelated variables, for ECT is the capacitance measurement, are also known as the principal components. ECT sensor data have been known are highly correlated due to overlapping sensing areas as depicted in

To implement the PCA technique for the ECT sensor data, raw ECT sensor data, C are first normalised to have zero mean and unity variance. Then using a mathematical technique called Singular Value Decomposition (SVD), along with the normalised ECT sensor data, N, mean and variance values the principal component are calculated. This generates a transformation matrix,

While pre-processing is employed on the training set, validation and testing sets have to be post-processed using the post-PCA technique before they can be used for MLP training. _{val/test}_{val/test}_{val/test}_{val/test}

The fully developed generic neural oil sensor needs to be assessed to verify its supremacy. The system is first evaluated based on test set which has been carried out during system development. The second evaluation involves executing the best-performed neural oil sensor using a verification dataset comprises 6,000 datapoints of various flows to further verify its performance. The verification data are ECT data obtained from θ values of 20.5°, 21.5°, 22.5°, 23.5°, 24.5° and 25.5°. These verification datasets are used to verify the robustness of generic neural oil sensor.

The search for the smallest possible MAE proceeds with MLPs trained with less than five input components (

^{®}. A correlation plot shows how close the MLP outputs are to the respective target outputs. From the figure, it can be seen that most data points gather closer to the y = x straight line going through the origin. This means that the MLP outputs have a high correlation with the target output values. The figure also shows that the MLP enhanced with the PCA technique has a higher correlation value, R of 0.9747, that the non-enhanced MLP with R = 0.9722.

Previous works in the literature only focused on developing a process parameter sensor based on fixed ECT sensor parameters. This paper has presented methods and experiments aimed at developing an intelligent oil fraction sensor system based on generic sensor parameters which is desirable in industry in order to save equipment and operation costs. The ANN technique is used to develop a generic neural oil fraction sensor using generic sensor parameters. The work proposed varying the primary electrode sizes towards developing the generic neural oil fraction sensor. It is also involved proposing a new generic normalization equation based on average empty and full ECT sensor data. Experimental methods have been devised for the investigation of the best developed neural sensor system. The results from the experiments have shown that it is feasible to develop a neural oil fraction sensor system for generic ECT data by employing MLP ANN. Furthermore, the application of PCA technique to the input data sets has shown its capability in improving the generalization performance of the MLP ANN. The MLP ANN generalizes better when the number of input datasets is optimum as the PCA technique has removed the correlated components. In the future, it would be interesting to consider other ANN architectures such as a Hybrid-MLP for the possibility of improving the performance of the neural oil fraction sensor system based on generic ECT sensor data. Besides that, the methods and experiments can be considered for the development of generic neural oil fraction sensor systems for three-component flows in a pipeline comprising gas, oil and water. This way the system could be made more useful as three-component flows are common in an oil industrial environment.

The authors wish to thank the Ministry of Science, Technology and Innovation (MOSTI), Malaysia for financially supporting this research under eScience fund grant No. 03-01-05-SF0134 and the Ministry of Higher Education (MOHE), Malaysia for the Fundamental Research Grant Scheme (FRGS), No. 203/PELECT/6071148.

The author declares no conflict of interest.

Schematic diagram of an ECT system.

ECT sensor model.

A schematic diagram of (

A multiple input neuron.

A schematic diagram of a Multi-Layer Perceptron (MLP) neural network.

Activation function: (

A schematic diagram of experimental method towards developing neural oil fraction sensors based on (

MLP training process to develop intelligent flow sensor.

Overlapping sensing region between electrode 1 and other electrodes.

Schematic diagram of pre-processing of ECT data.

Schematic diagram of post-processing procedure.

Test set MAE values and training times of MLP trained with five to 60 principal components.

Test set MAE values and training times of MLP trained with one to four principal components.

Test set MAE values and training times of MLP trained with six to nine principal components.

Correlation output display of oil fraction neural sensor supported by MATLAB.

Number of flow patterns generated for each θ.

Full | 1 |

Stratified | 1,224 |

Bubble | 1,436 |

Annular | 199 |

Core | 199 |

Empty | 1 |

The MAE values for MLP oil fraction sensors trained using different training algorithms and hidden activation functions.

| |||
---|---|---|---|

LM | Tansig | 4.13 | 3.73 |

Logsig | 4.18 | 3.91 | |

| |||

BR | Tansig | 4.40 | 3.89 |

Logsig | 4.41 | 3.99 |

The MAEs for different intelligent oil fraction sensors.

MLP_20° | 7.56 |

MLP_21° | 7.26 |

MLP_22° | 7.84 |

MLP_23° | 7.08 |

MLP_24° | 8.46 |

MLP_25° | 8.38 |

MLP_26° | 6.30 |

Generic MLP | 3.73 |

MAEs of best-performed MLPs with and without PCA technique.

MAE (%) | 3.73 | 3.05 |

MAEs of the best MLP using optimal principal components tested on ECT data based on various θs.

| ||||||
---|---|---|---|---|---|---|

Stratified | 2.63 | 2.62 | 3.02 | 3.74 | 5.34 | 8.16 |

Bubble | 1.20 | 1.25 | 1.31 | 1.52 | 2.69 | 3.12 |

Core | 0.99 | 1.15 | 0.95 | 1.07 | 1.63 | 1.52 |

Annular | 1.50 | 1.19 | 1.03 | 0.96 | 1.48 | 2.26 |

| ||||||

Average | 1.65 | 1.63 | 1.7 | 1.99 | 3.03 | 4.15 |