Optimal Artificial Neural Network Model For Prediction of Oil and Gas Pipelines Defect Length

Abstract

Magnetic flux leakage (MFL) signals are used for estimating not only the size, but also the shape of the faults causes by deteriorating metal that makes the oil and gas pipelines. Such defects, such as rust, which, if left undetected and poorly handled, may have devastating effects, both in terms of environmental degradation and loss of life, and also millions of dollars in repair costs to be incurred by the ownership firms. Algorithms focused on machine learning have demonstrated the ability to solve the issue by identifying and measuring the size and shape of such defects effectively. In particular, artificial neural networks (ANN) have shown great potential to generate high precision results. In this article, the optimization of ANN was carried out by using noisy and noiseless measurements of MFL signals. ANN optimization was conducted for the training function (12 separate training functions), the hidden neurons numbered between 1 and 100, and hidden layers numbered between 1 and 10. The output was calculated by root mean square (RMSE) error. It has been found that gradient descent momentum (GDM) and gradient descent (GD) exhibit bad performance outcomes than all other studied training functions, whereas all other studied training algorithms showed equal and comparable performance outcomes. The highest output outcomes have been found in the range of 1, 10 and 20 to 35 with regard to the number of hidden neurons. Network output deteriorates as the number of hidden neurons deviates from the optimal range observed. With regard to the optimal number of hidden layers, it has been observed that ANN yields better output results with 1,2,5 and 8 hidden layers for noiseless MFL signals and the best results with 1,2,5 and 7 hidden layers are observed for noisy datasets. To evaluate the oil and gas pipeline defects, the optimal inferred parameter range may be used to train the ANN model for improved performance outcomes.