A novel metaheuristic-based approach for prediction of corrosion characteristics in offshore pipelines

Corrosion is recognized as the primary cause of structural damage and collapse. While numerous studies can be found that imply a decent local prediction, they lack generalizability and extensive application, or they exhibit out of bound errors, making development of an accurate risk-based inspection planning unfeasible. Machine Learning-based approaches opened new windows in the field; however, they are frequently optimized by gradient-based algorithms which may have various issues and restrictions, such as the need for continuous differentiable function. Metaheuristic optimizers have already managed to address some of these concerns, showing superiority over others due to their unique characteristics, including easy-to-understand procedure, easy-to-implement approach, and derivative-free mechanism. Therefore, the aim of this study is to investigate how powerful these optimizers can be in prediction of corrosion using a machine learning-based approach. Four well-known swarm-based optimizers such as Grey Wolf Optimizer, Particle Swarm Optimizer, Slime Mould Optimization, and BAT optimization will be utilized in this study. In addition, a novel metaheuristic optimizer will be employed to enhance performance of optimizers. Their performance will be evaluated through optimizing 1D functions and multi-dimensional functions in terms of different metrics, such as mean and standard deviation of optima and run time. The algorithms will then be integrated into the architecture of a two-hidden-layer feedforward neural network to optimize the corresponding network's weight and bias values. A Sobol indices-based sensitivity analysis will be carried out to identify the most substantial parameters on the corrosion depth. Moreover, to investigate the robustness of the network, white Gaussian noise at different signal-to-noise ratios will be added to the output and change in the output will be measured and compared against gradient decent-based feedforward neural network. The hybrid network will then be integrated into a performance-based engineering framework for estimation of corrosion-induced risk on the pipe. The results demonstrate that the novel approach facilitates development of an automated risk-informed decision-making framework.