Bayesian network is employed to estimate a risk-based life cycle cost of corrosion for assets. It has been highly recognized that inclusion of mechanistic models to a Bayesian network can increase the confidence in estimation of corrosion rates. However, coefficients of mechanistic models are often unknown, especially when complex rate processes are involved, which discourages the usage of the model. A methodology is proposed here, to introduce a mechanistic model as a bias to a regressive machine learning (ML) algorithm. No attempts have been made to obtain phenomenological coefficients of the mechanistic model. Instead, a methodology is proposed to obtain a highly tuned parameter vector for a ML algorithm from a learning set of corrosion rate data.
A constant challenge persists among corrosion engineers to estimate and predict field corrosion rates despite the huge advancements in corrosion science. This situation has compelled the corrosion engineers to opt for the machine learning (ML) algorithms for corrosion prediction. However, the "blackbox" ML algorithms are not appreciated in high stakes decisions because they use arbitrary fitting models rather than scientific principles.1 Learning achieved by such an algorithm is confined to itself and no useful knowledge can be acquired from it. Hence, it is necessary to include mechanistic models into machine learning algorithms for more confident prediction of corrosion rates.
According to the latest impact report by NACE, global economic loss due to corrosion was estimated to be $2.5 trillion in 2013, of which 15 to 35% could be saved by implementing proper corrosion control and management practices.1 Bayesian network (BN), a probabilistic learning algorithm with cause-consequence type structure, is utilized in calculation of indirect costs from the failure, especially in cases when localized failure mechanisms are expected.1 Learning, and consequently the reliability, of BN depends on input corrosion data, which can be generated by mechanistic models, mined from expert knowledge, or simply being available from the field.2 Among these, the mechanistic models are the most preferable to generate corrosion rate data used in preparing conditional probability tables for BN.2,3
