Learning with Corrosion Feature: For Automated Quantitative Risk Analysis of Corrosion Mechanism

Abstract

This work attempts to address an important issue in the process of quantitative risk analysis. In particular, an automated approach to identify corrosion mechanism that may happen given a set of readings from ship or chemical plant is developed. Given some records where each consists of a set of measurements (design and operating conditions) and label (type of corrosion mechanism expected to happen), learning through Support Vector Machines is performed. After learning, prediction to identify corrosion mechanism expected to happen can be done via the trained SVM classifier. The methodology starts with transforming each record into corresponding mathematical representation, in feature space known as Corrosion Feature. Supervised learning with points obtained from the process of representation and corresponding label can be performed subsequently. In 16-dimensional space, learning with non-linear kernel has demonstrated encouraging performance on a dataset with 4, 423 records created by expert in the industry.