Application of Machine Learning (ML)-based multi-classifications to identify corrosion fatigue cracking phenomena in Naval steel weldments

Abstract

Corrosion fatigue (CF) crack detection in welded ship hull structures is quite crucial as well as challenging for ensuring their structural integrity. This is conventionally done by offline ship hull survey and inspection methods, which have large time and cost implications. Data-driven approaches of ship hull crack detection are promising in this regard. Corrosion fatigue crack growth rate (CFCGR) behaviour of shipbuilding steel was studied for its base metal (BM), fatigue-prone heat affected zones (HAZs) of Submerged Arc Welding (SAW), and shielded metal arc welding (SMAW) joints of Naval-grade steel. The results indicated the highest CFCGR for BM, intermediate for SAW, and smallest for SMAW welds. This was attributed to 44% lath martensite observed in the SMAW HAZ microstructure as compared to 33% in SAW HAZ. A novel ML-based method has been proposed for CF crack location identification by a multi-classification approach using CFCGR data for structural health monitoring (SHM) applications. Prediction of crack location (either BM, SMAW HAZ, or SAW HAZ) was investigated using K-nearest neighbour (KNN), linear support vector machine classifier (LSVC), Naïve Bayes (NB), decision tree (DT), random forest (RF), and artificial neural network (ANN) models. DT and RF models were found to perform very well even without scaling and hyper-parameter tuning operations and exhibited 98–99% classification accuracy. Scaling and tuning operations resulted in significant improvements in accuracy of KNN models from 43% to 84% (minmax scaling) and 98% (standard scaling). Significant improvement in the accuracy from 37% to 98% ANN model was achieved by standard scaling. The proposed method shall be useful in data-driven real-time crack location identification in ship hull structures.