Review of State-of-the-Art Machine Learning Research for Structural Integrity Assessment of Welded Joints

Abstract Machine Learning (ML) complements traditional scientific methods of physical experimentation and computer simulation and is a field of growing interest in material science and engineering. Whilst traditional methods are used for engineering design and analysis, additional insights from ML may help to prevent the failure of critical components which adversely affects plant operations, triggering emergency shutdowns, unplanned maintenance and reduced service-life performance. ML utilizes data and algorithms for predictive analyses that could influence manufacturing decisions and life-cycle management. An emerging field is the application of ML to structural welds due to its potential to reveal additional insights into design and integrity compared with traditional methods. Using welded joints as a case study, this paper reviews the state-of-the-art ML research spanning the areas of weld quality classification, welding parameter optimization, non-destructive testing (NDT) and residual stress assessment. This paper describes the stages of ML model development and application in existing weld research, including the stages of data collection, training algorithms and evaluation metrics for various weld integrity assessments. A key lesson from current research is the need for ML models to be developed using a wider array of datasets, i.e. more input features and entities, thereby highlighting the need for a datacentric approach for the lifecycle management of engineering components that could have considerable safety and cost benefit for the design, operation, and decommissioning of engineering plant.