Parametric Study of Anomaly Detection Models for Defect Detection in Infrared Thermography

In the current NDT 4.0 revolution, machine learning and artificial intelligence have emerged as the major enablers for non-destructive testing and evaluation (NDT&E) of industrial components. However, recent developments in active thermal NDT (TNDT) support its use as a practical method for checking a range of industrial components. Additionally, recent post-processing research in TNDT has developed several machine learning models to replace human interaction and offer automatic defect detection. However, the smaller area of the flaws and their related few thermal profiles than the wide sound area, leading to imbalanced datasets, make it difficult to train a supervised deep neural. Recently added to TNDT are anomaly detection models and one-class classifiers, both of which are commonly applied machine learning models to real-world issues. The accuracy and other important metrics in autonomous defect detection are influenced by the hyper-parameters of these models, such as contamination factor, volume of training data, and initialization parameter of the relevant model. The current paper investigates how initialization parameters affect these models' TNDT capabilities for automated flaw detection. Using quadratic frequency modulated thermal wave imaging (QFMTWI), a carbon fiber-reinforced polymer specimen with variously sized artificially produced back-holes at different depths is examined. A good hyper-parameter for automatic flaw identification is chosen after qualitatively comparing testing accuracy, precision, recall, F-score, and probability.