A Comparative Study of Transfer Learning on CNN-Based Models for Fault and Anomaly Detection in Industrial Processes

Early anomaly detection in production processes is crucial to mitigate financial losses and safety risks associated with the production of defective goods and equipment failures. Traditional manual inspection methods are fraught with challenges like unreliability, inaccuracy, safety concerns, and higher costs. To address some of these issues, the paper explores the application of deep learning (DL) and transfer learning techniques in automating anomaly detection, thereby enhancing accuracy, safety, and objectivity. The study evaluated CNN-based pre-trained models focusing on two specific case studies: 3D printing of cylindrical and cubic objects, and corrosion detection in equipment and pipelines. The methodology involves analyzing, cleaning, and augmenting image datasets from these scenarios using seven different well-known models, followed by a comparative evaluation, based on effectiveness and efficiency, with a custom-defined model to ascertain the most optimal solution. The results show that transfer learning is a very effective solution in detecting anomalies in a broad spectrum of activities involving both product and manufacturing plant facility issues.