Autoencoders for Quality Control and Anomaly Detection in Smart and Autonomous Systems

Anomaly detection plays a crucial role in the evolution of modern industry, where interconnected systems, pervasive sensing, and real-time analytics are central to achieving efficiency and resilience. In recent years, deep learning has been deeply studied and widely applied in the fields due to its powerful feature learning ability. The deep learning-based detection method integrates feature learning into the process of building model, so that the model automatically learns deeper features through massive data instead of designing custom feature extractor. This work is aimed to study, design and develop appropriate deep learning models to implement effective anomaly detection in industry 4.0 context, with a focus on robotics applications. The neural network model most extensively used in this work is based on the autoencoder architecture, which has proven effective across all the contexts in which it has been tested, due to its ability to perform dimensionality reduction and feature learning, both in unsupervised settings, where it captures intrinsic data representations without labeled information, and in supervised scenarios.