The United Nations identify three pillars in sustainable development: the economic pillar, the social pillar and the environmental pillar. This thesis is a collection of four papers with each paper presenting an advancement in machine learning for the sustainability of manufacturing systems. The first paper considers the application of machine learning to address the environmental pillar, the second and third paper consider advances for the economic pillar, the fourth paper considers both the environmental and the economic pillars. Traditional manufacturing has two main flaws in terms of sustainability: it relies on raw materials extracted from finite natural reservoirs and it labels as ``waste'' any product that has lost its usefulness. To address both the issues, the first paper identifies an emerging computer-vision-enabled material monitoring technology and it provides a survey of works relevant for its development. Several greedy algorithms for unsupervised variable selection have been proposed in the past without a systematic benchmarking. Therefore, in the second paper we review and provide a comparative study of these methods. Their application is of interest in the semiconductor manufacturing industry to reduce the cost of silicon wafer monitoring. The third paper primarily focuses on the same semiconductor application as second one. However, noting that unsupervised dimensionality reduction and variable selection can be both performed through an autoencoder, the second paper proposes a novel framework to design autoencoders with reduced computational complexity. In the fourth paper we improve the generalization and reduce the training time of the multi-batch L-BFGS to train small convolutional network classifiers implementing a development-based increase of the memory size. Two target manufacturing applications are considered: anomaly detection during silicon wafer etching and material recognition of waste objects. Finally, this thesis discusses the potential impact of machine learning on the three sustainability pillars and presents a framework for a deeply sustainable deployment of the machine learning systems here proposed.
