Lectura de tesis

Título: Study, Design and Evaluation of Neuromorphic and Edge Artificial Intelligence Systems for Predictive Maintenance

Doctorando: Juan Manuel Montes Sánchez.

Directores: Ángel Fco. Jiménez Fernández y Saturnino Vicente Díaz.

Tutor: Ángel Fco. Jiménez Fernández.

As technology becomes more complex thanks to the advances in computing and automation, industry balances between the cost of running these newly developed systems and the advantages they offer. In the last few years, the spread of Artificial Intelligence (AI) helped to integrate new strategies that aim to reduce failures and make maintenance interventions more efficient. Predictive Maintenance (PdM) is already starting to change the life cycle of every industrial equipment. However, the use of powerful AI algorithms is rising some concerns due to their high consumption in resources. Industry can still benefit from these methods without compromising its improve in efficiency thanks to some modifications that make AI more specific and easy to run into power efficient devices, minimising data transfer while offering similar accuracy and results. This is commonly known as Edge AI, and it is starting to be used in PdM applications. In this thesis, we studied the current state of art of Edge AI in PdM scenarios. After some contributions in the development of a novel and patented biomedical robot, we identified some unresolved maintenance issues related to one specific part of it that could potentially benefit from PdM techniques. We prepared a recording scenario to obtain properly labelled multi-sensor datasets that were made public. Using that data, we explored several ways of implementing AI-based algorithms that could detect failure, with efficiency as our main objective. In our first approach, we tested a method based on Recurrent Neural Networks (RNNs) in which we optimised not only the network itself, but also the nature of the input data and the selection of a target deployment device. We also made a totally different approach based in audio signals in which we processed the data with a power efficient neuromorphic audio sensor (NAS) to see if we could extract the audio features for PdM sound classification. During this process, a new neuromorphic audio dataset for PdM was recorded and published. Finally, since industry also made its move in Edge AI with several new software tools, we designed a comparison between the most popular ones. In this comparison, we tested cost, features, performance, and usability of each one when used in the same PdM scenario we studied before. The results and conclusions of each step of this work are already published in different forms: one patent, two journal articles, two conference proceedings and two public datasets. We also present them here as a single document, with some additional content which serves as a link between them for better understanding.