Predictive maintenance of induction motors using ultra-low power wireless sensors and compressed recurrent neural networks

In real-world applications - to minimize the impact of failures - machinery is often monitored by various sensors. Their role comes down to acquiring data and sending it to a more powerful entity, such as an embedded computer or cloud server. There have been attempts to reduce the computational effort related to data processing in order to use edge computing for predictive maintenance. The aim of this paper is to push the boundaries even further by proposing a novel architecture, in which processing is moved to the sensors themselves thanks to decrease of computational complexity given by the usage of compressed recurrent neural networks. A sensor processes data locally, and then wirelessly sends only a single packet with the probability that the machine is working incorrectly. We show that local processing of the data on ultra-low power wireless sensors gives comparable outcomes in terms of accuracy but much better results in terms of energy consumption that transferring of the raw data. The proposed ultra-low power hardware and firmware architecture makes it possible to use sensors powered by harvested energy while maintaining high confidentiality levels of the failure prediction previously offered by more powerful mains-powered computational platforms.