Noise-aware AI methods for robust acoustic monitoring of bearings in industrial machines

Traditionally, companies have relied on vibration based condition monitoring technologies to implement condition based maintenance strategies. However, these technologies have drawbacks, such as the requirement of contact accelerometers. As an alternative, acoustic condition monitoring is non-invasive and allows for easy deployment. Furthermore, the use of microphones potentially enables the monitoring of multiple components using a single sensor, making the monitoring system scale better with machine or production complexity. However, microphone signals typically show a low signal-to-noise ratio (SNR), impacted by the high level of background noise which is often present in industrial environments. Particularly, the traditional method for monitoring the health condition of rolling element bearings, based on assessing whether the squared envelope spectrum of the bearing signal exceeds a given threshold at the fault frequencies, cause too many false positives when applied directly to microphone signals. It is therefore crucial to develop strategies to increase the robustness of acoustic monitoring methods. In this paper, we present and evaluate two data-driven strategies to robustly diagnose bearing faults from a microphone signal. Our proposed strategies are noise weighting based on the detection of background noise, and an artificial intelligence (AI) model that uses as input a combination of the traditional bearing fault frequencies and the mel spectrum of the microphone signal. These methods leverage both domain knowledge and data-driven techniques to increase the detection robustness. Our approach is implemented as a model trained and tested on bearing accelerated lifetime tests performed in the Smart Maintenance Lab setup at Flanders Make. Our results show that the use of our proposed strategies leads to significant improvements in diagnostic performance and time to first detection over noise-unaware acoustic monitoring methods.