Neuromorphic Audio Dataset for Predictive Maintenance in Peristaltic Pumps

Predictive maintenance (PdM) applications are becoming more and more popular, in part because Artificial Intelligence (AI) algorithms are becoming more powerful and affordable. However, this method requires a lot of data which need to be specific to the application. Only a few audio datasets on the topic are available, most of them are focused on generic systems, and none to the best of our knowledge has been created from a neuromorphic engineering approach, which offers benefits in terms of power consumption and robustness among others. We used a Neuromorphic Auditory Sensor (NAS) to record a neuromorphic audio dataset that covers two different scenarios of PdM in actual equipment from a commercial device. We focused on faults due to deterioration of the peristaltic pumps caused by aging and also on detecting the presence of undesired air bubbles in the hydraulic circuit. The audio information was processed in Address Event Representation (AER) format, which was then converted into a set of labeled sonogram images of 0.5 s each. We used a window overlap of 50%, which enlarges the number of samples and improves training results. All of these data are now available as a public dataset. We did a classification test for both scenarios by training a DenseNet121 model that achieved a F1 weighted average of 0.99 for the aging scenario and 0.71 for the bubble scenario. Further processing should be done in the future to improve the results on the bubble scenario.