Embedded acoustic fault monitoring for water pumps

Maintaining pumps, especially waste water pumps, is quite a cost-intensive task. The proper operation must be guaranteed under all circumstances. Accessing the pumps, however, is not easily done, as they are submerged in waste water. This paper describes the development of a fault classification system based on acoustic signals, with the focus on finding an optimal feature space and an efficient classifier in terms of energy and memory footprint. Those characteristics are especially important when the classifier has to run on a resource-constrained platform like an embedded system. In this paper, we show how the combination of a dimensionality reduction and a feature selection can be used to reduce the memory footprint of the entire system by 79%, with no significant loss in test set accuracy. With this strategy, a neural network with thirty input features was deployed on an embedded system with a memory footprint for the classification parameters of only 22.94 kB.