Pressure Pulse Classification for Flow Disturbance Detection in Micro Diaphragm Pumps

Micropumps have the potential to complement or even replace infusion pumps used in hospital and home care environments. For safe and reliable operation, error detection must be available for these devices. In this work we present a machine learning based approach that classifies pressure pulses created by a micropump into several operation states (normal operation, air-in-line, upstream- and downstream occlusion). An automated setup to generate these flow states and to automatically read and label the corresponding pressure pulse sensor data was built. Using this setup, a dataset consisting of 12.000 pressure pulses from five micropumps was recorded. Characteristic statistical features for the four flow states were extracted from this dataset using TSfresh. Using these statistical features, a decision tree model was trained. The model achieved a classification accuracy of 93%.