Options
2020
Master Thesis
Titel
Investigation and Implementation of Machine Learning Algorithms for Condition Monitoring of Piezoelectrically Driven Micropumps
Abstract
Micropumps offer a wide range of applications. Especially in medical technology, such as drug dosing applications, there are high requirements to dosing accuracy and reliability of the micro-fluidic systems. To monitor the dosing accuracy, pressure or flow sensors are classically used. In this thesis piezoelectrically driven micropumps are investigated. These pumps consist of a piezoceramic disk mounted on a diaphragm as the actuator. Since cause and effect can be interchanged in the piezoelectric effect, such a disk naturally also possesses sensor properties in addition to the actuator capability. From the theory of electromechanical coupling it can be deduced that a sensor component, which provides information about the pressure curves in the pump chamber, can be measured in the decaying drive signal of the micropumps. The investigations in this thesis have shown that this sensor signal can be used by means of intelligent algorithms to enable condition monitoring of the pump. The advantage of this approach compared to the classical approaches with external sensors is that it allows for a truly collocated, concurrent and single-stroke monitoring of the pump. This condition monitoring here includes scenarios in which back-pressures in the range of 0 to 250 mbar were applied at the pump outlet. For the support vector machine (SVM) classification algorithm used, features in the signals first had to be identified manually. This was not necessary for the neural networks, as these allow classification on the raw data. With the hand-selected features and the SVM, an accuracy of over 90% on the test data was achieved. The best convolutional neural network achieved an accuracy of over 80%. With the successful implementation of the condition monitoring, the foundation for a wider-ranging monitoring of the pump parameters has been laid.
ThesisNote
München, TU, Master Thesis, 2020
Verlagsort
München