Options
2021
Master Thesis
Titel
Flow Disturbance Detection in Micro Diaphragm Pumps: Automated Data Acquisition Setup and Time Series Classification with Machine Learning
Abstract
Miniaturised infusion pumps with active dosage control are used as insulin pumps or as intrathecal pain pumps. Furthermore, they are evaluated as a general drug delivery system for metronomic cancer therapy or used for drug research with laboratory animals. The requirements for reliability are high. Flow disturbances like an occlusion or air in the infusion line can lead to unexpected dosage behaviour and thus impair the safety of miniaturized infusion pumps. Based on a literature review about flow disturbance detection in infusion pumps, it is decided to detect disturbances with time series classification of transient pressure response measurements. In this work, an infusion pump based on a diaphragm micropump with a piezoelectric actuator is considered. The pulse response of a single pump stroke is measured by acquiring the pressure drop across a flow restricting capillary with a differential pressure sensor. To generate a comprehensive data set an automated measurement setup is built. It simulates flow disturbances with electrically actuated valves, detects air in the system and measures the pressure drop across the flow restricting capillary. Data logs representative of four classification categories, which are normal operation, upstream and downstream occlusion and air in line, can be acquired. This way time series data representative of four classification categories for five pumps under test is generated. In total the data set used for machine learning contains 12000 time series. With this data two time series classification models are trained and evaluated. A feature based decision tree that uses the median, maximum and minimum of the time series achieves a test set classification accuracy of 93 %. A computationally more complex time series forest for classification achieves a 99.6 % test set classification accuracy. The feature based decision tree is implemented as an online classification running on the automated measurement setup. To realise time series classification for disturbance detection in a real life use case the measurements have to be repeated with the specific drug and setup. Additionally the robustness of the disturbance detection with respect to environmental factors, like a change in back pressure has to be quantified.
ThesisNote
München, TU, Master Thesis, 2021
Author(s)
Verlagsort
München