Overview
The findings of this project demonstrate compelling results in identifying pipetting anomalies in pressure time-series data using various machine learning techniques. However, the methods employed were unable to detect rare normal instances such as partial coagulation. To address this issue, more advanced and refined approaches are needed.
The objective of this master's thesis is to explore the integration of physical models of the pipetting system into neural networks to enhance anomaly detection and achieve a more comprehensive understanding of the pipetting process.
In the initial phase, a physical model of an air-pressure pipetting system will be constructed based on the Bernoulli equation. This Bernoulli equation is intended to be integrated into a neural network to estimate the resulting pipetting volume based on a pressure curve and piston motion. The resultant model should be capable of distinguishing between abnormal and rare normal instances.
In the subsequent phase, we aim to investigate the extent to which the resulting physically-informed model can provide deeper insights into the pipetting process. Furthermore, additional techniques and methods will be explored to enhance the interpretability of the pipetting process for non-experts.
Ultimately, a Minimum Viable Product will be introduced to showcase a potential MLOps architecture for the development of pipetting monitoring algorithms and to illustrate how the resulting models can be employed to devise diagnostic tools.
