Pump Diagnostics using Machine Learning

Finished: 2020-05-13

MSc assignment

This assignment is a collaboration between Controllab Products B.V. and the Robotics and Mechatronics (RaM) department of the University of Twente. Controllab has developed and implemented a controller for a pump for a customer using model-based design. This pump has proven to be very successful. However, sometimes a pump is sent back to the customer. At the moment a log file with important values (such as internal pressures and positions of motors) can be read from the pump. A limited number of experts is able to use this log file to assess where there is a potential problem in the pump. It appears that the pump is often returned incorrectly or that the problem can also be solved on location. Controllab has now been asked to extend the developed software with a extensive diagnostics to detect possible problems and accompanying instructions for solving them.

The purpose of the assignment is to develop machine learning models that are able to recognize problem situations of the pump in real-time, estimate the seriousness of the problem and come up with a set of instructions on how to solve the possible problem. With the help of the 20-Sim model of the pump, that was developed by Controllab, all kinds of error situations can be replayed and log data can be generated. With this generated log data, machine learning models can be trained. In this way it should be possible for untrained users of the pump to recognize the problems and solve them using the given instruction set.