Development and Construction of a Parameterizable Condition Classification System for Electromagnetic Proportional Valves using Neuronal Networks

In this paper the development of a compact condition classification system for electromagnetic proportional valves is shown. It allows the generation of training data as well as a fast testing and comparison of different trained neuronal networks. By using quantization and pruning, a neuronal network with drastically reduced complexity has been created, so a FPGA implementation was possible. The developed and implemented network shows a very high classification rate and can distinguish 12 different false reasons of the valves. The system requires the measurement of the supply current only, which allows a simple integration of such a false detection circuitry into existing systems. In the future, the system can be modified easily, e.g. to use and test a hardware based AI accelerator instead of the FPGA implementation.