Repozytorium publikacji - Politechnika Gdańska

This paper describes a condition-based monitoring system estimating DC motor degradation with the use of an autoencoder. Two methods of training the autoencoder are evaluated, namely backpropagation and extreme learning machines. The root mean square (RMS) error in the reconstruction of successive fragments of the measured DC motor angular-frequency signal, which is fed to the input of autoencoder, is used to determine the health indicator (HI). A complete test bench is built using a Raspberry Pi system (i.e., motor driver controlling angular frequency) and Jetson Nano (i.e., embedded compute node to estimate motor degradation) to perform exploratory analysis of autoencoders for condition-based monitoring and comparison of several classical artificial intelligence algorithms. The experiments include detection of degradation of DC motor working in both constant and variable work points. Results indicate that the HI obtained with the autoencoders trained with the use of either training method is suitable for both work points. Next, an experiment with multiple autoencoders trained on each specific work point and running in parallel is reviewed. It is shown that, in this case, the minimum value of RMS error among all autoencoders should be taken as HI. Furthermore, it has been shown that there is a nearlinear relationship between HI and the difference between measured and reconstructed angular-frequency waveforms.

Autorzy

• Krzysztof Włódarczak,
• mgr inż. Łukasz Grzymkowski,
• dr hab. inż. Tomasz Stefański link otwiera się w nowej karcie