Colloquium: Maurice de Bruin (C&O)

This study presents a multi-source domain adaptation (MSDA) model for predicting the remaining useful life (RUL) of bearings across different operating conditions and bearing types using only frequency domain data. Advanced RUL prediction models often rely on time, frequency, and time-frequency domain data, but some systems only collect frequency domain data, limiting the available degradation information. To address this, the proposed MSDA model uses a common feature extractor to capture domain-invariant features and domain-specific regressors for RUL prediction. By minimizing the maximum mean discrepancy between source and target domain features and aligning domain-specific RUL predictions, domain adaptation can be achieved. Experimental results demonstrate that, even with only frequency domain data, the MSDA model achieves RUL prediction performance (in terms of RMSE and score) comparable to advanced models that utilize both frequency and time domain data.

Supervisor: I.I. De Pater