Non-stationary condition monitoring through event alignment

We present an event alignment framework which enables change detection in non-stationary signals. Classical condition monitoring frameworks have been restrained to laboratory settings with stationary operating conditions, which are not resembling real world operation. In this paper, we apply the technique for non-stationary condition monitoring of large diesel engines based on acoustical emission sensor signals. The performance of the event alignment is analyzed in an unsupervised probabilistic detection framework based on outlier detection with either principal component analysis or Gaussian processes modeling. We are especially interested in the true performance of the condition monitoring performance with mixed aligned and unaligned data, e.g. detection of fault condition of unaligned examples versus false alarms of aligned normal condition data. Further, we expect that the non-stationary model can be used for wear trending due to longer and continuous monitoring across operating condition changes