Non-stationary Condition Monitoring of large diesel engines with the AEWATT toolbox

We are developing a specialized toolbox for non-stationary condition monitoring of large 2-stroke diesel engines based on acoustic emission measurements. The main contribution of this toolbox has so far been the utilization of adaptive linear models such as Principal and Independent Component Analysis, as combined modeling and feature reduction methods.

These models describe the, say 1024 or 2048, acoustic emission samples per engine revolution, i.e. data are in the crank angle domain. In this framework we have applied unsupervised learning using only one feature – the log-likelihood of an example given the trained linear model. The setup is semi unsupervised, as model parameters are learnt from normal condition data only, thus the system is not directly capable of error identification.

However, it should be noticed that the adaptive linear models allow for some diagnosis based on the angular location of residual energy. Also, the framework can be extended, for instance by post modeling of repeated faults. Furthermore, we have investigated the problem of non-stationary condition monitoring when operational changes induce angular timing changes in the observed signals.

Our contribution, the inversion of those angular timing changes called “event alignment”, has allowed for condition monitoring across operation load settings, successfully enabling a single model to be used with realistic data under varying operational conditions-