Condition monitoring with mean field independent components analysis

We discuss condition monitoring based on mean field independent components analysis of acoustic emission energy signals. Within this framework, it is possible to formulate a generative model that explains the sources, their mixing and the noise statistics of the observed signals. Using a novelty detection approach based on normal-condition examples only, we detect faulty examples with high precision. The detection is done by evaluating the likelihood that the model, trained with normal examples, generated the signals, compared to a threshold obtained with normal examples. Acoustic emission energy signals from a large diesel engine are used to demonstrate this approach. The experiment show that mean field independent components analysis detects the induced fault with higher accuracy than principal components analysis, while at the same time selecting a more compact model.