Signal processing informed deep learning for failure detection in a fleet of multi-stage planetary gearboxes with limited knowledge about characteristic frequencies

Condition monitoring of multi-stage planetary gearboxes is a complex challenge given the fact that gears the large number of rotating subcomponents. Typically, the large number of gears creates many harmonic excitations masking bearing signatures. Different state-of-the-art harmonic removal methods, e.g. cepstrum liftering, are available. Such methods have been shown to be automatable. However, exact characteristic frequency values are not always known for such gearboxes in commercial systems. Estimation of gear teeth numbers has been shown in literature. Bearing frequency determination is much more challenging. Deep learning methods can offer a solution. Once the harmonic content is removed, focus can be on the detection of modulations linked to bearing problems. Spectral coherence methods have shown to be highly reliable for such detection. However, if no info is available about normal behaviour in the coherence maps it is essential to detect which modulations are changing over time. This paper investigated the use of deep learning auto-encoders trained on spectral coherence maps as core component in an anomaly detection framework to identify changes in modulations. The auto-encoders are trained with large sets of healthy data. In this way we maximally use available data and avoid the need of large sets of labelled failure data. Typically, such data is not available for most operators. To illustrate the methodology data of six offshore wind turbines is used.