New statistical learning perspective for design of a physically interpretable prototypical neural network for machine condition monitoring

Abstract

Deep and shallow neural networks have lots of weights and hyper-parameters to make them magical and mysterious and their applications for machine condition monitoring appear in exponentially growing publications. In the domain of machine condition monitoring, it is a good time to rethink one critical question existing in state of art neural networks. How to construct a prototypical neural network and physically interpret its weights and correlate them with physically interpretable fault features to support machine health conditions identified from the prototypical neural network? This short communication starts with a statistical learning perspective to artificially design a physically interpretable prototypical neural network that has an exactly same structure with a standard one-hidden-layer neural network. Subsequently, it is shown that weights of the physically interpretable prototypical neural network are highly correlated with fault characteristic frequencies and informative frequency bands. Following, this short communication timely answers the aforementioned question and hopes to motivate readers to specifically design more unique and powerful neural networks for physical condition monitoring and classifications to enrich the domain of machine condition monitoring. Simultaneously, this short communication highlights an essential difference between machine learning algorithms in the domain of machine condition monitoring and those in other domains is that physically interpretable weights are necessary fault features in the domain of machine condition monitoring to scientifically support machine health conditions identified from neural networks.