| Abstract |
Predictive maintenance strives to maximize the
availability of engineering systems. Over the last decade, machine
learning has started to play a pivotal role in the domain to predict
failures in machines and thus contribute to predictive maintenance. Ample approaches have been proposed to exploit machine
learning based on sensory data obtained from engineering systems. Traditionally, these were based on feature engineering from
the data followed by the application of a traditional machine
learning algorithm. Recently, also deep learning approaches
that are able to extract the features automatically have been
utilized (including LSTMs and Convolutional Neural Networks),
showing promising results. However, deep learning approaches
need a substantial amount of data to be effective. Also, novel
developments in deep learning architectures for time series have
not been applied to predictive maintenance so far. In this paper,
we compare a variety of different traditional machine learning
and deep learning approaches to a representative (and modestly
sized) predictive maintenance dataset and study their differences.
In the deep learning approaches we include a novel approach that
has not been tested for predictive maintenance yet: the temporal
convolutional neural network. We compare the approaches over
different sizes of the training dataset. The results show that,
when the data is scarce, the temporal convolutional network
performs better than the common deep learning approaches
applied to predictive maintenance. However, it does not beat the
more traditional feature engineering based approaches. |