Incorporating Uncertainty into Unsupervised Machine Learning for Cyber-Physical Systems

摘要

In the field of Cyber-Physical Systems (CPS), the early detection of anomalies is crucial to avoid future faulty behaviors, e.g. preventing downtimes or decreasing product qualities. As a solution, unsupervised machine learning can be used to learn models of the historic system behavior and consequentially detect deviations from these models. Since CPS data are high dimensional time series, suitable approaches such as Long Short-Term Memory (LSTM) neural networks are good solution candidates.CPSs have specific requirements for such machine learning algorithms. Learned models must be especially useful in closed control loops, i.e. without human supervision. For this, it is essential that the uncertainty about model predictions is also part of the learned models. In order to incorporate such uncertainties, we modify the mean squared error loss function used by LSTM. This paper also analyses the solution on artificial and real data.