A Novel Deep Soft Clustering for Unsupervised Univariate Times Series

Electromechecanical actuators in the aerospace industry are gradually replacing hydraulic ones. In these circumstances, prognostics and health management are innovative frameworks to ensure better safety on board, especially in flight controls where jamming is dreaded. It allows the user to assess and predict system health in real-time. The first step is to collect temporal data from the monitored actuator and perform a data mining procedure to gain insight into its current health. Clustering encompasses several data-driven methods used to reveal patterns. However, getting a set of classes usually requires providing the algorithm with prior knowledge, such as the number of groups to seek. To avoid this drawback, we have developed a clustering algorithm using a deep neural network, as its core, to get the number of groups in data associated with their likelihood. Temporal sequences are reshaped into pictures to be fed into an artificially trained neural network: U-NET. The latter outputs segmented images from which one-dimensional information is extracted and filtered, without any need for parameter selection. A kernel density estimation finally transforms the signal into a candidate density. This new method provides a robust clustering result coupled with an empirical probability to label the times series. It lays the groundwork for future training of diagnosis and prognosis structures in the PHM framework.