Autoencoder-Based Iterative Modeling and Multivariate Time-Series Subsequence Clustering Algorithm

This paper introduces an algorithm for the detection of change-points and identification corresponding subsequences in transient multivariate time-series data (MTSD). The analysis such has become increasingly important due to growing availability many industrial fields. Labeling, sorting or filtering highly measurement training Condition-based Maintenance (CbM) models is cumbersome error-prone. For some applications it can be sufficient filter measurements by simple thresholds finding based on changes mean value variation. But a robust diagnosis component within group example, which complex non-linear correlation between multiple sensor values, approach would not feasible. No meaningful coherent data, could used CbM model, emerge. Therefore, we introduce that uses recurrent neural network (RNN) Autoencoder (AE) iteratively trained incoming data. scoring function reconstruction error latent space information. A model identified subsequence saved recognition repeating as well fast offline clustering. evaluation, propose new similarity measure curvature more intuitive clustering metric. comparison with seven other state-of-the-art algorithms eight datasets shows capability increased performance our cluster MTSD online conjunction mechatronic systems.