Diagnostic spatio-temporal transformer with faithful encoding

This paper addresses the task of anomaly diagnosis when underlying data generation process has a complex spatio-temporal (ST) dependency. The key technical challenge is to extract actionable insights from dependency tensor characterizing high-order interactions among temporal and spatial indices. We formalize problem as supervised discovery, where ST learned side product multivariate time-series classification. show that positional encoding used in existing transformer works serious limitation capturing higher frequencies (short time scales). propose new with theoretical guarantee, based on discrete Fourier transform. also discovery framework, which can provide readily consumable diagnostic information both directions. Finally, we demonstrate utility proposed model, DFStrans (Diagnostic Fourier-based Spatio-temporal Transformer), real industrial application building elevator control.