Anomaly score for rotational machines using conditional Variational Auto-encoder adaptable to speed changes

The basic way for detecting anomalies in rotating equipment would be to use the frequency spectrum of vibration. These days, however, rotating devices are generally driven by inverters, and changes in the rotational speed cause significant changes in the vibration spectrum. In order to detect anomalies regardless of dependence on operating conditions, AI technology that performs learning to capture normal conditions is useful. As a specific method, Conditional Variational Auto-encoder (CVAE), which takes rotational speed information as conditional inputs, is promising. We are trying CVAE using publicly available dataset provided by the University of Ottawa to search for optimal conditions. We report methods used and results obtained, including limiting training samples, increasing the number of conditions, interpolation and extrapolation of condition inputs.