Improved anomaly diagnosis of production facilities by combining Autoencoder with spectral characteristics

In a previous study, we had proposed a method for detecting abnormalities using an Autoencoder. It considers the amplitude value of each frequency spectrum while performing the FFT analysis of time-series data. In this study, we applied this method to detect abnormalities in pressure washers. In particular, we verified its detection accuracy on the artificially-generated anomaly data. The results showed a deteriorated detection performance for a varying spectrum amplitude near the resonance frequency. Therefore, in addition to the conventional autoencoder, the proposed method further improves anomaly detection accuracy by treating the spectrum as a lumped spectrum in a predetermined frequency range. The effectiveness of the proposed method was verified using data obtained from equipment anomalies. In conclusion, the proposed anomaly-detection method can robustly cope with frequency fluctuations near the peaks and detect anomalies with a accuracy higher than that of the conventional anomaly detection method, which uses only an autoencoder. Thus, the proposed method can detect anomalies even before factory become aware of them.