Noise-robust Early Detection of Cooling Fan Deterioration with a Variational Autoencoder-based Method

Electrical equipment usually requires air cooling by fans, and it is important to detect early signs of fan deterioration. However, when using sensors to detect deterioration of fans in the real world, the sensor signal is often mixed with external noise. In this paper, we analyze the deterioration trends in the operating sounds of cooling fans actually used for 3 and 7 years in uninterruptible power supplies. Furthermore, we propose a deep learning-based method for noise-robust early detection of deterioration based on fan operating sounds called variational autoencoder-based deterioration estimation (VAE-DE). Existing anomaly detection methods based on autoencoders often react to external noise mixed into deterioration trend, which causes false positives. While another conventional method using variational autoencoders has the property of being robust against noise, it is not reactive to slight sound changes that accompany fan deterioration. If a deterioration trend is known, noise-robust early detection of deterioration can be achieved through learning using a loss function that includes a term for separation from normal sounds. Experiments investigating deterioration sounds in a fan showed changes in latent variables for early deterioration sounds and corresponding increases in the anomaly score. We also show that anomaly scores can be robustly estimated even in the presence of noise. The results reveal that the proposed method performs better than conventional methods.