Incremental perceptual vibration hashing autoencoder for machine condition monitoring

Feature extraction is of significant importance for machine condition monitoring, which extracts the potential machine condition information from monitored data. Autoencoder can extract the inherent structural feature in data, thus widely used for feature extraction from vibration signals. However, the training of autoencoder is a time-consuming process. Especially when facing a practical condition monitoring application in which numerous abnormal conditions emerge successively, an autoencoder needs to be retrained to adapt to the new faulty data, thus the training time multiplies. This paper develops the incremental perceptual vibration hashing autoencoder (IPVHAE), dedicated to mapping the inherent machine condition information in vibration signals into a compact, low-dimensional machine condition hash (MCH). An incremental learning approach based on parameter constraints is employed to facilitate the retraining of the IPVHAE on new data. The parameters of an autoencoder only need to be updated with the new data instead of all the data under available machine conditions. The effectiveness of the developed algorithm is verified using the Case Western Reserve University bearing dataset. Results show that the machine condition hash generated by the IPVHAE can balance low dimensionality with high discriminability, meanwhile minimizing the time and computing cost during the incremental learning process.