Development of Fault Diagnosis Model based on Semi-supervised Autoencoder

The maintenance paradigm based on PHM (Prognostics and Health Management) technology, utilizing big data to predict process conditions through manufacturing intelligence, is rising. However, in most industries, there is lack of accurate labeling of sensor data, posing challenges in data utilization due to the significant cost of labeling tasks. Consequently, recent research has focused on semi-supervised learning methodologies, which are applicable in label-absent scenarios. Especially, there is a growing emphasis on semi-supervised autoencoder, which learns both labeled and unlabeled data simultaneously. Also, there is a demand for the development of fault diagnosis models for essential components, such as bearings in most mechanical systems. Vibrational data is actively being integrated with artificial intelligence for application in bearing fault diagnosis frameworks. Nonetheless, diagnosing the condition of bearings inside machine systems, especially within the machine tool spindle, remains challenging, as the labeling of collected data causes significant costs. Therefore, this paper aims to develop a fault diagnosis model for unlabeled bearings in machine tool spindle using a semi-supervised autoencoder. Initially, a monitoring system of bearing simulator that imitates a machine tool spindle bearing was constructed, and vibration signals from both normal and fault bearings were collected based on this system. Subsequently, a semi-supervised autoencoder model was developed to construct a fault diagnosis model using labeled simulator data and unlabeled machine tool spindle bearing data. To evaluate the model, additional data of normal and fault bearings in machine tool spindle were collected, and the performance of the model was compared with a conventional fault diagnosis model based on 1D-CNN.