Fault Diagnosis of Semiconductor Handlers using Multiple Autoencoders

This paper introduces a denoising autoencoder designed to distinguish between normal vibration data and various types of fault vibration data generated in semiconductor handlers. Unlike previous studies that windowed all time-series data based on fixed intervals, this study proposes a method of windowing data at the unit of action, using semiconductor handler time stamps for initial segmentation. The data are then uniformly standardized via linear interpolation. Additionally, to simplify the time-series vibration data patterns, the Hilbert transform and Butterworth low-pass filter are applied. The deep-learning model is structured with parallel components with both time and frequency domains, facilitating detection across various fault scenarios. In contrast to studies processing time-series data into fixed sizes, this approach demonstrates strong performance despite employing a relatively simple deep-learning model.