A Deep Learning Model for Robust Wafer Fault Monitoring With Sensor Measurement Noise

Standard fault detection and classification (FDC) models detect wafer faults by extracting features useful for fault detection from time-indexed measurements of the equipment recorded by in situ sensors (sensor signals) and feeding the extracted information into a classifier. However, the preprocessing-and-classification approach often results in the loss of information in the sensor signals that is important for detecting wafer faults. Furthermore, the sensor signals usually contain noise induced by mechanical and electrical disturbances. In this paper, we propose the use of a stacked denoising autoencoder (SdA), which is a deep learning algorithm, to establish an FDC model for simultaneous feature extraction and classification. The SdA model can identify global and invariant features in the sensor signals for fault monitoring and is robust against measurement noise. Through experiments using wafer samples collected from a work-site photolithography tool, we confirmed that as the sensor measurement noise severity increased, the SdA's classification accuracy could be as much as 14% higher than those of the twelve models considered for comparison, each of which employed one of three feature extractors and one of four classifiers.