The detection of hair cracks is one of the key challenges to improve wafer-processing stability. Contrary to other defects on the wafer-edge, hair cracks have a very small geometric footprint, making them hard to detect for measurement systems. This raises the demand for a powerful data analysis tool, which can extract the relevant information even in low signal-to-noise ratio scenarios. In this paper, we investigate an approach for hair crack detection using a laser-based wafer edge inspection device and deep neural networks to analyze and classify the measured data. We propose different pre-processing methods for the raw measurement data, to improve the learning behavior of the networks. The results show that a substantial improvement, in both detection rate and false positive rate, can be achieved by appropriate pre-processing of the measured data.
展开▼
