A Hybrid DCNN ECOC-SVM System for Detecting Semiconductor Wafer Defects Using Histogram Equalization

In the semiconductor manufacturing sector, identifying defective wafers is a challenging task. Several techniques are conducted for detecting defective and non-defect semiconductor wafers. However, these developed techniques are not able to detect tiny defects in defective wafer maps accurately. Also, these techniques are unable to identify complex patterns in semiconductor wafers precisely. To overcome these limitations this paper proposes a hybrid approach with histogram equalization. The hybrid system comprises the deep convolutional neural network (DCNN) and error-correcting output code-based support vector machines (ECOC-SVM) classifier model. In this proposed method, histogram equalization is performed on the real wafer map dataset of WM811K to improve the visibility of tiny defects in wafer maps. The DCNN architecture extracts valuable features from histogram-equalized wafer map images. The swish function is used in the DCNN architecture and DCNN extracts features from complex defect patterns of wafer maps. The extracted features are used for training and testing purposes of the classifier model. The error-correcting capability of the ECOC-SVM classifier model improves the overall robustness and accuracy of the hybrid system for detecting eight types of wafer defects and one non-defect type wafer. The testing accuracy of the hybrid model is 96.1%. In this paper, comparative studies are given between the proposed system and another hybrid system that does not use the histogram equalization technique and swish function in the DCNN. The findings demonstrate that the proposed system accurately detects the tiny defects in defective wafers and identifies the complex patterns in semiconductor wafer maps. This paper also provides a comparative analysis of the proposed system with relative research methods.