OES structural feature based fault detection method for plasma etching

Optical Emission Spectra (OES) is a widely used signal in plasm etching. In this paper, an OES structural feature based fault detection method is proposed. Firstly, a template of normal OES curves is extracted via non-negative matrix factorization. Then singular points of the curves are detected by local matching based on the template. The magnitude and occurrence time of these singular points form a structural feature vector, which is a quantitative and simplified description of the curve´s shape. Lastly, one-class SVM is introduced for generate a fault detection model based on structural feature vectors from normal OES curves. Experiments on an industrial benchmark dataset show that the proposed method is effective.