Smart manufacturing systems are equipped with various sensors to collect engineering variables for monitoring of manufacturing process quality, such as fault detection. Multiple fundamental difficulties, however, arise in the fault detection including high nonlinearity, high dimensionality, unequal batch length, unsynchronized batch trajectory and imbalance in the ratio of fault data and normal data in the sensor time series recordings. To capture the dynamics, nonlinearity and non-Gaussianity in the recordings, we design a feature extraction method using bispectral features and Principal Component Analysis (PCA); to address the class-imbalance issue, we use one-class classifiers. We evaluate the proposed method using recall, precision, G-means and Area Under ROC Curve (AUC) on a publicly available benchmark of a real semiconductor etching dataset for fault detection. It is demonstrated that bispectral features processed further by PCA can significantly improve the fault detection accuracy and lower the false alarm rate of one-class classifiers. The k-Nearest Neighbor (k-NN) can achieve the 100% fault detection rate and zero false alarm in the 10-fold cross validation test. Published by Elsevier Ltd.