Aiming at the difficulty of fault classification caused by the low utilization rate of effective information in high-dimensional fault datasets， linear principal component discriminant analysis （LPCDA） is proposed to reduce the dimension of the fault datasets. The feature of this method is to integrate the idea of discriminability between classes and principal component calculation into the linear discriminant analysis（LDA） algorithm. Through these two ideas， the algorithm has the ability to eliminate relevant information and redundant features. Thus the valuable fault state information and the main components of the characteristics that can reflect the running state of the machine can be retained better. The algorithm has the characteristics of eliminating the relevant information， redundancy feature and the main components of the retained feature in the high-dimensional fault datasets， which has the function of reducing the difficulty of fault classification and improving the accuracy of automatic identification. This research can provide a theoretical reference for effectively reducing the scale of high-dimensional fault datasets， improving the ability to mine effective information and reducing the difficulty of fault classification.