A Fast SVM Method Based on Two-dimensional Energy Features for Turbopump Fault Detection

A fast support vector machine (SVM) method based on two-dimensional energy features is presented for real-time fault detection of a turbopump. Primitive train samples set is calculated in steps of 70, with the time-domain energy and energy change rate as the fault features. The conditional positive definite kernel function is used to calculate distance between normal and fault samples in primitive train samples set, and boundary train samples set is selected to calculate support vectors and construct the decision function. Training on 1000 normal samples and 1000 fault samples took only 0.42 s. For the detection data, the alarm is 3.02 s in advance of the shutdown moment. The training speed and classification speed are improved with good real-time performance and accuracy.