Abstract: In order to improve one-class support vector machine (OCSVM)’s performance, that is OCSVM’s training efficiency and decision precision, a double threshold OCSVM online detection (DTOOD) algorithm is proposed. In DTOOD, the OCSVM detection model with two-layer thresholds can separate outliers into non-margin support vectors and real abnormal samples. And the detection model can be updated online adaptively without real abnormal samples as they are omitted in future training sets. Meanwhile, sequential minimal optimization algorithm for OCSVM is introduced to improve the training efficiency. DTOOD is applied to the analysis of a liquid rocket engine turbopump historical vibration data, and the results show that DTOOD can detect the faults of the turbopump very well without any false alarm. And the computation is fast enough to assure DTOOD’s ability of real time fault detection.

Key words: Novelty detection, One-class support vector machine, Sequential minimal optimization, Turbopump fault detection