In order to solve the problem of imbalance between positive and negative samples in liquid rocket engine fault diagnosis, and to enable adaptive fault detection during engine steady working state, a anomaly detection model based on fast incremental one-class support vector machine was established. Feature engineering method was adopted to extract features from sensor-obtained multivariate time series. Through incremental leaning, the one-class support vector machine model was improved and applied to liquid rocket engine anomaly detection. The one-class support vector machine detection model was endowed with adaptability for various engine individuals and multiple working conditions, while increasing computing speed. The analysis results of multiple hot test data show that the model is effective, fast-training and practically valuable.