To validly detect the anomalies of parameters in the engine test, a fault detection algorithm of engine based on time series data mining is proposed. The parameter time series are transformed into symbolic strings by a representation method based on shape features. The stable states and transition states are extracted from the parameter time series according to symbolic semantics. Meanwhile, the detection algorithm of abnormal pattern from the stable states is realized by similarity measurement between time series based on statistic features, combined with the most unusual pattern discovery method. The results of numerical experiments show that the new method validly detects the fault of engine and has the better robustness than the traditional method.