Abstract: Considering the fact that modeling a complex system such as the launch vehicle is particularly difficult, anomaly detection methods based on data driven are studied. This paper proposes a multi-strategy anomaly detection method based on statistics of probability density distribution. The unsupervised learning method is used to excavate the model for normal correlation between different parameters of the rocket engine. As the early anomaly data of the rocket engine will cause damage to the normal correlation model, the multi strategy anomaly detection and evaluation mechanism is introduced to effectively shield the influence of parameter measurement faults on fault detection of the system, so that system anomaly can be detected more accurately. We use the historical data of engine tests as the training samples to train the characteristic models, and experimental verification of anomaly detection of engine test data in real time is carried out by using the one-element model, the multi-element model and the model based on hybrid probability density distribution. The result shows that compared with the traditional thresh and norm based anomaly detection methods, the multi-strategy anomaly detection method can give not only the normal ad anomaly states of the system, but also the anomaly scores of each parameter and the whole system, thus providing a more flexible reference for further fault diagnosis of launch vehicles.

Key words: fault diagnosis, anomaly detection, data mining, probability density distribution, intelligence test