基于多源数据密度估计自编码器的液体火箭发动机无监督异常检测

Liquid rocket engine hot fire tests represent a pivotal phase in the engine manufacturing and delivery process. Conducting high-accuracy anomaly detection during hot fire tests is a crucial method to ensure the secure and dependable operation of the engine, as well as the successful execution of manned lunar missions. To address the technical challenges associated with the high false positive and false negative rates in anomaly detection caused by the interference of multiple component operational responses and the nonlinear system with differential degradation trends among engine components, an unsupervised learning method of multi-source data density estimation autoencoder is proposed based on liquid rocket engine health operation status monitoring data. Fine-grained feature representations are learned using micro patch feature embedding with a global structured information modeling encoder, and then reconstructed by a decoder to learn health state data distributions. Meanwhile, utilizing gaussian mixture density estimation network constraints enables the autoencoder to achieve improved feature representation and information reconstruction. Finally, the anomaly detection evaluation score based on the dual scales of data feature space representation and data distribution is constructed, the normal situation constraint boundaries under the healthy state operation of the liquid rocket engine are defined, and the intelligent detection of anomalies at the engine system level under the multi-source monitoring data is finally realized. The results of the multi-source data anomaly detection task for the ground thermal test of liquid rocket motors under the same type and across types prove the effectiveness and superiority of the proposed method.