.png){kind=icon}
.png){kind=icon}
Return
Retentive multimodal scale-variable anomaly detection framework with limited data groups for liquid rocket engine
.png){kind=icon}
PRE
AI
DOI：10.1016/j.measurement.2022.112171.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Anomaly detection (AD) plays a significant role in the safe and stable operation of liquid rocket engine (LRE). Facing the fact that LRE operating data are in multiple modalities and extremely scarce, existing methods, which ignore the modal properties, cannot detect anomalies effectively. With a limited number of data groups, they hinder the AD during LRE research and development (R & D) especially in the early stage. A retentive multimodal scale-variable framework is proposed for AD with limited LRE operating data groups. Then the framework utilizes multimodal scale-variable structure to extract and fuse the distinct features of different modalities, which pays more attention to modal properties. To suppress the overfitting and get stronger generalization ability, the framework constructs a memory matrix to fuse it with test data at feature level, which has a retentive memory of samples encountered. To verify the effectiveness of this framework, three experimental cases are designed for research and discussion according to different stages of R & D process. Results show that this framework could recognize the state of LRE with F1-scores over 0.977 in three experimental cases, verifying it has reliable performance and strong potential of AD in the whole R & D cycle of LRE.
Keywords:
Anomaly detection
Liquid rocket engine
Multimodal learning
Scale-variable structure
Limited data