Fault detection for reusable liquid rocket engines based on stacked wavelet autoencoder

Research on fault detection methods is crucial because the safe operation of reusable liquid rocket engines directly impacts rocket flight safety and reusability efficiency. This paper proposes an unsupervised fault detection method that combines a stacked wavelet autoencoder (SWAE) with an isolation forest (IF). The method is trained solely on normal operational data and leverages SWAE to integrate the time-frequency analysis capability of wavelet transform with the feature learning advantages of autoencoders (AE), enabling hierarchical extraction of noise-robust temporal features. Additionally, IF is introduced to exploit its ability to rapidly isolate anomalous samples, thereby achieving effective unsupervised fault detection. The suggested approach performs better in cross-operating-condition generalization, multi-type fault identification, and early gradual failure detection in three common engineering case studies. Compared with traditional approaches such as adaptive thresholding, the proposed method achieves the highest accuracy, recall rate, and F₁ score.