Dynamic sparse vector autoregressive moving average model for online fault detection in liquid rocket engines

Online fault detection in liquid rocket engines (LREs) is the essential stage in the advancement towards autonomous engine health management. Nevertheless, the most recent recursive vector autoregressive moving average (VARMA) model-based online fault detection method in LREs is constrained by the issue of numerical computational instability in its recursive process. To address this challenge, a novel dynamic sparse VARMA (dsVARMA) model is proposed for online fault detection of LREs. This is achieved by unifying adaptive dictionary learning with the recursive iterations of VARMA to enhance the numerical computational stability. Next, a hardware-in-the-loop (HIL) test bench was constructed and the proposed method was implemented in the embedded hardware for validation. Subsequently, the proposed method was tested on a simulation dataset with injected faults and a hot-fire testing dataset by means of HIL tests. The experimental results demonstrate that the proposed method outperforms other widely used online FD methods of LREs in terms of both detection accuracy and detection time. The proposed method is effective in detecting failure modes characterized by abrupt changes in slowly varying sensor measurements, as exemplified by leakage faults, valve stuck faults, and turbopump rotor stuck faults in LREs. Concurrently, experimental findings further demonstrated that the proposed method exhibits considerable potential for real-world engineering applications.