ODDformer: odd–even de-stationary and decomposition techniques transformer for aircraft engine remaining useful life prediction

Abstract In the aerospace industry, accurately predicting the remaining useful life (RUL) of aircraft engines is critical to reduce maintenance costs and increase safety. Existing RUL prediction algorithms fail to account for global temporal factors, overlook the non-stationary nature of monitored data, and neglect critical trends and seasonal characteristics. These factors directly affect the sensitivity of the forecast model to changes in the system state. In light of this, this study introduces an innovative end-to-end deep learning model, called odd–even de-stationary and decomposition transformer (ODDformer), specifically designed for accurate RUL prediction. By incorporating global time embedding, our model demonstrates improved temporal awareness. We propose an innovative odd–even sequence normalization technique, enhancing data stability. Our method incorporates advanced odd–even de-stationary attention to capture crucial dynamic features, deepening model understanding of data evolution. Simultaneously, our channel-independent series decomposition modules extract reliable trend and seasonal features for each sensor. Finally, the two feature sets are fused to obtain the final prediction results. Experimental results on the N-CMPASS dataset demonstrate a 50.89% reduction in RMSE for ODDformer compared to the baseline and a 59.08% reduction for Score. Ablation experiments have validated the efficacy of these components. Our findings offer promising potential for improving tasks like fault diagnosis and anomaly detection in prognostics and health management.