A multimodal deep learning assisted uncertainty-aware MPC framework for fault tolerant control of aviation piston pumps

Abstract

Aviation piston pumps are critical components of aircraft hydraulic systems. The progressive wear can significantly degrade system reliability, control accuracy and operational safety of piston pumps. Conventional fault-tolerant control (FTC) methods generally rely on fixed system models and discrete fault assumptions, limiting their adaptability under continuously evolving degradation and uncertain operating conditions. To address these limitations, this paper proposes a novel uncertainty-aware multimodal deep learning-assisted model predictive control (UA-DL-MPC) framework for intelligent fault-tolerant control of aviation piston pumps operating under multi-mode wear conditions. The proposed framework integrates multimodal deep learning-based fault classification, degradation severity estimation, and uncertainty quantification with an adaptive model predictive controller. The predicted degradation severity continuously adapts the degradation-aware control model, while the predictive uncertainty adaptively scales the auxiliary fault-compensation signal to regulate controller aggressiveness according to diagnostic confidence. A nonlinear auxiliary compensation mechanism is further incorporated to improve disturbance rejection under severe wear conditions. Experimental evaluation was conducted using a real aviation piston pump wear dataset containing Healthy, Valve Wear, Servo Wear, and Swashplate Wear conditions. The proposed framework achieved 99.49% fault classification accuracy on the independent test set and 99.32 ± 0.99% average accuracy during five-fold cross-validation. The severity estimation module achieved MAE = 0.0111, RMSE = 0.0124, R² = 0.9805, and Pearson correlation = 0.9992. Furthermore, the proposed controller achieved the lowest tracking error (RMSE = 0.558), representing approximately 63% and 27.7% reductions compared with conventional MPC and DL-MPC, respectively, while maintaining practical control effort and robust constraint satisfaction. These results demonstrate the effectiveness of integrating uncertainty-aware deep learning with predictive control for intelligent fault-tolerant operation of safety-critical aviation hydraulic systems.