Turbofan engine remaining useful life prediction with liquid neural networks: A comparative analysis

Turbofan engines are a primary power source in civil aviation, and their unexpected failures can have severe consequences in terms of both safety and cost. Within the scope of predictive maintenance strategies, Remaining Useful Life (RUL) prediction aims to optimize maintenance scheduling by forecasting the time until component failure. In the literature, discrete-time deep learning architectures such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN), and Transformer have been widely employed for this problem; however, despite engine degradation being inherently a continuous-time physical process, the potential of continuous-time neural network architectures for modeling this process remains largely unexplored. In this thesis, RUL prediction was performed on the NASA C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) dataset using a Closed-form Continuous-time (CfC) based Liquid Neural Network (LNN) model. The input-dependent time constants of the CfC cell offer the potential to adapt to the varying dynamics of the degradation process. The model hyperparameters were systematically optimized through a two-stage Bayesian optimization strategy based on the Tree-structured Parzen Estimator (TPE) algorithm implemented via the Optuna framework. In the first stage, architectural and training parameters such as hidden state size, learning rate, and mini-batch size were optimized; in the second stage, data pipeline parameters including sequence length, RUL ceiling, and correlation-based sensor selection threshold were sequentially tuned. Experimental results demonstrated that the CfC-based LNN model achieved strong performance on single operating condition sub-datasets. On FD001, the model attained 8.40 RMSE and a NASA score of 107, while on FD003 it reached 9.71 RMSE. These results correspond to lower RMSE values than those reported for the compared LSTM, DCNN, BiLSTM+Attention, CNN-LSTM-Attention, and STAR Transformer models. On multi-operating condition sub-datasets, the model achieved 13.50 RMSE on FD002 and 20.97 RMSE on FD004; although these RMSE values are higher than those reported by some classical approaches, they remain competitive with state-of-the-art Transformer-based architectures. It should be noted that these comparisons are based on values reported by different studies whose preprocessing and experimental protocols are not strictly identical. The two-stage Bayesian optimization produced dataset-specific configurations, enabling performance levels unattainable with a single uniform setting.