A fully multilayer perceptron-based model with controllable parameters for remaining useful life prediction

Abstract End-to-end data-driven prediction methods are essential for effective prognostics and health management, enabling accurate and timely predictions of remaining useful life (RUL) in complex systems. These methods reduce reliance on manual features and expert knowledge, making them crucial for industries like aerospace and manufacturing. Existing deep learning-based RUL prediction methods demonstrate strong predictive performance by integrating spatiotemporal features from time-series data. However, as the depth of these models increases, their complexity and computational demands grow rapidly, and these models have their own limitations. Moreover, for RUL prediction tasks, these models may not always be necessary, especially when simpler architectures can deliver competitive results. Models based on the simple architecture of multilayer perceptrons (MLPs) often perform poorly in RUL prediction tasks due to limitations, such as their inability to capture spatial features. In light of these challenges, we propose the recurrent mixing MLP (RMixMLP), a fully MLP-based block structure designed to capture both channel and temporal information while addressing the limitations of deeper models. The RMixMLP comprises two components: (1) the mixing MLP (MixMLP), which integrates a series-MixMLP to extract temporal features and a channel-MixMLP to combine channel features, and (2) a self-recurrent framework that allows parameter sharing within the MixMLP block, ensuring data integrity and controlling model complexity. Data processed through stacked RMixMLP blocks are fed into a fully connected layer for RUL prediction, eliminating the need for other deep learning methods. The proposed technique was validated on the company-modular aero-propulsion system simulation aircraft turbofan engine dataset, with experimental results demonstrating that RMixMLP improves prediction accuracy by 9.7% compared to recent state-of-the-art models. At the same time, compared with the traditional MLP-based model, RMixMLP not only reduces the number of model parameters, but also improves the prediction performance by 20%.