DTFFNet: A dual-branch time–frequency feature fusion network for remaining useful life prediction of mechanical equipment

Abstract

With its exceptional feature extraction and modeling capabilities, deep learning has emerged as a cornerstone technology in the field of remaining useful life (RUL) prediction for mechanical equipment. However, time-domain-based deep learning methods face significant limitations in capturing global features, hindering their ability to fully model the complexities of equipment degradation processes. To address these challenges, we propose DTFFNet, a dual-branch time–frequency interactive integration framework that combines time-domain and frequency-domain signal analysis for efficient feature interaction and fusion. Specifically, the time-domain branch employs a CNN-Transformer structure to extract local features and model long-term dependencies within sequences. The frequency-domain branch utilizes a frequency-domain Multilayer Perceptron (MLP) to extract global dependencies from the real and imaginary components of Fourier-transformed signals. Additionally, a multi-head cross-attention mechanism is introduced to construct a time–frequency cross-domain interactive fusion module, enabling information exchange and complementary representation learning across time and frequency domains, enhancing feature fusion depth and effectiveness. To meet the specific demands of RUL prediction, a joint learning strategy is proposed that combines weighted and constrained MSE losses, enabling targeted optimization during critical degradation phases and along specific error directions to enhance prediction accuracy. Finally, experimental analysis on widely used turbofan engine dataset and PHM 2012 bearing dataset demonstrates that DTFFNet not only significantly improves prediction accuracy compared to mainstream RUL prediction methods but also exhibits enhanced robustness in noisy environments. These results establish DTFFNet as a reliable and effective solution for equipment health management in complex industrial settings.