Feature Information Fusion-Enhanced Representation Network for Rotating Machinery Remaining Useful Life Prediction

The degradation process of rotating machinery during its service life is highly nonlinear and exhibits significant variability due to multi-physics coupling effects, leading to substantial challenges in remaining useful life (RUL) prediction accuracy. To address this issue, we propose a novel RUL prediction framework integrating feature information fusion enhanced representation with a temporal convolutional attention feature calibration network (TCAFCN). The methodology consists of three key innovations: First, vibration signals are decomposed through improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) to generate multi-resolution component signals. Subsequently, an improved refined composite hierarchical fuzzy dispersion entropy (IRCHFDE) is developed to hierarchically decompose these components, constructing a high-dimensional multi-resolution entropy feature space (HMEFS). Second, a linear local tangent space alignment (LLTSA) algorithm is implemented to generate low-dimensional fusion features by eliminating interference information while preserving critical degradation characteristics. Third, the proposed TCAFCN-BiLSTM architecture synergistically combines temporal convolutional layers with attention mechanisms to capture local feature correlations, while bidirectional long short-term memory (BiLSTM) modules extract global temporal dependencies of calibrated features. Extensive experimental validation on two industrial datasets demonstrates the proposed method’s superior performance. The results significantly advance the precision of prognostic approaches for complex rotating machinery systems.