DST-SGR: A Mamba-Inspired Dual-Stream Spectro-Temporal Network for Remaining Useful Life Prediction of Rotating Machinery

Abstract

The prediction of Remaining Useful Life (RUL) for rolling bearings is essential in predictive maintenance to avoid unexpected failures and losses. High-frequency vibration signals pose certain problems since impulses can be found easily in time-domain, while harmonics are more noticeable by time–frequency analysis. Traditional approaches have difficulty dealing with the modeling problem due to LSTM network issues, such as vanishing gradient, and high computational cost of transformer models of quadratic complexity (O(L²)). This paper introduces the Dual-Stream Spectro-Temporal network DST-SGR, a Mamba-inspired architecture in which selective temporal modelling is realized through optimized gated recurrent (GRU-based) state interactions with linear O(L) computational cost, providing the input-dependent selectivity of selective state space models in a fully cross-platform PyTorch implementation. While the temporal branch detects impulses from raw time-series data, the spectral branch learns harmonic signatures from an STFT spectrogram. Three Spectro-Temporal Interaction Blocks (STIBs) couple the streams through bidirectional cross-gated fusion. On the NASA IMS bearing dataset under a leakage-free, file-level evaluation protocol with piecewise RUL labels, DST-SGR attains a test-set MAE of 0.152 and RMSE of 0.191, with 5-fold file-level cross-validation MAE of 0.128 ± 0.004 and generalizes to IMS Tests 2 and 3 under changed fault locations. The model contains 131,137 trainable parameters and trains in approximately 0.97 hours on an NVIDIA RTX 4060 GPU.