Rolling-bearing remaining useful life prediction under high-intensity white Gaussian noise conditions

Abstract To improve RUL prediction fidelity for rolling bearings operating under severe acoustic contamination, this paper proposes a multi-scale quadratic convolutional network with broadcast multi-head self-attention-enhanced LSTM. The proposed approach synthesizes a multi-scale feature ensemble by fusing temporal, spectral, and higher-order statistical descriptors, enabling comprehensive characterization of the bearing degradation process from multiple perspectives. To capture intricate non-linear signatures across multiple degradation regimes, we devise a multi-scale quadratic convolutional backbone that progressively refines feature granularity. Concurrently, the LSTM gates are re-parameterized through a broadcast, multi-head self-attention module that dynamically re-weights hidden states, enabling robust modeling of extended temporal correlations. Comprehensive validation on three run-to-failure bearing datasets confirms that the proposed approach delivers superior RUL forecasting accuracy and resilience across varying operating regimes. Moreover, under comparative experiments with added noise at −3 dB, −6 dB, and −9 dB, the model maintains low prediction error and good trend-fitting ability, indicating strong noise resistance and promising application potential. In addition, ablation experiments further verify the effectiveness of the proposed method.