To address the challenge of data distribution shift induced by varying operating conditions in cross-condition axle box bearing fault diagnosis, where conventional methods struggle to learn fault representations with sufficient generalizability, this paper proposes a Real Temporal Slice Spiking Mamba (RTS‑Spiking Mamba) model. The proposed model first employs a real temporal slice spiking encoding scheme to partition the raw vibration signal into consecutive time slices, and performs local feature extraction and spike encoding on each slice individually, yielding spike feature sequences that preserve the genuine temporal evolution of the vibration signal. Then, a Spiking Mamba module is constructed, in which multi-level Leaky Integrate-and-Fire (LIF) neurons are applied to the projected and locally convolved features, while an additional LIF neuron converts the continuous-valued output of the selective state-space scan into spike representations. This design couples spike-based nonlinear transformations with the long-range sequence modeling capability of Mamba, enabling the model to learn fault representations with strong cross‑condition generalization. Eight cross‑condition domain generalization tasks are set up using bogie axle box bearing dataset from Beijing Jiaotong University for validation. Experimental results demonstrate that the proposed method outperforms state-of-the-art cross-condition learning algorithms and an existing Spiking Mamba variant.

Bearing fault diagnosis; Domain generalization; Spiking neural network; Mamba; State space model; Real temporal slice