Spiking Parametric Multi-View Fusion Mixer for energy-efficient aero-engine bearing fault diagnosis

Accurate and timely fault diagnosis of rolling bearings in aero-engines is critical for ensuring flight safety, yet deploying deep learning models on resource-constrained edge devices remains a major challenge. This paper proposes the Spiking Parametric Multi-View Fusion Mixer (SPMF-Mixer), a fully spike-driven framework that achieves high diagnostic accuracy with a favorable accuracy–energy trade-off. The core of the architecture is the Multi-Branch Transient Dynamics Aggregator (MTDA), an attention-free module that replaces conventional self-attention with three parallel branches: a Contextual Branch capturing local morphological features via depthwise convolutions, an Interactive Branch facilitating cross-channel information fusion through point-wise convolutions, and a Transient Branch employing a Spiking Haar Wavelet Transform to isolate high-frequency fault impulses. To overcome the limited dynamic response of standard spiking neurons, we further propose the Adaptive Parametric Leaky Integrate-and-Fire (APLIF) neuron, which incorporates bio-inspired inhibitory gating, learnable membrane decay, and spike-dependent reset modulation to enhance sensitivity to transient fault signatures while maintaining efficient spiking activity. Extensive experiments on the DIRG high-speed bearing dataset and the HIT inter-shaft bearing dataset demonstrate that SPMF-Mixer achieves the best performance among all compared methods and maintains a clear advantage under severe noise. Pareto frontier analysis further shows that it achieves a superior balance between energy consumption and diagnostic performance. These results suggest that SPMF-Mixer is a promising solution for edge-side aero-engine bearing monitoring in noisy and resource-constrained operating environments.