MLBA-Net: A Lightweight CNN–Mamba–LSTM Network for Few-Shot Cross-Domain Bearing Fault Diagnosis
Year: 2026
Type: preprint
Abstract: Deep learning has substantially advanced vibration-based fault diagnosis of rolling bearings, yet two obstacles continue to impede industrial deployment: the performance degradation that models suffer when transferred across test rigs (domain shift), and the scarcity of labeled fault data for newly commissioned machinery. This paper first establishes a unified, leakage-aware, multi-seed benchmark that spans three datasets (CWRU, Paderborn (PU) and SEU), all six cross-dataset transfer directions and eight modern baseline architectures, and subsequently proposes MLBA-Net, a lightweight hybrid network tailored to the extreme few-shot cross-domain scenario. MLBA-Net transforms vibration segments into three-channel continuous wavelet transform (CWT) images, extracts local time–frequency patterns through a convolutional stem, reorganizes the feature maps into frequency-band tokens that preserve the ordered harmonic structure of bearing fault signatures, and models them with a gated parallel bidirectional Mamba and bidirectional LSTM pair, followed by band attention and a multi-query pooling head. With merely 0.549 M parameters, MLBA-Net attains 98.26±0.43% accuracy on CWRU over ten seeds. In the central 1-shot CWRU→PU experiment it reaches a group Macro-F1 of 0.321±0.068—significantly ahead of both pure state-space models and the CNN baseline PFFNet, statistically tied with TimesNet, while trailing the two Transformer baselines—and it achieves the best zero-shot transfer in both PU-source reverse directions. The pure Mamba architecture fine-tunes worst of all models at every shot count, whereas the hybrid escapes this failure mode, suggesting that complementary inductive biases may improve adaptation stability under extremely limited target-domain supervision. Controlled studies further show that noise-augmented training, rather than an architectural denoising module, drives noise robustness, and that replacing the Mamba branch with a BiGRU yields statistically equivalent few-shot performance—an honest boundary on the architecture’s claims.
Source: Preprints.org
Authors: Chenhui Zhao, Wai Yie Leong
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214682302