Dynamic inductive bias in liquid neural networks for robust few-shot bearing fault diagnosis

Abstract Achieving reliable bearing fault diagnosis under small-sample conditions remains a persistent challenge in intelligent maintenance, since most existing methods depend heavily on large-scale labeled data and show limited robustness under noisy and complex conditions. To overcome this limitation, we propose a dynamic inductive bias (DIB) liquid neural network, which explicitly incorporates inductive biases aligned with the dynamics of vibration signals into deep architectures. The central concept lies in introducing a dynamic time-constant mechanism that adaptively regulates neuronal sensitivity to multi-scale temporal patterns, effectively constraining the hypothesis space and mitigating overfitting under data-scarce scenarios. Moreover, the framework integrates convolutional neural network-based local feature extraction with attention-driven global information aggregation, thereby enabling hierarchical and complementary representations of vibration signals. Extensive experiments conducted on both a public benchmark and a custom dataset demonstrate that the proposed framework consistently surpasses state-of-the-art deep learning methods under small-sample and high-noise conditions. These results underscore the effectiveness, robustness, and practical industrial potential of embedding DIBs into neural architectures for bearing fault diagnosis.