Traditional acoustic emission (AE) detection methods and learning frameworks often struggle to achieve reliable diagnostic performance under small-sample conditions, especially when signals have weak energy, high noise interference, and severe attenuation. This limitation stems from their reliance on large labeled datasets and limited noise robustness. To address these challenges, a novel kernel-enhanced Mel-dual-attention (KMDA) model is proposed for AE fault diagnosis under low signal-to-noise ratio (SNR) and small-sample conditions. A Mel-guided sparse kernel learning mechanism is employed, together with a three-stage denoising strategy, to enable unsupervised pretraining of convolutional kernels that capture the statistical structure of AE signals encoded in enhanced Mel spectrograms. This method mitigates the drawbacks of random kernel initialization and provides informative priors that facilitate subsequent training with limited data. Furthermore, a cross-channel consensus–divergence fusion module is introduced to suppress redundant information across sensors while reinforcing weak signal components. To validate the proposed framework, a metal stress-corrosion cracking experimental platform was developed to collect a low-SNR, small-sample dual-sensor AE dataset. The experimental results demonstrate that the KMDA model significantly outperforms state-of-the-art fault diagnosis methods under low-SNR and small-sample conditions.