A cascaded method of SDP visual representation and ProtoVBL-Net for few-shot cross-domain diagnosis

Intelligent fault diagnosis of electric machines is often challenged by data scarcity and poor generalisation across varying operating conditions in few-shot, cross-domain scenarios. To address this, we propose an innovative two-stage cascaded diagnosis framework named Prototype Vision – Broad Learning Network (ProtoVBL-Net). In the first stage, we employ a novel multi-axis Symmetrized Dot Pattern (SDP) fusion technique to efficiently transform raw multi-dimensional vibration signals into RGB ‘fault fingerprint’ images. This process effectively decouples and enhances cross-domain features, providing a robust visual representation for subsequent diagnosis. In the second stage, departing from conventional nearest-prototype classification, our key contribution is the use of a prototype distance vector as a high-level meta-feature. This vector is fed into a non-iterative Broad Learning System (BLS) for adaptive boundary calibration and final classification. This mechanism synergistically combines the sample efficiency of metric learning with the powerful non-linear classification capabilities of BLS. Extensive experiments across multiple operating conditions and eight fault categories demonstrate that ProtoVBL-Net achieves superior generalisation, with a peak diagnosis accuracy of 95.12%, significantly outperforming state-of-the-art baselines. The results validate its methodological novelty and strong potential for data-constrained industrial applications.