Structure-guided hierarchical decision framework for edge-side fault diagnosis of rolling bearing surface defects

Reliable fault-state decision-making for rolling bearings is constrained by limited samples, weak surface evidence and strict edge-side latency requirements. This paper proposes a structure-guided hierarchical decision framework for bearing surface fault diagnosis. An annular region of interest (ROI) is isolated using rotational symmetry, radial scanning and deterministic circle fitting. Dual-driven reconstruction then fuses coarse regions with multi-directional gradient evidence to recover low-contrast boundaries. A ten-dimensional geometric-textural representation is classified by a prototype-grouping adaptive binary tree support vector machine (PABT-SVM), using linear routing and radial basis function (RBF) terminal nodes. Unlike flat or predefined tree-based SVMs, PABT-SVM derives an asymmetric hierarchy from fault-state prototypes and optimises individual node parameters. The reconstruction stage achieves a 96% detection rate, while the final diagnosis reaches a 93.61% macro-averaged F1-score (Macro-F1) with an average representation-and-decision time of 36.28 ms per sample. These results demonstrate a favourable balance between diagnostic performance, interpretability and computational efficiency.