Adaptive spectral emphasis and geometric metric learning for acoustic fault diagnosis

Acoustic fault diagnosis of rotating machinery under variable-speed conditions remains challenging because acoustic signals are highly sensitive to speed variations, propagation effects, and environmental interference, which often lead to unstable spectral representations and non-Euclidean feature distributions. Conventional prototypical learning methods, which typically rely on uniformly processed features and fixed Euclidean metrics, therefore exhibit limited robustness and poor cross-condition generalization. To address these issues, this paper proposes an artificial-intelligence-based acoustic fault diagnosis framework that implements geometry-aware few-shot prototypical learning. First, an adaptive low-frequency emphasis mechanism is introduced in the log-Mel domain to identify structurally dominant low-frequency regions through cumulative spectral analysis and enhance their relative salience. Second, prototype matching is extended to Euclidean, spherical geodesic, and manifold geodesic metrics, and a dataset-geometry analysis module automatically selects the most appropriate metric according to spherical conformity and manifold curvature. The framework is evaluated on the University of Ottawa electric motor dataset and the Machinery Fault Database under varying speed and load conditions. In 5-way 5-shot cross-condition tasks, it improves average accuracy from 74.97% to 87.06% and from 77.23% to 87.17%, respectively, while outperforming representative metric-learning methods.