Improved convex polyhedron support vector machine modelling for fault diagnosis in mechanical systems with multi-source uncertainty

Accurate mathematical modelling is essential for reliable fault diagnosis and safe operation of mechanical systems. In practice, however, diagnostic decisions are often degraded by multi-source uncertainty arising from sensor noise, operating-condition variations, and environmental disturbances, which weakens the credibility of conventional point-based classifiers in engineering applications. To address this issue, this paper proposes an improved convex polyhedron support vector machine for uncertainty-aware fault diagnosis using bounded statistical features. Starting from the classical support vector machine formulation, the proposed framework explicitly represents uncertain samples as convex polyhedrons rather than deterministic points. A hybrid -volume-similarity-distance index is introduced to remove potential outliers during convex polyhedron construction and to obtain refined uncertain features. The geometric information of the resulting convex polyhedrons is then incorporated into a modified cost-sensitive learning scheme, in which the separating hyperplanes are iteratively updated through vertex searching. Experiments on three mechanical fault diagnosis datasets demonstrate that the proposed method outperforms six benchmark methods, achieving average diagnostic accuracies of 97.69%, 97.54%, and 97.43%, respectively. These results show that the proposed model can effectively characterize bounded uncertainty in statistical features and improve the reliability of fault classification under noisy operating conditions.