Abstract
The recognition of defects is a critical component in monitoring the structural integrity of metal components. Accurate discrimination between surface and subsurface defects is crucial for assessing the structural safety of metal components, yet remains challenging due to interference between defect depth and classification. This paper proposes a novel defect recognition framework integrating Improved Tensor Locality Preserving Projection (ITLPP) with multi-feature fusion. To address the limitations of single-feature approaches, our method systematically extracts multi-domain features to construct an analytic domain × feature variable matrix. The ITLPP algorithm then performs dimension reduction by preserving the spatial correlations in feature matrix and sample class information. Subsequently, a fusion matrix mechanism enhances category-sensitive features, and finally achieves effective defect classification through improved k-nearest neighbors (KNN). Experimental results demonstrate our proposed method outperforms conventional single-eigenvalue methods and other methods in terms of detection performance, achieving a recognition accuracy for surface/subsurface defects of 98.1%. Notably, the method can also be used to identify defects with different cross-sectional shapes, which shows that the method is suitable for natural defects.