Research on Metal Surface Defect Evaluation Based on DualBranch-SSMNet and Supervised Contrastive Learning

Abstract

Surface defect assessment of metallic workpieces is critical for industrial production and structural safety. Although eddy current testing provides non-contact and nondestructive inspection advantages, existing deep learning methods still suffer from insufficient long-range dependency modeling, inadequate utilization of frequency-domain information, and limited discrimination of fine-grained defects in complex eddy current signals. To address these limitations, this study proposes DualBranch-SSMNet, a Dual-Branch framework integrating a selective state space model and supervised contrastive learning. In the 1D branch, a selective state space model is introduced to efficiently model long-sequence eddy current signals with linear computational complexity, thereby enhancing global dependency capture. In addition, a lightweight joint time–frequency representation branch is constructed based on the outer product of frequency-domain amplitudes and autocorrelation sequences to improve the characterization of periodic and spectral features. Furthermore, supervised contrastive learning is employed to optimize the feature space and enhance the discriminability of similar defect categories. Experimental results on the MDDECT dataset demonstrate that the proposed method achieves an accuracy of 96.98% and an F1-score of 96.99% in the four-class classification task, indicating its superior performance and strong generalization capability for complex eddy current signal analysis.