Open Access

Peng ChenORCiD, Ruijin Zhang, Changbo HeORCiD, Yaqiang Jin, Shuai Fan, Junyu QiORCiD, Chengning ZhouORCiD, Chun Zhang

• Traditional industrial surface defect detection using CCD/CMOS cameras faces limitations in detecting minute defects on aluminum substrates in dynamic industrial scenes. While event cameras can capture small defects through event information flow (EIF), they struggle with noise-related challenges. Additionally, current contrastive learning frameworks employ simplified loss computation mechanisms that limit their capability to discover and characterize novel categories. This study addresses these challenges through two main innovations: (1) a novel event stream imaging technique that combines equal time window segmentation, overlapping windows, and Gaussian filtering to enhance data quality, and (2) Progressive Contrastive Representation Learning (PCRL), a sophisticated framework for handling both known and unknown fault classes. The PCRL framework overcomes traditional limitations through a four-stage loss function strategy and structured analysis of intra-cluster and inter-cluster relationships, enabling better feature extraction and model convergence for previously unseen patterns. Through systematic validation, our approach demonstrates superior performance in noise reduction, unknown fault detection, and precise surface defect classification, offering a robust solution for industrial applications. Through systematic validation, our methodology exhibits exceptional efficacy in mitigating signal noise, identifying novel fault patterns, and executing high-precision surface anomaly categorization, yielding a comprehensive accuracy rate of 93.51% across all test cases and demonstrating remarkable performance with 95.36% accuracy in previously unencountered fault categories, thereby presenting a sophisticated solution for industrial implementations while establishing a foundation for subsequent scholarly investigations in this field.