Study on appearance anomaly detection methods for large-scale nuclear fuel assemblies

Anomaly detection of large-scale nuclear fuel assemblies is crucial for nuclear safety. Traditional manual inspection has problems such as low efficiency, strong subjectivity, and high missed detection rates. However, existing machine vision methods often misjudge normal textures as anomalies due to discontinuous component distribution and significant texture differences between grids and end plugs. In addition, these methods also have difficulties in capturing tiny surface defects. This paper proposes an anomaly detection method for nuclear fuel assemblies based on the UNet semantic segmentation network and Patchcore network. The framework performs hierarchical extraction and quantification of local patch features of normal and abnormal samples, and constructs a dynamic benchmark library that integrates the diversity of normal patterns and prototypes of abnormal features for multi-dimensional anomaly scoring. Feature segmentation enables semantic-level localization of different components, thereby adopting customized detection strategies for the textures of structures such as grids and end plugs to accurately identify abnormal regions. In simulated fuel assembly detection scenarios, compared with traditional methods, this framework significantly reduces the false positive rate of grid and end plug components, while greatly improving the detection rate of tiny defects. It effectively solves the problem of semantic ambiguity in block detection of complex structures and provides a highly robust solution for industrial automated inspection of nuclear fuel assemblies. This method significantly improves detection efficiency and reliability, and has important engineering application value for the safe operation and maintenance of nuclear power equipment.