Defect Detection for SRF Cavities Using Hybrid Deep Learning Model

Superconducting Radio Frequency (SRF) cavities are crucial component of linear accelerator. Surface defects in niobium cavities lead to field emissions which can disturb required cryogenic temperature and negatively impact performance of linear accelerator. It is of vital importance to detect defects in cavities during inspection before installation in linear accelerator. Current manual process of inspection using borescope is very time consuming. This work proposes transformer encoder and convolutional layers based model for automatic surface defect classification which can significantly reduce time and effort required for inspection. In order to automate cavity defect inspection, models with lesser number of parameters are optimal to allow deployment on local economical hardware and keep running cost low. Our model utilizes convolutional filters of different sizes to cover range of surface defects in cavities. Point-wise convolution is used for channel attention and transformer encoder for spatial attention. Additionally, the model adopts weighted cross-entropy loss to minimize overfitting and improve results. For SRF cavity defect detection, the proposed model achieves a similar or better recall and f1 score with 33% fewer parameters than the standard image classification model.