Abstract: In order to monitor the quality of TIG welding for blocked tube welding of high-temperature gas-cooled reactor steam generators under the constraints of operation space and radiation environment, a real-time monitoring system based on a fiber optic spectrometer for TIG welding process was developed for monitoring the depth of penetration during welding. This study used the system to collect arc spectra and utilized Principal Component Analysis to obtain the spectral principal components of different weld penetration depths. An innovative ATT-L2R-BiLSTM deep learning model was proposed to achieve classification and recognition of weld penetration depth during blocked tube TIG welding. The results show that the model achieved an accuracy of 92.61% under laboratory conditions, which is 5.11% higher than that of the Bi-LSTM network. The model was tested and verified on a blocked tube verification platform for nuclear power steam generators, achieving an accuracy of 99.26%. Finally, deep mining of welding quality features and precise evaluation of weld penetration depth during TIG welding were achieved under incomplete spectral information.