An improved training framework in neural network model fordisruption prediction and its application on EXL-50

Abstract

Abstract A neural network model with classical annotation method has been used on EXL-50 
tokamak to predict the impending disruptions. However, the results revealed issues of
overfitting and overconfidence in predictions caused by the inaccurate labeling. To mitigate
these issues, an improved training framework has been proposed. In this approach, soft labels
from previous training serve as teachers to supervise the further learning process, which has 
demonstrated its significant improvement in predictive model performance. Notably, this
enhancement is primarily attributed to the coupling effect of the soft labels and correction 
mechanism. This improved training framework introduces an instance-specific label smoothing 
method, which reflects a more nuanced model’s assessment on the likelihood of a disruption. 
It presents a possible solution to effectively address the challenges associated with accurate 
labeling across different machines