Fault prediction of gyrotron system on test bench using a deep learning algorithm

The gyrotron system is an essential component of the Electron Cyclotron Resonance Heating (ECRH) system, generating high-power millimeter waves that can be utilized for plasma heating and current drive in magnetically confined nuclear fusion. During high-power long-pulse operation, the gyrotron is prone to faults such as arcing, mode jump, and overcurrent, leading to interruptions in continuous system output. These faults not only endanger the safety of the gyrotron but also reduce the overall efficiency and performance of the system. Addressing this issue, we have employed deep learning methods to predict faults that may occur during the gyrotron's long pulse operation. Using experimental data collected from the megawatt-level continuous wave (CW) gyrotron test bench since 2019, we constructed a dataset comprising 52,772 data points, including 23,241 instances of faulty data. We trained a model based on the Transformer encoder architecture and compared its performance against classical CNN and LSTM models. The results demonstrated that the Transformer model exhibited superior performance in this context, achieving an AUC value of 0.869. The True Positive Rate (TPR) and False Positive Rate (FPR) reached 85.2% and 18.2%, respectively. Subsequent offline testing verified that the trained model could successfully predict faults in advance with second-level accuracy, laying the groundwork for future gyrotron operation management. In conclusion, deep learning has demonstrated its potential application in the prediction of operational faults in the gyrotron system.