Robust learning-based fault prediction method for gyrotron system

During the high-power long-pulse operation of the gyrotron, radio frequency (RF) oscillation faults may occur, which can affect the stable operation of the electron cyclotron resonance heating system. Recently, some research has begun to use deep learning to diagnose RF oscillation faults in the gyrotron. However, these learning-based methods have overlooked the useless or erroneous data, which we refer to as noisy data, in the gyrotron experiment data. Noisy data are inevitably generated during the data acquisition process, which would lead to the performance degradation of the learning-based methods. To address this, we propose a novel data processing framework based on the Gaussian mixture model (GMM) and multi-layer perceptron (MLP), which allows the model to learn more robust data representations and can identify noisy data. In particular, we inject the representations learned by the MLP model into the GMM for modeling. We then use the GMM to calculate the posterior probability of these representations and filter out the noisy data based on this. Extensive experiments show that this method achieves good performance in the presence of noisy data. Moreover, as the proportion of noisy data increases, our method maintains stable results in gyrotron fault diagnosis, demonstrating the robustness of our approach.