Assessment of Autoencoders for signal reconstruction in high power negative ion based plasma sources

To support the development of reliable heating systems for ITER, neutral beam injectors are tested at dedicated facilities such as the RF-driven negative ion source, ROBIN (RF-Operated Beam source in India for Negative ions). ROBIN uses an Inductively Coupled Plasma (ICP) source and is equipped with diagnostics including extraction current monitors, acceleration voltage probes, and beam current sensors. However, the presence of high-power RF radiation and high-voltage transients introduces significant noise, leading to degradation in signal quality — especially in current measurements critical for control and analysis. In this work, we address the reconstruction of noisy or distorted current signals in ROBIN using machine learning-based autoencoder (AE) models. These models are well-suited for this task due to their ability to learn non-linear relationships between input and target signals and to extract latent representations that are robust to noise. Training of these models to reconstruct the beam or extraction current is performed using correlated input signals such as RF forward power, plasma grid voltage, and filter current waveforms. We explore different AE architectures — including standard feedforward AEs, LSTM-based AEs, and Variational Autoencoders (VAEs)—and evaluate their performance in capturing the time-varying structure of current signals. The models are assessed based on reconstruction accuracy and their ability to recover temporal features relevant to pulse dynamics. The reconstructed output can support improved monitoring and control of the ion source under high-noise operating conditions thereby supporting data driven sensor development.