Stage-level prefire localization in linear transformer drivers via mask-augmented residual temporal convolutional networks

Gas-switch prefire is a frequent failure mode in linear transformer drivers (LTDs), and stage-level fault diagnosis is complicated by noisy signals, missing channels, and sparse instrumentation in large facilities. This work proposes a mask-augmented residual temporal convolutional network that ingests water-transmission-line (WTL) voltages together with binary channel-mask indicators, enabling robust inference under incomplete and non-ideal sensing. The network is trained exclusively on physics-based simulations from a Preisach-based 12-stage LTD circuit model with 1440 labeled waveforms; 40 dB Gaussian noise is added to approximate prototype measurements. Five-fold cross-validation yields 98.5% accuracy with 25% random channel removal. On independent prototype shots weakly labeled using a physically interpretable WTL-polarity rule, the classifier achieves 100% correct stage localizations with Softmax confidence >0.91. Moreover, a sparse-sensor variant using only three group WTL voltages plus masks retains 89.6% simulated accuracy, indicating a cost-effective deployment path for large LTDs. By explicitly encoding missing channels and leveraging long-receptive-field temporal convolutions, the proposed approach provides a data-driven fault diagnosis pathway that is resilient to non-ideal measurements and scalable to sparse probe layouts in next-generation pulsed-power facilities.