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

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Pulsed Power Systems

Background:

• Gas-switch prefire is a common failure in linear transformer drivers (LTDs).
• Diagnosing faults in large LTD facilities is challenging due to noisy signals, missing data, and limited instrumentation.
• Existing methods struggle with incomplete and non-ideal sensor measurements.

Purpose of the Study:

• To develop a robust fault diagnosis method for LTDs that can handle incomplete and noisy sensor data.
• To improve the accuracy and reliability of stage-level fault detection in large-scale pulsed-power facilities.
• To propose a scalable and cost-effective solution for real-time fault diagnosis.

Main Methods:

• A mask-augmented residual temporal convolutional network was proposed.
• The network ingests water-transmission-line (WTL) voltages and binary channel-mask indicators.
• Training was performed using physics-based simulations of a 12-stage LTD circuit with added Gaussian noise.

Main Results:

• Achieved 98.5% accuracy with 25% random channel removal during cross-validation.
• Demonstrated 100% correct stage localization on prototype shots with high confidence (>0.91).
• A sparse-sensor variant with only three group WTL voltages retained 89.6% accuracy, showing cost-effectiveness.

Conclusions:

• The proposed mask-augmented network offers a data-driven, resilient pathway for LTD fault diagnosis.
• The approach effectively handles non-ideal measurements and is scalable for future pulsed-power facilities.
• Explicitly encoding missing channels and using temporal convolutions enhance diagnostic robustness.