Direct drive solenoid valves (DDSVs) serve as critical actuators in aviation hydraulic systems, facing challenges in fault diagnosis due to electromagnetic interference and noise coupling, with traditional data-driven methods lacking robustness. In order to solve this issue, this work suggests a reliable defect diagnostic technique based on an attention-aware weighting mechanism: projection gradient descent (PGD) simulates complicated electromagnetic interference by creating perturbation samples that represent actual operating conditions. Deep neural networks (DNN) and an attention perception module (APM) undergo collaborative training, dynamically adjusting sample weights to enhance model robustness. An experimental platform for aviation solenoid valve fault diagnosis validation was established. According to experimental results, this approach outperforms other weighting strategies in metrics like diagnostic accuracy and convergence stability under interference conditions, avoids extreme weight concentration, and fully utilizes both original and perturbed samples to improve interference resistance.