Validation-Selected Constrained Incremental Learning for Simulation-Based UAV Subsystem Fault Prediction

Le résumé fourni par la source

Fixed offline prognostic models can lose accuracy when sensor-data distribution changes after model development. This paper presents a validation-selected incremental-learning framework for time-before-failure (TBF) regression in three simulated unmanned-aerial-vehicle (UAV) subsystems: avionics, power, and airframe structure. According to the dataset provider, the subsystem simulation records used in this paper were generated with a customized quadrotor model on the RflySim framework, and the original source arrays used in the experiments are publicly available through the research repository specified in the Data Availability Statement. A leakage-controlled protocol partitions 200 parameterized degradation realizations per subsystem before window use, fits normalization statistics on source-training realizations only, and creates a controlled target domain through standardized sensor bias, gain, noise, and contiguous dropout. CNN–LSTM, TCN–CNN, and Transformer–CNN candidates are compared over three random seeds; CNN–LSTM is selected for all three subsystems by the source-validation RMSE. The incremental benchmark compares no update, fine tuning, elastic weight consolidation (EWC), replay, learning without forgetting (LwF), and an L2 parameter constraint. Each trainable method receives four predeclared candidates, the same three seeds and five-epoch budget, and selection by the same old-plus-new validation score before held-out testing. The resulting stability–plasticity balance is subsystem and storage-condition dependent: replay gives the strongest balance for power and airframe when source storage is permitted, whereas tuned fine tuning gives the lowest forgetting for avionics among the replay-free updates. L2 remains a compact replay-free update but is not universally superior to the other tuned methods. Near-failure errors and threshold warning rates provide an additional maintenance-oriented acceptance check. The evidence establishes controlled simulation feasibility; real-flight, bench-test, and hardware-in-the-loop validation remain outside the present scope.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.