Aircraft Neural Network Training With Probability-Filtered Failure Modes

The patent extracts only the most relevant operational modes from the complete data space by filtering based on probability thresholds. The filtering process removes operational modes with extremely low probabilities (below theoretical threshold) and separates high-probability modes (above real-world threshold) into a different dataset, keeping only the intermediate probability modes for training. This extraction approach maintains the neural network's ability to handle diverse operational modes while significantly reducing the computational burden of training.

The patent applies local quality by creating different training datasets with different probability characteristics. The training operational mode data space specifically targets intermediate probability modes that represent edge cases and unusual scenarios. Meanwhile, operational modes above the real-world threshold are handled separately, ensuring that both common and unlikely scenarios are covered. This localized approach to data selection ensures that the neural network is specifically trained on the types of operational modes that are computationally feasible to include while maintaining reliability.

The patent applies preliminary action by pre-filtering and stratifying the operational mode data space before training. Instead of randomly selecting modes after the fact, the system first categorizes all operational modes by their probability characteristics, identifying which modes fall into the intermediate probability range. This preliminary organization ensures that the training dataset has optimal coverage of operational modes before training begins, eliminating the need for random selection and post-training validation adjustments.