Accurate recognition of helicopter flight conditions is vital for operational safety and envelope protection, yet real-time classification of multivariate time series remains challenging. This paper introduces Dynamic Multi-Stream Network with Confidence Gating (DMSN-CG), a two-stage neural framework for online classification of helicopter flight time series. The first stage (DMSN) employs a custom encoder–decoder design with multiple parallel prediction streams, aggregated by an adaptive soft mixture-of-experts weighting through a lightweight linear fusion that acts as an implicit regularizer, yielding ensemble-like, stable predictions. The second stage (CG) applies a learned confidence gate to suppress low-confidence outputs and refine predictions via a hierarchical dilated-convolution stack fused with input-conditioned attention; a custom loss further penalizes erroneous class transitions within prediction windows. Evaluated on 490 flights with 11,078 labeled maneuvers across 49 mixed regimes, DMSN-CG achieves 97.46% accuracy (precision 0.98, recall 0.97, F1 0.97), outperforming prior neural and non-neural approaches not relying on specific feature engineering. Our model reduces memory by 98% to 45.7 MB (50.4 × representation efficiency) and achieves 17.5 ms GPU and 60.1 ms CPU inference latency, well below our 80 ms acquisition cycle, enabling online deployment. Test-time compute scaling via accumulation and majority voting allows adjustable accuracy–computation trade-off. These results demonstrate DMSN-CG is a flexible, highly accurate, memory-efficient, and responsive solution for envelope protection, automatic recovery, and predictive maintenance in safety-critical aerospace systems. • Introduces DMSN-CG, a two-stage neural architecture with multi-stream decoding and learned confidence gating for real-time flight condition recognition. • Achieves state-of-the-art accuracy of 97.46% on 490 helicopter flights across 49 mixed flight regimes without feature engineering. • Reduces model size by 98% to 45.7 MB and meets online constraints with sub-80 ms CPU inference latency. • Enables deployment flexibility through causal inference and test-time compute scaling via accumulation and majority voting.
Restifo et al. (Wed,) studied this question.