Artificial neural network model based on Kolmogorov-Arnold representation theorem and retention mechanism for real-time aircraft flight phases classification

In modern aviation, accurate classification of flight phases is crucial for both operational safety and efficiency, as well as for reliable air traffic modeling and prediction. This paper introduces KARMA (Kolmogorov-Arnold Retention Memory Aware), a novel artificial neural network that uniquely integrates the Kolmogorov-Arnold representation theorem with a multi-scale retention mechanism for real-time flight phase classification. KARMA incorporates Kolmogorov-Arnold Network (KAN) layers for flexible function approximation, and employs convolutional feature extraction to enhance temporal data processing. The model also features a memory aggregation block that maintains context across sequential predictions. Benchmarking against state-of-the-art models—including Transformers, Retentive Networks, Extended Long-Short Term Memory, and State Space Models—demonstrates that KARMA achieves superior accuracy. The KARMA model achieved an accuracy of 0.93, a precision of 0.93 and a recall of 0.96. Additionally, the research introduces a custom, high-quality dataset collected with the developed device, enabling real-time, expert-labeled sensor data. KARMA's compact design ensures suitability for embedded avionics systems, setting a new standard for flight phase classification.