Prediction of flutter onset by an LSTM neural network from measured time-variable responses of a randomly-tested airfoil using Lyapunov exponents for flutter classification

Abstract

Flutter, a self-excited oscillation due to energy transfer from the flow to the structure, can cause catastrophic failures in many aerospace structures if uncontrolled. Mostly, predictions of flutter states rely on model-based evaluations under restrictive conditions, such as constant Mach numbers and altitude, which are challenging to replicate outside laboratories. To counter this problem, we investigated flutter prediction using artificial intelligence, specifically long short-term memory (LSTM) neural networks on dynamically varied operational data to simulate real-world conditions. A novel test rig of wing model in a closed circular wind tunnel with controlled airflow velocity was used for flutter simulations under variable conditions. Hundreds of vibration records, captured at critical trigger levels, formed a robust dataset for flutter classification and prediction. Average divergence and Lyapunov largest exponent methods were used to classify stability and chaos in the system, which provided valuable input data for training
artificial intelligence. Analysis of results demonstrated the efficacy of neural networks in rapidly identifying flutter onset, which could contribute to advancements in flutter monitoring airborne structures under diverse operational conditions.