Deep Learning-Based Identification of Cavitation Instabilities in Turbopump Inducers Using Raw Visualization Data

Abstract

Abstract. Conventional signal decomposition of flow visualization data with a limited field of view (FOV) fails to accurately identify cavitation instabilities in turbopump inducers. To overcome this limitation, a new deep learning-based framework is proposed to directly analyze visualization data and identify such instabilities. The framework employs a hybrid convolutional-recurrent neural network to extract spatiotemporal features of instability modes. The trained model correctly identifies two physically distinct instability modes—alternate blade cavitation (ABC) in a two-bladed inducer and supersynchronous rotating cavitation (RC) in a three-bladed inducer—consistent with ground-truth pressure measurements. Sensitivity analyses show that the model correctly classifies instability modes with input durations as short as one-sixth to one-half of the training sequence length and maintains over 90% accuracy when the field of view is reduced to approximately 80% of the original extent. These degraded input data conditions are not explicitly included during training, yet the model sustains its performance without retraining, demonstrating the robustness of the deep learning-based approach to limited observation conditions.