Multimodal identification of ramjet instability via time-frequency feature fusion

Supersonic inlet unstart poses critical safety hazards to ramjet engines and hypersonic aerial vehicles. This phenomenon encompasses multiple distinct instability modes, including little buzz, big buzz, and mixed oscillatory patterns, each imposing differentiated aerodynamic loads and control demands. However, existing identification methods fail to simultaneously achieve high accuracy and broad Mach-range applicability across multiple instability modes, and no mature multi-modal discrimination scheme is currently available. This study proposes a multi-modal ramjet instability identification framework based on a Time-Frequency Fusion CNN-LSTM (TFF-CL) architecture with a collaborative feature injection strategy. An improved K-means clustering algorithm incorporating a weighted hybrid distance metric is developed to enable accurate annotation of dispersed multi-condition data samples, achieving a labeling accuracy of 99.32% and representing an improvement of 35.45 percentage points over the standard K-means method. Continuous Wavelet Transform (CWT) is applied to convert one-dimensional pressure time-series signals into two-dimensional time-frequency representations, from which a Convolutional Neural Network (CNN) extracts oscillation waveform features. A collaborative feature injection strategy is then employed to introduce these features into the Long Short-Term Memory layer, enabling joint characterization of transient oscillation morphology and temporal dynamic information. Under wide-range variable Mach number conditions (Mach 3.0-5.0), TFF-CL achieves a validation accuracy of 95.65%, outperforming four representative comparison methods. In the online recognition validation, TFF-CL achieves an overall accuracy of 98.7% with no cross-modal misclassification. At an input sequence length of 200Δ t , the dimensionless computation time is 0.4153, well below the real-time engineering threshold of 1.0. The proposed TFF-CL framework provides an effective and generalizable solution for multi-modal inlet instability identification.