Deep Acoustic Classification of Aircraft and Engine Types Using CNN–LSTM Models: Spectral–temporal Analysis, Cross-validation, and Performance Benchmarking

Detecting aircraft engine types has a wide range of applications in both defense and civilian sectors. In military settings, identifying the engine type of an aircraft helps distinguish between passenger platforms and potential threats, enabling appropriate situational response. Recognizing the aircraft category from its acoustic signature is also crucial for real-time situational awareness and sound-based monitoring. On the civilian side, airports and urban authorities use acoustic identification to monitor overhead aircraft, manage air traffic, and assess environmental noise impacts. In this study, we propose a deep learning–based model to classify aircraft into four categories: passenger airplane, drone, fighter jet, and helicopter. Log-frequency cepstral coefficients (LFCCs) are extracted as input features and fed into a custom hybrid architecture combining Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) layers. To evaluate performance, we applied ۵-fold cross-validation and compared the proposed model with several baselines, including standard CNN, CNN-LSTM, CNN-GRU, CNN-RNN, and Support Vector Machine (SVM) classifiers. Experimental results show that our model outperforms all baselines in classification accuracy. The proposed model achieved an overall accuracy of ۸۲.۶۵% for the four-class task. After applying temporal consistency smoothing as a post-processing step, overall accuracy improved to ۸۳.۴۰%, with passenger airplane accuracy increasing from ۷۰.۷% to ۷۳.۴۴% and fighter accuracy from ۸۱.۵% to ۸۴.۶۲%. The binary engine-type classification task achieved ۸۴.۱۵% accuracy, which remained unchanged after smoothing. An ablation study confirmed the synergistic effect of combining CNN and LSTM components, as the CNN-only and LSTM-only models achieved only ۶۱.۴% and ۵۳.۹% accuracy, respectively.

کلیدواژه ها:

Aircraft engine acoustics ، Log-frequency cepstral coefficients features ، CNN-LSTM ، Engine type classification ، Acoustic signal processing ، Temporal smoothing