Multi-scale 1차원 C-LSTM을 사용한 수중 파쇄 음향의 분류

Abstract

In this study, a hybrid deep learning model was developed to classify sounds for real-time monitoring of underwater construction operations. The proposed model utilizes a multi-scale CNN with an attention mechanism for acoustic feature extraction and a LSTM Network for time-series classification. The model’s performance was evaluated using acoustic data recorded from actual underwater crushing experiments. The results showed that the proposed model’s F1-score was 10 points higher compared to classical classification method, which is RNN classifier with MFCC feature extraction. The results also showed that the model reduced misclassifications and improved continuity across state transitions. Finally, the computation time remained under 4ms on a CPU, confirming the model’s suitability for real-time application on teleoperation.