Investigation of Time-Frequency Feature Combinations with Histogram Layer Time Delay Neural Networks

2024-09-20

arXiv.org

AI Summary

Conventional time-frequency (T-F) representations exhibit limited discriminative power for underwater acoustic signal recognition under low signal-to-noise ratio (SNR) conditions. Method: This paper proposes the Histogram-layer Time-Delay Neural Network (HL-TDNN) framework, which systematically evaluates and fuses multiple T-F features—including STFT, Constant-Q Transform (CQT), and Mel-spectrogram—at the feature level. A learnable histogram layer replaces conventional frame-level statistics to enhance robustness and interpretability, and ablation studies rigorously validate the efficacy of feature combinations. Contribution/Results: We identify, for the first time, that specific multi-scale T-F fusion strategies—particularly CQT+Mel—significantly improve model discriminability. On a real-world underwater target recognition task, the optimal fusion yields a 5.2% absolute accuracy gain. These results demonstrate the critical role of acoustics-informed, synergistic T-F feature design for low-SNR modeling, establishing an interpretable, high-performance, and lightweight paradigm for weak underwater signal recognition.

Technology Category

• Machine Learning: Representation Learning
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Planning, Routing, and Scheduling: Activity and Plan Recognition

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

While deep learning has reduced the prevalence of manual feature extraction, transformation of data via feature engineering remains essential for improving model performance, particularly for underwater acoustic signals. The methods by which audio signals are converted into time-frequency representations and the subsequent handling of these spectrograms can significantly impact performance. This work demonstrates the performance impact of using different combinations of time-frequency features in a histogram layer time delay neural network. An optimal set of features is identified with results indicating that specific feature combinations outperform single data features.

Problem

• Optimizing time-frequency feature combinations for neural networks
• Improving underwater acoustic signal processing performance
• Identifying optimal feature sets for enhanced model accuracy

Innovation

• Combines time-frequency features for neural networks
• Uses histogram layer time delay neural networks
• Optimizes feature combinations for acoustic signals