Multiple features can be extracted from time–frequency representation of signals for the purpose of acoustic event detection. However, many underwater acoustic signals are formed by multiple events (impulsive and tonal), which generates difficulty on the high-resolution TFR for each component. For the characterization of such different events, we propose an anisotropic chirplet transform to achieve the TFR with high energy concentration. Such transform applies a time–frequency varying Gaussian window to compensate the energy of each component while suppressing unwanted noise. Using a set of directional chirplet ridges from the obtained TFR, a structure-split-merge algorithm is designed to reconstruct a multimodal sparse representation, which provides instantaneous frequency and time features. Specifically, a pulsed-to-tonal ratio, based on these features, is computed to distinguish two acoustic signals. The presented method is validated using shallow water experimental underwater acoustic communication signals and large sequences of harmonics and pulsed bursts from common whales.

中文翻译：

---

水下声信号的多峰稀疏时频表示

可以从信号的时频表示中提取多个特征，以进行声音事件检测。但是，许多水下声信号是由多个事件（冲动和音调）形成的，这给每个组件的高分辨率TFR带来了困难。为了表征这些不同的事件，我们提出了一种各向异性的Chirplet变换，以实现具有高能量浓度的TFR。这种变换应用时频变化的高斯窗，以补偿每个分量的能量，同时抑制不想要的噪声。使用从获得的TFR中获得的一组定向chirplet脊，设计了一种结构拆分合并算法来重构多模态稀疏表示，该表示提供了瞬时频率和时间特征。具体来说，是脉冲对音调的比率，基于这些特征，进行计算以区分两个声学信号。通过浅水实验水下声通​​信信号以及来自大鲸鱼的大量谐波和脉冲爆发序列对所提出的方法进行了验证。