A multitask convolutional neural network (CNN) is trained to localize the instantaneous position of a motorboat throughout its transit past a wide aperture linear array of hydrophones located 1 m above the sea floor in water 20 m deep. A cepstrogram database for each hydrophone and a cross-correlogram database for each pair of adjacent hydrophones are compiled for multiple motorboat transits. Cepstrum-based and correlation-based feature vectors (along with ground-truth source bearing and range data) form the inputs to train three CNNs so that they can predict the instantaneous source range and bearing for other “unseen” motorboat transits. It is shown that CNNs operating on multi-sensor cepstrum-based feature maps are able to predict the instantaneous range and bearing of a transiting motorboat, even when the source is near an endfire direction. Also, multi-sensor generalised cross correlation-based feature maps are able to predict the range and bearing of a transiting motorboat in the presence of interfering multipath arrivals. When compared with the cepstrum-only CNN, cross correlation-only CNN, and the conventional model-based method of passive ranging by wavefront curvature, the combined cepstrum-cross correlation CNN is shown to provide superior source localization performance in a multipath underwater acoustic environment.

中文翻译：

---

使用水听器阵列对传输的宽带源进行声学定位的多任务卷积神经网络

训练多任务卷积神经网络 (CNN) 以定位汽艇在其通过位于海底上方 1 m 处、水深 20 m 的大孔径线性水听器阵列的整个过程中的瞬时位置。每个水听器的倒谱数据库和每对相邻水听器的互相关数据库被编译用于多个摩托艇过境。基于倒谱和相关性的特征向量（连同地面实况源方位和距离数据）构成了训练三个 CNN 的输入，以便它们可以预测其他“看不见的”摩托艇过境的瞬时源范围和方位。结果表明，在多传感器倒谱上运行的 CNN基于特征图能够预测过境摩托艇的瞬时范围和方位，即使源位于端火方向附近。此外，基于多传感器广义互相关的特征图能够在存在干扰多径到达的情况下预测经过的摩托艇的范围和方位。与仅倒谱的 CNN、仅互相关的 CNN 和传统的基于模型的基于波前曲率的被动测距方法相比，组合的倒谱-互相关CNN 在多径水声环境中可提供卓越的声源定位性能.