A Deep Neural Network and Acoustic Target Voiceprint Feature Extraction Method

Abstract

A deep neural network and an underwater acoustic target voiceprint feature extraction method, the deep neural network includes an input layer, a hidden layer and an output layer, used for the extraction of underwater acoustic target voiceprint features, and the number of nodes in the input layer is the underwater acoustic target signal The number of frequency points of the original signal spectrum, the number of frequency points of all frequencies within the value range of the fundamental frequency and the sum of harmonic orders, the number of nodes in the output layer is the number of frequency points of the original signal spectrum, and the number of nodes in the hidden layer is less than the number of nodes in the input layer number; the underwater acoustic target voiceprint feature extraction method includes the signal acquisition step, the fundamental frequency and harmonic acquisition step and the reconstruction step. The accurate extraction of the fundamental frequency and harmonics and the reconstruction of the original signal spectrum weaken the noise line spectrum contained in the original signal spectrum, purify the original signal spectrum, and reduce the impact of the interference line spectrum on the final identification of the final ship target individual. And can adapt to frequency drift.

Description

technical field [0001] The invention relates to a neural network and a feature extraction method, in particular to a deep neural network and an underwater acoustic target voiceprint feature extraction method, belonging to the field of target recognition. Background technique [0002] In the underwater acoustic signal, there are voiceprint features that can distinguish individuals like fingerprints. Ship radiated noise is mainly generated by sound sources such as generators, propulsion systems and auxiliary equipment on board, and can be detected and acquired by detection equipment. There are differences in the detected underwater acoustic signal corresponding to its multiple sound sources. Compared with the voiceprint features of other types of ships, the voiceprint features include simple features and complex features. The line spectrum in the voiceprint features is a simple feature. These feature lines Spectrum can be described by frequency, amplitude, and width, while th...

AI Technical Summary

Problems solved by technology

In the field of neural network research, stochastic gradient descent and error backpropagation algorithms are typical algorithms for traditional training of multi-layer networks. It is difficult to achieve ideal results when training neural networks with multiple hidden layers.

Among them, one of the main difficulties stems from the ubiquity of the local optimum of the non-convex objective function of the deep network, which makes the randomly initialized network easy to fall into a poor local optimum during training.

Images