基于动态参数HMM的水声信号线谱轨迹提取方法  

Line Spectrum Trajectory Extraction Method of Underwater Acoustic Signal Based on Dynamic Parameter HMM

作　　者：罗昕炜[1] 李磊 沈子涵 LUO Xinwei;LI Lei;SHEN Zihan(Key Laboratory of Underwater Acoustic Signal Processing,Ministry of Education,Southeast University,Nanjing 210096,China;Shenyang Aircraft Design and Research Institute,Shenyang 110035,China)

机构地区：[1]东南大学水声信号处理教育部重点实验室,南京210096 [2]航空工业沈阳飞机设计研究所,沈阳110035

出　　处：《电子与信息学报》2022年第6期1956-1965,共10页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(12174053,91938203);中央高校基本科研业务费专项基金(2242021k30019)。

摘　　要：针对传统隐马尔可夫模型(HMM)方法提取时变线谱与多线谱的能力较弱以及动态规划过程计算量过大的问题,该文提出一种基于动态参数的1维隐马尔可夫模型(1D-HMM)的方法用于水声信号低频分析与记录(LOFAR)图中的线谱轨迹提取。该方法将时变频率状态建模为1阶马尔可夫过程,利用Viterbi算法循环提取多条线谱轨迹。在动态规划的迭代过程中,通过实时计算序列的1阶导数动态调整HMM中的状态转移概率矩阵,提升了对线谱轨迹的提取能力和多线谱的分辨能力;设计了一种基于动态滑动窗口的功率谱累积方法估计线谱的生灭,剔除虚假的线谱轨迹并判断线谱轨迹提取的结束。同时,该方法在实现过程中设计了对LOFAR图数据的块处理策略,大大减少了计算量。仿真和实际数据的处理结果表明,该方法在低信噪比条件下能够有效地检测和跟踪复杂时变频谱的频率状态,并有较好运行效率,为声呐设备的弱信号检测提供了良好的技术支持。In view of the weak ability of traditional Hidden Markov model(HMM)method to extract timevarying line spectrum and multi line spectrum and the large amount of calculation in dynamic programming process,a One-Dimensional Hidden Markov Model(1D-HMM)method based on dynamic parameters is proposed to extract line spectrum trajectory in LOw Frequency Analysis and Recording(LOFAR)diagram of underwater acoustic signal.In this method,the time-varying frequency state is modeled as a first-order Markov process,and the Viterbi algorithm is repeated several times to extract multiple frequency trajectories with the largest a posteriori probability.In the iterative process,the state transition probability matrix in HMM is dynamically adjusted by the first derivative of the sequence calculated in real time,which improves the extraction ability of line spectrum trajectory and the resolution ability of multi line spectrum;A power spectrum accumulation method based on dynamic sliding window is designed to estimate the birth and death of line spectrum and eliminate false line spectrum trajectories.At the same time,the block processing strategy is designed for LOFAR graph data in the implementation process,which reduces greatly the amount of calculation.The simulation and actual data processing results show that the method can effectively detect and track the frequency state of complex time-varying spectrum with good operational efficiency under low signalto-noise ratio conditions,which provides good technical support for the detection of weak signals in sonar devices.

关 键 词：线谱轨迹提取 频率轨迹 隐马尔可夫模型 动态搜索 

分 类 号：TN911.7[电子电信—通信与信息系统]