Bayes匹配场地声参数反演:多步退火Gibbs采样算法

Geoacoustic Parameters Inversion of Bayes Matched-field:A Multi-annealing Gibbs Sampling Algorithm

为实现海洋地声环境参数的快速高效反演,提出多步退火Gibbs(Multi-AG)采样算法,可消除因参数搜索空间设置对参数反演结果的影响,并有效解决Bayes匹配场高维参数反演过程中常见的运算量大、旁瓣高等问题。分析待反演地声参数对匹配场处理器的敏感性,用以制定多步反演与退火方案,利用Gibbs采样算法反演敏感性级别最高的参数,计算其均值并代入后续反演步骤,进而采用退火Gibbs采样算法逐步反演后续参数;利用数值仿真实验对比Metropolis-Hastings算法、Gibbs采样算法、快速Gibbs采样算法和Multi-AG采样算法的反演效果。实验结果表明,与其他3种算法相比,Multi-AG采样算法可通过最小的计算量得到均方差最小、精度最高的参数反演结果。

A multi-annealing Gibbs （muhi-AG） sampling algorithm is developed to obtain a fast, accu- rate inversion result of ocean geoacoustic parameters. The proposed algorithm can deal well with huge computation load and high side lobe in multi-dimensional inversion of Bayes matched-field, and also e- liminate the effects from the sampling bounds. The sensitivity of geoacoustic parameters to the matched- field processor is analyzed, which contributes to establish the multi-step inversion and annealing scheme. The Gibbs sampling algorithm is used to invert the highest sensitive parameters, which mean value is nec- essary to the following inversion steps. The inversion of remain parameters can be operated with annealing Gibbs sampling algorithm step by step. The inversion effects of Metropolis-Hastings, Gibbs, FGS, and multi-AG algorithms are compared through numerical experiment, and the research shows that multi-AG sampling algorithm can be used to obtain the inversion results with the smallest mean square deviation and the highest precision, while costing the least algorithm computation.