1.State Key Laboratory of Acoustics, Institute of Acoustics, Chinese Academy of Sciences, Beijing 100190, China 2.School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049, China
Abstract:In order to improve the computational efficiency of algorithms while exploring the method to overcome the ambiguity problems in underwater geo-acoustic inversion, we use the data of transmission losses at the broadband sound frequencies and multiple propagating distances with the matrix of polynomial chaos expansion coefficients of transmission losses to invert the speed (c), attenuation (α) of compression sound wave and the density ratio of seabed to seawater (ρ) in their prior searching intervals. When approximating the transmission loss with the polynomial chaos expansion, the expansion coefficients are the functions of parameters including sound frequency, source and hydrophone’s position while the polynomial bases are functions of the above geo-acoustic parameters which are uniformly distributed in their respective intervals. The expansion coefficients are calculated by embedding the orthogonal polynomial bases into the acoustic wide-angle parabolic equation model. After that, the coefficients are deduced using the Galerkin projection and least angel regression. Under the situations of low sound frequency, short or medium sound propagation distance and short or medium length of intervals of geo-acoustic parameters, the polynomial chaos expansion can approximate the transmission losses accurately with the relatively error less than 1%. In the simulation case, with the high signal to noise ratio and the low errors of relative distances between source and receivers, the geo-acoustic parameters can be inverted accurately when the appropriate truncated powers are chosen. And the time cost is reduced by at least an order of magnitude compared with that of traversal grids searching procedure. Keywords:geo-acoustic inversion/ polynomial chaos expansion/ acoustic wide-angle parabolic equation
研究得到, (14)式中当声速、吸收率的分布区间, 随机多项式展开的截断幂次一定时, 平面波复声压的相对误差正比于频率、水平传播距离; 误差上限一定时, 随机多项式展开截断幂次越大, 可预报的频率f越高、距离r越远. 图1给出随机多项式展开在不同截断幂次下, 频率-距离平面上复声压相对误差等于1%的等值线图. 可以看出, 误差小于1%的最远距离r (单位为km)与最高频率f近似满足$ r\propto {f}^{-1} $, 如频率$ f=50\;{\rm{Hz}} $, 误差小于1%的最远距离r为2.9 km. 图 1 不同截断幂次下, 在“频率f-距离r”平面上随机多项式展开平面波声压相对误差1%的等值线. 其中声速的范围是1645?1655 m/s, 吸收率的范围是0.55?0.65 dB/λ Figure1. Isolines of 1% relative error about sound pressure for plane wave expanded by the polynomial chaos in frequency-range space. The intervals of sound speed and attenuation are 1645?1655 m/s and 0.55?0.65 dB/λ, respectively.
表1Pekeris波导的水文环境和声源参数 Table1.Hydrological conditions of Pekeris waveguide and acoustic source parameters.
在展开截断幂次N = 4时, Pekeris波导中地声参数取区间中值时的距离-传播损失曲线、地声参数区间内随机多项式展开近似传播损失的距离-验证集平均误差曲线分别[30]如图2和图3所示. 在传播水平距离10 km以内, 随机多项式展开近似传播损失的误差小于1%; 在趋势上误差仍随着水平传播距离r的增加而增加, 但不再单调, 如图2和图3中圈点部分, 传播损失的部分极大值点位置上近似误差出现极大值. 图 2 Pekeris波导中100 m深度不同水平距离上的声传播损失, 圈点为传播损失的部分极大值点 Figure2. Acoustic transmission loss at different horizontal ranges with the depth of 100 m in Pekeris waveguide. Circled points are partial local maximum points.
图 3 Pekeris波导中100 m深度不同水平距离上随机多项式展开近似声传播损失的验证集误差, 圈点为部分误差极大值点 Figure3. Validation set errors of acoustic transmission loss expanded by polynomial chaos at different horizontal ranges with the depth of 100 m in Pekeris waveguide. Circled points are partial local maximum points.
在计算效率方面, 嵌入随机多项式的声学抛物方程的计算复杂度与算法在深度、水平方向的差分网格数和随机多项式展开截断幂次有关. 若声场计算中, 在深度方向上差分离散的网格数为正整数Z, 水平方向上差分离散的网格数为正整数X, 随机多项式展开基个数为Q (对应随机多项式展开截断幂次为 N, Q与N的关系见(2)式), 算法的时间复杂度为
图4为遍历方案中地声参数的搜索网格数Y取值为50, 随机多项式展开截断幂次N取不同值时遍历搜索法与随机多项式展开法运算复杂度之比随抛物方程算法在深度上的离散网格数Z的变化曲线. 可见, 若遍历搜索法的参数搜索网格数$Y$一定, 随机多项式展开方法的截断幂次N和抛物方程深度方向上的离散网格数$Z$越小时, 随机多项式展开法的计算效率优势越明显. 在浅海、低频环境下, 宽角抛物方程声场计算在深度方向上的离散网格数$Z$较小, 随机多项式展开法地声反演的计算效率比遍历搜索法会有极大提升. 如在截断幂次$ N\leqslant 6, Z\leqslant 6000 $时, 随机多项式方法相比遍历法的计算效率提高至少一个数量级. 图 4 不同随机多项式截断幂次N、不同深度方向差分网格数Z下, 遍历法与随机多项式展开法计算复杂度之比 Figure4. Calculation complexity ratio of the ergodic method and the polynomial chaos expansion method under different polynomial truncated power N and difference grid number in depth direction Z.
24.2.地声反演仿真计算与评估 -->
4.2.地声反演仿真计算与评估
仿真采用浅海负梯度声速剖面的声传播环境, 海水声速剖面如图5所示. 声源深度50 m, 采用20—200 Hz频段, 1/3倍频程间隔共计12个中心频点的数据. 接收水听器位于海底100 m深度, 收发水平距离为2000—5000 m, 均匀间隔共计61个接收水听器. 海底纵波声速、吸收率、比重的搜索区间为 图 5 仿真计算使用的海水声速剖面 Figure5. Sound speed profile used in the simulation case.
表3收发水平距离不同误差下反演误差均值 Table3.Average inversion errors under different horizontal distances.
在信号存在噪声、收发水平距离存在随机误差时, 存在最优的展开幂次$ {N}_{\rm{b}} $使得反演的误差最小. 图6给出了无噪声、收发距离准确, 以及信噪比15 dB、收发距离存在 ± 20 m以内的随机误差时, 反演误差均值与随机多项式截断幂次N的变化. 图 6 不同随机多项式展开截断幂次下, (a)海底声速、(b)吸收率、(c)比重反演结果的误差 Figure6. Errors of geo-acoustic inversion results for (a) sound speed, (b) attenuation and (c) ratio of density under different truncated powers of polynomial chaos.
在不考虑信号噪声和收发位置随机因素时, 反演结果随着随机多项式展开截断幂次的增加而愈加准确; 但是噪声和收发水平距离误差使得反演存在最优截断幂次$ {N}_{\rm{b}} $. 1) 当截断幂次$ N < {N}_{\rm{b}} $时, 随机多项式展开不足以准确计算整个待反演区间内的传播损失. 以声速为例, 如图7所示, 展开截断幂次$ N=1 $时, 只能准确反演区间中值附近的声速, 对反演区间的边缘位置, 由于展开幂次不足, 模型本身对传播损失的计算不准确, 导致反演结果失真. 图 7 随机多项式展开截断幂次$ N=1 $时的海底声速的反演结果 Figure7. Geo-acoustic inversion results of sound speed of seabed when the truncated power of polynomial chaos N =1.
2) 当截断幂次$N > {N}_{\rm{b}} $时, 随着截断幂次N的增加, 随机多项式展开基的个数快速增加, 系数矩阵$ \overline{\overline {\boldsymbol{I}}} ({\boldsymbol{f}}, {\boldsymbol{z}}) $的条件数快速增长, 方程(13)的解对传播损失的误差愈加敏感, 反演方法的鲁棒性丧失, 使得反演结果不准确. $ {N}_{\rm{b}} $的选择受到待反演物理量、信噪比、收发水平距离误差大小的影响. 仿真算例中, 信噪比15 dB、收发水平距离误差为 ± 20 m时, 海底声速、吸收率、比重的最优截断幂次$ {N}_{\rm{b}} $分别为$ {N}_{\rm{b}}=3, 2, 3 $. 在反演的区间内随机抽取声速、吸收率、比重的三者测试样本200组, 反演结果与95%置信区间如图8所示. 在此条件下, 海底声速、吸收率、比重的反演平均误差为1.34 m/s, 0.02 dB/λ, $ 0.03 $. 在实际的反演过程中, 为了保证反演结果的可靠性, 可以采用后验方法评估反演的结果, 从采用不同截断幂次的反演结果中挑选出泛化效果最优的结果. 图 8 信噪比为15 dB, 收发水平距离误差为 ± 20 m时, (a)模拟声速、(b)吸收率、(c)比重的反演结果及其置信区间 Figure8. Geo-acoustic inversion results and its confidential intervals of (a) sound speed, (b) sound attenuation and (c) ratios of density between seabed and sea water when the signal to noise ratio is 15 dB and error of horizontal distance is ± 20 m.