Deduction model for low-frequency sound field of anechoic tank based on cLSTM-AM

基于cLSTM-AM消声水池低频声场推演模型

¹ Key Laboratory of High Performance Ship Technology(Wuhan University of Technology), Ministry of Education, Wuhan 430063, China
² School of Naval Architecture, Ocean and Energy Power Engineering, Wuhan University of Technology, Wuhan 430063, China
³ The 710th Research Institute of China Shipbuilding Industry Corporation, Yichang 443003, China

Received: 21 November 2024

Abstract

To address the challenges of reverberation and low measurement accuracy under non-free field conditions in the low-frequency sound field measurement of anechoic tank, a long short-term memory network model (cLSTM-AM) with convolution layer and attention mechanism is proposed to realize the accurate deduction of acoustic radiation and acoustic scattering characteristics of the cylindrical shell structure in the low frequency band. Initially, the paper delineates the fundamental principles of the cLSTM-AM model, the components and operations of the convolutional layer and the attention mechanism embedding model, and the indicators for evaluating the performance of the deduction model. Subsequently, a methodology combining the simulation with the background noise of underwater test is used to meet the free field conditions of the ideal anechoic tank test, and the sound pressure data set of the deduced model is obtained. Secondly, the cLSTM-AM model with convolution layer and attention mechanism, the LSTM-AM model without convolution layer and the cLSTM model without attention mechanism are compared and evaluated. It is found that the cLSTM-AM model among them has the best prediction performance. After the parameter adjustment and optimization, the mean absolute error (MAE), root mean square error (RMSE) and Pearson correlation coefficient(COR) achieved by using the cLSTM-AM model are 0.019, 0.035 and 0.933, respectively, which verifies that the method has a high accuracy in deducing the acoustic radiation characteristics of cylindrical shells from very low frequency to low frequency in anechoic tank. Finally, the data obtained by using the cLSTM-AM model training are compared with the acoustic scattering experimental data of the cylindrical shell in the anechoic tank, which proves that the low-frequency deduction model is also effective for the acoustic scattering characteristics. After the parameter optimization, the MAE, RMSE and Pearson correlation coefficient are 0.008, 0.012 and 0.993, respectively.

摘要

针对消声水池低频段声场测量存在混响、非自由场条件下测量精度低等问题，提出了一种含卷积层和注意力机制的长短期记忆网络模型(cLSTM-AM)以实现在低频段对水下结构物(以圆柱壳为典型代表)声辐射和声散射特性的准确推演。介绍了cLSTM-AM模型的算法原理、卷积层和注意力机制嵌入模型的组件和操作，以及评价推演模型性能的指标。采用仿真模拟结合水下试验背景噪声方法满足理想消声水池试验的自由场条件，获取推演模型的声压数据集。对含有卷积层和注意力机制的cLSTM-AM模型、无卷积层的LSTM-AM模型和无注意力机制的cLSTM模型进行了对比评估，发现cLSTM-AM模型具有最好的预测性能，参数调整和优化后，其平均绝对误差、均方根误差和皮尔逊相关系数分别为0.019, 0.035和0.933，验证了使用该方法对消声水池中甚低频至低频段圆柱壳声辐射特性推演具有较高的准确率。将cLSTM-AM模型训练得到的数据与消声水池中圆柱壳声散射试验数据进行对比，得到该低频推演模型对圆柱壳在低频段的声散射特性同样奏效，参数优化后，平均绝对误差、均方根误差和皮尔逊相关系数分别为0.008, 0.012和0.993。