Geoacoustic Inversion Based On Machine Learning And Graph Signal Processing With Distributed Sensor Network

Underwater acoustic detection and sound field prediction rely on marine environmental parameters,and seabed geoacoustic parameters provide support for marine resource exploration.It is difficult to measure the geoacoustic parameters directly.Therefore,the distributed sensor network(DSN)is carried out to study geoacoustic inversion,using its spatial diversity to improve the inversion performance.Further,DSN is combined with graph signal processing to improve the accuracy of the inversion parameters.Based on the principle of acoustic reciprocity,the moving sound source and distributed fixed nodes are equivalent to the receiving experimental structure of a fixed sound source and a synthetic horizontal array.Geoacoustic inversion is performed based on the matching field processing with the synthetic horizontal array data,and the frequency coherent objective function is used to avoid the problem of synchronization of the data received by the distributed nodes.Then,the graph signal processing is applied to the inversion results of each node to improve the accuracy of geoacoustic parameters.Machine learning method improves the performance of geoacoustic inversion with large dataset.Based on geoacoustic inversion model,the mapping relationship between sound field features and geoacoustic parameters is constructed through the network.The performance of the neural network regression model,the random forest regression model and the XGBoost regression model for geoa-coustic inversion are studied respectively,and the sensitivity of the attenuation coefficient of the sediment has been improved.Among the three regression models for geoacoustic inversion,the XGBoost regression model has the smallest amount of computation.The DSN is combined with graph signal processing based on machine learning inversion re-sults.Using the spatial diversity characteristics of DSN,the performance of geoacoustic inversion is further improved through graph filtering algorithm.The graph signal is constructed from the inversion results on each node of the DSN while the graph topology is constructed in combina-tion with the node position information,and the inversion results are optimized by graph filtering.Graph signal processing can reduce abnormal values caused by interference and noise,improving the smoothness of graph signals and the reliability of inversion results.On the basis of numerical simulation,the sea experiment is designed.And the experimental data is used to verify the inversion framework,which combines distributed sensor network and graph signal processing to improve the performance of geoacoustic inversion.The results of the synthetic horizontal linear array inversion are close to the results of the vertical array inversion,and the inversion performance of distributed sensor network is comparable to that of vertical array.Moreover,the dispersion of geoacoustic parameters is further reduced after graph signal processing.