| Direction of arrival(DOA)estimation of underwater targets has always been the focus and difficulty of military and civilian applications in the underwater navigation,battlefield reconnaissance and ocean development domains all over the world.Using the outputs of underwater acoustic array,it is generally necessary to realize the perception of the underwater acoustic environment and the interested targets through the design of various inversion technologies.However,the application of inversion technology often requires varying strong assumptions and constraints on the application scene,such as the specific seabed depth distribution,the temporal and spatial distribution of seawater temperature and salinity,et al.In practice,the temporal and spatial distribution of ocean parameters required for inversion is often complex and variable,which is always seriously mismatched with the inversion physical models.Obviously,the uncertainty of the inversion model seriously restricts the performance of the sonar array.On the other hand,modern artificial intelligence technology based on big data analysis and machine learning technology has shown significantly better performance than traditional model driven methods in the fields of computer vision,Internet,finance,transportation,et al.However,due to the complexity of underwater acoustic signals,the research progress of modern artificial intelligence technology in the analysis and application of underwater acoustic array signals is still relatively slow.The main reason is that it is always difficult to collect enough sample data with known truth labels required by artificial intelligence technology in the field of underwater acoustic.In order to address this problem in a certain extent,this dissertation takes the construction of DOA estimation machine learning model and its practical application extrapolation as the core idea,takes multi-model spatial spectral complementary information mining as the key entry point,and takes the high accuracy of DOA estimation in practical underwater environments as the final goal,the following innovative research contents have been studied:(1)For the problem that the performance of classical source number estimation methods decreases sharply under the low Signal-to-Noise(SNR)condition,two different source number and SNR joint estimation methods are proposed based on machine learning models constructed in the multi-task machine learning framework.Array response in ideal environment is the basis for realizing environmental parameter perception and target detection,and the source number is the basic constraint of most high-performance DOA estimation methods.Under the actual low SNR conditions,the performance of existing source number estimation methods will decline sharply,which will seriously affect the performance of DOA estimation methods.Therefore,firstly,this dissertation makes a theoretical analysis on the problem of source number estimation under the condition of low SNR,and formulate the problem to be the multi-task problem of joint estimation of source number and SNR.Therefore,under the guidance of two typical multi-task machine learning frameworks,joint source number and SNR estimation methods based on serial cascade and parallel cascade of multi-machine learning units are proposed respectively.And these two methods can all be used to provide a reliable source number support for the subsequent DOA estimation procedure.(2)For the problem that there is always contradiction between the computational complexity,DOA estimation performance and the environmental adaptability of DOA estimation method,a novel DOA estimation based on stacking meta-learning method with complementary spatial spectra is proposed.The DOA estimation method based on strict physical constraints needs to significantly increase the computational complexity to ensure its performance under harsh conditions such as low SNR.At the same time,additional assumptions or computational models will affect the generality of the algorithm to a certain extent.On the other hand,it is found that the representation ability of typical general low computational complexity DOA estimation algorithms is not zero under complex conditions such as low SNR,and the spatial spectrum information given by various general low computational complexity DOA estimation algorithms is significantly complementary to DOA information with the support of reliable source number information.Therefore,in this dissertation low computational complexity DOA estimation algorithms are regarded as the "soft sensor" for DOA information perception.Consequently,using the idea of heterogeneous sensor network joint analysis and based on simulated big data,a multi-spatial spectrum joint DOA estimation method based on metaagent fusion is proposed.Experimental results illustrated that the proposed method can make full use of the DOA information representation ability of typical low complexity DOA estimation methods,and at the same time significantly improves the performance of DOA estimation results under the condition of low SNR.(3)For the problem that,in the practical application,there is certainly model mismatch of the proposed machine learning models for DOA estimation,a practical application-oriented array data enhancement method based on target and environment joint spatial-temporal representation is proposed.The data quality is therefore enhanced by extracting the target components,which effectively solves the mismatching problem mentioned above.In the actual underwater acoustic environment,the array signal received by the receiving array in the active sonar system will be affected by a variety of environmental disturbances such as non-Gaussian environmental noise,interface reverberation and water body reverberation,which cannot meet the assumption of Gaussian white noise.Furthermore,both the joint estimation method of source number and SNR based on the serial or parallel cascade of multi machine learning units and the DOA estimation method based on stacking meta-agent fusion with multiple spatial spectrum are designed with the assumption of white Gaussian noise.Therefore,for the designed DOA estimation related machine learning models,it is necessary to whiten the actual data received by the active sonar system to complete the matching between the data and the models.In order to achieve the above purpose,under the weak constraint of local stationary of underwater acoustic environment,the optimal decomposition of ideal target component and environment component in target data segment is realized by analyzing the spatial-temporal spectrum multi-view joint representation of target data segment and nearest environment data segment.As a result,the DOA estimation related machine learning models can be used for DOA estimation in practice.Based on the above research results,several underwater target detection experiments of practical active sonar detection system have been carried out.Among them,the test results of lake trial illustrated that through the joint target-environment array data enhancement,the DOA estimation related machine learning models proposed in this dissertation can achieve the single angle accuracy which is significantly better than the existing methods.The results of sea trial experiment shown that the machine learning method proposed in this dissertation can realize accurate response of echoes from targets both in water and buried in seabed,which reflects that the superior performance of stable high-resolution DOA estimation for weak targets under strong interference.In summary,results of large number of simulation and actual data tests have shown that this dissertation presents a new idea to employ machine learning methods to improve the actual underwater target DOA estimation performance,which makes it can be an important technical reference to the practical application of artificial intelligence technology based on big data analysis and machine learning in underwater acoustic array signal analysis. |
