| In the field of underwater acoustics,the intelligent underwater acoustic target recognition is a difficult technical challenge,which is of great significance to the implementation of maritime power strategy and the exploitation of deep-sea resources.The technical research is mainly carried out based on two directions,one is the extraction and optimization of multi-domain features,and the other is the construction of recognition model.In the traditional recognition methods,the single feature or simple serial and parallel combination features are used,which makes it difficult to accurately describe the characteristics of the target.These recognition algorithms are also based on probabilistic framework,which makes it impossible to process the uncertain information in the sample data effectively.In order to solve these problems,a new intelligent underwater acoustic target recognition method based on multi-domain feature combination optimization and evidence classification is proposed in this paper.The main works and innovations of the thesis are as follows.(1)For the problem that underwater acoustic target signal is very weak in strong noise environment,the additive noise removal algorithm based on subspace transformation and the convolution noise removal algorithm based on homomorphic filtering are studied respectively.These two kinds of noise suppression techniques complement each other,realizing the enhancement of the target signal and the removal of interference information,thus solving the problem that the underwater acoustic target signal is usually covered by the marine ambient noise during long-distance detection.(2)For the problem that it is difficult to fully describe the characteristics of underwater acoustic targets for a single feature,a multi-domain feature extraction and selection method is proposed in this paper.Firstly,four kinds of characteristics such as time domain waveform structure,wavelet packet decomposition energy,MFCC and PLP auditory spectrum are extracted.And the combination and supplement of these features ensure a complete and accurate characterization of the essential characteristics of the target signal.Then,the feature fusion method based on the canonical correlation analysis is studied for the problem that the series and parallel combinations can cause mutual interference between the features.Finally,a feature subset selection algorithm based on belief functions(FSBBF)is proposed to optimize the combination feature.The experimental results based on the measured data of the underwater acoustic target show that the classification time of the optimized feature can be shortened by 3~6 seconds,and the recognition rate can be increased by 5%,which effectively solves the problem of the redundant feature information disturbing the classification performance.(3)For the problem that EK-NN is susceptible to noise and special samples,a new NEK-NN classification algorithm is proposed.A compound rule called “Dempster & PCR5” is constructed in the NEK-NN.The Dempster rule is used to combine the neighbor evidence in same class as to improve the efficiency,and the PCR5 is used to combine the different kinds of evidences as to achieve accurate distribution of conflict information.In order to avoid the influence of noise data and special samples,the weight coefficient of each evidences are also set according to the number of samples in each class.The NEK-NN algorithm effectively solves the problem that the recognition information is rich in uncertainty due to the complex and changeable marine environment.(4)In order to solve the problem that the traditional underwater acoustic target recognition algorithm can not deal with the uncertain information effectively,a new evidence recognition algorithm based on NEK-NN is proposed.In the new algorithm,the global mass function of each target data is constructed by NEK-NN,so as to achieve the accurate measurement of its class attributes and uncertainties.The sample data with a relatively small proportion(less than 0.5)of the uncertain information is discriminated to the class with the highest degree of confidence and added to the corresponding training sample database to enrich the sample data of the underwater acoustic target.The sample data with a large proportion of uncertain information can be discriminated based on the distance from the updated center of each class,thus avoiding the interference of uncertain information and improving the recognition accuracy.The new algorithm is based on the framework of evidence theory to realize the analysis of the target data,to a certain extent,it can solve the problem that the underwater acoustic target sample data is scarce and is rich in uncertainty information.(5)For the problem that there is no training sample but the number of target clssses is known,a new clustering recognition algorithm called TRACE is proposed.The NEK-NN and ECM are combined in TRANE,which makes it possible to further improve the recognition accuracy of the target data.The experimental results of underwater acoustic targets show that the recognition accuracy of TRANE can be 10% higher than that of the traditional recognition algorithm.To a certain extent,the recognition problem that without training samples but with high requirements for recognition accuracy can be solved.For the problem that there is no training sample and the number of target clssses is unknown,an adaptive evidence clustering recognition algorithm(ECNEK-NN)is proposed.The initial basic belief assignment and the number of classes of the target are given at random and then updated cyclically by the NEK-NN until it no longer changes,thus achieving full adaptive clustering of underwater acoustic target data.When the number of test samples is sufficient(greater than 1000),the recognition accuracy can usually reach more than 90%.When the number of test samples is poor(less than 100),the recognition accuracy can still be above 80%.To some extent,ECNEK-NN can solve the problem of underwater target recognition without any prior knowledge. |
PDF Full Text Request