Research On Information Fusion Method For Underwater Fast Target Recognition

With the emergence of fast underwater targets,due to their small,fast and flexible characteristics,fast underwater targets have gradually become an important underwater weapon.The underwater environment is becoming more and more complex,and traditional methods are difficult to identify well.Therefore,how to identify fast underwater targets well is a new and important task for protecting marine rights.Information fusion improves the accuracy of target recognition by combining multi-sensor data.In the context of this article,fusion of signals measured by multiple sonars achieves the purpose of improving the accuracy of underwater rapid target recognition.This paper is used weighted feature fusion to fuse multi-sensor data.This article mainly studies the following aspects:First,the Relief F feature selection method based on de-redundancy is used.Because of the complexity of the underwater environment,it is necessary to extract the characteristics of the underwater fast target from multiple aspects as much as possible to better identify the underwater fast target to form a comprehensive feature set,but this will also cause the problem of too high dimensionality.At this time,feature selection is required.Aiming at the problem that Relief F cannot remove redundant features,the feature set is de-redundant.Firstly,the redundant features are judged according to the distance correlation coefficient.Aiming at the problem of information loss caused by traditional redundant feature selection methods,an autoencoder is used to complete the task of redundant feature fusion.Secondly,in view of the inability to determine whether the features after the autoencoder fusion are helpful for classification,a classifier is used as an auxiliary task and an undercomplete autoencoder together to form a multi-task model,and the undercomplete autoencoder is assisted by the task of classification So that the hidden layer of the self-encoder can learn features that are more conducive to classification.Finally,Relief F is used to select features from the redundant feature set.Then calculate the weight to complete the feature weighted fusion.In view of the frequent changes in the orientation of fast underwater targets,dynamic weights are proposed.Compared with the traditional static weight,the dynamic weight proposed in this paper takes into account the change of the target orientation,and is more suitable for the underwater environment.Aiming at the problem that the single-criterion method of calculating weights is too one-sided and unstable,a method of calculating weights based on a multi-criteria index tree is proposed.The weight is determined according to the multi-criteria method,which is more comprehensive and improves the stability than the single criterion.Finally,the speed feature weight is proposed for the problem that the underwater fast target is not always uniform.Since the underwater fast target is a process of acceleration first and then uniform speed,and the sound signal in the acceleration process is more meaningful for recognition,a speed feature weight and the weight calculated by the previous multi-criteria are proposed to form a second-order Weight,and then perform feature weighted fusion on the feature set after feature selection.Finally,it is verified through experiments.According to comparative experiments,some of the improved methods proposed above are tested,and it is verified that the optimized feature selection method and the calculation of weights in feature weighted fusion are effective in the context of underwater rapid target recognition.