Research On Key Techniques Of Underwater Target Attribute Recognition Based On Information Fusion

The intrusion target attribute recognition under complex situations is one of the key fields of unmanned underwater detection system study, and it is important to the following situation and threat assessment, and attack strategy formulation. Recognition method based on information fusion aims to take advantage of the association, correlation, combination of data and information from multi-sources to achieve a more comprehensive and accurate estimation of the target, which then enhances the performance of the recognition system in many aspects including accuracy, robustness, universality, spoof attacks, etc. Recently, this technique has gained increasing popularity in the last few years.This thesis mainly focuses on several key techniques study of the target classification methodology for detecting devices and networks based on information fusion. The main research contents include:feature extraction methods of target acoustic and magnetic signals based on adaptive time-frequency analysis, underwater target recognition technology based on feature-level fusion for equipment monomer, and intrusion target detection and recognition technology based on underwater network’s decision-level fusion. The corresponding experiments based on sea trial recording data have been done and show the effectiveness of the proposed methods. The main contributions of this paper can be summarized as follows:1. A Wigner-2.5 time-frequency representation algorithm based on adaptive optimal parabola kernel (termed as, Wigner-2.5 AOPK TFR) is proposed. The generally utilized bilinear T-F analysis methods always present weak performance in resisting Gaussian noise and suppressing cross-term interference. In this paper, the higher-order T-F representation (Wigner-2.5 TFR) which could outperform the bilinear T-F distribution in noisy conditions is introduced as the basic signal T-F processing method. The optimal parabola kernel (OPK), which could greatly pass the auto-terms and cancel the cross-terms in the ambiguity plane of Wigner-2.5 TFR, is obtained by using adaptive optimizing technology. And, the OPK is then utilized to modify the Wigner-2.5 TFR to obtain the Wigner-2.5 AOPK TFR, which enhances the performance in cross-term and noise suppression and signal adaptability. The proposed Wigner-2.5 AOPK TFR is utilized to obtain the high order T-F matrix and T-F spectrogram of the target acoustic and magnetic signals. And, Feature datasets, which represent the macroscopic statistical properties of the target acoustic and magnetic signals’ T-F components distributed in the T-F plane, are then extracted and constructed. The extracted feature datasets include:the local time-varying and frequency-varying features of the target T-F matrix’s singular vectors and the macro shape features of the T-F spectrogram.2. A fast entropy assisted complete ensemble empirical mode decomposition algorithm (termed as, FEACEEMD) is proposed. Recently, the empirical mode decomposition (EMD) and the improved complete ensemble EMD with adaptive noise (Improved CEEMDAN) are known as two major adaptive methods to decompose a complex and non-linear signal into a set of oscillation scales (named as modes, or IMF), but they all have their own advantages and disadvantages. In this paper, with the help of IMF’s intermittent and random assessment index, the Improved CEEMDAN and the EMD are seamlessly serial fused to bring in the FEACEEMD method. The proposed method possesses the purpose of overcoming the traditional’mode mixing’ problem, and as well as guaranteeing the real-time calculation. Additionally, in order to faster select the real and useful modes from all modes obtained by FEACEEMD, this paper also gives an effective mode sequential selection rule based on IMF’s cross correlation index and variance contribution rate. The proposed FEACEEMD method and the effective mode sequential selection rule are utilized to obtain the effective modes contained in the target acoustic and magnetic signals. The obtained effective modes have a close relationship with the target’s structure components, and which are then studied to extract the detailed nuance differencing features of the target acoustic and magnetic structure. The extracted mode nuance feature datasets include:the IMFs’linear predictive coefficients feature (AR model coefficients, ARDC), and the IMFs’ discrete parametric features, such as variation degree, energy distribution ratio, and instantaneous spectral centroid.3. The original feature datasets, obtained from the target’s acoustic and magnetic signals, can not be directly fused because of the presence of noise and interference. In this paper, a fuzzy rough feature selection algorithm based on improved artificial fish swarm optimization method (termed as, AFFRFS) is studied. In order to find a robust、optimal、real、globally and minimal reduction of the original feature datasets, the AFFRFS redefines and optimizes the behaviors of the traditional artificial fish swarm optimization method to faster find the candidate feature subsets from the original high dimensional feature datasets, and then introduces the fuzzy rough theory to evaluate the candidate feature subset’s classification performance to determine the final optimal feature subset. The reduction experiments show that the selected feature subsets can be better used in the following fusion process than the original feature datasets, and the proposed AFFRFS is effective in feature selection application.4. In order to obtain a better recognition performance, a feature-level fusion algorithm based on the generalized discriminative multi-set canonical correlation analysis (termed as, GDMCCA) is proposed, and the kernelized algorithm of GDMCCA (re-termed as, KGDMCCA) used for multi-nonlinear feature subsets fusion is further presented based on the kernel optimization trick. GDMCCA and KGDMCCA are presented as measures of association among multiple sets of random variables (feature datasets), but they embody more class information of the classification samples. From mathematical point, the paper gives a detailed solution of the proposed two algorithms and then presents the feature fusion steps based on them. The proposed fusion steps lay a theoretical foundation for mine monomer target classification based on multiple physical fields and multiple feature subsets. In terms of three fusion scenarios, including homologous acoustic multiple feature-subsets, homologous magnetic multiple feature-subsets and heterologous acoustic-magnetic multiple feature-subsets, this research conducts a variety of feature fusion classification experiments to verify the proposed fusion algorithm’s effectiveness. Results show that the proposed algorithms are more effective and robust than other related feature fusion methods under the same fusion target classification scenarios.5. For a given underwater network with so many particular constraints in doing underwater target detection, a novel distributed target fusion detection algorithm is proposed based on the optimal window statistics. In this algorithm, the traditional hypothesis testing of a ’point’ target is extended to the hypothesis testing of an optimal target acoustic irradiation window. The paper gives the detail description of the optimal window statistics and provides the detailed description of the fusion target detection rule. The false alarm probability in system-level is approximately derived and the detection probability is also estimated by using Monte Carlo statistical experiments. The results show that the proposed fusion detection rule can perform better in system-level detection compared with the existing ones. Subsequently, using the optimal PCR6 evidence fusion rule, a target fusion classification model, which gets its group evidence from the sensor nodes covered by the above-mentioned MNs’ optimal window, is proposed. The usage of PCR6 involved in the model is then optimized and simplified based on the group evidences’ pre-processing method. Simulation experiments verify the effectiveness of the proposed target fusion classification model.This paper deeply investigates several key techniques involved in the classification methods of unmanned underwater detection system based on information fusion. The proposed algorithms are verified by experiments and the results are positive. The research achievements of this paper lay a solid foundation for the future underwater target detection system development.