Battlefield Sound Based On Mathematical MorphologyResearch On The Recognition Method Targets

This dissertation studied the battlefield acoustic target recognition method based on mathematical morphology which can be applied on the anti-tank and anti-helicopter intelligent mines.Since mathematical morphology fractal dimension has a single feature and low discrimination in the signal characteristics,three improved feature extraction methods were proposed to obtain the nonlinear characteristics of battlefield acoustic signals effectively.Then,the battlefield sound targets were identified based on morphological neural network.Main work of this dissertation includes the following aspects.(1)A method combining the Variational Mode Decomposition(VMD)with fractal dimensions was presented to quickly identify different armored acoustic target,such as tank acoustic signal and helicopter acoustic signal.Firstly,the number of Intrinsic Mode Functions(IMFs)decomposed by VMD was determined according to the frequency domain energy distribution of the two kinds of acoustic signals.Then,the capacity dimensions and the information dimensions of the IMFs for each kind of sound were calculate by the mathematical morphologic method.Finally,the fractal dimensions were used as the feature vectors by SVM to recognize the acoustic targets.The proposed method showed a higher recognition rate and operation speed compared with traditional EMD and the means of box counting.To obtain more nonlinear characteristics,the acoustic recognized method based on morphological multifractal of double dimensions changed was proposed according to the multifractal characteristics of battlefield acoustic target.This method defined double dimensions changed distributed function.The regression analysis was used to show that the accuracy of fitting with the function was high and the slope of two points could be used as the fractal dimension.The best scale group was selected based on the criterion of speed and recognition rate.The simulation results showed that the algorithm was faster than the previous method of measurement in morphological multifractal obviously.The calculated multifractal dimensions were used as feature input and the support vector machine was used for acoustic target recognition.The results showed that the acoustic target recognition rate was increased by 23.5% compared with the previous method.Therefore,the proposed method could be a better choice for battlefield acoustic target recognition using the nonlinear characteristic of the signal.(2)Since the fractal feature had few extracted eigenvalues on nonlinear feature extraction and the problem of insufficient signal nonlinearity was reflected,singular value decomposition and mathematical method based on SVD and mathematical morphology fractal dimension spectrum(SVD-MMFDS)is introduced.The Hankel matrix was constructed using the collected acoustic signals,and the constructed matrix was decomposed by the SVD decomposition method.Then,the signal component was reconstructed according to the relationship between the singular value and the frequency component who were arranged according to the magnitude of the amplitude.The reconstruction steps were as follows.Firstly,the fractal dimension of the first frequency component was calculated.Then the second frequency component was superimposed and the fractal dimension of the superimposed signal was calculated.Once superimposed,the fractal dimension was calculated once again until the complete original signal was formed.Finally,a fractal dimension spectrum reflecting the nonlinearity of the signal,namely SVD-MMFDS,was obtained.Compared with the EMD-box counting method,the nonlinear features extracted by SVD-MMFDS had better discrimination,and successfully solved the problem that the fractal dimension could not fully reflect the signal nonlinearity.(3)The Constructive Morphology Neural Network(CMNN)and the Fuzzy Lattice Constructive Morphology Neural Network(FL-CMNN)were used to identify battlefield acoustic targets.According to the constructive neural network model of Peter Sussner,a training method based on single-dimensional box was proposed,which improved the computational efficiency and did not generate redundant intervals.The fuzzy morphological neural network was introduced to solve the problem that the data points in the classification were outside the singledimensional box.The neural network training process is the same as the morphological neural network,but the test process of fuzzy morphological neural nfetwork is more scientific.The multi-fractal dimension extracted in Chapter 3 and the SVD-MMFDS extracted in Chapter 4 were used as feature inputs.Compared with other classifiers,such as SVM and BP neural network,CMNN and FL-CMNN had an excellent performance in training efficiency,test efficiency and classification accuracy.