Study On Theory Of Passive Acoustic Target Recognition

Based on the application background of intelligent mine and analysis of classical acoustic generation mechanism & characteristics, the key technology of passive acoustic target detection and recognition is studied, corresponding theories and algorithms is proposed. The achievements in this paper exist as a theories development and engineering operation reference of passive acoustic recognition.On the basis of current technology development, the characteristics of classical acoustic in battlefield are studied, and the signal generation mechanism and characteristics of four classical acoustic targets (tank, helicopter, fighter and armored vehicle) are summarized and analysed.Signal preprocessing is one important technical way to improve the accuracy of recognition rate. The denoising algorithm in this paper is based on wavelet transform and EMD. An adaptive noise cancelling algorithm based on EMD and two denoising algorithms of multi-microphone information fusion based on time delay vector close rule (TDVCR) are proposed:1) A new reference signal extraction algorithm is proposed. Whereas the characteri-stics of EMD is adaptive, the higher IMF components are considered as the reference input of adaptive noise canceling implement. The denoising results are compared with the signlas filtered by wavelet global thresholds and layered thresholds, the exprements represent the EMD adaptive noise cancelling algorithm can reach better denoising property;2) According to the characteristics of multi-microphone time delay estimation, time delay vector close rule is proposed. Triangular time delay vector errors are put forward combining with information fusion and TDVCR of wavelet coefficients, and are weighted by time delay error thresholds defined by multi-microphone integrated support degree. The denoising signals are reconstructed by above steps. The multi-microphone denoising algorithm based on EMD is developed, using for reference of weighted IMF components. The time delay vector errors are calculated by TDVCR and also weighted by time delay thresholds defined by multi-microphne integrated support degree, and the weighting matrix of IMF function is got. The denoising signal is reconstructed with IMF function and weighting matrix. The theoretical analysis and experiments represent two denoising algorithms here show better filtering properties.Feature extraction and feature selection is another key technology of acoustic recognition. Five algorithms of feature extraction are studied, category separability norm of all engivectors are given, then the global separability and single target separability and two-target separability are researched. Based on the zero-crossing theory and AR model parameter and kernel fisher criterion analysis, one feature selection algorithm and two feature extraction algorithms are proposed:1) Remarkable feature selection based on distance separability measure is proposed. The engivectors are processed by distance separability measure. The remarkable function is constructed and used for selecting separability value. The engivectors corresponding to the selected separability value are considered as effective feature.2) Feature extraction based on EMD and power ratio is proposed. The IMF components of signals are processed by FFT, the spectrum of IMFs are got. Normalized the power ratio of IMF spectrum and original signal spectrum, and making the power ratio as engivector of signals. The category separability norm and recognition rate show that the feature extraction algorithm based on EMD and power_ratio is effective.3) Feature extraction based on multiscale frequency division is proposed. The multiscale frequency division is put forward based on multiscale theory. The signals are normalize at different frequency section, and created new engivectors. The experimental results represent the feature extraction algorithm based on multiscale frequency division is simple & effective and can be used for engineering application.At the section of classifier design for acoustic recognition, a model matching classifer based on comparability coefficients is proposed. The models of engivectors and corresponding weights are decided by offline traiing corresponding to the proper comparability function selected, and are used for classifying the target by setting thresholds. The particle neural network and SVM are applied to the acoustic target classification and recognition. The experimental results are given in the paper.In order to verify the effectivity and engineering feasibility of recognition algorithm, the recognition hardware system is designed and developed. The measurement experiments of sound velocity inside and outside are designed. The signal collecting experiments are carried by multi-microphone array inside and outside, which provide lots of signals for recogniton. The simplicity recognition algorithm is used in this hardware system. The experiements represent the hardware system has advantages of engineering application, which is prepare for arming the weapon.