The Feature Extraction And Classification Of Active Sonar Echoes Based On Timbre Parameters

The decisive factor of different types of underwater targets classification correctly is thefeature effective extraction. There are a large number of feature extracted methods in timedomain, frequency domian and time-frequency combined domian, and those get a goodrecognition results. But in practice, it still depends on the auditory system of trained sonarmember to judge the kind of underwater target. The loudness, pitch, timbre, and duration areeffective characteristic parameters, which can describe the hearing attribute of a sound,among them, loudness, pitch, timbre and duration with more mature calculation model, havebeen applied in the feature classification and recognition of active sonar echoes. Because ofthe amplitude of the loudness is serious impacted by the distance between target and workplatform, while the pitch and the timbre is merely impacted by the distance, and those twoparameters can reflect the qualitative difference of different kinds of targets. Therefore, thetimbre characteristics is extracted in this paper to realize the effectively separated of the targetechoes, noise and reverberation.Based on the research on timbre dimensions of the predecessors. The auditoryparameters which related to the structure of the spectrum and the energy distribution arechosen in this paper in order to subdivising the power spectrum, and analysis of scatteringcharacteristics of the target and the bottom of the sea which can be reflected by thedistribution of the auditory parameters. HOS can restrain some noise with special probabilitydensity function, and it can detect the signal with frequency coupling, in recently years, HOSwas widely applied in many field. The auditory perception characteristics were extracted onthe third and fourth order diagonal slices spectrum in this paper for better classification effect.Therefore, it derives1.5dimentions spectrum and2.5dimentions spectrum of the CW andLFM pulse contain information of signal in the paper. Gaussian-mixture model was used tofitting the instantaneous value distribution of the target echoes, noise and the reverberation,and analyzing HOS functions of the two kinds of non-gaussian distribution. At last, SupportVector Machine (SVM)was used to classify target signals, noise and reverberation.The simulation analysis verify the feasibility of the auditory perception extracted infeature classification. Data processing for a lake try, auditory perception features extracted onthe data’s power spectrum, third and fourth order cumulant diagonal slice of spectrum wasused by echoes recognition. Experimental data processing results show that, in the differentperception feature combinated space, the highest recognition rate is about96.87%, and the average recognition rate is more than80%. Moreover, the recognition rate on2.5dimentionsspectrum is higher than that on power spectrum, and the recognition is bad on1.5dimentionsspectrum.