Research On Aircraft Type Recognition Based On Shortwave Speech Communication

Aircraft type recognition based on shortwave speech communication, is a new topic innon-cooperation field, which is different from radar and image means. The former has moredetection operating range. However, the way to identify aircraft type mainly is by experiencedprofessionals and no open relevant literatures are reported about it. Combined with theNational Natural Science Foundation ‘Research on aircraft type recognition based onshortwave speech communication’, based on the research of the target signal, the relatedresearch work on aircraft type recognition is conducted.Firstly, to identify the target signal, Welch algorithm is used to analyze the backgroundsignal/aircraft signal in this paper. The result shows that there exists a lot of differencebetween the ground signal and aircraft signal, which realizes the location of target signal inthe aircraft type recognition. The physical characteristics of aircraft cockpit background soundis the primary task of studied based on shortwave speech communication. Ensemble mean andensemble variance are defined and this paper applys them to frequency domain and cepstrum,which analyzes physical characteristics of background sound. The result reveals that differenttype aircraft has the different characteristics in spectral peak, which is the preparation workfor speech suppression and feature extraction.Secondly, the target signal is the background sound of aircraft cockpit and not the pilots’speech. Here the speech becomes serious interference which must be removed. Speechdetection algorithm in noise conditions based on spectral entropy is used to get speechsegment. For suppressing speech, For suppressing speech, two algorithms, empirical modedecomposition and wavelet transform(EMD_WT) and the algorithm of ensemble empiricalmode decomposition and wavelet packet (EEMD_WP) was proposed. In the process ofspeech suppression, EMD_WT produces extra noise pollution and brought hidden trouble.EEMD_WP keeps the aircraft cockpit background sound and weakens speech greatly.Comparision tests prove that EEMD_WP has better performance both in time domain and infrequency domain.Thirdly, the feature extraction is an important work in aircraft type recognition. In theview of the auditory perception, feature is extracted by bark wavelet packet. According to the procedure of Mel cepstrum coefficient and physical characteristics, nodes of wavelet packetdecomposition are selected by diagonal slice mode. The above features are combined withtime sequential Delta feature of MFCC and PLP. Pricipal component analysis is used toreduce the redundancy of high order features. The performance of all features components areestimated by Fisher criterion which produces brief and multi-angle features. The experimentsshow that the The mixed features have better classification performance than signal featureand implement the modeling of the aircraft cabin background sound, which makes preparationfor the further research on aircraft type recognition.Finally, to overcome the problem in parameter redundancy of Gaussian mixturemodel(GMM), general gradient learning algorithm is used to implement BayesianYing-Yang(BYY) harmony learning and model selection. In the process, number of Gaussiancomponent and parameter estimation are finished automatically. However, general gradientlearning algorithm easily traps into a local maximum value of the harmony function so thatthe modeling or clustering result is not reasonable. In order to overcome this disadvantage,natural gradient learning algorithm based on Riemann manifold was proposed to implementBYY harmony learning. The method can realize automated model selection and number ofGMM component can be pruned in various degrees effeciently. Moreover, the speed ofconverges is more quickly and accurately than general gradient learning algorithm. Furthermore, it makes space distribution fittting better for target signal and improve recognitionperformance in very few information.