Research On Key Technologies For Time Varying Signal Processing

Signal is the carrier of information. The key issue for signal processing are the extraction and analysis of main attribute information. Time-varying signal(TVS) as various audio signals and vital signs signal, etc., is the important kind of nonlinear non-stationary characteristics signal, which covering the main information extraction difficult and hot issues. In the most of current studies, TVS was treated as short-term smooth and linear, and both of its time and frequency domain information were focused on. While, its time-varying information has been ignored,which not only seriously impacted information extraction accuracy, but also greatly limited the practical application.This paper deals with the challenges of TVS, which focus on the core issues as the signal segment standard, the signal component analysis, signal representation and feature extraction, and classification. The main research contents include:(1) Real-time accurate segmentation is the premise of TVS signal analysis. In order to deal with segment standard and fast robust TVS segment algorithm problem, the Hilbert spectrum(HS) based adaptive signal segmentation technique has been proposed. Through deeply studied the description and recovery method of nonlinear distortion produced during non-stationary process, the signal energy features on both HS time domain and frequency domain have been extracted accurately. To enhance the distinguish ability on HS temporal characteristics, energy intensity index based on the logarithm of time-domain energy has been proposed, and used in the initial segmentation process. Then, within the inspection of the energy difference in adjacent and same time length segments, the differences in TVS frequency bands have been used in split points detection method of TVS. Compared with other main segment techniques, the segmentation results by our method was more accurate, and with fewer insertion and undetected errors. In speech recognition applications, this method gave the reasonable segmentation results in more than 87% of cases.(2) For the automatical analysis problem on active component of TVS, a method of the TVS component analysis has been propose, which integrated independent component analysis(ICA) and Hilbert Huang Transform. In the real and complex environment, TVS obtained is usually mixed from a variety of source signals together. these hidden ingredients could be separate by ICA. However, ICA has the uncertain independent component(IC) order and uncertain magnitude. To this end, to achieve a more accurate expression of the signal source, empirical mode decomposition(EMD). With the characteristic that trend term of IC could effectively characterize the differences of the signal, automatic identification of interest IC has been implemented; the more precise ingredients have been obained by frequency domain filtering under calculating the instantaneous frequency of IC. To deal with Electroencephalogram(EEG), a typical TVS with multi-component mixing and strong interference, the capability of signal separation method proposed in this paper and component extraction capability have been verified. For EOG artifact effects, this method could effectively remove 99.9% of eye artifact in the EEG signal. With the comparation to the classic method based on ICA and regression analysis, this method could obained a better correlation result 0.84 from the classic ICA 0.69 and regression analysis 0.71, at the same time, retained more information about cerebral cognitive activity.(3) TVS usually has high-dimensional feature which could be extracted only through efficient modeling. Sparse model is simple, efficient expression of TVS. This paper presented a sparse modeling and feature extraction method based on the intrinsic mode function based dictionary. The most essential component of the signal by EMD decomposition is intrinsic mode function(IMF). We proposed a creative idea that building the dictionary learning with IMF, and introducing hierarchical sparse modeling. orthogonal matching track method has been used in TVS modeling, to greatly eliminate inefficiencies in modeling and at the same time ensure the effectiveness. Then, feature extraction based on compressive sensing(CS) method has been proposed, which only contained a small amount of core sensing features coefficient(SFC). Experimental results showed that: at the same sparsity, the signal to noise ratio(SNR) of sparse model reconstructed this article, was at least 2d B higher than other mainstream methods; at the same SNR, the minimum number of atoms needed in the modeling sparse; to audio signals, this method could not only maximizely reveal the semantic information, but also suitable for audio emotion recognition.(4) The most important part of the practical application construction is a highperformance, robust TVS classifier. To deal with the complex data distribution of practical application, diverse categories, drift changes and the emergence of new classes and other issues, this paper focused on the adaptive resonance theory(ART) neural network technology, and proposed semi-supervised ART2 network(SS-ART2), and studied the operational mechanism of SS-ART2 based on similarity and confidence. SS-ART2 was good at unknown category sample by activating a new output mode, which greatly reduced the impact of emerging situations on network performance; the introduction of semi-supervised learning mechanism, made full use of marked and unmarked samples which greatly enhanced the efficiency of network training; the usage of network sample weights update method based on generalized similarity and confidence solved the labeling error, noise, bizarre samples and pattern drift and other issues; by adopting the method of local alert threshold, network shock by global adjustment could be effectively avoid. Experiments showed that the proposed SS-ART2 network structure was reasonable, and the output mode clusters could reflect the actual data and can obtained the best classification accuracy.The method proposed in this paper could synthetically utilize the difference between its own characteristics and signal system which could usually obain better performance for multi-channel TVS. While single-channel became a big challenge for the system. To systematically evaluate these methods, we selected a singlechannel 10 hours continuous news corpus, and randomly added a impulse noise. The result showed that the methods proposed in this dissertation could effectively remove all kinds of interference, and ensure the accuracy of acoustic event recognition were not affected. Moreover, by extracting signal characteristics and differences between them, accuracy of speech emotion recognition reached at 70.2%, which was about 6% higher than other mainstream methods.In this dissertation, the key issues of TVS process have been deeply and meticulously studied, mainly focused on solving precision segmentation in the realtime TVS processing system, component analysis, feature extraction and reliable identification. The good results have been achieved in difficult problems of TVS such as EEG and audio signal processing, etc., which has laid a solid foundation for a variety of TVS process and an important instructive in similar fields.