The Research Of ECG Signals' Recognition Based On Time-frequency Characteristics

ECG (electrocardiograms) signal could reflect pathological features, and it's an important basis for heart disease's diagnosing, but it's a kind of weak, non-stationary signal which is easily influenced by noises, so it has great difficulties in recognition process. This article focuses on the research of time-frequency characteristics according to ECG signals'features, mainly researching on the detection of QRS waves, feature extraction and recognition.The main contents are listed as follows:First, aiming at the QRS waves'detection algorithm, I put forward the QRS waves detection algorithm based on three-order B-spline wavelet and self-adaption threshold, I chose 3-order B-Spline wavelet as mother wavelet which filter has a small quantity of coefficients and be good at detecting singularity under noise environment, then combining the self-adaptation threshold method to improve the speed of detecting wavelet transform's modulus maximum pairs. Simulation results in the MIT-BIH database improves that this method could detect the ECG signals with high noises and base-line drift, which also satisfying the real-time detection demand.Second, I put forward an algorithm to recognize late potentials based on wavelet energy entropy.This algorithm extracts wavelet energy entropy characteristics according to multiresolution analysis of wavelet transform, which amplifying the differences between normal and abnormal ECG signals, restraining noises.In addition, this algorithm doesn't need to localize the QRS waves pulse by pulse.The final simulation results prove the validity of this algorithm.Third,I put forward another recognition algorithm based on wavelet packets'entropy's extraction, this algorithm further studying wavelet packets characteristics which has finer time-frequency structure than wavelet characteristics.For further quantitative clustering analysis,then I input the wavelet packets parameters into multi-class support vector machine classifier, and compared with other parameters like wavelet parameters,other classifier like RBF(Radial Basis Function) neural network.The simulation results show that wavelet packets characteristics applied to multi-class SVM reaches a high recognition ratio about 97.14%, better than other compared approaches.At last, I researched on the arrhythmia signals analysis based on time-frequency distribution.I adopted several advanced time-frequency approaches to distinguish arrhythmia signals, aiming at improving time-frequency concentration and observability.This part mainly concludes two approaches. One approach is High-order spectra of WVD, this method combines advantages of WVD and High-order spectra that effectively reducing noises and improving time-frequency concentration.Another is an advanced approach based on PWVD's Hough transform, which applies Hough Transform to PWVD.This method concentrates and highlights signals,restraining noises to some extent,which improves time-frequency maps'observability.