An Atomic Time Scale Algorithm Using Wavelet Decomposition

An new atomic time scale algorithm and the method for processing atomic clock signals are studied in this thesis. The Neural Network is applied to estimate atomic clock signals and detach clock noise for the first time. The main contributions of the author抯 work are listed as follows: (1) The evaluation of frequency stability is studied. When we discuss the stability of atomic clock in time domain, but frequency domain, Allan variance shown great leakage. In this thesis, the wavelet variance is presented and applied to characterize diverse frequency stability in Different Frequency Rang (DFR). The practical equations of wavelet variance are deduced, too. (2) The disadvantages of traditional algorithms for atomic time scale are discussed. Firstly, when only one set of weight is used, the stability of the time scale calculated with such an algorithm can be increased in one frequency range with one sort of noise being decreased, while all the other noises in the DFR can抰 be deduced. Secondly, one set of weights can be determined only according to the stability of all clocks in one frequency range, therefore one clock which is none-stable in this frequency range but be very stable in other frequency ranges has to be wasted. Thirdly, these algorithms are limited to estimate the lately behavior of the clocks because by the weights are based on the historical character of clocks. (3) A new atomic time scale algorithm Wavelet Decomposition Algorithm is presented. All frequency stability of atomic clocks are considered in wavelet domain. The noise of time scale calculated with this algorithm is smaller than that of the classical algorithms. (4) The effect of abnormal point in clock signal predicting is studied for the first time in this thesis. The AR model and Kalman model are greatly affected by abnormal point and the interference can last for a long time. A new method of Neural Network (NN) is applied to and used to estimate and predict the clock signal, it can greatly restrain the interference of abnormal point. (5) The detachment of clock noise is analyzed. Because the noise is non-stationary, it is difficult to be detached. The Forward Neural Network with two latent layers can detach complex objecis, so, it can .be used to detach noise. This method is simple and works very well in detaching the five-sort noise of atomic clock The only thing one must do is to train the NN by using simulated noise. All programs of the new algorithms in this paper are given in MATLAB.