Abstract: An approach based on time-frequency analysis to recognize outliers in measured data series by spectrogram is proposed, which discovers the time-frequency character of measured data series. To obtain the best estimation of anti-jamming for measured series, the rule of weighted-square error is introduced to optimize the parameters of estimating model. Example shows that diagnosing outliers by spectrogram is accurate and reliable, and wavelet neural network of weighted error-energy function has excellent approximation ability and fast convergence speed. Moreover, the outliers’ influence on the estimate result can be eliminated efficiently by this way.

Key words: Outlier, Spectrogram, Wavelet neural network, Weighted error-energy function