Frequency Estimation using Spectral Techniques with the Support of a Deep Learning Method

In the case of damage detection, it is important to estimate the frequencies accurately. DFTbased methods provide us with amplitude-frequency pairs, but displayed frequencies carry important errors in the case of short signals. On the other hand, the amplitudes displayed for a sinusoidal signal with different time lengths describe approximately a sinc function. When involving interpolation to find the maxima of the sinc function, to which the real amplitude of the signal corresponds, it is necessary to ensure the existence of at least three points on the main lobe of the sinc function. To this aim, we apply zero-padding to the original signal in such a way that its length is doubled. The frequency estimation method proposed in this paper involves an artificial neural network (ANN). The three amplitudes taken from the main lobe determined for all considered signal lengths are the input values used to train the network. The correction term that allows us to evaluate the frequency will be the target. Following the simulations made, it is found that using normalized data sets we can estimate any frequencies, irrespective of the frequency with which the network was trained. It applies to any signal length, and any signal amplitude.