A multi-layer feed-forward perceptron for microwave signals processing

This paper investigates the processing of radar signals using artificial neural networks. Today, the use of FMCW radar is considered to control the agricultural implements working depth, in order to overcome the limitations of sensors based on optical or ultrasound devices towards agricultural environment (dust, rain, etc.). The objective is to determine the radar-target distance R with a direct identification of the discrete-time radar signal S/sub b/[n]. The neural network structure in a multi-layer feed-forward perceptron. Using simulation studies, we illustrate the capability of this neural network in determining the relation between R and S/sub b/. The first training strategy uses non-disturbed input data, but this solution leads to an important overfitting when noise is added to the input signals. The second strategy reduces this sensitivity by carrying out the training phase with noisy input data. A comparison of accuracy and computing times is achieved between neural network and spectral analysis (FFT).