Feature Selection for Robust Classification of Crack and Drop Signals

This paper presents a study to the problem of features selection for accurate recognition of crack signals in raining conditions, using a multilayer perceptron and a radial basis function neural network. The features extraction process is accomplished for two time frames: in the first time frame the presence of the signal reflection is minimal; in the later a wider window include signal reflections at the specimen edges. An extensive set of 90 features (41 of them are novel), 67 in the time domain and 23 in the frequency domain, are extracted from the normalized signals and are sorted according to Fisher ratio (F-ratio). The signals database consists of over than 20,000 simulated cracks and drops signals. The NNs classification accuracy of a single crack signal in rain conditions using the most robust features ranges from 80.08% to 99.08%, which remains in the same levels when white-Gaussian noise up to 0 dB signal-to-noise ratio is added to the signal.