Prediction of Degradation of Concrete Surface Layer Using Neural Networks Applied to Ultrasound Propagation Signals

To study the process of concrete degradation, the so-called spatiotemporal waveform profiles were obtained, which are sets of ultrasonic signals acquired by step-by-step surface profiling of the concrete surface. The recorded signals were analyzed and informative areas for three fre-quencies were identified. The type of the created neural network is a multilayer perceptron. Stochastic gradient descent was chosen as the learning algorithm. Measurement datasets (test, training and validation) were created to determine two factors of interest - the degree of materi-al degradation and the thickness (depth) of the degraded layer. The article proves that the train-ing datasets are quite sufficient to obtain acceptable results. It is shown that the results for the Fourier amplitude spectra are significantly worse than the results of neural networks built on the basis of information about the measured signals themselves.