APPLICATION OF NEURAL NETWORK ALGORITHMS IN THE DEVELOPMENT OF ACOUSTIC METHODS OF CONCRETE STRENGTH CONTROL

Abstract

The paper illustrates the possibility of determining the strength of concrete using the standard protocol of acoustic signal recording and reproduction used in the vast majority of conventional portable devices, such as Android phones (and others). We developed a convolutional neural network capable of perceiving sound signals (from mechanical impacts on concrete with a hammer) pre-transformed into image-spectrograms. These spectrograms are compared to strength value established by any of the standard methods. In the long run, this approach may be the most reliable and simple method of strength monitoring in the field, requiring only a simple phone and a hammer. Moreover, this method can be adapted to many other problems where the physical and mechanical properties of a material are in some way related to the acoustic properties of an elasto-plastic body. The process of development and training of the neural network is described, and statistical evaluation of the quality of the obtained results is performed.