Neural network classification of granular flows from audio signals: preliminary results

Granulometric analysis is a key element in the study of geological granular flows. This category of flows includes some of the most dangerous phenomena known in geology. Knowing the size of the clasts flowing in a geological granular flow is extremely important to predict its behavior and hazard. This work combines the potential of neural networks (NN) with the task of classifying from monodisperse granular flows to determine the type of material and the granular size, directly from audio recordings. For this purpose, three NN models are proposed: one to classify the material (in this case, either dacite or bearing balls) and the other two to predict the granular size of each material. In addition, a data augmentation method based on surrogate audio signals is proposed to increase the amount of data. The results of these NN models are promising, as almost 99% of the 35,100 size predictions are located within 1 phi from the true size and more than 92% of the 21,600 material classifications are correct. Although in nature geological granular flows are polydisperse and our studies refer only to monodisperse materials, they represent a very important starting point for future studies where audio recordings from granular flows could be used as a discriminative data, with strong implications for civil protection.