Modular Artificial Neural Network to Classify Materials and Determine Water Content in Soil Samples

In the construction industry, it is necessary to know the soil type and its water content to ensure compliance with required specifications. Existing solutions involve expensive equipment and require significant time to deliver reliable results. This article focuses on the application of modular neural networks to automate the analysis of measurement data for five soil types. The data analyzed were obtained using resistive and capacitive sensors, as well as the bulk volume mass of materials. A modular architecture consisting of 14 neural networks was designed. One sequential network specialized in material classification with an Adam optimizer. The other 13 neural networks were trained using evolutionary strategies by material type and water content range. The results show that modular architecture improves response time and reliability for individual network models, achieving an accuracy of 94.69%. The modular system was validated using 20% of the database and 10-fold cross-validation. For water content determination, the accuracy in the material with the highest variability was −0.1770% with a standard deviation of 0.6239%. The use of this modular system reduces operating and analysis times in material classification and water content determination through its real-time application. It validates its use in soil analysis processes for construction and can be used in educational settings.