Neural network models for predicting soil behavior for intelligent zero-cycle construction machinery control systems

Abstract

This study is devoted to the development and application of neural network models for predicting soil behavior to create intelligent control systems for construction machinery with a zero work cycle. It has been established that traditional methods for predicting deformations and stability of soil massifs during zero-cycle excavation work are not accurate enough due to the complex nonlinear nature of the interaction of soil with the working organs of machinery. The paper proposes a multi-level architecture of a neural network model that integrates convolutional and recurrent layers with the attention mechanism, which demonstrates increased prediction accuracy by 27,8% compared to traditional machine learning models. The developed approach takes into account both spatial and temporal correlation between indicators of soil condition and parameters of functioning of construction equipment. Experimental studies conducted at 12 construction sites with different types of soils have shown that the implementation of the developed system reduces fuel consumption by 18,5%, increases the productivity of earthmoving operations by 23,7% and reduces the risks of emergency situations by 32,1%. The results obtained indicate a high potential for integrating neural network technologies into construction machinery management systems to optimize zero-cycle construction processes.