ESTRATEGIA DE SELECCIÓN DE HARDWARE Y ALGORITMOS DE INTELIGENCIA ARTIFICIAL EN ENTORNOS 'EDGE COMPUTING'

Abstract

Currently most of the data collected in companies and industrial manufacturing environments through IoT devices is processed in the cloud [1]. Given the large volume of data that each company manages due to the emergence of IoT, cloud computing is not the best option for certain sectors such as automotive [3]. Within this sector, the quality perceived by the end customer is closely linked to the assembly line. These assembly lines collect a high number of variables (temperatures, pressures, pumps, etc.) and a real-time prediction by means of small digital twins in this process would avoid both material and labor costs. Today, to perform such anticipation through artificial intelligence models in real time is unfeasible due to the latency that exists with cloud processing. Therefore, there is an imminent need to develop applications that are deployed at the edge of the network for the automotive manufacturing and painting process that enable the generation of digital twins. In this regard, the proposal of the 'Edge Computing' concept [2] seeks to mitigate in part this situation. For this reason, this paper has focused not only on demonstrating that splitting the data of a large model by generating small 'Edge' models maintains the predictive capacity of each algorithm, but also stress tests (limits, constraints, etc.) have been performed on the 'Edge Computing' hardware technology currently available on the market that allows generating and processing this type of models. Key Words: Edge Computing, Internet of things, Machine Learning, Industry 4.0