Resumen

PEREZ RAMOS, Yandy; FERNANDEZ-ABALLI ALTAMIRANO, Carlos; CARDENAS BARRERA, Julian L.  y  HERRERA FERNANDEZ, Francisco. Distributed architecture for fault detection in industrial equipment with improved Precision Score and Robustness Index. EAC [online]. 2023, vol.44, n.3, pp. 26-40.  Epub 01-Feb-2024. ISSN 1815-5928.

Creating algorithms and systems that can process and store large amounts of data represents a great scientific, economic, and practical challenge. The application of machine learning (ML) to these problems is not trivial, and even less so if the processing of these algorithms needs to be distributed to handle the large computational load of data analysis and decision making. This paper presents a distributed and robust architecture to train, deploy, and execute distributed failure detection algorithm pipelines improving their Robustness and Precision. The solution is based on Smart Operational Realtime Bigdata Analytics (SORBA), a patented distributed architecture. The architecture combines the metrics of Robustness and Precision to automatically optimize the selection of industrial failure detection machine learning algorithm pipelines and their hyperparameters. A system of modules is developed for the acquisition, normalization, data conditioning, training, deployment, and online execution of machine learning algorithm pipelines. The solution was validated by comparing the Machine Learning (ML) results of two use cases: an industrial motor and a locomotive battery, with those achieved with Spark. The experiments showed an average improvement on the Precision Score of 28.76% and Robustness Index of 10.9%. The solution streamlines the implementation of successful applications and improves the performance of these indicators with respect to the solutions currently available in the Spark MLlib.

Palabras clave : Industrial failure detection; distributed architecture; Machine learning; Industrial data processing; Edge Computing..

        · resumen en Español     · texto en Español     · Español ( pdf )