Detección de fallas en un sistema hidráulico utilizando redes neuronales artificiales

Abstract

The objective of this research project was to detect faults and conditions in a hydraulic system using artificial neural networks. It began with the investigation of the factors that influenced intelligent models for fault detection. In addition, the condition monitoring database of hydraulic systems from the repository of Carolina University was analyzed. Consequently the Python programming language was used for the approach of an intelligent model based on Deep Learning, with artificial neural networks creating a structure of visualization and data cleansing. Then, the data is dimensioned and normalized for a test set and training 80-20 of learning to ANN. An artificial neural network based on the dense Theano-Keras library was implemented as a classifier together with the Adam optimizer based on stochastic decay gradients with the sigmoid-softmax function activator. The prediction of the conditions of the hydraulic system that is shown in several confusion matrices was evaluated, resulting in an accuracy of 97% efficiency showing the development and the level of times or iterations of learning, demonstrating an acceptable level of confidence for fault detection in hydraulic systems. It was concluded that accuracy is the final metric used to measure the accuracy of the intelligent model. Before starting the data analysis with the intelligent method, it is recommended to obtain adequate and reliable information about Machine Learning and Deep Learning since all the artificial intelligence methods cannot be applied to all the intelligent models.