Clasificación de fallas en rodamientos utilizando datos de análisis de ultrasonido basado en una máquina de aprendizaje extremo

Abstract

The main objective of this research work was to classify bearing failures from a database with ultrasound parameters by applying the extreme learning machine (ELM) algorithm, for this purpose, first the database obtained from a curricular integration work available in the repository of Escuela Superior Politécnica de Chimborazo was acquired. Next, an exploratory analysis of the database was carried out by filtering the parameters referred to ultrasound where it was evidenced the existence of missing data in each state of bearing failure, so a pre-treatment was made by increasing the database in each state of failure by means of the oversampling technique. Before applying ELM, a partition of the total data into training data and test data was made to subsequently extract a set of characteristics that are presented when analyzing the data and thus be able to enter the algorithm to be trained; then the ELM class was created by adjusting the main hyper parameters such as the number of neurons in the hidden layer, activation function among others that are necessary for training. Once the model was trained with this data set, it was tested by entering the test data and the results were evaluated allowing to classify with the maximum effectiveness obtained the bearings found in the different failure states. The accuracy in the training set resulted in 94.08% and the accuracy in the test set resulted in 90.43%. With this it can be concluded that is that the limit of the model performance was reached with the parameters and data provided in this learning model what would be recommended to have more initial parameters for better performance of the applied ELM model.