Uma metodologia experimental para avaliar abordagens de aprendizado de máquina para diagnóstico de falhas com base em sinais de vibração Vitória, ES

Abstract

This work presents a systematic procedure to fairly compare experimental performance values for machine learning approaches for fault diagnosis based on vibration signals. In the vast majority of related scientific publications, the estimate of accuracy and similar performance criteria are the only quality parameters presented. The methodology that was used to generate the results of these publications is predominantly biased, based on validation methods that are too simple. In addition, all methods, in general, recycle identical patterns to estimate the best hyperparameters, introducing additional overfitting in the results. To repair this problem, the conditions used in the training, validation and test division were critically analyzed and an algorithm was proposed that allows a well-defined comparison of the experimental results. To illustrate the work’s ideas, the Case Western Reserve University Bearing Data benchmark is used as a case study. Four distinct classifiers are compared experimentally, under more difficult generalization tasks using the proposed evaluation structure: K-Nearest Neighbors, Support Vector Machine, Random Forest and One-dimensional Convolutional Neural Network. An extensive review of the literature suggests that most research work at Case Western Reserve University Bearing Data uses similar standards for training and testing, making classification an easy task.