Comparison of motor fault diagnosis performance using RNN and K-means for data with disturbance

Abstract

Maintenance of an industrial electric motor is very important. The most commonly used algorithm for deep learning motor diagnosis using deep learning is CNN, which is one of the representative supervised learning algorithms. However, the failure diagnosis algorithm made with the CNN algorithm is vulnerable to this data. For this reason, an algorithm that complements this has been proposed, and that is to use the RNN and K-means algorithms. The method using RNN has a cyclic neural network structure, so it can grasp the similarity of data. K-means also uses the Euclidean distance method to grasp the similarity between data and classify the data using it. Due to the characteristics of these two algorithms, even if a disturbance is an input, if the similarity of data is high, it is determined as similar data. In this paper, two algorithms were used to perform fault diagnosis and two experiments were conducted to understand the differences and characteristics of the two algorithms. As a result of experiment 1 classifying only normal failures, experiment 2 experimented by increasing the number of failures to be classified. In the case of RNN, the results of experiments 1 and 2 showed similar accuracy. However, in the case of the algorithm using K-means, the accuracy decreased as the number of classifications increased.