Diagnosis of multiple rotor bar faults in asynchronous motors using convolutional neural networks approach

Asynchronous motors, which are frequently used in today's industrial drive systems, have high efficiency, robust construction, power and size diversity. Broken rotor bar failures often occur in these motors, depending on environmental and operating conditions. Since such malfunctions directly affect the working efficiency, studies on diagnosis are especially important. Although thermal, acoustic and angular velocity signals are used in the studies, it is more common to use the stator current signal and motor vibration signal. In order for fault diagnosis to be made automatically, the necessity of classification process arises. In recent studies, artificial intelligence is used in addition to some signal processing techniques for classification processes. In this study, two basic approaches are tried for diagnosis and classification of broken rotor bar failure by using machine learning and deep learning methods within the scope of artificial intelligence. Both empirical mode decomposition (EMD) and continuous wavelet transform (CWT) techniques are used to extract fault features from signals. First, a phase current and motor vibration signal from the open access faulty motor dataset is filtered and enveloped. In the first approach tried, these signals were separated into 5 intrinsic mode functions (IMF) by EMD method and spectral entropy and instantaneous frequency features were obtained. These features are added end-to-end and a new feature vector is created and classified by support vector machine (SVM), k nearest neighbor (KNN) and decision tree (DT) machine learning methods. In the second approach tried, the same current and vibration signals were converted into pictures with CWT and their features were extracted and classified with the ResNet18 model. In the last study for the second trial, the current and vibration properties extracted with ResNet18 were combined and classified with SVM. In this classification study, the highest accuracy was obtained with 100%. At the end of the study, the tried approaches and the methods made for the same data set in the literature were compared, and the most successful approach was the combination of CWT and ResNet18 used with transfer learning.