Machine learning based fault analysis for electrical motors

In modern industrial facilities, it is crucial for mass production to continue uninterrupted and efficiently. Any breakdowns that may occur on production lines can lead to extended downtimes, negatively impacting production efficiency (Overall Equipment Effectiveness - OEE). Planned maintenance and predictive maintenance practices are crucial in preventing these failures. In addition to traditional maintenance methods, predictive maintenance practices have gained prominence in maintenance activities, especially with the influence of modern industrial technologies, including Industry 4.0. Through this method, also known as condition-based maintenance, potential faults requiring maintenance in production systems can be predicted in advance, allowing them to be addressed before they occur and without long system shutdowns. In this thesis study, fault analyses were performed on the starch transfer system of a food production facility using vibration and temperature values obtained from spiral motor and gearbox as well as ambient temperature and humidity parameters. through different models. As a result of the fault analyses conducted, the ability to predict long downtimes caused by faults is crucial for preventing breakdowns that could lead to production, time, and cost losses. In this analysis study, machine learning algorithms, which are one of the artificial intelligence methods, have been used. Within the scope of machine learning algorithms, the performance data of the systems have been compared using supervised learning methods such as k-nearest neighbors (KNN), decision tree (DT), Naive Bayes (NB), support vector machines (SVM), random forest (RF), and XGBoost algorithms In this study, the input data for the models included ambient temperature and humidity, motor temperature and vibration, as well as gearbox temperature and vibration parameters. The model has four outputs: no failure, short stop failure, long stop failure close, and major failure during the long stop. These data were analyzed using some supervised learning algorithms in a Python program. As a result of the study, it was found that the XGBoost and DT algorithms provided higher accuracy rates.