Anomaly detection using machine learning algorithms in predictive maintenance processes

This thesis investigates the effectiveness of using machine learning and deep learning algorithms for anomaly detection in the predictive maintenance processes of industrial motors. The study focuses on multivariate time series data obtained from various sensors, including RMS acceleration in the X and Z axes, slip angle, temperature, peak acceleration in the Z axis, high-frequency RMS acceleration, spectral kurtosis, and peak factor. These data provide valuable information about the motor's mechanical and thermal performance, as well as critical conditions such as vibration and wear. The research is based on a dataset enriched with failure durations observed on specific dates, including a detailed analysis of failures classified as "Major" and "Minor." Major failures refer to serious and large-scale impact failures, while minor failures refer to less serious or smaller-scale impact failures. This classification is important for prioritizing failures and determining maintenance strategies. Our application consists of two stages. In the first stage, classification algorithms were used to classify the failures in the dataset according to their importance; in the second stage, the aim was to predict the failure times using multivariate LSTM models. The primary objective of this thesis is to assess the effectiveness of machine learning and deep learning methods in the early detection of motor failures and to demonstrate the potential contribution of these techniques to industrial predictive maintenance strategies. The research aims to make a significant contribution to the existing literature on early detection of motor failures and to assist in the improvement of industrial maintenance practices.