Classification of bearing faults using machine learning techniques

Machine health and performance are directly related to the bearings in the machine. The performance of the bearings, which are essential parts of the machines, directly affect the performance of the machine. Failure of the bearing can reduce the efficiency of the machine. All this means not only financial, but also time, labor and occupational health losses. Due to these losses, it is very important to monitor the bearing condition and to determine the bearing failure in advance. Finding the cause of bearing failure is of great importance. The bearing failure is directly related to the Fault Modes. Determining the Bearing Failure Mode is very important for determining the cause of the fault. In this thesis, acoustic signals of ball bearings with different faults are analyzed and fault modes are classified with different machine learning techniques. In order to collect the acoustic signals, an experimental test rig was designed consisting of a motor, a load generating rotor and a recording device capable of recording stereo sound with a phase difference of 90 degrees. By using the same type of ball bearings, different types of artificial errors in accordance with the ISO15243 standard have been created on them. Acoustic data of faulty bearings were collected using a test set and a data set was formed. Transformation of acoustic data from Time Domain to Frequency domain was realized by Fast Fourier Transform using Matlab program and a software was developed to extract functional features from data in Frequency Domain. Using the obtained features, bearing failure classifications were made by applying different Machine Learning Techniques and their performance rates were compared.