Application of Deep Learning Method for Condition Monitoring and Fault Diagnosis from Vibration Data in Bearings

Since bearings are machine elements that are frequently used in several industry due to their load carrying capacity, they are subjected to wear or breakage such as adhesion, abrasion and creep under overloading conditions. For this reason, condition monitoring and fault detection are an important issue for sustainability, high performance and reliability. Feature selection is a difficult task, hence, some features may change due to changing working conditions. Therefore, in this study, convolutional neural networks (ESA), which is a deep learning method in which features are determined by internal dynamics, are used for the detection of healthy bearings (SR) and bearing failures (outer ring failure-AR1, inner ring failure-AR2, rolling element failure-AR3). In order to train ESA approaches with different architectures, spectrograms of vibration signals using Short-Time Fourier Transform were obtained. The results of GoogleNet, ResNet-50, EfficientNet-B0 and AlexNet approaches that are trained with spectograms are comparatively examined. It has been seen that ESAs with complex architectures (GoogleNet, ResNet-50, EfficientNet-B0 ) detect failures with 100% accuracy and AlexNet with 90% accuracy, but it has been observed that the training time increases as the network structure changes and the number of layers increases. It is observed that the results of the study are far better than the similar papers in the literature. As a result, it is seen that the convolutional neural network method with different approaches provides high classification accuracy in the most basic bearing fault detection and is a promising method for fault diagnosis.