Development of condition monitoring and predictive maintenance methods using deep learning for industry 4.0

Industry 4.0 aims to bring together devices and information elements that communicate with each other, defined as the Internet of Things (IoT), used in the industrial field. Thanks to the low energy consumption of the devices and the capacity to generate large data, condition monitoring methods can be developed in devices used in the industrial field. Maintenance work performed with the traditional method is carried out periodically or after the occurrence of the malfunction. In such approaches, problems such as the routine replacement of an undamaged part or the loss of time and production due to post-failure replacement. For this reason, a study that can perform fault detection and condition monitoring by using deep learning algorithms is presented. Deep learning approach can draw meaningful results from the big data obtained. In this method, feature extraction is made using large amounts of tagged training data. The reason it is called deep learning is because of the use of multi-layered neural network. This method has many benefits. The fault diagnosis method, which is a preventive maintenance type, can detect the remaining life of the parts in the instantly monitored system before the system fails and does not require periodic maintenance. The machine learning method gives successful results in problems such as image recognition and classification. The use of this method in the analysis and classification of signals obtained in the industrial environment offers a modern approach to fault detection. It is possible to analyze vibration signals and perform bearing fault detection without using machine learning or deep learning method. Among the proposed methods, a classification method was applied by making use of the amplitude ratios of vibration signals. In the thesis study; Predictive maintenance methods were created by using machine learning method, vibration signal amplitude ratios and classification and using different deep learning algorithms. It is aimed to classify the data correctly by applying different algorithms on the data sets used. AlexNet, SqueezeNet, Resnet50 and GoogleNet algorithms were used in the study and their performances were compared. In addition, machine learning and fault classification methods were applied according to the amplitude ratios of vibration signals. It has been seen that the presented thesis study is less costly than traditional maintenance methods and detects faults with high accuracy.