Predictive maintenance on turbofan aircraft engine and bearing data with cascade deep learning models

With the developing sensor technology and the Industry 4.0 revolution, a large amount of data is collected from each step of the production lines and systems. These obtained data are used for both quality estimation and estimation of maintenance times of systems and equipment. With Industry 4.0, maintenance activities have also evolved, and predictive maintenance is being worked on before malfunctions occur. Maintenance processes have become more important today due to the high number of sensitive products and reliability in production. Maintenance activities performed at the right time both increase quality and reduce maintenance costs. With predictive maintenance applications, systems are monitored before failures occur and in a way to prevent unnecessary parts replacement. In addition, in these studies carried out with the collected data, the source and type of failure can be predicted within the data to reduce downtime, and costs can be further reduced. Within the scope of this thesis, Remaining Useful Life studies were carried out within the scope of predictive maintenance on the data obtained from systems and equipment. Deep learning structures, a sub-branch of machine learning, were used in these prediction models. Within the scope of this thesis, the Remaining Useful Life estimation was studied on the Commercial Aero-Propulsion System Simulation aircraft engine dataset and bearing vibration data. In these studies, it is aimed to develop a data-driven model. Deep learning-based prediction models are presented. Self-Attention Layer-based structures, which have come to the forefront in recent years, are proposed. Thus, it is aimed to focus on the points that will affect the result both in sensor data and in the required attributes. In the aircraft engine dataset, a model that makes more precise predictions with the help of the Self-Attention layer and the layer containing shortcuts has been developed. In addition, an asymmetric loss function is proposed for this prediction problem. The model is tested on four different scenarios and sub-datasets, and the successful results are reported. With the proposed model, 11.017, 23.14, 12.62, and 26.39 values were obtained in the Root Mean Squared Error of evaluation metric in FD001, FD002, FD003, and FD004 sub-datasets, respectively. In the score function proposed in the data competition within the scope of this problem, 157.19, 2490.09, 218.6, and 3800.15 values were obtained, respectively. Within the scope of the study, the effect of specific layers of the proposed model on the results was also analyzed. In the bearing vibration dataset, a new deep learning model is presented based on the Transformer structure, which has given superior performance results in the fields where it has been used in recent years. In this dataset, vibration data with high sampling frequency is made available by preprocessing the information contained in the frequency domain. The data brought to the Mel frequency domain were trained with a model with a deep structure. A new score evaluation function is proposed to compare the performance of the prediction models in the data competition where the dataset is proposed. The results obtained on the test bearings gave a value of 0.433 in this score function. The results obtained on each bearing in the test dataset and the overall score value are presented in comparison with other models in the literature, and the superiority in performance is indicated. More accurate Remaining Useful Life predictions can be made with the increasing amount of data and the development of deep learning techniques. In particular, introducing new models in prediction studies and applying these structures with deep structures in data such as vibration and aircraft engines give successful results. Applying these structures to systems has paved the way for establishing more reliable and traceable structures.