Defence in mechanical engineering M.Sc. (Tech) Jesse Miettinen

Opponent Professor Robert X. Gao, Case Western Reserve University, U.S.A.

Custos Professor Raine Viitala, Aalto University, School of Engineering, Department of Mechanical Engineering

AI can decrease the workload of condition monitoring experts) Vibration-based condition monitoring can be useful for planning the maintenance of rotating machines. Vibration patterns often reveal information of faulty components and performance of the machine. However, vibration analysis still requires much manual effort from experts. Numerous previous studies have shown that deep learning is a promising approach for automating many laborious vibration analysis tasks. Deep learning models can be optimised to extract features from vibration data, and then, to process the extracted features effectively. However, vibration patterns are very sensitive to the operating conditions of the rotating machine, such as the rotating speed and load. These deviations to the vibration patterns decrease the performance of deep learning models, especially if the models have been trained with sparse vibration data. Moreover, explaining the decision-making process of deep learning models is difficult, which decreases trust towards these models in condition monitoring tasks. These factors have largely hindered the industrial adoption of deep learning for vibration analysis. To this end, the thesis experimentally demonstrates best practices related to the vibration data acquisition, model optimisation and model structuring for condition monitoring purposes. The results show that the training data should include vibration patterns acquired under the operating conditions that the rotating machines are typically used. Moreover, the results suggest that torsional vibration measured with torque transducers is a very potent source of data for deep learning-based fault diagnosis models. In addition, the results indicate that deep learning models perform more accurately in condition monitoring tasks if they are optimised to process simultaneously sampled vibration data from many sensors. Finally, 1D convolutional networks are excellent models for analysing the vibration data in the time series format. Furthermore, the fault diagnosis decisions of these types of models can be explained, and the results show that they can learn to analyse vibration on a reasonable basis. In practice, these results are very useful for any research and development project related to the application of deep learning for vibration-based condition monitoring.