Operational modal analysis and prediction of remaining useful life for rotating machinery

The significance of rotating machinery spans areas from household items to vital industry sectors, such as aerospace, automotive, railway, sea transport, resource extraction, and manufacturing. Hence, our technologised society depends on efficient and reliable operation of rotating machinery. To contribute to this aim, this thesis leverages measurable quantities during its operation for structural-mechanical evaluation employing Operational Modal Analysis (OMA) and the prediction of Remaining Useful Life (RUL). Modal parameters determined by OMA are central for the design, test, and validation of rotating machinery. This thesis introduces the first open parametric simulation dataset of rotating machinery during an acceleration run. As there is a lack of similar open datasets suitable for OMA, it lays a foundation for improved reproducibility and comparability of future research. Based on this, the Averaged Order-Based Modal Analysis (AOBMA) method is developed. The novel addition of scaling and weighted averaging of individual machine orders in AOBMA alleviates the analysis effort of the existing Order-Based Modal Analysis (OBMA) method by providing a unified set of modal parameters with higher accuracy. As such, AOBMA showed a lower mean absolute relative error of 0.03 for damping ratio estimations across compared modes while OBMA provided an error value of 0.32 depending on the processed order. At excitation with high harmonic contributions, AOBMA also resulted in the highest number of accurately identified modes among the compared methods. At a harmonic ratio of 0.8, for example, AOBMA identified an average of 11.9 modes per estimation, while OBMA and baseline OMA followed with 9.5 and 9 modes, respectively. Moreover, it is the first study, which systematically evaluates the impact of excitation conditions on the compared methods and finds an advantage of OBMA and AOBMA over traditional OMA regarding mode shape estimation accuracy. While OMA can be used to evaluate significant structural changes, Machine Learning (ML) methods have seen substantially greater success in condition monitoring, including RUL prediction. However, as these methods often require large amounts of time and cost- intensive training data, a novel data-efficient RUL prediction methodology is introduced, taking advantage of distinct healthy and faulty condition data. When the number of training sequences from an open dataset is reduced to 5%, an average prediction Root Mean Square Error (RMSE) of 24.9 operation cycles is achieved, outperforming the baseline method with an RMSE of 28.1. Motivated by environmental considerations, the impact of data reduction on the training duration of several method variants is quantified. When the full training set is utilised, the most resource-saving variant of the proposed approach achieves an average training duration of 8.9% compared to the baseline method.