2024

Junttila, Jukka; Raunio, Kalle

The NL ROSE project aims to reach the next level (NL) of real-time operational state estimation (ROSE) of rotating machines by extending the state estimation capabilities from constant speed to variable-speed operation. Real-time operational state estimation is a fundamental building block of mechanical digital twins (DTs) which in turn are seen to create additional value to the machinery industry. Feature extraction and machine learning (ML) methods based on measured acceleration vibration data were previously developed and studied for the purpose of the operational state estimation of a Wärtsilä generating set operating at constant speed as part of the project called DigiBuzz funded by Business Finland. During the DigiBuzz project it was shown that the demand to produce large amounts of high-quality data for the purposes of ROSE model development is high. Thus, creating the ability of efficient in-house data production is key for this project as well as future work to accelerate the development process of DTs and reach our goals faster.<br/>The demonstration targets of this project were two variable-speed internal combustion engines (ICEs) manufactured by ACGO Power. The vibration acceleration measurements of the two ICEs were carried out during emission tests performed at VTT engine laboratory. The applicability of the previously developed feature extraction and ML methods as well as other two openly available feature extraction methods for time series, namely MiniRocket and TSFEL, was validated using the measured data. As was expected based on previous work, the accuracies of all models were relatively high in general. However, the measurement point location proved to have a significant effect on the model accuracy. The models based on the vibration accelerations measured at the engine block drive-end side (MP4) were the most accurate at large.<br/>The models based on the previously developed feature extraction method performed similarly as the models in the previous studies. In fact, the models of this study were somewhat more accurate. Especially the models based on the accelerations measured at MP4, for which the lowest accuracy was 96.699 % and the rest were over 99.9 % accurate.<br/>The two openly available feature extraction methods performed extremely well at the task. The models based on the features extracted with either of the two methods were in most cases 100 % accurate, and 99.74 % accurate in the worst case. Feature selection showed that the number of features can be reduced radically without significant effect on the accuracy. 100 % or close to 100 % accuracy can be achieved in most cases by using only the best 25 of 9996 MiniRocket features or the best ten of the over 800 TSFEL features.<br/>The data measured in this study are not enough to cover the whole range of operation of the studied engines. However, the results show that the ROSE of variable-speed ICEs could be possible if data were available sufficiently. Thus, further development of the studied subject requires measurements covering the operational range of ICEs more comprehensively.