It is vital to maintain the stability and longevity of mechanical systems in the crucial field of vibration monitoring. The thesis introduces three sets of algorithms specifically tailored to fulfill the function of detecting or predicting vibration faults in the Water Injection Systems of Floating Production Storage and Offloading Unit (FPSO). These methodologies include PCA-based Prognosis, LSTM + One-class SVM, and VARIMA + One-class SVM. Notably, the LSTM + One-class SVM algorithm exhibits superior detective performance and robust resistance to data fluctuations, surpassing the other two approaches. Furthermore, it becomes evident that applying Principal Component Analysis for dimensionality reduction can adversely affect the discerning abilities inherent to both LSTM- and VARIMA-related algorithms.

In the context of predictive functionality, given the absence of definitive indicators, the development of predictive models remains contingent upon the existence of robust detective models. Both LSTM- and VARIMA-related algorithms (excluding PCA) demonstrate their efficacy in fulfilling this prerequisite. While the performance of PCA-based Prognosis continues to lag behind, its distinctive capacity to delve into the intricate patterns of equipment operational states hints at the possibility of unearthing richer insights compared to the aforementioned techniques. Consequently, the potential of PCA-based Prognosis to evolve into a viable predictive model should not be underestimated.
... Read more