DATA-DRIVEN REMAINING USEFUL LIFE (RUL) PREDICTION OF A HIGH-PRESSURE COMPRESSOR PACKING BY INTEGRATING PCA, SVD AND PARTICLE-FILTER (PF) METHOD

Innovative machine learning-based prognostic strategies are developed in this paper for predicting the remaining useful life (RUL) of high-pressure packing in plunger-type hyper compressors. Principal component analysis (PCA)-based order reduction and singular value decomposition (SVD)-based health state assessment methods are applied first. By using PCA, the number of relevant sensors can be reduced from 192 to 3. The last three relevant sensors are the leakage flow, x and y directional vibrations measured on the plunger. Correlation coefficients obtained from chorological singular-value matrices constructed using leakage-flow data are chosen to establish sum-energy-based HI model. Thirteen run-to-failure data provided by a local chemical factory facilitate the developments of three similarity-based and one similarity-PF-based RUL prediction algorithms. The resultant RUL prediction schemes are validated on four testing datasets. The latter are online datasets, therefore, no HI values can be retrieved, although the correlation coefficient, R, can be evaluated online. Two averages obtained from sum energy of the thirteen data sets were used for similarity-based RUL estimations, whereas another similarity-based scheme applies the parameters of the nearest neighbour (in Euclidean-distance sense) according to each day’s HI estimate and correlation-coefficient related parameters, from those of training datasets. The latter is also called daily best-fit similarity scheme. The similarity-PF-based approach integrates the daily best-fit similarity scheme and PF algorithm to smooth the RUL prediction. The RUL predicting results obtained from three similarity-based and similarity-PF-based approaches are compared with those obtained using FNN scheme. The comparisons indicate that the predictions obtained from the schemes proposed in this study are consistent with those of FNN. The latter have been proven to be reliable.