A RUL model based on KICA and maximum likelihood estimation with Particle Filters

Residual Useful Life (RUL) prediction model based on stochastic filtering is an important component of Condition Based Maintenance. Currently there are many difficult problems on stochastic filtering model. The first is how to use the historical oil analysis data and the new monitoring information effectively in the prediction model. For metal concentration information is a cumulant and usually affected by oil changes, the total metal concentration can not reflect the degeneration process of an item accurately. The second is how to extract the feature of data accurately when with a great deal of non-linear and non- Gaussian condition monitoring information and then make a dimensional reduction. To the problems above, firstly, a stochastic filtering prediction model based on the increment features of oil data was established. Secondly used Kernel independent component analysis (KICA) algorithm to accomplish the dimensional reduction process of high-dimension monitoring data, thirdly introduced a parameter estimation method which combining the gradient-free maximum likelihood parameter estimation and the Particle Filter(PF) algorithm together in evaluating the unknown parameters´ value and finally prognostics the RUL of an item. At last, the feasibility and performance of the proposed prediction model through experimental oil samples was assessed.