Evaluation of Prediction Models for Smart Oil Filters

Abstract

In this master thesis, different modelling techniques for the remaining useful lifetime (RUL) of oil filter are discussed, implemented and tested.

The filter is not only crucial for cleaning the oil, but it also helps in maintaining oil pressure. The filter that is clogged can impact oil pressure as well as the overall performance of the engine. The clogged oil filter can cause oil pressure to drop unexpectedly. The best strategy is to estimate the remaining useful lifetime of such critical and essential components of machines. This estimation will provide the possibility to overcome the machine failure in advance and can save many troubles in the field.

The main idea of this study is to identify parameters that can help to find trends and predict the remaining useful life of oil filters. Moreover, different predictive models are identified, build and tested. In the light of literature review and laboratory experiments, there are recommendations of the predictive model(s) for oil filters.

Predicting RUL using threshold data leads to estimation models based on the following techniques:

1. Statistics ­ 2. Kalman filters 3. Particle filter ­ 4. Degeneration based model

The Kalman filter did not provide satisfactory results because of the non-linear behaviour of data; however, noise modelling can improve the accuracy of the results.

The extended Kalman filter provided better results than the Kalman filter, but accuracy is still very low that might improve with a better autoregressive technique for modelling the system state.

The Particle filter results were satisfactory; however, confidence period is high towards the failure threshold point which is due to particle degeneration. Random resampling is used to limit the effect of particle degeneration; however, efficient resampling technique can provide better results.

The Degeneration based model results were most promising as they are in 5% confidence period. However, there is a drawback that predictions can be started after half of the failure threshold is reached. More research study is required in this area to improve the overall predictions.