146
7.5 Model Comparison
To show the comparison of this new model with the model developed in Chapter 6,
Figure 7-3 and Figure 7-4 below show the plots of distribution of residual life at the
last monitoring point for each test dataset.
Figure 7-3: Case 1 – pdf residual time of engine 83000128 at the last observation point
Figure 7-4: Case 2 – pdf residual time of engine 83000130 at the last observation point
147 From these figures, it can be seen that the new model gives a new prediction that is
significantly better than the previous one. Thus, it is concluded that the more condition- monitoring data is used, the better the model estimation becomes.
7.6 Summary
This chapter has presented a new development of conditional residual time prediction, in which the previous approach is enhanced with data that we assume will increase or
decrease with the residual time. This was done by first assuming that the residual time is a function of observed reflective monitoring data,
i
y
. Then, we set up a relationship indicating how the residual time will change with the responsive monitoring data,
i
z
. Issues such as the possibility of the responsive variables being correlated with each
other were solved by using ICA. By providing the data required by the model, we then estimated the model parameters using the maximum likelihood approach. The results
from parameter estimation were sufficiently good to justify proceeding with fitting the model to the data. The results of this approach were satisfactory, showing better
prediction than the previous model with the same dataset but without
i
z
. It is concluded that more information leads to a better prediction.
148
8 CHAPTER 8: A WEAR PREDICTION MODEL
BASED ON SPECTROMETRIC OIL ANALYSIS PROGRAMME USED IN DIESEL ENGINES
8.1 Introduction
In a recent parallel study conducted by Wang 2006, the deterioration process of a system is presented by a generic term called ―wear‖. In general, the deterioration
process is assumed to be described by discrete variables as shown in Chapter 3. Wang 2006 however, proposed that a continuous random variable could also be used to
represent the deterioration process of a system and formulated its transitional probability. One of the advantages of this approach is that it eliminates the need of a
threshold level of the wear that is commonly used in other wear prediction models Pandey et al., 2005. In general, the model is formulated within a stochastic filtering
approach and hidden Markov framework, which can be used to predict the current and future wear in a system given past observed information to date. Thus, having these
predictions, it would enable us to develop appropriate maintenance decisions for a system subject to condition monitoring information.
In his recent work, Wang 2006 had used simulated data for the modelling development and later conducted a test for the model using an oil based monitoring
dataset from aircraft engines. However, it has been reported that many factors could affect the wear of the system such as engine age, type of service, environmental
conditions, engine metallurgy, etc., Macian et al., 2003. Here, we have the opportunity to study the performance of this proposed wear prediction model by re-applying it to a
different set of observed oil monitoring data. The monitoring information that is made available for this study are several sets of complete life data of diesel engines used in
ships along with the history of their monitored SOAP information, see Chapter 6 for detail of the datasets.
The aim of the study is to test, validate and compare the wear prediction model as proposed and to see whether it could be used to generalize the deterioration process of
oil based monitoring systems. A modelling methodology for the wear prediction model that is suited with the observed monitoring data will then be presented.
149
8.2 Modelling Development