The Bayesian Network which is derived from the casebase using Bayesian Knowledge Discoverer is given in Fig 5.3. The network structure is separated in two parts, one with failure mode which is connected to failure description , failure consequence and failure remark which again is connected to design class code . In the other substructure, inventory id is connected to operational time and maintenance on sub units . The latter is again connected to the down time of the unit.
It seems reasonable that the network contains two substructures, because the first one deals mainly with features that describe a failure, and the second one deals with features that describe equipment. The only exception is the design class code which intuitively should have been connected to the inventory id . This might be because the majority of the objects in the casebase have the feature failure remark missing. (See plot of feature values in appendix G).
It also seems reasonable that the two features failure mode and inventory id are the ones who are the roots of each substructure. The failure mode feature states what is wrong with the equipment, if it is ``no function'', ``erroneous function'', ``failure indication'' or ``other''. It is likely that these vary with the different types of failure states and with the failure consequence, and with the failure remark. The operational time, down time and maintenance on equipment is likely to be influenced by the inventory item in inventory id .
The long time spent on computing the networks with BKD is probably due to the high number of features with a large range of values. The feature INVENTORY_I_ID has a range of 20, and MAIN_EVENT_SU_CODE has a range of 30. The range of these features could have been reduced, or the features replaced by others to create the network faster, and also to make the network representation more mathematically sound.