CHAPTER 5 FAULT PATTERN EXTRACTION
5.2 T HE SEVERITY OF A FAULT PATTERN AND THE CRITICALITY OF A FAULT STATE
in the no-fault states, while at least one pattern among the set must be found in each fault state, the objective of fault detection using fault pattern is as follow (Definition 2)
Definition 2 (Fault detection using the fault pattern): to extract a set of fault patterns that are found in the fault sates but not in the no-fault states,, and at least one pattern among the set should be found in each fault state.
๐ฆ๐ข๐ง ๐๐ ๐๐ข๐โ ๐๐๐๐ข๐ where
๐๐ ๐๐ข๐: the total number of the given fault states;
๐๐๐๐ข๐: the number of discernible fault states obtained by Algorithm 5.2.
Algorithm 5.2 Discernible fault states
Require: ๐(๐) (a set of event codes), ๐๐๐ก๐ก๐๐๐๐๐๐๐ข๐๐ก (the pth fault pattern), ๐๐ ๐๐ข๐ (the total number of the given fault states)
1: ๐๐ข๐๐๐๐_๐๐_๐๐๐ ๐๐๐๐๐๐๐๐_๐๐๐ข๐๐ก_๐ ๐ก๐๐ก๐๐ โ 0
2: for (๐ = ๐; ๐ < ๐๐ ๐๐ข๐+ ๐; ๐ + +) do // For all fault states
3: If there exists any ๐๐๐ก๐ก๐๐๐๐๐๐๐ข๐๐ก โ for any ๐ such that it is found in the qth fault state 4: ๐๐ข๐๐๐๐_๐๐_๐๐๐ ๐๐๐๐๐๐๐๐_๐๐๐ข๐๐ก_๐ ๐ก๐๐ก๐๐ + +
5: End if 6: End for
5.2.1 The severity degree of a fault pattern
In terms of online fault detection, the extracted fault patterns will be compared with the event codes of the current time segment of the monitored sensor signals. In this procedure, we want to know how severe the monitored pattern is. If a specific fault pattern is discovered in every fault state, it is reasonable to define it as a strongest reference for online fault detection. Unfortunately, if a certain pattern is found from an individual state, it is not reasonable to employ it as only one reference pattern for detecting every fault occurrence. However, it should not be neglected because a fault can be caused by various root causes, and the weak pattern can represent a minor root cause of a fault state.
Therefore, it is reasonable to employ a set of all extracted fault patterns for online detection, and provide a commensurate score with its frequency of the corresponding patternโs occurrence as a weight. That is, a severity degree of a fault pattern is assessed in proportional to the found number of the fault pattern, and consequently it will be used to quantitatively estimate the effect of the fault pattern.
Algorithm 5.3 explains how to calculate the severity degree of each fault pattern. Because the severity degree is calculated based on the number of the corresponding pattern occurrences in every fault state, a higher value indicates that the fault pattern is found in many fault states and can therefore be considered as significant. If the same fault pattern is discovered multiple times in a single fault
Algorithm 5.3 A severity degree of a fault pattern
Require: ๐(๐) (a set of event codes), ๐๐๐ญ๐ญ๐๐ซ๐ง๐๐๐ฎ๐ฅ๐ญ (fault patterns), ๐๐๐ก๐ก๐๐๐๐๐๐๐ข๐๐ก (the pth fault pattern), ๐๐ ๐๐ข๐ (the total number of the given fault states)
1: for (๐ = ๐; ๐ < |๐๐๐ญ๐ญ๐๐ซ๐ง๐๐๐ฎ๐ฅ๐ญ| + ๐; ๐ + +) do
2: ๐๐๐ข๐๐๐๐๐๐๐ (๐๐๐ก๐ก๐๐๐๐๐๐๐ข๐๐ก) โ 0 // ๐๐๐ก๐ก๐๐๐๐๐๐๐ข๐๐ก is the pth element of ๐๐๐ญ๐ญ๐๐ซ๐ง๐๐๐ฎ๐ฅ๐ญ 3: for (๐ = ๐; ๐ < ๐๐๐๐๐+ ๐; ๐ + +) do
4: If there exists ๐๐๐ก๐ก๐๐๐๐๐๐๐ข๐๐ก that it is found in the qth fault state 5: ๐๐๐๐ข๐๐๐๐๐๐๐ (๐๐๐ก๐ก๐๐๐๐๐๐๐ข๐๐ก) + +
6: End if5 7: End for
8: ๐ ๐๐ฃ๐๐๐๐ก๐ฆ(๐๐๐ก๐ก๐๐๐๐๐๐๐ข๐๐ก) โ ๐๐๐๐๐๐๐๐ง๐๐ก๐๐๐(๐๐๐ข๐๐๐๐๐๐๐ (๐๐๐ก๐ก๐๐๐๐๐๐๐ข๐๐ก), ๐๐ ๐๐ข๐) 9: End for
state, it is only counted once because it describes only one fault state. After counting the number of the fault pattern occurrence, then the occurrence number is divided by the total number of the given fault state as a normalization step.
For example, Figure 5.2-(b) shows the computational procedure for calculating the severity degree of the extracted fault patterns. The first fault pattern โ211โ (i.e., [๐๐๐ ๐๐๐ ๐๐๐]๐ป) has the highest severity degree (= 1.00) because it is found in every fault state. The other two patterns, โ312โ and
โ231โ, have smaller severity degrees (0.25) in the example, however, the degrees are larger than 0 because a pattern should correspond to at least one defect state according to Definition 1.
5.2.2 The criticality of a fault state
Although multiple fault states are given to analyze, each fault state can show different hazardous level to the system, such as critical, major, minor, warning, or indeterminate. This is because the current dangerous degree is an important factor for finding optimal maintenance actions, scheduling repair procedure, and further analyzing root causes of defects (D. Goyal & B. S. Pabla, 2015). Therefore, we also introduce a method to identify the criticality of a fault state quantitatively to answer how critical the current fault state is to the system.
Figure 5.3 Examples of the defect patternโs degree of severity and the criticality of the defect states:
(a) a matrix of discretized state vectors, D(X), in which the number of sensors is 3 and the recording time is 1โ33, (b) the procedure for the severity degree of defect patterns, and (c) the computation of the criticality of each defect state (the weight of the most severe pattern is 0.8).
If multiple patterns are identified in a fault state, it is reasonable to classify the target system as more dangerous than those with fewer patterns. Similarly, if a more severe pattern (i.e., a fault pattern which shows high severity) is identified in a fault state, the state can be classified as more dangerous than other states with a less severe pattern. Considering both the characteristics, the criticality of a fault state is assessed by an exponentially weighted average of the severity of extracted defect patterns, as the following equation:
๐๐๐๐ก๐๐๐๐๐๐ก๐ฆ(๐ญ๐บ๐) = ๐ผ ร ๐ ๐๐ฃ๐๐๐๐ก๐ฆ(๐๐๐ก๐ก๐๐๐,1๐๐๐ข๐๐ก) + โ(1 โ ๐ผ)๐โ1ร ๐ ๐๐ฃ๐๐๐๐ก๐ฆ(๐๐๐ก๐ก๐๐๐๐,๐๐๐๐ข๐๐ก)
๐๐
๐=2
where ๐๐ is the total number of extracted fault patterns in the qth fault state, ๐๐๐ก๐ก๐๐๐,๐๐๐๐ข๐๐ก is the rth severe pattern in the qth state, and ๐ผ is a user-defined weight for the most severe pattern.
Assume that the weight for the most severe pattern is set to 0.8 and the matrix of discretized state vectors is given as shown in Figure 5.3-(a). For the first fault state which is from time segment 5 to 10, โ211โ is only found, therefore, the criticality of the first fault state is calculated as 0.8 ร 1 = 0.8.
In contrast, the largest number of fault patterns, โ211โ, โ322โ, โ312โ. โ231โ, is extracted in the third fault state. The weights of the four patterns are assigned to be 0.8, 0.2, 0.04, and 0.008 respectively (as exponential decreasing), and then, the final criticality of the third state is 0.912, as shown in Figure 5.3-(c).