FASE IV: Comunicación de resultados
2.4. Indicadores para la auditoría de gestión.
The first step towards model compilation for CME is a compiled version of Sherlock, called Miniature Mode Estimation (Mini-ME) [Chung, 2001].
This engine compiled component mode constraints into conflicts, and used these conflicts in an online mode estimation algorithm to determine mode estimates for the system. The online mode estimation algorithm is similar to the candidate generation step of GDE, and uses probabilities to generate likely mode estimates similar to Sherlock. The conflicts are used to generate a kernel diagnosis that satisfies all conflicts, and this kernel diagnosis is extended to a mode estimate by ensuring that all components in the system have an assigned mode. The architecture of the Mini-ME engine is shown below in Figure 3-1.
Mini-ME Offline
Dissent Generator
Online Partial Diagnosis
Trigger
Best-first Kernel Diagnosis Generator
Monitors Partial
Diagnosis Rule Generator System
Model
Most Likely Diagnosis
Dissents Partial Diagnosis Conflicts
Rules
Discrete Observations
Continuous Observations
Figure 3-1 - Architecture of the Mini-ME Engine
The architecture denotes the generation of dissents in an offline process.
Dissents are a mapping of observations to conflicts. The dissents are
transformed by Mini-ME offline into partial diagnoses. These partial diagnoses have a similar representation to the constituent diagnoses in GDE, so the term constituent diagnosis is used to refer to these partial diagnoses. This offline transformation enables Mini-ME to avoid performing this step online. In the online portion, Mini-ME only needs to determine the appropriate sets of constituent diagnoses to use given the current observations. The final step to generating a consistent diagnosis is to determine the smallest set of component mode assignments, kernel diagnoses, that are a minimal set covering of the constituent diagnoses. By choosing assignments in the constituent diagnoses, Mini-ME reconstructs the diagnosis from the conflicts, enabling the assignments chosen to satisfy all conflicts and be consistent with the observations. Mini-ME uses component mode probabilities to generate the most likely kernel diagnoses, and then extends the kernel to a full diagnosis.
3.2.1 Mini-ME Example
The diagnostic process of Mini-ME is best demonstrated by example using the NEAR Power storage system described in Chapter 1. Focusing on the interaction of the switch and redundant chargers with the observation variables of the bus-voltage, Figure 3-2 depicts the system.
Switch-voltage
Switch - voltage Bus-Voltage
Figure 3-2 - NEAR Power Storage System Example The modes of the components are given below (note that the unknown mode is not shown):
switch
(charger-1, p=0.49), (charger-2, p=0.49), (stuck-charger-1, p=0.01), (stuck-charger-2, p=0.01)
charger-1, charger-2
(full-on, p = 0.39), (trickle, p = 0.39), (off, p = 0.2), (broken, p = 0.02) bus-voltage : { zero, low, nominal }
The following are some of the relevant dissents:
[ ] ⇒ ¬[ SWITCH = STUCK-CHARGER-1 ∧ CHARGER-2 = FULL-ON]
[ ] ⇒ ¬[ SWITCH = STUCK-CHARGER-1 ∧ CHARGER-2 = TRICKLE]
[ ] ⇒ ¬[ SWITCH = CHARGER-2 ∧ CHARGER-1 = FULL-ON]
[ ] ⇒ ¬[ SWITCH = CHARGER-2 ∧ CHARGER-1 = TRICKLE]
[ ] ⇒ ¬[ SWITCH = CHARGER-1 ∧ CHARGER-2 = FULL-ON]
[ ] ⇒ ¬[ SWITCH = CHARGER-1 ∧ CHARGER-2 = TRICKLE]
[ BUS-VOLTAGE = LOW ] ⇒ ¬[ SWITCH = CHARGER-1 ∧ CHARGER-1 = FULL-ON ]
[ BUS-VOLTAGE = LOW ] ⇒ ¬[ SWITCH = CHARGER-1 ∧ CHARGER-1 = OFF ]
[ BUS-VOLTAGE = NOMINAL ] ⇒ ¬[ SWITCH = CHARGER-1 ∧ CHARGER-1
= OFF ]
These dissents express the links between switch and charger modes so that only one charger is on at any time, and that the charger that is on corresponds to the position of the switch. The dissents make this explicit.
For instance, in the third and fourth dissents, note that the component modes that are inconsistent are the switch = charger-2 and the mode charger-1 = full-on or trickle. This limits the modes of the charger-1 to be either off, broken or unknown.
The first step in Mini-ME is to use the current observations to determine the relevant dissents, and their consequents, the conflicts. Consider the observation that the bus-voltage = nominal. Mini-ME triggers those dissents that mention the observable bus-voltage = nominal, and any that do not mention an observable. The following constituent diagnoses represent the first two dissents:
[ SWITCH = CHARGER-1, SWITCH = CHARGER-2, SWITCH = STUCK-CHARGER-2, CHARGER-2 =
TRICKLE, CHARGER-2 = OFF, CHARGER-2 = BROKEN]
[ SWITCH = CHARGER-1, SWITCH = CHARGER-2, SWITCH = STUCK-CHARGER-2, CHARGER-2 = FULL-
ON, CHARGER-2 = OFF, CHARGER-2 = BROKEN]
The remaining sets of constituent diagnoses are not shown here for brevity.
Mini-ME uses these sets of constituent diagnoses to generate kernel diagnoses, which represent a minimal set covering of the constituent diagnoses. This process is similar to the GDE process of ‘candidate generation’. The generation of kernel diagnoses is guided by the probability of component mode assignments. The set covering begins by determining the most likely component mode assignment in the first set of constituent diagnoses. In this case, this results in:
switch = charger-1, p = 0.49
To perform the minimal set covering, Mini-ME determines the sets of constituent diagnoses that mention this assignment as a constituent diagnosis. Additionally, Mini-ME chooses a set of constituent diagnoses that this one does not appear. For instance, the assignment switch = charger-1 would not appear in the set of constituent diagnoses derived from dissents 5 and 6. The sets of constituent diagnoses for dissent 5 are:
[SWITCH = CHARGER-2, SWITCH = STUCK-CHARGER-1, SWITCH = STUCK-CHARGER-2, CHARGER-2 =
TRICKLE, CHARGER-2 = OFF, CHARGER-2 = BROKEN]
Mini-ME uses this set of constituent diagnoses to choose a mode assignment for charger-2 that is the most probable. This corresponds to the mode assignment (charger-2 = trickle) with p = 0.39. This results in the set of assignments { (switch = charger-1), (charger-2 = trickle) } with p = 0.1911. Mini-ME would however recognize that this set of assignments is infeasible because of the 6th dissent that says that the two are infeasible.
Mini-ME would then choose another constituent diagnosis from the constituent diagnoses for dissent 5. The next most likely component mode assignment is charger-2 = off with p = 0.2. The combination of assignments results in a p = 0.098. This set of assignments does satisfy the current dissents for this observation. This results in Mini-ME extending this kernel diagnoses to a full diagnosis by choosing the most likely component mode for charger-1, which results in full-on with p = 0.49.
The mode estimate determined by Mini-ME is the most likely of all possible mode estimates since the search for it was guided by probabilities.
Mini-ME determines the most likely diagnosis using the dissents that pertain to the current observations. This diagnosis is guaranteed to be consistent with the observations because the set of conflicts are sufficient to generate all diagnoses, as shown by GDE and Sherlock. What remains is to develop the process of mode compilation to generate dissents offline.