08.11.2014 Views

Probabilistic Performance Analysis of Fault Diagnosis Schemes

Probabilistic Performance Analysis of Fault Diagnosis Schemes

Probabilistic Performance Analysis of Fault Diagnosis Schemes

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

lating between the achieved performance values. (By Fact 3.7, the points (0,0) and (1,1)<br />

should also be included in this lower bound.) However, as mentioned above, the set V is<br />

too abstract to make this approach practical. Therefore, in the next section, we consider an<br />

extended definition <strong>of</strong> the roc that applies to more concrete sets <strong>of</strong> tests.<br />

Extending the ROC to Specific Families <strong>of</strong> Tests<br />

In Definition 3.9, the roc is defined with respect to the set <strong>of</strong> all tests, including randomized<br />

tests. This definition allowed us to characterize the roc in terms <strong>of</strong> likelihood ratio tests, via<br />

the Neyman–Pearson Lemma (Lemma 3.10), or in terms <strong>of</strong> Pareto optimality. In practice,<br />

however, we want to be able to evaluate the performance <strong>of</strong> a given test or a given family <strong>of</strong><br />

tests. For example, consider the parameterized family <strong>of</strong> fault detection schemes<br />

ˆ V = { V ε ∈ V : V ε = (F,δ ε ) and ε > 0 } , (3.18)<br />

where the residual generator F is fixed and δ ε is a threshold function defined as<br />

⎧<br />

⎨0, if |r | < ε,<br />

δ ε (r ) :=<br />

⎩1, otherwise.<br />

Clearly, V ε → V yes as ε → 0, regardless <strong>of</strong> the choice <strong>of</strong> F . Similarly, V ε → V no as ε → ∞.<br />

Hence, the set <strong>of</strong> achievable performance points is a curve passing through (0,0) and (1,1)<br />

(see Figure 3.4). Using randomization, as in Example 3.5, the tests in V ˆ can be used to<br />

achieve any performance point between this curve and the diagonal (i.e., any point in the<br />

convex hull <strong>of</strong> the curve). Hence, we have the following natural extension <strong>of</strong> the definition<br />

<strong>of</strong> the roc.<br />

W ˆ k ⊂ W k to be the set <strong>of</strong> performance<br />

points that are achieved by some test in V ˆ . The upper boundary <strong>of</strong> the set Wˆ<br />

k is<br />

Definition 3.11. Let ˆ V ⊂ V be some subset <strong>of</strong> tests. Define<br />

called the receiver operating characteristic (roc) for the class <strong>of</strong> tests ˆ V at time k.<br />

3.5 Certifying and Visualizing <strong>Performance</strong><br />

3.5.1 Bounds on <strong>Performance</strong> Metrics<br />

Given a fault detection scheme V , the system G θ is said to be available at time k if no fault<br />

has occurred and no false alarm has been issued. Hence, the probability <strong>of</strong> availability is<br />

given by the performance metric P tn,k . In a physical system affected by wear and deterioration,<br />

Q 1,k → 1 as k → ∞, so P tn,k → 0 as k → ∞. Therefore, any bound on P tn,k can only<br />

be enforced over a specified time window. Given N ∈ N and a > 0, one criterion for system<br />

35

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!