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.

scalar-valued, we can write ˆr ⋆ k<br />

as the solution <strong>of</strong> the optimization<br />

ˆr ⋆ k = −min {<br />

min<br />

ρ ∈P (•)<br />

− ˆr k (ρ),<br />

min<br />

ρ ∈P (•)<br />

}<br />

ˆr k (ρ) .<br />

This problem is convex if P (•) is a convex set and both ˆr k (ρ) and − ˆr k (ρ) are convex functions<br />

<strong>of</strong> ρ (i.e., ˆr k (ρ) is affine in ρ). Once optimum values k ⋆ and ρ ⋆ have been obtained, the<br />

worst-case probability <strong>of</strong> false alarm is given by<br />

P ⋆ f = 1 − P( |r k ⋆(ρ ⋆ )| < ε k ⋆ | θ 0:k ⋆ = 0 0:k ⋆)<br />

.<br />

To summarize, the problem <strong>of</strong> computing Pf<br />

⋆<br />

and ˆr k is affine in ρ, for all k.<br />

is a convex optimization if P (•) is a convex set<br />

Minimizing the Probability <strong>of</strong> Detection<br />

Suppose that Assumptions 1–3 hold. Let ϑ be a fault parameter sequence such that ϑ N ≠ 0,<br />

and let k f be the fault time, as defined in equation (5.2). The worst-case probability <strong>of</strong><br />

detection is<br />

Pd ⋆ = 1 − max min P( )<br />

|r k (ρ)| < ε k | θ 0:k = ϑ 0:k .<br />

ρ ∈P (•) k f ≤k≤N<br />

By Proposition 5.2, optimum values <strong>of</strong> ρ and k are obtained by solving<br />

ˆµ ⋆ = min<br />

ρ ∈P (•)<br />

max<br />

k f ≤k≤N<br />

| ˆr k (ρ)|<br />

<br />

Σk<br />

.<br />

If we define the vector<br />

and the diagonal matrix<br />

⎡ ⎤<br />

ˆr k f<br />

(ρ)<br />

ˆr<br />

ˆR(ρ) :=<br />

k f +1(ρ)<br />

⎢ ⎥<br />

⎣ . ⎦<br />

ˆr N (ρ)<br />

{<br />

1<br />

W := diag ,<br />

Σ k f<br />

1<br />

Σ k f +1<br />

}<br />

1<br />

,..., , (5.6)<br />

Σ N<br />

then we may write<br />

ˆµ ⋆ = min<br />

ρ ∈P (•)<br />

∥ ∥W 1/2 ˆR(ρ) ∥ ∥<br />

∞<br />

.<br />

Since the matrix W is fixed, taking the ∞-norm is equivalent to taking the weighted pointwise<br />

maximum <strong>of</strong> ˆr k f<br />

(ρ),..., ˆr N (ρ). Because the pointwise maximum <strong>of</strong> convex functions is<br />

convex [5], computing P ⋆ d is a convex optimization if P (•) is convex and each ˆr k is a convex<br />

function <strong>of</strong> ρ, for k = k f ,..., N. Once an optimum value ρ ⋆ has been computed, let k ⋆ be<br />

85

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

Saved successfully!

Ooh no, something went wrong!