Subspace-based methods for eigenstructure identification and diagnosis


The problem is to identify, or diagnose changes in, the eigenstructure of a linear system of the form

X(n)  =  F X(n-1) + V(n)
Y(n)  =  H X(n-1) + W(n)
where V is the input excitation noise, W is the measurement noise, X is the state, and Y is the measurement. Only Y is observed. The eigenstructure consists of the collection of eigenvalues of F, and associated eigenvectors premultiplied by the observation matrix H.

Eigenstructure identification and diagnosis is, for instance, of interest in vibration mechanics , where this eigenstructre represents the modal information about the structure, and output only identification and monitoring is considered, i.e., the input excitation is natural and not measured.

Subspace algorithms for eigenstructure identification[BAB2000] use the covariance Hankel matrix H of the observation as a starting point:

Hij = cov( Y(n), YT(n-i-j) )
and uses the fact that H and the observability matrix of the pair (H,F)  possess identical left kernel space (hence the name of the technique). Since the left kernel of pair (H,F)  characterize the latter up to a change of basis in the state space, it is enough to compute this left kernel, and this can be performed by using the Hankel matrix H. More precisely, an abstract presentation of the method can be the following :
  1. estimate H based on available output samples ;
  2. compute a matrix S of maximal rank, satisfying STH= 0 ;
  3. find a pair (H,F)  such that its observability matrix O(H,F) satisfies also  STO(H,F) = 0.
Subspace algorithms for eigenstructure monitoring and diagnosis [BAB2000] are derived from the former one as follows :
  1. a nominal pair (H,F) is assumed, with its associated observability matrix O(H,F) ;
  2. compute a matrix S of maximal rank, satisfying  STO(H,F) = 0 ;
  3. estimate H based on current available output samples ;
  4. confront H and S by computing STH, and check whether or not STH = 0  holds.
Since the Hankel matrix H contains empirical covariance estimates, it is contaminated by noise and therefore the quantity STH  is a random variable and checking whether or not STH = 0  holds must be regarded as a statistical testing problem. We handle this by showing that, for a long sample, the quantity STH  is approximately Gaussian, with a known covariance (reflecting the uncertainty due to excitation and measurement noises), and finally it remains to test if this Gaussian quantity is zero mean or not.

Important properties of subspace algorithms are listed now :