Robust multiple-fault detection filter (Q4543804)

Consider the time-varying linear system \[ \dot{x}(t)= A(t) x(t) + B_u(t) u(t) + B_w(t) w(t) + \sum_{i=1}^q F_i(t) \mu_i(t), \] with the output equation \(y(t) = C(t)x(t) + v(t)\). Here, \(x\), \(u\), \(w\), \(y\), and \(v\) denote the state, control input, process noise, output, and sensor noise, respectively. The last term in the state equation denotes the faults which may occur in the system. The fault directions, \(F_i\), are assumed to be known, whereas the fault magnitudes, \(\mu_i\), are unknown. When there are no faults in the system, then the \(\mu_i\)'s are zero. The objective is to find a filter gain \(L\) for the linear observer \[ \dot{\widehat{x}}(t) = A(t) \widehat{x}(t) + B_u(t) u(t) + L(t)[y(t) - C(t) \widehat{x}(t)] \] and projectors \(\widehat{H}_i\) such that \(\widehat{H}_i(y-C\widehat{x})\) is only/primarily effected by \(\mu_i\). The authors solve this problem by formulating it as a LQG problem with indefinite cost. Hence they assume that \(v, w, \mu_i\), and \(x(t_0)\) are independently, normally distributed, random variables, and minimise the cost criterion \[ J = \frac{1}{t_1-t_0} \int_{t_0}^{t_1} {\text{ tr}}\left[ \sum_i \left[ \gamma^{-1}_i E[\widehat{h}_i(t)\widehat{h}_i(t)^T] + E[\bar{h}_i(t)\bar{h}_i(t)^T]- E[h_i(t)h_i(t)^T] \right] \right] dt. \] Here the last term represents the transition from \(\mu_i\) to \(\widehat{H}_iC\widehat{x}\), whereas the first terms represents the influence of the other terms on \(\widehat{H}_iC\widehat{x}\). Furthermore, the authors show that when all \(\gamma_i\)'s converge to zero, then their filter approximates the RDD filter from \textit{M. Massoumnia} [IEEE Trans. Autom. Control 31, 839-846 (1986; Zbl 0599.93017)].