A Robust Fault Detection Filter for Linear Time-Varying System with Non-Gaussian Noise

Zhemeng Zhang, Yifei Nie and Le Yin

arXiv: 2504.17648v2 · PDF · 1 September 2026
Categories: math.OC
Licence: CC BY 4.0 The site shows the text.

Abstract

This paper addresses the problem of robust fault detection filtering for linear time-varying (LTV) systems with non-Gaussian noise and additive faults. The conventional generalized likelihood ratio (GLR) method utilizes the Kalman filter, which may exhibit inadequate performance under non-Gaussian noise conditions. To mitigate this issue, a fault detection method employing the H_{\infty} filter is proposed. The H_{\infty} filter is first derived as the solution to a regularized least-squares (RLS) optimization problem, and the effect of faults on the output prediction error is then analyzed. The proposed approach using the H_{\infty} filter demonstrates robustness in non-Gaussian noise environments and significantly improves fault detection performance compared to the original GLR method that employs the Kalman filter. The effectiveness of the proposed approach is illustrated using numerical examples.

Items 6

Notation

F8Y6

\left( \,\cdot \,\right)^T and \left( \, \cdot \, \right)^{-1} are the transposition and inverse of a matrix, respectively. \mathbb{E}\left[ \,\cdot \,\right] represents the expectation operator. \mathbf{I} denotes the identity matrix with appropriate dimension. \mathbb{R}^n stands for the Euclidean space of n-dimensional. \left\{ x_0\,,\,\, x_1 \,, ... \,,\,\,x_k \right\} is represented as x_{0:k}. Given a vector x and a positive-definite matrix A \succ 0, {\lVert x \rVert}_A^2 = x^TAx. Given a cost function J(x), \hat{x} = \arg\min J(x) means that \hat{x} is the value at which J(x) attains its minimum.

Remark

Remark 1

MB4S

In practical applications, S_k is usually set as an identity matrix, which indicates that the same interest is assigned to the states in z_k. Note that blackthe parameter \alpha governs the trade-off between the robustness and sensitivity of the filter. When \alpha \rightarrow 0, the additive regularization term in cost function (?) vanishes, and the H_{\infty} filter reduces to the Kalman filter.

Remark

Remark 2

7TQ5

By assumption, \theta is a constant, and C_k and \varGamma_k are known at time step k. Hence, the covariance \varSigma_k^0 of the innovation \epsilon_k^0 in the fault-free system is equal to the covariance \varSigma_k in (?).

Proposition

Proposition 1

FFUJ

The GIR defined in (?) can be expressed in terms of E_k and d_k as follows

h_k = d_k^TE_k^{-1}d_k

Proof of Proposition 1

\begin{aligned} h_k &= \sum_{j = 1}^{k}\lVert \epsilon_j \rVert^2_{\varSigma_j^{-1}} - {\sum_{j = 1}^{k}\lVert \epsilon_j^0 \rVert^2_{\varSigma_j^{-1}}} \\ &= \sum_{j = 1}^{k}\epsilon_j^T\varSigma_j^{-1}\epsilon_j \\ &\quad - \sum_{j = 1}^{k} {\left( \epsilon_j - C_j\varGamma_j\hat{\theta}_k \right)}^T \varSigma_j^{-1} \left( \epsilon_j - C_j\varGamma_j \hat{\theta}_k \right) \\ &= \sum_{j = 1}^{k} \left( 2\epsilon_j^T\varSigma_j^{-1}C_j\varGamma_j\hat{\theta}_k - \hat{\theta}_k^T\varGamma_j^T C_j^T\varSigma_j^{-1}C_j\varGamma_j\hat{\theta}_k\right) \\ \end{aligned}

where \theta is assumed to be a constant fault vector in (?), E_k is a symmetric matrix. From (?), so the (Proof) can be rewritten

\begin{aligned} h_k &= 2\left( \sum_{j = 1}^{k} \epsilon_j^T\varSigma_j^{-1}C_j\varGamma_j \right)E_k^{-1}d_k \\ &\quad - d_k^TE_k^{-1} \left( \sum_{j = 1}^{k} \varGamma_j^T C_j^T\varSigma_j^{-1}C_j\varGamma_j \right) E_k^{-1}d_k \\ &= d_k^TE_k^{-1}d_k \end{aligned}

Remark

Remark 3

7W7I

It is important to note that if the H_{\infty} filter is replaced with the Kalman filter, the GIR method will degenerate into the conventional GLR method in [zhang2014].

Source and licence

The text on this page is from arXiv:2504.17648v2 by Zhemeng Zhang, Yifei Nie and Le Yin, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.

The record of each item is in analysis/objects/2504/2504.17648.jsonl.