betaFeedback

WorksMathematics

Sparse Fault Detection in Linear Systems: Equivalence of GLR Tests and ℓ1\ell_1-regularized Least Squares

Abstract

This paper studies the detection of an impulsive fault of unknown time and input channel in a linear system with unknown initial state and Gaussian measurement noise. First, we show that the generalized likelihood ratio test, after the initial state is eliminated by projection onto the parity space, is the zero-solution test of a weighted lasso with an unpenalized nuisance block; the regularization parameter is the test threshold, and for unknown noise variance the same holds for the square-root lasso. Second, we show that the null distribution depends on the system only through a Gram matrix, so that the threshold is computed exactly and offline, and we compare it with closed-form bounds. Third, we characterize the fraction of a fault that survives the projection: we give it in closed form through observability Gramians, we show that it converges to a geometrically decaying profile for strictly stable systems, and that, for systems with unit-modulus modes, it grows linearly with the fault time as a fraction of the record length. Numerical studies, including a two-input, two-output example, illustrate the results.

Introduction

Model-based fault and attack detection is a mature subject with renewed relevance in cyber-physical infrastructure, where an actuator or a sensor channel can be compromised while the plant state is not measured directly. We consider the following basic setup: a linear system is observed through noisy outputs over a finite window; the state at the beginning of the window is unknown; and at an unknown time, through an unknown input channel, the system receives a short impulsive fault. The task is to decide whether such a disturbance occurred, with a false-alarm probability that is guaranteed and does not depend on the unknown initial state.

Practical motivation. Faults and attacks in industrial cyber-physical systems, such as power systems and heavy machinery, often act for a short duration: a device malfunction, or a signal injected for a few samples. Such an event can occur at any time and through any actuator, and the plant state at the start of the record is unknown. The detector must decide from a finite record whether an event occurred, with a guaranteed false-alarm probability.

Theoretical interest. Detectors for disturbances or unknown inputs (in our context attacks and faults) are mostly chi-square detectors [1, 2], which treat the disturbance as an arbitrary vector. Unknown forces in practice are rarely arbitrary. Johnson Johnson [3] models them as a known waveform, such as a step, a ramp, a pulse or a sinusoid, whose coefficients jump at a few unknown times, that is, as the output of a known linear generator driven by sparse impulses. A fault is then a sparse combination of known signatures, one for each onset time and input channel, and the natural detector is a sparse GLR test. We study the simplest such structure, a single impulse, and ask what these detectors look like and what they gain over the chi-square detector.

Contributions. For the fault model discussed above, this paper considers generalized likelihood ratio (GLR) tests and makes the following contributions:

  1. Lasso form. The invariant GLR detects a fault exactly when a weighted lasso with an unpenalized nuisance block returns a nonzero solution, and the regularization parameter is the threshold. For unknown noise variance the same holds for the square-root lasso (Lemma 3).

  1. Exact calibration. The null distribution depends on the system only through the Gram matrix of the projected atoms, and for unknown noise variance neither on the initial state nor on the variance. Thus the threshold is computed exactly and offline (Theorem 1).

  1. Visibility profile. The fraction of a fault that survives the data projection has a closed form through observability Gramians (Theorem 2). It decays geometrically for strictly stable systems and equals (τ+1)/N(\tau+ 1)/N asymptotically for systems with unit-modulus modes (Theorem 3), where τ\tau is the time at which fault occurs.

Simulations, including a two-input, two-output example, confirm the predictions: the proposed tests detect far more often than the chi-square detector, and removing poorly visible atoms does not improve detection, while identifying the faulty channel is harder than detecting the fault. Appendix § 1 collects the proofs and Appendix II additional simulations.

Related literature. The unknown initial state is a nuisance parameter. In the analytical redundancy literature it is removed by projecting the outputs onto the parity space, the orthogonal complement of the range of the observability matrix [4, 5]; Fouillard and Nikiforov Fouillard and Nikiforov [6, 7] show that tests on the projected data are invariant to the initial state, and derive optimal ones for a fault at a known time. For an unknown time, the generalized likelihood ratio (GLR) of Willsky and Jones Willsky and Jones [8] maximizes over the onset, and the threshold of the resulting maximum is usually set by simulation. In the security literature, Pasqualetti, Dörfler and Bullo Pasqualetti, Dörfler and Bullo [9] characterize undetectable attacks as inputs whose output signature coincides with the free response of some initial state, and Fawzi, Tabuada and Diggavi Fawzi, Tabuada and Diggavi [10] use ℓ1\ell_1 decoding to handle sparse attacks. Independently, Fuchs Fuchs [11] observed that for a dictionary of signal atoms in Gaussian noise the ℓ2\ell_2–ℓ1\ell_1 criterion (the lasso) detects a signal exactly when the GLR test does, and that the regularization parameter is the threshold. Finally, selective inference [12, 13, 14] provides exact pp-values for the first variable entering a lasso path.

Problem Setup and Preliminaries

Notation. For a matrix MM, range(M)(M) is its range and ∥M∥\lVert M\rVert its spectral norm; for a symmetric matrix, λmin⁡\lambda_{\min} and λmax⁡\lambda_{\max} are its extreme eigenvalues and M⪯M′M \preceq M' is the Loewner order. εj\varepsilon_j is the jjth unit vector, 1N\mathbf{1}_N is the all-ones vector, and Q(c)=Pr⁡[z≥c]Q(c) = \Pr[z \ge c] is the right tail probability of the standard univariate normal N(0,1)\mathcal{N}(0,1). The null distribution of a statistic is its distribution under H0H_0.

System dynamics and attack model

Consider the standard discrete time linear system:

xk+1=Axk+Bukyk=Cxk+ek,\begin{aligned} x_{k+1} &= Ax_k + Bu_k \\ y_k &= Cx_k + e_k, \end{aligned}

where xk∈Rnx_k \in\mathbb{R}^n, uk∈Rmu_k \in\mathbb{R}^m, and yk∈Rpy_k \in\mathbb{R}^p. The initial state x0x_0 is unknown and the measurement errors ek∼N(0,σ2Ip)e_k \sim\mathcal{N}(0,\sigma^2 I_p) are i.i.d. The triple (A,B,C)(A,B,C) is known and (A,C)(A,C) observable. The nominal input is zero.

A fault or an attack is an impulse of unknown amplitude μ≠0\mu\ne0 that hits an unknown input channel j∈{1,…,m}j \in\{1,\ldots,m\} at an unknown time τ∈{0,…,N−2}\tau\in\{0,\ldots,N-2\}, that is,

uk={μεj,k=τ,0,k≠τ,u_k = \begin{cases} \mu\varepsilon_j, & k=\tau,\\ 0, & k\ne\tau, \end{cases}

where εj∈Rm\varepsilon_j \in\mathbb{R}^m is the jjth unit vector. We observe y0,…,yN−1y_0,\ldots,y_{N-1} with N≥nN \ge n, so the fault can occur at any τ∈{0,…,N−2}\tau\in\{0,\ldots,N-2\}. We call the pair a=(τ,j)a=(\tau,j) a fault index and write A\mathcal{A} for the set of all fault indices, so ∣A∣=m(N−1)\lvert\mathcal{A}\rvert=m(N-1).

Stack the outputs y=(y0,…,yN−1)∈RNpy=(y_0,\ldots,y_{N-1}) \in\mathbb{R}^{Np} and errors e=(e0,…,eN−1)∼N(0,σ2I)e=(e_0,\ldots,e_{N-1}) \sim\mathcal{N}(0,\sigma^2 I) to get the static linear model:

y=Ox0+μda+e,O=[CCA⋮CAN−1]∈RNp×n.(1)y = \mathcal{O}x_0 + \mu d_a + e,\qquad \mathcal{O} = \begin{bmatrix} C\\ CA\\ \vdots\\ CA^{N-1} \end{bmatrix} \in\mathbb{R}^{Np\times n}. \tag*{(1)}

Here O∈RNp×n\mathcal{O} \in\mathbb{R}^{Np\times n} is the observability matrix, which has full column rank because (A,C)(A,C) is observable and N≥nN \ge n. The vector da∈RNpd_a \in\mathbb{R}^{Np} is the output response to a unit impulse uτ=εju_\tau=\varepsilon_j, namely

(da)k={CAk−τ−1bj,k>τ,0,k≤τ,(d_a)_k = \begin{cases} CA^{k-\tau-1}b_j, & k>\tau,\\ 0, & k\le\tau, \end{cases}

where bjb_j is the jjth column of BB. We call dad_a the signature of the fault aa.

Problem statement

Since x0x_0 is unknown, Ox0\mathcal{O}x_0 can be any vector in range(O)(\mathcal{O}). A fault with signature dad_a close to range(O)(\mathcal{O}) can be explained, almost entirely, by a different x0x_0. This is the central difficulty of the problem, and the rest of the paper quantifies it. We want to decide whether a fault has occurred.

Problem 1 (Invariant fault detection). Given yy in (1), test

H0: μ=0,H_0:\ \mu=0,
H1: μ≠0 for some a∈A,H_1:\ \mu\ne0\text{ for some }a\in\mathcal{A},

without knowing x0x_0 and, in a variant, without knowing σ\sigma. Design a statistical test such that

(i) its false-alarm probability is at most a prescribed value α\alpha for every x0x_0 (and every σ\sigma in the variant), and

(ii) its threshold can be computed offline from (A,B,C,N)(A,B,C,N), without data.

Moreover, quantify for each a∈Aa\in\mathcal{A} the part of the fault that the x0x_0 hides. Requirements (i) and (ii) are addressed in Section IV, through the structure and calibration of the GLR test. The hidden part of a fault is discussed in Section V.

In Problem 1 “for some a∈Aa\in\mathcal{A}” means that, under H1H_1, exactly one of the ∣A∣=m(N−1)\lvert\mathcal{A}\rvert=m(N-1) faults, each a pair of input channel and fault time, has occurred. The alternative H1H_1 is therefore composite: it contains one model y=Ox0+μda+ey=\mathcal{O}x_0+\mu d_a+e for each index aa, and the amplitude μ≠0\mu\ne0 is unknown too.

Performance. A test δ\delta declares a fault when δ(y)=1\delta(y)=1. We measure it by its false-alarm probability PFA=Pr⁡[δ(y)=1∣(x0,σ2),H0]P_{\mathrm{FA}}=\Pr[\delta(y)=1\mid(x_0,\sigma^2),H_0] and probability of detection PD=Pr⁡[δ(y)=1∣(x0,σ2),H1]P_{\mathrm{D}}=\Pr[\delta(y)=1\mid(x_0,\sigma^2),H_1]. Here H1H_1 refers to the fault parameters (μ≠0,a∈A)(\mu\ne0,a\in\mathcal{A}). Requirement (i) of Problem 1 says that PFA≤αP_{\mathrm{FA}}\le\alpha for every x0x_0 (and σ\sigma).

Generalized Likelihood Ratio Tests

We consider GLR tests for known and for unknown noise variance σ\sigma. We first eliminate x0x_0 from the data.

Eliminating the initial state

Let P⊥P^\perp be the orthogonal projector onto the parity space range⁡(O)⊥\operatorname{range}(\mathcal{O})^\perp, which has dimension r=Np−nr=Np-n. Since P⊥O=0P^\perp\mathcal{O}=0, projecting yy (1) eliminates x0x_0:

P⊥y=μP⊥da+P⊥e,(2)P^\perp y=\mu P^\perp d_a+P^\perp e, \tag*{(2)}

where P⊥e∼N(0,σ2P⊥)P^\perp e\sim\mathcal{N}(0,\sigma^2P^\perp). Moreover, P⊥yP^\perp y is a maximal invariant1 for the shifts y↦y+Ovy\mapsto y+\mathcal{O}v, v∈Rnv\in\mathbb{R}^n [6].

Only the part of a fault signature that no initial state x0x_0 can imitate remains in the projected data P⊥yP^\perp y. For each a∈Aa\in\mathcal{A} we define its visible length and visible fraction,

wa=∥P⊥da∥andρa2=∥P⊥da∥2∥da∥2∈[0,1],(3)w_a=\lVert P^\perp d_a\rVert\quad\text{and}\quad \rho_a^2=\frac{\lVert P^\perp d_a\rVert^2}{\lVert d_a\rVert^2}\in[0,1], \tag*{(3)}

and call a↦ρa2a \mapsto\rho_a^2 the visibility profile. If wa=0w_a = 0, then da∈range⁡(O)d_a \in\operatorname{range}(\mathcal{O}) and hence the fault is indistinguishable from x0x_0. We exclude this case in Assumption 1.

For wa>0w_a > 0, the projected atom za=P⊥da/waz_a = P^\perp d_a/w_a is a unit vector, and the projected data

P⊥y=(μwa)za+P⊥e.(4)P^\perp y = (\mu w_a)z_a + P^\perp e. \tag*{(4)}

Thus, compared with the model with x0x_0 known, the overall fault magnitude is scaled down from ∥da∥\lVert d_a\rVert to wa=ρa∥da∥w_a = \rho_a\lVert d_a\rVert. So the problem becomes detecting an unknown multiple of one atom zaz_a from the dictionary {za}a∈A\{z_a\}_{a\in\mathcal{A}}, with aa unknown.

Lemma 1 (Geometry of visibility). Let θa∈[0,π/2]\theta_a \in[0,\pi/2] be the angle between dad_a and the subspace range⁡(O)\operatorname{range}(\mathcal{O}), that is, the smallest angle between dad_a and a nonzero vector of range⁡(O)\operatorname{range}(\mathcal{O}). Then ρa=sin⁡θa\rho_a = \sin\theta_a.

Proof. Split da=p+P⊥dad_a = p+P^\perp d_a with p=(I−P⊥)da∈range⁡(O)p=(I-P^\perp)d_a \in\operatorname{range}(\mathcal{O}) and P⊥da⊥range⁡(O)P^\perp d_a \perp\operatorname{range}(\mathcal{O}). For every s∈range⁡(O)s \in\operatorname{range}(\mathcal{O}) we have ⟨da,s⟩=⟨p,s⟩\langle d_a,s\rangle= \langle p,s\rangle, so by the Cauchy–Schwarz inequality cos⁡∠(da,s)≤∥p∥/∥da∥\cos\angle(d_a,s) \le\lVert p\rVert/\lVert d_a\rVert, with equality when s=ps=p. Hence cos⁡θa=∥p∥/∥da∥\cos\theta_a = \lVert p\rVert/\lVert d_a\rVert. Since pp and P⊥daP^\perp d_a are orthogonal and da=p+P⊥dad_a=p+P^\perp d_a, the Pythagorean identity gives

∥P⊥da∥2=∥da∥2−∥p∥2=∥da∥2(1−cos⁡2θa)=∥da∥2sin⁡2θa.\begin{aligned} \lVert P^\perp d_a\rVert^2 &= \lVert d_a\rVert^2-\lVert p\rVert^2 \\ &=\lVert d_a\rVert^2(1-\cos^2\theta_a)=\lVert d_a\rVert^2\sin^2\theta_a. \end{aligned}

Dividing by ∥da∥2\lVert d_a\rVert^2 yields ρa2=sin⁡2θa\rho_a^2=\sin^2\theta_a, and since ρa≥0\rho_a\ge0 and sin⁡θa≥0\sin\theta_a\ge0 for θa∈[0,π/2]\theta_a\in[0,\pi/2], we conclude ρa=sin⁡θa\rho_a=\sin\theta_a. □

Figure 1(a) illustrates the lemma. The unknown x0x_0 moves the data yy only along range⁡(O)\operatorname{range}(\mathcal{O}), so it can mimic the part of a fault da∈range⁡(O)d_a\in\operatorname{range}(\mathcal{O}). The visible part is the perpendicular component, whose length relative to ∥da∥\lVert d_a\rVert is sin⁡θa\sin\theta_a. Two corner cases are worth studying. First, if da∈range⁡(O)d_a\in\operatorname{range}(\mathcal{O}), a change in x0x_0 produces exactly the same data, and ρa=0\rho_a=0. Second, if da∈range⁡(O)⊥d_a\in\operatorname{range}(\mathcal{O})^\perp, no x0x_0 can mimic it, and ρa=1\rho_a=1. The early fault da′d_{a'} shown in Figure 1 is close to the former, and the late fault dad_a is closer to the latter.

Figure 1. Geometry of the detection problem. (a) In RNp\mathbb{R}^{Np} the unknown initial state moves the data freely along range⁡(O)\operatorname{range}(\mathcal{O}). Of a fault atom dad_a only the component orthogonal to range⁡(O)\operatorname{range}(\mathcal{O}) is visible; its length is wa=ρa∥da∥w_a=\rho_a\lVert d_a\rVert with ρa=sin⁡θa\rho_a=\sin\theta_a. (b) In the parity space the normalized visible components are unit vectors ±za\pm z_a. The shaded region is the set of projected data for which the lasso of Section IV returns zero, that is, the no-alarm region of the GLR test with threshold hh.

Assumption 1 (Distinguishable faults). For every a∈Aa\in\mathcal{A}, we impose (i) wa>0w_a>0; and (ii) for a≠a′a\ne a', ∣zaTza′∣<1\lvert z_a^\mathsf{T}z_{a'}\rvert<1.

Both conditions can be checked from the matrices (A,B,C)(A,B,C) and do not rely on the fault parameters (μ,τ)(\mu,\tau).

Condition (i) states that no fault can be explained by x0x_0 alone. Thus, da∉range⁡(O)d_a\notin\operatorname{range}(\mathcal{O}), that is, θa>0\theta_a>0 in Figure 1(a). It is the finite-horizon, single-impulse version of an undetectable input in [9], and if (i) fails, the data yy carry no information about that fault. For τ≥n−1\tau\ge n-1 it is equivalent to da≠0d_a\ne0: if Ox=da\mathcal{O}x=d_a, then CAkx=0CA^kx=0 for k≤τk\le\tau, so x=0x=0 by observability and hence da=0d_a=0. Note that wa>0w_a>0 does not make a fault easy to detect. A small ρa\rho_a means that only a small fraction of the fault is visible, and § 5 quantifies this.

Condition (ii) states that no two projected atoms are parallel after projection. Two faults with the same signature cannot be distinguished, and would be merged into one atom; this happens, for instance, if b2b_2 is a multiple of b1b_1.

The GLR test

Under H1H_1, with fault index aa and amplitude μ≠0\mu\ne0, the density of yy is p(y∣x0,μ,σ2,H1)=N(x0+μda,σ2I)p(y\mid x_0,\mu,\sigma^2,H_1)=\mathcal{N}(x_0+\mu d_a,\sigma^2I), and under H0H_0 the density p(y∣x0,σ2,H0)=N(x0,σ2I)p(y\mid x_0,\sigma^2,H_0)=\mathcal{N}(x_0,\sigma^2I). Since H1H_1 is composite and x0x_0 is unknown, we resort to the generalized likelihood ratio (GLR) statistic, in logarithmic form:

ℓ(y)=log⁡max⁡a∈Asup⁡x0,μ,σ2p(y;x0,μ,a,σ2)sup⁡x0,σ2p(y;x0,0,σ2).(5)\ell(y)=\log\frac{\max_{a\in\mathcal{A}}\sup_{x_0,\mu,\sigma^2}p(y;x_0,\mu,a,\sigma^2)}{\sup_{x_0,\sigma^2}p(y;x_0,0,\sigma^2)}. \tag*{(5)}

where σ2\sigma^2 is maximized only in the unknown-variance case, and the supremum over μ≠0\mu\ne0 equals the supremum over μ∈R\mu\in\mathbb{R} because pp is continuous in μ\mu. The GLR test is

δ(y)={1 (alarm),ℓ(y)≥η,0,ℓ(y)<η.(6)\delta(y)= \begin{cases} 1\ \text{(alarm)}, & \ell(y)\ge\eta,\\ 0, & \ell(y)<\eta. \end{cases} \tag*{(6)}

Here η≥0\eta\ge0 is such that PFA=Pr⁡[ℓ(y)≥η∣(x0,σ),H0]≤αP_{\mathrm{FA}}=\Pr[\ell(y)\ge\eta\mid(x_0,\sigma),H_0]\le\alpha. Because H1H_1 is composite, optimal tests usually do not exist, and the GLR test is sub-optimal. To evaluate ℓ(y)\ell(y) we need the residuals of the least-squares fits under each hypothesis.

Lemma 2 (Residuals). For each fault index a∈Aa\in\mathcal{A},

min⁡x0,μ∥y−Ox0−μda∥2=∥P⊥y∥2−(zaTy)2,min⁡x0∥y−Ox0∥2=∥P⊥y∥2.\begin{aligned} \min_{x_0,\mu}\lVert y-\mathcal{O}x_0-\mu d_a\rVert^2&=\lVert P^\perp y\rVert^2-(z_a^\mathsf{T}y)^2,\\ \min_{x_0}\lVert y-\mathcal{O}x_0\rVert^2&=\lVert P^\perp y\rVert^2. \end{aligned}

Proof. Consider the first minimization problem. For a fixed μ\mu, least-squares minimization over x0x_0 gives

min⁡x0∥y−Ox0−μda∥2=∥P⊥(y−μda)∥2=∥P⊥y−μwaza∥2,\min_{x_0}\lVert y-\mathcal{O}x_0-\mu d_a\rVert^2=\lVert P^\perp(y-\mu d_a)\rVert^2=\lVert P^\perp y-\mu w_a z_a\rVert^2,

where the final equality follows from (4). Because zaz_a is a unit vector and wa>0w_a>0, the amplitude μ⋆=(zaTP⊥y)/wa\mu^\star=(z_a^\mathsf{T}P^\perp y)/w_a minimizes ∥P⊥y−μwaza∥2\lVert P^\perp y-\mu w_a z_a\rVert^2. From these results we get

min⁡x0,μ∥y−Ox0−μda∥2=∥P⊥y∥2−(zaTP⊥y)2.\min_{x_0,\mu}\lVert y-\mathcal{O}x_0-\mu d_a\rVert^2=\lVert P^\perp y\rVert^2-(z_a^\mathsf{T}P^\perp y)^2.

Because za∈range⁡(O)⊥z_a\in\operatorname{range}(\mathcal{O})^\perp, we have zaTP⊥y=zaTyz_a^\mathsf{T}P^\perp y=z_a^\mathsf{T}y, proving the first identity. For the fault-free model, setting μ=0\mu=0 in the least-squares problem above gives the second identity. □

We evaluate (5). Let ℓproj(y)\ell_{\mathrm{proj}}(y) be defined as in (5) with the projected data P⊥yP^\perp y in place of yy. The densities of P⊥yP^\perp y under H0H_0 and H1H_1 are Gaussian and do not depend on x0x_0. The next proposition evaluates ℓproj\ell_{\mathrm{proj}} and ℓ\ell, and shows that projecting the data loses no information for the GLR test.

Proposition 1 (The GLR test). Let Assumption 1(i) hold, and let r=Np−n≥2r=Np-n\ge2. Define the statistics

T=max⁡a∈A(zaTy)2σ2,U=max⁡a∈A∣zaTy∣∥P⊥y∥∈[0,1].(7)T=\max_{a\in\mathcal{A}}\frac{(z_a^\mathsf{T}y)^2}{\sigma^2},\qquad U=\max_{a\in\mathcal{A}}\frac{\lvert z_a^\mathsf{T}y\rvert}{\lVert P^\perp y\rVert}\in[0,1]. \tag*{(7)}

(i) (Known σ\sigma): ℓ=ℓproj=T/2\ell=\ell_{\mathrm{proj}}=T/2.

(ii) (Unknown σ\sigma): ℓproj=−r2log⁡(1−U2)\ell_{\mathrm{proj}}=-\frac{r}{2}\log(1-U^2) and ℓ=Nprℓproj\ell=\frac{Np}{r}\ell_{\mathrm{proj}}.

By Proposition 1, ℓproj(y)\ell_{\mathrm{proj}}(y) is an increasing function of TT when σ\sigma is known and of UU when σ\sigma is unknown. Hence the GLR test of the projected data, ℓproj(y)≥η\ell_{\mathrm{proj}}(y)\geq\eta, is

δ(P⊥y)={1{T≥γ},σ known, γ=2η,1{U≥u},σ unknown, u=1−e−2η/r.(8)\delta(P^{\perp}y)= \begin{cases} \mathbb{1}\{T\geq\gamma\}, & \sigma\ \text{known},\ \gamma=2\eta,\\ \mathbb{1}\{U\geq u\}, & \sigma\ \text{unknown},\ u=\sqrt{1-e^{-2\eta/r}}. \end{cases} \tag*{(8)}

Proposition 1 shows that the log-likelihood ratios of yy and of P⊥yP^{\perp}y are the same for known σ\sigma, and for unknown σ\sigma they are equal up to the scaling factor Np/rNp/r. Hence the GLR test statistics are the same TT and UU for the unprojected data, and the only change is in the thresholds determined by η\eta in (6). In practice we choose γ\gamma and uu directly as the smallest values with Pr⁡[T≥γ]≤α\Pr[T\geq\gamma]\leq\alpha and Pr⁡[U≥u]≤α\Pr[U\geq u]\leq\alpha under H0H_0 (Section IV). See Remark 1 for how UU relates to TT.

Example 1 (Integrator). Let A=B=C=1A=B=C=1, so O=1N\mathcal{O}=\mathbf{1}_N and dτ=(0,…,0,1,…,1)Td_\tau=(0,\ldots,0,1,\ldots,1)^{\mathsf T} with q=N−τ−1q=N-\tau-1 ones. Set t=τ+1t=\tau+1. Then P⊥=I−1N1NT/NP^{\perp}=I-\mathbf{1}_N\mathbf{1}_N^{\mathsf T}/N, wτ2=∥dτ∥2−q2/N=tq/Nw_\tau^2=\lVert d_\tau\rVert^2-q^2/N=tq/N, hence ρτ2=t/N\rho_\tau^2=t/N, and

zτTy=tqN(yˉ>τ−yˉ≤τ),z_\tau^{\mathsf T}y=\sqrt{\frac{tq}{N}}(\bar{y}_{>\tau}-\bar{y}_{\leq\tau}),

where yˉ>τ\bar{y}_{>\tau} and yˉ≤τ\bar{y}_{\leq\tau} are the sample means after and up to time τ\tau. Under H0H_0, zτTy/σz_\tau^{\mathsf T}y/\sigma is the standardized difference of two sample means, so TT is the classical scan statistic for a change in mean [5]. An early fault is poorly visible because only tt samples remain to estimate the unknown pre-fault level.

Remark 1 (The two statistics). Under H0H_0, σ^2=∥P⊥y∥2/r\hat{\sigma}^2=\lVert P^{\perp}y\rVert^2/r is an unbiased estimate of σ2\sigma^2, and rU2=max⁡a(zaTy)2/σ^2rU^2=\max_a(z_a^{\mathsf T}y)^2/\hat{\sigma}^2. So rU2rU^2 is TT with σ\sigma replaced by σ^\hat{\sigma}, and UU is the cosine of the angle between P⊥yP^{\perp}y and the closest atom ±za\pm z_a. It is invariant to scaling of yy, so its null distribution does not depend on σ\sigma (Theorem 1).

Remark 2 (Comparison with χr2\chi_r^2 fault/attack detectors). In the well-known χr2\chi_r^2 attack detector for cyber-physical systems [1, 2], the test statistic is ∥P⊥y∥2/σ2\lVert P^{\perp}y\rVert^2/\sigma^2. This is the value of 2ℓ2\ell in (5) if the fault/attack is allowed to be any vector of the parity space range⁡(O)⊥\operatorname{range}(\mathcal{O})^{\perp} instead of an impulse fault (our case). Here r=Np−nr=Np-n is the number of degrees of freedom, the dimension of the parity space range⁡(O)⊥\operatorname{range}(\mathcal{O})^{\perp}. Our fault/attack model is an impulse, a single atom of that space. The χr2\chi_r^2 detector therefore spreads an impulse over all rr dimensions and cannot return (τ,j)(\tau,j), whereas TT matches one atom and returns the maximizing pair. With an arbitrary mean, σ\sigma is not identifiable from P⊥yP^{\perp}y, so the χr2\chi_r^2 detector needs σ\sigma given, whereas the impulse structure lets UU handle unknown σ\sigma. Detectors built on Kalman innovations assume a prior on x0x_0, and the false-alarm probability is guaranteed only when x0x_0 follows it; ours holds for every x0x_0.

Structure and Calibration of the Test

The test statistics TT and UU depend on the model (1) only through the unit atoms zaz_a, so the results in this section hold for any collection of atoms, that is, a finite frame [15]. The system dynamics enters only in Section V. We show that the correlations among the atoms, collected in the Gram matrix G=[zaTza′]G=[z_a^{\mathsf T}z_{a'}], determine the geometry of the test as a lasso, the null distribution of its statistics, and hence the thresholds.

Lasso form

Let D=[da]a∈A∈RNp×∣A∣D=[d_a]_{a\in\mathcal{A}}\in\mathbb{R}^{Np\times|\mathcal{A}|} be the matrix whose columns are the fault signatures in (1). For h,λ>0h,\lambda>0, consider the lasso problems

(P1) argmin⁡x0,θ 12∥y−Ox0−Dθ∥22+h∑awa∣θa∣,\underset{x_0,\theta}{\operatorname{argmin}}\ \frac{1}{2}\lVert y-\mathcal{O}x_0-D\theta\rVert_2^2+h\sum_a w_a|\theta_a|, (9)

(P2) argmin⁡x0,θ ∥y−Ox0−Dθ∥2+λ∑awa∣θa∣.\underset{x_0,\theta}{\operatorname{argmin}}\ \lVert y-\mathcal{O}x_0-D\theta\rVert_2+\lambda\sum_a w_a|\theta_a|. (10)

Let (x^0,θ^)(\hat{x}_0,\hat{\theta}) denote a solution of (P1) or (P2). Here Dθ=∑aθadaD\theta=\sum_a\theta_a d_a, so θa\theta_a is the weight on the fault dad_a. The variable x0x_0 is not penalized, and waw_a is the visible length in Figure 1(a). Our model (1), with a single fault aa of amplitude μ\mu, is the special case θ=μea\theta=\mu e_a, where eae_a is the aa-th standard basis vector. A solution with θ^≠0\hat{\theta} \ne0 declares a fault. Lemma 3 below shows that this reproduces the GLR tests of Proposition 1.

Lemma 3 (GLR test in (8) as a lasso). Under Assumption 1, θ^=0\hat{\theta}=0 is a solution of (P1) if and only if h≥max⁡a∣za⊤y∣h \ge\max_a |z_a^\top y|, that is, T≤h2/σ2T \le h^2/\sigma^2. If P⊥y≠0P^\perp y \ne0, θ^=0\hat{\theta}=0 is a solution of (P2) if and only if λ≥U\lambda\ge U.

Hence the GLR tests in (8) reject H0H_0 exactly when θ^≠0\hat{\theta} \ne0, with h=σγh=\sigma\sqrt{\gamma} in (P1) and λ=u\lambda=u in (P2). Geometrically (see Figure 1(b)), since za⊤y=za⊤P⊥yz_a^\top y=z_a^\top P^\perp y, the test declares no fault exactly when P⊥yP^\perp y lies in the set {v∈range⁡(O)⊥:∣za⊤v∣≤h ∀a}\{v\in\operatorname{range}(\mathcal{O})^\perp:|z_a^\top v|\le h\ \forall a\}, which is a polytope when the atoms span range⁡(O)⊥\operatorname{range}(\mathcal{O})^\perp.

Remark 3 (Insight and limits). The lemma gives the penalty a false-alarm meaning: with h=σγh=\sigma\sqrt{\gamma} or λ=u\lambda=u calibrated as in Section IV-B, the lasso has false-alarm probability at most α\alpha for every x0x_0, in place of cross-validation. The free block x0x_0 gives invariance to the initial state, and the weights waw_a treat all atoms alike: with unit weights, θ^=0\hat{\theta}=0 is a solution of (P1) if and only if h≥max⁡awa∣za⊤y∣h\ge\max_a w_a|z_a^\top y|, so the rule thresholds ∣za⊤y∣|z_a^\top y| at the level h/wah/w_a. The lemma does not make the test cheaper, since the maximum is one pass over the atoms whereas a lasso solver iterates. The gain is that the same problem also returns θ^\hat{\theta}, and that standard lasso solvers and theory apply.

Exact threshold from the Gram matrix

We show that under H0H_0 the laws of TT and UU are determined by the Gram matrix G=[za⊤za′]a,a′∈AG=[z_a^\top z_{a'}]_{a,a'\in\mathcal{A}} of the atoms, and that the law of UU does not depend on x0x_0 or σ\sigma, so the thresholds of the GLR tests can be computed once offline.

Theorem 1 (Null distribution). Let V∈RNp×rV\in\mathbb{R}^{Np\times r} be an orthonormal basis of range⁡(O)⊥\operatorname{range}(\mathcal{O})^\perp and define the vector za=V⊤ζaz_a=V^\top\zeta_a. Let g∼N(0,Ir)g\sim\mathcal{N}(0,I_r). Under H0H_0, for every x0x_0 and σ\sigma,

T=dmax⁡a(ζa⊤g)2,U=dmax⁡a∣ζa⊤v∣,T \mathrel{\overset{d}{=} } \max_a(\zeta_a^\top g)^2,\qquad U \mathrel{\overset{d}{=} } \max_a|\zeta_a^\top v|,

where vv is uniform on the unit sphere of Rr\mathbb{R}^r.

Proof. Under the hypothesis H0H_0, P⊥y=P⊥eP^\perp y=P^\perp e because P⊥O=0P^\perp\mathcal{O}=0. Put g=V⊤e/σ∼N(0,Ir)g=V^\top e/\sigma\sim\mathcal{N}(0,I_r). Then za⊤y=za⊤P⊥e=σζa⊤gz_a^\top y=z_a^\top P^\perp e=\sigma\zeta_a^\top g, which gives the statement for TT. Also ∥P⊥y∥=σ∥g∥\|P^\perp y\|=\sigma\|g\|, so U=max⁡a∣ζa⊤g∣/∥g∥U=\max_a|\zeta_a^\top g|/\|g\|. A standard Gaussian is rotation invariant, so v=g/∥g∥v=g/\|g\| is uniform on the sphere.

By Theorem 1, the threshold uu with Pr⁡[U≥u]≤α\Pr[U\ge u]\le\alpha under H0H_0 gives false-alarm probability at most α\alpha for all x0x_0 and σ\sigma, and so does the threshold γ\gamma of TT for known σ\sigma. For a Gaussian mean test or a χ2\chi^2 test, such a threshold is the inverse of the Q(⋅)Q(\cdot) function or of the χ2\chi^2 cdf. Here TT and UU are maxima over correlated variables whose cdf has no closed form, so we compute the thresholds offline by sampling (Algorithm 1), up to Monte Carlo error.

Remark 4 (Closed-form test thresholds). Sampling can be avoided with closed-form thresholds cc with the false alarm probability Pr⁡[max⁡a∣za⊤y∣/σ≥c]≤α\Pr[\max_a|z_a^\top y|/\sigma\ge c]\le\alpha under H0H_0, which give γ=c2\gamma=c^2. This test declares a fault when at least one of the ∣A∣|\mathcal{A}| single-atom tests ∣za⊤y∣/σ≥c|z_a^\top y|/\sigma\ge c fires, and each single-atom test has false-alarm probability 2Q(c)2Q(c). Bonferroni correction

Algorithm 1 Offline thresholds γ\gamma and uu

Input: atoms ζa\zeta_a, a∈Aa\in\mathcal{A}; level α\alpha; number of samples MM

  1. for k=1,…,Mk=1,\ldots,M do

  2. draw gk∼N(0,Ir)g_k\sim\mathcal{N}(0,I_r)

  3. Tk←max⁡a(ζa⊤gk)2,Uk←max⁡a∣ζa⊤gk∣/∥gk∥T_k\leftarrow\max_a(\zeta_a^\top g_k)^2,\quad U_k\leftarrow\max_a|\zeta_a^\top g_k|/\|g_k\|

  4. end for

  5. γ←\gamma\leftarrow value above which a fraction α\alpha of T1,…,TMT_1,\ldots,T_M lie

  6. u←u\leftarrow value above which a fraction α\alpha of U1,…,UMU_1,\ldots,U_M lie

Output: γ\gamma for TT (known σ\sigma) and uu for UU

[16] adds these probabilities over the atoms and solves for 2∣A∣Q(c)=α2|\mathcal{A}|Q(c)=\alpha. The Šidák bound in [17], the independence approximation used by Fuchs [11], multiplies them as if the atoms were independent and solves for 1−(1−2Q(c))∣A∣=α1-(1-2Q(c))^{|\mathcal{A}|}=\alpha. Finally, the Rice-type tube bound [18] uses the correlation between neighboring atoms for one input channel and solves 2Q(c)+(κ0/π)e−c2/2=α2Q(c)+(\kappa_0/\pi)e^{-c^2/2}=\alpha, where κ0=∑aarccos⁡(za⊤za+1)\kappa_0=\sum_a\operatorname{arccos}(z_a^\top z_{a+1}) is the length of the polygonal curve through consecutive atoms on the sphere. All three are conservative when the atoms are correlated. An exact alternative that needs no sampling, the spacing test, is described in Appendix I-E.

Probability of Detection

The thresholds of Section IV fix the false-alarm probability. We now ask how often the test detects a fault. We show that the unknown initial state interferes with the fault, and that this scales the signal-to-noise ratio in the probability of detection (Section V-A). We then give the visible fraction in closed form through observability Gramians, and show that for long records it is a fixed-length window for stable systems and a fixed fraction of the record for systems with unit-modulus modes (Section V-B).

Interference from the initial state

Projecting out the unknown initial state x0x_0 loses nothing for the GLR test (Proposition 1), but x0x_0 interferes with the fault: the component of dad_a in range⁡(O)\operatorname{range}(\mathcal{O}) is indistinguishable from Ox0\mathcal{O}x_0 and is removed, so by (2) only P⊥daP^\perp d_a remains. A fault of amplitude μ\mu has visible length μwa=μρa∥da∥\mu w_a=\mu\rho_a\|d_a\|, so the visible fraction ρa\rho_a of Section III is the fraction of the fault that survives, a geometric quantity by Lemma 1. The next proposition states the consequence for the probability of detection PD(a,μ)P_D(a,\mu), defined in Section II-B.

Proposition 2 (Interference and probability of detection). Let σ\sigma be known and let δ(P⊥y)=1{max⁡a∣za⊤y∣/σ≥c}\delta(P^\perp y)=\mathbf{1}\{\max_a|z_a^\top y|/\sigma\ge c\}. For every x0x_0, every fault aa and every amplitude μ\mu,

PD(a,μ)≥Q(c−sa)+Q(c+sa).P_D(a,\mu)\ge Q(c-s_a)+Q(c+s_a).

where Q(⋅)Q(\cdot) is right tail probability of a standard Gaussian and sa=(μwa)/σ=ρa(μ∥da∥)/σs_a=(\mu w_a)/\sigma=\rho_a(\mu\|d_a\|)/\sigma.

Proof. Since za∈range⁡(O)⊥z_a\in\operatorname{range}(\mathcal{O})^\perp and P⊥da=wazaP^\perp d_a=w_a z_a, za⊤y=za⊤P⊥y=μwa+za⊤ez_a^\top y=z_a^\top P^\perp y=\mu w_a+z_a^\top e, which does not depend on x0x_0. Hence za⊤y/σ∼N(sa,1)z_a^\top y/\sigma\sim\mathcal{N}(s_a,1). The event {max⁡a′∣za′⊤y∣/σ≥c}\{\max_{a'}|z_{a'}^\top y|/\sigma\ge c\} contains {∣za⊤y∣/σ≥c}\{|z_a^\top y|/\sigma\ge c\}, whose probability is Q(c−sa)+Q(c+sa)Q(c-s_a)+Q(c+s_a). ∎

The raw signal-to-noise ratio μ∥da∥/σ\mu\lVert d_a\rVert/\sigma is the length of the fault in units of the noise standard deviation, which is what the test would see if x0x_0 were known. The interference scales it by the visible fraction to sa=ρaμ∥da∥/σs_a=\rho_a\mu\lVert d_a\rVert/\sigma, so the bound decreases as ρa\rho_a decreases: a poorly visible fault is detected less often, and its amplitude must be 1/ρa1/\rho_a times larger to reach the same bound. The bound is conservative, because atoms other than aa can also raise the alarm.

Visibility profile

We now compute ρa2\rho_a^2 from (A,B,C)(A,B,C). Let Wi=∑k=0i−1(Ak)⊤C⊤CAkW_i=\sum_{k=0}^{i-1}(A^k)^\top C^\top CA^k be the observability Gramian over ii samples; then O⊤O=WN≻0O^\top O=W_N\succ0, and Wi⪯Wi′W_i\preceq W_{i'} for i≤i′i\le i', since each summand is positive semidefinite.

Theorem 2 (Visibility profile). For a=(τ,j)a=(\tau,j) let t=τ+1t=\tau+1 and q=N−tq=N-t. Then

1−ρa2=bj⊤WqAtWN−1(At)⊤Wqbjbj⊤Wqbj.(9)1-\rho_a^2=\frac{b_j^\top W_qA^tW_N^{-1}(A^t)^\top W_qb_j}{b_j^\top W_qb_j}. \tag*{(9)}

For n=1n=1 and A=a0≠0A=a_0\ne0 the formula (11) reduces to ρτ2=∑k=02τa02k/∑k=0N−1a02k\rho_\tau^2=\sum_{k=0}^{2\tau}a_0^{2k}/\sum_{k=0}^{N-1}a_0^{2k}, which equals (τ+1)/N(\tau+1)/N for the integrator of Example 1.

The next result describes the profile for long records. Call an atom poorly visible if ρa2<ρ02\rho_a^2<\rho_0^2 for a fixed ρ0∈(0,1)\rho_0\in(0,1).

Theorem 3 (Visibility for long records). Let AA be Schur stable, ∥At∥≤Mλt\lVert A^t\rVert\le M\lambda^t with λ<1\lambda<1, and let W∞W_\infty solve W∞=A⊤W∞A+C⊤CW_\infty=A^\top W_\infty A+C^\top C. Then: (i) for each fixed tt and jj, lim⁡N→∞(1−ρ(t−1,j)2)=γt,j:=bj⊤W∞AtW∞−1(At)⊤W∞bjbj⊤W∞bj\lim_{N\to\infty}(1-\rho_{(t-1,j)}^2)=\gamma_{t,j}:=\frac{b_j^\top W_\infty A^tW_\infty^{-1}(A^t)^\top W_\infty b_j}{b_j^\top W_\infty b_j}; (ii) for every N≥nN\ge n and every atom, 1−ρa2≤κ∞M2λ2(τ+1)1-\rho_a^2\le\kappa_\infty M^2\lambda^{2(\tau+1)}, with κ∞=λmax⁡(W∞)/λmin⁡(Wn)\kappa_\infty=\lambda_{\max}(W_\infty)/\lambda_{\min}(W_n); (iii) the number of poorly visible atoms per input channel is at most ⌈ln⁡(κ∞M2/(1−ρ02))2ln⁡(1/λ)⌉\left\lceil\frac{\ln(\kappa_\infty M^2/(1-\rho_0^2))}{2\ln(1/\lambda)}\right\rceil, whatever NN is.

Now let AA be diagonalizable with all eigenvalues on the unit circle, (A,C)(A,C) observable, and bj≠0b_j\ne0 for all jj. Then: (iv) Wˉ=lim⁡i→∞Wi/i\bar W=\lim_{i\to\infty}W_i/i exists and is positive definite, and

max⁡τ∣ρ(τ,j)2−τ+1N∣⟶0(N→∞).(10)\max_\tau\left|\rho_{(\tau,j)}^2-\frac{\tau+1}{N}\right|\longrightarrow0\qquad(N\to\infty). \tag*{(10)}

In particular, the number of poorly visible atoms per channel is ρ02N+o(N)\rho_0^2N+o(N) for every system in this class, whatever (A,B,C)(A,B,C) is.

Remark 5 (Reading Theorem 3). A stable plant forgets its initial state, so only the first few samples are ambiguous: after the transient, an impulse and an initial condition are distinguishable. A plant with an integrator or an oscillator remembers its initial state, so that whether the first ρ02N\rho_0^2N samples contain a new impulse or only a different initial condition cannot be resolved, and ρ2→ρ02\rho^2\to\rho_0^2 over that window regardless of the system. Theorem 3(iv) thus says that the integrator of Example 1 is representative of the whole class. The bound in Theorem 3(ii) is loose and meant qualitatively; the exact decay is given by (i), and for n=1n=1 it is γt=a02t\gamma_t=a_0^{2t}, which gives the poorly visible window

Visibility profiles for marginal systems

Figure 2. Visibility profiles. Top: marginal systems (integrator and undamped rotation, 30∘30^\circ per step); the curves for N=100N=100 and N=400N=400 coincide as functions of τ/N\tau/N, as Theorem 3(iv) states. Bottom: strictly stable scalar systems on a log scale; the approximation a02(τ+1)a_0^{2(\tau+1)} (dotted) is indistinguishable from the exact profile, so the poorly visible window has fixed length (Theorem 3(i)–(iii)).

ln⁡(1/(1−ρ02))/(2ln⁡(1/∣a0∣))\ln(1/(1-\rho_0^2))/(2\ln(1/|a_0|)). Jordan blocks are not covered by Theorem 3(iv).

Numerical studies

All tests have α=0.05\alpha=0.05 and N=100N=100 unless stated, and probabilities are estimated by Monte Carlo. The max⁡\max test is TT with the exact threshold of Algorithm 1 and known σ\sigma, and the UU test is UU with its exact threshold and unknown σ\sigma. A fault has raw signal-to-noise ratio μ∥da∥/σ=5\mu\lVert d_a\rVert/\sigma=5, so by (4) the test sees 5ρa5\rho_a. The statistics depend on γ\gamma only through P−ν\mathcal{P}_{-\nu}, so faulty runs use x0=0x_0=0 and σ=1\sigma=1. The systems are the integrator A=1A=1, the system A=0.9A=0.9 (both with B=C=1B=C=1), and the rotation A=0.9R(π/6)A=0.9R(\pi/6) with B=(1,0)⊤B=(1,0)^\top, C=(1,0)C=(1,0), where R(θ)R(\theta) is the planar rotation by θ\theta. More simulations are in Appendix II.

Calibration

The closed-form thresholds are conservative. Bonferroni, Šidák and Rice give false-alarm probabilities between 0.014 and 0.041 against the nominal 0.05, because adjacent projected atoms are strongly correlated (0.67 to 0.98 for τ≥10\tau\ge10), so the 99 statistics behave like far fewer independent ones. The exact threshold gives 0.049 to 0.051 (10510^5 runs). The UU test gives 0.049 to 0.052 for σ∈{0.1,1,10}\sigma\in\{0.1,1,10\} and

x0∼N(0,502I)x_0 \sim\mathcal{N}(0,50^2 I), as Theorem 1 states, whereas a Bonferroni threshold designed for σ=1\sigma=1 and used at σ=2\sigma=2 gives 0.72 to 0.99.

Probability of detection

Figure 3 shows the probability of detection against the fault time τ\tau (4000 runs per point). Detection follows visibility. For the integrator at τ=5\tau=5, ρa2=0.06\rho_a^2=0.06 and the max test detects with probability 0.11, barely above α\alpha; at τ=50\tau=50 (ρa2=0.51\rho_a^2=0.51) it detects with probability 0.81. For A=0.9A=0.9 both saturate at 0.95 to 0.98 after about ten steps. The bound of Proposition 2 holds at every point, is within 0.04 of the max test for ρa2≥0.8\rho_a^2 \ge0.8, and is loose elsewhere, by at most 0.17. The UU test loses at most 0.05. The χr2\chi_r^2 detector of Remark 2 is far weaker, with probability of detection 0.23 at τ=50\tau=50 and 0.45 at τ=90\tau=90 for the integrator, because it spreads the impulse over r=99r=99 dimensions.

Figure 3. Probability of detection against fault time τ\tau for the integrator (top) and A=0.9A=0.9 (bottom), raw signal-to-noise ratio 5. Markers: max test and UU test. Dashed: lower bound of Proposition 2. Dotted: χr2\chi_r^2 detector, closed form. The horizontal line is α\alpha.

Two inputs, two outputs

Two unit masses in a chain are connected to a wall by unit springs, with Rayleigh damping βK\beta K, β=0.3\beta=0.3, and

K=[2−1−11].K=\begin{bmatrix}2 & -1 \\ -1 & 1\end{bmatrix}.

The outputs are the two positions, a force impulse can act on either mass, the discretization is zero-order hold with step 0.5, and N=200N=200 (398 atoms). For a

fault on channel 1 (Figure 4), the UU test detects with probability 0.94 at τ=55\tau=55, but names the right channel in only 73% to 77% of the detecting runs for τ≤170\tau\le170. Atoms of the two channels at the same time have projected correlation 0.89 at τ=10\tau=10 and 0.95 for τ≥50\tau\ge50, as large as between neighboring times of one channel. It is easier to say that something hit the plant than which actuator did.

Figure 4. Fault on channel 1, N=200N=200, β=0.3\beta=0.3. Probability that the UU test alarms, and that it alarms with the maximizing atom on channel 1. The horizontal line is α\alpha.

Conclusion

We studied the detection of an impulsive fault of unknown time and channel in a linear system with unknown initial state. The invariant GLR test is the zero-solution test of a weighted lasso with a free nuisance block, and for unknown variance of a square-root lasso whose parameter is calibrated exactly. The threshold is computed exactly from the Gram matrix of the projected atoms. The part of a fault hidden by the initial state has a closed form through observability Gramians; it is a fixed-length window for strictly stable systems and a fraction (τ+1)/N(\tau+1)/N of the record for systems with unit-modulus modes. Open problems are several simultaneous faults, for which the covariance test [12] or a detection-boundary analysis is the natural tool, step, ramp and sensor faults, which change only the atoms, Jordan blocks, for which our numerics suggest a power law in τ/N\tau/N, and networked systems with structured setts. The simulations use small examples with N≤800N \le800.

Proofs

Proof of Proposition 1

Proof of Proposition 1. Split y=P⊥y+(I−P⊥)yy=P^\perp y+(I-P^\perp)y. Since P⊥da=wazaP^\perp d_a=w_a z_a and the two components are orthogonal,

∥y−Ox0−μda∥2=∥P⊥y−μwaza∥2+∥(I−P⊥)(y−μda)−Ox0∥2.\lVert y-\mathcal{O}x_0-\mu d_a\rVert^2=\lVert P^\perp y-\mu w_a z_a\rVert^2+\lVert(I-P^\perp)(y-\mu d_a)-\mathcal{O}x_0\rVert^2.

The matrix O\mathcal{O} has full column rank, so for each μ\mu the second term is zero at x0=O†(y−μda)x_0=\mathcal{O}^{\dagger}(y-\mu d_a). Maximizing over x0x_0 therefore removes the second term under both hypotheses, and the profile likelihoods are

sup⁡x0p(y;x0,μ,a,σ2)=(2πσ2)−Np/2exp⁡(−∥P⊥y−μwaza∥22σ2)=(2πσ2)−n/2pproj(P⊥y;μ,a,σ2),(11)\sup_{x_0} p(y;x_0,\mu,a,\sigma^2)=(2\pi\sigma^2)^{-Np/2}\exp\left(-\frac{\lVert P^\perp y-\mu w_a z_a\rVert^2}{2\sigma^2}\right) =(2\pi\sigma^2)^{-n/2}p_{\mathrm{proj}}(P^\perp y;\mu,a,\sigma^2), \tag*{(11)}

with μ=0\mu=0 under H0H_0, where pprojp_{\mathrm{proj}} is the density of P⊥yP^\perp y. The factor (2πσ2)−n/2(2\pi\sigma^2)^{-n/2} is the same under both hypotheses.

(i) For known σ\sigma the factor cancels in the ratio in (5), so ℓ=ℓproj\ell=\ell_{\mathrm{proj}}. By Lemma 2, maximizing (13) over μ\mu turns the exponent into −(∥P⊥y∥2−(zaTy)2)/(2σ2)-(\lVert P^\perp y\rVert^2-(z_a^\mathsf{T}y)^2)/(2\sigma^2) under atom aa and −∥P⊥y∥2/(2σ2)-\lVert P^\perp y\rVert^2/(2\sigma^2) under H0H_0. The log-ratio for atom aa is (zaTy)2/(2σ2)(z_a^\mathsf{T}y)^2/(2\sigma^2), and maximizing over aa gives T/2T/2.

(ii) Let R0=∥P⊥y∥2R_0=\lVert P^\perp y\rVert^2 and Ra=∥P⊥y∥2−(zaTy)2R_a=\lVert P^\perp y\rVert^2-(z_a^\mathsf{T}y)^2 be the residuals of Lemma 2. For m>0m>0 and R>0R>0,

sup⁡σ2(2πσ2)−m/2e−R/(2σ2)=(2πeR/m)−m/2.\sup_{\sigma^2}(2\pi\sigma^2)^{-m/2}e^{-R/(2\sigma^2)}=(2\pi eR/m)^{-m/2}.

With m=Npm=Np, the density in the first line of (13) gives the log-ratio Np2log⁡(R0/Ra)\frac{Np}{2}\log(R_0/R_a) for atom aa, and with m=rm=r, the density pprojp_{\mathrm{proj}} gives r2log⁡(R0/Ra)\frac{r}{2}\log(R_0/R_a). Both residuals are positive almost surely because r≥2r\ge2. Finally log⁡(R0/Ra)=−log⁡(1−(zaTy)2/∥P⊥y∥2)\log(R_0/R_a)=-\log(1-(z_a^\mathsf{T}y)^2/\lVert P^\perp y\rVert^2), and maximizing over aa gives −log⁡(1−U2)-\log(1-U^2).

Proof of Lemma 3

Proof of Lemma 3. Step 1: eliminate x0x_0. The penalties do not involve x0x_0. For fixed θ\theta, the minimum of ∥y−Dθ−Ox0∥2\lVert y-D\theta-\mathcal{O}x_0\rVert^2 over x0x_0 equals ∥P⊥(y−Dθ)∥2\lVert P^\perp(y-D\theta)\rVert^2, as in Lemma 2. So (9) is equivalent to min⁡θ12∥P⊥y−P⊥Dθ∥2+h∑awa∣θa∣\min_\theta\frac{1}{2}\lVert P^\perp y-P^\perp D\theta\rVert^2+h\sum_a w_a|\theta_a|, and (10) to the same problem with the norm unsquared and λ\lambda in place of hh.

Step 2: normalize. Since wa>0w_a>0, φa=waθa\varphi_a=w_a\theta_a is a bijective change of variables; P⊥Dθ=∑aφaza=ZφP^\perp D\theta=\sum_a\varphi_a z_a=Z\varphi with Z=[za]Z=[z_a], and the penalty becomes h∥φ∥1h\lVert\varphi\rVert_1. We obtain a standard lasso in the unit-norm dictionary ZZ with data y~=P⊥y\widetilde y=P^\perp y.

Step 3: optimality of zero. The objective 12∥y~−Zφ∥2+h∥φ∥1\frac{1}{2}\lVert\widetilde y-Z\varphi\rVert^2+h\lVert\varphi\rVert_1 is convex, so φ=0\varphi=0 is optimal if and only if 0∈−ZTy~+h∂∥φ∥1∣φ=00\in-Z^\mathsf{T}\widetilde y+h\partial\lVert\varphi\rVert_1|_{\varphi=0}, that is, ∥ZTy~∥∞≤h\lVert Z^\mathsf{T}\widetilde y\rVert_\infty\le h. Since za∈range⁡(O)⊥z_a\in\operatorname{range}(\mathcal{O})^\perp, ZTy~=ZTyZ^\mathsf{T}\widetilde y=Z^\mathsf{T}y, and the condition reads max⁡a∣zaTy∣≤h\max_a|z_a^\mathsf{T}y|\le h.

Step 4: square-root version. If y~≠0\widetilde y\ne0, φ↦∥y~−Zφ∥\varphi\mapsto\lVert\widetilde y-Z\varphi\rVert is differentiable at 00 with gradient −ZTy~/∥y~∥-Z^\mathsf{T}\widetilde y/\lVert\widetilde y\rVert, and the same subgradient argument gives ∥ZTy~∥∞/∥y~∥≤λ\lVert Z^\mathsf{T}\widetilde y\rVert_\infty/\lVert\widetilde y\rVert\le\lambda, that is, U≤λU\le\lambda.

Proof of Theorem 2

Proof of Theorem 2. The blocks of dad_a with k≥tk\ge t are CAibjCA^i b_j, i=k−t=0,…,q−1i=k-t=0,\ldots,q-1, and the earlier blocks vanish. The corresponding blocks of O\mathcal{O} are CAtAiCA^tA^i. Hence ∥da∥2=∑i<q∥CAibj∥2=bjTWqbj\lVert d_a\rVert^2=\sum_{i<q}\lVert CA^i b_j\rVert^2=b_j^\mathsf{T}W_qb_j and OTda=∑i<q(At+i)TCTCAibj=(At)TWqbj\mathcal{O}^\mathsf{T}d_a=\sum_{i<q}(A^{t+i})^\mathsf{T}C^\mathsf{T}CA^i b_j=(A^t)^\mathsf{T}W_qb_j. The projection of dad_a on range⁡(O)\operatorname{range}(\mathcal{O}) is O(OTO)−1OTda\mathcal{O}(\mathcal{O}^\mathsf{T}\mathcal{O})^{-1}\mathcal{O}^\mathsf{T}d_a, with squared norm daTO(OTO)−1OTda=bjTWqAtWN−1(At)TWqbjd_a^\mathsf{T}\mathcal{O}(\mathcal{O}^\mathsf{T}\mathcal{O})^{-1}\mathcal{O}^\mathsf{T}d_a=b_j^\mathsf{T}W_qA^tW_N^{-1}(A^t)^\mathsf{T}W_qb_j. This equals ∥da∥2−∥P⊥da∥2\lVert d_a\rVert^2-\lVert P^\perp d_a\rVert^2; dividing by ∥da∥2\lVert d_a\rVert^2 gives (11).

Proof of Theorem 3

Proof of (i)–(iii) of Theorem 3. (i) For fixed tt, q=N−t→∞q=N-t\to\infty, so Wq→W∞W_q\to W_\infty and WN→W∞W_N\to W_\infty, and W∞≻0W_\infty\succ0 by observability. The right-hand side of (11) is continuous in (Wq,WN)(W_q,W_N) at (W∞,W∞)(W_\infty,W_\infty), since its denominator tends to bjTW∞bj>0b_j^\mathsf{T}W_\infty b_j>0.

(ii) Let x=(At)TWqbjx=(A^t)^\mathsf{T}W_qb_j and let Wq1/2W_q^{1/2} be the symmetric square root. Then x=(At)TWq1/2(Wq1/2bj)x=(A^t)^\mathsf{T}W_q^{1/2}(W_q^{1/2}b_j), the numerator of (11) is xTWN−1x=∥WN−1/2(At)TWq1/2(Wq1/2bj)∥2x^\mathsf{T}W_N^{-1}x=\lVert W_N^{-1/2}(A^t)^\mathsf{T}W_q^{1/2}(W_q^{1/2}b_j)\rVert^2, and the denominator is ∥Wq1/2bj∥2\lVert W_q^{1/2}b_j\rVert^2. Therefore

1−ρa2≤∥WN−1/2∥2∥At∥2∥Wq1/2∥2=λmax⁡(Wq)λmin⁡(WN)∥At∥2.1-\rho_a^2\le\lVert W_N^{-1/2}\rVert^2\lVert A^t\rVert^2\lVert W_q^{1/2}\rVert^2=\frac{\lambda_{\max}(W_q)}{\lambda_{\min}(W_N)}\lVert A^t\rVert^2.

By monotonicity, Wq⪯W∞W_q\preceq W_\infty and WN⪰WnW_N\succeq W_n for N>nN>n, so λmax⁡(Wq)≤λmax⁡(W∞)\lambda_{\max}(W_q)\le\lambda_{\max}(W_\infty) and λmin⁡(WN)≥λmin⁡(Wn)\lambda_{\min}(W_N)\ge\lambda_{\min}(W_n). With ∥At∥≤Mλt\lVert A^t\rVert\le M\lambda^t this is (ii).

(iii) If an atom is poorly visible, 1−ρa2>1−ρ021-\rho_a^2>1-\rho_0^2, so by (ii) κ0M2λ2t>1−ρ02\kappa_0M^2\lambda^{2t}>1-\rho_0^2, which holds only for t<ln⁡(κ0M2/(1−ρ02))/(2ln⁡(1/λ))t<\ln(\kappa_0M^2/(1-\rho_0^2))/(2\ln(1/\lambda)). This bound on t=τ+1t=\tau+1 does not involve NN.

Proof of (iv) of Theorem 3. Step 1: the Gramian average. Let A=∑λλPλA=\sum_\lambda\lambda P_\lambda be the spectral decomposition over the distinct eigenvalues λ\lambda, with PλP_\lambda the spectral projectors. For real xx, CAkx=∑λλkCPλxCA^kx=\sum_\lambda\lambda^kCP_\lambda x, so ∥CAkx∥2=∑λ,μ(λμ)k(CPλx)T(CPμx)\lVert CA^kx\rVert^2=\sum_{\lambda,\mu}(\lambda\mu)^k(CP_\lambda x)^\mathsf{T}(CP_\mu x). For λ≠μ\lambda\ne\mu on the unit circle, λμ≠1\lambda\mu\ne1 and the geometric sum ∑k<i(λμ)k\sum_{k<i}(\lambda\mu)^k is bounded, so its average over ii vanishes. Hence xTWix/i→∑λ∥CPλx∥2=xTW‾xx^\mathsf{T}W_i x/i\to\sum_\lambda\lVert CP_\lambda x\rVert^2=x^\mathsf{T}\overline{W}x with W‾=∑λPλCTCPλ\overline{W}=\sum_\lambda P_\lambda C^\mathsf{T}CP_\lambda, and by polarization Wi/i→W‾W_i/i\to\overline{W}.

Step 2: W‾≻0\overline{W}\succ0. If xTW‾x=0x^\mathsf{T}\overline{W}x=0, then CPλx=0CP_\lambda x=0 for every λ\lambda. Each PλxP_\lambda x is zero or an eigenvector of AA lying in ker⁡C\ker C, which the PBH test excludes by observability. So Pλx=0P_\lambda x=0 for all λ\lambda, and x=∑λPλx=0x=\sum_\lambda P_\lambda x=0. In particular bjTW‾bj>0b_j^\mathsf{T}\overline{W}b_j>0.

Step 3: shift invariance. ATWiA=Wi+(Ai)TCTCAi−CTCA^\mathsf{T}W_iA=W_i+(A^i)^\mathsf{T}C^\mathsf{T}CA^i-C^\mathsf{T}C, and the last two terms are bounded in ii, because sup⁡t∈Z∥At∥=:M<∞\sup_{t\in\mathbb{Z}}\lVert A^t\rVert=:M<\infty. Dividing by ii, ATW‾A=W‾A^\mathsf{T}\overline{W}A=\overline{W}, hence (At)TW‾At=W‾(A^t)^\mathsf{T}\overline{W}A^t=\overline{W} and (A−t)TW‾A−t=W‾(A^{-t})^\mathsf{T}\overline{W}A^{-t}=\overline{W} for all t∈Zt\in\mathbb{Z}.

Step 4: the limit of the profile. Fix δ∈(0,1)\delta\in(0,1) and consider atoms with q=N−t≥δNq=N-t\ge\delta N. Write Wq=qW‾+EqW_q=q\overline{W}+E_q and WN=N(W‾+FN)W_N=N(\overline{W}+F_N) with ∥Eq∥/q→0\lVert E_q\rVert/q\to0 and ∥FN∥→0\lVert F_N\rVert\to0. With s=A−tbjs=A^{-t}b_j, ∥s∥≤M∥bj∥\lVert s\rVert\le M\lVert b_j\rVert, Step 3 gives x=(At)TWqbj=qW‾s+ex=(A^t)^\mathsf{T}W_qb_j=q\overline{W}s+e with ∥e∥≤M∥Eq∥∥bj∥\lVert e\rVert\le M\lVert E_q\rVert\lVert b_j\rVert, and sTW‾s=bjTW‾bjs^\mathsf{T}\overline{W}s=b_j^\mathsf{T}\overline{W}b_j. Hence xTWN−1x=q2NbjTW‾bj(1+o(1))x^\mathsf{T}W_N^{-1}x=\frac{q^2}{N}b_j^\mathsf{T}\overline{W}b_j(1+o(1)) and bjTWqbj=qbjTW‾bj(1+o(1))b_j^\mathsf{T}W_qb_j=qb_j^\mathsf{T}\overline{W}b_j(1+o(1)), with o(1)o(1) uniform over tt with q≥δNq\ge\delta N, since ∥s∥\lVert s\rVert is bounded and Eq,FNE_q,F_N do not depend on tt. By (11), 1−ρa2=qN(1+o(1))1-\rho_a^2=\frac{q}{N}(1+o(1)), that is, ρa2−t/N=o(1)\rho_a^2-t/N=o(1).

Step 5: late atoms. For q<δNq<\delta N the argument of part (ii) gives 1−ρa2≤λmax⁡(Wq)M2/λmin⁡(WN)≤Kq/N1-\rho_a^2\le\lambda_{\max}(W_q)M^2/\lambda_{\min}(W_N)\le Kq/N with K=2M4∥C∥2/λmin⁡(W‾)K=2M^4\lVert C\rVert^2/\lambda_{\min}(\overline{W}) for large NN, so ∣ρa2−t/N∣≤(K+1)δ|\rho_a^2-t/N|\le(K+1)\delta. Letting N→∞N\to\infty and then δ→0\delta\to0 proves (12). The count of poorly visible atoms follows, because atoms with t/N<ρ02−ϵt/N<\rho_0^2-\epsilon are poorly visible and those with t/N>ρ02+ϵt/N>\rho_0^2+\epsilon are not, for NN large.

The spacing test

The spacing test is exact and needs no simulation, but it requires σ\sigma known, and we take σ=1\sigma=1. Among the signed atoms szas z_a with s=±1s=\pm1, let λ1=max⁡(a,s)sza⊤y\lambda_1=\max_{(a,s)}s z_a^\top y be the largest correlation with yy, attained at (a∗,s∗)(a^*,s^*). Let V−V^- be the smallest value λ1\lambda_1 could take, with the component of yy orthogonal to za∗z_{a^*} held fixed, while (a∗,s∗)(a^*,s^*) is still selected. Explicitly, with ρa,s=ss∗za⊤za∗\rho_{a,s}=ss^*z_a^\top z_{a^*} and ra,s=(sza−ρa,ss∗za∗)⊤yr_{a,s}=(s z_a-\rho_{a,s}s^*z_{a^*})^\top y,

V−=max⁡{0,max⁡(a,s)≠(a∗,s∗)ra,s1−ρa,s},p=Q(λ1)Q(V−).(12)V^-=\max\left\{0,\max_{(a,s)\ne(a^*,s^*)}\frac{r_{a,s}}{1-\rho_{a,s}}\right\},\qquad p=\frac{Q(\lambda_1)}{Q(V^-)}. \tag*{(12)}

Under H0H_0 and Assumption 1, pp is uniform on [0,1][0,1] for every x0x_0 and NN, so rejecting when p≤αp\le\alpha has false-alarm probability exactly α\alpha. The price is that conditioning on V−V^- discards information: with correlated atoms V−V^- is close to λ1\lambda_1, so pp is rarely small and the test detects less often than the exact threshold of Section IV-B.

Proof of the exactness of the spacing test. Fix a candidate (a0,s0)(a_0,s_0) and let u=s0za0u=s_0z_{a_0}, ξ=u⊤y\xi=u^\top y. By the proof of Theorem 1, ξ∼N(0,1)\xi\sim\mathcal{N}(0,1), since uu is a unit vector in range⁡(O)\operatorname{range}(O). For every other (a,s)(a,s), sza⊤y=ρξ+rs z_a^\top y=\rho\xi+r with ρ=sza⊤u\rho=s z_a^\top u and r=(sza−ρu)⊤yr=(s z_a-\rho u)^\top y. The vector (ra,s)(r_{a,s}) is jointly Gaussian with ξ\xi and Cov⁡(r,ξ)=(sza−ρu)⊤u=ρ−ρ=0\operatorname{Cov}(r,\xi)=(s z_a-\rho u)^\top u=\rho-\rho=0, so it is independent of ξ\xi. The pair (a0,s0)(a_0,s_0) is the maximizer if and only if ρξ+r≤ξ\rho\xi+r\le\xi for all other (a,s)(a,s), that is, (1−ρ)ξ≥r(1-\rho)\xi\ge r. Since ρ<1\rho<1 for every other signed atom (Assumption 1; for the antipode (a0,−s0)(a_0,-s_0), ρ=−1\rho=-1 and r=0r=0, giving ξ≥0\xi\ge0), this is the event ξ≥V−(r)\xi\ge V^-(r), on which λ1=ξ\lambda_1=\xi. Therefore, conditionally on {(a∗,s∗)=(a0,s0)}\{(a^*,s^*)=(a_0,s_0)\} and on rr, λ1\lambda_1 is a standard normal truncated to [V−,∞)[V^-,\infty), with survival function t↦Q(t)/Q(V−)t\mapsto Q(t)/Q(V^-). A variable evaluated at its own survival function is uniform, so pp is conditionally uniform; the conditional distribution does not depend on (a0,s0)(a_0,s_0) or rr, hence pp is uniform unconditionally. □

Additional Simulations

Visibility versus horizon

Table I counts the poorly visible atoms {τ:ρτ2<0.9}\{\tau:\rho_\tau^2<0.9\}. For strictly stable systems the count does not depend on NN and matches the window ln⁡(10)/(2ln⁡(1/∣a0∣))\ln(10)/(2\ln(1/|a_0|)) of Remark 5: 10.9 for A=0.9A=0.9 (10 counted) and 114.6 for A=0.99A=0.99 (114 at N=400N=400; for N≤200N\le200 the horizon is shorter than the window). For unit-modulus modes the count grows like 0.9N0.9N. For the integrator it equals the number of τ\tau with (τ+1)/N<0.9(\tau+1)/N<0.9, namely 89, 179, 359, and a four-state, two-output system with two planar rotations (by 0.7 and 1.9 radians) gives the same counts, which is the universality of Theorem 3(iv). For the rotation with ρ=1\rho=1, max⁡τ∣ρτ2−(τ+1)/N∣\max_\tau|\rho_\tau^2-(\tau+1)/N| over the atoms with N−τ−1≥0.1NN-\tau-1\ge0.1N is 0.015, 0.0074, 0.0037, 0.0019 for N=100,200,400,800N=100,200,400,800, a rate of 1/N1/N. On a random stable four-state system (p=2p=2, spectral radius 0.92), formula (11) matches the direct projection to 7×10−167\times10^{-16}, and 1−ρa21-\rho_a^2 stays below κ(WN)∥AT∥2\kappa(W_N)\|A^T\|^2, with largest ratio 1.7×10−31.7\times10^{-3}: the bound of Theorem 3(ii) is correct and loose.

systemN=100N=100N=200N=200N=400N=400
A=1A=1 (integrator)89179359
A=0.99A=0.9975107114
A=0.9A=0.9101010
rotation, ρ=1\rho=191180360
rotation, ρ=0.95\rho=0.95232323

Table I. Poorly visible atoms {τ:ρτ2<0.9}\{\tau:\rho_\tau^2<0.9\}.

Thresholds

Table II gives the false-alarm probability of the max test at nominal value 0.05 (10510^5 runs, σ=1\sigma=1). The Rice bound is

systemBonferroniŠidákRiceexact
A=1A=10.0140.0140.0310.049
A=0.9A=0.90.0280.0290.0390.051
rotation, ρ=0.9\rho=0.90.0410.0410.0340.051

Table II. False-alarm probability of the max test at nominal value 0.05.

not uniformly tighter than Bonferroni, and the independence approximation of Fuchs (Šidák) coincides with Bonferroni to within 0.001.

Guard band and spacing test

Figure 5 adds two tests to Figure 3. The guard band test is the max test on the atoms with ρa2≥0.5\rho_a^2\ge0.5 (49 atoms for the integrator, 96 for A=0.9A=0.9), with the exact threshold recomputed. It lowers the threshold from 3.06 to 2.84 for the integrator but gains at most 0.023 in the probability of detection, and loses up to 0.26 (integrator, τ=20,35\tau=20,35) and 0.30 (A=0.9A=0.9, τ=1\tau=1) at the faults it excludes. Removing poorly visible atoms does not help. The spacing test of Appendix I-E has false-alarm probability 0.049, 0.052, 0.049 for the three systems, with uniform pp-values (Kolmogorov distance at most 0.010), but detects less: 0.46 against 0.81 at τ=50\tau=50 for the integrator, and it saturates at 0.78 against 0.98 for A=0.9A=0.9, because V−V^- is close to λ1\lambda_1 for correlated atoms.

Figure 5. As Figure 3, with the guard band test and the spacing test added.

Two inputs, two outputs

Table III uses the system of Section VI-C with β∈{0.3,1}\beta\in\{0.3,1\}, ρ(A)=1.0.972,0.909)\rho(A)=1.0.972,0.909). For β>0\beta>0 the count per channel does not depend on NN, and for β=0\beta=0 it grows like 0.9N0.9N. The slowest-mode window is 40.2 for β=0.3\beta=0.3 and 12.1 for β=1\beta=1, against 37 to 38 and 13 to 14 counted. The UU test has false-alarm probability 0.048 to 0.053 with σ=3\sigma=3 unknown and x0∼N(0,2I4)x_0\sim\mathcal{N}(0,2I_4), while Bonferroni and Šidák give 0.014 to 0.018.

Figure 6. Two inputs, two outputs, N=200N=200. Top: visible fraction ρτ2\rho_{\tau}^{2} of a fault on channel 1 for β∈{0,0.3,1}\beta\in\{0,0.3,1\}. Bottom: as Figure 4.

β=0\beta=0β=0.3\beta=0.3β=1\beta=1
N=100N=10090/9037/3813/14
N=200N=200179/18037/3813/14
N=400N=400360/36137/3813/14
UU test0.0480.0530.053
Bonferroni / Šidák0.0170.0140.018

Table III. POORLY VISIBLE ATOMS ON CHANNEL 1 / CHANNEL 2, AND FALSE-ALARM PROBABILITIES.

References

  1. [1]Y. Mo, S. Weerakkody, and B. Sinopoli, “Physical authentication of control systems: Designing watermarked control inputs to detect counterfeit sensor outputs,” IEEE Control Systems Magazine, vol. 35, no. 1, pp. 93–109, 2015.DOI
  2. [2]C. Murguia and J. Ruths, “Characterization of model-based detectors for CPS sensor faults,” arXiv preprint arXiv:1710.04407, 2017.arxiv.org/abs/1710.04407
  3. [3]C. D. Johnson, “Identification of unknown, time-varying forces/moments in dynamics and vibration problems using a new approach to deconvolution,” Shock and Vibration, vol. 5, no. 3, pp. 181–197, 1998.DOI
  4. [4]E. Y. Chow and A. S. Willsky, “Analytical redundancy and the design of robust failure detection systems,” IEEE Transactions on Automatic Control, vol. 29, no. 7, pp. 603–614, 1984.
  5. [5]M. Basseville and I. V. Nikiforov, Detection of Abrupt Changes: Theory and Application. Englewood Cliffs, NJ: Prentice Hall, 1993.
  6. [6]M. Fouladirad and I. Nikiforov, “Optimal statistical fault detection with nuisance parameters,” Automatica, vol. 41, no. 7, pp. 1157–1171, 2005.DOI
  7. [7]M. Fouladirad, L. Freitag, and I. Nikiforov, “Optimal fault detection with nuisance parameters and a general covariance matrix,” International Journal of Adaptive Control and Signal Processing, vol. 22, no. 5, pp. 431–439, 2008.DOI
  8. [8]A. S. Willsky and H. L. Jones, “A generalized likelihood ratio approach to the detection and estimation of jumps in linear systems,” IEEE Transactions on Automatic Control, vol. 21, no. 1, pp. 108–112, 1976.
  9. [9]F. Pasqualetti, F. Dörfler, and F. Bullo, “Attack detection and identification in cyber-physical systems,” IEEE Transactions on Automatic Control, vol. 58, no. 11, pp. 2715–2729, 2013.DOI
  10. [10]H. Fawzi, P. Tabuada, and S. Diggavi, “Secure estimation and control for cyber-physical systems under adversarial attacks,” IEEE Transactions on Automatic Control, vol. 59, no. 6, pp. 1454–1467, 2014.DOI
  11. [11]J.-J. Fuchs, “The generalized likelihood ratio test and the sparse representations approach,” in Image and Signal Processing (ICISP 2010), ser. Lecture Notes in Computer Science, vol. 6134. Springer, 2010, pp. 245–253.DOI
  12. [12]R. Lockhart, J. Taylor, R. J. Tibshirani, and R. Tibshirani, “A significance test for the lasso,” The Annals of Statistics, vol. 42, no. 2, pp. 413–468, 2014.
  13. [13]J. Taylor, J. Loftus, and R. J. Tibshirani, “Inference in adaptive regression via the Kac–Rice formula,” The Annals of Statistics, vol. 44, no. 2, pp. 743–770, 2016.DOI
  14. [14]R. J. Tibshirani, J. Taylor, R. Lockhart, and R. Tibshirani, “Exact post-selection inference for sequential regression procedures,” Journal of the American Statistical Association, vol. 111, no. 514, pp. 600–620, 2016.arxiv.org/abs/1401.3889
  15. [15]M. R. Sherriff and D. Chatterjee, “On a frame theoretic measure of quality of LTI systems,” in IEEE Conference on Decision and Control, 2017, pp. 4012–4017.
  16. [16]O. J. Dunn, “Multiple comparisons among means,” Journal of the American Statistical Association, vol. 56, no. 293, pp. 52–64, 1961.DOI
  17. [17]Z. Šidák, “Rectangular confidence regions for the means of multivariate normal distributions,” Journal of the American Statistical Association, vol. 62, no. 318, pp. 626–633, 1967.DOI
  18. [18]J. Sun, “Tail probabilities of the maxima of Gaussian random fields,” The Annals of Probability, vol. 21, no. 1, pp. 34–71, 1993.DOI

Paper details

vlnc:W_fgttwdkz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents