In this paper
AbstractIntroductionProblem Setup and PreliminariesSystem dynamics and attack modelProblem statementGeneralized Likelihood Ratio TestsEliminating the initial stateThe GLR testStructure and Calibration of the TestLasso formExact threshold from the Gram matrixProbability of DetectionInterference from the initial stateVisibility profileNumerical studiesCalibrationProbability of detectionTwo inputs, two outputsConclusionProofsProof of Proposition 1Proof of Lemma 3Proof of Theorem 2Proof of Theorem 3The spacing testAdditional SimulationsVisibility versus horizonThresholdsGuard band and spacing testTwo inputs, two outputsReferencesWorksMathematics
Sparse Fault Detection in Linear Systems: Equivalence of GLR Tests and -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:
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).
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).
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 asymptotically for systems with unit-modulus modes (Theorem 3), where 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 decoding to handle sparse attacks. Independently, Fuchs Fuchs [11] observed that for a dictionary of signal atoms in Gaussian noise the – 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 -values for the first variable entering a lasso path.
Problem Setup and Preliminaries
Notation. For a matrix , range is its range and its spectral norm; for a symmetric matrix, and are its extreme eigenvalues and is the Loewner order. is the th unit vector, is the all-ones vector, and is the right tail probability of the standard univariate normal . The null distribution of a statistic is its distribution under .
System dynamics and attack model
Consider the standard discrete time linear system:
where , , and . The initial state is unknown and the measurement errors are i.i.d. The triple is known and observable. The nominal input is zero.
A fault or an attack is an impulse of unknown amplitude that hits an unknown input channel at an unknown time , that is,
where is the th unit vector. We observe with , so the fault can occur at any . We call the pair a fault index and write for the set of all fault indices, so .
Stack the outputs and errors to get the static linear model:
Here is the observability matrix, which has full column rank because is observable and . The vector is the output response to a unit impulse , namely
where is the th column of . We call the signature of the fault .
Problem statement
Since is unknown, can be any vector in range. A fault with signature close to range can be explained, almost entirely, by a different . 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 in (1), test
without knowing and, in a variant, without knowing . Design a statistical test such that
(i) its false-alarm probability is at most a prescribed value for every (and every in the variant), and
(ii) its threshold can be computed offline from , without data.
Moreover, quantify for each the part of the fault that the 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 ” means that, under , exactly one of the faults, each a pair of input channel and fault time, has occurred. The alternative is therefore composite: it contains one model for each index , and the amplitude is unknown too.
Performance. A test declares a fault when . We measure it by its false-alarm probability and probability of detection . Here refers to the fault parameters . Requirement (i) of Problem 1 says that for every (and ).
Generalized Likelihood Ratio Tests
We consider GLR tests for known and for unknown noise variance . We first eliminate from the data.
Eliminating the initial state
Let be the orthogonal projector onto the parity space , which has dimension . Since , projecting (1) eliminates :
where . Moreover, is a maximal invariant1 for the shifts , [6].
Only the part of a fault signature that no initial state can imitate remains in the projected data . For each we define its visible length and visible fraction,
and call the visibility profile. If , then and hence the fault is indistinguishable from . We exclude this case in Assumption 1.
For , the projected atom is a unit vector, and the projected data
Thus, compared with the model with known, the overall fault magnitude is scaled down from to . So the problem becomes detecting an unknown multiple of one atom from the dictionary , with unknown.
Lemma 1 (Geometry of visibility). Let be the angle between and the subspace , that is, the smallest angle between and a nonzero vector of . Then .
Proof. Split with and . For every we have , so by the Cauchy–Schwarz inequality , with equality when . Hence . Since and are orthogonal and , the Pythagorean identity gives
Dividing by yields , and since and for , we conclude . □
Figure 1(a) illustrates the lemma. The unknown moves the data only along , so it can mimic the part of a fault . The visible part is the perpendicular component, whose length relative to is . Two corner cases are worth studying. First, if , a change in produces exactly the same data, and . Second, if , no can mimic it, and . The early fault shown in Figure 1 is close to the former, and the late fault is closer to the latter.
Figure 1. Geometry of the detection problem. (a) In the unknown initial state moves the data freely along . Of a fault atom only the component orthogonal to is visible; its length is with . (b) In the parity space the normalized visible components are unit vectors . 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 .
Assumption 1 (Distinguishable faults). For every , we impose (i) ; and (ii) for , .
Both conditions can be checked from the matrices and do not rely on the fault parameters .
Condition (i) states that no fault can be explained by alone. Thus, , that is, in Figure 1(a). It is the finite-horizon, single-impulse version of an undetectable input in [9], and if (i) fails, the data carry no information about that fault. For it is equivalent to : if , then for , so by observability and hence . Note that does not make a fault easy to detect. A small 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 is a multiple of .
The GLR test
Under , with fault index and amplitude , the density of is , and under the density . Since is composite and is unknown, we resort to the generalized likelihood ratio (GLR) statistic, in logarithmic form:
where is maximized only in the unknown-variance case, and the supremum over equals the supremum over because is continuous in . The GLR test is
Here is such that . Because is composite, optimal tests usually do not exist, and the GLR test is sub-optimal. To evaluate we need the residuals of the least-squares fits under each hypothesis.
Lemma 2 (Residuals). For each fault index ,
Proof. Consider the first minimization problem. For a fixed , least-squares minimization over gives
where the final equality follows from (4). Because is a unit vector and , the amplitude minimizes . From these results we get
Because , we have , proving the first identity. For the fault-free model, setting in the least-squares problem above gives the second identity. □
We evaluate (5). Let be defined as in (5) with the projected data in place of . The densities of under and are Gaussian and do not depend on . The next proposition evaluates and , 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 . Define the statistics
(i) (Known ): .
(ii) (Unknown ): and .
By Proposition 1, is an increasing function of when is known and of when is unknown. Hence the GLR test of the projected data, , is
Proposition 1 shows that the log-likelihood ratios of and of are the same for known , and for unknown they are equal up to the scaling factor . Hence the GLR test statistics are the same and for the unprojected data, and the only change is in the thresholds determined by in (6). In practice we choose and directly as the smallest values with and under (Section IV). See Remark 1 for how relates to .
Example 1 (Integrator). Let , so and with ones. Set . Then , , hence , and
where and are the sample means after and up to time . Under , is the standardized difference of two sample means, so is the classical scan statistic for a change in mean [5]. An early fault is poorly visible because only samples remain to estimate the unknown pre-fault level.
Remark 1 (The two statistics). Under , is an unbiased estimate of , and . So is with replaced by , and is the cosine of the angle between and the closest atom . It is invariant to scaling of , so its null distribution does not depend on (Theorem 1).
Remark 2 (Comparison with fault/attack detectors). In the well-known attack detector for cyber-physical systems [1, 2], the test statistic is . This is the value of in (5) if the fault/attack is allowed to be any vector of the parity space instead of an impulse fault (our case). Here is the number of degrees of freedom, the dimension of the parity space . Our fault/attack model is an impulse, a single atom of that space. The detector therefore spreads an impulse over all dimensions and cannot return , whereas matches one atom and returns the maximizing pair. With an arbitrary mean, is not identifiable from , so the detector needs given, whereas the impulse structure lets handle unknown . Detectors built on Kalman innovations assume a prior on , and the false-alarm probability is guaranteed only when follows it; ours holds for every .
Structure and Calibration of the Test
The test statistics and depend on the model (1) only through the unit atoms , 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 , determine the geometry of the test as a lasso, the null distribution of its statistics, and hence the thresholds.
Lasso form
Let be the matrix whose columns are the fault signatures in (1). For , consider the lasso problems
(P1) (9)
(P2) (10)
Let denote a solution of (P1) or (P2). Here , so is the weight on the fault . The variable is not penalized, and is the visible length in Figure 1(a). Our model (1), with a single fault of amplitude , is the special case , where is the -th standard basis vector. A solution with 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, is a solution of (P1) if and only if , that is, . If , is a solution of (P2) if and only if .
Hence the GLR tests in (8) reject exactly when , with in (P1) and in (P2). Geometrically (see Figure 1(b)), since , the test declares no fault exactly when lies in the set , which is a polytope when the atoms span .
Remark 3 (Insight and limits). The lemma gives the penalty a false-alarm meaning: with or calibrated as in Section IV-B, the lasso has false-alarm probability at most for every , in place of cross-validation. The free block gives invariance to the initial state, and the weights treat all atoms alike: with unit weights, is a solution of (P1) if and only if , so the rule thresholds at the level . 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 , and that standard lasso solvers and theory apply.
Exact threshold from the Gram matrix
We show that under the laws of and are determined by the Gram matrix of the atoms, and that the law of does not depend on or , so the thresholds of the GLR tests can be computed once offline.
Theorem 1 (Null distribution). Let be an orthonormal basis of and define the vector . Let . Under , for every and ,
where is uniform on the unit sphere of .
Proof. Under the hypothesis , because . Put . Then , which gives the statement for . Also , so . A standard Gaussian is rotation invariant, so is uniform on the sphere.
By Theorem 1, the threshold with under gives false-alarm probability at most for all and , and so does the threshold of for known . For a Gaussian mean test or a test, such a threshold is the inverse of the function or of the cdf. Here and 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 with the false alarm probability under , which give . This test declares a fault when at least one of the single-atom tests fires, and each single-atom test has false-alarm probability . Bonferroni correction
Algorithm 1 Offline thresholds and
Input: atoms , ; level ; number of samples
for do
draw
end for
value above which a fraction of lie
value above which a fraction of lie
Output: for (known ) and for
[16] adds these probabilities over the atoms and solves for . The Šidák bound in [17], the independence approximation used by Fuchs [11], multiplies them as if the atoms were independent and solves for . Finally, the Rice-type tube bound [18] uses the correlation between neighboring atoms for one input channel and solves , where 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 loses nothing for the GLR test (Proposition 1), but interferes with the fault: the component of in is indistinguishable from and is removed, so by (2) only remains. A fault of amplitude has visible length , so the visible fraction 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 , defined in Section II-B.
Proposition 2 (Interference and probability of detection). Let be known and let . For every , every fault and every amplitude ,
where is right tail probability of a standard Gaussian and .
Proof. Since and , , which does not depend on . Hence . The event contains , whose probability is . ∎
The raw signal-to-noise ratio is the length of the fault in units of the noise standard deviation, which is what the test would see if were known. The interference scales it by the visible fraction to , so the bound decreases as decreases: a poorly visible fault is detected less often, and its amplitude must be times larger to reach the same bound. The bound is conservative, because atoms other than can also raise the alarm.
Visibility profile
We now compute from . Let be the observability Gramian over samples; then , and for , since each summand is positive semidefinite.
Theorem 2 (Visibility profile). For let and . Then
For and the formula (11) reduces to , which equals for the integrator of Example 1.
The next result describes the profile for long records. Call an atom poorly visible if for a fixed .
Theorem 3 (Visibility for long records). Let be Schur stable, with , and let solve . Then: (i) for each fixed and , ; (ii) for every and every atom, , with ; (iii) the number of poorly visible atoms per input channel is at most , whatever is.
Now let be diagonalizable with all eigenvalues on the unit circle, observable, and for all . Then: (iv) exists and is positive definite, and
In particular, the number of poorly visible atoms per channel is for every system in this class, whatever 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 samples contain a new impulse or only a different initial condition cannot be resolved, and 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 it is , which gives the poorly visible window

Figure 2. Visibility profiles. Top: marginal systems (integrator and undamped rotation, per step); the curves for and coincide as functions of , as Theorem 3(iv) states. Bottom: strictly stable scalar systems on a log scale; the approximation (dotted) is indistinguishable from the exact profile, so the poorly visible window has fixed length (Theorem 3(i)–(iii)).
. Jordan blocks are not covered by Theorem 3(iv).
Numerical studies
All tests have and unless stated, and probabilities are estimated by Monte Carlo. The test is with the exact threshold of Algorithm 1 and known , and the test is with its exact threshold and unknown . A fault has raw signal-to-noise ratio , so by (4) the test sees . The statistics depend on only through , so faulty runs use and . The systems are the integrator , the system (both with ), and the rotation with , , where is the planar rotation by . 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 ), so the 99 statistics behave like far fewer independent ones. The exact threshold gives 0.049 to 0.051 ( runs). The test gives 0.049 to 0.052 for and
, as Theorem 1 states, whereas a Bonferroni threshold designed for and used at gives 0.72 to 0.99.
Probability of detection
Figure 3 shows the probability of detection against the fault time (4000 runs per point). Detection follows visibility. For the integrator at , and the max test detects with probability 0.11, barely above ; at () it detects with probability 0.81. For 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 , and is loose elsewhere, by at most 0.17. The test loses at most 0.05. The detector of Remark 2 is far weaker, with probability of detection 0.23 at and 0.45 at for the integrator, because it spreads the impulse over dimensions.
Figure 3. Probability of detection against fault time for the integrator (top) and (bottom), raw signal-to-noise ratio 5. Markers: max test and test. Dashed: lower bound of Proposition 2. Dotted: detector, closed form. The horizontal line is .
Two inputs, two outputs
Two unit masses in a chain are connected to a wall by unit springs, with Rayleigh damping , , and
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 (398 atoms). For a
fault on channel 1 (Figure 4), the test detects with probability 0.94 at , but names the right channel in only 73% to 77% of the detecting runs for . Atoms of the two channels at the same time have projected correlation 0.89 at and 0.95 for , 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, , . Probability that the test alarms, and that it alarms with the maximizing atom on channel 1. The horizontal line is .
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 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 , and networked systems with structured setts. The simulations use small examples with .
Proofs
Proof of Proposition 1
Proof of Proposition 1. Split . Since and the two components are orthogonal,
The matrix has full column rank, so for each the second term is zero at . Maximizing over therefore removes the second term under both hypotheses, and the profile likelihoods are
with under , where is the density of . The factor is the same under both hypotheses.
(i) For known the factor cancels in the ratio in (5), so . By Lemma 2, maximizing (13) over turns the exponent into under atom and under . The log-ratio for atom is , and maximizing over gives .
(ii) Let and be the residuals of Lemma 2. For and ,
With , the density in the first line of (13) gives the log-ratio for atom , and with , the density gives . Both residuals are positive almost surely because . Finally , and maximizing over gives .
Proof of Lemma 3
Proof of Lemma 3. Step 1: eliminate . The penalties do not involve . For fixed , the minimum of over equals , as in Lemma 2. So (9) is equivalent to , and (10) to the same problem with the norm unsquared and in place of .
Step 2: normalize. Since , is a bijective change of variables; with , and the penalty becomes . We obtain a standard lasso in the unit-norm dictionary with data .
Step 3: optimality of zero. The objective is convex, so is optimal if and only if , that is, . Since , , and the condition reads .
Step 4: square-root version. If , is differentiable at with gradient , and the same subgradient argument gives , that is, .
Proof of Theorem 2
Proof of Theorem 2. The blocks of with are , , and the earlier blocks vanish. The corresponding blocks of are . Hence and . The projection of on is , with squared norm . This equals ; dividing by gives (11).
Proof of Theorem 3
Proof of (i)–(iii) of Theorem 3. (i) For fixed , , so and , and by observability. The right-hand side of (11) is continuous in at , since its denominator tends to .
(ii) Let and let be the symmetric square root. Then , the numerator of (11) is , and the denominator is . Therefore
By monotonicity, and for , so and . With this is (ii).
(iii) If an atom is poorly visible, , so by (ii) , which holds only for . This bound on does not involve .
Proof of (iv) of Theorem 3. Step 1: the Gramian average. Let be the spectral decomposition over the distinct eigenvalues , with the spectral projectors. For real , , so . For on the unit circle, and the geometric sum is bounded, so its average over vanishes. Hence with , and by polarization .
Step 2: . If , then for every . Each is zero or an eigenvector of lying in , which the PBH test excludes by observability. So for all , and . In particular .
Step 3: shift invariance. , and the last two terms are bounded in , because . Dividing by , , hence and for all .
Step 4: the limit of the profile. Fix and consider atoms with . Write and with and . With , , Step 3 gives with , and . Hence and , with uniform over with , since is bounded and do not depend on . By (11), , that is, .
Step 5: late atoms. For the argument of part (ii) gives with for large , so . Letting and then proves (12). The count of poorly visible atoms follows, because atoms with are poorly visible and those with are not, for large.
The spacing test
The spacing test is exact and needs no simulation, but it requires known, and we take . Among the signed atoms with , let be the largest correlation with , attained at . Let be the smallest value could take, with the component of orthogonal to held fixed, while is still selected. Explicitly, with and ,
Under and Assumption 1, is uniform on for every and , so rejecting when has false-alarm probability exactly . The price is that conditioning on discards information: with correlated atoms is close to , so 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 and let , . By the proof of Theorem 1, , since is a unit vector in . For every other , with and . The vector is jointly Gaussian with and , so it is independent of . The pair is the maximizer if and only if for all other , that is, . Since for every other signed atom (Assumption 1; for the antipode , and , giving ), this is the event , on which . Therefore, conditionally on and on , is a standard normal truncated to , with survival function . A variable evaluated at its own survival function is uniform, so is conditionally uniform; the conditional distribution does not depend on or , hence is uniform unconditionally. □
Additional Simulations
Visibility versus horizon
Table I counts the poorly visible atoms . For strictly stable systems the count does not depend on and matches the window of Remark 5: 10.9 for (10 counted) and 114.6 for (114 at ; for the horizon is shorter than the window). For unit-modulus modes the count grows like . For the integrator it equals the number of with , 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 , over the atoms with is 0.015, 0.0074, 0.0037, 0.0019 for , a rate of . On a random stable four-state system (, spectral radius 0.92), formula (11) matches the direct projection to , and stays below , with largest ratio : the bound of Theorem 3(ii) is correct and loose.
| system | |||
| (integrator) | 89 | 179 | 359 |
| 75 | 107 | 114 | |
| 10 | 10 | 10 | |
| rotation, | 91 | 180 | 360 |
| rotation, | 23 | 23 | 23 |
Table I. Poorly visible atoms .
Thresholds
Table II gives the false-alarm probability of the max test at nominal value 0.05 ( runs, ). The Rice bound is
| system | Bonferroni | Šidák | Rice | exact |
| 0.014 | 0.014 | 0.031 | 0.049 | |
| 0.028 | 0.029 | 0.039 | 0.051 | |
| rotation, | 0.041 | 0.041 | 0.034 | 0.051 |
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 (49 atoms for the integrator, 96 for ), 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, ) and 0.30 (, ) 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 -values (Kolmogorov distance at most 0.010), but detects less: 0.46 against 0.81 at for the integrator, and it saturates at 0.78 against 0.98 for , because is close to 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 , . For the count per channel does not depend on , and for it grows like . The slowest-mode window is 40.2 for and 12.1 for , against 37 to 38 and 13 to 14 counted. The test has false-alarm probability 0.048 to 0.053 with unknown and , while Bonferroni and Šidák give 0.014 to 0.018.
Figure 6. Two inputs, two outputs, . Top: visible fraction of a fault on channel 1 for . Bottom: as Figure 4.
| 90/90 | 37/38 | 13/14 | |
| 179/180 | 37/38 | 13/14 | |
| 360/361 | 37/38 | 13/14 | |
| test | 0.048 | 0.053 | 0.053 |
| Bonferroni / Šidák | 0.017 | 0.014 | 0.018 |
Table III. POORLY VISIBLE ATOMS ON CHANNEL 1 / CHANNEL 2, AND FALSE-ALARM PROBABILITIES.
References
- [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]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]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]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]M. Basseville and I. V. Nikiforov, Detection of Abrupt Changes: Theory and Application. Englewood Cliffs, NJ: Prentice Hall, 1993.
- [6]M. Fouladirad and I. Nikiforov, “Optimal statistical fault detection with nuisance parameters,” Automatica, vol. 41, no. 7, pp. 1157–1171, 2005.DOI
- [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]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]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]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]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]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]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]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]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]O. J. Dunn, “Multiple comparisons among means,” Journal of the American Statistical Association, vol. 56, no. 293, pp. 52–64, 1961.DOI
- [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]J. Sun, “Tail probabilities of the maxima of Gaussian random fields,” The Annals of Probability, vol. 21, no. 1, pp. 34–71, 1993.DOI




