Choose your reading route
eduKateSG · Why Mathematics? · Generalized Likelihood-Ratio Fault Detection for Reliability Sensors: Change Time, Fault Magnitude, Log-Likelihoods and Thresholds
Start with the question, audit the worked example, or jump to the assumptions and practice. Every chapter stays open for uninterrupted reading.
Understand the question
Define the quantity and follow the mechanism.Route 2 · Chapters 6–10
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 11–15
Check assumptions
See where design and calibration matter.Route 4 · Chapters 16–20
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 21–24
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why mathematics is important becomes clear when a detector must consider both whether a fault occurred and which fault parameters best explain the data. A generalized likelihood-ratio (GLR) test compares a no-fault model with a fault model, maximising the alternative likelihood over unknown quantities such as change time or fault magnitude. In innovation-based fault detection, this turns residual patterns into explicit competing explanations.
For a simple standardized normal mean-shift example, suppose the alternative shift is δ=1 and observations are 0.2, 0.7 and 1.4. The log-likelihood-ratio increment is δy−δ²/2=y−0.5. Summing from the first observation gives 0.8; starting at the second gives 0.2+0.9=1.1; starting at the third gives 0.9. The maximised candidate is 1.1 with change time before the second observation. This teaching calculation fixes δ; a full GLR may also maximise over δ.
Did You Know? A GLR detector can estimate when a change began as part of the same optimisation that decides whether the change is credible.
This article answers one intent: The mechanism compares fault and no-fault explanations through log-likelihood evidence while searching over unknown change parameters.
GLR detection is a structured comparison between models. The alternative becomes flexible because unknown change time or magnitude is estimated from the data. That flexibility can improve detection and also increase overfitting and false-alarm opportunity. The suffix-sum example shows why the best onset is selected. Later chapters add fault signatures, unknown inputs, multiple hypotheses and recursive computation. Learners practise logs, maxima and cumulative evidence; practitioners practise explicit hypotheses, bounds and validation.
Full chapter index
Chapters 1–5 · Understand the question
Chapters 6–10 · Reproduce the example
Chapters 11–15 · Check assumptions
Chapters 16–20 · Explore applications
Chapters 21–24 · Study and verify
Begin with the object being measured, because a detector name is less useful than the statistic it changes. Likelihood ratios compare how well the same observations fit two stated models. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A sensor-bias model competes with the nominal filter. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
A high ratio supports only the included alternative. List plausible fault families. Transfer the idea by changing one input at a time and explaining the direction before computing.
A useful audit begins by writing the claim in plain language, then identifying the observation, reference model and comparison rule. In this chapter, the central claim is: Likelihood ratios compare how well the same observations fit two stated models. Test it first on a short sequence where every intermediate value can be inspected. Then change one condition and explain why the conclusion does or does not survive.
The mathematical value appears when a pattern is connected to one reproducible calculation. H0 represents no fault; H1 specifies a fault form with unknown parameters. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
H1 may be an additive bias after an unknown time. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Poorly chosen H1 misses other failures. State both hypotheses. Keep raw evidence beside the transformed statistic so diagnosis remains possible.
For transfer, sketch a before-and-after case and label what is measured, what is predicted and what is inferred. The concrete case here is: H1 may be an additive bias after an unknown time. That example is illustrative rather than evidence about a particular school, learner or machine. A sound explanation keeps calculation, interpretation and action in separate sentences.
Separate the mechanism from the interpretation so a number does not become unsupported certainty. Kalman innovations provide a convenient sequence with predicted covariance under H0. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A bias produces a structured mean response. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Nonwhite nominal residuals invalidate the simple likelihood. Check baseline diagnostics. Compare with a simple baseline and an alternative cause before deciding.
Reliability work needs a counterexample as well as a successful example. Here the important boundary is: Nonwhite nominal residuals invalidate the simple likelihood. Ask what other mechanism could create the same observed pattern, what additional measurement would separate the explanations, and how uncertainty should be recorded. This turns a formula into a defensible decision process.
Keep timing, units and assumptions visible before applying a threshold. Logs turn products into sums and improve numerical stability. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
For unit variance and shift δ, the increment is δy−δ²/2. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Density constants cancel only when models are formed consistently. Derive the expression. Explain the key equation aloud and point to the term carrying each meaning.
A student can practise this idea without specialised equipment. Start with a tiny table, reproduce the stated operation, and predict how the output changes when one entry is doubled, delayed or removed. The practical next step is: Derive the expression. Explain the result aloud, including the assumption that would most change the answer.
Make the example auditable by predicting direction, calculating and checking a nearby case. Evaluate candidate start times and retain the largest evidence. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
The example compares starts 1,2 and 3. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Repeated maximisation increases false-alarm opportunity. Calibrate the search. Report observation, assumption, calculation and authorised action as separate layers.
A useful audit begins by writing the claim in plain language, then identifying the observation, reference model and comparison rule. In this chapter, the central claim is: Evaluate candidate start times and retain the largest evidence. Test it first on a short sequence where every intermediate value can be inspected. Then change one condition and explain why the conclusion does or does not survive.
Begin with the object being measured, because a detector name is less useful than the statistic it changes. Increments −0.3,0.2,0.9 give suffix sums 0.8,1.1,0.9. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
The best fixed-shift candidate starts at observation two. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Short samples provide unstable timing. Report a range or uncertainty. Transfer the idea by changing one input at a time and explaining the direction before computing.
For transfer, sketch a before-and-after case and label what is measured, what is predicted and what is inferred. The concrete case here is: The best fixed-shift candidate starts at observation two. That example is illustrative rather than evidence about a particular school, learner or machine. A sound explanation keeps calculation, interpretation and action in separate sentences.
| Candidate start | Included increments | Log-likelihood sum |
|---|---|---|
| 1 | −0.3, 0.2, 0.9 | 0.8 |
| 2 | 0.2, 0.9 | 1.1 |
| 3 | 0.9 | 0.9 |
The mathematical value appears when a pattern is connected to one reproducible calculation. A full GLR maximises likelihood over the allowed magnitude. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
Estimated bias can be derived by weighted least squares in simple Gaussian cases. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Unbounded magnitude can overfit one outlier. Constrain physically plausible values. Keep raw evidence beside the transformed statistic so diagnosis remains possible.
Reliability work needs a counterexample as well as a successful example. Here the important boundary is: Unbounded magnitude can overfit one outlier. Ask what other mechanism could create the same observed pattern, what additional measurement would separate the explanations, and how uncertainty should be recorded. This turns a formula into a defensible decision process.
Separate the mechanism from the interpretation so a number does not become unsupported certainty. Signal when maximised evidence crosses a calibrated boundary. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
Higher thresholds reduce false alarms and increase delay. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
A likelihood value is not automatically a posterior probability. Use simulation or theory. Compare with a simple baseline and an alternative cause before deciding.
A student can practise this idea without specialised equipment. Start with a tiny table, reproduce the stated operation, and predict how the output changes when one entry is doubled, delayed or removed. The practical next step is: Use simulation or theory. Explain the result aloud, including the assumption that would most change the answer.
Keep timing, units and assumptions visible before applying a threshold. Dynamic systems spread a fault through time according to a signature. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A sensor step creates a known innovation response shape. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Wrong dynamics distort the signature. Validate impulse responses. Explain the key equation aloud and point to the term carrying each meaning.
A useful audit begins by writing the claim in plain language, then identifying the observation, reference model and comparison rule. In this chapter, the central claim is: Dynamic systems spread a fault through time according to a signature. Test it first on a short sequence where every intermediate value can be inspected. Then change one condition and explain why the conclusion does or does not survive.
Make the example auditable by predicting direction, calculating and checking a nearby case. Different fault locations produce different residual signatures. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A sensor bias and actuator loss may affect channels differently. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Some faults are not identifiable from available outputs. Check structural distinguishability. Report observation, assumption, calculation and authorised action as separate layers.
For transfer, sketch a before-and-after case and label what is measured, what is predicted and what is inferred. The concrete case here is: A sensor bias and actuator loss may affect channels differently. That example is illustrative rather than evidence about a particular school, learner or machine. A sound explanation keeps calculation, interpretation and action in separate sentences.
Begin with the object being measured, because a detector name is less useful than the statistic it changes. Run a bank of GLR correlations for several fault types or locations. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
The highest credible statistic suggests an isolation hypothesis. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Multiple tests inflate false discovery. Calibrate the family. Transfer the idea by changing one input at a time and explaining the direction before computing.
Reliability work needs a counterexample as well as a successful example. Here the important boundary is: Multiple tests inflate false discovery. Ask what other mechanism could create the same observed pattern, what additional measurement would separate the explanations, and how uncertainty should be recorded. This turns a formula into a defensible decision process.
The mathematical value appears when a pattern is connected to one reproducible calculation. Disturbances can mimic fault signatures. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A load change produces innovation evidence similar to actuator loss. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Ignoring unknown inputs creates false diagnoses. Use decoupling or joint models. Keep raw evidence beside the transformed statistic so diagnosis remains possible.
A student can practise this idea without specialised equipment. Start with a tiny table, reproduce the stated operation, and predict how the output changes when one entry is doubled, delayed or removed. The practical next step is: Use decoupling or joint models. Explain the result aloud, including the assumption that would most change the answer.
Separate the mechanism from the interpretation so a number does not become unsupported certainty. Use the full covariance in the likelihood. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A common-mode shift across sensors should not be counted as independent evidence. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Diagonal approximations exaggerate evidence. Whiten with stable factors. Compare with a simple baseline and an alternative cause before deciding.
A useful audit begins by writing the claim in plain language, then identifying the observation, reference model and comparison rule. In this chapter, the central claim is: Use the full covariance in the likelihood. Test it first on a short sequence where every intermediate value can be inspected. Then change one condition and explain why the conclusion does or does not survive.
Keep timing, units and assumptions visible before applying a threshold. A step, ramp, pulse and intermittent fault require different alternatives. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A pulse may produce evidence then disappear. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
One GLR form cannot be optimal for all. Choose a small justified library. Explain the key equation aloud and point to the term carrying each meaning.
For transfer, sketch a before-and-after case and label what is measured, what is predicted and what is inferred. The concrete case here is: A pulse may produce evidence then disappear. That example is illustrative rather than evidence about a particular school, learner or machine. A sound explanation keeps calculation, interpretation and action in separate sentences.
Make the example auditable by predicting direction, calculating and checking a nearby case. Limit candidate change times to a recent window for computation and adaptation. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A 100-sample window bounds memory. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Old slow faults may leave the window. Match window to detection horizon. Report observation, assumption, calculation and authorised action as separate layers.
Reliability work needs a counterexample as well as a successful example. Here the important boundary is: Old slow faults may leave the window. Ask what other mechanism could create the same observed pattern, what additional measurement would separate the explanations, and how uncertainty should be recorded. This turns a formula into a defensible decision process.
Begin with the object being measured, because a detector name is less useful than the statistic it changes. After a signal, confirm, estimate, isolate and decide when to restart. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
Repair creates a new baseline. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Immediate reset can cause repeat alarms. Define lifecycle rules. Transfer the idea by changing one input at a time and explaining the direction before computing.
A student can practise this idea without specialised equipment. Start with a tiny table, reproduce the stated operation, and predict how the output changes when one entry is doubled, delayed or removed. The practical next step is: Define lifecycle rules. Explain the result aloud, including the assumption that would most change the answer.
The mathematical value appears when a pattern is connected to one reproducible calculation. Recursive sufficient statistics can avoid recalculating every candidate from scratch. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
Cumulative sums yield suffix evidence efficiently. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Optimisation can become costly for nonlinear models. Profile and bound computation. Keep raw evidence beside the transformed statistic so diagnosis remains possible.
A useful audit begins by writing the claim in plain language, then identifying the observation, reference model and comparison rule. In this chapter, the central claim is: Recursive sufficient statistics can avoid recalculating every candidate from scratch. Test it first on a short sequence where every intermediate value can be inspected. Then change one condition and explain why the conclusion does or does not survive.
Separate the mechanism from the interpretation so a number does not become unsupported certainty. GLR detection can support earlier diagnosis and targeted inspection. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
Estimated bias and onset guide a sensor check. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
It does not guarantee the named fault is real. Seek independent confirmation. Compare with a simple baseline and an alternative cause before deciding.
For transfer, sketch a before-and-after case and label what is measured, what is predicted and what is inferred. The concrete case here is: Estimated bias and onset guide a sensor check. That example is illustrative rather than evidence about a particular school, learner or machine. A sound explanation keeps calculation, interpretation and action in separate sentences.
Keep timing, units and assumptions visible before applying a threshold. Measure false-alarm rate, detection delay, magnitude error, time error and isolation accuracy. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A detector can signal quickly but identify the wrong channel. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
One metric hides trade-offs. Report scenarios separately. Explain the key equation aloud and point to the term carrying each meaning.
Reliability work needs a counterexample as well as a successful example. Here the important boundary is: One metric hides trade-offs. Ask what other mechanism could create the same observed pattern, what additional measurement would separate the explanations, and how uncertainty should be recorded. This turns a formula into a defensible decision process.
Make the example auditable by predicting direction, calculating and checking a nearby case. The topic connects normal distributions, logs, maxima and cumulative sums. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
A learner can reproduce suffix scores 0.8,1.1,0.9. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Maximum evidence is not certainty. Explain the competing models. Report observation, assumption, calculation and authorised action as separate layers.
A student can practise this idea without specialised equipment. Start with a tiny table, reproduce the stated operation, and predict how the output changes when one entry is doubled, delayed or removed. The practical next step is: Explain the competing models. Explain the result aloud, including the assumption that would most change the answer.
Begin with the object being measured, because a detector name is less useful than the statistic it changes. Specify H0,H1, parameter bounds, covariance, candidate times, threshold, reset and confirmation. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
Unit tests include nominal, step, ramp and impulse cases. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
A detector without a hypothesis record is unauditable. Archive settings. Transfer the idea by changing one input at a time and explaining the direction before computing.
A useful audit begins by writing the claim in plain language, then identifying the observation, reference model and comparison rule. In this chapter, the central claim is: Specify H0,H1, parameter bounds, covariance, candidate times, threshold, reset and confirmation. Test it first on a short sequence where every intermediate value can be inspected. Then change one condition and explain why the conclusion does or does not survive.
The mathematical value appears when a pattern is connected to one reproducible calculation. CUSUM accumulates evidence for a targeted shift; GLR also maximises unknown parameters. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
GLR can estimate onset and magnitude. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
Extra flexibility increases computation and false-alarm calibration burden. Choose by uncertainty. Keep raw evidence beside the transformed statistic so diagnosis remains possible.
For transfer, sketch a before-and-after case and label what is measured, what is predicted and what is inferred. The concrete case here is: GLR can estimate onset and magnitude. That example is illustrative rather than evidence about a particular school, learner or machine. A sound explanation keeps calculation, interpretation and action in separate sentences.
Separate the mechanism from the interpretation so a number does not become unsupported certainty. GLR means generalised over unknown parameters, not universally correct. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
Change δ and recompute the example. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
The winning hypothesis may be the least wrong in a limited set. State excluded alternatives. Compare with a simple baseline and an alternative cause before deciding.
Reliability work needs a counterexample as well as a successful example. Here the important boundary is: The winning hypothesis may be the least wrong in a limited set. Ask what other mechanism could create the same observed pattern, what additional measurement would separate the explanations, and how uncertainty should be recorded. This turns a formula into a defensible decision process.
Keep timing, units and assumptions visible before applying a threshold. Willsky surveys dynamic-system failure detection; later GLR work treats unknown inputs; NASA documents innovation-based GLR schemes. The importance of mathematics here is that the statement becomes a quantity or logical test that another reader can reproduce.
Recent eduKate articles cover CUSUM, NIS and multiple models. Follow the order of operations, keep the time index and units, and check whether the direction makes sense before trusting the number.
References do not set one operational threshold. Validate with real fault signatures. Explain the key equation aloud and point to the term carrying each meaning.
A student can practise this idea without specialised equipment. Start with a tiny table, reproduce the stated operation, and predict how the output changes when one entry is doubled, delayed or removed. The practical next step is: Validate with real fault signatures. Explain the result aloud, including the assumption that would most change the answer.
The nearest published companions are Student’s t filtering, Huber robust Kalman filtering, maximum correntropy Kalman filtering and normalized innovation squared. The eduKateSG Mathematics Learning Hub, eduKateSingapore Mathematics World, eduKate Sengkang Mathematics Hub, Bukit Timah Tutor Mathematics Hub and eduKate Punggol Mathematics Pathway provide wider learning routes. Primary sources: https://doi.org/10.1016/0005-1098(76)90041-8, https://doi.org/10.1080/00207729608929330, https://ntrs.nasa.gov/api/citations/19820005272/downloads/19820005272.pdf?attachment=true.
Frequently asked questions
What is generalised in a GLR test?
Unknown alternative parameters such as fault magnitude or change time are estimated by maximising likelihood.
Can GLR estimate when a fault began?
Yes, when change time is included among candidate parameters, the maximising candidate provides an onset estimate.
Is a log-likelihood ratio a probability?
No. It is evidence comparing specified models; converting to probability requires additional assumptions and priors.
Why can one outlier dominate?
Flexible magnitude and timing can fit a single extreme value, so bounds, persistence and robust alternatives matter.
How is GLR different from CUSUM?
CUSUM targets a specified shift recursively; GLR also searches unknown parameters and usually needs more computation and calibration.
Extended practice laboratory
Practice case 1: Compute increments
Setup: Use δ=1 and three observations.
Checking target: Get −0.3,0.2,0.9.
Practice case 2: Suffix search
Setup: Sum from starts 1,2,3.
Checking target: Find maximum 1.1.
Practice case 3: Change δ
Setup: Use δ=0.5.
Checking target: Recompute evidence.
Practice case 4: Add an outlier
Setup: Replace final value by 4.
Checking target: Observe overfit risk.
Practice case 5: Estimate magnitude
Setup: Fit a mean shift after time two.
Checking target: Compare with fixed δ.
Practice case 6: Threshold sweep
Setup: Try several thresholds.
Checking target: Compare delay and false alarms.
Practice case 7: Ramp fault
Setup: Generate gradual bias.
Checking target: Test a step alternative.
Practice case 8: Unknown input
Setup: Add a load disturbance.
Checking target: Inspect confusion.
Practice case 9: Two sensors
Setup: Create correlated residuals.
Checking target: Use full covariance.
Practice case 10: Decision memo
Setup: Report detected onset and bias.
Checking target: Include uncertainty and alternatives.