Why Mathematics? | Kalman Filtering for Condition Monitoring: State-Space Models, Prediction, Update, Gain, Covariance and Sensor Fusion

Start with a hidden state prediction. Wear, temperature bias or health may be latent. Sensors observe noisy proxies. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. A state is a model construct, not automatically a physical variable. Define it.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Save state definitions, transition, process covariance, measurement map and sensor covariance. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use hidden condition state as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? The Kalman gain is derived from uncertainty, so a noisier sensor automatically receives less weight under the model. That is why a careful report includes the construction, not only the final scalar.

Contents · Next chapter

Now compare it with a noisy measurement. A matrix or equation predicts how state evolves. Wear increases from one interval to the next. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Unmodelled regime changes break the transition. Check residuals.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Rebuild every predict and update step from timestamped inputs. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use state transition as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The gain updates both mean and covariance. Process covariance represents uncertainty in state evolution. Random loads perturb degradation. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. It is not simply sensor error. Estimate or elicit it separately.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Check innovation distributions and covariance positive semidefiniteness. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use process noise as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Start with a hidden state prediction. A sensor maps hidden state to observed data. Voltage relates linearly to condition. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Calibration drift changes the mapping. Maintain sensor models.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Separate filtering, smoothing and forecasting results. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use measurement equation as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Primary research: https://doi.org/10.1115/1.3662552. Read the authors’ definition and conditions before carrying the method into a censored reliability model.

Contents · Previous chapter · Next chapter

Now compare it with a noisy measurement. Measurement covariance describes sensor uncertainty. A precise gauge has variance one. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Correlated sensors require off-diagonal terms. Model dependence.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Validate nonlinear or correlated-noise assumptions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use measurement noise as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The gain updates both mean and covariance. Propagate the previous posterior mean through the state equation. Mean condition advances to ten. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Nonlinear dynamics need extended or other filters. Use the correct algorithm.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Save state definitions, transition, process covariance, measurement map and sensor covariance. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use prediction mean as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? The Kalman gain is derived from uncertainty, so a noisier sensor automatically receives less weight under the model. That is why a careful report includes the construction, not only the final scalar.

Contents · Previous chapter · Next chapter

Start with a hidden state prediction. Propagate prior uncertainty and add process noise. Variance becomes four. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Forgetting process noise makes the filter overconfident. Audit the equation.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Rebuild every predict and update step from timestamped inputs. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use prediction covariance as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now compare it with a noisy measurement. Subtract predicted observation from the new measurement. Twelve minus ten gives residual two. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. A residual can reflect noise or model error. Monitor its distribution.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Check innovation distributions and covariance positive semidefiniteness. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use innovation as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The gain updates both mean and covariance. Combine predicted observation uncertainty with measurement noise. Four plus one equals five in the scalar example. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Covariances add only under stated independence. State assumptions.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Separate filtering, smoothing and forecasting results. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use innovation covariance as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Start with a hidden state prediction. Gain balances predicted uncertainty against measurement uncertainty. Four divided by five gives 0.80. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Gain is not an arbitrary learning rate. Derive it.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Validate nonlinear or correlated-noise assumptions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use kalman gain as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now compare it with a noisy measurement. Multiply gain 0.80 by innovation two and add to ten. The updated mean is 11.6. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Using the measurement twice double-counts information. Follow the recursion.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Save state definitions, transition, process covariance, measurement map and sensor covariance. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use worked update mean as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? The Kalman gain is derived from uncertainty, so a noisier sensor automatically receives less weight under the model. That is why a careful report includes the construction, not only the final scalar.

Contents · Previous chapter · Next chapter

The gain updates both mean and covariance. Multiply prior variance four by one minus gain. The updated variance is 0.8. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Roundoff can harm matrix covariance symmetry. Use stable forms.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Rebuild every predict and update step from timestamped inputs. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use worked update variance as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Second authoritative source: https://www.cs.cmu.edu/~motionplanning/papers/sbp_papers/k/Kalman1960.pdf. Compare its reference distribution, diagnostics and finite-simulation cautions with this worked example.

Contents · Previous chapter · Next chapter

Start with a hidden state prediction. The update is a precision-weighted compromise in the scalar direct-measurement case. Lower variance receives more weight. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. This simple view changes with indirect multivariate sensors. Keep the model context.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Check innovation distributions and covariance positive semidefiniteness. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use precision interpretation as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now compare it with a noisy measurement. Stack multiple measurements and their covariance. Vibration and temperature update one state. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Assuming independence when errors share a source overstates information. Model correlations.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Separate filtering, smoothing and forecasting results. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use sensor fusion as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The gain updates both mean and covariance. Prediction can continue without an update. A sensor packet is lost. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Uncertainty should grow during the gap. Do not impute certainty.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Validate nonlinear or correlated-noise assumptions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use missing measurements as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Start with a hidden state prediction. Transition matrices and process noise may depend on elapsed time. A two-hour gap differs from ten minutes. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Fixed-step code can misstate uncertainty. Use actual timestamps.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Save state definitions, transition, process covariance, measurement map and sensor covariance. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use irregular timing as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? The Kalman gain is derived from uncertainty, so a noisier sensor automatically receives less weight under the model. That is why a careful report includes the construction, not only the final scalar.

Contents · Previous chapter · Next chapter

Now compare it with a noisy measurement. A backward pass uses later data to refine earlier states. Post-event analysis improves historical estimates. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Smoothing is not available in real-time at that earlier moment. Separate online and retrospective results.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Rebuild every predict and update step from timestamped inputs. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use smoothing as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The gain updates both mean and covariance. Posterior state distributions yield probabilities of crossing a limit. Condition above twelve triggers review. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Threshold on mean ignores uncertainty. Use tail probabilities.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Check innovation distributions and covariance positive semidefiniteness. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use failure thresholds as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Start with a hidden state prediction. Innovations should match predicted covariance under the model. Standardised residuals expose drift. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Good residuals do not prove physical truth. Check multiple regimes.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Separate filtering, smoothing and forecasting results. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use model diagnostics as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now compare it with a noisy measurement. Extended, unscented or particle filters address nonlinear or non-Gaussian settings. A nonlinear degradation law bends the update. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Each approximation can fail differently. Benchmark on simulation.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Validate nonlinear or correlated-noise assumptions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use nonlinear extensions as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The gain updates both mean and covariance. Square-root or Joseph-form updates preserve covariance properties better. Long runs avoid negative eigenvalues from roundoff. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Equivalent algebra can differ numerically. Use tested libraries.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Save state definitions, transition, process covariance, measurement map and sensor covariance. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use numerical stability as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? The Kalman gain is derived from uncertainty, so a noisier sensor automatically receives less weight under the model. That is why a careful report includes the construction, not only the final scalar.

Contents · Previous chapter · Next chapter

Start with a hidden state prediction. Filtered state and uncertainty support inspection timing. Intervene when expected loss warrants it. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. A filtered estimate is not a maintenance policy. Add costs and constraints.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Rebuild every predict and update step from timestamped inputs. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use decision implications as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now compare it with a noisy measurement. Save state definitions, transition, Q, measurement map, R, initial distribution and time convention. A reviewer reproduces 11.6 and 0.8. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Gain alone is incomplete. Archive innovation logs.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Check innovation distributions and covariance positive semidefiniteness. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use reproducible reporting as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The gain updates both mean and covariance. Kalman filtering connects weighted averages, matrices, variance and feedback. Verify gain and update. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. More data are useful only through their uncertainty model. Draw predict-update arrows.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Separate filtering, smoothing and forecasting results. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use parent and student learning route as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Useful next reading: approach to learning mathematics, Bukit Timah Mathematics curriculum overview, cross-subject learning approach, previous Why Mathematics article, Mathematics Learning Hub, eduKate Mathematics Learning System. Choose the destination that matches the next question rather than treating links as a checklist.

Contents · Previous chapter · Next chapter

Start with a hidden state prediction. The Kalman filter does not automatically handle nonlinear dynamics, bad sensors or wrong noise models. Its optimality is conditional on specific assumptions. This step matters because sequential estimation, uncertainty updating and reliability sensor fusion require the mathematical object, prediction target and reliability question to stay aligned. Diagnostics and validation remain essential. Read Kalman’s original paper.

A reproducible audit should record variables, units, observation boundaries, censoring codes, model version and software settings. Validate nonlinear or correlated-noise assumptions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to guide reliability work.

Try a small learning experiment: calculate gain four over five, update mean ten with measurement twelve, and update variance four. Use common misconceptions and next reading as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Quick answers

What must be fixed before checking a model or algorithm? Define the dataset unit, censoring mechanism, time scale, parameterisation, simulation design and diagnostic rule.

Does a satisfactory diagnostic prove the model is true? No. It supports a particular check under stated assumptions and simulation conditions.

Which mathematics matters? Probability defines the reference distribution, simulation makes it observable, proportions summarise tail events, variance describes uncertainty and graphs expose structured mismatch.

How should a student practise? Preserve intermediate quantities, predict directions before calculating, change one input at a time and explain the result in ordinary language.

What follows an internal check? Sensitivity analysis, independent computation, held-out or later-data prediction and a separate decision analysis when cost or safety matters.

Contents · Previous chapter · Mathematics Learning Hub