Why Mathematics? | Bayesian Online Change-Point Detection for Condition Monitoring: Run Lengths, Hazard Functions, Predictive Evidence and False Alarms

A change detector asks whether the recent data still belong to the same segment. A change point marks a shift in parameters governing the observations, not simply one unusual value. Mean vibration rises and remains at a new level. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. An outlier, trend and regime change are different hypotheses. Define the type and persistence of change.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use what counts as a change point as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Did You Know? BOCPD updates every plausible time since the last change at once; its central state is a distribution over run length, not a single alarm score. This is a useful reminder that an algorithm's name is shorter than the assumptions needed to use it well.

Contents · Next chapter

Run length turns an unknown last-change time into a probability distribution that can be updated online. Run length records how many observations have occurred since the most recent change. Run length zero represents a change at the current step. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. It counts observations under the chosen convention, not necessarily elapsed hours. State indexing and timestamps.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use run length as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

The alert is built from predictive evidence, not from surprise alone. The algorithm carries probability mass over many possible run lengths rather than one estimated last-change time. Both a recent switch and a long stable run remain plausible. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Summarising only the most probable run can hide uncertainty. Plot the full distribution or key credible sets.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use run-length distribution as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

Sequential monitoring makes decision calibration as important as the recursion itself. The hazard gives the prior probability that a new segment begins after a run of a given length. A constant hazard implies a geometric prior on segment length. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. A convenient constant is not automatically physically justified. Use domain knowledge and sensitivity checks.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use hazard function as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Primary research: https://arxiv.org/abs/0710.3742. Check the authors' model, recursion and scope before transferring the method to a maintenance stream.

Contents · Previous chapter · Next chapter

A change detector asks whether the recent data still belong to the same segment. Each run-length hypothesis has a posterior predictive distribution based on data in its candidate segment. A stable Gaussian segment predicts the next vibration reading. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Independent-identically-distributed segments may be unrealistic for autocorrelated sensors. Choose an appropriate within-segment model.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use predictive model within a run as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Did You Know? BOCPD updates every plausible time since the last change at once; its central state is a distribution over run length, not a single alarm score. This is a useful reminder that an algorithm's name is shorter than the assumptions needed to use it well.

Contents · Previous chapter · Next chapter

Run length turns an unknown last-change time into a probability distribution that can be updated online. If no change occurs, probability mass moves from run length r to r plus one and is weighted by predictive evidence. A reading consistent with the current segment grows long-run hypotheses. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Evidence depends on parameter uncertainty, not only distance from a point estimate. Use the full predictive density.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use growth probabilities as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

The alert is built from predictive evidence, not from surprise alone. If a change occurs, mass from all previous run lengths resets to zero through the hazard and reset predictive model. Several histories contribute to the same new-segment hypothesis. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Reset predictions require a defensible prior for new-segment parameters. Run prior predictive checks.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use change probability as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

Sequential monitoring makes decision calibration as important as the recursion itself. Multiply continuation and change priors by their respective predictive likelihoods. Masses 0.18 and 0.08 normalise to 0.6923 and 0.3077. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. A likelihood ratio alone is not the posterior probability of change. Include prior odds.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use worked two-hypothesis update as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

A change detector asks whether the recent data still belong to the same segment. The sum across all growth and reset paths is the predictive probability of the new observation. The teaching total is 0.26. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Tiny evidence can cause numerical underflow across long sequences. Use log-sum-exp or scaling.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use normalising evidence as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Did You Know? BOCPD updates every plausible time since the last change at once; its central state is a distribution over run length, not a single alarm score. This is a useful reminder that an algorithm's name is shorter than the assumptions needed to use it well.

Contents · Previous chapter · Next chapter

Run length turns an unknown last-change time into a probability distribution that can be updated online. Conjugate models can update sufficient statistics for each run length efficiently. Normal observations with unknown mean use running counts and sums. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Conjugacy is a computational convenience, not proof of fit. Check tails, autocorrelation and variance shifts.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use conjugate updates as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

The alert is built from predictive evidence, not from surprise alone. Student-t predictive distributions arise when mean and variance are integrated under common conjugate priors. Short runs produce wider predictions than long runs. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Prior hyperparameters strongly affect early detection. Report them and perform sensitivity analysis.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use unknown mean and variance as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

Sequential monitoring makes decision calibration as important as the recursion itself. Implementation may discard very small run-length probabilities or cap the maximum run to control cost. A thousand-step stream keeps only material hypotheses. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Aggressive pruning can eliminate the true long-run history. Audit discarded mass.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use pruning and truncation as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Second authoritative source: https://www.cs.princeton.edu/~rpa/pubs/adams2007changepoint.pdf. Use it to compare notation, algorithmic conditions and the limits of the teaching example.

Contents · Previous chapter · Next chapter

A change detector asks whether the recent data still belong to the same segment. Repeated online checking creates alerts even under stable operation unless thresholds and decision rules are calibrated. A 0.30 change probability may occur occasionally by chance. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Posterior probability is not the same as a guaranteed long-run false-alarm rate. Evaluate simulations and stable historical periods.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use false alarms as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Did You Know? BOCPD updates every plausible time since the last change at once; its central state is a distribution over run length, not a single alarm score. This is a useful reminder that an algorithm's name is shorter than the assumptions needed to use it well.

Contents · Previous chapter · Next chapter

Run length turns an unknown last-change time into a probability distribution that can be updated online. Higher thresholds or lower hazards can reduce false alerts while delaying detection. A safety-critical shift may justify faster review. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. There is no universally optimal threshold. Choose with consequence and workload data.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use detection delay as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

The alert is built from predictive evidence, not from surprise alone. A robust predictive model can absorb isolated extremes while sustained evidence builds a change hypothesis. One sensor spike is followed by normal readings. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Too-heavy tails can hide genuine abrupt shifts. Test both contamination and true-change scenarios.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use outliers versus persistent changes as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

Sequential monitoring makes decision calibration as important as the recursion itself. Joint predictive distributions combine evidence across channels and represent their dependence. Temperature and vibration shift together. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Multiplying marginal likelihoods assumes conditional independence. Model covariance or use structured features.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use multiple sensors as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

A change detector asks whether the recent data still belong to the same segment. Load, maintenance and operating mode can explain changes that should not be treated as unexplained faults. A scheduled speed increase changes vibration. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Ignoring known interventions creates avoidable alarms. Condition on recorded context.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use covariates and interventions as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Did You Know? BOCPD updates every plausible time since the last change at once; its central state is a distribution over run length, not a single alarm score. This is a useful reminder that an algorithm's name is shorter than the assumptions needed to use it well.

Contents · Previous chapter · Next chapter

Run length turns an unknown last-change time into a probability distribution that can be updated online. Later data can confirm, revise or reject an online alert and refine the estimated change time. A persistent shift validates the earlier warning. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Retrospective certainty was unavailable at alert time. Preserve the original information set.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use retrospective verification as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Contents · Previous chapter · Next chapter

The alert is built from predictive evidence, not from surprise alone. A change probability becomes useful when combined with inspection costs, hazard consequences and alternative explanations. Route a moderate alert for review rather than automatic shutdown. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. Detection does not identify root cause by itself. Connect alerts to a documented response plan.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use maintenance decisions as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Useful next reading: condition monitoring, thresholds and false alarms, hidden-state sequence inference, Mathematics Learning Hub, eduKate Mathematics Learning System, approach to learning mathematics, Bukit Timah Mathematics curriculum overview, cross-subject learning approach. Each destination answers a different next question; choose by need rather than opening a link list at random.

Contents · Previous chapter · Next chapter

Sequential monitoring makes decision calibration as important as the recursion itself. BOCPD joins Bayes' rule, conditional probability, recursion, predictive distributions and decision thresholds. Students can reproduce the 0.18 and 0.08 update before considering many run lengths. The quantity being updated must retain its definition, units and time boundary; otherwise a numerically tidy result can answer the wrong reliability question. A high change probability is model-conditional evidence, not proof that equipment is failing. Explain the prior, evidence, alternative models and action separately.

Here is the audit path. First, write the conditional statement in words. Second, list the inputs that were known at this step. Third, calculate or simulate the update without using later information. Fourth, inspect whether probabilities sum to one, covariance has the right units, and the result moves in a physically sensible direction. Save the hazard function, run-length support, within-run predictive model, priors, pruning rule and alert threshold. This record lets another reader distinguish evidence, approximation and judgement.

For a student, the transferable skill is to slow the calculation down. Use practice, misconceptions and next reading as the focus, then multiply continuation probability 0.90 by likelihood 0.20 and change probability 0.10 by reset likelihood 0.80, then normalise 0.18 and 0.08 to obtain 0.6923 and 0.3077. Predict which term should increase before touching a calculator. Change one defensible input and explain why the result changes. Finish by saying what the number supports and what it cannot establish. That final sentence prevents mathematics from being mistaken for certainty.

Quick answers

Why is mathematics important here? It makes the hidden assumptions, probability updates, uncertainty and decision consequences explicit enough to inspect.

Does the algorithm reveal the true physical condition? It estimates a model-defined state from data. Validation and engineering evidence are still required.

Which school mathematics transfers? Ratios, weighted averages, functions, graphs, probability, matrices and careful use of units all reappear.

What should a beginner calculate first? Reproduce the small numerical example, retain every intermediate value and explain the direction of the update.

What should a practitioner report? Data boundaries, model equations, priors or starting conditions, algorithm settings, diagnostics, sensitivity checks and the decision rule.

How should parents support learning? Ask the student to explain one symbol, one assumption and one limit. Clear explanation is evidence of understanding, not an extra task after the mathematics.

Contents · Previous chapter · Mathematics Learning Hub