WHY MATHEMATICS? • TWO VIEWS OF ONE SYSTEM
Use cross-block relationships to validate sensors with care
Why is mathematics important in sensor validation? Canonical correlation analysis turns two multivariable views of the same events into matched score pairs. It asks which linear combination of one block is most strongly related to which linear combination of the other, then uses departures from that learned relationship as evidence for review.
This matters wherever no single reading tells the whole story. Electrical, thermal, mechanical and chemical measurements can check one another only after rows are aligned, units are handled, covariances are estimated and assumptions are tested. The benefits of learning mathematics appear in the discipline of that chain: each transformation has a purpose, each threshold needs evidence and each conclusion has limits.
Students and parents can read this as a bridge from school algebra and statistics to engineering, computing and AI. The examples are educational, not a substitute for certified calibration or safety procedures.
- Route 1 — Pair the blocks: rows, centring, scaling and covariance.
- Route 2 — Build canonical pairs: weights, whitening, SVD and correlations.
- Route 3 — Validate: residuals, regularisation, pair selection and thresholds.
- Route 4 — Transfer: missingness, dynamics, interpretation and practice.
Browse all chapters · See the canonical-pair calculation · Student and parent guide
The mathematical mechanism in one sentence
CCA finds vectors a and b that maximise corr(Xa,Yb), then finds further pairs under orthogonality constraints. This optimisation converts a collection of pairwise relationships into ordered cross-block modes. A validation workflow can monitor canonical scores, prediction residuals or disagreement statistics, provided these quantities are tested on independent data.
The method is powerful because it respects the fact that each block has its own internal covariance. It is limited because correlation is linear, sample-dependent and associational. Good mathematics keeps both truths visible.
Chapter index
| Chapters | Question answered | Start |
|---|---|---|
| 1–6 | How are two sensor blocks prepared and connected? | Pair the evidence |
| 7–12 | How do canonical pairs and residuals work? | Build the pairs |
| 13–16 | How is the model stabilised and validated? | Test reliability |
| 17–20 | How do missingness, dynamics and interpretation change the result? | Use responsibly |
CHAPTER 1 OF 20 · Back to index
1. Why validate one sensor block against another
Canonical correlation analysis, or CCA, studies linear relationships between two sets of variables measured on the same observations. For sensor validation, one block can check whether another block still follows the learned cross-system pattern.
Concrete sensor example
A heating system may provide an electrical block with voltage and current and a thermal block with inlet temperature, outlet temperature and heat-flow estimates.
What could go wrong
CCA detects association between blocks; it does not prove that either block is calibrated or causally correct.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 2 OF 20 · Back to index
2. Rows must refer to the same events
CCA assumes row i in block X and row i in block Y describe the same case. Synchronisation, alignment and sampling rules therefore come before matrix algebra.
Concrete sensor example
If an electrical sample at 10:00 is paired with a thermal sample from 10:05, a real dynamic lag may be mistaken for sensor disagreement.
What could go wrong
Time stamps can be inaccurate, and the correct physical delay may vary with operating conditions.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 3 OF 20 · Back to index
3. Centring gives deviations a common origin
Subtracting each training mean forms centred blocks. Canonical relationships then describe co-variation around the reference operating centre instead of absolute offsets.
Concrete sensor example
A temperature of 65°C becomes a deviation of +3°C when the training mean is 62°C; the stored mean must be used for future data.
What could go wrong
Centring can hide meaningful level shifts if analysts forget to monitor the means separately.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
Did you know? CCA is unchanged by invertible linear re-expression within either block in its ideal population form, yet finite-sample estimation and regularisation can make the practical result depend strongly on preprocessing.
CHAPTER 4 OF 20 · Back to index
4. Scaling decides what variation counts
Standardising by training standard deviations prevents high-number variables from dominating simply because of units. Other physical scaling may be more appropriate when measurement tolerances differ.
Concrete sensor example
Voltage in millivolts and flow in litres per minute need a declared scale before their covariance matrices can be compared.
What could go wrong
A nearly constant but noisy sensor can be exaggerated by standardisation, so scaling needs engineering judgement.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 5 OF 20 · Back to index
5. Within-block covariance describes each view
Σ_XX and Σ_YY summarise variation inside the two blocks. Their diagonal terms are variances and off-diagonal terms are covariances among variables in the same block.
Concrete sensor example
Voltage and current may rise together inside X, while several temperature readings co-move inside Y. Those internal patterns affect which combinations are stable.
What could go wrong
Covariance estimates become unreliable when the sample is small, variables are redundant or operating modes are mixed.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 6 OF 20 · Back to index
6. Cross-covariance connects the blocks
Σ_XY records how deviations in X align with deviations in Y. CCA uses this bridge rather than studying either block in isolation.
Concrete sensor example
If higher electrical power normally accompanies greater heat flow, relevant cross-covariance entries are positive under comparable conditions.
What could go wrong
Cross-covariance can be created by a shared operating command, so it should not be interpreted automatically as direct causation.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 7 OF 20 · Back to index
7. Canonical variates are weighted combinations
CCA chooses a weight vector a for X and b for Y, creating u=Xa and v=Yb. The first pair maximises their correlation subject to variance constraints.
Concrete sensor example
A canonical electrical score may combine voltage and current, while its thermal partner combines temperature rise and flow.
What could go wrong
Weights depend on scale and correlation; a large coefficient is not automatically the most important physical sensor.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 8 OF 20 · Back to index
8. Whitening separates internal scale from cross-link
Whitening transforms each block so its fitted covariance becomes the identity. The remaining matrix expresses cross-block structure in comparable coordinates.
Concrete sensor example
If whitened cross-correlation is diag(0.9,0.4), singular values 0.9 and 0.4 are the canonical correlations.
What could go wrong
Whitening can amplify directions with tiny variance and becomes unstable when covariance matrices are nearly singular.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
Did you know? CCA is unchanged by invertible linear re-expression within either block in its ideal population form, yet finite-sample estimation and regularisation can make the practical result depend strongly on preprocessing.
CHAPTER 9 OF 20 · Back to index
9. Singular value decomposition finds the pairs
After whitening, singular value decomposition produces matched directions in the two blocks. Orthogonal pairs are ordered by decreasing canonical correlation.
Concrete sensor example
The first pair may represent overall load, while the second describes a weaker balance between two subsystems.
What could go wrong
Mathematical orthogonality does not guarantee simple physical interpretation.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 10 OF 20 · Back to index
10. Canonical correlations quantify linear agreement
Each canonical correlation lies between zero and one and measures the fitted linear correlation between a matched pair of scores.
Concrete sensor example
Correlations 0.9 and 0.4 correspond to squared values 0.81 and 0.16, often described cautiously as squared correlations rather than causal variance shares.
What could go wrong
A high training correlation can result from overfitting, drift or common control signals, and does not guarantee independent validation.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 11 OF 20 · Back to index
11. Residual disagreement creates a validation signal
When paired scores normally move together, a difference or prediction residual can flag observations that violate the reference relationship.
Concrete sensor example
For u=(1,0.5) and v=(0.8,-0.5), first-pair disagreement is 0.2 while second-pair disagreement is 1.0, suggesting the weaker mode deserves inspection.
What could go wrong
Residual scale depends on score normalisation and covariance; fixed raw differences may not be comparable across pairs.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 12 OF 20 · Back to index
12. A worked two-pair example
A diagonal whitened cross-correlation matrix provides a transparent calculation: its singular vectors are coordinate axes and its singular values are the two canonical correlations.
Concrete sensor example
With diag(0.9,0.4), the first canonical pair preserves the stronger cross-block relationship; retaining only it discards the second mode.
What could go wrong
Real matrices are rarely diagonal, so software must be checked against dimensions, centring and a small hand-verifiable case.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 13 OF 20 · Back to index
13. Regularised CCA stabilises high dimensions
Adding ridge terms to within-block covariance estimates prevents inversion from exploding when variables are numerous or collinear.
Concrete sensor example
Duplicated temperature channels make Σ_YY singular, while Σ_YY+λI remains invertible and limits extreme weights.
What could go wrong
Regularisation changes the target; λ must be selected on held-out groups rather than chosen solely for numerical convenience.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
Did you know? CCA is unchanged by invertible linear re-expression within either block in its ideal population form, yet finite-sample estimation and regularisation can make the practical result depend strongly on preprocessing.
CHAPTER 14 OF 20 · Back to index
14. Choosing the number of pairs
Analysts can compare held-out correlation, residual detection, stability and interpretability when deciding how many canonical pairs to retain.
Concrete sensor example
Keeping one pair may capture overall load but miss a sensor imbalance expressed by the second pair.
What could go wrong
Retaining every fitted pair can preserve noise and make thresholds harder to estimate.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 15 OF 20 · Back to index
15. Train and test without leakage
Means, scales, covariance estimates, regularisation and weights must be learned from training data only. Validation and test blocks use the frozen transformations.
Concrete sensor example
Splitting adjacent time points randomly allows nearly identical operating periods into both sets and overstates transfer.
What could go wrong
Use splits by asset, campaign, batch or time period so the test asks whether the relationship generalises.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 16 OF 20 · Back to index
16. Thresholds need false-alarm evidence
A residual statistic becomes actionable only after its reference distribution and decision threshold are validated. Quantiles, parametric approximations or resampling may be used cautiously.
Concrete sensor example
A 99th-percentile training threshold is not automatically a 1% future false-alarm rate if observations are dependent or conditions drift.
What could go wrong
Thresholds need independent evaluation, persistence rules and a documented response.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 17 OF 20 · Back to index
17. Missing data changes the paired sample
CCA requires paired information. Dropping rows, imputing values or fitting models for missingness each changes the evidence and must be reported.
Concrete sensor example
If one thermal channel fails exactly during high-load events, complete-case analysis removes the most important operating region.
What could go wrong
Imputation can create artificial agreement between blocks when it uses variables that also define the canonical relation.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 18 OF 20 · Back to index
18. Dynamics and lagged CCA
Physical responses may be delayed, so lagged blocks can test whether present X aligns with future Y or vice versa.
Concrete sensor example
Electrical input may precede temperature change by several minutes; comparing aligned lags can reveal the strongest relationship.
What could go wrong
Searching many lags can overfit, and a chosen lag may vary across regimes.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
Did you know? CCA is unchanged by invertible linear re-expression within either block in its ideal population form, yet finite-sample estimation and regularisation can make the practical result depend strongly on preprocessing.
CHAPTER 19 OF 20 · Back to index
19. Interpretation and root-cause limits
CCA identifies combinations that co-vary across blocks. It helps localise disagreement, but weights and residuals are not causal proof of a failed component.
Concrete sensor example
A drifting temperature sensor can break an electrical-thermal relation, but so can a heat leak, changed controller or timing error.
What could go wrong
Diagnosis needs physical checks, redundancy, maintenance evidence and alternative hypotheses.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
CHAPTER 20 OF 20 · Back to index
20. Student pathway and transferable reasoning
Students can learn CCA by moving from paired scatterplots to covariance, linear combinations, matrix whitening, SVD and residual monitoring.
Concrete sensor example
Begin with two single variables, then two two-variable blocks with a diagonal cross-correlation matrix, and finally a small real dataset.
What could go wrong
A model card should record row alignment, units, scaling, splits, retained pairs, thresholds, limits and escalation steps.
Audit the step by writing matrix dimensions, identifying the training-only quantities and checking one special case. Then ask whether the result would survive a new asset, a new season or a changed operating range. This turns the calculation into evidence rather than an attractive plot.
A useful learning routine is to explain the same idea three ways: in words, with a small diagram or table, and with one equation. If the three explanations disagree, pause and repair the model before adding more data or complexity.
Worked example: whitening, singular values and disagreement
Suppose each block has already been centred and whitened using training data. The fitted cross-correlation matrix is R=diag(0.9,0.4). Because R is diagonal and non-negative, its singular vectors are the coordinate axes and its singular values are 0.9 and 0.4. These are the two canonical correlations.
| Pair | Canonical correlation | Squared correlation | Cautious reading |
|---|---|---|---|
| 1 | 0.90 | 0.81 | Strong fitted linear agreement |
| 2 | 0.40 | 0.16 | Weaker fitted linear agreement |
For a new observation, suppose X produces canonical scores u=(1,0.5) and Y produces v=(0.8,−0.5). Simple paired differences are d=(0.2,1.0). The second disagreement is larger even though the second training correlation was weaker. A proper monitoring statistic would scale residuals using their fitted covariance and would compare the result with a threshold validated on independent normal operation.
The example is deliberately transparent. Real covariance matrices may require regularisation; the canonical directions rotate away from coordinate axes; sign choices are arbitrary; and residuals are correlated. Software output should be checked for centring, scaling, dimensions, solver settings and sign-consistent interpretation.
Practice and transfer lab
Recompute after a scale change
Multiply one raw X variable by 100. Explain why a correctly standardised analysis should represent the same information, then identify circumstances in which regularisation changes the result.
Introduce a timing error
Shift Y by one row and compare the first canonical correlation. Decide whether a lower correlation indicates a faulty sensor or merely bad alignment.
Test an operating-mode split
Fit on low-load data and test on high-load data. Record whether the relationship transfers and whether one model or separate models is defensible.
Design residual escalation
Create green, review and alarm zones for a scaled disagreement statistic. State the repeat measurement and human check required in each zone.
Compare with pairwise correlation
Explain what a multivariable canonical pair can reveal that a table of individual sensor correlations may miss, and what interpretability it sacrifices.
Student and parent guidance
Students should begin with paired scatterplots and ordinary correlation. Next, combine two variables in each block and calculate how a weighted sum changes the relationship. Whitening and singular value decomposition then have a clear job: remove within-block scale and find the strongest cross-block alignment.
Parents can support understanding by asking simple but deep questions: Do these rows describe the same event? Which quantities were learned from training data? What would count as a real validation failure? Can a high correlation still be misleading? These prompts build problem-solving skills without turning the conversation into formula recall.
Frequently asked questions
Is CCA just many ordinary correlations?
No. It optimises linear combinations from both blocks jointly while accounting for within-block covariance. Pairwise tables remain useful diagnostics, but they do not solve the same optimisation.
Does a canonical correlation of 0.9 mean 90% accuracy?
No. Correlation measures linear association between scores, not classification accuracy. Its square is 0.81, but even that is a squared association, not a guarantee or causal percentage.
Why are canonical weight signs sometimes reversed?
If (a,b) is a solution, (−a,−b) describes the same correlation. Software may choose either sign, so interpret matched patterns rather than treating sign orientation as unique.
Can CCA certify calibration?
No. It can reveal consistency or disagreement between blocks, but metrological calibration requires traceable standards, documented procedures and uncertainty assessment.
What if variables outnumber observations?
Sample covariance matrices become singular or unstable. Regularised CCA, dimension reduction or more data can help, but validation must reflect the chosen remedy.
Can nonlinear relationships be handled?
Kernel and deep variants exist, but they add tuning and validation risks. First establish whether transformations, regimes or physical features make a linear analysis adequate.
Primary source and selected next reading
CCA originates with Harold Hotelling, “Relations Between Two Sets of Variates,” Biometrika 28(3–4), 321–377 (1936), DOI 10.1093/biomet/28.3-4.321. The publisher record verifies the title, volume, pages and publication date.
- Canonical variate analysis for dynamic condition monitoring
- Partial least squares monitoring for quality-relevant faults
- Independent component analysis for non-Gaussian sensor monitoring
- Slow feature analysis for process monitoring
- How Mathematics Works
For a useful comparison, ask how CCA differs from PLS. CCA maximises correlation after accounting for both blocks’ variance, while PLS commonly seeks directions with strong covariance and a predictive purpose. The choice should follow the decision, sample size and validation evidence.