WHY MATHEMATICS? • FROM WHOLE-PROCESS ALARM TO SUBSYSTEM CLUE
Keep process structure visible while predicting quality
Why is mathematics important in engineering fault localisation? Multiblock partial least squares gives a concrete answer. It respects the way a complex process is divided into subsystems, learns latent directions related to quality and lets an analyst trace a global alarm back toward blocks and variables that deserve inspection.
The mathematical mechanism is hierarchical. Variables form block scores; block scores form a super score; latent scores support quality prediction; and deviations can be decomposed for diagnosis. This structure develops problem-solving skills because every level has a different meaning and a different possible error.
This guide uses a transparent two-block calculation and then builds toward realistic validation, dynamics and deployment. The examples explain mathematics in careers without promising that one model guarantees safety, employment or fault-free production.
- Route 1 — Organise the system: blocks, responses, alignment and scaling.
- Route 2 — Build the latent model: covariance, scores, weights and loadings.
- Route 3 — Monitor and diagnose: predictions, residuals and contributions.
- Route 4 — Deploy responsibly: components, missingness, dynamics, drift and review.
Choose a reading route · Open the two-block calculation · Use the practice studio
Why multiblock thinking matters
A whole-process model can say that something changed without showing where to look. Separate models can miss how subsystems interact. Multiblock PLS sits between those extremes: it builds a shared predictive model while preserving named groups of variables. The result can support earlier, more organised investigation when the block definitions, data alignment and validation are sound.
For students, this is the importance of mathematics in action. Weighted sums, covariance, regression and residuals become a language for a system with parts. The same habit transfers to finance, health data, environmental monitoring and AI: keep source groups visible, define the target and separate predictive evidence from causal explanation.
Chapter index
| Chapters | System layer | Read |
|---|---|---|
| 1–5 | Blocks, response and covariance | Frame the process |
| 6–12 | Scores, super weights, prediction and contributions | Follow the hierarchy |
| 13–16 | Components, validation, missingness and dynamics | Test the model |
| 17–20 | Regimes, causality, operations and learning | Transfer responsibly |
CHAPTER 1 OF 20 · Back to contents
1. Why divide a process into blocks
Large systems often have natural sections: feed preparation, reactor, separator, utilities and product testing. Multiblock partial least squares keeps those sections visible while modelling their joint relation to quality.
Process example
A production line may group flow and composition in block 1, reactor temperatures in block 2, and downstream pressures in block 3.
Boundary of the claim
Blocks should reflect a defensible process structure; arbitrary grouping can manufacture a convenient but misleading diagnosis.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
Did you know? A block can have modest influence on a global score yet a large residual, because modelled shared variation and unexplained variation answer different questions. Good diagnosis checks both.
CHAPTER 2 OF 20 · Back to contents
2. The response block defines relevance
PLS uses one or more response variables Y, so directions in the process data are chosen partly for their relation to quality or another target.
Process example
Product strength, purity or energy use can serve as Y when the goal is to monitor variation relevant to that outcome.
Boundary of the claim
A measured response may be delayed, noisy or available only for a subset of batches, limiting what the model can learn.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 3 OF 20 · Back to contents
3. Each block needs aligned observations
Rows across every X block and Y must describe the same run, batch or time window. Alignment errors break the cross-block interpretation before modelling begins.
Process example
A laboratory quality result for batch 42 must be paired with process trajectories from batch 42 rather than the next batch released from storage.
Boundary of the claim
Residence times and asynchronous sampling can make the correct pairing a modelling question rather than a clerical one.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 4 OF 20 · Back to contents
4. Centring and scaling are block decisions
Variables are centred and often scaled using training data. Block scaling may also prevent a large block from dominating because it contains more variables.
Process example
A utilities block with fifty sensors can outweigh a chemistry block with five unless its overall scale is controlled.
Boundary of the claim
Equal block scaling is not always physically sensible; the chosen rule should match the question and be validated.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 5 OF 20 · Back to contents
5. PLS seeks predictor-response covariance
In ordinary PLS, a predictor score t=Xw and response score u=Yc are chosen to share strong covariance. This connects latent directions to a predictive purpose.
Process example
A combination of temperature and flow that moves with purity can be more useful than the direction of largest process variance alone.
Boundary of the claim
Strong covariance can arise from a common operating schedule and need not identify a causal mechanism.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
Did you know? A block can have modest influence on a global score yet a large residual, because modelled shared variation and unexplained variation answer different questions. Good diagnosis checks both.
CHAPTER 6 OF 20 · Back to contents
6. Block scores preserve subsystem views
Multiblock PLS forms scores for each X block and combines them into a super score. The model can therefore show both whole-process movement and subsystem participation.
Process example
Feed, reactor and separator blocks each produce a score; their coordinated combination predicts the quality score.
Boundary of the claim
Different multiblock algorithms normalise scores and weights differently, so software conventions must be documented.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 7 OF 20 · Back to contents
7. Super weights combine block scores
Super weights describe how block scores contribute to a common latent component. They make the hierarchy explicit: variables form block scores, then blocks form a process score.
Process example
With two block scores t1=3/√2 and t2=4/√2 and equal super weights 1/√2, the super score is 3.5.
Boundary of the claim
A large super weight does not alone prove that a block caused a fault; it is part of a fitted predictive representation.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 8 OF 20 · Back to contents
8. Loadings reconstruct measured variation
Loadings map latent scores back toward the original variables. Residuals record the variation not represented by retained components.
Process example
A reactor-temperature loading pattern can show which temperatures typically rise together with a quality-relevant score.
Boundary of the claim
Loadings are conditional on preprocessing and correlations; unstable variables can exchange apparent importance.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 9 OF 20 · Back to contents
9. Regression predicts quality
The retained scores and response loadings yield a prediction Ŷ from X. Prediction residuals compare observed and predicted quality when measurements become available.
Process example
If a transparent illustrative coefficient is c=0.8 and T=3.5, the predicted quality score is 2.8.
Boundary of the claim
A latent score and a physical quality unit are not interchangeable unless the centring and scaling are reversed correctly.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
Did you know? A block can have modest influence on a global score yet a large residual, because modelled shared variation and unexplained variation answer different questions. Good diagnosis checks both.
CHAPTER 10 OF 20 · Back to contents
10. Monitoring combines score and residual views
Score-space statistics flag unusual movement inside the learned model, while residual statistics flag variation not explained by it. Both can be evaluated globally and by block.
Process example
A feed composition shift may move a retained score, while a failed sensor may appear mainly as a large residual in one block.
Boundary of the claim
One threshold cannot distinguish every failure mode; detection and diagnosis are separate tasks.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 11 OF 20 · Back to contents
11. Block contributions support localisation
A global alarm can be decomposed into block-level and variable-level contributions. The purpose is to prioritise inspection, not to claim automatic root cause.
Process example
In the equal-weight example, block contributions to T are 1.5 and 2.0, or about 42.86% and 57.14% of the signed total.
Boundary of the claim
Contribution percentages can be unstable, can cancel by sign and depend on the statistic being decomposed.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 12 OF 20 · Back to contents
12. A complete two-block calculation
A small calculation reveals the nested structure more clearly than a software dashboard: variable values create two block scores, block scores create T, and T predicts Y.
Process example
For x1=(2,1), t1=(2+1)/√2; for x2=(1,3), t2=(1+3)/√2; equal super weights give T=7/2.
Boundary of the claim
This is an educational projection with chosen weights, not a fitted multiblock PLS model.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 13 OF 20 · Back to contents
13. Deflation builds later components
After one component is extracted, modelled structure is removed so later components can represent remaining covariance patterns.
Process example
The first component may capture production rate, while a second separates temperature balance from composition balance.
Boundary of the claim
Deflation choices differ across algorithms and can make later components difficult to compare.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
Did you know? A block can have modest influence on a global score yet a large residual, because modelled shared variation and unexplained variation answer different questions. Good diagnosis checks both.
CHAPTER 14 OF 20 · Back to contents
14. Choose component count by validation
Too few components underfit quality-relevant structure; too many absorb noise. Grouped cross-validation can compare prediction, monitoring and stability.
Process example
Validation by entire batches or campaigns tests transfer better than randomly scattering adjacent rows.
Boundary of the claim
The component count that minimises prediction error may not give the best fault detection or interpretability.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 15 OF 20 · Back to contents
15. Regularisation and missing blocks
High-dimensional or missing-block data may require regularisation, sparse weights or algorithms designed for incomplete views.
Process example
A laboratory block may be absent online while process blocks remain available, requiring a clear deployment model.
Boundary of the claim
Imputation can leak outcome information or create artificial cross-block agreement.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 16 OF 20 · Back to contents
16. Dynamic processes need time structure
Static block models may miss lags and trajectories. Lagged variables, batch unfolding or dynamic latent models can represent time relationships.
Process example
A feed change can affect reactor temperature now and product quality much later, so aligned windows should reflect residence time.
Boundary of the claim
Adding many lags increases dimension and can blur which physical section deserves attention.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 17 OF 20 · Back to contents
17. Regimes and drift challenge one model
Grade changes, seasons and equipment ageing alter covariance and prediction relationships. A single global model may create systematic false alarms.
Process example
A block contribution that is unusual for low-load operation may be normal at full production.
Boundary of the claim
Separate regime models or guarded adaptation can help, but transitions and update rules need validation.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
Did you know? A block can have modest influence on a global score yet a large residual, because modelled shared variation and unexplained variation answer different questions. Good diagnosis checks both.
CHAPTER 18 OF 20 · Back to contents
18. Contributions are clues, not causes
Block and variable contributions identify where an alarm statistic accumulated. They do not prove which component initiated the event.
Process example
A separator block may show the largest deviation because it responds downstream to a feed disturbance.
Boundary of the claim
Root-cause analysis combines time order, process diagrams, interventions and maintenance evidence.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 19 OF 20 · Back to contents
19. Operational deployment needs an escalation path
A useful monitor specifies what happens after a global alarm: confirm data quality, inspect block contributions, check persistence, compare known scenarios and escalate.
Process example
A transient single-sample alarm may trigger observation; a persistent reactor-block residual may trigger an engineer review.
Boundary of the claim
Without ownership and response rules, even an accurate monitor becomes ignored or misused.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
CHAPTER 20 OF 20 · Back to contents
20. Student pathway connects mathematics to systems
Students can build from weighted averages and regression to vectors, covariance, latent scores, residuals and hierarchical diagnosis.
Process example
Start with the two-block calculation, then fit an ordinary PLS model, then restore block labels and compare diagnostics.
Boundary of the claim
Keep units, data lineage, split logic, thresholds and limitations in a model card so every result remains explainable.
Make this step auditable by naming the row unit, the block, the target, the training-only estimates and the output. Then trace how one changed sensor value moves through the hierarchy. A correct number with an unclear path is difficult to trust and almost impossible to debug.
For learning transfer, compare the latent calculation with a direct physical balance or simple baseline. If the sophisticated model adds no stable information, the simpler explanation may be better. Mathematics helps by making that comparison explicit rather than rewarding complexity for its own sake.
Worked calculation: two blocks, one super score
Let block 1 contain x₁=(2,1) and block 2 contain x₂=(1,3). Choose simple within-block weight vectors w₁=w₂=(1,1)/√2. The block scores are t₁=x₁w₁=3/√2≈2.1213 and t₂=x₂w₂=4/√2≈2.8284.
Now choose equal super weights a=(1,1)/√2. The super score is T=t₁/√2+t₂/√2=3/2+4/2=7/2=3.5. If an illustrative response loading or regression coefficient is c=0.8, the predicted centred quality score is ŷ=Tc=2.8.
| Quantity | Exact value | Decimal | Role |
|---|---|---|---|
| Block score t₁ | 3/√2 | 2.1213 | Block 1 summary |
| Block score t₂ | 4/√2 | 2.8284 | Block 2 summary |
| Block 1 term in T | 3/2 | 1.5 | 42.86% of signed total |
| Block 2 term in T | 4/2 | 2.0 | 57.14% of signed total |
| Super score T | 7/2 | 3.5 | Whole-process latent score |
| Predicted quality score | 0.8×3.5 | 2.8 | Illustrative prediction |
The block shares are 1.5/3.5=3/7≈42.86% and 2.0/3.5=4/7≈57.14%. These describe contributions to this signed score under the chosen normalisation. They are not probabilities and do not prove that block 2 caused a fault. In a real fit, weights come from a stated MB-PLS algorithm and the entire pipeline is validated.
Practice studio
Rebalance the blocks
Double every variable in block 1. Compare results with and without block scaling and explain which geometry each choice creates.
Add a response residual
If observed centred quality is 3.4 while prediction is 2.8, calculate the residual. List at least three explanations besides process failure.
Create a cancelling contribution
Change one block score to be negative. Explain why signed contributions can cancel and why absolute percentages may answer a different question.
Split by campaign
Design a train-validation-test scheme for six production campaigns. Keep all rows from each campaign together and justify the allocation.
Draw the escalation map
Start with a global alarm, then add data-quality confirmation, block contributions, variable checks, persistence and human review. Assign an owner to each step.
Guidance for students and parents
Students can begin with weighted averages and a two-block table. Once the hierarchy is clear, connect each operation to vectors and matrices. Keep one notebook page for dimensions: X_b has observations by variables, w_b has variables by component, and t_b has observations by component. Dimensional checks catch many errors before code runs.
Parents can ask the learner to identify the system parts, the quality outcome and the difference between a contribution and a cause. Those questions show why mathematics education matters beyond examinations: a responsible analyst must explain where a number came from and how far it can travel.
Frequently asked questions
How is multiblock PLS different from ordinary PLS?
Ordinary PLS can model all predictors together. Multiblock methods retain an explicit grouping and provide block-level scores, weights or diagnostics, with exact details depending on the algorithm.
Does the block with the largest contribution contain the root cause?
Not necessarily. It is the largest contributor to a chosen statistic under a fitted model. Downstream responses, correlations and scaling can move contribution without identifying the initiating mechanism.
Why not fit one model per subsystem?
Separate models can miss shared variation and interactions that matter for quality. They may still be useful baselines or local diagnostics, so compare rather than assume.
Can blocks have different numbers of variables?
Yes, but size can affect influence. Variable scaling and block scaling should be declared and validated so a large block does not dominate merely by count.
What happens when quality is measured late?
Rows must be paired according to process timing and material genealogy. Delayed laboratory results may require lagging, batch matching or soft-sensor design.
Is MB-PLS suitable for novel faults?
It can flag unusual scores or residuals, but named fault classification requires additional evidence. A new event may appear as a departure without receiving a correct label.
Primary research and next reading
The foundational process-monitoring reference is J. F. MacGregor, C. Jaeckle, C. Kiparissides and M. Koutoudi, “Process Monitoring and Diagnosis by Multiblock PLS Methods,” AIChE Journal 40(5), 826–838 (1994), DOI 10.1002/aic.690400509. The publisher abstract documents monitoring charts for process subsections, earlier detection and block/variable diagnosis in the studied setting.
- Partial least squares monitoring for quality-relevant faults
- Multiway PCA for batch process monitoring
- Recursive PCA for adaptive monitoring
- Sparse PCA for interpretable sensor diagnosis
- How Mathematics Works
Next, compare block contributions with time order. If the feed block moves first and the separator block reacts later, the temporal evidence changes the diagnosis. This is a strong example of mathematics and engineering working together: structure, prediction and causality remain distinct but connected.