Choose your reading route
eduKateSG · Why Mathematics? · Transfer Learning for Reliability Across Fleets: Domain Shift, Feature Alignment, Fine-Tuning and Negative Transfer
Start with the question, audit the worked example, or jump to the assumptions and practice. Every chapter stays open for uninterrupted reading.
Understand the question
Define the quantity and follow the mechanism.Route 2 · Chapters 8–14
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 15–21
Check assumptions
See where design and calibration matter.Route 4 · Chapters 22–27
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 28–32
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why is mathematics important in transfer learning for reliability across fleets? A model trained on one machine, site or operating regime often meets a new distribution at deployment. Transfer learning tries to reuse representations while controlling the gap between source and target data. Similarity can help when target failures are scarce, but a poor transfer can be worse than starting again. This guide develops domain discrepancy, feature alignment, negative transfer, calibration and practical validation through concrete examples.
Full chapter index
Chapters 1–7 · Understand the question
Chapters 8–14 · Reproduce the example
Chapters 15–21 · Check assumptions
Chapters 22–27 · Explore applications
Chapters 28–32 · Study and verify
The source supplies a dataset, feature distribution and prediction task; the target is where the model will be used. Different loads, sensors, component variants or maintenance policies can shift both inputs and outcomes. Name the source, target and target label availability before selecting a transfer method.
Empirical risk minimisation assumes training and future data are sufficiently similar. A classifier or survival model can learn site-specific shortcuts. When vibration ranges, sampling rates or failure coding change, the same function may be miscalibrated or rank incorrectly. Transfer learning makes the similarity question explicit.
Under covariate shift, p(X) changes while p(Y|X) stays stable. Importance weighting can reweight source observations toward the target input distribution. This assumption is strong and testable only indirectly. A new component design can change the outcome mechanism as well, making simple weighting inadequate.
Label shift assumes class or event prevalence changes while feature distributions within labels remain stable. In reliability, maintenance policy or environment can alter failure prevalence. Estimate prevalence carefully and recalibrate. Survival outcomes add censoring and horizon dependence, so a single class-prior correction may be insufficient.
Concept shift changes p(Y|X): the same sensor pattern has a different relationship with failure. Firmware, design and intervention changes can cause it. Feature alignment alone can hide the shift rather than fix it. Target labels, mechanistic understanding and conservative deployment are needed.
Source and target may use different sensors, units or sampling rates. Map physical quantities and calibrate units before learning an abstract representation. A neural adapter cannot recover a missing measurement concept. Create a feature contract with units, ranges, timestamps and missingness semantics.
Start with source-trained parameters and update some or all layers on target data. Freeze early layers when they capture general signal patterns; unfreeze more when target differences are larger. Learning rate and target sample size control forgetting and overfitting. Compare against target-only and no-adaptation baselines.
Suppose two source groups have losses 0.10 and 0.30 with target importance weights 0.5 and 1.5. A normalised weighted mean is [0.5(0.10)+1.5(0.30)]/(0.5+1.5)=0.25. The unweighted mean is 0.20. Weights change which source regions represent the target; they do not prove the conditional model is stable.
| Group | Loss | Weight |
|---|---|---|
| 1 | 0.10 | 0.5 |
| 2 | 0.30 | 1.5 |
| Normalised weighted mean | 0.25 | — |
MMD measures distance between source and target distributions through kernel mean embeddings. Training can penalise MMD between learned representations. Zero empirical discrepancy does not guarantee equal outcome relationships. Kernel choice, bandwidth and sample size affect the measure. Report predictive target validation beside alignment.
A feature extractor is trained to predict the task while confusing a domain classifier. The goal is representation that carries task information but not easy source-target identity. If domain identity is connected to genuine failure mechanisms, removing it can hurt. Tune the trade-off and inspect negative transfer.
Correlation alignment matches feature covariance matrices across domains. Standardisation and covariance estimates matter, especially with few target units. Matching first two moments does not match tails, multimodality or labels. Use it as one diagnostic or regulariser, not a certificate of equivalence.
A source or combined autoencoder can learn compressed features before fault or survival prediction. Reconstruction rewards common signal variance, which may ignore rare failure information. Evaluate whether latent features retain event-relevant structure. A low reconstruction error is not a reliability metric.
Physics simulators can generate labelled faults that are rare in real fleets. The simulation gap includes sensor noise, unmodelled mechanisms and operating policy. Randomise plausible parameters, calibrate against real signals and hold out real failures for evaluation. Synthetic volume does not replace target validity.
Several fleets can contribute different regimes. Weight sources by similarity, sample size and label quality rather than pooling blindly. A large but mismatched source can dominate. Leave-one-source-out validation reveals whether the method can transfer to an unseen fleet.
A small labelled target set can calibrate or fine-tune the model. Keep a separate target test set; otherwise every scarce label becomes training and performance remains unknown. Use nested or repeated splits with wide uncertainty. Sometimes calibration alone is safer than full representation updates.
Unlabelled target signals reveal p(X), missingness and support but not p(Y|X). They can support alignment or self-supervision. Avoid claiming improved reliability without labelled or mature target outcomes. Monitor whether aligned features preserve physically meaningful separation.
Importance weights and adaptation fail where the target contains regions absent from the source. Estimate overlap, visualise distances and define abstention. A model cannot learn target failures for a new operating regime from no evidence. Collect data or narrow deployment.
Transfer is negative when adapted performance is worse than target-only or source-only baselines. It can arise from mismatched mechanisms, over-alignment or noisy source labels. Measure it explicitly across target folds and subgroups. A transfer method should be allowed to decline transfer.
Fine-tuning on a small target can erase useful source behaviour. This may matter if one model serves several fleets. Regularisation, replay or multi-domain training can help, but shared models also create governance complexity. Evaluate every supported domain after updates.
Source and target fleets can have different observation windows and retirement policies. A model may appear different because target outcomes are less mature. Use comparable horizons and censoring-aware metrics. Domain alignment of sensors cannot correct an informative follow-up process.
Adapted scores need target calibration. Evaluate calibration-in-the-large, slope and curves at operational horizons. A representation can transfer ranking while baseline risk changes. Recalibration may be enough when ordering transports; it is not enough when concept shift changes ordering.
Use target concordance or time-dependent AUC with uncertainty and mature labels. Small target event counts make estimates unstable. Report source, target and pooled performance separately. Never use pooled success to hide a failed target domain.
Track physical units, feature means, variances, missingness, embeddings and outcome rates. Statistical distances are screening tools, not final decisions. Link every alert to investigation and response. Some shifts are harmless seasonal changes; others invalidate the model.
Compare source-only, target-only, pooled, fine-tuned and adapted models under identical target test data. Include simple baselines and calibration. Repeat across target sites and time. This hierarchy shows whether transfer adds value beyond extra complexity.
| Model | Question |
|---|---|
| Source-only | Does knowledge transfer unchanged? |
| Target-only | What can local data do? |
| Fine-tuned | Does adaptation help? |
| Calibration-only | Is baseline shift sufficient? |
Adaptation strength, frozen layers, learning rate and alignment kernel can be tuned until the target test improves. That makes the test training data. Use inner target validation or a development domain, then evaluate once on a held-out target cohort.
Compare feature effects and explanations before and after transfer. A model may preserve performance by switching to site proxies. Explanations do not certify causality, but large changes can reveal unwanted dependencies. Use the same reference and output when comparing.
CHAPTER 27 OF 32 · Explore applications
27. Common misconception: similar embeddings mean safe transfer
An adversarial or MMD objective can make domains visually indistinguishable while erasing outcome-relevant distinctions. Only target performance, calibration and domain evidence can support deployment. Alignment is a means, not the goal.
CHAPTER 28 OF 32 · Study and verify
28. Common misconception: more source data always helps
A massive mismatched source can overwhelm a small relevant target. Quality, mechanism and overlap matter more than row count. Weight, select or reject sources and report negative-transfer checks. More data can add bias as well as precision.
Create source values near 0 and target values shifted by 2. Standardise separately and together, then calculate a simple weighted loss. Add a target region with no source points and explain why weighting cannot solve it. This connects averages, distributions and model limits.
Start with units, graphs and weighted averages, then study probability distributions and optimisation. Ask the learner what changed between source and target and whether the outcome rule also changed. This question is central to responsible AI and everyday comparison.
A transfer model can improve target ranking while worsening probability calibration. Representation and baseline risk move differently. This is why target calibration is measured separately rather than assumed from accuracy.
Read Wen, Gao and Li's fault-diagnosis transfer study at https://doi.org/10.1109/TSMC.2017.2754287. Connect it to https://edukatesg.com/2026/11/03/why-mathematics-conformal-survival-prediction-reliability-lower-bounds-censoring-coverage-distribution-shift/ and https://edukatesg.com/2026/11/03/why-mathematics-deep-survival-learning-reliability-neural-hazards-censoring-regularisation-validation/. Transfer learning matters because mathematics measures what is shared, what shifted and where evidence ends.
Extended practice laboratory
Begin with a domain contract rather than an adaptation algorithm. Identify what changed in measurements, populations, follow-up and outcome mechanisms. Test overlap, then compare source-only, target-only, pooled and adapted baselines on untouched target outcomes. Alignment, fine-tuning and weighting address different shift assumptions; none replaces target calibration. Record negative transfer and allow the system to decline reuse. A reliable transfer claim is narrow: this source helped this target, under these sensors, horizons and policies. Mathematics makes the reuse testable by measuring discrepancy, support, weighted risk and target performance.
Students should compare source and target graphs before discussing neural networks. Parents can ask whether the rule connecting inputs to outcomes also changed. That distinction is the heart of responsible transfer learning.
A capstone transfer dossier should state source and target domains, physical feature mapping, overlap, shift hypothesis, adaptation method, target labels and untouched evaluation. Compare source-only, target-only, pooled, calibration-only and adapted models. Show target calibration, negative-transfer results and unsupported regions. Repeat on another held-out fleet to test whether the method generalises beyond one pair. Add a rollback plan and a rule for refusing transfer. The strongest outcome may be a narrow deployment or a decision not to reuse the source model. Mathematics helps because it turns similarity from an impression into measured assumptions and target evidence.
Define domains before parameters
Source and target fleets may differ in equipment, sensors, loads, maintenance, environment and follow-up. Write these differences before adaptation. Feature names do not prove equivalence: temperature may be measured elsewhere, sampled differently or stored in another unit. Outcome definitions can shift when one fleet records component failure and another any removal. A data contract must establish comparable inputs, horizon, event and decision context. Otherwise alignment can make incomparable records look close without valid evidence.
A weighted-risk audit
With target-similarity weights 0.5 and 1.5 and losses 0.10 and 0.30, normalised weighted loss is (0.5×0.10+1.5×0.30)/(0.5+1.5)=0.25. Importance weighting aims to make source evaluation resemble target covariates under assumptions about overlap and outcome mechanism. Large weights reveal little source support and inflate variance. Display weight tails and validate on real target labels. If target operation lies outside all source ranges, reweighting cannot manufacture the missing mechanism.
Fine-tuning needs baselines
Compare source-only, target-only, pooled, calibration-only and fine-tuned models on untouched target data. Target-only shows what local labels can do. Calibration-only is useful when ranking transfers but baseline risk shifts. Choose frozen layers inside target development data. Report uncertainty across target splits or fleets. If adaptation worsens a proper score or calibration against target-only, call it negative transfer and keep the fallback rather than celebrating source performance.
Alignment can erase signal
MMD and adversarial domain classifiers can make representations harder to distinguish, but lower discrepancy is not the reliability objective. If a feature separates fleets because target machines truly run hotter and heat affects failures, erasing it can hurt prediction. Evaluate target outcome performance and calibration across alignment strengths. A domain classifier near chance may mean alignment, weak testing or collapsed features. Mechanism review and target validation decide which explanation is credible.
Follow-up shift differs from sensor shift
Source assets may have mature outcomes while target assets are newly deployed. Apparent target failure is then lower because many future events are censored. Sensor alignment cannot repair this. Align time origins, entry rules, event definitions and horizons; use censoring-aware validation and wait for mature outcomes where needed. Check whether maintenance removes high-risk assets before recorded failure. A model can transfer signal shapes and still misstate probabilities because baseline hazard or observation policy changed.
Overlap and abstention
Plot source and target operating ranges, nearest-source distance or density-ratio diagnostics. Mark target regions with weak support and evaluate them separately. An embedding picture can conceal absent physical regimes. Establish an abstention or local-model rule before deployment. Collect target labels in unsupported regions or narrow the scope. Confidence from a neural network is not evidence of support. The transfer claim should name the fleets, sensors, ranges and horizons actually tested.
A safe fleet contract
Specify approved fleets, sensor versions, ranges and horizons. Add unit checks, an overlap score, abstention and target calibration monitoring. Re-evaluate source fleets after fine-tuning to detect forgetting. Hold out a fleet or later period when possible and report results separately. Create rollback criteria for drift and negative transfer. The strongest result may be limited reuse or rejection. Mathematics turns a vague impression of similarity into measured assumptions and a bounded decision.
Target evidence and staged rollout
Begin with a shadow evaluation in which predictions are stored but do not change maintenance. This reveals target calibration, latency, missing fields and unsupported cases under the real pipeline. Next use a limited, reviewed rollout with predefined stopping rules. Compare performance by fleet, operating regime and data freshness rather than relying on one pooled score. Monitor whether technicians respond differently to alerts, because the policy can change future labels and censoring. Keep a target-only challenger and a calibration-only option live for comparison. When new target failures mature, update the transfer decision with a time-stamped analysis; do not repeatedly tune on the same holdout. A transfer model should be promoted only when target evidence shows useful, stable improvement over honest baselines and when the system can abstain outside support. This staged process keeps options open for students and engineers: reuse is a hypothesis to test, not a shortcut guaranteed by a large source dataset.
Questions that reveal negative transfer
Ask whether target error rises where source density is highest or where it is lowest. Check whether adaptation improves ranking but worsens calibration, and whether one fleet benefits while another loses. Compare failures before and after unit conversion, representation alignment and fine-tuning. Inspect cases rejected by the overlap rule. If the adapted model is better only after pooling the same target cases used for tuning, the evaluation is circular. Students can learn a powerful scientific habit here: a method needs a condition under which it will be rejected. Predeclared baselines and stopping rules give transfer learning that condition.
A classroom domain-shift experiment
Create two synthetic fleets with the same two sensor names but change units, operating range and the relationship between heat and failure. Train on the source, then test direct reuse, unit repair, recalibration and fine-tuning on a small target sample. Plot target calibration and identify unsupported cases. Students should predict which change each method can repair before seeing results. The experiment shows why covariate shift, baseline-risk shift and concept shift are different. Parents can ask what evidence comes from the target itself. That single question prevents a large source dataset from being mistaken for local proof.
Final self-check
Before deployment, verify units, feature meaning, outcome definition, horizon, overlap and target follow-up. Confirm that the adapted model beats source-only and target-only baselines on untouched target data, with calibration as well as ranking. Record unsupported regions and rollback rules. These checks keep transfer claims narrow, reproducible and open to rejection when new target evidence disagrees.
Practice case 1: Domain inventory
Setup: Compare source and target units.
Task: Map only comparable features.
Checking target: Semantic alignment comes first.
Write source and target feature summaries with physical units, ranges, sampling rates and missingness meanings. Identify whether differences are calibration issues, real operating shifts or feature mismatches. Map only comparable quantities. A learned adapter cannot fix Celsius interpreted as Fahrenheit or a target sensor that measures a different phenomenon. The domain inventory should precede any embedding plot.
Practice case 2: Weighted risk
Setup: Use two losses and weights.
Task: Calculate target-weighted mean.
Checking target: Inspect overlap and weight tails.
Use source losses 0.10 and 0.30 with target importance weights 0.5 and 1.5. Calculate the normalised weighted loss 0.25 and compare the unweighted 0.20. Then add a target region with no source observations: its density ratio cannot be estimated reliably. This shows both what reweighting does and where overlap ends. Inspect weight tails and effective sample size.
Practice case 3: Fine-tuning
Setup: Freeze different layers.
Task: Compare target validation.
Checking target: Transfer must beat honest baselines.
Train source-only, target-only and fine-tuned models with the same target test cohort. Freeze different layer sets and record performance, calibration and uncertainty. A source representation may help with few labels, while full fine-tuning can overfit or forget. Select settings inside target development data. Transfer earns its name only if it improves against honest baselines.
Practice case 4: MMD
Setup: Measure representation discrepancy.
Task: Compare alignment and target performance.
Checking target: Lower discrepancy is not the goal alone.
Compute a representation discrepancy such as MMD before and after adaptation. Then check target label performance. A lower discrepancy with worse predictions reveals over-alignment. Vary the kernel bandwidth and report sensitivity. Distribution matching is an optimisation term, not proof that outcome mechanisms match. Keep the target test outcomes outside alignment tuning.
Practice case 5: Adversarial alignment
Setup: Confuse a domain classifier.
Task: Check useful target structure.
Checking target: Do not erase outcome mechanisms.
Create a domain-adversarial model and inspect what the domain classifier can still detect. Remove an operating-regime feature that genuinely changes failure and observe potential negative transfer. The representation should not erase useful target structure merely to confuse site labels. Compare subgroup calibration and explanations before deployment. Domain invariance is a tool, not an absolute objective.
Practice case 6: Follow-up shift
Setup: Align censoring and horizons.
Task: Evaluate mature target outcomes.
Checking target: Sensor alignment cannot fix labels.
Align source and target follow-up horizons and censoring rules before comparing models. A target fleet with shorter observation can look safer because fewer failures have matured. Use censoring-aware metrics at supported horizons and report event counts. Sensor alignment does not repair outcome-data shift. Delay claims until target labels are mature enough for the intended horizon.
Practice case 7: Fleet holdout
Setup: Rotate an unseen target fleet.
Task: Measure negative transfer.
Checking target: Allow rejection and fallback.
Hold out an entire fleet as a target and repeat across fleets. Record when transfer helps, does nothing or harms. Analyse overlap, design differences and label quality for each case. A method that succeeds on one convenient target is not a general transfer system. Include an abstention or target-only fallback when similarity is weak.
Practice case 8: Dashboard
Setup: Monitor shift and calibration.
Task: Link alerts to actions.
Checking target: State narrow validated scope.
Create a deployment dashboard with domain distances, feature support, target calibration, missingness, prediction range and negative-transfer comparison. Link alerts to actions: recalibrate, fine-tune, narrow use, collect labels or roll back. Preserve source and target versions. The final claim should state exactly which fleet, horizon and adaptation were validated rather than saying the model is universally transferable.
Practice case 9: Unit contract
Setup: Mix Celsius and Fahrenheit sources.
Task: Detect and repair unit mismatch.
Checking target: Semantic alignment comes before representation alignment.
Convert units using documented sensor metadata and compare distributions before and after. A neural network may absorb the mismatch in training yet fail when one pipeline changes. Add automated range and unit checks. Feature names alone do not establish comparability across fleets.
Practice case 10: Overlap
Setup: Plot source and target operating ranges.
Task: Identify unsupported target regions.
Checking target: No algorithm transfers from absent evidence.
Estimate nearest-source distance or density ratios and mark target cases outside support. Evaluate performance separately there and implement abstention. Alignment may move embeddings together visually while physical operating conditions remain novel. Collect target labels or narrow deployment rather than hiding extrapolation.
Practice case 11: Negative transfer
Setup: Compare adapted and target-only models.
Task: Calculate the performance difference.
Checking target: Transfer must be allowed to fail.
Use repeated target splits and uncertainty. If adaptation worsens proper score or calibration, call it negative transfer even if source performance is strong. Investigate mechanism mismatch, over-alignment and noisy source labels. Keep a target-only fallback and publish where transfer was rejected.
Practice case 12: Fine-tuning depth
Setup: Freeze different layer blocks.
Task: Compare target validation and forgetting.
Checking target: More adaptation is not always better.
Early layers may encode general signal shapes while later layers capture source-specific decisions, but this is an empirical claim. Test several freeze patterns inside target development data. Re-evaluate all served domains after updating. A shared model that forgets the source may require domain-specific heads or separate deployment.
Practice case 13: Calibration only
Setup: Keep source ranking and recalibrate target probabilities.
Task: Compare with full adaptation.
Checking target: Baseline shift may not require representation change.
If target discrimination remains strong but probabilities are offset, a target calibration layer can be simpler and safer than fine-tuning. Validate on later target outcomes. If ranking also fails, calibration cannot repair ordering. This comparison separates baseline-risk shift from deeper concept shift.
Practice case 14: Simulation gap
Setup: Train on simulated faults and test real signals.
Task: Audit mismatched mechanisms.
Checking target: Synthetic volume does not guarantee realism.
Randomise plausible loads, noise and component parameters, then compare feature and outcome distributions with real data. Hold out real failures for final assessment. When simulation misses a mechanism, adaptation can align superficial statistics and retain wrong physics. Use experiments and domain review to improve the simulator or narrow claims.