Measurement Analytics, AI, and Decision Support

Introduction and Context

Electrical measurement has traditionally focused on obtaining a reliable value of voltage, current, power, energy, frequency, temperature, vibration, insulation resistance, or another engineering variable. Modern electrical systems, however, generate measurements continuously through smart meters, phasor measurement units, intelligent electronic devices, digital oscilloscopes, protection relays, distributed sensors, and supervisory control systems. The resulting challenge is no longer only how to measure, but also how to convert large, heterogeneous, uncertain data sets into trustworthy engineering decisions.

Measurement analytics is the systematic use of statistical, physical, computational, and domain-specific methods to transform measurement data into information and actionable knowledge. Artificial intelligence (AI) extends this capability through algorithms that learn patterns, classify operating states, estimate unmeasured quantities, detect anomalies, and support decisions. Decision support is the final stage: evidence is presented in a form that enables an engineer or operator to select an appropriate action while retaining accountability for safety and performance.

In electrical engineering, analytics must respect physical laws, instrument characteristics, uncertainty, sampling constraints, cybersecurity requirements, and the consequences of incorrect decisions. A highly accurate predictive model is not sufficient if its input sensors are poorly calibrated, its training data do not represent the operating environment, or its output cannot be explained to a protection engineer.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Explain the measurement-to-decision lifecycle and distinguish data, information, inference, prediction, and decision.
2. Characterize electrical measurement data, including uncertainty, sampling, missing values, outliers, and drift.
3. Apply descriptive, diagnostic, predictive, and prescriptive analytics to instrumentation problems.
4. Compare statistical, physical-model, machine-learning, and hybrid approaches.
5. Formulate and interpret common performance measures for detection, estimation, and classification.
6. Design an auditable AI-enabled measurement system with suitable validation and governance.
7. Evaluate a practical condition-monitoring and decision-support application.

1. The Measurement-to-Decision Lifecycle

A useful analytical architecture contains six linked stages:

1. Sensing and acquisition: Physical variables are converted into electrical signals and digitized.
2. Data quality assurance: Time alignment, calibration status, range checks, missing-data handling, and plausibility checks are applied.
3. Feature construction: Raw waveforms or time series are transformed into informative quantities such as RMS current, harmonic distortion, temperature gradients, or sequence components.
4. Inference and prediction: Models estimate states, identify faults, forecast demand, or detect abnormal behaviour.
5. Decision support: Results are compared with operating limits, costs, risks, and available actions.
6. Feedback and learning: The outcome of an intervention is recorded and used to improve the instrument, model, or operating policy.

The distinction between an observation and a decision is fundamental. A sensor may report a high winding temperature; analytics may infer that cooling performance is deteriorating; a decision-support system may recommend load reduction or inspection. These are different statements with different uncertainty and responsibility.

1.1 Data, information, knowledge, and action

• Data: Recorded numerical or categorical observations, including timestamps and metadata.
• Information: Data organized to reveal trends, relationships, or operating states.
• Knowledge: Validated interpretation based on physical understanding, historical evidence, and models.
• Decision: A selected action under stated objectives, constraints, and risk tolerance.

Measurement analytics should preserve the chain of evidence. Each result should be traceable to the data used, preprocessing applied, model version, calibration records, and decision rule.

2. Measurement Data in Electrical Systems

Electrical measurement data are commonly classified by their structure.

2.1 Sampling and aliasing

If a signal contains frequency components up to f_max, the ideal uniform sampling frequency must satisfy the Nyquist condition

where f_s is the sampling frequency. This condition is necessary but not always sufficient in practice. An anti-aliasing filter is normally required because real signals contain noise and components beyond the intended bandwidth. Transients may require a substantially higher sampling rate than steady-state measurements.

For a sinusoidal signal x(t) = A sin(2πft + φ), A is amplitude, f is frequency in hertz, and φ is phase angle. Inadequate sampling can change the apparent amplitude and frequency and may cause a high-frequency disturbance to appear as a false low-frequency component.

2.2 Uncertainty and data quality

A measured result should be regarded as an estimate rather than an exact value. A simple measurement model is

where y is the recorded value, x is the measurand, b is systematic bias, and ε is random error. Bias may arise from calibration error, sensor loading, temperature dependence, or aging. Random error may arise from electrical noise, quantization, and repeatability limitations.

For independent input quantities x₁, x₂, ..., x_n and output y = f(x₁, x₂, ..., x_n), the first-order combined standard uncertainty is commonly approximated by

where u(xᵢ) is the standard uncertainty of input xᵢ. This expression assumes small uncertainties, a locally linear model, and independent inputs. If inputs are correlated, covariance terms must be included:

where u(xᵢ,xⱼ) is the covariance. Uncertainty budgets are especially important when an AI system uses measurements close to a regulatory or protection threshold.

2.3 Missing values, outliers, and drift

Missing data should not be silently replaced. The reason for missingness—communication failure, sensor saturation, maintenance, or a real zero—affects the correct treatment. Interpolation may be acceptable for a short gap in a slowly varying temperature series, but it is unsafe for a protection waveform or a rapidly changing fault event.

An outlier is not necessarily an error. It may represent a genuine switching transient or incipient fault. Robust procedures therefore combine statistical tests with engineering limits and event records. Sensor drift is a gradual change in response. It can be identified by reference checks, redundant sensors, comparison with a physical model, or recalibration history.

3. Layers of Measurement Analytics

3.1 Descriptive analytics

Descriptive analytics answers: What happened? Examples include mean load, maximum demand, voltage-distribution plots, harmonic spectra, and event counts. For observations x₁, x₂, ..., x_n, the arithmetic mean is

The sample variance is

The denominator n−1 provides an unbiased estimator of population variance under the usual independent, identically distributed assumptions. For time-series data with autocorrelation, effective information is less than the sample count, so uncertainty estimates require care.

For an AC waveform with N equally spaced samples, the discrete RMS estimate is

This formula estimates the RMS over the selected window. The window must be appropriate to the waveform and system frequency; a nonintegral-cycle window can bias harmonic and RMS estimates.

3.2 Diagnostic analytics

Diagnostic analytics answers: Why did it happen? Correlation, regression, spectral analysis, event sequence analysis, and fault-tree reasoning can identify contributing factors. Correlation alone does not establish causation. For example, transformer temperature and load may be correlated because both rise during high-demand periods, but ambient temperature and cooling-fan status may also be causal variables.

A linear regression model can be written as

where y is the response, x_j are explanatory variables, β_j are coefficients, and ε is unexplained error. The model is useful only when its assumptions—appropriate functional form, reasonably independent errors, and adequate data coverage—are examined.

3.3 Predictive analytics

Predictive analytics estimates an unknown, future, or unobserved quantity. Applications include feeder-load forecasting, remaining useful life estimation, fault classification, and virtual sensing. A forecast should include uncertainty, not only a point value. A prediction interval describes a range expected to contain a future observation at a stated confidence level, subject to model assumptions.

3.4 Prescriptive analytics and decision support

Prescriptive analytics evaluates possible actions. A simple decision rule may compare predicted risk with an intervention cost. More advanced systems solve constrained optimization problems. A generic formulation is

Here a is an action vector, J(a) is expected cost, C_energy and C_maintenance are operating and maintenance costs, C_failure is failure consequence, P_failure(a) is action-dependent failure probability, and g and h represent inequality and equality constraints. The formulation is meaningful only when costs, probabilities, and constraints are defensible. Optimization does not remove engineering judgement.

4. Artificial Intelligence for Measurement Systems

4.1 AI problem categories

AI should be selected after defining the measurement objective and failure consequences. A complex model is not automatically better than a calibrated physical model or a threshold rule.

4.2 Supervised and unsupervised learning

In supervised learning, a model is trained using input-output pairs. A regression loss frequently used in training is mean squared error:

where yᵢ is the observed value and ŷᵢ is the prediction. Root mean square error is

and has the same units as the response. Mean absolute error is

MAE is less sensitive to large errors than MSE. Metrics must be interpreted in relation to engineering tolerances; a numerically small error may still be unacceptable near a protection boundary.

In classification, a confusion matrix distinguishes true positives, true negatives, false positives, and false negatives. Precision and recall are

where TP, FP, and FN have their usual meanings. Accuracy can be misleading for rare faults. A classifier that labels every sample as “healthy” may have high accuracy but zero useful fault recall.

4.3 Feature engineering for electrical measurements

Features should reflect both signal behaviour and physical meaning. Examples include:

• RMS, peak, crest factor, and waveform asymmetry;
• fundamental magnitude and phase;
• total harmonic distortion;
• negative- and zero-sequence components;
• rate of temperature rise;
• moving averages and variability;
• spectral energy in selected frequency bands;
• breaker-operation timing and contact resistance;
• ratios such as load current to rated current.

For a voltage waveform with fundamental RMS component V₁ and harmonic RMS components V₂, V₃, ..., V_n, total harmonic distortion is commonly defined as

The harmonic range and measurement bandwidth must be stated. Different standards and instruments may use different reporting conventions.

4.4 State estimation and the Kalman-filter idea

Many electrical states cannot be measured directly or economically. State estimation combines a dynamic model with noisy measurements. A linear discrete model is

where xₖ is the state, uₖ is a known input, zₖ is the measurement, A, B, and H are model matrices, and wₖ and vₖ represent process and measurement noise. A Kalman filter recursively predicts the state and corrects it using the new measurement. Its classical optimality depends on linear dynamics, unbiased Gaussian noise, known covariance matrices, and an adequately specified model. Extended and unscented variants address certain nonlinear cases but do not eliminate modelling limitations.

4.5 Physics-informed and hybrid models

Purely data-driven models may fail when operating conditions differ from training data. Physics-informed models constrain learning using circuit equations, energy balances, thermal laws, or known protection logic. A hybrid system may use a physical transformer thermal model to generate a baseline and an AI model to learn residual error:

The residual model r_AI should be trained only where the physical model is known to be incomplete. Hybrid methods can improve extrapolation, interpretability, and fault analysis, although they require careful integration of model uncertainties.

5. Validation, Explainability, and Trust

5.1 Correct data partitioning

Randomly splitting adjacent time-series samples can produce leakage: nearly identical samples from one event may appear in both training and test sets. A stronger design separates data chronologically, by asset, by site, or by operating regime. The test set should represent future deployment conditions and should remain untouched until final evaluation.

Validation should test:

• accuracy and calibration;
• robustness to noise, missing data, and sensor drift;
• performance across loads, seasons, and equipment types;
• false-alarm and missed-event costs;
• latency and computational requirements;
• behaviour outside the training domain.

5.2 Explainability

An operator needs to know why an alarm was issued, which measurements contributed, how certain the result is, and what action is recommended. Explanations may include feature importance, similar historical events, residual plots, confidence intervals, physical constraints, and counterfactual statements such as “the alarm would disappear if the temperature trend returned below the specified rate.” Explanations should not be treated as proof of causality; some post-hoc explanations approximate model behaviour rather than reveal its internal reasoning.

5.3 Calibration and uncertainty

A probability estimate should be calibrated: among cases assigned probability 0.8, approximately 80% should occur under comparable conditions. Confidence in an AI output should reduce when sensors fail, input variables leave their known range, or the operating regime changes. Out-of-distribution detection and abstention—referring a case to a human rather than forcing a prediction—are valuable safety mechanisms.

6. Decision-Support Architecture

A practical architecture contains:

1. sensors and instrument interfaces;
2. time synchronization and data acquisition;
3. edge preprocessing and local protection logic;
4. a secure historian or data platform;
5. analytics and model services;
6. visualization, alarm management, and workflow integration;
7. human approval and maintenance feedback.

Edge analytics reduces communication latency and bandwidth and can continue operating during loss of a central connection. Cloud or central analytics offers greater computational and fleet-level learning capability. Safety-critical protection should not depend solely on a remote AI service; deterministic protection functions and fail-safe states must remain available.

Decision interfaces should distinguish:

• measured value from estimated value;
• alarm from diagnosis;
• recommendation from command;
• model confidence from physical certainty;
• advisory action from automatically executed action.

Alarm rationalization is essential. Excessive alarms create alarm flooding and encourage operators to ignore warnings. Each alarm should have a defined consequence, priority, response time, responsible role, and rational set point.

7. Worked Example: Anomaly Detection in a Three-Phase Motor

Consider an illustrative motor-monitoring system measuring phase current, winding temperature, vibration, and supply voltage. The goal is to identify developing mechanical or electrical faults early enough to schedule inspection.

The monitoring window contains 60 one-minute observations. The measured RMS current has mean 42 A and sample standard deviation 2 A during a stable baseline. A new five-minute sequence is 47, 48, 49, 50, and 51 A. The mean of the sequence is

The baseline mean is 42 A, so the shift is 7 A. If the baseline is approximately stable and normally distributed, a rough standardized shift is

This is a strong statistical indication of abnormality, but it is not by itself a diagnosis. The current increase could result from increased mechanical load, low voltage, phase imbalance, bearing friction, or a sensor problem.

The system therefore checks correlated features. Suppose temperature is rising at 0.18 °C/min, vibration has increased by 40%, and supply voltage has remained within its normal range. A rule-based layer issues an “inspection recommended” advisory rather than an immediate trip. A classifier trained on labelled historical events estimates:

• mechanical-load anomaly: 0.72;
• electrical supply anomaly: 0.11;
• sensor fault: 0.17.

Because the probabilities do not prove causation, the operator is shown the contributing trends and asked to verify load and sensor status. Maintenance inspection confirms increased bearing friction. The motor is serviced during a planned outage rather than after failure.

The example illustrates several principles: statistical deviation detects a change; multichannel evidence improves diagnosis; model probabilities support but do not replace engineering verification; and the recommended action must reflect consequence, urgency, and uncertainty.

8. Detailed Applied Case: Illustrative Distribution-Substation Monitoring

Background

A distribution utility operates oil-immersed transformers equipped with top-oil temperature, load current, ambient temperature, dissolved-gas analysis at periodic intervals, and circuit-breaker event records. Unexpected transformer failures cause customer interruption and expensive emergency replacement.

Problem

A threshold-only maintenance policy produces two problems. Conservative thresholds generate unnecessary inspections, while slow-developing faults may remain below a single limit until damage is advanced. The utility seeks a decision-support system for inspection prioritization without allowing AI to override protection functions.

Analysis

Engineers first establish data quality. Timestamps are synchronized, sensor ranges are checked, calibration records are linked to each channel, and periods of known maintenance are labelled. The dataset is divided by transformer and by calendar period so that the test set represents unseen equipment and future operation.

A thermal baseline estimates expected top-oil temperature from load and ambient temperature. The residual is

where T_measured is observed top-oil temperature and T_expected is the model estimate. A persistent positive residual suggests abnormal heat transfer, cooling degradation, or a changed thermal condition. The analytics layer combines residual persistence, loading, gas trends, and breaker events.

Approach and solution

The deployed design uses three levels:

1. Deterministic safeguards: Existing protection relays, trip logic, and statutory limits remain independent of the AI service.
2. Hybrid analytics: A thermal model produces expected temperature; a statistical detector identifies persistent residuals; a supervised model ranks likely causes when labelled events are available.
3. Human-centred workflow: The dashboard displays evidence, confidence, recommended inspection priority, and data-quality warnings. Automatic control is not permitted for uncertain diagnoses.

A risk score can be expressed conceptually as

where P(F | D) is the estimated probability of failure F given data D, and C(F) is the consequence of failure. The score is useful for prioritization only when probabilities and consequences are defined consistently. It is not a universal physical quantity and should not be compared across assets without a common scale.

Outcome

The utility reports fewer unnecessary inspections, earlier identification of cooling-system degradation, and improved maintenance scheduling. The main benefit is not the AI model alone; it is the integration of calibrated measurements, physical expectations, transparent evidence, and a controlled work process. Periodic review shows that performance declines when a new sensor type is introduced, leading to retraining and model-version control.

Learning points

• Asset-specific baselines are often more informative than universal thresholds.
• Data provenance and calibration status are part of the model input.
• Risk-based prioritization must include consequences, not only predicted probability.
• Protection and safety functions require independent fail-safe mechanisms.
• A model must be monitored after deployment for drift and changing operating conditions.

9. Implementation Methodology and Best Practices

A robust project should proceed in the following order:

1. Define the engineering decision, not merely the available data.
2. Specify the measurand, accuracy, latency, sampling rate, and acceptable uncertainty.
3. Audit sensors, calibration, synchronization, communication, and metadata.
4. Establish a simple physical or statistical baseline.
5. Select features and algorithms appropriate to the decision consequence.
6. Separate training, validation, and future test data without leakage.
7. Evaluate false alarms, missed events, latency, robustness, and cost.
8. Conduct pilot deployment with human review and documented overrides.
9. Monitor data drift, model drift, cybersecurity events, and maintenance outcomes.
10. Retire or retrain models under formal change control.

A baseline is indispensable. If a sophisticated AI system does not improve a transparent threshold, regression, or physical model under realistic cost-weighted evaluation, its complexity is not justified.

10. Challenges, Ethics, Governance, and Regulation

Technical limitations

Measurement analytics is limited by sensor accuracy, unrepresentative training data, nonstationary loads, rare fault labels, communication loss, computational latency, and changing equipment. Fault data are often scarce because healthy operation is common and failures are costly. Synthetic data can help, but simulated faults may not reproduce real sensor behaviour.

Cybersecurity

Connected instruments increase the attack surface. Threats include spoofed measurements, replay attacks, denial of service, unauthorized model changes, and poisoned training data. Controls should include authenticated devices, least-privilege access, encryption, network segmentation, secure update procedures, immutable audit logs, and independent plausibility checks. A cybersecurity alarm should not be confused with an equipment fault.

Ethical and professional responsibility

AI may influence personnel safety, service continuity, and financial decisions. Engineers remain responsible for defining safe operating envelopes and escalation procedures. Systems should support non-discrimination where human work allocation is affected, protect operational data, document limitations, and provide a means to challenge or override an automated recommendation. Automatic action requires a higher assurance case than advisory analytics.

Standards and conformity

Projects should align with applicable measurement, functional-safety, power-system, cybersecurity, data-protection, and AI-governance requirements. The exact standard set depends on the equipment, jurisdiction, and safety classification. Compliance cannot be inferred merely because a model has high predictive accuracy.

11. Contemporary Developments

Important developments include edge AI for low-latency monitoring, federated learning across utilities without centralizing raw data, digital twins combining physical and data-driven models, self-supervised learning for systems with few fault labels, and uncertainty-aware neural models. Foundation models for time series are being investigated, but their usefulness in safety-critical instrumentation depends on traceability, domain adaptation, and rigorous validation.

Digital twins are particularly relevant to electrical measurement. A twin may combine asset geometry, operating history, sensor data, and simulation. Its value depends on synchronization with the physical asset and explicit treatment of model discrepancy. Federated learning can reduce data-sharing barriers, but differences in sensor calibration and operating practice can make local data statistically non-equivalent.

Chapter Summary

Measurement analytics transforms instrument output into evidence for engineering action. Its lifecycle begins with sensing and data-quality assurance, continues through feature construction and inference, and ends with a controlled decision and feedback loop. Sampling, calibration, uncertainty, missing data, outliers, and drift are foundational concerns rather than preprocessing details.

Descriptive, diagnostic, predictive, and prescriptive analytics answer progressively more demanding questions. AI methods support classification, regression, anomaly detection, forecasting, and state estimation, but their value depends on representative data, proper validation, explainability, and physical consistency. Hybrid physics-informed approaches are often attractive in electrical systems because they combine domain knowledge with the ability to learn residual behaviour.

Decision-support systems should distinguish measurement from inference and recommendation from automatic control. Safety-critical protection must retain independent deterministic safeguards. Effective deployment requires cybersecurity, model governance, calibration traceability, drift monitoring, human oversight, and cost- or risk-based evaluation. The central professional principle is that AI should strengthen measurement-based engineering judgement, not obscure or replace it.

Key Terms

Measurement analytics: Systematic transformation of measurement data into information and decisions.

Measurand: Quantity intended to be measured.

Measurement uncertainty: Quantified doubt associated with a measurement result.

Data provenance: Record of data origin, processing, timing, and custody.

Aliasing: Misrepresentation of frequency content caused by inadequate sampling.

Feature: Derived variable used to represent informative signal behaviour.

Descriptive analytics: Analysis of what has occurred.

Diagnostic analytics: Analysis of causes or contributing factors.

Predictive analytics: Estimation of future, unknown, or unmeasured outcomes.

Prescriptive analytics: Evaluation or recommendation of actions.

Supervised learning: Learning from labelled input-output examples.

Unsupervised learning: Learning structure from primarily unlabelled data.

Anomaly detection: Identification of observations that differ from expected behaviour.

State estimation: Inference of hidden system states from models and measurements.

Data leakage: Unintended use of test or future information during model development.

Model drift: Degradation caused by changes in data or operating conditions.

Explainability: Ability to communicate the basis and limitations of a model output.

Digital twin: Dynamic digital representation linked to a physical asset.

Decision support: Presentation of evidence and recommendations for human or controlled action.

Review Questions

Short Answer Questions

1. Distinguish measurement data, information, knowledge, and decision.
2. State the Nyquist sampling condition and explain why anti-aliasing filtering is used.
3. What is the difference between a false positive and a false negative in fault classification?
4. Why can random train-test splitting be inappropriate for time-series measurements?
5. Name four data-quality problems that can affect an AI-enabled instrument system.

Long Answer Questions

1. Explain the complete measurement-to-decision lifecycle for a smart electrical substation.
2. Discuss uncertainty propagation and its importance when an analytical result is compared with a protection or regulatory limit.
3. Compare descriptive, diagnostic, predictive, and prescriptive analytics using electrical-engineering examples.
4. Explain the advantages and limitations of physics-informed and purely data-driven models.
5. Design a governance and cybersecurity framework for an AI-based condition-monitoring system.

Analytical / Application Questions

1. A current sensor has a baseline mean of 30 A and standard deviation of 1.5 A. A new five-sample mean is 34.5 A. Calculate the standardized shift and explain why it is not, by itself, a fault diagnosis.
2. Develop a validation plan for a classifier detecting transformer cooling-system faults when only a small number of labelled failures are available.
3. For a distribution transformer, propose a decision-support workflow that combines load, ambient temperature, top-oil temperature, and maintenance cost. Identify where human approval and independent safety functions should remain.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes predictive analytics from prescriptive analytics?

A. Predictive analytics estimates what may occur, while prescriptive analytics evaluates possible actions.

B. Predictive analytics calibrates sensors, while prescriptive analytics samples waveforms.

C. Predictive analytics removes all uncertainty, while prescriptive analytics measures voltage.

D. Predictive analytics applies only to historical data, while prescriptive analytics applies only to analog instruments.

Correct Answer: A. Predictive analytics estimates what may occur, while prescriptive analytics evaluates possible actions.

Question 2

If the highest relevant signal frequency is 5 kHz, which sampling frequency satisfies the ideal Nyquist condition?

A. 5 kHz

B. 8 kHz

C. 10 kHz exactly

D. 12 kHz

Correct Answer: D. 12 kHz

Question 3

For observed values 2, 4, and 6, what is the arithmetic mean?

A. 3

B. 4

C. 5

D. 6

Correct Answer: B. 4

Question 4

Which metric is especially useful for assessing the proportion of actual fault cases detected by a classifier?

A. Recall

B. Specific energy

C. Sampling frequency

D. Mean operating voltage

Correct Answer: A. Recall

Question 5

Why can accuracy be misleading for rare-fault classification?

A. Accuracy has units of amperes.

B. A classifier can obtain high accuracy by labelling nearly all cases as healthy.

C. Accuracy cannot be calculated from a confusion matrix.

D. Accuracy always penalizes false positives more than false negatives.

Correct Answer: B. A classifier can obtain high accuracy by labelling nearly all cases as healthy.

Question 6

In a measurement model y = x + b + ε, what does b represent?

A. Sampling frequency

B. Random error only

C. Systematic bias

D. The measured physical quantity

Correct Answer: C. Systematic bias

Question 7

Which practice most directly helps prevent time-series data leakage?

A. Randomly duplicate all samples

B. Use chronological or asset-based separation of training and test data

C. Remove timestamps before splitting

D. Train and test on the same operating event

Correct Answer: B. Use chronological or asset-based separation of training and test data

Question 8

For a waveform with fundamental RMS voltage 100 V and second-harmonic RMS voltage 5 V only, the stated THD is approximately:

A. 0.05%

B. 5%

C. 20%

D. 105%

Correct Answer: B. 5%

Question 9

What is a principal advantage of a hybrid physics-informed model?

A. It guarantees zero measurement noise.

B. It combines physical knowledge with data-driven correction or inference.

C. It eliminates the need for calibration.

D. It permits protection systems to ignore deterministic safeguards.

Correct Answer: B. It combines physical knowledge with data-driven correction or inference.

Question 10

Which action is most appropriate when an AI model receives inputs far outside its training domain?

A. Force the highest-confidence class

B. Disable all physical protection permanently

C. Use an out-of-distribution warning or refer the case for human review

D. Delete the input without recording the event

Correct Answer: C. Use an out-of-distribution warning or refer the case for human review