13. Reliability Analytics, AI and Decision Support

Introduction and Context

Maintenance management has progressively moved from repairing failed equipment to managing the probability, consequences, and economics of failure. Reliability analytics represents the next stage in this development. It uses operational, maintenance, process, and contextual data to understand how assets perform, identify changing risk, anticipate failures, and support better decisions about inspection, intervention, redesign, spares, and capital investment.

The availability of computerized maintenance management systems (CMMS), enterprise asset management (EAM) platforms, historians, sensors, industrial Internet of Things (IIoT) devices, and cloud computing has greatly expanded the volume of information available to maintenance organizations. Artificial intelligence (AI) and machine learning can detect patterns that are difficult to identify through manual analysis. However, sophisticated algorithms do not automatically produce reliable decisions. Data may be incomplete, sensors may be poorly calibrated, failure labels may be inconsistent, and a statistically accurate model may still recommend an economically or operationally inappropriate action.

Reliability analytics must therefore be understood as a managerial and engineering discipline rather than merely a technical exercise. Its purpose is to connect evidence with decisions. The central question is not simply, “Can the model predict failure?” but also, “What should the organization do, when should it act, what will the action cost, and what risk will remain?”

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the role of reliability analytics in maintenance and production management.

2. Distinguish descriptive, diagnostic, predictive, and prescriptive analytics.

3. Interpret reliability trends, survival curves, hazard rates, and common performance indicators.

4. Design a data-to-decision workflow for predictive maintenance.

5. Compare statistical reliability methods with machine-learning approaches.

6. Evaluate dashboards and decision-support systems using operational and economic criteria.

7. Identify problems involving data quality, explainability, cybersecurity, privacy, and model drift.

8. Recommend responsible governance practices for AI-enabled maintenance.

From Maintenance Records to Reliability Intelligence

Reliability analytics begins with the conversion of operational observations into information that can support action. A useful analytical system normally integrates several data sources: asset registers, work orders, failure codes, inspection findings, sensor readings, production conditions, environmental variables, operator reports, and parts or labor costs. These sources often use different asset identifiers and time conventions. Data integration is consequently an engineering problem as well as an information-technology problem.

Four forms of analytics are especially relevant:

• Descriptive analytics explains what has happened. Examples include mean time between failures, equipment availability, backlog age, and monthly maintenance cost.
• Diagnostic analytics investigates why an event occurred. It may compare failure modes, operating conditions, maintenance practices, or supplier batches.
• Predictive analytics estimates what is likely to happen. Examples include a probability of failure within thirty days or an estimate of remaining useful life (RUL).
• Prescriptive analytics recommends what should be done. It may prioritize inspections, schedule a planned intervention, alter operating conditions, or hold a particular spare part.

These categories are connected. A predictive model built on poorly understood historical failures can mislead decision-makers, while a prescriptive recommendation requires knowledge of constraints such as production schedules, technician availability, safety rules, and spare-parts lead time.

Important terminology includes the following. An asset is an item of equipment, system, or component whose performance is managed. A failure event is the loss of required function. A failure mode describes how the failure occurs, while a failure mechanism describes the physical or chemical process causing it. A failure label is the coded representation used in a database. Censoring occurs when an asset has not failed by the end of the observation period or leaves observation for another reason. RUL is the estimated operating time remaining before a defined failure or performance limit. A feature is an input variable used by an analytical model, such as vibration amplitude, discharge pressure, load, or operating hours.

Measuring Reliability Through Time

Traditional reliability measures remain essential because advanced analytics should complement, not replace, sound reliability engineering. For a non-repairable item, the reliability function is

R(t) = P(T > t)

where T is the random time to failure. The associated failure distribution is F(t) = 1 − R(t). The probability density function describes the distribution of failure times, while the hazard function represents the instantaneous failure rate conditional on survival until time t:

h(t) = f(t) / R(t)

The hazard function is particularly useful for understanding whether risk is increasing, decreasing, or approximately constant. The familiar “bathtub curve” depicts early-life failures, a period of relatively stable random failures, and a wear-out period in which failure risk increases. It is a conceptual model, not a universal law. Many modern electronic and mechanical systems do not follow a simple bathtub pattern, and different failure modes of the same asset may have different hazard behavior.

For repairable assets, mean time between failures (MTBF), mean time to repair (MTTR), availability, and failure rate are frequently used. In a simplified steady-state setting, inherent availability may be approximated by:

A = MTBF / (MTBF + MTTR)

This relationship must be interpreted carefully. MTBF can improve because failures are genuinely reduced, because minor failures are not recorded, or because operating hours are calculated incorrectly. Similarly, a high availability figure can conceal serious production consequences if a small number of failures are extremely disruptive.

Trend Analysis

Trend analysis examines how indicators change across time, assets, locations, or operating conditions. Useful measures include failures per operating hour, emergency work percentage, repeat failures, overdue preventive tasks, bad-actor concentration, inspection compliance, and cost per unit produced. Rates should generally be normalized by exposure. Comparing two machines using raw failure counts is misleading if one operates twice as many hours or under substantially heavier loads.

A trend may be represented using moving averages, control charts, regression lines, seasonal decomposition, or cumulative plots. The analyst should distinguish a real change in reliability from changes in reporting practices, production volume, maintenance policy, or asset population. A sudden increase in recorded failures may indicate deterioration, improved reporting, a revised failure code, or more sensitive detection.

Segmentation is often more informative than an overall average. Reliability should be examined by asset family, manufacturer, age, duty cycle, failure mode, shift, product grade, and environmental condition where such variables are relevant. Pareto analysis can identify a small number of failure modes or assets that account for a large share of risk, but the “vital few” should be defined using business impact rather than frequency alone.

Survival Analysis and Reliability Distributions

Survival analysis is valuable when assets have different ages and when many assets have not yet failed. Simply averaging observed failure times can produce biased results because it ignores censored observations. The Kaplan–Meier estimator provides a non-parametric estimate of the survival function and incorporates both failures and censored records. It can be used to compare survival among equipment groups, provided that the groups and observation process are appropriately defined.

Parametric models such as exponential, Weibull, lognormal, and gamma distributions are also widely used. The exponential model assumes a constant hazard rate and is useful in limited circumstances, but it is often too restrictive for aging equipment. The Weibull model is flexible because its shape parameter can represent decreasing, constant, or increasing hazard. A Weibull shape parameter greater than one generally indicates increasing failure risk, a value near one suggests approximately constant risk, and a value below one suggests decreasing risk. These interpretations concern the fitted failure population and should not be treated as proof of a physical mechanism.

Survival curves support decisions such as warranty analysis, inspection interval design, replacement policy, and comparison of maintenance strategies. They do not, by themselves, establish causation. If one group has better survival, the difference may result from load, operating discipline, maintenance quality, or selection effects rather than equipment design alone.

Predictive Models for Maintenance

Predictive maintenance models estimate a future condition, event, or time horizon. The target may be binary, such as failure within fourteen days; continuous, such as bearing temperature; or time-to-event, such as RUL. The appropriate model depends on the decision, the data structure, and the cost of errors.

Statistical approaches include logistic regression for event probabilities, linear or generalized linear models for measurable outcomes, proportional-hazards models for time-to-event analysis, and state-space or autoregressive models for time-series behavior. These methods often provide useful interpretability and can perform well when relationships are reasonably structured.

Machine-learning approaches include decision trees, random forests, gradient-boosted trees, support-vector machines, neural networks, and autoencoders. Classification models can identify high-risk cases; regression models can estimate a numerical quantity; anomaly-detection methods can identify behavior that differs from a learned normal state. Deep learning can be effective for high-frequency signals such as vibration or acoustic emissions, but its benefits depend on sufficient, representative, and well-labelled data.

A sound modelling process includes:

1. Defining the decision and prediction horizon.

2. Establishing the asset population and unit of analysis.

3. Creating reliable failure and maintenance labels.

4. Aligning sensor readings with operating and maintenance events.

5. Removing leakage, in which information unavailable at prediction time enters the model.

6. Splitting data chronologically or by asset to represent real deployment.

7. Evaluating discrimination, calibration, false alarms, missed failures, and business impact.

8. Piloting the model with human review before broad deployment.

Accuracy alone is inadequate. A model with high accuracy may be useless when failures are rare. Precision measures the proportion of predicted failures that actually fail; recall measures the proportion of actual failures that are detected. A precision–recall curve is often more informative than accuracy in an imbalanced maintenance dataset. Calibration asks whether predictions stated as probabilities correspond to observed frequencies. A model predicting a 70 percent failure probability should be correct approximately 70 percent of the time within comparable cases, subject to sampling variation.

The costs of false positives and false negatives are usually unequal. A false positive may cause an unnecessary shutdown or inspection, whereas a false negative may cause a safety incident, major production loss, or environmental release. Thresholds should therefore be selected with engineering, safety, and economic consequences in mind.

Dashboards as Decision Interfaces

A reliability dashboard is not simply a collection of charts. It is an interface through which different users interpret risk and decide what to do. A plant manager may need production exposure, financial risk, and major bad actors; a reliability engineer may need failure-mode trends and Weibull parameters; a planner may need actionable work orders, lead times, and labor requirements.

An effective dashboard has a defined audience, purpose, refresh frequency, data owner, and escalation rule. It should distinguish leading indicators, such as inspection compliance and increasing vibration, from lagging indicators, such as failures and downtime. Visual design should show uncertainty, data freshness, thresholds, and missing-data conditions. A red status without an explanation or recommended response encourages either overreaction or disregard.

A practical dashboard may combine:

• asset health and risk ranking;
• failure and downtime trends normalized by operating exposure;
• survival or hazard information for critical populations;
• predicted failure probabilities and confidence ranges;
• recommended actions, due dates, and responsible roles;
• spare-part and labor constraints; and
• realized outcomes, including avoided failures, false alarms, and model performance.

Dashboards should not become substitutes for field verification. A predicted bearing problem should lead to an appropriate inspection or test, not automatically to replacement unless the decision rule and consequences justify it.

Prescriptive Decision Support

Prediction becomes valuable only when connected to a decision. Prescriptive analytics evaluates alternative actions under constraints. For example, if three pumps have elevated risk but only one maintenance crew is available, the system must prioritize work using probability of failure, consequence, intervention cost, production schedule, safety implications, and confidence in the evidence.

A simple expected-loss formulation is:

Expected risk = Probability of failure × Consequence of failure

An intervention can be justified when its cost, including disruption, is lower than the expected avoidable loss, adjusted for uncertainty and risk tolerance. More advanced systems use optimization, simulation, Markov decision processes, mixed-integer programming, or reinforcement learning. These methods can schedule inspections and replacements while considering workforce, spares, access windows, and production commitments.

Prescriptive systems should present alternatives rather than conceal trade-offs. A recommendation might state: inspect during the next planned opportunity; replace immediately; continue operation with enhanced monitoring; or reduce load temporarily. Each option should show assumptions, expected cost, residual risk, and the evidence supporting the recommendation. Human decision-makers remain accountable for safety-critical choices.

Data Quality and the Reliability of the Analysis

The quality of an analytical result is constrained by the quality of the data-generating process. Common problems include duplicate assets, missing operating hours, inconsistent time stamps, vague failure codes, work orders closed without cause identification, sensor drift, unrecorded temporary repairs, and changes in maintenance policy. “Garbage in, garbage out” is incomplete: biased or selectively recorded data can produce convincing but systematically wrong conclusions.

Data governance should define ownership, naming standards, critical fields, validation rules, retention requirements, and correction procedures. Failure coding should distinguish symptoms, failure modes, causes, and consequences. Sensor data should include calibration status, units, sampling frequency, and quality flags. Analysts should preserve raw data and document transformations so that results can be audited.

Missing data require investigation rather than automatic imputation. A missing reading may indicate random communication loss, a failed sensor, a machine being offline, or deliberate removal of an alarming value. Each situation has different analytical implications. Similarly, maintenance interventions can change future data: a successful repair may reduce risk, while an ignored recommendation may create a selection bias in model evaluation.

Explainability, Governance, and Responsible AI

Explainability is the ability to communicate why a model produced an output and how that output should be used. In maintenance, explanations may include the variables contributing most strongly to a risk score, comparison with a normal operating baseline, the relevant time trend, and similar historical cases. Global explanations describe general model behavior; local explanations describe an individual asset or prediction.

Explanations should not be confused with proof of causality. A feature may be highly predictive because it is correlated with an unmeasured condition or because it records a consequence that appears shortly before failure. Engineering review is needed before changing operating practice or redesigning equipment.

Responsible AI requires clear accountability. Organizations should specify who approves models, who can override recommendations, how overrides are recorded, and how incidents are investigated. Safety-critical models should undergo validation, change control, and periodic review comparable to other engineered systems. Procurement contracts should address access to data, model documentation, cybersecurity responsibilities, audit rights, and termination or portability arrangements.

Privacy may arise when systems combine operator identity, location, work patterns, wearable devices, or performance records with equipment data. Data collection should be proportionate to a legitimate purpose, access should be controlled, retention should be limited, and employees should receive transparent explanations. An equipment-health system should not quietly become an individual-surveillance system.

Cybersecurity, Drift, and Operational Resilience

Connected maintenance systems increase the attack surface of operational technology. Threats include unauthorized access, manipulation of sensor readings, ransomware, malicious model inputs, denial of service, and compromise of vendor connections. A false normal signal may be more dangerous than an obvious system outage because it can delay intervention.

Controls include network segmentation, least-privilege access, multifactor authentication, secure remote access, patch and vulnerability management, encryption, logging, backup, incident response, and tested manual fallbacks. Model pipelines should be protected against poisoned training data and unauthorized changes. Cybersecurity must be coordinated with production safety because aggressive IT changes can themselves disrupt control systems.

Model drift occurs when the relationship between inputs and outcomes changes. Data drift concerns changes in the distribution of inputs; concept drift concerns changes in the relationship between inputs and failure. Drift may follow equipment replacement, process changes, seasonal conditions, software updates, sensor relocation, a new product mix, or a revised maintenance policy.

Monitoring should include input distributions, missingness, alert rates, calibration, precision, recall, false-alarm burden, and realized maintenance outcomes. A model should be retrained only after the cause of deterioration is understood. Retraining on unexamined data can reproduce a new problem. Version control, approval records, rollback capability, and champion–challenger testing support controlled evolution.

Illustrative Case: Predictive Support for a Process Pump Fleet

Background

The following is an illustrative case. A chemical-processing plant operates forty-eight centrifugal pumps. Unplanned pump failures cause production interruptions, but the plant has limited maintenance windows. The existing CMMS contains work orders and failure codes, while the process historian records pressure, flow, motor current, temperature, and vibration for selected pumps.

Situation and Problem

Management initially requested an AI system that would “predict pump failure.” A review found that failure codes were inconsistent: some work orders recorded the failed component, others recorded only the observed symptom, and several planned replacements were coded as failures. Pumps also operated under different flow regimes. A model trained without correcting these issues would likely generate unreliable alerts.

Analysis and Approach

The reliability team first defined the decision: identify pumps requiring inspection within the next fourteen days, not declare the exact failure date. It created a common asset hierarchy, reconciled historian tags with CMMS identifiers, and classified events as confirmed failure, planned intervention, inspection finding, or non-failure work. Operating hours and flow exposure were added to normalize comparisons.

A survival analysis showed that seal-related events increased with a combination of high temperature and unstable operating conditions, while bearing-related events were more closely associated with vibration and motor-current patterns. The team then developed a transparent gradient-boosted classification model and compared it with a logistic-regression baseline. Data were divided chronologically so that evaluation represented future use. The model was assessed using recall at a fixed inspection capacity, calibration, false alarms per pump-month, and estimated avoided downtime.

The decision-support layer ranked alerts using predicted failure probability, consequence category, inspection cost, production schedule, and spare availability. It did not automatically generate replacement work orders. Instead, it created an inspection recommendation showing the dominant signals, recent trend, data quality status, and suggested verification method.

Outcome

During a three-month pilot, technicians confirmed several developing defects and rejected some alerts caused by a faulty vibration sensor. The rejected alerts led to a new sensor-quality rule rather than immediate model retraining. The plant reduced emergency interventions during the pilot period, but the team avoided claiming that every improvement was caused by the model because production mix and maintenance practices also changed. The model was retained with monthly monitoring and quarterly engineering review.

Learning Points

The case illustrates that the principal work was not algorithm selection. Asset identification, label quality, decision definition, exposure normalization, human verification, and outcome measurement were equally important. It also demonstrates that an explainable recommendation with a clear inspection action may be more valuable than an opaque prediction of an exact failure time.

Common Failure Modes in Analytics Programmes

Organizations often begin with technology rather than a decision problem. They purchase sensors without defining what action a signal will trigger, or they create dashboards filled with indicators that no role owns. Other common mistakes include optimizing for model accuracy while ignoring consequence and cost, using random data splits that allow future information into training, treating maintenance completion as proof that a prediction was correct, and failing to distinguish an alert from a confirmed failure.

Additional risks include excessive alarm frequency, automation bias, overconfidence in small datasets, unrecognized changes in asset population, vendor lock-in, and the loss of engineering knowledge when experts are excluded from model design. A model can also create perverse incentives: if teams are judged only on the number of avoided failures, they may perform unnecessary interventions or classify uncertain events in ways that improve apparent performance.

Best practice is to begin with a critical decision, establish a baseline, improve data discipline, and introduce analytics in stages. A staged pathway may move from trusted descriptive measures to diagnostic studies, then to narrowly scoped predictive pilots and finally to prescriptive optimization. Each stage should have success criteria involving safety, reliability, cost, usability, and trust.

Current Developments and Emerging Directions

Several developments are shaping reliability analytics. Edge computing permits faster processing near machines and can reduce bandwidth and latency. Digital twins combine asset models, real-time data, and simulation, although their value depends on model validity and maintenance of the underlying representation. Physics-informed machine learning combines engineering relationships with data-driven methods and may improve performance when failure data are scarce.

Generative AI is increasingly used to search maintenance records, summarize inspection reports, assist troubleshooting, and translate technical questions into database queries. Its outputs require verification because language fluency does not guarantee factual accuracy. Federated learning and privacy-preserving methods may allow organizations to learn from distributed data without centralizing all raw records, but they introduce additional governance and technical complexity.

The likely direction is not the complete replacement of reliability engineers. Rather, analytical systems will augment expert judgment by integrating evidence, identifying weak signals, testing alternatives, and making assumptions visible. Organizations that combine data literacy, engineering knowledge, disciplined work management, and responsible governance will gain more value than those that treat AI as an isolated software purchase.

Chapter Summary

Reliability analytics connects maintenance data with engineering and managerial decisions. Descriptive and diagnostic analysis establish what has happened and why; predictive models estimate future risk; prescriptive systems compare actions under operational and economic constraints. Trend analysis, survival methods, hazard functions, and normalized indicators provide essential foundations for advanced modelling.

Effective predictive maintenance depends on a clearly defined decision, reliable asset and failure data, appropriate validation, and evaluation of false positives, missed failures, calibration, and business impact. Dashboards should be designed as decision interfaces rather than collections of charts. AI-enabled systems must address explainability, human accountability, privacy, cybersecurity, model drift, and responsible governance. The central professional principle is that analytics should improve informed action while making uncertainty and residual risk explicit.

Key Terms

Reliability analytics: The systematic use of data and analytical methods to understand, predict, and improve asset reliability.

Descriptive analytics: Analysis that summarizes historical events and performance.

Diagnostic analytics: Analysis intended to identify causes or contributing conditions.

Predictive analytics: Estimation of future events, conditions, or failure probabilities.

Prescriptive analytics: Analysis that recommends actions among feasible alternatives.

Reliability function: The probability that an item performs its required function beyond time t.

Hazard function: The instantaneous conditional rate of failure at a specified time.

Survival analysis: Statistical analysis of time-to-event data that can incorporate censored observations.

Censoring: A condition in which the event of interest is not observed during the available observation period.

Weibull distribution: A flexible reliability distribution capable of representing changing hazard rates.

MTBF: Mean operating time between failures of a repairable asset or population.

MTTR: Mean time required to restore an asset after failure.

Remaining useful life: Estimated operating time before a defined failure or performance limit.

Feature: An input variable used by an analytical model.

Data leakage: Use of information during model development that would not be available when making a real prediction.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Model drift: Deterioration or change in model behavior caused by changing data or relationships.

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

Data governance: Policies and responsibilities controlling data quality, access, use, retention, and accountability.

Decision support: Information, analysis, or recommendations designed to improve a human or organizational decision.

Review Questions

Short Answer Questions

1. Distinguish descriptive, diagnostic, predictive, and prescriptive analytics.

2. Why should failure rates often be normalized by operating exposure?

3. What is the purpose of a hazard function in reliability analysis?

4. Define data leakage and explain why it creates an unrealistic model evaluation.

5. What is model drift, and how can it be monitored?

Long Answer Questions

1. Explain how survival analysis supports maintenance and replacement decisions, including the significance of censoring.

2. Discuss the stages of developing and validating a predictive maintenance model.

3. Describe the design characteristics of an effective reliability dashboard for different managerial users.

4. Explain why data quality, explainability, cybersecurity, and privacy are essential elements of responsible AI in maintenance.

5. Compare predictive analytics with prescriptive decision support, using an industrial maintenance example.

Analytical / Application Questions

1. A plant reports that failures increased by 25 percent after installing a new CMMS coding scheme, while production hours also increased. Design an analysis to determine whether reliability actually deteriorated.

2. A model detects 90 percent of failures but generates four false alarms for every confirmed failure. Recommend whether and how it should be deployed for a safety-critical pump system, considering costs, thresholds, and human verification.

3. A predictive model’s performance declines after a process changes from one product mix to another. Develop a governance and technical response that distinguishes data drift from concept drift and prevents unsafe continued use.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes predictive analytics from prescriptive analytics in maintenance?

A. Predictive analytics summarizes past costs, whereas prescriptive analytics records work orders.

B. Predictive analytics estimates future risk, whereas prescriptive analytics recommends actions under constraints.

C. Predictive analytics measures only sensor accuracy, whereas prescriptive analytics measures MTTR.

D. Predictive analytics replaces engineering judgment, whereas prescriptive analytics eliminates uncertainty.

Correct Answer: B. Predictive analytics estimates future risk, whereas prescriptive analytics recommends actions under constraints.

Question 2

Why is censoring important in survival analysis?

A. It proves that all assets have identical failure distributions.

B. It removes assets that have not failed so that averages are simpler.

C. It represents cases where the event has not been observed during the available observation period.

D. It converts a repairable asset into a non-repairable asset.

Correct Answer: C. It represents cases where the event has not been observed during the available observation period.

Question 3

A model predicts a 70 percent failure probability for a group of comparable pumps. Calibration would require that:

A. Approximately 70 percent of comparable pumps eventually fail within the defined prediction context.

B. The model has an accuracy of exactly 70 percent on every individual prediction.

C. The model detects 70 percent of all failures regardless of false alarms.

D. Every pump receives the same risk score for operational consistency.

Correct Answer: A. Approximately 70 percent of comparable pumps eventually fail within the defined prediction context.

Question 4

Which practice most directly prevents data leakage during predictive-model development?

A. Using future maintenance outcomes as input features because they are highly correlated with failure.

B. Splitting data chronologically so that information from the future does not enter training.

C. Deleting all records for assets that have not yet failed.

D. Increasing the number of model variables without reviewing their availability at prediction time.

Correct Answer: B. Splitting data chronologically so that information from the future does not enter training.

Question 5

A sudden rise in recorded failures follows a revision of failure-code definitions. What should analysts do first?

A. Conclude that equipment reliability has deteriorated.

B. Discard all post-revision records.

C. Investigate the coding change and reconcile definitions before comparing trends.

D. Replace the reliability model with a neural network.

Correct Answer: C. Investigate the coding change and reconcile definitions before comparing trends.

Question 6

Which dashboard feature most improves decision usefulness?

A. Displaying many indicators without ownership or escalation rules.

B. Showing risk scores together with data freshness, uncertainty, and recommended verification actions.

C. Using red status for every asset whose sensor value is missing.

D. Automatically issuing replacement work orders for every high-risk prediction.

Correct Answer: B. Showing risk scores together with data freshness, uncertainty, and recommended verification actions.

Question 7

What is the principal concern when a model has high accuracy but failures are very rare?

A. Accuracy may conceal poor detection of the minority failure class.

B. High accuracy proves that the model is causally correct.

C. Rare failures make calibration unnecessary.

D. The model must have excessive recall by definition.

Correct Answer: A. Accuracy may conceal poor detection of the minority failure class.

Question 8

Which example best represents concept drift?

A. The average sensor value changes because a different sensor scale is installed.

B. The proportion of missing readings increases after a communications failure.

C. The relationship between vibration and failure changes after equipment operating conditions are redesigned.

D. The dashboard color scheme is changed by the user interface team.

Correct Answer: C. The relationship between vibration and failure changes after equipment operating conditions are redesigned.

Question 9

Why should a predictive maintenance system normally retain human verification before replacement?

A. Predictions are never useful for maintenance decisions.

B. Human verification can test the physical condition and assess context, consequences, and data quality.

C. Human verification guarantees that the model cannot drift.

D. Replacement decisions do not involve cost or safety considerations.

Correct Answer: B. Human verification can test the physical condition and assess context, consequences, and data quality.

Question 10

Which control is most relevant to protecting connected operational technology from unauthorized model or sensor manipulation?

A. Network segmentation and least-privilege access.

B. Removing all audit logs to improve system speed.

C. Allowing unrestricted vendor access for convenience.

D. Using only historical averages in every decision.

Correct Answer: A. Network segmentation and least-privilege access.