12. Data Analytics, AI and Condition-Based Intelligence

Introduction and Context

Mechatronic systems generate information continuously. Sensors measure vibration, temperature, pressure, current, position, acoustic emission, speed and load; controllers record commands and responses; production systems retain alarms, maintenance histories and quality data. The engineering challenge is no longer simply to acquire these signals, but to convert them into reliable knowledge about the state, performance and future behaviour of a machine.

Condition-based intelligence is the use of measured evidence to determine whether equipment is operating normally, degrading, or approaching failure, and to support an appropriate engineering decision. It extends condition monitoring from periodic inspection to continuous or event-driven assessment. Data analytics supplies the methods for cleaning, transforming and interpreting data. Artificial intelligence (AI) contributes models capable of classifying operating states, detecting unusual patterns, estimating remaining useful life and assisting decisions. These technologies are valuable only when connected to physical understanding, trustworthy data and a maintenance process.

A model that predicts a bearing fault accurately in a laboratory may fail on a production line because of changing loads, sensor mounting, machine variants, environmental noise or a different failure mechanism. Consequently, the central concern of this chapter is not the selection of an impressive algorithm, but the design of an evidence-based intelligence chain: sensing, data quality, signal processing, feature engineering, modelling, validation, explanation, deployment and action.

Learning Objectives

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

1. distinguish condition monitoring, condition-based maintenance, predictive maintenance and prognostics;

2. assess the quality, relevance and representativeness of an industrial dataset;

3. apply appropriate signal-processing and feature-engineering methods to mechatronic signals;

4. explain supervised, unsupervised and physics-informed approaches to diagnosis and prognostics;

5. formulate system-identification and anomaly-detection problems;

6. evaluate machine-learning models using leakage-resistant validation and decision-relevant metrics;

7. interpret the roles of computer vision, digital twins and edge-cloud architectures;

8. identify limitations involving explainability, drift, cybersecurity, safety and governance; and

9. design a practical condition-based intelligence workflow for industrial deployment.

From Measurements to Maintenance Decisions

Condition monitoring observes variables associated with equipment health. A vibration spectrum, for example, may contain components associated with shaft rotation, gear meshing or bearing defects. Condition-based maintenance uses such observations to decide when inspection, adjustment or replacement is justified. Predictive maintenance goes further: it estimates the probability or time of a future failure, thereby supporting maintenance before functional failure occurs. Prognostics is the systematic estimation of future health or remaining useful life (RUL), usually with uncertainty.

These terms describe increasing levels of decision support, not necessarily separate technologies. A temperature threshold is condition monitoring; a rule that schedules inspection after a persistent threshold crossing is condition-based maintenance; a model estimating failure risk over the next week is predictive maintenance; and a model estimating the distribution of time to loss of function is prognostics.

A useful abstraction is the condition-intelligence loop:

physical asset → sensors and operating context → data conditioning → health indicators → inference → maintenance decision → intervention and feedback.

Operating context is essential. The same motor current may be normal at high load and abnormal at low load. Speed, torque, product type, ambient temperature and control mode should therefore be recorded with the measured signal. Without context, an intelligent system can confuse a change in operation with a change in health.

Data Quality as an Engineering Problem

Industrial datasets are rarely clean experimental datasets. They contain missing samples, duplicated records, sensor saturation, time-clock errors, calibration changes, communication dropouts, maintenance interruptions and periods in which the machine was not operating. Labels are particularly difficult. A failure may be discovered during an inspection rather than at its true initiation time, while “normal” data may include undetected degradation.

Data quality should be examined along several dimensions:

• Validity: Does the value lie within physically credible and instrument-specific limits?
• Completeness: Are required variables and time intervals present?
• Consistency: Do timestamps, units, sampling rates and machine identifiers agree?
• Accuracy: Has the sensor been calibrated and installed correctly?
• Timeliness: Is the data available early enough to support action?
• Representativeness: Do the data cover loads, speeds, environments, machine types and failure modes expected in practice?
• Traceability: Can each record be linked to an asset, operating state, maintenance event and source system?

Data cleaning must not erase the phenomenon being studied. Replacing a genuine transient with an average value may make a model appear stable while removing the evidence of incipient failure. Missingness itself can be informative: repeated communication loss may indicate vibration, electromagnetic interference or a failing sensor. Imputation, filtering and deletion should therefore be documented and evaluated against engineering consequences.

The unit of analysis also matters. Randomly splitting individual samples from a continuous time record into training and test sets causes temporal leakage because neighbouring samples may contain nearly identical information. A more defensible strategy is to split by time, machine, production batch or complete run, depending on the intended deployment. If the model will be used on unseen machines, machine-wise validation is more meaningful than a random row-wise split.

Signal Processing for Mechatronic Health Data

Raw signals are often unsuitable for direct interpretation. Signal processing seeks to separate informative behaviour from measurement noise and to represent the signal at a useful time and frequency resolution.

Sampling frequency must be selected in relation to the highest frequency of interest. Sampling below twice that frequency can produce aliasing, in which high-frequency content appears incorrectly at a lower frequency. Anti-aliasing filtering before sampling is therefore important. In practice, the required rate is also influenced by transients, filter characteristics and the diagnostic bandwidth of the sensor.

Time-domain analysis describes amplitude behaviour. Common measures include mean, root-mean-square (RMS), variance, standard deviation, peak value, peak-to-peak value, crest factor and kurtosis. RMS often reflects overall vibration or energy, whereas crest factor is sensitive to isolated peaks. Kurtosis can increase in the presence of impulsive events, but none of these statistics is a universal fault indicator. Load changes, sensor resonance and signal clipping can produce similar changes.

Frequency-domain analysis uses the Fourier transform or power spectral density to identify periodic components. It is valuable for rotating machinery, where characteristic frequencies can be related to rotational speed, gear mesh or bearing geometry. Order tracking expresses frequency relative to shaft rotational order and is preferable when speed varies. A fixed-frequency spectrum can smear components during acceleration or deceleration.

Time-frequency methods address non-stationary behaviour. Short-time Fourier transforms provide a sequence of local spectra, while wavelet transforms offer multiresolution analysis and can represent short-lived impulses. Envelope analysis demodulates high-frequency resonance excited by repetitive bearing impacts. The method must match the failure physics; applying a sophisticated transform to an irrelevant frequency band does not create useful information.

Signals from different sensors may also be fused. Early fusion concatenates features before modelling; intermediate fusion combines learned representations; late fusion combines separate model decisions. Synchronisation and calibration are necessary in every case. A camera image, vibration segment and controller state must refer to the same physical event if their relationship is to be interpreted causally.

Feature Engineering and Health Indicators

Feature engineering converts conditioned signals into variables that are informative, stable and computationally practical. Features may be statistical, spectral, temporal, spatial, image-based or process-derived. Examples include band energy, spectral kurtosis, sideband amplitude, temperature rise over ambient, motor-current harmonics, pressure pulsation and the rate of change of a measurement.

A good feature should respond to degradation while remaining comparatively insensitive to nuisance variables such as load or ambient temperature. Normalisation can be performed using operating-condition groups, reference baselines or regression models. For example, temperature may be compared with expected temperature at the current load rather than with a fixed absolute threshold.

Dimensionality reduction methods such as principal component analysis can compress correlated variables and reveal dominant variation. However, a principal component is not automatically a health indicator; its physical meaning and stability must be examined. Feature selection may use domain knowledge, correlation analysis, mutual information, regularisation or model-based importance. Feature extraction through neural networks can be effective with large datasets, but it generally requires stronger validation and creates more difficult explanations.

A health indicator is a scalar or low-dimensional representation intended to track condition. Desirable properties include monotonicity during degradation, trendability across assets and robustness to operating changes. In real equipment, degradation is rarely perfectly monotonic. A health indicator should therefore be interpreted with confidence bounds, smoothing assumptions and knowledge of intervention history.

Modelling Strategies: Identification, Diagnosis and Prognostics

System identification estimates a mathematical model of a dynamic system from measured inputs and outputs. In mechatronics it may be used to identify motor, actuator, thermal or process dynamics. A model such as

y(k) = f[y(k−1), u(k−1), θ] + e(k)

relates current output to previous outputs, inputs, parameters and error. Changes in estimated parameters, residuals or dynamic response can indicate degradation. Model-based monitoring is especially useful when physical laws and control inputs are known, although unmodelled disturbances may produce false alarms.

In supervised diagnosis, labelled examples are used to map observations to fault classes or health states. Classification models include logistic regression, support vector machines, decision trees, random forests and neural networks. Regression models estimate a continuous quantity such as temperature, efficiency or RUL. Supervised learning is powerful when failure labels are reliable and sufficiently diverse, but industrial failure data are often rare and imbalanced.

Unsupervised and semi-supervised methods learn normal behaviour and flag deviations. Clustering, density estimation, autoencoders and one-class classifiers are common approaches. They are useful where failures are uncommon or labels are unavailable. Their central assumption is that training data represent healthy operation. If degraded data enter the reference set, the model may learn that degradation is normal.

Semi-supervised approaches combine abundant normal data with a small number of labelled events. Transfer learning can adapt a model from one machine or operating regime to another, but transfer is justified only after checking that the underlying physical relationships are sufficiently similar.

Physics-informed and hybrid models combine conservation laws, mechanistic degradation models or digital-twin simulations with data-driven residual models. The physical model can constrain implausible predictions and improve performance under limited data. Conversely, an inaccurate physical model may introduce systematic bias. Hybrid modelling is therefore not a substitute for validation.

Prognostics requires a definition of failure, a prediction horizon and a representation of uncertainty. RUL should not be reported as a single precise number when the evidence supports only a broad range. Sources of uncertainty include sensor noise, unknown operating conditions, model uncertainty, variation between assets and ambiguity in the failure threshold. Prediction intervals, probability distributions or risk bands are more responsible outputs than unsupported point estimates.

Anomaly Detection and Decision Thresholds

An anomaly is an observation or sequence that differs materially from an established reference pattern. Detection is not equivalent to diagnosis. An anomaly score may indicate that something has changed without identifying the component or failure mechanism.

Thresholds can be fixed, adaptive, statistical or decision-based. A fixed threshold is easy to implement but may not account for speed and load. An adaptive threshold may be based on a moving baseline, quantiles, control charts or a model of expected behaviour. Threshold selection should consider false positives, false negatives, alarm persistence and the cost of inspection or failure.

For rare failures, accuracy is a poor metric because a model can achieve high accuracy by predicting “healthy” almost always. Precision, recall, specificity, F1 score, precision-recall curves and the false-alarm rate per operating hour are more informative. In prognostics, useful measures include absolute error at a specified horizon, calibration of prediction intervals and the timeliness of warning. Evaluation must reflect the maintenance decision: an alarm that arrives after a required lead time is operationally unsuccessful even if its classification label is correct.

Alarm logic often improves when it requires persistence, repeated evidence or corroboration from multiple sensors. Hysteresis can prevent rapid switching around a threshold. Escalation levels can distinguish observation, inspection, planned intervention and immediate shutdown. The logic must be tested against nuisance events and documented so that operators understand the meaning of each alarm.

Computer Vision in Condition Intelligence

Vision-based inspection is increasingly applied to surface defects, weld quality, leakage, corrosion, tool wear, assembly verification and thermal anomalies. A vision pipeline normally includes controlled illumination, image acquisition, calibration, preprocessing, segmentation or detection, feature extraction and decision logic. Convolutional neural networks and modern object-detection or segmentation models can learn complex visual patterns, but their reliability depends strongly on image quality and coverage of expected variation.

Industrial vision differs from ordinary image classification. Defect prevalence may be extremely low; camera position, lighting, reflectivity and contamination can change; and a missed defect may have safety implications. Dataset splits should be made by part, batch, time period or production line to avoid near-duplicate images in training and testing. Synthetic images and augmentation can improve coverage, but they should not conceal a shortage of real examples.

Explainability may involve saliency maps, defect masks, nearest reference images or geometric measurements. Such tools help engineers inspect whether a model is attending to the defect or to an irrelevant background feature. They are aids to investigation, not proof that the model is causally correct.

Validation, Explainability and Deployment

A credible validation plan begins with the intended use. Define the asset population, operating envelope, prediction horizon, failure definition, action threshold and acceptable risk before selecting a model. The dataset should be divided into training, validation and final test data without temporal or asset leakage. Hyperparameters must be selected using the training and validation data; the final test set should remain untouched until the design is fixed.

Cross-validation is useful for estimating variation, but ordinary random cross-validation is unsuitable for dependent time-series observations. Rolling-origin evaluation, blocked time splits and leave-one-asset-out tests better reflect deployment. A baseline rule, such as a calibrated threshold or persistence model, should be included. A complex model is justified only if it provides measurable operational benefit.

Explainability has several levels. A global explanation describes overall relationships, feature effects and operating limits. A local explanation describes why a particular alarm or prediction was produced. Mechanistic validation asks whether the detected features correspond to credible failure physics. Counterfactual analysis asks what change in input would alter the decision. Explanations can be misleading when features are correlated, when the model is unstable or when a post hoc method approximates rather than reveals the model’s actual reasoning.

Deployment introduces constraints absent from an offline experiment. Edge inference may reduce latency and protect bandwidth, while cloud systems provide greater storage and computing capacity. A practical architecture often performs filtering, feature extraction and immediate alarms at the edge, with selected data and model updates managed centrally. Version control, configuration management, rollback, cybersecurity, access control and audit trails are essential engineering functions.

Models require monitoring after deployment. Data drift occurs when input distributions change; concept drift occurs when the relationship between input and outcome changes. Sensor replacement, controller retuning, process redesign and seasonal conditions can all cause drift. Performance should be reviewed using delayed maintenance outcomes, alarm rates, missingness, calibration and operator feedback. Retraining must be governed rather than automatic, particularly where safety or regulatory compliance is involved.

An Illustrative Case: Predictive Monitoring of a Variable-Speed Pump

Background

Consider a fictional process plant using several variable-speed centrifugal pumps. Each pump has vibration sensors on the drive-end bearing, motor current measurement, discharge pressure, speed, temperature and a controller event log. Maintenance has historically been reactive. A bearing failure can interrupt production, but excessive preventive replacement also creates cost and alignment risk.

Situation and Problem

The plant initially proposes a neural network to predict bearing failure thirty days in advance. Historical records contain only a small number of confirmed failures, while many maintenance work orders have vague descriptions such as “noise” or “routine replacement.” Operating speed varies widely, and vibration amplitude increases naturally with flow and speed. A random sample split produces apparently excellent test accuracy.

Analysis

An engineering review identifies leakage: adjacent windows from the same pump and failure episode occur in both training and test sets. It also finds that some “healthy” windows were collected shortly before undocumented replacements. Sensor timestamps are not aligned, and vibration data are sampled at different rates after a monitoring-system upgrade. The problem is reformulated as two linked tasks: first, detect deviation from condition-adjusted normal behaviour; second, estimate whether the deviation is consistent with bearing degradation and requires inspection.

Approach and Solution

The team synchronises signals using controller timestamps, removes impossible values without deleting informative transients, and records missingness. Data are grouped by pump and operating regime. A baseline model predicts expected vibration features from speed, load-related pressure and temperature. The residuals are converted into health indicators using band energy, envelope amplitude, crest factor and trend statistics.

A semi-supervised anomaly detector is trained primarily on verified healthy periods. Alarm persistence is required over several operating windows, and a second evidence channel—temperature rise or characteristic spectral content—must corroborate a high-priority alarm. A small labelled set is used to assess whether detected anomalies correspond to known bearing events. Validation uses time-blocked splits and leave-one-pump-out testing. The system reports an anomaly score, contributing features, operating context and a confidence band rather than a categorical claim of certain failure.

The edge device calculates features and issues local warnings. Daily summaries and selected waveforms are sent to a central server for trend analysis. Maintenance engineers review high-priority cases, and their findings are linked back to the data. A model registry stores the deployed version, training data range and threshold configuration.

Outcome

In this illustrative case, the most important improvement is not a particular AI algorithm. It is the removal of leakage, adjustment for operating conditions and integration of alarms with inspection practice. The system produces fewer nuisance alarms than the original fixed threshold and provides earlier, more credible evidence for planned bearing inspection. Its predictions remain advisory until sufficient field evidence is accumulated.

Learning Points

The case demonstrates that asset identity, timestamp integrity, labels, operating context and maintenance feedback are as important as model architecture. It also shows why anomaly detection and diagnosis should be separated, why uncertainty should be visible, and why deployment must include human review, governance and monitoring.

Professional and Managerial Implications

Condition-based intelligence changes responsibilities across an organisation. Instrumentation engineers define observability and calibration requirements; controls engineers interpret dynamics and operating regimes; data scientists construct and validate models; maintenance engineers define failure modes and actions; production managers determine acceptable downtime and access windows; cybersecurity specialists protect interfaces; and managers govern risk, cost and accountability.

The business case should not be expressed only as model accuracy. Relevant outcomes include avoided downtime, reduced secondary damage, maintenance labour, spare-parts inventory, inspection burden, energy performance and safety exposure. A high-performing detector may still be uneconomic if each alarm requires an expensive shutdown. Conversely, a modest model may be valuable when inspections are cheap and failures are costly.

Challenges, Risks and Common Mistakes

Important failure modes in analytics projects include:

• confusing correlation with a physical failure mechanism;
• training on post-failure data while claiming early prediction;
• allowing data leakage through random sampling of time windows;
• ignoring class imbalance and reporting accuracy alone;
• treating maintenance records as perfectly reliable labels;
• deploying a model outside its validated operating envelope;
• suppressing unusual data as “outliers” without engineering investigation;
• using an opaque model where an operator must justify a safety decision;
• failing to monitor drift after sensor, software or process changes; and
• generating alarms without a defined response, owner or escalation time.

AI can amplify systematic bias in the data. A model trained mainly on one machine family may underperform on another. Vision systems can be biased by lighting or supplier-specific appearance. Automated shutdown decisions can create hazards if the model fails, communication is lost or the sensor is compromised. Safe design therefore requires fail-safe behaviour, independent protection where necessary, manual override under controlled procedures and clear separation between advisory analytics and certified safety functions.

Ethical, Governance and Regulatory Considerations

Industrial data may reveal production rates, operator practices, supplier information or proprietary designs. Access should follow least-privilege principles, and data retention should be justified. Personnel should not be subjected to undisclosed individual surveillance through machine or performance data. Models affecting work allocation or maintenance responsibility require transparent governance and an opportunity for review.

Traceability is essential: the organisation should be able to identify the data version, preprocessing steps, model version, threshold, decision and human response associated with an alarm. For safety-critical applications, AI should not be treated as a replacement for hazard analysis, independent protection layers, verification and validation. Professional judgement remains accountable for decisions made using uncertain evidence.

Emerging Directions

Current development is moving toward self-supervised learning, in which large volumes of unlabelled sensor data support representation learning; federated learning, in which models are trained across sites without centralising all raw data; and foundation models adapted to multimodal industrial signals. Digital twins increasingly combine simulation, live measurements and parameter estimation. Edge AI is expanding because low latency, bandwidth limits and data sovereignty matter in factories.

These developments do not remove the central engineering questions. A larger model does not guarantee better transfer across machines, and synthetic data do not automatically represent real degradation. Future systems are likely to combine mechanistic constraints, uncertainty quantification, causal reasoning and human-centred interfaces. The most useful intelligence will be that which is robust, auditable and connected to an achievable maintenance action.

Best Practices for Implementation

A practical implementation should:

1. begin with a specific failure mode, decision and measurable business outcome;

2. perform failure-mode and effects analysis to determine what must be observed;

3. establish sensor, timestamp, calibration and asset-identity standards;

4. document operating regimes, maintenance events and label uncertainty;

5. create a simple engineering baseline before adopting complex AI;

6. split data according to the intended deployment and prevent leakage;

7. evaluate rare-event performance, lead time, calibration and alarm burden;

8. expose context, evidence and uncertainty to users;

9. pilot the system with human review and a defined escalation process;

10. monitor drift, data quality and operational outcomes after release; and

11. control model versions, cybersecurity, retraining and retirement through formal governance.

Chapter Summary

Data analytics and AI transform mechatronic measurements into condition-based intelligence through a chain of sensing, conditioning, feature engineering, inference and maintenance action. Reliable practice begins with data quality, operating context and physically meaningful failure definitions. Time- and frequency-domain processing, time-frequency analysis, system identification, supervised learning and anomaly detection provide complementary tools. Prognostics must communicate uncertainty, while validation must prevent temporal and asset leakage and use decision-relevant metrics. Computer vision, edge-cloud architectures, digital twins and hybrid physics-data models extend capability but introduce new requirements for calibration, explainability and governance. Deployment success depends less on algorithmic novelty than on robust evidence, clear alarm responses, drift monitoring, cybersecurity and integration with professional maintenance practice.

Key Terms

Condition monitoring: Measurement and interpretation of variables that indicate equipment condition.

Condition-based maintenance: Maintenance scheduled or triggered using evidence of actual equipment condition.

Predictive maintenance: Maintenance planning supported by predictions of future degradation or failure risk.

Prognostics: Estimation of future health or remaining useful life, including uncertainty.

Remaining useful life (RUL): Expected time or usage until a defined failure or performance limit.

Aliasing: Misrepresentation of high-frequency signal content caused by insufficient sampling.

Order tracking: Analysis of rotating-machine signals relative to shaft rotational speed.

Envelope analysis: Demodulation technique used to reveal repetitive impacts carried by a resonance.

Feature engineering: Construction or selection of variables that represent informative signal behaviour.

Health indicator: A variable intended to track equipment degradation or condition.

System identification: Estimation of a system’s dynamic model from measured inputs and outputs.

Anomaly detection: Identification of observations that differ from an established normal pattern.

Data leakage: Unintended use of information during training that would not be available at deployment.

Concept drift: Change over time in the relationship between observations and outcomes.

Data drift: Change in the distribution of model inputs over time.

Explainability: The extent to which a model’s outputs and influencing evidence can be understood.

Digital twin: A computational representation linked to a physical asset through data and models.

Edge inference: Execution of analytics near the physical equipment rather than exclusively in a remote cloud.

Review Questions

Short Answer Questions

1. Distinguish condition monitoring from prognostics.

2. Why is operating context important when interpreting vibration or temperature?

3. What is aliasing, and how can it be prevented?

4. Define data leakage in a time-series maintenance dataset.

5. Why is accuracy insufficient for evaluating a rare-failure detector?

Long Answer Questions

1. Explain a complete condition-intelligence loop from sensing to maintenance action.

2. Compare time-domain, frequency-domain and time-frequency signal analysis for rotating machinery.

3. Discuss supervised, unsupervised and hybrid physics-data approaches to equipment diagnosis.

4. Develop a validation strategy for an RUL model used on several different machines.

5. Explain the requirements for responsible deployment of AI in an industrial maintenance system.

Analytical / Application Questions

1. A motor’s RMS vibration rises whenever speed increases, but no fault is found. Design an analysis that separates operating-condition effects from degradation.

2. A vision model reports 99% accuracy on a defect dataset, yet misses defects on a new production batch. Identify likely causes and propose a validation and data-improvement plan.

3. Design alarm logic for a pump anomaly detector where false alarms are costly but failure causes major production loss. Specify evidence, persistence, escalation and human response.

22. Multiple Choice Questions with Answers

Question 1

Which situation best represents prognostics rather than simple condition monitoring?

A. Recording bearing temperature every second

B. Displaying a vibration spectrum to an operator

C. Estimating the probability distribution of time until bearing failure

D. Activating a warning when temperature exceeds a fixed limit

Correct Answer: C. Estimating the probability distribution of time until bearing failure

Question 2

Why can a random row-wise split produce misleading performance for time-series maintenance data?

A. It always removes all failure observations

B. It may place highly similar neighbouring windows in both training and test sets

C. It prevents the use of frequency-domain features

D. It increases the sensor sampling rate

Correct Answer: B. It may place highly similar neighbouring windows in both training and test sets

Question 3

Which technique is most appropriate for analysing rotating machinery whose speed changes substantially during operation?

A. Order tracking

B. Fixed categorical encoding

C. Random oversampling of timestamps

D. Image thresholding

Correct Answer: A. Order tracking

Question 4

A detector is trained using mostly healthy data and flags deviations without assigning a fault class. What type of approach is this?

A. Unsupervised or semi-supervised anomaly detection

B. Multiclass supervised classification

C. Closed-loop system identification only

D. Regulatory compliance testing

Correct Answer: A. Unsupervised or semi-supervised anomaly detection

Question 5

Which metric is especially informative when failures are rare and false alarms impose a significant burden?

A. Accuracy alone

B. Number of model parameters

C. Precision-recall performance and false alarms per operating hour

D. Training-set size alone

Correct Answer: C. Precision-recall performance and false alarms per operating hour

Question 6

In the pump case, why were vibration features adjusted for speed and pressure?

A. To eliminate all variation from the data

B. To distinguish operating-condition effects from possible degradation

C. To guarantee that every failure would be predicted

D. To replace the need for maintenance records

Correct Answer: B. To distinguish operating-condition effects from possible degradation

Question 7

What is concept drift in a deployed condition-intelligence model?

A. A change in the relationship between inputs and outcomes

B. A temporary missing sensor sample only

C. A deliberate increase in sampling frequency

D. A fixed alarm threshold

Correct Answer: A. A change in the relationship between inputs and outcomes

Question 8

Which statement best describes a responsible RUL output?

A. A single exact failure date with no qualification

B. A prediction interval or risk distribution reflecting uncertainty

C. A binary label that ignores operating conditions

D. A value obtained only from training data

Correct Answer: B. A prediction interval or risk distribution reflecting uncertainty

Question 9

What is a principal danger of deleting every statistical outlier during preprocessing?

A. The model will necessarily become nonlinear

B. Genuine transient evidence of incipient failure may be removed

C. The sampling theorem will no longer apply

D. The machine identity will be duplicated

Correct Answer: B. Genuine transient evidence of incipient failure may be removed

Question 10

Which deployment practice most directly supports auditability?

A. Changing thresholds informally whenever an alarm occurs

B. Storing model version, preprocessing, threshold and decision history

C. Using only the most complex available neural network

D. Disabling operator feedback

Correct Answer: B. Storing model version, preprocessing, threshold and decision history