Diagnostics, Prognostics and Maintenance Analytics

Introduction

Modern maintenance management has moved from repairing failed equipment to managing the evolving risk of failure. This transition is enabled by diagnostics, prognostics and maintenance analytics. Diagnostics identifies abnormal conditions and their likely causes; prognostics estimates how a fault will develop and when functional failure may occur; and maintenance analytics converts condition, process and maintenance data into decisions about inspection, intervention, spares and production risk.

The central managerial challenge is not simply to predict failure. It is to make a technically credible and economically sound decision under uncertainty. A prediction is valuable only when it changes an action—such as reducing load, scheduling an outage, replacing a component or increasing inspection frequency.

Learning Objectives

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

1. Distinguish detection, diagnosis, prognosis and maintenance decision-making.
2. Apply statistical process control and trend analysis to condition-monitoring data.
3. Explain anomaly-detection methods and their limitations.
4. Construct and interpret degradation models and remaining-useful-life estimates.
5. Quantify uncertainty in forecasts and maintenance decisions.
6. Design a practical analytics workflow for industrial assets.
7. Evaluate technical, economic, ethical and governance issues in predictive maintenance.

1. The Maintenance Analytics Chain

A useful analytical chain is:

measurement → data preparation → detection → diagnosis → prognosis → decision → feedback

A sensor measurement is not itself a diagnosis. For example, increasing vibration may result from imbalance, misalignment, looseness, resonance, bearing damage or a change in operating speed. Similarly, an anomaly score is not a failure probability. It indicates departure from a reference pattern, but the operational meaning must be established.

Important terminology includes:

• Condition indicator: a calculated variable associated with equipment health, such as vibration RMS, oil-particle count or motor current.
• Fault: a physical or functional abnormality.
• Failure: loss of the required function, usually defined against explicit performance criteria.
• Detection: recognition that current behaviour differs from expected behaviour.
• Diagnosis: identification of the fault type, location or cause.
• Prognosis: estimation of future health, failure risk or remaining useful life.
• Remaining useful life (RUL): time or usage from the current observation until a specified failure criterion is reached.
• Health indicator: a variable designed to represent degradation, ideally with a monotonic relationship to health loss.
• Concept drift: change in the data-generating process over time, such as a product mix or operating regime change.

A reliable system defines the failure threshold before modelling. “Failure” may mean a trip, a safety limit breach, unacceptable product quality, excessive energy use or a maintenance policy threshold. Different thresholds produce different RUL values.

2. Data Foundations and Measurement Discipline

Analytics quality is constrained by data quality. Maintenance records often contain inconsistent failure codes, missing completion dates, duplicated work orders and replacement events that are not true failures. Sensor data may contain calibration errors, communication gaps, changing sampling rates and confounding effects caused by speed, load, temperature or product type.

A professional data pipeline should include:

1. Asset identity: a stable identifier and equipment hierarchy.
2. Time alignment: consistent time zones, clocks and sampling intervals.
3. Operating context: speed, load, temperature, recipe, environment and production state.
4. Event labels: inspection, adjustment, partial repair, component replacement and functional failure.
5. Data-quality flags: missingness, sensor saturation, implausible values and calibration status.
6. Separation of training and test periods: future observations must not leak into model development.

For a measured signal x₁, x₂, …, xₙ, the arithmetic mean is:

Here x̄ is the sample mean and n is the number of observations. The sample standard deviation is:

The denominator n − 1 is appropriate when estimating population variability from a sample. These statistics assume observations are sufficiently comparable. Mixing measurements from different speeds or loads can create a false trend.

Feature engineering must have physical meaning. Common features include RMS vibration, kurtosis, crest factor, temperature slope, pressure deviation, motor-current harmonics, lubricant contamination and counts of alarms. A feature should be traceable to a failure mechanism or validated empirically.

3. Statistical Process Control for Condition Monitoring

Statistical process control (SPC) distinguishes routine variation from evidence of a changed process. In maintenance, the “process” may be bearing vibration, hydraulic pressure, cutting-tool force or energy consumption.

For independent observations from a stable process, a basic individuals chart uses a centre line μ and limits:

where σ is the process standard deviation. In practice, μ and σ are estimated from a stable baseline. The limits are probability-based warning boundaries, not physical safety limits. A value beyond a control limit indicates unusual statistical behaviour; it does not prove imminent failure.

For subgroup means, the standard error is σ/√n, so the approximate limits for a mean chart are:

The subgroup size n must be defined consistently. When data are autocorrelated, ordinary limits can generate excessive false alarms. Time-series methods or residual charts should then be considered.

Typical SPC signals include:

• one point beyond a control limit;
• several successive points on one side of the centre line;
• a sustained upward or downward run;
• a systematic cycle;
• increasing spread;
• a trend approaching a physical or operational limit.

Control limits should not be confused with specification limits. Specification limits arise from engineering, safety or quality requirements; control limits arise from observed process variation. A process may be statistically stable but operate outside specification, or statistically unstable while still remaining within specification.

Worked Example: A Vibration Control Chart

Suppose a pump’s vibration RMS has a stable baseline mean of 2.0 mm/s and standard deviation 0.20 mm/s. The three-sigma limits are:

A reading of 2.55 mm/s is inside the control limits, but six consecutive readings increasing from 1.95 to 2.55 mm/s may indicate degradation. The correct response is not automatic replacement. The engineer should confirm speed and load, inspect for process changes, check sensor installation and examine the vibration spectrum.

4. Trend Analysis and Change Detection

Trend analysis estimates whether a condition indicator is changing systematically. A simple linear model is:

where yₜ is the condition indicator at time t, β₀ is the intercept, β₁ is the rate of change per unit time, and εₜ is unexplained variation. A positive β₁ suggests deterioration only when higher y represents poorer health and operating conditions are comparable.

The least-squares slope is:

The units of β̂₁ are units of the indicator per unit time. A trend should be accompanied by confidence or prediction intervals, not presented as a certain trajectory.

Trend methods include moving averages, exponentially weighted moving averages, robust regression, Theil–Sen slopes, cumulative sums (CUSUM) and exponentially weighted moving average (EWMA) charts. Smoothing reduces random noise but can delay detection and obscure abrupt changes. Robust methods are useful when occasional outliers are caused by transients or measurement errors.

A change-point method seeks a time τ at which the statistical properties change. For example, a mean-shift model is:

The change point may represent a repair, process change, sensor replacement or actual degradation. Therefore, event logs must be integrated with statistical analysis.

5. Anomaly Detection and Fault Diagnosis

Anomaly detection is particularly useful when labelled failure data are scarce. Three broad strategies are common:

In supervised classification, labelled examples of fault classes are used. Examples include logistic regression, decision trees, random forests and support-vector machines. Performance should be evaluated with time-aware validation and a cost-sensitive confusion matrix.

In unsupervised detection, a model of normal behaviour is developed. For a multivariate observation x with mean vector μ and covariance matrix Σ, the squared Mahalanobis distance is:

This measures distance while accounting for scale and correlation. It requires an appropriate covariance estimate and is unreliable when variables are strongly non-normal, the sample is small or Σ is nearly singular. Regularisation or dimension reduction may be needed.

Autoencoders, isolation forests and clustering can detect unusual patterns, but an unusual pattern may be a new normal operating mode rather than a fault. The model should therefore receive contextual variables and be reviewed after production changes.

Diagnosis combines sensor evidence with engineering knowledge. A practical diagnostic workflow is:

1. Confirm that the anomaly is real and not a sensor or communication problem.
2. Check operating regime and recent process changes.
3. Compare multiple independent indicators.
4. Examine frequency-domain, spatial or transactional signatures where relevant.
5. Generate candidate fault modes.
6. Rank candidates using evidence, failure-mode knowledge and consequence.
7. Recommend a confirmatory inspection or test.

A diagnostic model should report evidence and confidence, not only a fault label. Explainability is especially important when a false diagnosis can cause an unnecessary shutdown or an unsafe continued operation.

6. Degradation Modelling

A degradation model describes how an asset’s health changes with time, usage or stress. Let z(t) be a latent health state and y(t) a measured indicator:

Here v is a degradation rate, w(t) represents process variation, h(·) maps health to the observed measurement and ε(t) is measurement noise. A linear model is useful when degradation is approximately steady. It is inappropriate when wear accelerates, repairs reset the condition, or operating stress changes substantially.

Common empirical forms include:

• Linear wear: y(t) = a + bt.
• Exponential growth: y(t) = aeᵇᵗ, often used for accelerating indicators.
• Power-law wear: y(t) = a + btᵖ, where p controls curvature.
• Stochastic models: state-space, Wiener-process or gamma-process models for random degradation.

A Wiener process model can be represented conceptually as:

where X(t) is degradation, μ is average drift, σ is diffusion intensity and W(t) is standard Brownian motion. The model assumes continuous stochastic evolution and is not automatically suitable for monotonic wear; a constrained or gamma-process model may be more appropriate for strictly increasing cumulative damage.

Model selection should be based on mechanism, residual diagnostics, predictive performance and operational usefulness—not merely the highest in-sample fit.

7. Remaining Useful Life Estimation

If a health indicator increases toward a failure threshold L and a fitted linear trajectory is y(t) = a + bt with b > 0, the point estimate of failure time is:

If the current time is t₀, the point estimate of RUL is:

This calculation assumes a stable slope, a valid threshold, no intervention and comparable future operating conditions. It should not be treated as a guaranteed date.

Worked Example: Linear RUL Estimate

A bearing health indicator is currently 6.0 units. Its estimated increase is 0.08 units per operating hour, and the intervention threshold is 8.0 units. The point estimate is:

If the slope estimate is uncertain, the RUL estimate is uncertain. Suppose plausible slopes range from 0.06 to 0.10 units per hour. Then:

This range is not a formal confidence interval; it is a scenario range. A formal interval should account for uncertainty in the intercept, slope, noise, threshold and future operating conditions.

A probabilistic prognosis estimates a distribution such as P(RUL ≤ r | data). Results may be presented as median RUL, prediction intervals or probabilities of surviving a planned operating period. Prediction intervals include future variation and are wider than confidence intervals for the average trajectory.

Survival analysis provides another framework. The survival function is:

where T is the random failure time. The hazard function is:

The hazard is an instantaneous conditional failure rate, not a probability by itself. Survival models are useful when some assets are right-censored because observation ends before failure.

8. Uncertainty and Decision-Making

Uncertainty has several sources:

• Aleatory uncertainty: inherent random variation in loading, environment or failure processes.
• Epistemic uncertainty: incomplete knowledge, limited data or model inadequacy.
• Measurement uncertainty: sensor noise, calibration and sampling effects.
• Scenario uncertainty: unknown future operating conditions or maintenance actions.

A decision should combine failure risk with consequences and action costs. A simplified expected-loss comparison is:

where C_failure is the consequence cost and C_action is the cost of inspection, repair, downtime or intervention. This expression is a simplified decision model; real applications may include production scheduling, safety constraints, salvage value, false alarms and multiple future periods.

For example, a planned replacement costing $4,000 may be justified when the probability of failure during a critical production window is sufficiently high and the failure consequence is $60,000. However, the calculation must include whether the replacement itself introduces installation risk and whether a lower-cost inspection can reduce uncertainty.

Value of information is important. An inspection is valuable when it changes the decision enough to justify its cost. Analytics should therefore support a decision policy rather than merely produce a forecast.

9. Illustrative Case: Predictive Maintenance of a Process Pump

Background

The following is a fictional illustrative case. A chemical-processing plant operates six centrifugal pumps. A pump failure causes approximately eight hours of production interruption. Each pump has accelerometers, bearing-temperature sensors, motor-current measurements and a historian containing speed, flow and load.

Problem

The maintenance team experiences nuisance vibration alarms during low-flow operation, while two bearing failures occurred without a sustained high temperature. The plant needs earlier warning without increasing unnecessary shutdowns.

Analysis

The team first separated data by speed and flow regime. A baseline was created using 90 days of verified healthy operation. Vibration RMS was normalised by operating regime, and sensor communication gaps were flagged rather than imputed blindly. SPC showed that a raw vibration limit generated many false alarms during low-flow recirculation.

A multivariate normal-operation model used vibration, temperature, current and flow. The anomaly score rose when vibration increased relative to speed and when temperature residuals increased relative to load. Spectral inspection indicated a defect-frequency pattern consistent with an outer-race bearing fault on one pump.

A degradation indicator combined the validated vibration-band amplitude and temperature residual. A monotonicity check showed that the indicator increased over the final 30 operating hours before confirmed replacement. A linear RUL model estimated 18 hours to the intervention threshold, with a conservative scenario range of 12–27 hours.

Approach and Solution

The plant adopted a three-level response:

• Advisory: verify sensor and operating context.
• Inspection: conduct an ultrasound and lubrication check at the next feasible opportunity.
• Action: schedule bearing replacement when the probabilistic forecast placed a high probability of crossing the threshold before the next planned outage.

The model was not allowed to issue an automatic shutdown. A rotating-equipment engineer reviewed high-consequence cases, and every intervention was recorded for later model evaluation.

Outcome

During a pilot period, nuisance alarms decreased because operating regime was incorporated. One pump was inspected and found to have early bearing damage; replacement was completed during a short planned stop. The plant avoided an emergency outage, but the team also identified a limitation: after lubrication, the vibration indicator temporarily decreased while the underlying bearing damage remained. The model was revised to recognise maintenance events and avoid interpreting temporary recovery as full health restoration.

Learning Points

1. Contextualisation is often more valuable than a complex algorithm.
2. Detection and diagnosis require different evidence.
3. RUL is conditional on the threshold, model and future operating regime.
4. Maintenance events must be represented explicitly.
5. Human review is appropriate when consequences are high.

10. Implementation Framework

A mature implementation can follow these stages:

1. Define the decision: What action will the forecast support?
2. Define failure and intervention thresholds: Use functional and consequence-based criteria.
3. Select critical assets: Consider safety, bottleneck status, repair cost, failure frequency and observability.
4. Map failure modes: Use FMEA, reliability analysis and subject-matter knowledge.
5. Assess data readiness: Check quality, labels, sensor coverage and operating context.
6. Establish a baseline: Use verified healthy data and document its conditions.
7. Develop detection and diagnosis: Begin with interpretable features and rules where appropriate.
8. Validate prognosis: Use time-based, asset-based and event-based validation; avoid data leakage.
9. Connect outputs to work management: Define alarm ownership, response times and escalation.
10. Monitor the model: Track false alarms, missed failures, calibration, drift and intervention outcomes.

Useful performance measures include precision, recall, false-alarm rate, detection lead time, calibration, RUL error and avoided consequence. Accuracy alone can be misleading when failures are rare. For a rare failure class, a model that predicts “healthy” always may have high accuracy but no maintenance value.

11. Limitations, Governance and Ethics

Predictive maintenance systems can fail through poor sensors, biased historical labels, unrepresentative training data, changing equipment design and overconfident forecasts. A model trained on one operating regime may be unsafe in another. Censored maintenance records can also bias failure-time analysis.

Governance should specify:

• data ownership and access control;
• cybersecurity for connected equipment;
• model versioning and audit trails;
• approval authority for automated actions;
• retention of raw data and maintenance decisions;
• procedures for model drift and incident investigation;
• clear communication of uncertainty.

Ethical concerns include surveillance of operators, unfair performance attribution and the risk that management uses predictions to defer necessary maintenance. Analytics must support safe work, not replace competent engineering judgement. Safety-critical decisions require documented validation, fail-safe design and compliance with applicable industry regulations and standards.

Artificial intelligence can improve feature extraction, classification and forecasting, but it does not remove the need for physical understanding. Physics-informed models, hybrid digital twins and interpretable machine learning are increasingly important because they combine data adaptability with engineering constraints. Contemporary research also addresses federated learning, edge analytics, uncertainty-aware deep learning and causal inference. These methods are promising, but their adoption should be justified by measurable decision value.

Chapter Summary

Diagnostics identifies the nature and likely cause of abnormal equipment behaviour, while prognostics estimates future degradation and remaining useful life. Maintenance analytics links these outputs to economically and operationally sound actions.

SPC provides a disciplined method for distinguishing common variation from unusual behaviour, but control limits are not equivalent to engineering limits. Trend analysis and change-point detection reveal evolving conditions, yet they must account for autocorrelation, operating regime and maintenance events. Anomaly detection may be rule-based, statistical or machine-learning based; each approach requires validation and contextual interpretation.

Degradation models may be linear, nonlinear or stochastic. RUL estimates are conditional forecasts, not guaranteed failure dates. Survival analysis is valuable when failure observations are censored, and uncertainty should be communicated through ranges, distributions or calibrated probabilities. Decisions should combine failure risk, consequence, action cost and the value of additional information.

Successful implementation depends on reliable data, explicit failure definitions, physical understanding, time-aware validation, integration with work management and strong governance. The most useful predictive-maintenance system is not necessarily the most complex; it is the one that improves safe and timely decisions.

Key Terms

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

Condition indicator: A measured or calculated variable associated with equipment health.

Control limit: A statistically derived boundary used to identify unusual process variation.

Concept drift: Change in the relationship between inputs, outputs and operating conditions over time.

Censoring: In survival analysis, incomplete observation of a failure time.

Degradation model: A mathematical representation of progressive health loss.

Diagnosis: Identification of a fault’s type, location or cause.

Failure threshold: A criterion defining unacceptable function or condition.

Health indicator: A variable intended to represent the state of an asset.

Hazard function: Conditional instantaneous rate of failure.

Maintenance analytics: Use of data and models to support maintenance decisions.

Mahalanobis distance: Correlation- and scale-adjusted distance from a reference distribution.

Prognosis: Forecast of future condition, failure risk or RUL.

Remaining useful life: Expected or probabilistic time until a defined failure or intervention threshold.

Residual: Difference between an observed value and a modelled or expected value.

Right-censoring: Observation ending before the event of interest occurs.

Statistical process control: Monitoring method for distinguishing common and special causes of variation.

Trend: Systematic directional movement in a condition indicator.

Uncertainty: Incomplete certainty arising from randomness, measurement, knowledge or future scenarios.

Review Questions

Short Answer Questions

1. Distinguish detection, diagnosis and prognosis.
2. Why should control limits not be confused with specification limits?
3. What is concept drift in a maintenance analytics system?
4. Define remaining useful life and explain why it depends on a threshold.
5. Identify two sources of uncertainty in RUL estimation.

Long Answer Questions

1. Explain how SPC and trend analysis can be combined to monitor a degrading asset.
2. Compare rule-based, statistical and machine-learning approaches to anomaly detection.
3. Discuss the assumptions and limitations of a linear degradation model.
4. Explain how survival analysis differs from a simple deterministic RUL calculation.
5. Develop an implementation framework for predictive maintenance in a process plant.

Analytical / Application Questions

1. A condition indicator is currently 5.4 units, the intervention threshold is 7.0 units, and the estimated degradation rate is 0.08 units per operating hour. Calculate the point RUL and identify two reasons why the result should not be treated as certain.
2. A vibration signal shows six consecutive increases but remains below its upper control limit. What additional checks should an engineer perform before recommending replacement?
3. A predictive model has high overall accuracy but misses most actual failures. Explain why this may occur and propose more appropriate evaluation measures.

Multiple Choice Questions with Answers

Question 1

Which activity primarily constitutes diagnosis?

A. Estimating the time until a threshold is crossed

B. Identifying the likely fault type or cause

C. Calculating the mean of a sensor signal

D. Scheduling all maintenance at fixed intervals

Correct Answer: B. Identifying the likely fault type or cause

Question 2

A process has a baseline mean of 10 units and standard deviation of 0.5 units. What is the approximate upper three-sigma control limit?

A. 10.5 units

B. 11.0 units

C. 11.5 units

D. 12.0 units

Correct Answer: C. 11.5 units

Question 3

Which statement correctly distinguishes control and specification limits?

A. Control limits are always safety limits

B. Specification limits are calculated only from sample variation

C. Control limits describe statistical behaviour, while specification limits represent requirements

D. They are two names for the same boundary

Correct Answer: C. Control limits describe statistical behaviour, while specification limits represent requirements

Question 4

For a health indicator currently at 6 units, a threshold of 8 units and a degradation rate of 0.1 unit per hour, the linear point RUL is:

A. 10 hours

B. 20 hours

C. 60 hours

D. 80 hours

Correct Answer: B. 20 hours

Question 5

What does the Mahalanobis distance account for that ordinary Euclidean distance does not?

A. Only time duration

B. Variable scale and correlation

C. Maintenance labour cost

D. Failure consequence

Correct Answer: B. Variable scale and correlation

Question 6

Which condition most directly creates a false maintenance alarm?

A. A validated fault signature under stable operating conditions

B. A sensor anomaly caused by communication loss

C. A confirmed failure threshold crossing

D. A repeated pattern supported by several sensors

Correct Answer: B. A sensor anomaly caused by communication loss

Question 7

In survival analysis, right-censoring occurs when:

A. A failure is observed twice

B. Observation ends before the failure event is observed

C. A sensor is installed on the right side of an asset

D. A model has no predictor variables

Correct Answer: B. Observation ends before the failure event is observed

Question 8

Which statement about RUL is most accurate?

A. It is always an exact failure date

B. It is independent of the selected failure threshold

C. It is conditional on data, model assumptions and future operating conditions

D. It can be calculated without defining failure

Correct Answer: C. It is conditional on data, model assumptions and future operating conditions

Question 9

Why is time-based validation important for predictive maintenance models?

A. It guarantees that every model is linear

B. It prevents future information from leaking into model development

C. It eliminates measurement noise

D. It makes failures occur more frequently

Correct Answer: B. It prevents future information from leaking into model development

Question 10

Which decision expression best represents a simplified expected-loss comparison?

A. Expected loss = mean + standard deviation

B. Expected loss = probability of failure × failure consequence + action cost

C. Expected loss = control limit − specification limit

D. Expected loss = RUL × sensor sampling rate

Correct Answer: B. Expected loss = probability of failure × failure consequence + action cost