Condition Monitoring and Predictive Maintenance

Introduction

Mechanical systems rarely fail without warning. Friction, imbalance, looseness, misalignment, fatigue, corrosion, contamination and overheating usually produce measurable changes before catastrophic failure. Condition monitoring is the systematic measurement and interpretation of these changes. Predictive maintenance uses the resulting information to decide what maintenance should be performed, when it should be performed and what resources are required.

For mechanical engineers, the subject combines machine dynamics, materials, lubrication, instrumentation, reliability engineering, data analysis and industrial communication. Artificial intelligence and the Internet of Things (IoT) extend these capabilities by enabling continuous sensing, automated anomaly detection and fleet-level analysis. They do not, however, replace engineering judgment: a model must be linked to physical failure mechanisms and validated against evidence.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Explain the relationship among corrective, preventive, condition-based and predictive maintenance.
2. Select suitable vibration, acoustic, oil and thermal measurements for rotating and thermal equipment.
3. Diagnose common faults in bearings, gears, pumps and shafts.
4. Apply reliability concepts, failure-mode analysis and maintenance decision logic.
5. Estimate remaining useful life (RUL) while recognizing uncertainty and model limitations.
6. Design an IoT-enabled condition-monitoring workflow and evaluate AI results responsibly.

1. Maintenance Strategies and the Condition-Monitoring Cycle

Corrective maintenance allows equipment to run until failure. It may be economical for non-critical, inexpensive components, but it can cause secondary damage, safety hazards and production loss. Preventive maintenance replaces or services equipment at fixed time or usage intervals. This reduces some failures but may waste useful life and can introduce maintenance-induced faults.

Condition-based maintenance schedules work when measured condition indicates deterioration. Predictive maintenance is a broader, data-supported approach that forecasts future condition or failure risk. The terms are sometimes used interchangeably, but predictive maintenance normally includes prognosis, uncertainty estimation and planning.

A practical cycle is:

1. Define the asset, function and consequences of failure.
2. Identify credible failure modes.
3. Select measurable physical indicators.
4. Acquire data using an appropriate sampling plan.
5. Clean, synchronize and validate the data.
6. Detect anomalies and diagnose likely causes.
7. Estimate severity, failure probability or RUL.
8. Plan and execute intervention.
9. Confirm the repair and update the model.

A useful system architecture contains sensors, signal conditioning, an edge device, a communications network, a data platform, analytics and a computerized maintenance-management system (CMMS). The CMMS should receive actionable recommendations rather than unexplained alarms.

2. Failure Modes and Reliability Foundations

A failure mode is the manner in which an item fails to perform its required function. Examples include bearing raceway fatigue, gear-tooth pitting, shaft imbalance, seal leakage, pump cavitation and motor overheating. Failure-mode and effects analysis (FMEA) records the cause, local effect, system effect, existing controls and recommended action. Criticality analysis then ranks failures according to safety, environmental, production and financial consequences.

Reliability is the probability that an item performs its required function for a stated time under stated conditions. If T is the random variable representing time to failure, reliability is:

where R(t) is dimensionless and t is time. The cumulative failure distribution is F(t) = P(T ≤ t), so:

The failure-rate or hazard function is:

where f(t) is the probability-density function of failure time. The hazard has units of inverse time and represents the instantaneous failure tendency conditional on survival to time t. It is not itself the probability of failure over an interval.

For a constant hazard λ, the exponential model gives:

and the mean time to failure is 1/λ. This model is appropriate mainly for a useful-life period with approximately random failures; it is not generally appropriate for fatigue-dominated components whose hazard increases with age.

A widely used fatigue-life model is the Weibull distribution:

where β is the shape parameter and η is the characteristic life, the time at which reliability is e⁻¹ ≈ 0.368. β < 1 suggests early-life failures, β ≈ 1 suggests an approximately constant hazard, and β > 1 suggests wear-out. Real components may require different distributions for different life phases.

Mean time between failures (MTBF) is often used for repairable equipment, while mean time to failure (MTTF) is used for non-repairable items. Neither metric alone describes safety, failure consequences, operating conditions or the distribution of failures. A high MTBF can coexist with rare but severe failures.

3. Vibration Monitoring of Rotating Machinery

Vibration is the most widely used condition indicator for rotating machinery because mechanical faults alter forces, stiffness, clearances or contact geometry. An accelerometer produces a signal x(t), commonly measured in acceleration. The root-mean-square (RMS) value for a signal sampled at N points is:

RMS measures overall energy-like severity but does not identify a fault. Crest factor is:

where x_peak is the largest absolute amplitude. A rising crest factor can indicate impulsive bearing damage, although it is influenced by speed, load, sensor mounting and signal bandwidth.

Frequency analysis uses the discrete Fourier transform (DFT) to represent periodic content. For a record of duration T_s, the frequency spacing is approximately:

Longer records improve frequency resolution but reduce the ability to track rapid changes. The sampling frequency f_s must exceed twice the highest frequency of interest to avoid aliasing, according to the Nyquist condition. An anti-aliasing filter is required before digitization.

Shaft speed and characteristic frequencies

If rotational speed is n revolutions per minute, rotational frequency is:

in hertz. A once-per-revolution component at f_r often indicates imbalance, eccentricity or misalignment. Harmonics at 2f_r and higher may indicate misalignment, looseness or nonlinear contact, but diagnosis must consider the machine structure and measurement direction.

For a rolling-element bearing, approximate characteristic frequencies are:

Here FTF is fundamental train frequency, BPFO is ball-pass frequency outer race, BPFI is ball-pass frequency inner race, BSF is ball-spin frequency, N_b is the number of rolling elements, d is rolling-element diameter, D is pitch diameter, and θ is contact angle. These expressions assume ideal geometry, constant speed and negligible slip. Manufacturer databases and actual operating conditions should be used when available.

Envelope analysis demodulates high-frequency resonance excited by repetitive impacts. It is particularly useful for early rolling-element bearing defects. A sensor must be mounted firmly, preferably at a load path close to the bearing. Magnetic or hand-held measurements can be useful for screening but may lack repeatability.

Diagnosing common vibration patterns

A spectrum is not a diagnosis by itself. The engineer should compare spectra over time, relate frequencies to machine geometry and check speed and load. Order tracking is preferable when speed varies because it expresses frequency as multiples of rotational speed.

4. Acoustic Emission and Ultrasound

Acoustic-emission sensors detect transient elastic waves generated by crack growth, friction, impacts or leakage. Ultrasound instruments commonly monitor frequencies above the audible range and convert them into an audible signal or a numerical level. These methods are valuable for slow-speed bearings, compressed-air leaks, steam traps and early lubrication problems where conventional vibration may have low energy.

Acoustic measurements are highly sensitive to attenuation, sensor coupling, background noise and structural transmission paths. They should be trended under comparable operating conditions. An isolated high reading can be caused by process noise rather than damage. Acoustic data are therefore strongest when combined with vibration, temperature, load and inspection evidence.

5. Oil and Wear-Debris Monitoring

Lubricant analysis provides information about both the lubricant and the machine. Important measurements include viscosity, water content, particle count, elemental spectroscopy, ferrous density, oxidation, acidity and additive depletion. Wear particles may indicate cutting wear, fatigue wear, sliding wear or contamination.

The interpretation must distinguish cause from consequence. For example, increased iron may result from gear wear, but it can also reflect a sampling error or contamination from another component. Sampling location, cleanliness, timing and laboratory method are essential. Samples should be taken from a well-mixed, active-flow location—not from the bottom of a stagnant reservoir—unless sediment analysis is specifically intended.

Oil analysis is especially effective for gearboxes, hydraulic systems, compressors and large bearings. It complements vibration: vibration can identify dynamic behavior, while oil debris can reveal wear before a strong mechanical signature appears.

6. Thermal Monitoring and Pump Diagnosis

Temperature is a direct indicator of friction, overload, poor lubrication, electrical loss, blocked flow and insulation problems. Infrared thermography detects surface temperature patterns without physical contact. Correct interpretation requires attention to emissivity, reflected radiation, viewing angle, atmospheric attenuation and environmental conditions.

A temperature alarm should be based on a baseline established at comparable speed, load and ambient temperature. Absolute limits may be supplied by the manufacturer, but rate of change and temperature difference between comparable components are often more informative.

For pumps, useful measurements include suction and discharge pressure, flow, motor current, vibration, bearing temperature and seal leakage. The hydraulic head developed by a pump is:

where H is head in metres, p is pressure, ρ is fluid density, g is gravitational acceleration, v is mean velocity and z is elevation. The equation assumes steady flow and compares corresponding sections; losses may need to be included when evaluating the complete system.

Cavitation occurs when local absolute pressure falls near or below the fluid vapour pressure. It commonly produces broadband noise, unstable flow, vibration and pitting. A simplified available net positive suction head is:

where p_s,abs is absolute suction pressure and p_v is vapour pressure. The pump requires NPSH_R specified by the manufacturer, and adequate margin is needed. Gauge pressure must not be used in place of absolute pressure in this calculation.

7. Data Acquisition and IoT Architecture

A condition-monitoring system must be designed from the decision backward. If the decision is whether to replace a bearing during the next planned shutdown, a daily low-frequency measurement may be adequate. If the decision concerns protection from a sudden overspeed event, high-speed real-time instrumentation is required.

The measurement chain includes:

• sensor selection and mounting;
• signal conditioning and analog-to-digital conversion;
• timestamping and synchronization;
• edge processing and local buffering;
• secure network transmission;
• storage of raw and derived data;
• analytics, visualization and work-order integration.

Sampling plans should record speed, load, process state, sensor identity, calibration status and environmental conditions. Data without context can generate misleading alarms. Time synchronization is important when vibration, temperature and process variables are fused.

IoT deployments should use device authentication, encrypted communication, least-privilege access, signed firmware and network segmentation. Edge analytics can reduce bandwidth and continue operation during network loss, but raw data should be retained periodically for audit and model improvement.

8. Artificial Intelligence for Detection and Diagnosis

AI methods are useful when data volume or complexity exceeds manual analysis. Common approaches include:

• threshold and control-chart methods for transparent monitoring;
• principal component analysis for multivariable dimensionality reduction;
• clustering for operating-state discovery;
• supervised classification when labelled fault data exist;
• autoencoders and other anomaly-detection models when failure labels are scarce;
• regression and survival models for prognosis.

Feature engineering may use RMS, kurtosis, crest factor, spectral amplitudes, envelope energy, temperature gradients, oil-particle counts and process variables. Features should be calculated using windows that preserve the relevant physical information.

A model must be evaluated using time-based splits. Randomly mixing observations from the same machine into training and test sets can cause leakage and produce unrealistically high accuracy. Useful metrics include precision, recall, false-alarm rate, missed-detection rate, detection lead time and calibration of predicted probabilities. Accuracy alone is inadequate for rare failures.

The model should distinguish anomaly detection from fault diagnosis. An anomaly means that behavior differs from the learned reference; it does not prove a specific failure mode. Diagnosis requires physical interpretation and, ideally, corroborating evidence.

9. Remaining Useful Life and Prognostics

RUL is the expected or predicted time until a defined failure criterion, conditional on current evidence. The failure criterion must be explicit: loss of performance, vibration limit, temperature limit, leakage rate or inability to meet a safety function.

A simple degradation model uses a health indicator y(t):

where a is the initial level, b is the degradation slope, and ε(t) is measurement and process noise. If failure occurs at threshold y_lim and b > 0, the estimated RUL from the current time t_c is:

This is a simplified extrapolation. It assumes a stable degradation trend, a meaningful health indicator, a known threshold and approximately constant slope. It should not be presented as a guaranteed failure time.

A probabilistic prognostic system produces a distribution p(RUL | data), not merely a point estimate. Uncertainty arises from sensor noise, unknown loads, model error, unit-to-unit variation and changing operating conditions. Maintenance decisions should use prediction intervals and consequence-based risk, not only the mean RUL.

10. Worked Example: Bearing-Health Trend

An accelerometer produces a bearing health indicator in mm/s RMS. At day 10 the value is 2.4 mm/s; at day 40 it is 3.6 mm/s. A warning threshold is 5.0 mm/s. Assume, only for this illustrative calculation, a linear trend and constant operating conditions.

The slope is:

At day 40, the estimated time to the threshold is:

Thus the estimated threshold crossing is around day 75. This does not mean that failure will occur exactly on day 75. The engineer should verify sensor mounting, speed and load, inspect the spectrum and envelope data, assess the bearing’s criticality and plan an intervention before the risk becomes unacceptable. If the slope accelerates, the linear estimate becomes optimistic.

11. Maintenance Decision-Making

A useful decision combines condition, consequence and intervention opportunity. A low-severity anomaly on a redundant fan may justify continued monitoring. The same anomaly on a turbine bearing may require immediate controlled shutdown.

Risk can be expressed conceptually as:

This relationship is a decision framework rather than a complete probabilistic model. Consequence may include safety, environmental, quality, production and repair costs. Maintenance optimization can compare the expected cost of intervention with the expected cost of failure, including downtime and secondary damage.

Alarm design should include at least three states: normal, alert and action. Alarm limits should be based on engineering standards, manufacturer guidance, historical baselines and the consequences of error. Alarm flooding is a common failure of poorly designed systems; each alarm should have a defined response and owner.

12. Illustrative Case: Cooling-Pump Bearing Degradation

Background

A process plant operates four centrifugal cooling pumps, three required for normal production. Each pump has rolling-element bearings and a variable-speed drive. Vibration and bearing temperature are collected every hour; oil samples are taken monthly.

Problem

Pump P-204 develops gradually increasing vibration. The overall RMS value rises from 1.8 to 3.1 mm/s over six weeks. The temperature increases by 8 °C relative to the parallel pumps, while flow and discharge pressure remain normal.

Analysis

The maintenance engineer first checks sensor mounting, speed, load and data continuity. The vibration spectrum shows a rising component near the calculated BPFO and the envelope spectrum shows repetitive impacts. Oil analysis reports increased ferrous debris but normal viscosity and water content. The evidence is consistent with outer-race bearing deterioration rather than cavitation, because pump hydraulic variables remain stable and no broadband cavitation signature is observed.

Approach and solution

The team assigns P-204 an action status because the pump is important but redundant capacity exists. A replacement bearing, seal kit, tools and qualified personnel are scheduled during the next production window. The parallel pumps are monitored more frequently. Before replacement, a spare pump is confirmed and a contingency shutdown plan is prepared.

After replacement, vibration returns to 1.7 mm/s, temperature matches the parallel pumps and follow-up oil analysis shows declining debris. Inspection of the removed bearing finds an outer-race fatigue spall. The team investigates lubrication quantity, alignment, installation practice and shaft fit rather than treating the bearing replacement as the complete solution.

Outcome and learning points

The intervention avoided an unplanned outage and secondary shaft damage. The case demonstrates that a credible diagnosis uses trend, frequency, thermal and oil evidence together. It also shows why a threshold alone is insufficient: the rising trend and bearing-specific frequencies provided earlier and more actionable information.

13. Challenges, Ethics and Governance

Condition-monitoring programs face false alarms, missed detections, sensor failure, changing operating regimes, insufficient failure labels and concept drift. A model trained on one machine may not transfer to another because of geometry, foundation stiffness, speed range or maintenance practice.

Engineers must not conceal uncertainty behind an AI-generated score. Automated recommendations should be traceable to data, features, model version and decision rules. Human approval is essential for safety-critical shutdowns unless the protection system has been separately validated to the relevant safety requirements.

Data governance includes ownership, retention, cybersecurity, access control and auditability. Worker-related data, such as operator actions or wearable measurements, may raise privacy and labor concerns. Monitoring should be proportionate, transparent and used primarily to improve equipment and safety rather than to impose unjustified individual surveillance.

Standards and regulations vary by industry and jurisdiction. Organizations should consult applicable vibration, machinery safety, functional-safety, electrical, pressure-equipment, environmental and cybersecurity requirements. Manufacturer limits and recognized standards should take precedence over arbitrary universal thresholds.

14. Best Practices and Contemporary Developments

Effective programs:

• begin with criticality and failure-mode analysis;
• measure only what supports a maintenance decision;
• establish clean baselines for operating states;
• combine complementary sensing modalities;
• trend data rather than relying on isolated values;
• validate alarms in the field;
• preserve raw data and metadata;
• use time-aware model validation;
• quantify uncertainty and define escalation rules;
• close the loop by confirming repairs.

Digital twins increasingly combine physics-based models with live sensor data. Physics-informed machine learning can constrain predictions to plausible mechanical behavior. Wireless energy-efficient sensors and edge AI improve deployment flexibility, while federated learning may allow multiple sites to improve models without centrally sharing all raw data. These developments remain subject to practical constraints: battery life, connectivity, cybersecurity, explainability and the scarcity of representative failure data.

Chapter Summary

Condition monitoring measures physical evidence of deterioration, while predictive maintenance converts that evidence into diagnosis, prognosis and maintenance action. Vibration is central to rotating-machinery diagnosis, with RMS, spectra, envelopes and characteristic frequencies supporting analysis of imbalance, misalignment, looseness, bearings and gears. Acoustic emission detects impacts and leakage; oil analysis reveals lubricant condition and wear debris; thermal methods identify friction, overload and flow-related problems. Pump diagnosis requires both mechanical signals and hydraulic reasoning, including correct use of absolute pressure in NPSH calculations.

Reliability functions, hazard rate, Weibull behavior, FMEA and criticality analysis provide the decision framework. RUL estimates are conditional predictions with uncertainty, not exact failure dates. IoT architecture must preserve measurement quality, context, cybersecurity and integration with maintenance workflows. AI can detect patterns and support prognosis, but physical validation, time-aware testing, human accountability and transparent governance remain essential.

Key Terms

Condition monitoring: Measurement and interpretation of equipment condition during operation.

Predictive maintenance: Maintenance planning supported by condition data and forecasts of future deterioration or failure.

Failure mode: The specific manner in which an item fails to perform its required function.

FMEA: A structured analysis of failure modes, causes, effects and controls.

Reliability: Probability that an item performs its required function for a stated time and condition.

Hazard rate: Instantaneous failure tendency conditional on survival to a given time.

Weibull distribution: A flexible life distribution whose shape parameter describes early-life, random or wear-out behavior.

RMS: Root-mean-square magnitude of a measured signal.

Crest factor: Ratio of peak amplitude to RMS amplitude.

Order tracking: Analysis expressed as multiples of rotational speed.

Envelope analysis: Demodulation method used to detect repetitive impacts, especially in bearings.

BPFO: Ball-pass frequency of the outer race.

BPFI: Ball-pass frequency of the inner race.

Acoustic emission: Elastic-wave energy generated by sudden deformation, cracking, impact or friction.

Wear debris: Particles produced by mechanical wear and carried in lubricant.

Thermography: Non-contact measurement and interpretation of infrared radiation.

Remaining useful life: Predicted time until a defined failure or performance threshold.

Concept drift: Change in data relationships over time that reduces model validity.

False alarm: Indication of a fault when the relevant fault is absent.

Criticality: Relative importance of an asset or failure based on consequences and risk.

Review Questions

Short Answer Questions

1. Distinguish corrective, preventive, condition-based and predictive maintenance.
2. What information does an envelope spectrum provide for a rolling-element bearing?
3. Why must absolute pressure be used in an NPSH calculation?
4. Explain the difference between RUL and failure time certainty.
5. State two reasons why an AI model may produce unreliable maintenance recommendations.

Long Answer Questions

1. Explain how vibration, oil, acoustic and thermal data can be combined to diagnose a gearbox or pump.
2. Derive the relationship between reliability and the cumulative failure distribution, and explain the meaning of hazard rate.
3. Describe a complete IoT condition-monitoring architecture from sensor to CMMS work order.
4. Compare threshold-based monitoring, supervised classification and anomaly detection for industrial equipment.
5. Discuss the engineering, ethical and cybersecurity requirements for deploying AI-based predictive maintenance.

Analytical / Application Questions

1. A bearing health indicator rises from 2.0 to 3.2 mm/s RMS over 24 days. If the action threshold is 5.0 mm/s and a constant linear trend is assumed, estimate the threshold-crossing time from the latest observation. State two reasons why the estimate may be inaccurate.
2. A pump shows increased broadband vibration, unstable discharge pressure and reduced flow, while suction pressure approaches vapour-pressure conditions. Diagnose the likely fault and recommend measurements or actions to confirm it.
3. A gearbox model gives excellent random-test accuracy but fails when deployed on a new machine and speed range. Analyze the likely causes and propose a revised validation and deployment strategy.

Multiple Choice Questions with Answers

Question 1

Which measurement is most directly associated with early repetitive impacts from a rolling-element bearing defect?

A. Envelope spectrum

B. Ambient humidity

C. Static shaft length

D. Paint thickness

Correct Answer: A. Envelope spectrum

Question 2

For a constant hazard rate λ, which reliability model is appropriate?

A. R(t) = e^(−λt)

B. R(t) = λt

C. R(t) = 1 + λt

D. R(t) = λ/t

Correct Answer: A. R(t) = e^(−λt)

Question 3

A shaft rotates at 1800 rpm. What is its rotational frequency?

A. 18 Hz

B. 30 Hz

C. 60 Hz

D. 1800 Hz

Correct Answer: B. 30 Hz

Question 4

In the Weibull model, a shape parameter β greater than 1 generally indicates:

A. Wear-out behavior with increasing hazard

B. Only random failures with constant hazard

C. Guaranteed absence of failure

D. A decreasing early-life hazard

Correct Answer: A. Wear-out behavior with increasing hazard

Question 5

Which condition is most characteristic of pump cavitation?

A. Stable flow and a single clean 1× component

B. Broadband noise, unstable flow and pressure near vapour pressure

C. Zero temperature variation under all loads

D. Only an increase in lubricant viscosity

Correct Answer: B. Broadband noise, unstable flow and pressure near vapour pressure

Question 6

If a measurement record lasts 2 seconds, its approximate Fourier frequency spacing is:

A. 0.5 Hz

B. 2 Hz

C. 20 Hz

D. 200 Hz

Correct Answer: A. 0.5 Hz

Question 7

Why is random mixing of time-series observations from the same machine a poor test strategy?

A. It can cause data leakage and overestimate generalization

B. It always removes all useful features

C. It prevents any model from being trained

D. It guarantees excessive false negatives

Correct Answer: A. It can cause data leakage and overestimate generalization

Question 8

A health indicator is 3.6 mm/s, the limit is 5.0 mm/s, and the estimated degradation slope is 0.04 mm/s per day. Under a linear assumption, the estimated RUL is:

A. 14 days

B. 25 days

C. 35 days

D. 125 days

Correct Answer: C. 35 days

Question 9

Which statement best describes an anomaly detected by an AI model?

A. It proves a specific bearing race has failed

B. It shows behavior differs from the learned reference and requires investigation

C. It eliminates the need for physical inspection

D. It is always equivalent to a safety trip

Correct Answer: B. It shows behavior differs from the learned reference and requires investigation

Question 10

Which practice most improves the engineering usefulness of condition-monitoring data?

A. Recording sensor values without speed or load information

B. Using one universal threshold for every machine

C. Synchronizing measurements with operating context and maintenance records

D. Discarding raw data after generating an alarm

Correct Answer: C. Synchronizing measurements with operating context and maintenance records