13. Condition Monitoring, Triboinformatics and AI

Introduction and Context

Tribological failures rarely occur without warning. Long before a bearing seizes, a gear tooth fractures, or a hydraulic actuator loses efficiency, changes usually appear in lubricant chemistry, wear-particle population, vibration, temperature, acoustic emission, or electrical behaviour. Condition monitoring seeks to detect and interpret these changes so that maintenance can be planned according to actual equipment condition rather than fixed operating time.

In advanced mechanical systems, no single measurement is normally sufficient. Oil analysis may reveal chemical degradation and metallic debris, while vibration identifies dynamic faults. Temperature indicates frictional or lubrication-related heat generation, acoustic emission can detect short-duration impacts, and motor-current or stray-voltage measurements may expose electromechanical and lubrication-related abnormalities. The integration of these sources creates a more reliable representation of tribological health.

The growth of sensors, cloud platforms, edge computing, digital twins, and machine learning has given rise to triboinformatics: the systematic collection, organisation, analysis, and application of data concerning friction, wear, lubrication, surfaces, materials, and operating conditions. Triboinformatics does not replace tribological knowledge. Rather, it combines physical understanding with data-driven methods to improve diagnosis, prognosis, design, and maintenance decisions.

This chapter examines the architecture of an integrated condition-monitoring system, methods of data interpretation, digital-twin and prognostic frameworks, and the opportunities and risks associated with artificial intelligence. Particular attention is given to data quality, explainability, cybersecurity, governance, and the continuing need for expert human judgement.

Learning Objectives

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

1. Explain the purpose and scope of condition monitoring in tribological systems.

2. Distinguish the information provided by oil, wear-debris, vibration, acoustic, thermal, and electrical measurements.

3. Design a multimodal monitoring workflow for machinery.

4. Describe triboinformatics, digital twins, fault diagnosis, and remaining-useful-life estimation.

5. Select suitable machine-learning approaches for labelled, unlabelled, and time-dependent data.

6. Evaluate data quality, uncertainty, explainability, and model drift.

7. Assess cybersecurity, ethical, regulatory, and organisational issues in AI-supported maintenance.

8. Formulate best practices for implementing a human-centred condition-monitoring programme.

From Periodic Inspection to Continuous Tribological Intelligence

Traditional maintenance strategies are often classified as reactive, preventive, or predictive. Reactive maintenance allows failure to occur; preventive maintenance replaces or services components at prescribed intervals; predictive maintenance uses measurements to estimate condition and intervene when evidence indicates an approaching failure. Prescriptive maintenance goes further by comparing possible actions and recommending an intervention that balances risk, cost, production, and safety.

A condition-monitoring system has four connected functions: sensing, interpretation, decision, and action. Sensing produces observations. Interpretation converts observations into indicators or diagnoses. Decision analysis determines whether the evidence justifies inspection, load reduction, lubrication, repair, or continued operation. Action closes the loop and should subsequently generate information about whether the intervention was effective.

The central distinction is between a measurement and a condition indicator. A raw vibration waveform is a measurement; root-mean-square acceleration, kurtosis, an envelope-spectrum peak, or a bearing-fault probability is an indicator derived from it. Indicators become useful only when they are related to operating state, baseline behaviour, failure mechanisms, and decision thresholds.

The Evidence Obtained from Different Sensing Modalities

Lubricant and wear-debris analysis

Lubricant analysis provides both a chemical history and a record of interactions between surfaces. Important tests include viscosity, acid number, base number, water content, oxidation or nitration indicators, additive depletion, particle count, ferromagnetic debris, elemental spectroscopy, ferrography, and filter-residue examination.

Viscosity changes can indicate contamination, oxidation, thermal degradation, or the use of an incorrect lubricant. Water may promote corrosion, additive depletion, and loss of film strength. Elemental analysis can identify the likely source of metals, although it may be less sensitive to large particles. Ferrography and analytical ferrography help distinguish rubbing wear, cutting wear, fatigue particles, and severe sliding wear through particle morphology and concentration.

Sampling practice is decisive. A sample taken from a stagnant drain point may not represent the lubricant circulating through a loaded bearing. Sample location, flushing procedure, container cleanliness, timing, operating temperature, and record of oil addition must be standardised. A trend from comparable samples is generally more informative than an isolated value.

Vibration and dynamic response

Vibration monitoring is particularly effective for rolling-element bearings, gears, shafts, couplings, and imbalance or misalignment problems. Time-domain measures include peak value, crest factor, kurtosis, and root-mean-square value. Frequency-domain analysis uses the discrete Fourier transform to reveal running speed, harmonics, gear-mesh frequencies, sidebands, and bearing characteristic frequencies. Envelope analysis demodulates repetitive impacts that may be hidden in the ordinary spectrum.

Interpretation must account for speed, load, sensor mounting, structural transmission paths, and resonance. A large amplitude does not automatically identify a fault, and a small amplitude does not guarantee safety. Order tracking is valuable when rotational speed varies because it represents components by rotational order rather than fixed frequency.

Acoustic emission

Acoustic-emission sensing captures transient elastic waves produced by events such as crack growth, asperity fracture, plastic deformation, and impacts. It can be more sensitive than conventional vibration to early-stage surface damage, especially at high frequencies. However, signals are strongly affected by attenuation, sensor coupling, background noise, and the structure through which waves travel. Acoustic emission therefore requires careful calibration and often benefits from event-rate and energy analysis rather than simple amplitude comparison.

Thermal measurements

Temperature is a useful integrative indicator of friction, lubrication, load, and heat rejection. Contact thermocouples, resistance sensors, infrared cameras, and thermal imaging can identify hot bearings, blocked lubrication passages, abnormal seals, or overloaded gear meshes. Temperature is comparatively easy to interpret but often changes after damage has become significant. Ambient temperature, airflow, emissivity, load, and measurement location must be considered.

Electrical measurements

Electrical signals provide indirect information about mechanical and tribological behaviour. Motor-current signature analysis can detect load modulation caused by eccentricity, misalignment, bearing damage, or gear faults. Electrical discharge machining-like damage may occur when shaft voltages pass through rolling bearings in variable-frequency-drive systems. Stray-current monitoring and insulated-bearing strategies are consequently relevant to modern electric drives.

Electrical sensing is not a substitute for mechanical evidence. Inverter switching, supply imbalance, control algorithms, and electromagnetic interference can produce patterns resembling mechanical faults. Electrical and vibration data are strongest when interpreted together with speed, load, and drive-control information.

Multimodal Fusion: From Signals to Decisions

Sensor fusion may be performed at three levels. In data-level fusion, synchronised raw or minimally processed signals are combined. This approach preserves information but requires compatible sampling rates, accurate time synchronisation, and substantial computational capacity. Feature-level fusion combines indicators such as viscosity, iron concentration, envelope energy, temperature slope, and motor-current sideband amplitude. Decision-level fusion combines independent diagnostic outputs, for example by weighted voting, Bayesian updating, or a probabilistic risk model.

A practical workflow begins with a failure-mode and effects analysis or a reliability-centred maintenance study. The engineer identifies credible failure mechanisms, their observable precursors, the time available for intervention, and the consequences of missed detection. Sensors are then selected according to failure physics rather than convenience. Data are timestamped with operating variables such as speed, torque, temperature, lubricant identity, maintenance history, and environmental conditions.

A baseline is established during known healthy operation. Thresholds may be absolute, statistical, model-based, or derived from standards and manufacturer guidance. Adaptive thresholds are useful where operating conditions vary, but they must not adapt so aggressively that a slowly developing fault becomes the new normal. Alarm logic should distinguish warning, action, and emergency states and should define the required human response.

Triboinformatics and the Organisation of Knowledge

Triboinformatics extends ordinary condition monitoring by treating tribological information as a structured, reusable body of knowledge. Its data may include material properties, surface topography, lubricant formulation, additive chemistry, contact pressure, sliding speed, temperature, wear rate, test configuration, field sensor records, maintenance actions, and failure photographs.

The value of a database depends on metadata. A wear-rate value without counterface material, roughness, load, speed, atmosphere, lubricant, test duration, and measurement uncertainty is difficult to compare with another value. Ontologies and controlled vocabularies can express relationships such as “lubricant protects steel under boundary lubrication” or “particle morphology is associated with fatigue wear.” Such structures support searching, knowledge graphs, cross-machine comparison, and reuse of experimental results.

Triboinformatics also requires attention to scale. Laboratory tribometers provide controlled evidence but may not reproduce contamination, vibration, thermal gradients, or maintenance practices in the field. Field data are realistic but often noisy and incomplete. Combining both requires domain knowledge and explicit representation of uncertainty rather than treating all observations as equivalent.

Digital Twins and Prognostic Reasoning

A digital twin is a continuously updated digital representation of a physical asset, its operating environment, and relevant degradation processes. A simple monitoring dashboard is not necessarily a digital twin. A twin should connect measurements with a model of behaviour, state estimation, and a defined relationship to the physical asset.

Digital twins may be physics-based, data-driven, or hybrid. A physics-based model can represent contact force, film thickness, thermal balance, stress, fatigue accumulation, or lubricant degradation. A data-driven model learns relationships from historical observations. A hybrid model uses physical equations to constrain or interpret machine-learning predictions. Hybrid approaches are especially attractive when failure data are scarce, because physical knowledge reduces implausible predictions.

State estimation methods, including Kalman filtering and particle filtering, combine imperfect measurements with a dynamic model to estimate unobserved variables such as wear depth, friction coefficient, or bearing health. Prognosis then estimates the probability distribution of remaining useful life (RUL), rather than presenting a single apparently precise date. RUL uncertainty increases when operating conditions change, the failure mode is unknown, or the model has not encountered comparable degradation.

A useful prognostic architecture includes: asset identity; sensor and maintenance history; operating-context data; health-index calculation; degradation model; uncertainty estimate; risk-based decision rule; and feedback after inspection or repair. The architecture should also record why an alarm was generated and which evidence supported the recommendation.

Machine Learning for Tribological Diagnosis

Supervised learning is appropriate when labelled examples of normal and faulty conditions are available. Classification models may identify imbalance, misalignment, lubrication starvation, bearing damage, or gear damage. Regression models estimate quantities such as wear rate, temperature, friction, or RUL. Suitable algorithms include logistic regression, support-vector machines, random forests, gradient-boosting methods, and neural networks.

Unsupervised and semi-supervised learning are often more realistic because failures are rare and labels may be uncertain. Clustering can identify operating regimes or previously unrecognised condition groups. Autoencoders and other representation-learning methods can learn normal behaviour and flag reconstruction errors. Novelty detection is useful when the question is not “which known fault is present?” but “has the machine departed from its healthy pattern?”

Time-series models must account for sequence, lag, seasonality, and changing load. Recurrent neural networks, temporal convolutional models, transformers, and state-space models can represent temporal dependencies, but greater complexity does not guarantee better engineering performance. A transparent model with stable performance across machines may be preferable to a highly accurate model that fails under domain shift.

Model evaluation must avoid leakage. Randomly splitting adjacent samples from the same degradation event into training and test sets can produce an unrealistically high score because nearly identical information appears in both sets. Better practice is to separate data by machine, time period, operating campaign, or failure event. Relevant metrics include precision, recall, false-alarm rate, missed-detection rate, calibration, detection lead time, and economic consequence—not accuracy alone.

Explainability, Uncertainty, and Human Oversight

A maintenance engineer must be able to challenge an AI recommendation. Explainability may involve feature importance, local surrogate explanations, counterfactual reasoning, saliency maps, trend plots, or comparison with a healthy baseline. These methods explain aspects of a model but do not prove causal validity. An important distinction exists between a model that identifies a correlation and a tribological explanation grounded in contact mechanics, lubrication regime, or material damage.

Uncertainty should be explicit. Sources include sensor noise, missing data, uncertain labels, variation between machines, model approximation, and future operating conditions. Prediction intervals, confidence or credibility measures, ensemble variation, and out-of-distribution detection can indicate when the system should defer to a human specialist.

Human oversight is not merely an approval step. Experts define failure modes, validate alarms, inspect ambiguous cases, and provide feedback for model improvement. A sound procedure specifies who may override an alarm, how overrides are recorded, when an independent inspection is mandatory, and how the model is withdrawn if it behaves unexpectedly.

An Illustrative Case: Integrated Monitoring of a Wind-Turbine Gearbox

Background

Consider a fictional offshore wind farm operator responsible for medium-speed gearboxes. Gearbox replacement requires a costly crane operation and prolonged production loss. The operator already collects supervisory-control data and periodic vibration measurements, but several failures have been detected only after substantial damage has occurred.

Situation and problem

A gearbox begins to show a small increase in bearing-envelope energy during periods of high torque. The temperature remains within the conventional alarm limit. An oil sample shows a gradual increase in ferrous debris, but the result is close to the historical range. Motor-current data show weak modulation at a frequency associated with the gearbox input stage. None of the individual indicators is decisive.

Analysis

The engineering team first verifies sampling quality and confirms that the vibration sensor has not loosened. They align the records by time and operating regime, separating low-, medium-, and high-torque periods. A health index is created from normalised envelope energy, debris concentration, temperature slope, and current-modulation features. A Bayesian decision model combines the evidence with the known consequence of gearbox failure. The result indicates elevated probability of an incipient bearing problem, with substantial uncertainty.

Approach and solution

The digital twin compares the observed trends with simulated load histories and previously observed degradation trajectories. It does not issue an automatic “replace” command. Instead, it recommends an intensified sampling interval, an inspection during the next planned access opportunity, and temporary avoidance of prolonged operation at the highest torque where production requirements permit. An engineer reviews the spectral evidence, the oil-particle morphology, and the model explanation.

Inspection confirms early surface fatigue and incipient spalling on a bearing raceway. The bearing is replaced during planned intervention. The operator then labels the event, updates the degradation database, and evaluates whether the warning lead time and false-alarm performance met the maintenance objective.

Outcome and learning points

The important result is not that AI identified the fault independently. It is that complementary evidence increased confidence before temperature or overall vibration reached a conventional alarm level. The case demonstrates the importance of operating-context normalisation, multimodal corroboration, uncertainty reporting, planned human inspection, and feedback into the data system. It also illustrates that a risk-based recommendation can be more valuable than a forced binary diagnosis.

Implementation Challenges and Failure Modes of the Monitoring System

Poor data quality is a principal threat. Common problems include sensor drift, incorrect calibration, inconsistent oil sampling, missing timestamps, changing sensor locations, duplicated records, communication failures, and unrecorded maintenance. Labels may also be unreliable: a component replaced “as a precaution” should not automatically be treated as a confirmed failure.

Class imbalance is unavoidable in many industrial datasets. Normal operation produces enormous quantities of data, while severe failures are rare. Oversampling or synthetic data generation can help, but it may also create unrealistic examples. Cost-sensitive learning, anomaly detection, physics-based augmentation, and carefully designed field trials are preferable to indiscriminate duplication.

Concept drift occurs when lubricant formulation, machine design, load profile, sensor hardware, software, or maintenance policy changes. A model must therefore be monitored after deployment. Performance dashboards should track missingness, feature distributions, alarm rates, calibration, lead time, and false alarms. Retraining should be governed by evidence, not by an arbitrary calendar alone.

Other common mistakes include installing sensors without defining a decision, using universal thresholds for unlike machines, confusing correlation with mechanism, ignoring operating context, treating a dashboard as a digital twin, and optimising a model for laboratory data while deploying it in the field. Excessive alarms create alarm fatigue and eventually cause personnel to disregard the system.

Cybersecurity, Governance, and Professional Responsibility

Connected monitoring systems enlarge the attack surface of industrial equipment. Threats include unauthorised access, manipulation of sensor data, ransomware, denial of service, malicious model updates, and theft of proprietary operational information. A false healthy signal could delay maintenance; a false alarm could cause unnecessary shutdown or be used to disrupt production.

Security should include asset inventory, network segmentation, least-privilege access, encrypted communication, secure authentication, signed software and model updates, backup and recovery procedures, logging, vulnerability management, and incident-response plans. Edge processing can reduce exposure of raw data, but it does not eliminate the need for secure architecture.

Governance must define data ownership, retention, access rights, model validation, auditability, and responsibility for decisions. Safety-critical actions should not depend on an opaque model without independent safeguards. Professional responsibility requires engineers to communicate uncertainty honestly, preserve inspection evidence, protect confidential data, and avoid presenting probabilistic predictions as guarantees.

Emerging Directions and Best Practices

Emerging systems are moving toward self-supervised learning from large quantities of unlabelled sensor data, federated learning across sites without centralising sensitive data, physics-informed neural networks, edge AI, and multimodal foundation models. Advanced imaging and automated particle classification may improve wear-debris interpretation. Digital thread technologies may connect design, manufacturing, service, and failure records throughout the asset life cycle.

These developments should be adopted selectively. A practical implementation normally follows this sequence:

• Define the failure mechanisms, decision consequences, and acceptable intervention lead time.
• Establish measurement procedures, metadata standards, calibration routines, and data ownership.
• Create a trustworthy healthy baseline across relevant loads, speeds, and environmental states.
• Combine complementary modalities rather than collecting sensors without a diagnostic purpose.
• Begin with interpretable indicators and transparent rules before introducing complex models.
• Validate on unseen machines, time periods, and operating regimes.
• Quantify uncertainty, false alarms, missed detections, and economic outcomes.
• Provide engineers with evidence, explanations, and escalation procedures.
• Monitor model and sensor performance after deployment.
• Secure the complete data and control chain and maintain human authority over safety-critical decisions.

Chapter Summary

Condition monitoring transforms physical evidence of tribological change into maintenance intelligence. Lubricant and debris analysis reveal chemical degradation and material removal; vibration describes dynamic behaviour; acoustic emission detects transient damage; thermal sensing indicates heat generation; and electrical measurements provide indirect evidence of mechanical and bearing-related faults. Their value increases when synchronised with speed, load, environmental, and maintenance information.

Triboinformatics organises these observations with experimental and materials knowledge, while digital twins connect measurements to asset models and degradation states. Machine learning supports classification, anomaly detection, regression, and remaining-useful-life estimation, but successful deployment depends on representative data, sound validation, uncertainty quantification, explainability, and monitoring for drift. Cybersecurity, governance, professional judgement, and human oversight are essential because an AI system participates in operational decisions but does not bear responsibility for them.

The most effective programmes are failure-mechanism-centred, multimodal, risk-based, and continuously improved through inspection feedback. Their objective is not simply to predict failure, but to support safe, timely, and economically defensible decisions.

Key Terms

Condition monitoring: Systematic observation of equipment condition to detect degradation and guide maintenance.

Triboinformatics: Organisation and analysis of data and knowledge concerning friction, wear, lubrication, surfaces, and materials.

Wear debris: Material particles generated by surface interaction and transported in or collected from a lubricant.

Ferrography: Separation and examination of magnetic and non-magnetic wear particles from lubricant.

Envelope analysis: Signal-processing method that reveals repetitive impact modulation in vibration data.

Acoustic emission: High-frequency elastic-wave activity generated by rapid deformation, cracking, or impact.

Motor-current signature analysis: Interpretation of electrical-current patterns to infer mechanical or electromechanical condition.

Health index: A combined numerical indicator representing the estimated condition of an asset.

Digital twin: An updated digital representation of a physical asset, its environment, and relevant behaviour.

State estimation: Estimation of unmeasured system states from models and imperfect observations.

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

Supervised learning: Machine learning using labelled examples to predict classes or continuous values.

Unsupervised learning: Learning patterns or structure from data without specified fault labels.

Concept drift: Change over time in the data-generating process or relationship learned by a model.

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

Explainability: Ability to describe the evidence or model behaviour supporting a prediction.

Model calibration: Agreement between predicted probabilities and observed frequencies.

Multimodal fusion: Combination of information from different sensing or data modalities.

Human oversight: Deliberate expert review, intervention, and accountability in an automated decision process.

Remaining-risk decision: Maintenance decision based on predicted condition, consequence, uncertainty, and available actions.

Review Questions

Short Answer Questions

1. Why is a trend from properly controlled oil samples generally more useful than one isolated oil-analysis result?

2. Distinguish a raw measurement from a condition indicator, giving one example of each.

3. What information can acoustic-emission sensing provide that may not be visible in ordinary vibration data?

4. Define concept drift in an industrial machine-learning application.

5. Why should remaining useful life be expressed as a distribution or interval rather than always as a single number?

Long Answer Questions

1. Explain how oil analysis, vibration, thermal, acoustic, and electrical measurements can be integrated into a tribological condition-monitoring programme.

2. Describe the architecture of a digital-twin-based prognostic system for a rolling-element bearing.

3. Compare supervised, unsupervised, and semi-supervised learning for industrial tribological diagnosis.

4. Discuss the principal data-quality problems that can invalidate an AI-based maintenance model and propose controls for them.

5. Explain why cybersecurity, explainability, governance, and human oversight are engineering requirements rather than optional additions.

Analytical / Application Questions

1. A gearbox shows rising ferrous debris and intermittent envelope-spectrum peaks, but its temperature remains stable. Design a measurement-validation and maintenance decision process, including uncertainty and escalation criteria.

2. A model achieves 98% accuracy after randomly splitting millions of adjacent vibration records into training and test sets. Critically assess this result and propose a more credible validation design.

3. Develop a multimodal health index for an electric-motor bearing system. Specify the inputs, operating-context variables, fusion method, alarm logic, and procedure for reviewing false alarms.

22. Multiple Choice Questions with Answers

Question 1

Which condition-monitoring practice most improves the interpretability of lubricant-analysis trends?

A. Taking samples at changing locations to increase variety

B. Standardising sampling location, procedure, timing, and metadata

C. Using only one sample after a component failure

D. Ignoring lubricant additions because they do not affect trends

Correct Answer: B. Standardising sampling location, procedure, timing, and metadata

Question 2

Why is envelope analysis useful for rolling-element bearing diagnosis?

A. It removes all effects of rotational speed

B. It reveals repetitive impact modulation that may be hidden in ordinary vibration

C. It directly measures lubricant viscosity

D. It prevents sensor mounting errors

Correct Answer: B. It reveals repetitive impact modulation that may be hidden in ordinary vibration

Question 3

Which example best represents feature-level sensor fusion?

A. Combining raw waveforms before any processing

B. Combining bearing-envelope energy, debris concentration, and temperature slope

C. Allowing an operator to inspect only one sensor

D. Replacing all sensors with a single camera

Correct Answer: B. Combining bearing-envelope energy, debris concentration, and temperature slope

Question 4

What is a major danger of randomly splitting adjacent time-series records into training and test sets?

A. The model cannot calculate vibration frequency

B. The test set may contain information nearly identical to the training set

C. The lubricant sample becomes contaminated

D. The machine automatically changes its load

Correct Answer: B. The test set may contain information nearly identical to the training set

Question 5

Which statement most accurately distinguishes a digital twin from a simple dashboard?

A. A digital twin displays more colours and graphics

B. A digital twin connects asset data with models, state estimation, and the physical asset

C. A dashboard always provides more accurate predictions

D. A digital twin requires no sensor data

Correct Answer: B. A digital twin connects asset data with models, state estimation, and the physical asset

Question 6

When failure labels are scarce, which approach is especially suitable for detecting departure from healthy behaviour?

A. Unsupervised novelty detection

B. Random guessing among known fault labels

C. Deleting all normal-operation data

D. Using only a fixed calendar interval

Correct Answer: A. Unsupervised novelty detection

Question 7

Why should a prognostic system report uncertainty in remaining-useful-life estimates?

A. Because uncertainty proves that the model is useless

B. Because future loads, measurements, degradation processes, and model assumptions are imperfectly known

C. Because maintenance decisions never require evidence

D. Because a single date is always prohibited

Correct Answer: B. Because future loads, measurements, degradation processes, and model assumptions are imperfectly known

Question 8

Which cybersecurity measure directly limits the privileges available to a user or software process?

A. Least-privilege access

B. Increasing sensor sampling frequency

C. Using a larger bearing

D. Changing the lubricant viscosity

Correct Answer: A. Least-privilege access

Question 9

In the illustrative gearbox case, why was high torque included in the analysis?

A. Operating context can change the meaning of sensor features

B. Torque has no relationship to gearbox condition

C. It replaces the need for oil analysis

D. It guarantees that every alarm is a failure

Correct Answer: A. Operating context can change the meaning of sensor features

Question 10

What is the most appropriate role of human experts in an AI-supported safety-relevant maintenance system?

A. Approve every prediction without examining evidence

B. Define failure mechanisms, review uncertain cases, and retain accountable decision authority

C. Remove all physical inspections after deployment

D. Override alarms without recording the reason

Correct Answer: B. Define failure mechanisms, review uncertain cases, and retain accountable decision authority