13. Reliability, Diagnostics and Intelligent Asset Management

Introduction and context

An electrical drive is not merely a motor connected to a converter and a mechanical load. It is an engineered asset whose performance depends on the interaction of electrical, electronic, thermal, mechanical, control, environmental and organizational factors. A drive may continue to rotate while its insulation is deteriorating, its bearings are damaged, its cooling system is obstructed, or its converter is accumulating switching stress. Reliability engineering and condition-based maintenance provide the methods needed to detect such degradation before it develops into an unsafe or costly failure.

The importance of this subject has increased with the adoption of variable-speed drives, high-power semiconductor converters, permanent-magnet machines, digitally controlled systems and networked industrial plants. These technologies improve efficiency and controllability, but they also introduce new failure mechanisms. Fast voltage transients may accelerate insulation ageing; common-mode currents may damage bearings; power semiconductor thermal cycling may produce solder fatigue; and poorly tuned control systems may impose excessive mechanical torque oscillations.

Intelligent asset management extends conventional maintenance by combining failure analysis, sensing, signal processing, statistical inference, prognostics, maintenance planning and digital technologies. Its purpose is not simply to collect data. The purpose is to support a defensible decision: whether to continue operation, inspect, derate, repair, replace or redesign an asset.

Learning objectives

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

1. Distinguish reliability, availability, maintainability, diagnostics and prognostics.

2. Identify common failure modes in motors, converters and mechanical loads.

3. Apply failure-mode and effects analysis and fault-tree reasoning to electrical drives.

4. Select appropriate condition-monitoring variables and interpret diagnostic signals.

5. Explain the stages of remaining-useful-life estimation and predictive maintenance.

6. Describe the architecture and limitations of a drive-system digital twin.

7. Evaluate machine-learning applications with attention to uncertainty, bias, cybersecurity and human responsibility.

8. Formulate an asset-management strategy linking technical evidence with operational decisions.

Reliability as a system property

Reliability is the probability that an item performs its required function for a specified period under stated conditions. The definition is deliberately conditional. A motor that is reliable at rated load and clean ambient temperature may be unreliable when operated with frequent overloads, blocked ventilation and repeated starts. Reliability therefore concerns both the equipment and its mission profile.

For a component with a constant failure rate, the reliability function may be represented by

\[R(t)=e^{-\lambda t}\]

where \(\lambda\) is the failure rate. Real drive components rarely have a constant failure rate throughout life. A bathtub curve is often used to describe early failures, a useful-life region with approximately random failures, and a wear-out region in which ageing mechanisms dominate. Electronic converters may show early manufacturing or installation defects, while bearings and insulation commonly exhibit wear-out behaviour.

Availability includes both reliability and restoration. A useful steady-state approximation is

\[A=\frac{MTBF}{MTBF+MTTR}\]

where MTBF is mean time between failures and MTTR is mean time to repair or restore. Increasing availability may involve improving reliability, reducing repair time, stocking critical spares, improving access, or introducing redundancy. Maintainability refers to the ease and speed with which an asset can be restored under specified conditions.

A drive should be analysed at several levels: component, subsystem and system. A gate-driver circuit, for example, may fail because of a component defect; the converter may then lose one switching leg; the motor may experience negative-sequence currents; and the production line may stop. System-level consequences cannot be inferred from component failure rates alone.

Failure modes in electric-drive systems

Failure-mode analysis begins by asking how an item can fail to perform its intended function. The answer should be specific. “Motor failure” is too broad; “stator insulation breakdown due to thermal ageing and repetitive voltage stress” is diagnostically useful.

Motors and insulation systems

Stator windings may suffer turn-to-turn faults, phase-to-phase faults, phase-to-ground faults, loose connections and partial discharge. Thermal ageing is strongly influenced by hot-spot temperature, while voltage rise time, repetition frequency and winding geometry affect electrical stress. Contamination and moisture reduce insulation resistance and may create surface tracking.

Rotor-related faults include broken rotor bars in induction motors, cracked end rings, eccentricity and demagnetization or mechanical damage in permanent-magnet machines. Rotor asymmetry often produces characteristic sidebands in current or vibration spectra, although the interpretation depends on slip, speed and load.

Bearings fail through fatigue, lubrication starvation, contamination, misalignment, incorrect mounting, electrical fluting and excessive loading. Inverter-fed motors may experience shaft voltages and bearing currents caused by common-mode switching voltage. A bearing-current problem cannot be solved reliably by replacing bearings alone; grounding, cable arrangement, filtering and insulation may also require attention.

Power converters and drives

Converter failure modes include semiconductor short circuit or open circuit, gate-driver malfunction, DC-link capacitor degradation, heat-sink or fan failure, busbar connection loosening, control-board failure and insulation breakdown. Electrolytic capacitors are sensitive to temperature, ripple current and drying of the electrolyte. Film capacitors may experience metallization loss and capacitance reduction.

Thermal cycling is a major stress on power modules. Differences in thermal expansion among semiconductor chips, solder layers, substrates and baseplates can create bond-wire lift-off, solder fatigue and increased thermal resistance. A converter may therefore fail after many cycles even when its instantaneous current remains within the nominal rating.

Mechanical loads and transmission systems

The load may be the source rather than the victim of abnormal drive behaviour. Pumps can suffer cavitation, blockage and seal failure; fans can experience imbalance or fouling; compressors may impose pulsating torque; conveyors may develop belt slip or misalignment; and gearboxes may develop tooth wear, pitting or lubrication failure. A control system that compensates for a deteriorating load can conceal the problem until a more severe failure occurs.

From failure modes to maintenance decisions

Failure Mode and Effects Analysis (FMEA) is a structured bottom-up method. For each function, the analyst identifies failure mode, local effect, system effect, causes, existing controls and recommended actions. Traditional risk priority numbers combine severity, occurrence and detectability, but such multiplication can disguise important differences. A low-occurrence failure with catastrophic consequences may deserve priority even if its numerical product is modest.

Fault Tree Analysis (FTA) works from an undesired top event, such as “drive unavailable,” and decomposes it using logical AND and OR relationships. For example, converter shutdown may result from overtemperature OR DC-link overvoltage OR protection-system malfunction. FTA is especially useful when the analyst needs to identify combinations of events or estimate system-level probability.

Reliability-centred maintenance (RCM) asks what functions the asset must perform, how it can fail, what consequences follow and which task is technically and economically justified. Typical tasks are:

• condition-directed tasks, performed when evidence shows deterioration;
• scheduled restoration or replacement, performed at a defined interval; and
• failure-finding tasks, used to reveal hidden failures in protective or standby systems.

If no scheduled or condition-based task is effective, redesign, redundancy, run-to-failure or operational change may be appropriate. Maintenance is therefore a risk-management decision, not a ritual based solely on calendar intervals.

Condition monitoring: observing degradation without dismantling

Condition monitoring converts physical behaviour into evidence. The best monitoring plan begins with a failure mechanism and asks which variable changes measurably before functional failure. Installing many sensors without this reasoning produces data but not necessarily knowledge.

Electrical measurements

Phase currents and voltages are available in most modern drives. Motor-current signature analysis can reveal rotor asymmetry, eccentricity, load oscillation and some power-quality problems. Voltage and current waveforms also support calculation of power, torque-related quantities, harmonic content and converter switching behaviour. DC-link voltage ripple can indicate capacitor ageing or rectifier abnormalities.

Insulation resistance, polarization index, surge comparison, partial-discharge measurements and leakage-current monitoring are useful for winding systems, although each test has a particular purpose and operating limitation. An offline insulation test cannot by itself represent all stresses occurring during inverter operation.

Thermal and environmental measurements

Winding resistance, embedded resistance-temperature detectors, thermistors, infrared thermography and thermal models provide information about temperature. Temperature should be interpreted with load, ambient conditions, cooling airflow and duty cycle. A high surface temperature may result from overload, blocked ventilation, a failed fan, poor heat transfer or a faulty sensor.

Humidity, dust, corrosive gases and water ingress are often essential contextual variables. Environmental data can transform an apparently random failure into an understandable pattern.

Mechanical measurements

Accelerometers measure vibration caused by imbalance, misalignment, looseness, bearing damage, gear-mesh problems and resonance. Time waveform, overall amplitude, crest factor, envelope spectrum and order analysis provide different perspectives. Order tracking is valuable in variable-speed machinery because fixed-frequency spectral interpretation becomes misleading when rotational speed changes.

Shaft torque, speed fluctuation, acoustic emission and oil debris analysis can supplement vibration. No single sensor is universally superior. Sensor bandwidth, mounting quality, sampling rate, aliasing, noise and synchronization must be considered.

Signal analytics and diagnostic reasoning

A diagnostic system commonly follows the chain: sensing, conditioning, sampling, feature extraction, inference and decision support. Anti-aliasing filters are required before sampling. Sampling frequency must be adequate for the highest frequency of interest, and timestamps from electrical and mechanical channels should be aligned when cross-domain analysis is intended.

Time-domain features include mean, root-mean-square value, variance, kurtosis, crest factor and peak-to-peak value. Frequency-domain methods use the Fourier transform to identify periodic components and sidebands. Time-frequency methods, such as wavelets or short-time Fourier transforms, are more appropriate for transients, start-ups and changing-speed operation.

A feature is useful only when it is sensitive to the target fault and sufficiently insensitive to normal operating variation. For example, vibration amplitude may increase with load even in a healthy machine. A normalized feature, a model-based residual or an operating-regime-specific baseline may be more informative.

Model-based diagnosis compares measured behaviour with an expected model. The residual

\[r(t)=y_{measured}(t)-y_{model}(t)\]

may indicate a fault when it exceeds a threshold. Thresholds should account for noise, uncertainty and operating conditions. A fixed threshold is simple but may generate false alarms during legitimate transients. Adaptive thresholds and statistical change detection can improve discrimination, provided that the baseline remains trustworthy.

Diagnostics should distinguish detection, isolation and identification. Detection asks whether abnormal behaviour exists. Isolation asks which subsystem is responsible. Identification estimates the fault type, severity or location. Confusing these stages leads to unjustified certainty.

Prognostics and remaining useful life

Prognostics estimates future degradation and remaining useful life (RUL), commonly defined as the time until a specified failure criterion under an assumed future usage profile. RUL is not a fixed property stored inside the machine. It changes with load, temperature, operating strategy and maintenance action.

Three broad approaches are used. Physics-of-failure models represent mechanisms such as insulation ageing, bearing fatigue or capacitor wear. Data-driven models learn relationships between features and outcomes from historical data. Hybrid models combine physical constraints with statistical learning. The choice depends on available failure data, model knowledge, operating variability and the cost of errors.

A prognostic workflow includes:

1. defining the failure threshold and mission conditions;

2. establishing a healthy baseline;

3. extracting degradation-sensitive indicators;

4. estimating the current health state;

5. forecasting future evolution;

6. quantifying uncertainty; and

7. converting the forecast into a maintenance decision.

Point estimates are inadequate for high-consequence decisions. Prediction intervals, confidence bounds or probability distributions should communicate uncertainty. A model that predicts RUL accurately on average but is overconfident can be more dangerous than a less accurate model that honestly expresses uncertainty.

Intelligent asset management and digital twins

An asset-management system links asset hierarchy, condition data, maintenance history, operating context, spares, personnel and business consequences. It should answer practical questions: Which asset requires attention? What evidence supports the recommendation? What is the consequence of waiting? What intervention is feasible during the next production opportunity?

A digital twin is a digital representation connected to a physical asset through data and models. For a drive, it may include electrical equations, thermal networks, loss models, mechanical-load models, control states, alarms and maintenance history. A useful twin need not reproduce every physical detail. It must represent the states and failure mechanisms relevant to the decision.

Digital twins support commissioning, anomaly detection, what-if analysis, virtual sensing, derating and maintenance scheduling. Their credibility depends on calibration, data quality, version control and explicit assumptions. A visually impressive dashboard is not a digital twin if it merely displays unverified measurements.

Machine learning can classify fault patterns, detect anomalies, estimate health indices and rank maintenance priorities. Supervised learning requires representative labelled data, which are often scarce because serious failures are rare. Unsupervised or semi-supervised methods can learn normal behaviour, but they may confuse a new operating mode with a fault. Domain adaptation is important when a model trained on one motor, site or load is transferred to another.

Responsible use requires explainability appropriate to the decision, protection against data leakage, monitoring of model drift, human review for high-impact actions and preservation of audit trails. A model should not recommend an immediate shutdown solely because its internal score is high if the evidence cannot be examined and the operational consequences are severe. Human expertise remains essential, especially for novel faults and conflicting indicators.

Applied case: predictive maintenance of a variable-speed pumping drive

Background

The following is an illustrative industrial case. A water-treatment plant operates several medium-voltage induction-motor pump drives. Each pump is critical during peak treatment demand, but redundant capacity exists at reduced efficiency. The existing maintenance practice is quarterly inspection and replacement of bearings on a fixed schedule.

Situation and problem

One pump begins producing intermittent high-vibration alarms. The alarm appears only at certain speeds and disappears when the pump is operated at a lower flow. Replacing the bearing does not permanently remove the symptom. Operators also report a small increase in motor current and occasional fluctuations in discharge pressure.

Analysis

A simple bearing-failure diagnosis is rejected because the vibration is speed-dependent and the new bearing shows no immediate damage. Synchronized measurements are collected: three-axis vibration, phase currents, shaft speed, discharge pressure, motor temperature and valve position. Order analysis reveals a strong component near a mechanical rotational order, while pressure data show periodic hydraulic pulsation. Current analysis indicates torque oscillation at the same operating region.

The investigation identifies partial-flow operation near a hydraulic instability region. The pump is experiencing a combination of flow recirculation and structural resonance. The control system maintains pressure by changing speed, repeatedly crossing the resonant region. The bearing is being loaded by the resulting vibration, but it is not the initiating fault.

Approach and solution

The plant creates a regime-based health model rather than using one global vibration threshold. The operating envelope marks the resonant speed band, and the controller is modified to pass rapidly through that band or avoid it where process constraints permit. A pressure sensor is added to improve detection of hydraulic instability, and vibration measurements are retained for confirmation. Maintenance personnel inspect alignment, coupling condition and foundation stiffness during the next planned outage.

The asset-management system records the evidence, intervention and subsequent performance. The bearing replacement interval is changed from a purely calendar-based task to a condition-directed task supported by vibration and operating-history trends.

Outcome

The revised strategy reduces nuisance alarms, prevents repeated bearing replacement without cause and makes the maintenance decision traceable. More importantly, it addresses the interaction among load, control, structure and motor rather than treating the motor as an isolated component.

Learning points

The case demonstrates that diagnosis requires operating context, not merely a high feature value. It also illustrates the difference between a symptom, a proximate damage mechanism and a root cause. A useful intelligent system combines sensors, engineering models, expert investigation and controlled operational action.

Challenges, risks and professional responsibilities

False positives cause unnecessary inspection, lost production and alarm fatigue. False negatives allow degradation to progress. Sensor failure, poor mounting, calibration drift, missing timestamps and changes in process configuration can corrupt a diagnostic system. Maintenance records may contain survivorship bias because failed assets are removed while healthy assets remain in service.

Cybersecurity is an engineering requirement for connected drives. Unauthorized alteration of set points, firmware or alarm thresholds may create physical hazards. Access control, network segmentation, secure updates, logging and recovery procedures should be designed with the drive’s safety function in mind.

Data governance must address ownership, retention, privacy where personnel data are involved, and the right to inspect automated recommendations. Safety-related decisions should comply with applicable electrical, machinery, functional-safety and occupational requirements. An algorithmic recommendation does not transfer professional accountability from the engineer or responsible organization.

Common mistakes include monitoring variables unrelated to credible failure modes, using arbitrary alarm limits, training and testing a model on highly similar records, ignoring changing speed and load, treating correlation as causation, and reporting a precise RUL without uncertainty. Another frequent error is optimizing maintenance cost while ignoring production, safety, environmental and reputational consequences.

Best-practice framework

A defensible programme can be organized as follows:

1. Define asset functions, criticality and unacceptable consequences.

2. Construct a failure-mode register using FMEA, FTA or RCM reasoning.

3. Map each important failure mechanism to observable variables.

4. Establish sensor quality, sampling, synchronization and data-management requirements.

5. Develop healthy baselines across realistic operating regimes.

6. Validate diagnostic features using engineering knowledge and field evidence.

7. Separate alarm, diagnosis, prognosis and maintenance authorization.

8. Report uncertainty and provide evidence that technicians can interpret.

9. Integrate recommendations with work orders, spares and production planning.

10. Review outcomes after each intervention and update the models under controlled governance.

Emerging practice is moving toward edge analytics, federated learning, physics-informed machine learning, virtual sensing and fleet-level comparison. These developments can reduce communication requirements and improve transferability, but they do not remove the need for validation. The central principle remains unchanged: intelligence is valuable only when it improves the reliability, safety and economic performance of a real asset.

Chapter Summary

Reliability and intelligent asset management treat the electrical drive as an interacting system of motor, converter, control, load, environment and maintenance process. FMEA, FTA and RCM connect failure mechanisms with consequences and maintenance tasks. Condition monitoring uses electrical, thermal, mechanical and environmental measurements, while signal analytics converts raw data into diagnostic evidence. Prognostics extends this process by estimating degradation and uncertain remaining useful life. Digital twins and machine learning can support anomaly detection, virtual sensing and maintenance planning, but their results require sound data, physical understanding, cybersecurity and human oversight. Effective practice is therefore a closed decision loop: understand failure, measure relevant behaviour, infer condition, act proportionately and learn from the outcome.

Key Terms

Reliability: Probability that an asset performs its required function for a specified time and condition.

Availability: Proportion of required time during which an asset is capable of performing its function.

Maintainability: Ease and speed with which an asset can be restored after a failure.

Failure mode: Specific manner in which an item fails to perform its intended function.

FMEA: Structured analysis linking failure modes, causes, effects, controls and actions.

Fault Tree Analysis: Top-down logical analysis of combinations leading to an undesired event.

Reliability-centred maintenance: Maintenance method based on functions, failure consequences and technically justified tasks.

Condition monitoring: Observation of asset variables to identify deterioration or abnormality.

Diagnostics: Detection, isolation and identification of faults.

Prognostics: Forecasting future degradation or failure behaviour.

Remaining useful life: Estimated time until a defined failure threshold under stated conditions.

Health indicator: Derived variable representing the condition or degradation state of an asset.

Residual: Difference between measured and model-predicted behaviour.

Order analysis: Spectral analysis referenced to rotational speed rather than fixed frequency.

Digital twin: Connected digital representation of a physical asset or process.

Model drift: Reduction in model validity as data, equipment or operating conditions change.

False positive: Alarm indicating a fault when the asset is healthy.

False negative: Failure to alarm when a relevant fault exists.

Thermal cycling: Repeated temperature variation that produces mechanical and material stress.

Asset criticality: Relative consequence of asset loss considering safety, production, environment and cost.

Review Questions

Short Answer Questions

1. Distinguish reliability from availability in an electrical-drive application.

2. Why can a motor bearing be a victim rather than the root cause of high vibration?

3. What is the purpose of a healthy baseline in condition monitoring?

4. Differentiate fault detection, fault isolation and fault identification.

5. Why should RUL estimates include uncertainty?

Long Answer Questions

1. Explain how FMEA, FTA and RCM can be combined to design a maintenance programme for a converter-fed motor.

2. Discuss electrical, thermal and mechanical condition-monitoring methods and their principal limitations.

3. Describe a complete prognostic workflow from failure definition to maintenance decision.

4. Explain the architecture, uses and validation requirements of a digital twin for an electrical drive.

5. Critically evaluate the use of machine learning for drive diagnostics, including data, explainability, cybersecurity and human oversight.

Analytical / Application Questions

1. A variable-speed fan shows high vibration only during acceleration, while current and temperature remain normal. Propose a measurement and analysis plan that distinguishes resonance, imbalance and sensor error.

2. A converter’s DC-link ripple gradually increases, but production cannot stop for two months. Formulate a risk-based decision using monitoring, derating, inspection and spare-planning options.

3. Design a condition-based maintenance strategy for a pump drive in which load, speed and ambient temperature vary substantially. Specify failure modes, sensors, features, alarm logic and decision rules.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes diagnostics from prognostics?

A. Diagnostics estimates repair cost, while prognostics measures availability

B. Diagnostics identifies present abnormality or fault, while prognostics forecasts future degradation or failure

C. Diagnostics applies only to motors, while prognostics applies only to converters

D. Diagnostics is offline testing, while prognostics is always online testing

Correct Answer: B. Diagnostics identifies present abnormality or fault, while prognostics forecasts future degradation or failure

Question 2

A motor-current feature rises whenever load increases but no fault is present. What is the most appropriate response?

A. Use the feature without modification because higher current always indicates failure

B. Remove all current sensors from the system

C. Normalize or model the feature with respect to operating regime

D. Set the alarm threshold equal to the highest measured current

Correct Answer: C. Normalize or model the feature with respect to operating regime

Question 3

Which failure mechanism is especially associated with repeated temperature excursions in power modules?

A. Solder fatigue and bond-wire degradation

B. Hydraulic cavitation

C. Rotor eccentricity caused by belt slip

D. Partial discharge in an unloaded cable only

Correct Answer: A. Solder fatigue and bond-wire degradation

Question 4

In fault-tree analysis, the top event is normally:

A. The most inexpensive component in the system

B. An undesired system-level event

C. The first sensor measurement collected

D. The maintenance technician’s preferred action

Correct Answer: B. An undesired system-level event

Question 5

Why is order analysis useful for variable-speed machinery?

A. It removes the need for vibration sensors

B. It expresses spectral components relative to rotational speed

C. It guarantees identification of the root cause

D. It measures insulation resistance during operation

Correct Answer: B. It expresses spectral components relative to rotational speed

Question 6

In the illustrative pumping-drive case, the repeated bearing replacement failed to solve the problem because:

A. The bearing was never connected to the shaft

B. The dominant initiating condition was hydraulic instability interacting with resonance

C. The motor had no temperature sensor

D. The pump was operated only at zero speed

Correct Answer: B. The dominant initiating condition was hydraulic instability interacting with resonance

Question 7

What is a principal risk of an overconfident RUL model?

A. It may communicate narrow predictions that conceal serious uncertainty

B. It always produces too many maintenance tasks

C. It cannot use any sensor data

D. It eliminates all false negatives by definition

Correct Answer: A. It may communicate narrow predictions that conceal serious uncertainty

Question 8

Which statement best describes a digital twin in this chapter?

A. A static equipment nameplate stored in a database

B. A connected digital representation using relevant data and models of a physical asset

C. A replacement for all physical inspections

D. A purely visual dashboard with no model or validation

Correct Answer: B. A connected digital representation using relevant data and models of a physical asset

Question 9

A model trained on one motor is deployed on a different motor and performance declines. This is most directly an example of:

A. Model drift or transfer-domain mismatch

B. Improved sensor calibration

C. Failure-mode elimination

D. Increased maintainability

Correct Answer: A. Model drift or transfer-domain mismatch

Question 10

Which practice best supports responsible automated maintenance decisions?

A. Allowing an opaque model to authorize every shutdown without review

B. Suppressing uncertainty to make recommendations appear precise

C. Maintaining audit trails, validating models and retaining human oversight for high-impact actions

D. Training only on successful repairs and ignoring unsuccessful predictions

Correct Answer: C. Maintaining audit trails, validating models and retaining human oversight for high-impact actions