13. Condition Monitoring, Diagnostics, and Prognostics

Introduction and Context

Mechanical systems rarely fail without producing detectable changes. A rolling-element bearing may develop impulsive contact forces before visible damage appears; a gearbox may generate sidebands around a gear-mesh frequency; a rotor may exhibit a changing synchronous amplitude as imbalance, misalignment, or looseness develops. Condition monitoring converts these physical changes into measured evidence, diagnostics interprets the evidence to identify the fault, and prognostics estimates how the condition may evolve and how long useful operation may remain.

These activities are central to modern mechanical-vibration engineering because maintenance decisions increasingly depend on evidence rather than fixed schedules. A purely reactive strategy accepts unexpected failure, while a time-based strategy may replace components that still possess substantial life. Condition-based and predictive maintenance seek a better balance among safety, availability, maintenance cost, energy use, and production requirements.

The task is not simply to compute a spectrum. It is a chain of reasoning: selecting sensors, acquiring trustworthy data, extracting features, recognizing abnormality, identifying the physical mechanism, estimating severity and progression, and making a defensible operational decision. The chain also includes uncertainty. A high vibration level is not automatically evidence of a dangerous fault, and a low level does not guarantee that a machine is healthy. Sensor location, operating condition, structural transmission paths, speed variation, signal quality, and the selected model all influence the conclusion.

Learning Objectives

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

1. Explain the distinction between condition monitoring, fault diagnosis, and prognostics.

2. Select and interpret time-domain, frequency-domain, time–frequency, envelope, cepstral, and order-domain features.

3. Relate vibration signatures to common rotating-machine faults.

4. Design a practical monitoring and diagnostic workflow.

5. Distinguish anomaly detection, fault classification, fault-severity estimation, and remaining-useful-life prediction.

6. Evaluate statistical-learning and physics-informed approaches for machinery health assessment.

7. Recognize uncertainty, data-quality problems, ethical issues, and common diagnostic errors.

8. Apply the concepts to maintenance planning through an industrial case.

From Measurement to Maintenance Decision

Condition monitoring is the systematic observation of measurable variables to assess equipment health. Vibration is often the principal variable, but temperature, lubricant debris, acoustic emission, motor current, pressure, speed, and process variables may provide complementary evidence. Diagnostics is the process of determining the type, location, cause, and severity of a fault. Prognostics predicts future degradation, usually through a health indicator, a degradation trajectory, or a distribution for remaining useful life (RUL).

A useful conceptual sequence is:

physical state → sensor response → processed signal → feature or health indicator → diagnosis → prognosis → maintenance decision.

The physical state includes clearance, stiffness, damping, surface damage, alignment, load, and lubrication. The sensor response is affected by the measurement chain and by structural transfer paths. Signal processing attempts to separate useful information from noise and irrelevant variation. A feature is a numerical representation of the signal, whereas a health indicator is a feature or model output designed to track condition over time. The final decision may be to continue operation, increase monitoring, reduce load, schedule inspection, or shut down immediately.

This sequence prevents a common mistake: treating an algorithmic output as the fault itself. A classifier identifies patterns associated with training data; it does not eliminate the need for physical interpretation and engineering judgment.

Measurement Architecture and Data Quality

Accelerometers are widely used because they capture dynamic forces over a broad frequency range. Velocity is often useful for assessing general machine vibration in a frequency band, while displacement is valuable for low-frequency shaft motion and relative rotor movement. Piezoelectric accelerometers are common for high-frequency measurements; industrial MEMS sensors can support distributed and wireless monitoring but may have different bandwidth, noise, and calibration characteristics. Tachometers or once-per-revolution pulses are essential when order analysis or phase-based diagnosis is required.

Sensor placement should be mechanically meaningful. Measurements are usually taken close to bearings, in horizontal, vertical, and axial directions where relevant. A poor mounting surface, magnetic base used at high frequency, loose stud, cable motion, or incorrect orientation can create misleading features. Sampling frequency must exceed twice the highest frequency of interest, with practical margin for anti-alias filtering. Record length determines frequency resolution: for a sampling frequency \(f_s\) and record length \(N\), the nominal spectral spacing is \(\Delta f=f_s/N\). A narrow spectral line cannot be resolved with an excessively short record.

The monitoring plan should also record operating context: rotational speed, load, temperature, product throughput, valve position, and recent maintenance. Data collected at different operating states should not be compared blindly. A speed-dependent vibration increase may be normal during a resonance crossing, whereas the same increase at constant speed may indicate deterioration.

Signal Representations and Their Interpretation

Time-domain features

The time waveform retains the temporal sequence of vibration. Common measures include mean, root-mean-square (RMS), variance, peak value, peak-to-peak value, crest factor, kurtosis, and skewness. For a discrete signal \(x_i\),

\[

x_{RMS}=\sqrt{\frac{1}{N}\sum_{i=1}^{N}x_i^2}, \qquad

CF=\frac{|x|_{peak}}{x_{RMS}}.

\]

RMS is related to overall energy and is useful for trending stable operating conditions. Crest factor can increase when isolated impacts appear, although it may fall again when severe damage produces nearly continuous roughness. Kurtosis, the normalized fourth central moment, is sensitive to impulsive events but also to isolated disturbances and nonstationary operation. No single statistic should therefore be interpreted independently.

Time-domain analysis is particularly useful for transient events, shock, looseness, impacts, and changes in overall severity. Its weakness is limited fault specificity: different faults can produce similar increases in RMS.

Frequency-domain analysis

The Fourier transform represents the distribution of signal energy with frequency. A power spectral density (PSD) estimates mean-square content per unit frequency and is useful for comparing vibration levels across bands. Spectral lines often correspond to shaft rotational frequency, harmonics, bearing defect frequencies, gear-mesh frequency, electrical excitation, or structural natural frequencies.

For a shaft rotating at frequency \(f_r\), imbalance commonly produces a strong component at 1X, while angular or parallel misalignment may produce significant 2X and axial vibration. Mechanical looseness can generate harmonics of running speed and sometimes subharmonics. Gear-mesh frequency is approximately

\[

f_{GM}=N_p f_r,

\]

where \(N_p\) is the number of teeth on the pinion. Modulation by eccentricity, backlash, or tooth damage creates sidebands spaced at a shaft or rotational frequency.

Frequency interpretation requires caution. A resonance can amplify a small excitation, and a structural mode can make a fault appear at a frequency unrelated to its origin. Spectral peaks should be checked against machine geometry, speed, phase, direction, and historical behavior.

Envelope analysis

Localized bearing defects often generate short impacts that excite a structural resonance at a higher frequency. The impacts may be difficult to identify in the raw waveform because of broadband noise. Envelope analysis commonly band-pass filters a high-frequency resonant band, demodulates the signal using the Hilbert transform, and computes the spectrum of the envelope. Characteristic frequencies include ball-pass frequency outer race (BPFO), ball-pass frequency inner race (BPFI), ball-spin frequency (BSF), and fundamental train frequency (FTF). Their approximate values depend on rolling-element diameter, pitch diameter, contact angle, number of rolling elements, and shaft speed.

Envelope spectra are powerful for early bearing detection, but they are sensitive to band selection, sensor location, speed variation, and transmission path. A false indication can result when another impact source excites the selected band.

Cepstral analysis

The cepstrum is obtained by taking the inverse Fourier transform of the logarithm of the spectrum, often using the real or power cepstrum. It transforms periodic spacing in the spectrum into a peak at a corresponding quefrency. This is useful when sidebands or harmonics are numerous and difficult to distinguish individually. Gear faults, for example, may produce a family of sidebands around gear-mesh frequency; periodic sideband spacing can become prominent in the cepstrum.

Cepstral analysis is not a replacement for spectral analysis. It is a complementary representation that emphasizes periodicity in the spectral structure. Windowing, logarithmic scaling, spectral noise, and speed variation affect interpretation.

Time–frequency analysis

A stationary spectrum can conceal when an event occurs. The short-time Fourier transform (STFT) divides a signal into overlapping windows and generates a spectrogram. Short windows provide better time localization but poorer frequency resolution; long windows do the opposite. Wavelet transforms use scale-dependent windows and are effective for transients and multi-scale impulsive behavior. Empirical mode decomposition and related adaptive techniques can isolate intrinsic oscillatory components, but mode mixing and end effects require attention.

Time–frequency methods are valuable during run-up, coast-down, variable-load operation, intermittent impacts, and rapidly evolving faults. Their outputs should be interpreted with knowledge of windowing, sampling, and the machine’s speed profile.

Order-domain analysis

When speed varies, fixed-frequency spectra smear components that are tied to shaft rotation. Order analysis normalizes frequency by rotational speed: 1 order corresponds to one cycle per revolution, 2 orders to two cycles, and so forth. Resampling using a tachometer or estimating instantaneous speed allows components to be tracked in the order domain. Run-up order maps can reveal resonant amplification, critical speeds, and speed-related fault components.

Order analysis is especially useful for engines, compressors, turbines, vehicle drivetrains, and variable-speed electric machines. Errors in tachometer pulses, torsional speed fluctuation, and inadequate angular resampling can create artificial orders.

Fault Diagnosis as an Engineering Inference Problem

Diagnosis combines pattern recognition with a causal model of the machine. A practical diagnostic framework has four stages: establish a baseline, detect deviation, identify the likely mechanism, and confirm through independent evidence.

Baseline data should represent healthy operation across expected speeds and loads. A fixed threshold is often inadequate because vibration distributions vary with operating state. Statistical limits may use a baseline mean and standard deviation, robust quantiles, or a model conditioned on speed and load. Thresholds should distinguish alert, alarm, and critical states, but these categories must be linked to machine-specific consequences rather than adopted mechanically.

Fault identification benefits from feature fusion. For example, a rising 1X radial component with stable phase may suggest imbalance; a strong axial component and 2X behavior may support misalignment; envelope peaks at bearing characteristic frequencies may indicate localized bearing damage; gear-mesh sidebands and a time-synchronous average may support gear deterioration. Temperature, lubrication analysis, inspection records, and process symptoms can confirm or challenge a vibration-based conclusion.

A diagnostic matrix is often useful. It maps observations to hypotheses, confidence, and recommended confirmation tests. Such a matrix should explicitly account for confounding faults. A loosened foundation can create harmonics and modulation resembling gear or bearing problems, while a soft foot can imitate misalignment.

Statistical Learning and Health Indicators

Statistical learning methods are useful when large, representative datasets are available. Unsupervised methods such as principal component analysis, clustering, autoencoders, and density estimation learn normal behavior and flag deviations. Supervised methods—including logistic regression, support-vector machines, random forests, gradient boosting, and neural networks—learn mappings from features to known fault classes. Regression models estimate severity or another continuous quantity.

The distinction between anomaly detection and diagnosis is important. An anomaly detector answers, “Does this observation differ from learned normal behavior?” A classifier answers, “Which known class is most consistent with it?” A system may detect a previously unseen fault but be unable to name it reliably.

Feature engineering remains valuable even when deep learning is used. Candidate features may include band energy, spectral kurtosis, envelope amplitudes, sideband ratios, order amplitudes, wavelet energy, temperature gradients, and trend slopes. Feature selection should consider redundancy, robustness, interpretability, and leakage. Randomly splitting sequential data can place nearly identical neighboring samples in both training and test sets, producing overly optimistic performance.

Evaluation should use machine-level or time-based separation, not merely sample-level accuracy. Precision, recall, false-alarm rate, detection delay, calibration, and cost-weighted consequences are often more meaningful than accuracy alone. A missed critical fault and an unnecessary inspection do not have equal cost.

Prognostics and Remaining Useful Life

Prognostics estimates future condition and RUL. RUL may be defined as the time until a health indicator reaches a failure threshold, performance falls below a requirement, or a safety limit is exceeded. These definitions must be stated explicitly because “failure” can mean functional failure, unacceptable vibration, loss of efficiency, or planned replacement.

A simple degradation model is

\[

h(t)=h_0+at+\varepsilon(t),

\]

where \(h(t)\) is a health indicator, \(a\) is a degradation rate, and \(\varepsilon(t)\) represents noise. More realistic models allow random effects between machines, nonlinear trajectories, regime changes, and covariate dependence on load and temperature. State-space models, Kalman or particle filters, proportional-hazard models, Gaussian processes, recurrent neural networks, and survival models are possible approaches.

An RUL estimate should be a distribution or interval when uncertainty matters, not merely a single date. Sources of uncertainty include sensor noise, unknown operating future, model inadequacy, variation among components, uncertain failure threshold, and sparse failure data. Prediction intervals should be checked for calibration: an interval advertised as 90% should contain the true outcome approximately 90% of the time under comparable conditions.

Physics-informed AI combines conservation laws, known kinematics, fatigue or contact models, and machine-learning flexibility. Examples include constraining a bearing model with defect-frequency relationships, embedding rotor dynamics in a surrogate model, or using a neural network to estimate unknown parameters while preserving a governing-equation residual. Such methods can improve extrapolation and reduce data requirements, but incorrect physics can bias the model. “Physics-informed” is not automatically synonymous with reliable.

An Illustrative Industrial Case: Gearbox Condition Assessment

Background

Consider an illustrative case involving a two-stage helical gearbox driving a process conveyor. The gearbox operates between 900 and 1,500 rpm, with load varying during production. It is equipped with triaxial accelerometers near the input and output bearings, a once-per-revolution tachometer, and oil-temperature monitoring.

Situation and problem

Over six weeks, the maintenance team observes intermittent increases in output-side vibration. Overall RMS rises only modestly, but production interruptions would be costly if a gear tooth or bearing failed unexpectedly. The team must determine whether the change is operational, sensor-related, or an early mechanical fault.

Analysis

First, the data are separated by speed and load. A comparison of unconditioned RMS values initially suggests a fault, but much of the variation corresponds to load changes. Order tracking then shows a growing component near the second-stage gear-mesh order and sidebands at the output-shaft order. Time-synchronous averaging strengthens the gear-mesh pattern. The envelope spectrum also shows a weak bearing-related component, but it is not stable across sensor positions.

Inspection of the sensor mounting reveals that the output-side accelerometer has been moved to a less rigid surface during a previous inspection. After remounting, the gear-mesh sidebands remain, whereas the apparent bearing feature largely disappears. Oil analysis shows no decisive increase in metallic debris. Phase measurements indicate that the dominant change is localized near the second-stage mesh rather than being a general imbalance condition.

Strategy and solution

The team constructs a health indicator from order-normalized sideband energy divided by total gear-mesh energy. A baseline model conditions the indicator on load and temperature. A change-point detector identifies a persistent increase rather than isolated peaks. A visual inspection is scheduled during the next planned outage, and the gearbox is operated at reduced peak load with increased monitoring until then.

Inspection finds localized wear and early pitting on one gear flank, without tooth breakage. The gear pair and lubricant are replaced during the planned outage. After replacement, the sideband indicator returns near its baseline distribution.

Outcome and learning points

The case demonstrates that diagnosis required more than an overall vibration threshold. Order analysis handled variable speed, sideband interpretation connected the pattern to gear modulation, sensor verification prevented an incorrect bearing diagnosis, and a load-conditioned health indicator supported a proportionate maintenance decision. The prognostic conclusion was deliberately conservative: the data justified scheduling and risk reduction, but not a precise claim of remaining life because no validated degradation-to-failure model existed for this gearbox.

Implementation, Governance, and Professional Practice

A monitoring program should define asset criticality, failure modes, measurable symptoms, sensor requirements, sampling plans, data ownership, alarm logic, and response actions. The best algorithm cannot compensate for missing maintenance records, inconsistent sensor placement, or unclear responsibility for acting on an alarm.

Common mistakes include using a universal threshold for every machine, ignoring speed and load, confusing correlation with causation, training on data from only one machine, reporting excessive decimal precision, and treating a black-box prediction as a maintenance order. Another error is “alarm fatigue”: too many false alerts cause operators to disregard genuine warnings. Models should therefore be evaluated in the context of workflow and consequences.

Cybersecurity and governance become important when sensors and analytics are connected to industrial networks. Access controls, versioned models, audit trails, secure updates, and retention policies help protect operational data and decision integrity. Automated shutdown recommendations require fail-safe design and clear authority. A model should not suppress a safety interlock merely because its predicted risk is low.

Ethically, engineers must communicate uncertainty and avoid presenting a probabilistic forecast as certainty. If an algorithm is trained mainly on one operating regime, its limitations must be disclosed before deployment elsewhere. Decisions affecting worker safety should remain reviewable, traceable, and supported by independent engineering evidence.

Current Developments and Best Practices

Contemporary systems increasingly combine edge computing, wireless sensing, cloud analytics, digital twins, and multimodal data. Edge processing reduces bandwidth and latency by transmitting features or events rather than raw signals continuously. Digital twins can connect measured behavior with a simulation or reduced-order physical model, although their usefulness depends on calibration and update quality. Self-supervised learning is attractive where labeled failure data are scarce: models learn regularities from large quantities of normal or weakly labeled data before being adapted to diagnostic tasks.

Useful practices include:

• Begin with failure modes and consequences, not with an available algorithm.
• Establish repeatable sensor locations and document mounting conditions.
• Record speed, load, temperature, and maintenance actions with every measurement.
• Use complementary representations rather than relying on one feature.
• Validate alarms on time-separated and machine-separated data.
• Use domain knowledge to constrain feature interpretation and model behavior.
• Report uncertainty, detection delay, false alarms, and missed detections.
• Reassess thresholds after maintenance, redesign, or operating-regime change.
• Link each alarm to a defined inspection or operating response.
• Preserve raw evidence for important events so that diagnoses can be audited.

Chapter Summary

Condition monitoring observes machine behavior; diagnostics identifies the likely fault and its severity; prognostics estimates future degradation and remaining useful life. Reliable practice begins with appropriate sensors, adequate sampling, consistent mounting, and operating-context data. Time-domain features describe impacts and overall energy, frequency spectra reveal characteristic components, envelope analysis demodulates bearing-related impacts, cepstra expose periodic spectral spacing, time–frequency methods describe transient behavior, and order analysis handles variable speed.

Diagnosis is strongest when signal features are interpreted through machine physics and confirmed with independent evidence. Statistical learning supports anomaly detection, classification, and health-index construction, but data leakage, regime variation, class imbalance, and poor calibration can undermine performance. Prognostic models must state the failure definition and represent uncertainty. Physics-informed AI can improve interpretability and extrapolation when its physical assumptions are valid. Ultimately, condition-monitoring technology is valuable only when it produces trustworthy, actionable, and governable maintenance decisions.

Key Terms

Condition monitoring: Continuous or periodic observation of measurable variables to assess equipment health.

Diagnostics: Determining the type, location, cause, and severity of a fault.

Prognostics: Predicting future degradation or remaining useful life.

Health indicator: A feature or model output intended to track condition over time.

RMS: Square root of the mean squared signal value, representing an energy-related magnitude.

Crest factor: Ratio of peak amplitude to RMS amplitude.

Kurtosis: A fourth-moment statistic sensitive to impulsive or heavy-tailed behavior.

Power spectral density: Distribution of signal mean-square content with frequency.

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

Cepstrum: Transform of the logarithm of a spectrum used to detect periodic spectral structure.

Spectrogram: Time-varying representation of spectral content, commonly based on the STFT.

Order analysis: Representation of vibration components relative to shaft rotational frequency.

Sideband: Spectral component offset from a carrier frequency by a modulation frequency.

Anomaly detection: Identification of observations that differ from learned normal behavior.

Remaining useful life: Predicted time or usage before a defined failure or performance limit.

Physics-informed AI: Learning method incorporating governing physical knowledge or constraints.

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

Prediction interval: Range intended to contain a future outcome with a stated probability.

Review Questions

Short Answer Questions

1. Distinguish condition monitoring, diagnostics, and prognostics.

2. Why must speed and load be recorded with vibration measurements?

3. What does a rising crest factor generally suggest, and why is it not conclusive?

4. State the purpose of envelope analysis in bearing monitoring.

5. Define order analysis and identify one application for it.

Long Answer Questions

1. Explain how time-domain and frequency-domain analysis complement one another in rotating-machine diagnosis.

2. Describe a complete workflow from sensor installation to maintenance decision.

3. Explain the causes of uncertainty in remaining-useful-life prediction and methods for representing it.

4. Discuss the value and limitations of statistical learning for machinery fault diagnosis.

5. Explain how physics-informed AI can be used in vibration-based prognostics.

Analytical / Application Questions

1. A variable-speed gearbox shows smeared gear-mesh peaks in fixed-frequency spectra. Propose an analysis procedure and explain why each step is required.

2. A bearing-monitoring system reports a high envelope amplitude after an accelerometer is remounted. How would you determine whether the indication is a real fault or a measurement artifact?

3. A classifier achieves 98% accuracy using randomly divided data from sequential measurements but performs poorly on a new machine. Analyze the likely causes and propose a more credible validation strategy.

22. Multiple Choice Questions with Answers

Question 1

Which representation is most appropriate for tracking a fault component tied to shaft rotation during a run-up?

A. Fixed-frequency RMS only

B. Order-domain analysis

C. Unfiltered temperature trend

D. Static time waveform mean

Correct Answer: B. Order-domain analysis

Question 2

Why is envelope analysis commonly used for localized rolling-element bearing defects?

A. It removes all effects of rotational speed

B. It reveals low-frequency modulation of high-frequency impact resonances

C. It directly measures lubricant viscosity

D. It replaces the need for sensor calibration

Correct Answer: B. It reveals low-frequency modulation of high-frequency impact resonances

Question 3

A gearbox spectrum contains gear-mesh peaks with sidebands spaced at the output-shaft frequency. This pattern most directly suggests:

A. Periodic modulation associated with the gear train

B. A purely thermal failure

C. A sensor with zero bandwidth

D. A constant DC offset only

Correct Answer: A. Periodic modulation associated with the gear train

Question 4

Which statement best distinguishes anomaly detection from fault classification?

A. Anomaly detection identifies deviation from normal behavior, whereas classification assigns a known fault class

B. Anomaly detection always predicts RUL, whereas classification never uses labels

C. Classification requires no training data

D. They are identical terms with no operational difference

Correct Answer: A. Anomaly detection identifies deviation from normal behavior, whereas classification assigns a known fault class

Question 5

For a sampled record of fixed sampling frequency, increasing the record length generally:

A. Increases nominal frequency resolution

B. Eliminates aliasing without filtering

C. Removes all structural resonances

D. Makes speed measurement unnecessary

Correct Answer: A. Increases nominal frequency resolution

Question 6

Which practice would most likely produce overly optimistic machine-learning performance in condition monitoring?

A. Testing on data from a later operating period

B. Separating training and testing by machine

C. Randomly splitting highly correlated neighboring time samples

D. Reporting false-alarm rate

Correct Answer: C. Randomly splitting highly correlated neighboring time samples

Question 7

A rising RMS value occurs only at higher loads and disappears after load normalization. The most defensible interpretation is:

A. A fault is proven regardless of operating condition

B. The variation may be operating-condition dependent and requires further evidence

C. The sensor must be disconnected

D. The machine has certainly reached its RUL limit

Correct Answer: B. The variation may be operating-condition dependent and requires further evidence

Question 8

Why should an RUL estimate often be reported as a distribution or interval rather than a single date?

A. Because uncertainty arises from noise, future operation, model error, and component variation

B. Because single dates cannot be calculated mathematically

C. Because all machines fail at random times with no physical causes

D. Because intervals eliminate the need for maintenance decisions

Correct Answer: A. Because uncertainty arises from noise, future operation, model error, and component variation

Question 9

In the illustrative gearbox case, remounting the accelerometer was important because it:

A. Changed the gear ratio

B. Helped distinguish a measurement artifact from a persistent gear-related signature

C. Guaranteed that no bearing fault existed

D. Eliminated the need for order tracking

Correct Answer: B. Helped distinguish a measurement artifact from a persistent gear-related signature

Question 10

What is a defining feature of physics-informed AI in vibration prognostics?

A. It ignores governing physics in favor of larger datasets

B. It incorporates physical relationships or constraints into the learning process

C. It can never be used with sensor data

D. It guarantees perfect extrapolation beyond training conditions

Correct Answer: B. It incorporates physical relationships or constraints into the learning process