13. Condition Monitoring, Diagnostics, and Digital Twins

Introduction and Context

Mechanical systems rarely fail without warning. Changes in vibration, temperature, lubrication condition, acoustic emission, electrical power, or operating performance often appear before a catastrophic failure. Condition monitoring is the systematic measurement and interpretation of these changes so that maintenance decisions can be based on the actual health of an asset rather than on elapsed time alone. Diagnostics goes further by identifying the likely fault, its location, severity, and progression. Prognostics estimates how the fault may develop and how long the machine can continue to perform its required function. Digital twins extend this logic by maintaining a computational representation of a physical asset and updating it with operational data.

These ideas are central to modern machine dynamics and controls. A rotating machine, for example, is not merely a collection of bearings, shafts, gears, and couplings. Its dynamic response reflects structural stiffness, damping, excitation forces, speed, load, lubrication, alignment, and control actions. Fault detection therefore requires both signal-processing competence and an understanding of machine physics. A spectral peak has meaning only when interpreted in relation to shaft speed, geometry, operating regime, and plausible failure mechanisms.

Condition-based maintenance is particularly valuable where failure has high consequences, spare parts have long lead times, or shutdowns are expensive. However, monitoring is not automatically beneficial. Poor sensor placement, inadequate sampling, unrepresentative training data, false alarms, cybersecurity weaknesses, and incorrect decisions can create cost without improving reliability. The objective is not to collect the largest possible quantity of data; it is to convert relevant data into timely, defensible maintenance and operational decisions.

Learning Objectives

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

1. Explain the relationship between condition monitoring, fault diagnosis, prognostics, predictive maintenance, and digital twins.

2. Select suitable sensors and sampling strategies for mechanical assets.

3. Apply time-domain, frequency-domain, and time–frequency methods to machinery signals.

4. Distinguish detection, diagnosis, prognosis, and remaining-useful-life estimation.

5. Design a practical workflow for data acquisition, feature extraction, model development, and maintenance decision-making.

6. Evaluate model performance using technically appropriate metrics and cost-based criteria.

7. Describe digital-twin architectures, virtual commissioning, and lifecycle optimization.

8. Recognize limitations, ethical responsibilities, cybersecurity risks, and implementation challenges.

From Measurement to Maintenance Decision

Condition monitoring begins with an observable variable. Common measurements include acceleration, velocity, displacement, temperature, pressure, flow, torque, motor current, oil debris, acoustic emission, and operating speed. The selected variable must be physically connected to the failure mechanism. Accelerometers are often effective for rolling-element bearing and gear faults because such faults generate dynamic impacts. Temperature may be more useful for lubrication failure or excessive friction. Motor-current analysis can provide indirect information about mechanical load, eccentricity, or broken rotor bars.

A useful monitoring architecture separates four tasks:

• Detection: deciding whether the current condition differs significantly from healthy operation.
• Diagnosis: determining the fault type, location, and severity.
• Prognosis: estimating future degradation and the time before a defined failure threshold is reached.
• Decision support: recommending inspection, derating, repair, replacement, or continued operation.

A simple alarm limit is not equivalent to diagnosis. For example, an increase in overall vibration may result from imbalance, misalignment, looseness, resonance, bearing damage, or a process change. Diagnosis requires additional evidence, such as spectral components, phase relationships, harmonics, envelope spectra, or comparison among measurement locations.

The maintenance decision should also consider consequence and uncertainty. A model may indicate a 70% probability that a bearing has entered a degraded state, but the appropriate action depends on the cost of an unnecessary shutdown, the consequence of failure, the availability of a spare, and the confidence interval around the predicted failure time. Engineering judgment remains essential even when analytics are advanced.

Sensors, Data Quality, and Experimental Design

Sensor selection should follow the physics of the asset. Accelerometers require suitable frequency range, mounting stiffness, and orientation. A loose magnetic mounting may distort high-frequency measurements. Temperature sensors must be placed where thermal changes are meaningful and must account for heat conduction and ambient variation. Tachometers or once-per-revolution references are important for order analysis when rotational speed varies. Pressure and flow sensors are essential in hydraulic and pneumatic systems, while strain gauges and torque transducers are valuable when load transmission is the primary concern.

Sampling frequency must satisfy the measurement objective. According to the Nyquist requirement, a signal sampled at frequency \(f_s\) cannot represent frequency content above \(f_s/2\) without aliasing. Anti-aliasing filters are therefore required before digitization. In practice, the sampling rate is selected above the highest frequency of interest, with allowance for filter transition bands. The record length determines frequency resolution:

\[

\Delta f = \frac{f_s}{N}

\]

where \(N\) is the number of samples. A long record improves frequency resolution but may blur transient changes if the machine operating condition is not stationary.

Data quality problems include missing samples, sensor drift, saturation, electrical interference, inconsistent speed, and changes in process load. Metadata should accompany every record: asset identity, sensor location, direction, speed, load, environmental conditions, maintenance status, and time stamp. Without such context, a machine-learning model may learn operating regime rather than health condition.

A robust monitoring program establishes a healthy baseline. Baseline data should cover the normal range of speed, load, temperature, and process conditions. Comparing a current signal with a single historical measurement is usually less reliable than comparing it with a distribution of healthy observations under comparable conditions.

Time-Domain Analysis

Time-domain analysis examines signal amplitude as a function of time. For a vibration signal \(x(t)\), common statistics include the mean, root-mean-square value, variance, peak, peak-to-peak value, crest factor, and kurtosis. The root-mean-square value is

\[

x_{rms}=\sqrt{\frac{1}{N}\sum_{i=1}^{N}x_i^2}

\]

and represents overall energy or severity for many vibration applications. Crest factor is the ratio of peak amplitude to RMS amplitude. It may increase when isolated impacts occur, as in an early bearing defect, although it can decrease again when damage becomes widespread and the signal becomes continuously energetic. Kurtosis measures the peakedness of a distribution and can support impact detection, but it is sensitive to outliers and operating conditions.

Time-domain measures are computationally efficient and suitable for continuous screening. Their weakness is limited fault specificity. Two different faults may produce similar RMS values, and a change in speed or load may alter the statistic without any damage. Time-domain indicators are therefore most effective when normalized by operating condition or combined with spectral and process features.

Trend analysis is important. A slowly increasing RMS value may be more informative than a single high measurement. However, trends should be interpreted with confidence bands and maintenance history. A sudden reduction after lubrication does not necessarily prove that a bearing was healthy; it may indicate temporary suppression of symptoms.

Frequency-Domain Analysis

The Fourier transform represents a time signal as a combination of sinusoidal components. The discrete Fourier transform is commonly evaluated using the fast Fourier transform (FFT). A power spectral density (PSD) describes how signal power is distributed over frequency and is useful for comparing energy levels across frequency bands.

Frequency-domain interpretation is based on characteristic frequencies. Rotational frequency is often written as \(f_r\). Imbalance commonly produces a strong component near 1× rotational frequency, while angular or parallel misalignment may generate elevated 1× and 2× components, often with significant axial vibration. Mechanical looseness can produce harmonics of rotational frequency. Gear-mesh frequency is approximately the number of teeth multiplied by rotational frequency, with sidebands reflecting modulation by shaft speed or load variation. Rolling-element bearing defect frequencies depend on bearing geometry, shaft speed, contact angle, and the location of the defect.

A spectrum must not be interpreted as a list of universal fault signatures. Structural resonances can amplify a forcing component, and changes in speed shift characteristic frequencies. Windowing reduces spectral leakage when the recorded segment does not contain an integer number of cycles. For variable-speed machines, order tracking expresses frequency relative to rotational speed. This prevents a speed change from being mistaken for a change in fault frequency.

Envelope analysis is widely used for localized bearing and gear impacts. A band-pass filter selects a high-frequency resonance excited by impacts; demodulation then extracts the amplitude envelope, and an FFT of the envelope reveals repetition frequencies. The method is powerful but dependent on suitable filter-band selection and adequate sensor bandwidth.

Time–Frequency Methods

Many machines operate under changing speed, load, or control commands. A single spectrum averages behavior and can conceal when a fault occurs. Time–frequency analysis preserves information about both occurrence time and frequency content.

The short-time Fourier transform (STFT) divides a signal into overlapping windows and computes a spectrum for each window. It produces a spectrogram. Short windows improve time localization but reduce frequency resolution; long windows do the opposite. This is the fundamental time–frequency trade-off.

Wavelet transforms use scale-dependent basis functions and are often effective for transients, impacts, and rapidly changing signals. Continuous wavelet transforms provide a detailed time–scale map, whereas discrete wavelet transforms support multiresolution decomposition and feature extraction. Empirical mode decomposition and the Hilbert–Huang transform can represent nonlinear and nonstationary signals, but mode mixing and end effects require careful treatment.

Time–frequency features may include energy in selected bands, ridge frequency, transient duration, spectral entropy, and the rate of change of a feature. These features are valuable in run-up, coast-down, intermittent rubbing, variable-speed gearboxes, and systems controlled by changing set points. Nevertheless, visual patterns in a spectrogram should be validated against machine events and controlled experiments rather than interpreted solely by appearance.

Diagnostic Reasoning and Fault Classification

Diagnosis combines signal evidence with a model of how the machine can fail. A physics-based approach uses equations, defect frequencies, dynamic response, and expert rules. A data-driven approach learns relationships from labeled or unlabeled data. Hybrid approaches use physical constraints to improve statistical models.

A diagnostic workflow generally includes signal conditioning, segmentation, feature extraction, feature selection, classification, and validation. Classifiers may include logistic regression, decision trees, support-vector machines, random forests, neural networks, and probabilistic models. For unlabeled data, clustering and autoencoders can identify deviations from normal behavior.

Classification accuracy alone is inadequate for imbalanced maintenance data. A confusion matrix distinguishes true positives, false positives, true negatives, and false negatives. Precision measures the proportion of alarms that are correct, while recall measures the proportion of actual faults detected. In safety-critical settings, a false negative may be much more costly than a false positive. Receiver operating characteristic curves and precision–recall curves help compare thresholds, but the final threshold should reflect maintenance consequences.

Explainability is especially important. Engineers need to know why a model issued an alarm, which sensors contributed, and whether the result is physically plausible. Feature-attribution methods can assist interpretation, but they do not replace physical validation.

Prognostics and Remaining Useful Life

Prognostics estimates future health or the time until a defined failure criterion. Remaining useful life (RUL) is not an inherent constant; it depends on the failure definition, required performance, operating conditions, and maintenance policy. A bearing may remain capable of rotation but no longer meet vibration or precision requirements.

A degradation model may use a health indicator \(h(t)\), such as envelope energy or temperature residual. A simple trend model is

\[

h(t)=h_0+at+\epsilon(t)

\]

where \(a\) is a degradation rate and \(\epsilon(t)\) represents uncertainty. More realistic models allow nonlinear degradation, random effects among assets, changing operating conditions, and process noise. State-space models, particle filters, Bayesian updating, survival models, and recurrent neural networks are among the available approaches.

A useful RUL result is probabilistic rather than a single date. It may be expressed as a probability distribution or prediction interval. Uncertainty arises from sensor noise, unknown loads, model inadequacy, variations among components, and ambiguous failure thresholds. The model should be recalibrated as new evidence arrives.

Prognostics can be divided into model-based, data-driven, and hybrid methods. Model-based methods are attractive when failure physics are understood and data are limited. Data-driven methods can capture complex patterns but require representative data and are vulnerable to distribution shift. Hybrid methods combine degradation equations, parameter estimation, and machine learning.

Predictive Maintenance and Decision Economics

Predictive maintenance uses condition information and forecasts to determine what action should be taken and when. It is not simply a prediction system. The output must connect to work orders, spare-parts planning, technician skills, production scheduling, and safety procedures.

A decision policy can compare expected costs. If an intervention costs \(C_m\), failure costs \(C_f\), and the estimated probability of failure before the next opportunity is \(P_f\), a simplified policy may intervene when the expected avoided failure cost exceeds intervention cost. In practice, downtime, quality loss, environmental consequence, safety risk, inventory cost, and lost production must also be considered. Maintenance optimization may use reliability-centered maintenance, risk-based inspection, Markov decision processes, or stochastic optimization.

The most suitable strategy is often mixed. Critical assets may receive continuous monitoring, less critical assets periodic route-based measurements, and low-consequence assets run-to-failure treatment. Monitoring should be justified by the failure mode being detected and the action that will follow an alarm.

Digital Twins and Virtual Commissioning

A digital twin is a persistent digital representation of a physical asset, process, or system that is connected to the physical counterpart through data and models. A static CAD model is not necessarily a digital twin. A useful twin includes a defined purpose, identity, state information, model logic, data interfaces, and a method for updating and validating its state.

A practical architecture contains:

1. The physical asset and its sensors or control system.

2. An edge layer for acquisition, filtering, synchronization, and rapid local decisions.

3. A data layer containing time-series data, asset metadata, maintenance records, and event histories.

4. A model layer containing physics-based, statistical, and machine-learning models.

5. An application layer for visualization, diagnosis, simulation, optimization, and work management.

6. Governance mechanisms for access control, versioning, cybersecurity, and model validation.

Digital twins may operate at component, machine, production-line, or fleet level. A bearing twin may estimate stiffness, clearance, temperature, and damage state. A machine twin may predict vibration under different speeds and loads. A plant twin may evaluate production schedules, energy use, bottlenecks, and maintenance windows.

Virtual commissioning uses a digital model to test control logic and automation sequences before physical installation. A controller can be connected to a simulated machine, allowing engineers to verify interlocks, timing, fault responses, and operating sequences. This reduces commissioning risk and enables parallel development of mechanical, electrical, and software systems. The model must nevertheless be validated; an inaccurate simulation can produce false confidence.

Lifecycle optimization extends the twin beyond maintenance. The same data can support design improvement, commissioning, operation, energy optimization, retrofit evaluation, and end-of-life decisions. Feedback from operation may reveal that a design is over-constrained, poorly cooled, difficult to maintain, or sensitive to a particular duty cycle.

Illustrative Case: Predictive Monitoring of a Variable-Speed Pump Train

Background

Consider an illustrative centrifugal-pump train driven by a variable-speed motor. The train includes a motor, flexible coupling, pump shaft, rolling-element bearings, and a process-control valve. The pump supplies a process line whose demand varies throughout the day.

Situation and Problem

Operators report intermittent vibration alarms during speed changes. Overall vibration is normal at steady operation but rises during some run-up events. Replacing the pump bearing after every alarm would be expensive, while ignoring the alarms could lead to a shaft or bearing failure.

Analysis

The monitoring team installs radial and axial accelerometers near the motor and pump bearings, a tachometer, discharge-pressure measurement, motor-current measurement, and bearing-temperature sensors. Data are collected across the normal speed and flow envelope. The team first verifies sensor mounting and synchronizes measurements with speed.

Time-domain analysis shows that the highest RMS values occur during run-up, not at constant speed. An order-tracked spectrum reveals a strong component near 2× rotational order in the axial direction, accompanied by coupling-related harmonics. The component moves with shaft speed, indicating a mechanical excitation rather than a fixed structural resonance. Pressure data show that the valve closes rapidly during the same events.

Envelope analysis does not show a persistent bearing-defect pattern, and temperature remains stable. Therefore, the evidence does not support bearing damage as the primary cause. A temporary misalignment force caused by rapid torque and hydraulic transients is more plausible. A finite-element and multibody simulation of the shaft–coupling system predicts increased axial response when the control sequence produces a rapid torque reversal.

Approach and Solution

The control engineer modifies the ramp rate and valve sequence to reduce transient torque. The maintenance team checks coupling alignment and soft-foot condition during the next planned shutdown. A diagnostic rule is implemented: alarms are escalated only when elevated axial 2× order persists during steady operation or when accompanied by temperature rise and bearing-envelope growth. A digital-twin dashboard compares measured order response with simulated response for current speed and valve position.

Outcome

Following the control modification, run-up vibration decreases and nuisance alarms become less frequent. No unnecessary bearing replacement is performed. The monitoring program now distinguishes transient response from progressive degradation and provides operators with an explanation linked to the control sequence.

Learning Points

The case illustrates that a high vibration value is not itself a diagnosis. Speed synchronization, operating context, multiple sensors, and dynamic modeling prevented a false conclusion. It also shows how controls, maintenance, and digital models can be combined rather than treated as separate activities.

Implementation Challenges, Risks, and Professional Responsibilities

A frequent mistake is installing sensors before defining failure modes and decisions. Another is training a model only on healthy data from one operating regime and then treating every new regime as a fault. Data leakage, inconsistent labels, overfitting, unexamined sensor drift, and changing machine configurations can invalidate apparently accurate models.

False alarms reduce trust, while missed alarms create safety and reliability risks. Alarm fatigue develops when thresholds are too sensitive or poorly prioritized. Every alarm should have a defined response, urgency category, responsible person, and escalation path. Model performance should be monitored after deployment because machine condition, process conditions, and sensor behavior change.

Digital systems introduce cybersecurity concerns. Connected assets require authentication, least-privilege access, network segmentation, secure updates, logging, and protection against manipulation of sensor data or control commands. Data ownership and access rights should be defined among equipment manufacturers, operators, contractors, and software providers. Maintenance records may also reveal production capability or commercially sensitive information.

Ethically, engineers must not present uncertain predictions as facts. A model should disclose its operating range, confidence, training limitations, and known failure modes. Safety-critical decisions require independent verification and fail-safe design. Automation should support competent human judgment rather than conceal accountability behind an opaque algorithm.

Current Developments and Best Practices

Current development is moving toward edge analytics, self-supervised learning, physics-informed machine learning, federated learning across fleets, and interoperable asset information models. Edge processing reduces communication latency and may protect sensitive raw data. Self-supervised methods exploit large quantities of unlabeled operational data, while physics-informed methods constrain predictions to physically plausible behavior. Fleet learning can improve rare-fault recognition, but differences among machines must be handled carefully.

Effective programs generally follow these practices:

• Begin with criticality analysis and failure-mode identification.
• Define the maintenance or operational decision before selecting analytics.
• Establish healthy baselines across relevant speed, load, and environmental conditions.
• Combine physics-based reasoning with statistical evidence.
• Synchronize signals with speed and preserve asset metadata.
• Validate models using time-separated and asset-separated data, not only random splits.
• Report uncertainty and use cost-appropriate alarm thresholds.
• Integrate alarms with work management and feedback from technicians.
• Review sensor health, model drift, cybersecurity, and performance after deployment.
• Treat the digital twin as a governed lifecycle system, not merely a visualization screen.

Chapter Summary

Condition monitoring converts measurements of machinery behavior into evidence about asset health. Time-domain statistics provide efficient screening, frequency-domain methods expose periodic and order-related fault signatures, and time–frequency methods address transients and nonstationary operation. Diagnosis identifies fault type and severity, while prognostics estimates future degradation and probabilistic remaining useful life. Predictive maintenance becomes effective only when analytical outputs are connected to maintenance decisions, costs, safety, and operational constraints.

Digital twins combine physical assets, data infrastructure, computational models, and governance. They support diagnosis, simulation, virtual commissioning, control improvement, and lifecycle optimization. Their value depends on data quality, model validity, cybersecurity, explainability, and integration with engineering practice. The central professional principle is to combine signal analytics with machine physics and disciplined decision-making.

Key Terms

Condition monitoring: Continuous or periodic measurement used to assess the health of machinery.

Fault detection: Identification that machine behavior has departed from an accepted healthy condition.

Fault diagnosis: Determination of the type, location, and severity of a fault.

Prognostics: Prediction of future degradation or failure progression.

Remaining useful life: Estimated time or usage before a defined performance or failure threshold is reached.

Predictive maintenance: Maintenance planned using condition information and forecasts rather than fixed intervals alone.

Order tracking: Analysis in which frequency is expressed relative to rotational speed.

Power spectral density: Distribution of signal power over frequency.

Envelope analysis: Demodulation method used to reveal repetitive impacts in a selected high-frequency band.

Spectrogram: Time–frequency representation produced by repeated spectral analysis of signal windows.

Aliasing: Misrepresentation of high-frequency content caused by insufficient sampling.

Health indicator: Measured or derived variable that represents degradation state.

Digital twin: Connected digital representation of a physical asset or system that is updated and used for analysis or decision-making.

Virtual commissioning: Testing control logic with a simulated machine or process before physical commissioning.

Model drift: Decline in model validity caused by changing data, equipment, or operating conditions.

False positive: Alarm indicating a fault when the relevant fault is absent.

False negative: Failure to issue an alarm when a relevant fault is present.

Physics-informed machine learning: Learning constrained or guided by physical laws, equations, or engineering relationships.

Lifecycle optimization: Coordinated improvement of design, operation, maintenance, energy use, and end-of-life decisions.

Review Questions

Short Answer Questions

1. Distinguish condition monitoring from fault diagnosis.

2. Why is a tachometer useful in variable-speed vibration analysis?

3. State two limitations of RMS vibration as a diagnostic indicator.

4. What is the purpose of envelope analysis?

5. Why should RUL generally be reported with uncertainty?

Long Answer Questions

1. Explain how time-domain, frequency-domain, and time–frequency analyses complement one another in machinery monitoring.

2. Describe a complete workflow for developing a predictive-maintenance system for a rotating machine.

3. Explain the architecture and engineering purposes of a digital twin.

4. Compare model-based, data-driven, and hybrid prognostic methods.

5. Discuss the technical, organizational, and cybersecurity challenges of deploying digital twins in industry.

Analytical / Application Questions

1. A gearbox shows increasing RMS vibration, but the increase occurs only when speed changes. Design an investigation using order tracking, time–frequency analysis, and operating metadata.

2. A classifier detects 95% of known faults but generates many false alarms in a new production regime. Analyze possible causes and propose a validation and deployment strategy.

3. Design a monitoring and decision policy for a pump whose failure has high production cost but low safety consequence. Include sensors, features, alarm logic, RUL uncertainty, and maintenance actions.

22. Multiple Choice Questions with Answers

Question 1

A vibration signal has a strong component at twice rotational speed, mainly in the axial direction, during steady operation. Which interpretation is most appropriate as an initial diagnostic hypothesis?

A. Possible coupling misalignment

B. Certain rolling-element bearing failure

C. Electrical aliasing only

D. Proof of cavitation

Correct Answer: A. Possible coupling misalignment

Question 2

Why is order tracking preferred over an ordinary fixed-frequency spectrum for a variable-speed machine?

A. It removes all measurement noise

B. It expresses components relative to rotational speed

C. It eliminates the need for a tachometer

D. It guarantees fault classification

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

Question 3

A monitoring model performs well on randomly divided data but poorly on a new machine. Which problem is most likely?

A. Excessive Nyquist frequency

B. Data leakage or lack of asset-independent validation

C. Perfect model generalization

D. Insufficient use of false positives

Correct Answer: B. Data leakage or lack of asset-independent validation

Question 4

What is the principal purpose of envelope analysis in bearing monitoring?

A. To measure average shaft displacement

B. To reveal repetitive impact frequencies after demodulation

C. To calculate process flow directly

D. To replace all temperature measurements

Correct Answer: B. To reveal repetitive impact frequencies after demodulation

Question 5

Which statement best distinguishes detection from diagnosis?

A. Detection estimates maintenance cost, while diagnosis measures sampling rate

B. Detection identifies abnormal behavior, while diagnosis identifies its likely cause

C. Detection requires a digital twin, while diagnosis never uses data

D. Detection predicts RUL, while diagnosis determines sensor bandwidth

Correct Answer: B. Detection identifies abnormal behavior, while diagnosis identifies its likely cause

Question 6

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

A. Failure thresholds and degradation conditions contain uncertainty

B. RUL is independent of operating conditions

C. Single dates are always mathematically impossible

D. Intervals eliminate the need for maintenance decisions

Correct Answer: A. Failure thresholds and degradation conditions contain uncertainty

Question 7

Which feature of a digital twin most clearly distinguishes it from a static CAD model?

A. It contains geometric information

B. It is connected to the physical asset and updated for analysis

C. It can be printed in three dimensions

D. It uses no engineering models

Correct Answer: B. It is connected to the physical asset and updated for analysis

Question 8

In the illustrative pump case, why were repeated bearing replacements rejected as the first response?

A. The pump had no sensors

B. Evidence indicated transient control-related excitation rather than persistent bearing damage

C. Bearing replacement is never useful

D. Temperature cannot be measured on pumps

Correct Answer: B. Evidence indicated transient control-related excitation rather than persistent bearing damage

Question 9

Which metric is especially important when missed faults are much more costly than unnecessary inspections?

A. Recall or sensitivity

B. File size

C. Sampling period alone

D. Mean signal value only

Correct Answer: A. Recall or sensitivity

Question 10

What is a major risk of training an anomaly-detection model on healthy data from only one operating regime?

A. It may classify normal changes in speed or load as faults

B. It necessarily improves transferability

C. It prevents all false negatives

D. It makes metadata unnecessary

Correct Answer: A. It may classify normal changes in speed or load as faults