Digital Monitoring, Automation, AI, and Cybersecurity

Introduction

Electrical power systems and rotating machines are increasingly operated through interconnected digital systems. A modern substation may contain intelligent electronic devices (IEDs), programmable logic controllers (PLCs), phasor measurement units (PMUs), protection relays, network switches, historians, engineering workstations, and supervisory control and data acquisition (SCADA) servers. In parallel, condition-monitoring sensors installed on transformers, generators, motors, and circuit breakers provide large quantities of operational data.

These technologies improve observability, reliability, safety, and maintenance planning. They also introduce new risks. Incorrect sensor calibration, poor time synchronization, inappropriate artificial-intelligence (AI) models, insecure remote access, or compromised control equipment can produce unsafe switching operations and widespread outages. Consequently, digitalization must be treated as an engineering discipline involving measurement, control, communication, data science, human factors, and cybersecurity.

This chapter develops the principles required to use digital monitoring and automation responsibly in electrical machines and power systems laboratories. It emphasizes the complete chain from physical measurement to operational decision:

sensor → data acquisition → communication → analysis → human or automatic action → verification

A digital system is useful only when every stage is sufficiently accurate, timely, secure, and understandable.

Learning Objectives

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

1. Explain the roles of SCADA, PLCs, IEDs, PMUs, and digital twins.
2. Distinguish supervisory control from fast automatic protection.
3. Interpret condition-monitoring data from electrical machines and transformers.
4. Explain synchrophasor measurement and time synchronization.
5. Apply basic data-quality checks and anomaly-detection methods.
6. Evaluate AI-based predictive maintenance with attention to explainability and uncertainty.
7. Design a laboratory automation architecture with appropriate cybersecurity controls.
8. Discuss privacy, safety, governance, and professional responsibilities in digital power engineering.

1. Digital Monitoring Architecture

1.1 From physical process to operational decision

A monitoring system measures physical variables, converts them into digital data, transports the data, and presents information to an operator or algorithm. Typical variables include voltage, current, active power, reactive power, frequency, winding temperature, bearing vibration, oil temperature, dissolved gases, breaker travel time, and partial-discharge indicators.

The quality of a measurement is not determined solely by the sensor. It also depends on signal conditioning, analogue-to-digital conversion, sampling rate, clock accuracy, communication reliability, database handling, and interpretation. A perfectly accurate sensor can still produce misleading information if its timestamp is wrong or if a scaling factor is incorrectly configured.

For a sampled sinusoidal signal, the sampling frequency must be sufficiently high to represent the highest frequency component of interest. Under the ideal Nyquist condition,

where f_s is the sampling frequency in hertz and f_max is the highest frequency component being measured. In practice, anti-aliasing filters and engineering margin are required. Aliasing occurs when higher-frequency content appears falsely as a lower-frequency component.

1.2 SCADA

SCADA systems supervise geographically distributed equipment. A typical arrangement contains field sensors and actuators, remote terminal units or PLCs, communication networks, a SCADA server, a historian, and human–machine interfaces (HMIs). SCADA commonly performs:

• acquisition of measurements;
• alarm generation;
• status indication;
• remote supervisory commands;
• event and trend recording; and
• operator reporting.

SCADA is generally not a substitute for dedicated protection. A protection relay must operate within a predictable time even when the supervisory network, server, or HMI is unavailable. SCADA commands should therefore be separated from essential protection functions and should require suitable authorization and interlocking.

1.3 PLCs and IEDs

A PLC is a rugged industrial controller designed to execute a programmed logic cycle. It reads inputs, executes logic, and updates outputs. PLCs are especially suitable for sequencing, interlocking, motor control, pump control, and laboratory automation.

An IED is a microprocessor-based device that performs one or more power-system functions. Examples include numerical protection relays, digital meters, transformer monitors, bay controllers, and fault recorders. Modern IEDs may support protection, control, measurement, event recording, and communication in a single device.

The engineering distinction is functional rather than merely physical. A PLC is commonly programmed for general-purpose control logic, while an IED is designed around power-system measurement, protection, and substation functions. Both require disciplined configuration management, testing, and cybersecurity.

1.4 Communication and interoperability

Digital systems exchange data through protocols such as Modbus, DNP3, IEC 60870-5-104, and IEC 61850. IEC 61850 supports structured models and fast peer-to-peer communication within substations. Protocol selection should consider determinism, interoperability, authentication, availability, engineering support, and the consequences of failure.

Interoperability does not mean that devices can be connected without engineering. Data points must have consistent names, units, scaling, quality flags, timestamps, and semantic meaning. For example, a value labelled “current” is incomplete unless the phase, location, direction convention, nominal ratio, unit, and timestamp are known.

2. Synchrophasors and Time-Aware Measurement

A conventional phasor represents the magnitude and phase angle of a sinusoidal quantity relative to a reference. A synchrophasor is a phasor whose time reference is synchronized, usually through a satellite-based or network-based clock. PMUs calculate voltage and current phasors, frequency, and rate of change of frequency with precise timestamps.

The complex representation of a sinusoidal voltage may be written as

where |V| is the rms magnitude and θ is the phase angle relative to the defined reference. The notation is meaningful only when the frequency, phase convention, measurement location, and time reference are specified.

Synchrophasors enable comparison of phase angles at geographically separated locations. Applications include oscillation monitoring, wide-area protection research, voltage-stability assessment, disturbance reconstruction, and state-estimation improvement. However, a PMU does not automatically provide a correct system-wide picture. Errors may arise from instrument transformers, GPS spoofing or loss of timing, communication delay, bad data, and inconsistent configuration.

A useful quality model separates:

1. measurement accuracy—how closely the value represents the physical quantity;
2. time accuracy—how correctly the timestamp represents the measurement instant;
3. availability—whether data arrive when needed;
4. integrity—whether data have been altered; and
5. latency—the delay between measurement and use.

These characteristics should be recorded with the data rather than hidden from the operator.

3. Condition Monitoring of Machines and Transformers

Condition monitoring estimates equipment health from measured physical or chemical indicators. It complements, but does not automatically replace, inspection, protection, and scheduled maintenance.

3.1 Rotating machines

Important indicators for induction and synchronous machines include:

• stator current and voltage imbalance;
• winding and bearing temperature;
• vibration amplitude and frequency spectrum;
• shaft speed and torque;
• insulation resistance and polarization index;
• partial discharge; and
• acoustic or stray-flux signatures.

Vibration analysis is particularly valuable for bearings, misalignment, looseness, unbalance, and some electromagnetic defects. A spectrum should be interpreted using shaft speed, pole number, bearing geometry, load, and machine construction. A peak at a frequency is not by itself a diagnosis.

For a balanced three-phase system, voltage unbalance can be expressed using symmetrical components. If V₁ is the positive-sequence voltage and V₂ is the negative-sequence voltage, the negative-sequence voltage unbalance factor is

This ratio is dimensionless. It is applicable when positive- and negative-sequence components are defined consistently and |V₁| is nonzero. Negative-sequence voltage produces a counter-rotating magnetic field in a motor and can cause additional rotor heating.

3.2 Transformers

Transformer monitoring may include top-oil temperature, winding temperature, load current, moisture, bushing power factor, dissolved-gas analysis, and partial discharge. Dissolved-gas analysis can indicate possible thermal faults, arcing, or paper insulation degradation, but interpretation requires trends, gas ratios, sampling quality, transformer type, and expert judgement. A single threshold should not be treated as a universal diagnosis.

Temperature-based loading assessment often uses thermal models. A simplified first-order model is

where T(t) is the temperature at time t, T_ambient is ambient temperature, ΔT_final is the eventual temperature rise, and τ is the thermal time constant. The model assumes a constant operating condition and a lumped thermal response. Real transformers may require separate oil and winding models, changing ambient conditions, and load-dependent losses.

3.3 Data fusion

Reliable condition assessment combines several evidence types: sensor values, inspection records, maintenance history, operating load, environmental conditions, and protection events. Data fusion reduces the danger of diagnosing a defect from one abnormal but harmless measurement.

4. Digital Twins

A digital twin is a computational representation linked to a physical asset through data, models, and updating procedures. It is more than a static simulation or a three-dimensional drawing. A useful twin specifies:

• the physical asset and its boundaries;
• the mathematical or empirical model;
• the data sources and update frequency;
• the validation method;
• uncertainty and confidence information; and
• the decisions the twin is intended to support.

A generator twin might simulate electromagnetic performance, thermal behaviour, vibration, and remaining useful life. A laboratory motor twin can compare measured current and temperature with expected values under changing load.

Models may be physics-based, data-driven, or hybrid. Physics-based models are often more interpretable and may extrapolate better outside the training data, but they can require detailed parameters. Data-driven models can capture complex relationships but may fail under operating conditions absent from the training set. Hybrid models combine physical constraints with statistical learning.

A twin must be validated against independent measurements. Calibration data used to fit a model should not be the only evidence used to claim accuracy. Model drift occurs when equipment ages, sensors are replaced, operating regimes change, or the environment differs from the development data.

5. AI for Anomaly Detection and Predictive Maintenance

5.1 Anomaly detection

An anomaly is an observation or pattern that differs significantly from expected behaviour. It is not necessarily a fault. Causes include sensor failure, communication errors, unusual but safe operating conditions, and genuine equipment degradation.

A simple standardized residual is

where x_t is the current observation, μ is the reference mean, and σ is the reference standard deviation. This method assumes that the reference population is reasonably stable and that σ is nonzero. A large absolute value of z_t may indicate an unusual observation, but it is not a proof of failure. For skewed or contaminated data, robust statistics such as the median and median absolute deviation may be more appropriate.

A moving average can reduce random noise:

where n is the number of observations in the window. Larger windows provide stronger smoothing but introduce more delay and may hide short events.

Supervised classification requires labelled examples of normal and fault conditions. Such labels are often scarce and biased toward known failures. Unsupervised or semi-supervised methods learn normal behaviour and flag deviations, but they still require engineering review.

5.2 Predictive maintenance

Predictive maintenance estimates whether intervention should be planned before failure. Common outputs include an anomaly score, fault class, probability of failure within a specified horizon, or remaining useful life. These outputs must include the prediction horizon and uncertainty.

Classification performance should not be judged by accuracy alone. If failures are rare, a model can achieve high accuracy by predicting “no failure” every time. A confusion matrix distinguishes true positives, false positives, true negatives, and false negatives. In safety-critical maintenance, the cost of a missed failure may be much higher than the cost of an unnecessary inspection.

A decision threshold should therefore reflect consequence, inspection cost, equipment criticality, and available redundancy. The final maintenance decision should remain traceable to evidence and engineering rules.

5.3 Explainability and human oversight

Explainability may involve feature importance, trend plots, local explanations, physical constraints, comparable historical cases, or a clear rule-based rationale. An explanation should answer: What evidence influenced the result? Over what time interval? How reliable is the evidence? What conditions were assumed?

Human oversight is essential when an AI output can initiate switching, trip equipment, inhibit protection, or defer safety-related maintenance. Autonomous action should be limited to functions that have been rigorously validated, bounded, and fail-safe.

6. Data Quality and Governance

A data-quality framework should address:

• completeness: are expected records present?
• validity: are values within physically meaningful ranges?
• consistency: do related measurements agree?
• timeliness: did the data arrive within the required interval?
• accuracy: does the measurement agree with a trusted reference?
• uniqueness: have duplicate records been removed?
• traceability: can the origin and transformation of each value be reconstructed?

Quality flags should distinguish valid, substituted, estimated, stale, out-of-range, and communication-failed values. Missing data should not silently be replaced by zeros. Imputation may be acceptable for exploratory analysis but must be labelled and generally excluded from automatic protection decisions.

Data governance defines ownership, access rights, retention, version control, calibration records, and permitted uses. Laboratory data may contain student identifiers, equipment access records, or security-sensitive network information. Privacy protection requires data minimization, access control, secure storage, and appropriate anonymization.

7. Cybersecure Operation

7.1 Threats and consequences

Power-system digital assets may be exposed to malware, credential theft, ransomware, denial of service, insider misuse, supply-chain compromise, protocol abuse, and manipulation of time sources. Consequences include loss of visibility, false alarms, incorrect control actions, equipment damage, safety hazards, and prolonged outage.

Cybersecurity objectives are commonly summarized as:

• confidentiality: preventing unauthorized disclosure;
• integrity: preventing unauthorized alteration; and
• availability: ensuring timely access to required functions.

In operational technology, availability and safety may take priority over confidentiality, but all three objectives remain relevant.

7.2 Defence-in-depth

Effective protection uses multiple layers:

1. asset inventory and network diagrams;
2. separation of corporate information technology from operational technology;
3. network segmentation and controlled conduits;
4. least-privilege accounts and multifactor authentication where feasible;
5. secure configuration and removal of unused services;
6. patch and vulnerability management with operational testing;
7. signed firmware and controlled engineering changes;
8. backups and recovery exercises;
9. security logging and anomaly monitoring; and
10. incident-response procedures and staff training.

Safety systems should fail in a predictable state. Remote commands should use confirmation, interlocks, authority checks, and independent local indications. Cybersecurity controls must not interfere with emergency shutdown or essential protection.

Relevant professional practice is informed by standards and guidance such as IEC 62443 for industrial automation and control-system security, IEC 62351 for security of power-system communication, ISO/IEC 27001 for information-security management, and NIST cybersecurity guidance. Standards should be applied according to the jurisdiction, asset class, and responsible authority; citing a standard does not replace engineering risk assessment.

8. Laboratory Implementation Workflow

A responsible laboratory project can follow this sequence:

1. Define the operational question, such as detecting bearing degradation.
2. Identify hazards, critical functions, and acceptable response times.
3. Select sensors and verify range, accuracy, bandwidth, isolation, and calibration.
4. Design acquisition, time synchronization, storage, and network segmentation.
5. Establish data-quality rules and label operating modes.
6. Collect baseline data across normal loads and environmental conditions.
7. Develop a transparent model before considering complex AI.
8. Test using data not used for model fitting.
9. Compare model output with independent inspection or known events.
10. Define human approval, alarm priorities, and fallback operation.
11. Document versions, assumptions, limitations, and cybersecurity controls.
12. Review performance periodically for drift and unintended consequences.

A laboratory should never create a fault merely to test an AI system unless the experiment has been risk-assessed and approved. Safe emulation, historical data, hardware-in-the-loop simulation, and controlled low-energy test rigs are preferred.

9. Illustrative Case Study: Condition Monitoring of a Laboratory Induction Motor

Background

An engineering laboratory operates a 7.5 kW induction motor driving a variable mechanical load. The motor has current transformers, a three-axis accelerometer, winding-temperature sensors, and a tachometer. Data are stored every second, while vibration waveforms are captured at a higher sampling rate during scheduled tests.

Problem

Operators observe increasing vibration and occasional temperature alarms. The laboratory must determine whether the cause is bearing degradation, shaft misalignment, overload, or a faulty sensor. The motor cannot be allowed to fail because it is required for several teaching experiments.

Analysis

First, the team checks sensor calibration, timestamp continuity, cable shielding, and data-quality flags. The temperature sensor is found to contain intermittent missing values, but the vibration and current channels are reliable. The team compares vibration features with shaft speed and load rather than using a fixed threshold.

A baseline is created from several weeks of normal operation. For each operating regime, the residual between measured vibration and expected vibration is calculated. A high residual is flagged only when it persists across multiple observations and is supported by a consistent spectral feature.

The team also checks voltage unbalance and current imbalance. If V₁ = 230 V and V₂ = 2.3 V, then

This value alone does not establish a motor fault; it is one item of evidence. The vibration spectrum shows a speed-related component and sidebands consistent with mechanical misalignment. A visual inspection confirms that the coupling alignment is outside the manufacturer’s recommended tolerance.

Approach and solution

The laboratory does not permit the AI model to trip the motor. Instead, the model generates a graded advisory alarm. A qualified technician verifies the alarm, inspects the coupling, corrects alignment, and repeats the test. The sensor-quality problem is corrected separately, and the model is retrained using labelled operating conditions while preserving the original version for audit purposes.

Outcome

Vibration decreases after alignment, and temperature returns to the normal range. The system remains useful because it combines physics-based checks, data-quality validation, trend analysis, and human inspection. The AI model supports prioritization but does not replace protection or professional judgement.

Learning points

• An anomaly is not automatically a failure.
• Operating context is essential for interpreting sensor data.
• Data-quality checks must precede AI analysis.
• Independent verification is required before maintenance action.
• Advisory systems are safer than uncontrolled autonomous actions in an educational laboratory.

10. Professional, Ethical, and Regulatory Considerations

Engineers are responsible for foreseeable consequences of digital systems. A model that is accurate on average may still be unacceptable if it misses rare but dangerous faults. Ethical engineering requires disclosure of uncertainty, limitations, data provenance, and known failure modes.

AI systems may reproduce bias when training data overrepresent one machine type, climate, manufacturer, or operating range. Privacy obligations arise when monitoring includes personnel identifiers, access logs, or video. Data should be collected for legitimate purposes, retained only as long as necessary, and protected from unauthorized use.

Accountability must remain clear. “The algorithm recommended it” is not an adequate engineering justification. Organizations should define who approves models, who authorizes control actions, who reviews alarms, and who responds to cyber incidents. Change management should cover software versions, model parameters, firmware, network rules, and sensor replacements.

Contemporary Developments and Limitations

Edge computing allows data to be processed near the equipment, reducing communication latency and bandwidth requirements. Federated learning can allow multiple sites to improve a model without directly pooling all raw data, although it does not eliminate privacy or model-poisoning risks. Physics-informed machine learning constrains models using known electrical or thermal relationships. Digital substations increasingly use process-bus architectures and sampled values, demanding careful synchronization and network engineering.

Nevertheless, digital technologies have limitations. Sensors fail, labels are incomplete, rare failures are difficult to model, and operating conditions change. Complex models may be difficult to validate or explain. Cybersecurity controls can introduce cost and operational constraints. Automation can also create automation bias, in which operators trust a computer-generated recommendation more than contradictory physical evidence. Good engineering therefore combines automation with independent protection, competent personnel, transparent procedures, and periodic review.

Best-Practice Checklist

Before deploying a digital monitoring or AI function, verify that:

• the engineering purpose and decision consequence are defined;
• sensor accuracy, calibration, isolation, and sampling are suitable;
• timestamps and units are reliable;
• missing and substituted data are explicitly identified;
• baseline data cover relevant operating conditions;
• the model has been tested on independent data;
• uncertainty and false-alarm costs are understood;
• human approval and fallback procedures exist;
• cybersecurity architecture and access rights are documented;
• changes are version-controlled and auditable; and
• performance is reviewed after deployment.

Chapter Summary

Digital monitoring connects physical electrical assets to information, analysis, and control. SCADA provides supervisory visibility and commands, while PLCs support programmed automation and IEDs combine power-system measurement, protection, control, and communication. PMUs extend measurement across the grid by providing time-synchronized phasors, but their value depends on measurement accuracy, timing, availability, integrity, and latency.

Condition monitoring uses electrical, thermal, mechanical, and chemical indicators to assess asset health. Digital twins link physical assets with computational models and live data. AI can detect anomalies and support predictive maintenance, but an anomaly is not necessarily a fault, and model performance must be evaluated with appropriate data, uncertainty, operating context, and consequence-based thresholds.

Data quality, explainability, privacy, and governance are essential. Cybersecure operation requires defence-in-depth, segmentation, least privilege, secure configuration, recovery planning, and trained personnel. In professional practice, AI should support—not obscure—engineering accountability. Safe digitalization combines validated models, independent protection, human oversight, and continuous review.

Key Terms

SCADA: Supervisory system for monitoring and controlling geographically distributed processes.

PLC: Programmable logic controller used for industrial sequencing and control.

IED: Intelligent electronic device used for digital measurement, protection, control, or communication.

PMU: Phasor measurement unit that produces time-synchronized electrical phasors.

Synchrophasor: A phasor measurement referenced to a common time standard.

Digital twin: A data-linked computational representation of a physical asset.

Condition monitoring: Measurement and interpretation of equipment condition indicators.

Anomaly detection: Identification of observations that differ from expected behaviour.

Predictive maintenance: Maintenance planning based on estimated future condition or failure risk.

Residual: Difference between an observed value and an expected or modelled value.

Data quality: Fitness of data for a defined engineering purpose.

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

Defence-in-depth: Use of multiple independent cybersecurity protection layers.

Integrity: Protection of data and systems against unauthorized alteration.

Availability: Ability to access a required function or data when needed.

Explainability: Ability to communicate how and why a model produced an output.

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

Edge computing: Processing data near the physical source rather than only in a central server.

Least privilege: Granting users and devices only the access necessary for their tasks.

Review Questions

Short Answer Questions

1. What are the principal functions of a SCADA system?
2. Why should SCADA supervisory control not be treated as a replacement for dedicated protection?
3. Define a synchrophasor and state one application of PMUs.
4. What is the difference between an anomaly and a confirmed equipment fault?
5. List four dimensions of data quality relevant to power-system monitoring.

Long Answer Questions

1. Explain the roles and differences of PLCs, IEDs, SCADA servers, and historians in a digital substation or laboratory.
2. Discuss how condition-monitoring data from an induction motor can be combined to assess bearing or alignment problems.
3. Explain the concept of a digital twin, comparing physics-based, data-driven, and hybrid models.
4. Describe a defence-in-depth cybersecurity architecture for a laboratory power-system automation network.
5. Discuss why explainability, uncertainty, and human oversight are necessary when AI is used for predictive maintenance.

Analytical / Application Questions

1. A motor has positive-sequence voltage V₁ = 400 V and negative-sequence voltage V₂ = 8 V. Calculate the negative-sequence voltage unbalance factor and explain why the result should not be used as the sole basis for declaring a motor fault.
2. A vibration model produces a reference mean of 3.0 mm/s and standard deviation of 0.5 mm/s. A new observation is 4.5 mm/s. Calculate the standardized residual and identify two additional checks required before issuing a maintenance alarm.
3. Design a laboratory workflow for an AI-based transformer condition-monitoring system. Identify the sensors, data-quality checks, cybersecurity controls, validation method, alarm response, and fallback procedure.

Multiple Choice Questions with Answers

Question 1

Which function is most characteristic of a SCADA system?

A. Supervisory monitoring and remote control

B. Replacing all local protection relays

C. Manufacturing transformer insulation

D. Removing the need for sensor calibration

Correct Answer: A. Supervisory monitoring and remote control

Question 2

What makes a synchrophasor different from an ordinary phasor measurement?

A. It is always measured without an instrument transformer

B. It is referenced to a synchronized time standard

C. It does not include phase angle

D. It can be used only in dc systems

Correct Answer: B. It is referenced to a synchronized time standard

Question 3

For V₁ = 230 V and V₂ = 2.3 V, the negative-sequence voltage unbalance factor is:

A. 0.01%

B. 0.1%

C. 1.0%

D. 10%

Correct Answer: C. 1.0%

Question 4

Which statement about an anomaly is correct?

A. Every anomaly proves that equipment has failed

B. An anomaly is a deviation that requires interpretation and verification

C. An anomaly can occur only in a cyberattack

D. An anomaly is identical to a protection trip

Correct Answer: B. An anomaly is a deviation that requires interpretation and verification

Question 5

What is a principal advantage of edge computing?

A. It eliminates all cybersecurity risks

B. It can reduce communication latency by processing data near the asset

C. It guarantees perfect sensor accuracy

D. It removes the need for model validation

Correct Answer: B. It can reduce communication latency by processing data near the asset

Question 6

A standardized residual uses the expression z_t = (x_t − μ)/σ. If x_t = 4.5, μ = 3.0, and σ = 0.5, what is z_t?

A. 0.33

B. 1.5

C. 3.0

D. 7.5

Correct Answer: C. 3.0

Question 7

Which cybersecurity principle grants a user only the access necessary for assigned duties?

A. Least privilege

B. Open trust

C. Maximum connectivity

D. Unrestricted delegation

Correct Answer: A. Least privilege

Question 8

Why is accuracy alone a poor metric for a rare-failure classifier?

A. Accuracy cannot be calculated for digital data

B. A model can appear accurate by predicting the common normal class every time

C. Accuracy is unrelated to observations

D. Accuracy applies only to analogue instruments

Correct Answer: B. A model can appear accurate by predicting the common normal class every time

Question 9

Which statement best describes a digital twin?

A. A static equipment photograph

B. A data-linked computational representation of a physical asset

C. A replacement for all physical inspections

D. A communication cable between two substations

Correct Answer: B. A data-linked computational representation of a physical asset

Question 10

Why should missing data not automatically be replaced by zero?

A. Zero is always outside the measuring range

B. The replacement can falsely imply a real physical zero and distort decisions

C. Zeros cannot be stored in historians

D. Replacement is prohibited in every form of analysis

Correct Answer: B. The replacement can falsely imply a real physical zero and distort decisions