Experimentation, Instrumentation and Data Analytics

Introduction

A major project in electrical engineering is not complete when a circuit, controller, machine, or embedded system appears to work. A defensible project must demonstrate how performance was measured, whether the measuring system was suitable, how data were processed, and how confidently conclusions can be stated. Experimentation converts an engineering question into observable evidence. Instrumentation provides the interface between a physical system and recorded data. Data analytics converts measurements into estimates, comparisons, models, and decisions.

A rigorous experimental process therefore links five activities: defining the measurement objective, designing a safe and controlled test, selecting and calibrating instruments, acquiring and cleaning data, and analysing and reporting uncertainty. Traceable records allow another engineer to reproduce the work and distinguish a genuine system effect from an instrument fault, environmental change, wiring error, or inappropriate statistical treatment.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. translate a project question into measurable variables and acceptance criteria;
2. design a safe, controlled, and repeatable electrical test setup;
3. select instruments according to range, resolution, accuracy, bandwidth, loading, and isolation;
4. explain calibration, verification, traceability, and correction;
5. acquire, timestamp, label, and preserve experimental data;
6. detect and treat missing values, outliers, noise, drift, and invalid readings;
7. calculate descriptive statistics and propagate measurement uncertainty;
8. select suitable plots and statistical comparisons;
9. maintain a laboratory record that supports reproducibility and auditability; and
10. recognise ethical, safety, cybersecurity, and data-governance responsibilities.

1. From Engineering Question to Measurement Plan

An experiment begins with a question, not with an available instrument. A useful measurement plan states:

• the system under test and its operating conditions;
• the independent variable that is deliberately changed;
• the dependent variable that is measured;
• controlled variables such as supply voltage, temperature, load, and switching frequency;
• the measurement range and expected magnitude;
• sampling rate and duration;
• the number of repetitions;
• acceptance criteria; and
• safety constraints and stopping conditions.

For example, “test the converter” is too vague. A measurable objective is: “Determine whether the converter efficiency remains at least 90% for input voltages from 18 V to 24 V and output currents from 0.5 A to 2.0 A, at an ambient temperature of 25 °C ± 2 °C.” This statement identifies the response, operating region, environmental condition, and criterion.

A variable should have a precise operational definition. If output power is required, it may be defined as the average of instantaneous power over a specified interval, rather than the product of separately averaged voltage and current when ripple is significant. The definition must also state whether transients, warm-up intervals, and fault events are included or excluded.

Experimental design principles

A controlled experiment changes one principal factor at a time when the purpose is diagnosis and simplicity is important. However, this approach can be inefficient and may miss interactions between factors. A small factorial design, in which two or more factors are deliberately combined, can reveal whether the effect of temperature depends on load current. At bachelor level, the essential requirement is to identify the factors and justify the chosen design rather than apply a design method mechanically.

Randomising the order of test conditions reduces the risk that time-dependent drift is mistaken for a treatment effect. Replicating measurements permits estimation of repeatability. A reference or baseline condition provides a comparison against which the modification is judged. Where possible, the operator should define the analysis before examining results; this reduces unconscious selection of favourable data.

2. Designing a Test Setup

A test setup should be electrically correct, measurable, safe, stable, and documented. A typical arrangement contains a source, protection, system under test, load, sensors, signal conditioning, data-acquisition hardware, and a computer or recorder. The signal path should be considered as a chain: each element can introduce attenuation, bandwidth limitation, noise, delay, offset, or isolation failure.

Electrical and physical considerations

Voltage measurements are connected in parallel and current measurements in series. An ammeter with significant burden voltage can alter the circuit being measured. A voltmeter with finite input resistance can load a high-impedance node. Differential probes and isolated inputs are required where the measurement reference is not at protective earth. Connecting an oscilloscope ground clip to a floating or high-side node may create a short circuit and expose the operator to hazardous potential.

The setup should include appropriately rated fuses, current limiting, emergency isolation, protective earth where required, insulated terminals, and barriers around exposed conductors. Ratings must consider nominal value, possible transient value, fault current, and energy. Lockout or isolation procedures are required for systems capable of dangerous voltage, current, stored energy, or mechanical motion.

Cable routing matters. Power cables and sensitive signal cables should be separated where practical. Shielding and a single-point signal reference can reduce electromagnetic interference, but shielding must not create unintended ground loops. Twisted pairs reduce magnetic pickup in differential measurements. Sensor leads should be mechanically secured to prevent intermittent contacts.

Measurement-system block diagram

Figure: A measurement chain from the physical system to recorded data, with traceability supporting every stage.

3. Selecting Instruments

Instrument selection is a specification-matching exercise. The most important criteria are summarised below.

Resolution, accuracy, and precision are different. A six-digit display may give excellent resolution while having poor accuracy if calibration is incorrect. Precision describes clustering of repeated values; accuracy describes closeness to the accepted value.

Sampling and aliasing

For a signal with highest significant frequency f_max, the sampling frequency f_s should exceed twice f_max to satisfy the ideal Nyquist condition:

Here, f_s and f_max are in hertz. This condition is necessary for band-limited signals but is not sufficient in practical systems: an anti-aliasing filter is required, and a margin above the theoretical minimum is normally selected. If a 10 kHz component is relevant, sampling at exactly 20 kS/s is a poor practical choice; a higher rate with suitable filtering is preferable.

Quantisation of an ideal n-bit ADC with input span V_span has an approximate step size:

where q is in volts if V_span is in volts. This represents quantisation resolution, not total measurement uncertainty. Offset, gain error, noise, reference instability, nonlinearity, and sensor error may dominate.

Instrument loading and four-wire measurement

When measuring a low resistance, lead resistance can be comparable to the resistance under test. A four-wire, or Kelvin, method drives current through one pair of leads and measures voltage with a high-impedance second pair. The estimated resistance is:

where V is the sense voltage in volts and I is the test current in amperes. Because the voltage leads carry negligible current, their voltage drop is much smaller than in a two-wire measurement. The method is useful only when the source is stable, contact resistance is controlled, and the measured device is not changing due to self-heating.

4. Calibration, Verification and Traceability

Calibration compares an instrument with a reference of known or accepted value under specified conditions. A calibration certificate normally states the instrument identity, range, environmental conditions, reference standards, results, uncertainty, date, and due date. Traceability means that the measurement result can be related through an unbroken documented chain of calibrations to recognised standards, with stated uncertainties.

Calibration is not the same as adjustment. Calibration determines performance; adjustment changes the instrument to reduce error. Verification is a check that the instrument remains suitable, often using a reference artefact before a test session. A daily check against a precision voltage source may detect damage, but it does not replace formal calibration.

A simple correction model is:

where x_indicated is the displayed value and C is a documented correction. If the calibration result shows a proportional gain error, a more suitable model may be:

where a is a dimensionless scale factor and b has the units of x. The correction is valid only over the range and conditions for which it was established.

Before use, inspect calibration status, leads, connectors, battery condition, ranges, zero, probe attenuation, and software configuration. Record instrument manufacturer, model, serial number, channel, range, firmware or software version, and calibration identifier. If an instrument is overdue, damaged, overloaded, or used outside its rated conditions, its results require documented evaluation and may need to be rejected.

5. Data Acquisition and Laboratory Records

Raw data should be preserved in an immutable or access-controlled form. Processed data should be stored separately, with a clear link to the raw file. Every record should identify:

• project and experiment number;
• date, time, operator, and location;
• system identification and hardware revision;
• instrument identification and calibration status;
• wiring or test-setup diagram;
• environmental conditions;
• test sequence and operating points;
• sampling rate, record length, filters, triggering, and averaging;
• units, channel names, scaling factors, and sensor orientations;
• anomalies, interruptions, rejected observations, and reasons; and
• software, scripts, and processing-version information.

A data dictionary prevents ambiguity. For example, “Iout” should state whether it is instantaneous current, RMS current, mean current, or a filtered value, and whether positive current denotes charging or discharging.

A useful file convention combines project, test, date, and run number, such as PV01_efficiency_2025-03-08_run03.csv. File names alone are not sufficient; metadata must be included in a laboratory notebook or machine-readable header.

6. Data Cleaning and Signal Preparation

Data cleaning must remove errors without hiding genuine behaviour. The original data should never be overwritten. First inspect the data visually and numerically. Check time monotonicity, duplicate timestamps, impossible values, saturation, discontinuities, missing records, unit conversion, and synchronisation between channels.

Missing values

A missing value should be labelled as missing, not silently replaced by zero. Interpolation may be acceptable for a short gap in a slowly varying signal, but it is inappropriate across a switching transient or fault event. The method, gap length, and reason must be documented. If missingness is related to the operating condition, deletion can bias the result.

Outliers

An outlier is an observation inconsistent with the assumed process; it is not automatically a bad observation. It may indicate a real transient, sensor saturation, loose connection, or data-entry error. Investigate the physical record, waveform, event log, and neighbouring channels before removal.

For a set of n observations x₁, x₂, …, xₙ, the sample mean is:

The sample standard deviation is:

The denominator n − 1 is used when s estimates the population standard deviation from a sample. The standard error of the mean is:

This describes uncertainty in the estimated mean under appropriate independent, identically distributed conditions; it is not the total uncertainty of the instrument or experiment.

The median is often preferable when distributions are skewed. The interquartile range, IQR = Q₃ − Q₁, describes the middle 50% of observations. A common exploratory rule flags values below Q₁ − 1.5 IQR or above Q₃ + 1.5 IQR. This is a screening rule, not proof that a value is erroneous.

Filtering

A moving average can reduce random noise but changes timing and can obscure transients. For a window of m equally spaced samples:

The window length and boundary treatment must be reported. Digital filters introduce amplitude and phase effects. A filter should be selected from the measurement objective, not chosen solely because it produces a smoother plot.

7. Uncertainty and Error Analysis

Error is the difference between a measured value and the accepted value; uncertainty quantifies incomplete knowledge about the measurand. Errors may be systematic, such as calibration offset or probe gain, or random, such as electrical noise and repeatability variation. Repeating a measurement reduces the random component of the mean but does not automatically remove systematic error.

For independent input quantities x₁, x₂, …, x_k and result y = f(x₁, x₂, …, x_k), the first-order combined standard uncertainty is:

This approximation assumes small uncertainties, a sufficiently smooth function, and independent inputs. If inputs are correlated, covariance terms must be included:

where u(x_i,x_j) is the covariance contribution. In many undergraduate applications, dominant sources are estimated separately and combined conservatively, but the assumptions must be stated.

For resistance R = V/I, with independent voltage and current uncertainties:

where u_r(R) is the relative standard uncertainty. The absolute standard uncertainty is u(R) = R u_r(R). This result follows from the general propagation equation and assumes V and I are nonzero and their uncertainties are small relative to their values.

An expanded uncertainty is commonly reported as U = k u_c, where k is a coverage factor selected for the required confidence interpretation. For approximately normal results and a large effective number of degrees of freedom, k ≈ 2 is often used for roughly 95% coverage, but the coverage statement and assumptions must be reported.

Worked example: efficiency measurement

An electrical converter has measured input voltage 24.00 V ± 0.05 V, input current 1.20 A ± 0.01 A, output voltage 12.00 V ± 0.03 V, and output current 2.20 A ± 0.02 A. Treating the stated uncertainties as standard uncertainties and assuming independence, define:

The nominal input power is 24.00 × 1.20 = 28.80 W. The nominal output power is 12.00 × 2.20 = 26.40 W. Therefore η = 26.40/28.80 = 0.9167, or 91.67%.

The relative standard uncertainty is:

The squared terms are approximately 0.00000625, 0.00008264, 0.00000434, and 0.00006944. Their sum is approximately 0.00016267, giving u_r(η) ≈ 0.01275. Thus u(η) ≈ 0.9167 × 0.01275 = 0.0117, or 1.17 percentage points. A k = 2 expanded result is approximately η = 91.7% ± 2.3 percentage points, subject to the stated independence and distribution assumptions. The conclusion should not claim that efficiency is exactly 91.67%; the uncertainty indicates the plausible measurement spread.

8. Statistical Analysis and Visualisation

Descriptive statistics should be matched to the question. Mean and standard deviation are useful for approximately symmetric measurements. Median and interquartile range are more robust for skewed data. Minimum and maximum values are useful for limits but are highly sensitive to sample size.

A scatter plot is appropriate for examining the relationship between two measured variables. A line plot is appropriate for time-series behaviour. A histogram shows distribution shape. A box plot compares groups compactly but does not show all observations. Error bars must be identified: they may represent standard deviation, standard error, confidence interval, or expanded uncertainty, and these are not interchangeable.

Correlation measures association, not causation. The Pearson sample correlation coefficient is:

It lies between −1 and +1 and is sensitive to outliers and nonlinear relationships. A strong correlation does not prove that changing x causes y to change.

For a straight-line model y = a + bx, b is the estimated change in y per unit change in x and a is the intercept. Regression residuals should be inspected for nonlinearity, unequal variance, and dependence. Extrapolation beyond the measured range is generally less reliable than interpolation.

Statistical significance and engineering significance differ. A very small change may be statistically detectable with many repeated observations but irrelevant to system performance. Conversely, a practically important difference may not be statistically conclusive with few noisy observations. Report effect size, uncertainty, sample size, and engineering acceptance criteria together.

9. Illustrative Case Study: Thermal Performance of a Power Converter

Background

The project team developed a 150 W DC–DC converter and needed to determine whether it could operate continuously at 100 W output without exceeding a component temperature limit of 85 °C.

Problem

Initial tests used a handheld thermometer and a single operating point. The readings varied substantially because airflow, warm-up time, sensor placement, and load resistance were not controlled. The team needed a repeatable method that related output power, efficiency, and temperature.

Analysis and approach

The team specified input voltage, output power, ambient temperature, fan setting, warm-up time, and sensor locations. A calibrated voltage and current measurement system recorded input and output quantities. Thermocouples were attached using the same mounting method at the switching device, magnetic component, and enclosure. A reference ambient sensor was placed away from heat sources.

Output power was stepped through 25 W, 50 W, 75 W, 100 W, and 125 W. Each condition was held for 15 minutes after a preliminary stabilisation period. Three runs were performed in randomised order. The acquisition rate was selected to capture average electrical quantities and temperature evolution; high-frequency switching waveforms were measured separately with an oscilloscope and suitable differential probe.

The raw files contained timestamps, set points, measured values, instrument identities, and an event log. A script checked missing samples, impossible temperatures, channel saturation, and time alignment. Temperature averages were calculated over the final two minutes of each hold period, while the maximum temperature was retained as a separate safety metric. No outlier was removed without an associated physical explanation.

Outcome

The analysis showed that efficiency decreased modestly above 75 W, while the switching-device temperature rose disproportionately because heat-sink performance depended on temperature difference and airflow. At 100 W, the mean temperature was below 85 °C, but the expanded uncertainty interval overlapped the limit under the warmest ambient condition. The team therefore did not claim unconditional compliance. It recommended a larger heat sink or restricted ambient operating range and identified the additional test required for final verification.

Learning points

The case illustrates that a credible conclusion depends on more than a plausible graph. Operating conditions, sensor placement, warm-up, calibration, repeated runs, uncertainty, and acceptance criteria must be connected. The conservative engineering decision was justified because uncertainty near a safety limit should not be ignored.

10. Professional, Ethical and Contemporary Considerations

Engineers have a duty to report inconvenient results, failed runs, and uncertainty. Selective deletion, changing axes to exaggerate an effect, reporting only favourable operating points, or presenting filtered data as raw data undermines professional integrity. Safety incidents and near misses should be recorded and investigated.

Laboratory data may include proprietary designs, personal information in operator logs, or network-connected equipment. Access control, backups, version control, malware protection, and clear ownership are therefore relevant. Automated scripts and AI-assisted analysis can accelerate plotting, anomaly detection, and code generation, but the engineer remains responsible for validation. AI-generated calculations must be checked against raw data, physical laws, units, and known limits. A model should not be used to invent missing measurements or conceal uncertainty.

Contemporary laboratories increasingly use remote instrumentation, synchronised distributed sensors, digital twins, and machine-learning diagnostics. These tools are valuable when their calibration, timestamp accuracy, training-data limitations, and failure modes are understood. Automation should include interlocks, watchdogs, range checks, and a human-accessible emergency stop.

Best-Practice Checklist

Before testing, confirm the objective, variables, limits, hazards, wiring, instrument ratings, calibration status, sampling plan, and stopping criteria. During testing, record configuration changes, environmental conditions, anomalies, and run identifiers. After testing, preserve raw files, verify units and timestamps, document cleaning decisions, calculate uncertainty, inspect plots and residuals, and compare results with predefined criteria. A final report should enable a competent engineer to reconstruct the setup and judge the strength of the evidence.

Chapter Summary

Experimentation in electrical engineering is a controlled method for producing defensible evidence. A measurement plan defines variables, conditions, repetitions, and acceptance criteria before data are collected. A sound test setup addresses loading, bandwidth, grounding, isolation, protection, and environmental control. Instrument selection must distinguish range, resolution, accuracy, repeatability, bandwidth, sampling rate, and safety rating.

Calibration compares instruments with references, while verification checks continuing suitability. Traceability requires documented links to recognised standards. Reliable data acquisition preserves raw observations and metadata, including instrument identity, configuration, units, timing, and anomalies. Cleaning should investigate missing values, outliers, saturation, and noise without concealing genuine system behaviour.

Descriptive statistics, plots, correlation, regression, and residual analysis help reveal patterns, but statistical association is not causation and statistical significance is not automatically engineering significance. Uncertainty analysis combines instrument, repeatability, and other relevant components under stated assumptions. Near a design or safety limit, uncertainty must be included in the decision. Ethical reporting, safe practice, version-controlled analysis, and reproducible records are essential professional outcomes.

Key Terms

Measurement measurand: The quantity intended to be measured.

Independent variable: A factor deliberately changed by the investigator.

Dependent variable: A response observed or measured during an experiment.

Instrumentation: Sensors, interfaces, instruments, and systems used to obtain measurements.

Resolution: The smallest increment an instrument can distinguish or display.

Accuracy: Closeness of a result to an accepted reference value.

Precision: Closeness of repeated results to one another.

Repeatability: Agreement under the same method, operator, equipment, and conditions.

Bandwidth: Frequency range over which a measurement system has specified response.

Aliasing: Misrepresentation of frequency content caused by inadequate sampling.

Calibration: Comparison of an instrument with a reference under specified conditions.

Verification: Evidence that an instrument or method remains suitable for use.

Traceability: Documented connection of a result to recognised standards through calibration chains.

Systematic error: Consistent bias caused by a persistent influence.

Random error: Unpredictable variation between repeated observations.

Uncertainty: Quantified doubt associated with a measurement result.

Outlier: An observation requiring investigation because it differs markedly from the assumed pattern.

Residual: Difference between an observed value and a model-predicted value.

Metadata: Information describing how, when, where, and with what configuration data were produced.

Reproducibility: Ability of another competent investigator to obtain and evaluate comparable results.

Review Questions

Short Answer Questions

1. Distinguish accuracy, precision, and resolution.
2. Why must an anti-aliasing filter normally accompany a practical sampling system?
3. What information should be included in a data dictionary?
4. Explain the difference between calibration and verification.
5. Why should raw data be preserved separately from processed data?

Long Answer Questions

1. Design a measurement plan for evaluating the efficiency of a battery charger over its operating range. Identify variables, controls, repetitions, safety provisions, and acceptance criteria.
2. Explain the main sources of loading, noise, grounding, and isolation error in an oscilloscope-based power-electronics test.
3. Describe a defensible workflow for cleaning a time-series data set containing missing samples, spikes, and sensor saturation.
4. Explain the propagation of uncertainty for a result calculated from measured voltage and current, including assumptions and limitations.
5. Discuss why a statistically significant result may still be of little engineering importance.

Analytical / Application Questions

1. A resistance is calculated from 5.00 V ± 0.02 V and 0.500 A ± 0.005 A. Calculate the nominal resistance and relative standard uncertainty, assuming independent small uncertainties.
2. A temperature limit is 80 °C. Three repeated tests give a mean of 78 °C with an expanded uncertainty of ±4 °C. Explain whether unconditional compliance can be claimed and propose an additional test or design action.
3. An engineer observes a strong correlation between switching frequency and measured temperature while load current also increases during the same test sequence. Explain why causation cannot yet be assigned and design a follow-up experiment.

Multiple Choice Questions with Answers

Question 1

Which property describes the closeness of repeated measurements to one another?

A. Accuracy

B. Precision

C. Bandwidth

D. Resolution

Correct Answer: B. Precision

Question 2

A voltage signal contains significant frequency components up to 8 kHz. Which sampling choice is most appropriate in practice?

A. 8 kS/s without filtering

B. 12 kS/s without filtering

C. 16 kS/s exactly, with no margin

D. A rate greater than 16 kS/s with suitable anti-aliasing filtering

Correct Answer: D. A rate greater than 16 kS/s with suitable anti-aliasing filtering

Question 3

What is the principal advantage of a four-wire resistance measurement?

A. It eliminates the need for calibration

B. It prevents all electrical noise

C. It reduces the effect of lead resistance in the voltage measurement

D. It allows resistance to be measured without a current

Correct Answer: C. It reduces the effect of lead resistance in the voltage measurement

Question 4

Which action best preserves experimental traceability?

A. Saving only the final graph

B. Recording instrument identity, settings, calibration status, units, and timestamps

C. Replacing rejected readings without explanation

D. Storing all files with the same generic filename

Correct Answer: B. Recording instrument identity, settings, calibration status, units, and timestamps

Question 5

For n independent repeated observations, the standard error of the sample mean is:

A. s × √n

B. s/n

C. s/√n

D. n/s

Correct Answer: C. s/√n

Question 6

An outlier in an electrical measurement should first be:

A. Deleted automatically

B. Replaced by the mean

C. Investigated against physical events, setup records, and neighbouring channels

D. Reported as proof that the instrument is inaccurate

Correct Answer: C. Investigated against physical events, setup records, and neighbouring channels

Question 7

Which statement about correlation is correct?

A. Correlation always proves causation

B. Correlation measures association and may be affected by confounding variables

C. Correlation can only be positive

D. Correlation is unaffected by outliers

Correct Answer: B. Correlation measures association and may be affected by confounding variables

Question 8

A displayed ADC has an input span of 10 V and 12 bits. Its ideal quantisation step is approximately:

A. 0.244 mV

B. 2.44 mV

C. 24.4 mV

D. 244 mV

Correct Answer: B. 2.44 mV

Question 9

If a measurement result is close to a safety limit and its uncertainty interval crosses that limit, the best engineering conclusion is to:

A. Round the result downward

B. Claim compliance based only on the mean

C. Ignore uncertainty because the instrument is calibrated

D. Avoid an unconditional compliance claim and perform further analysis or testing

Correct Answer: D. Avoid an unconditional compliance claim and perform further analysis or testing

Question 10

Which statement correctly distinguishes systematic and random error?

A. Systematic error is persistent bias, whereas random error varies unpredictably

B. Systematic error is removed by taking one measurement

C. Random error is always caused by incorrect wiring

D. They are identical terms for instrument resolution

Correct Answer: A. Systematic error is persistent bias, whereas random error varies unpredictably