Integrated Industrial Project and Sustainable Decision-Making

Introduction

Thermal and manufacturing laboratories traditionally develop individual technical skills: measuring temperature, determining efficiency, analysing surface finish, estimating uncertainty, or designing a process. Industrial engineering practice, however, requires these skills to be integrated into a decision. A technically efficient process may produce excessive scrap; a low-cost process may increase emissions; a high-quality design may be unreliable or impossible to regulate. The professional task is therefore to identify the system boundary, generate reliable evidence, evaluate competing objectives, and recommend an implementable course of action.

An integrated industrial project combines experimental design, measurement, statistical analysis, engineering economics, sustainability assessment, risk analysis, and communication. This chapter presents a structured method for a team-based thermal or manufacturing investigation. The approach is applicable to projects such as selecting machining parameters, improving a heat-treatment cycle, reducing furnace energy consumption, comparing cooling methods, or evaluating additive and conventional manufacturing routes.

The central principle is that sustainability is not an additional calculation performed after technical design. It is a decision criterion that must be incorporated from problem definition through operation, maintenance, end-of-life, and reporting.

Learning objectives

After studying this chapter, students should be able to:

1. Define an industrial laboratory problem, system boundary, constraints, and measurable responses.
2. Plan a team investigation using design of experiments (DOE), calibration, uncertainty analysis, and safe operating procedures.
3. Analyse technical, quality, reliability, cost, energy, emissions, waste, and life-cycle evidence.
4. Apply multi-criteria decision-making (MCDM) transparently and test the robustness of a recommendation.
5. Recommend lean, circular, resilient, and digitally enabled improvements.
6. Present defensible engineering conclusions while recognising ethical, regulatory, and data-governance obligations.

1. From laboratory exercise to industrial decision

An industrial project begins with a decision statement rather than a measurement statement. “Measure furnace efficiency” is incomplete. A stronger statement is: “Select an operating condition that achieves the required product hardness and dimensional quality while reducing energy per acceptable part, subject to safety and emissions limits.”

A useful project charter contains:

The system boundary must be explicit. For a machining project, it may include raw material, machine electricity, cutting fluid, tooling, scrap, inspection, and transport. For a thermal process, it may include fuel or electricity, warm-up, holding, cooling, rejected products, and emissions. Comparisons are valid only when alternatives are evaluated using the same functional unit, such as one acceptable component meeting specified tolerances.

A process map should distinguish inputs, transformations, outputs, and losses. Typical inputs include material, electricity, fuel, compressed air, water, tooling, and operator time. Outputs include product, heat, noise, wastewater, airborne emissions, chips, rejected parts, and data. This map helps prevent burden shifting—for example, reducing electricity while increasing water consumption or scrap.

2. Team organisation and project governance

A team should allocate responsibilities without isolating knowledge. Typical roles are project coordinator, experimental lead, measurement and calibration lead, data analyst, sustainability and cost analyst, and safety or compliance lead. Every critical result should be independently checked by another member.

The team should maintain a version-controlled project record containing the charter, risk assessment, raw data, calibration certificates, revised procedures, analysis scripts, decisions, and meeting records. Raw observations must never be overwritten by cleaned data. Corrections should be traceable, dated, and justified.

A stage-gate structure is effective:

1. Define: approve the decision question, boundary, success criteria, and risks.
2. Design: select factors, levels, responses, replicates, and measurement methods.
3. Pilot: test feasibility, detect unsafe conditions, and estimate variability.
4. Experiment: collect data under controlled and documented conditions.
5. Analyse: estimate effects, uncertainty, trade-offs, and constraints.
6. Decide: compare feasible alternatives and test sensitivity.
7. Implement and learn: conduct a confirmation run and establish monitoring.

A decision log is as important as a final graph. It records why a factor was excluded, why a boundary changed, or why a low-scoring alternative was rejected.

3. Design of experiments for thermal and manufacturing systems

DOE changes process factors deliberately so that their effects can be estimated efficiently. Factors may be continuous, such as cutting speed, temperature, feed rate, pressure, or holding time, or categorical, such as tool material, coolant type, furnace atmosphere, or machine.

Responses should be operationally defined. Examples include surface roughness R_a, dimensional error, hardness, thermal efficiency, cycle time, energy per acceptable part, tool wear, defect rate, and carbon-dioxide-equivalent emissions. A response definition should specify the instrument, sampling location, time window, unit, and acceptance rule.

Experimental design choices

A full factorial design studies every combination of factor levels. With k two-level factors, the number of treatment combinations is 2^k. This becomes expensive as k increases. A fractional factorial design reduces runs but may confound, or alias, effects. Screening designs identify influential factors; response-surface designs then model curvature and locate an operating region.

Randomisation reduces the risk that time-related disturbances are mistaken for factor effects. Blocking controls known nuisance variation, such as different batches of material, days, operators, or machine warm-up states. Replication estimates experimental variability and distinguishes a real effect from random fluctuation. Replicating only measurement readings does not replace independent process replication when the process itself varies.

A simple additive model for a response is:

Here, y is the response, x₁ and x₂ are coded factor values, β₀ is the intercept, β₁ and β₂ are main effects, β₁₂ is the interaction coefficient, and ε is unexplained error. The model assumes that the chosen functional form is adequate over the investigated region, observations are appropriately independent, and residual behaviour is sufficiently compatible with the analysis method. Coded factors are dimensionless; engineering predictions must be converted back to physical units.

An interaction means that the effect of one factor depends on another. For example, increased cutting speed may reduce cycle time at low feed but sharply increase tool wear at high feed. Ignoring interactions can produce an apparently attractive but unsafe recommendation.

Measurement quality and uncertainty

Before experimentation, instruments should be calibrated or checked against a traceable reference. Resolution is not the same as accuracy. A temperature sensor may display 0.1 °C while having a much larger systematic error.

For independent input quantities x_i, a first-order combined standard uncertainty is commonly estimated as:

where y = f(x₁, x₂, …) is the measurand, u(x_i) is the standard uncertainty of input x_i, and ∂f/∂x_i is the sensitivity coefficient. This propagation assumes small uncertainties, a sufficiently smooth model, and negligible covariance. If inputs are correlated, covariance terms must be included.

For example, if a power meter reads P = 2.00 ± 0.05 kW and a process runs for t = 3.00 ± 0.02 h, energy is E = Pt = 6.00 kWh. Assuming independence, the relative standard uncertainty is approximately:

Thus, u(E)/E = √[(0.05/2.00)² + (0.02/3.00)²] ≈ 0.026, giving u(E) ≈ 0.16 kWh. The result should not be reported with unjustified digits.

Thermal and manufacturing performance measures

Useful general measures include:

Efficiency η is dimensionless and should be clearly defined; thermal efficiency, for example, may use the enthalpy increase of a workpiece as useful output, while a plant-level measure may include auxiliary equipment.

The functional output should be acceptable parts, not merely total parts, when defects are significant. If a process consumes 12 kWh and produces 20 parts, of which 2 are rejected, its specific energy per acceptable part is 12/18 = 0.667 kWh per acceptable part.

For a batch, a simple first-law energy balance is:

The terms must refer to the same time interval and system boundary. In practice, unmeasured losses, standby consumption, and sensor limitations may cause the measured terms not to close exactly. The imbalance should be reported, not silently assigned to a convenient category.

Quality measures include yield, defect rate, capability, dimensional error, roughness, hardness, and reliability. If n units are produced and d are defective, the observed defect proportion is:

This estimate is meaningful only when the sampling process and defect definition are stated. A small sample can make a low observed defect rate statistically uncertain.

4. Sustainability and life-cycle thinking

Sustainability analysis should address environmental, economic, and social dimensions. It should not equate sustainability with energy reduction alone.

Energy and emissions

For electricity consumption, emissions may be estimated as:

where E is electricity use in kWh, EF is the applicable emission factor in kg CO₂e/kWh, and m_CO₂e is emissions in kg CO₂e. The factor must be geographically and temporally appropriate, and its source must be cited. If fuels, refrigerants, or purchased materials are included, their contributions must be added using consistent boundaries.

A product comparison should use emissions per functional unit:

This prevents a process with low emissions per cycle but poor yield from appearing sustainable.

Waste and circularity

Waste assessment distinguishes prevention, reduction, reuse, remanufacture, recycling, recovery, and disposal. Recycling is not automatically preferable if it requires large energy inputs or causes quality loss. Circular design seeks to retain material and product value for as long as possible.

Relevant indicators include material yield, recycled content, reclaim rate, repairability, component lifetime, and waste intensity. A manufacturing team should ask whether chips can be segregated by alloy, whether cutting fluid can be filtered and reused, whether heat can be recovered, and whether components can be remanufactured rather than discarded.

Life-cycle assessment

A simplified life-cycle inventory records resource and emission flows across defined stages: raw material, manufacture, distribution, use, maintenance, and end-of-life. Life-cycle impact assessment converts flows into impact categories such as climate change, water use, acidification, or resource depletion. A laboratory project may not justify a complete certified assessment, but it should state exclusions, use transparent data sources, and avoid presenting screening estimates as precise facts.

The functional unit is crucial. Comparing a lightweight component with a heavier one requires equal service, lifetime, safety, and performance—not merely equal mass. Boundary changes and uncertain data should be tested through sensitivity analysis.

5. Multi-criteria decision-making

Industrial alternatives commonly conflict. One setting may minimise energy, another may maximise quality, and a third may reduce cost or improve resilience. MCDM makes priorities visible rather than hiding them in an informal judgement.

The process is:

1. Define alternatives and mandatory constraints.
2. Select criteria linked to the project decision.
3. Classify each criterion as beneficial or non-beneficial.
4. Measure or estimate criterion values with units and uncertainty.
5. Normalise values so unlike units can be compared.
6. Assign weights using a documented method.
7. Calculate scores and rank feasible alternatives.
8. Perform sensitivity and robustness analysis.

For a beneficial criterion, min–max normalisation can be written as:

For a non-beneficial criterion:

Here, x_ij is the observed value of alternative i on criterion j, and r_ij lies between 0 and 1 when the denominator is nonzero. If all alternatives have the same value for a criterion, that criterion cannot discriminate and should be handled explicitly rather than divided by zero.

A weighted-sum score is:

where w_j is the weight of criterion j, with w_j ≥ 0 and ∑w_j = 1. This method assumes that criteria have been made commensurable and that additive trade-offs are acceptable. It is inappropriate to use a high score to compensate for violation of a safety or legal requirement; mandatory constraints must be screened first.

Weights can be obtained through stakeholder discussion, direct rating, pairwise comparison, or an analytic hierarchy process. No weighting method removes value judgement. Therefore, the report should show who supplied weights, what assumptions were made, and how rankings change when weights vary.

A robust decision is not necessarily the alternative ranked first under one exact set of assumptions. It is one whose feasibility and relative advantage persist across plausible changes in weights, data, and operating conditions.

6. Reliability, risk, resilience, and regulation

Reliability is the probability that an item performs its required function for a specified time under stated conditions. For a constant failure rate λ, the exponential reliability model is:

This model assumes a constant hazard rate and is most suitable for a limited period during the useful-life phase; it is not a universal model for infant mortality or wear-out. If the failure rate is not constant, a more general reliability expression is required based on the cumulative hazard.

Risk assessment combines the possibility of an event with its consequence. A risk matrix may rank hazards, but numerical scores can create false precision. Engineering teams should identify hazards, existing controls, additional controls, responsible persons, and verification evidence. Hierarchy-of-control principles favour elimination and engineering controls over warnings or personal protective equipment alone.

Resilience concerns the ability to absorb disruption, maintain critical function, and recover. A process may be efficient under normal conditions but fragile because it depends on one supplier, one specialist, one machine, or unstable energy availability. Resilience indicators include recovery time, capacity redundancy, critical-spare availability, supplier diversity, data backup, and the ability to operate safely in degraded modes.

Regulatory obligations may concern machine guarding, electrical safety, pressure systems, thermal exposure, chemicals, waste handling, emissions, worker exposure, product conformity, and environmental reporting. Requirements depend on jurisdiction and industry. Students must consult current institutional and legal sources rather than assume that a laboratory procedure is automatically compliant. A recommendation that violates a permit, standard, or safety limit is not an acceptable optimisation.

7. Digital tools and artificial intelligence

Digital sensors, programmable logic controllers, machine vision, manufacturing execution systems, digital twins, and cloud dashboards can improve traceability and control. However, more data does not guarantee better evidence. Sensor drift, timestamp errors, missing data, cyberattack, incompatible formats, and poor data ownership can undermine a digital project.

Artificial intelligence may help classify defects, detect anomalous energy consumption, estimate tool wear, or search experimental designs. It should support—not replace—engineering judgement. Training data must represent the intended operating conditions, and model performance should be evaluated on independent data. A model should not be used to make safety-critical recommendations without validation, explainable acceptance criteria, human oversight, and a fallback procedure.

Data governance should address access control, privacy, intellectual property, audit trails, retention, cybersecurity, and responsible use. Students should distinguish measured data, calculated data, assumptions, and model predictions in every report.

8. Illustrative case: selecting a low-energy machining route

Background and problem

The following is an illustrative case. A workshop must choose among three operating conditions for machining an aluminium component. The component must meet dimensional and surface-quality requirements. The team measures energy per cycle, acceptable-part yield, cycle time, tool wear, and estimated emissions. Safety and dimensional conformity are treated as mandatory constraints.

Experimental approach

A pilot study identifies cutting speed and feed rate as influential factors. A two-factor design with replication is used, while material batch is blocked. Power is recorded at a fixed sampling interval, tool wear is measured after each batch, and surface roughness is measured at three specified locations. The team randomises run order within each block and performs a confirmation run at the selected setting.

Three feasible alternatives remain after screening:

The table demonstrates why single-metric optimisation is unsafe. Alternative C has the lowest cycle energy and shortest time, but its poor yield and higher tool cost reduce its industrial attractiveness.

Decision analysis

The team chooses four criteria: yield (beneficial), cycle time (non-beneficial), tool cost (non-beneficial), and emissions per acceptable part (non-beneficial). It assigns weights of 0.35, 0.20, 0.20, and 0.25, respectively. Dimensional conformity has already been screened as a constraint.

Using min–max normalisation, the scores are approximately:

For example, the normalised time score for B is (8.0−6.5)/(8.0−5.4) ≈ 0.577. The weighted score for B is approximately 0.35×0.857 + 0.20×0.577 + 0.20×1.000 + 0.25×0.444 = 0.736. B is recommended for further confirmation.

Outcome and implementation

The confirmation run checks that B maintains yield and roughness when operated by another trained operator and after normal machine warm-up. The team recommends B, together with segregated chip collection, tool-life monitoring, standby shutdown, and a control chart for yield and roughness. The recommendation is conditional: if electricity prices, emission factors, or the value of production time change substantially, the decision should be revisited.

Learning points

The case illustrates five principles: constraints precede scoring; acceptable-part output is more meaningful than total output; efficiency and sustainability require a common functional unit; weights express priorities rather than objective truth; and implementation requires monitoring, not merely a final ranking.

9. Reporting and professional communication

A strong report connects evidence to action. Its structure may include an executive decision statement, background, charter and boundary, risk assessment, method, measurement system, DOE, results, uncertainty, sustainability inventory, MCDM assumptions, sensitivity analysis, recommendation, implementation plan, limitations, and appendices containing raw data.

Graphs should show units, sample size, uncertainty or variation where appropriate, and acceptance limits. A regression line is not proof of causation unless the design supports causal interpretation. Outliers must be investigated and documented; they should not be deleted merely because they weaken the preferred conclusion.

Recommendations should be specific: identify the operating window, responsible owner, required control, expected benefit, capital or training requirement, and verification metric. A useful implementation plan includes a baseline, target, deadline, pilot, review point, and escalation rule.

Challenges and limitations

Industrial laboratory projects face limited sample sizes, measurement error, uncontrolled environmental conditions, operator effects, equipment constraints, and incomplete life-cycle data. DOE models may be valid only within the tested region. Normalisation can exaggerate small differences, while weighted sums can conceal disagreement among stakeholders. Screening-level emissions estimates may depend strongly on emission factors and excluded stages.

Teams should therefore avoid unsupported generalisation. State confidence and practical significance separately. A statistically detectable improvement may be too small to justify implementation; a practically important improvement may require more data to establish statistical certainty. Where evidence is weak, recommend a controlled pilot rather than claiming final optimisation.

Best-practice checklist

• Define the functional unit and boundary before collecting data.
• Treat safety, legal, and product requirements as constraints.
• Calibrate instruments and document uncertainty.
• Use randomisation, blocking, replication, and confirmation runs.
• Preserve raw data and maintain an auditable analysis trail.
• Evaluate acceptable output, not only nominal production.
• Include energy, material, waste, emissions, cost, reliability, and resilience.
• Separate measured facts, assumptions, and predictions.
• Test rankings under plausible changes in weights and data.
• Design controls and monitoring for implementation.
• Report limitations and conflicts of interest honestly.

Chapter Summary

An integrated industrial project transforms laboratory measurements into a defensible engineering decision. The process begins with a clear charter, functional unit, system boundary, constraints, and measurable responses. DOE provides an efficient structure for studying factors, interactions, variability, and operating regions, while calibration and uncertainty analysis establish the credibility of measurements. Sustainable evaluation includes energy, emissions, waste, circularity, cost, quality, reliability, resilience, and life-cycle impacts.

MCDM enables alternatives with conflicting advantages to be compared transparently. Normalisation and weighted scoring are useful only when assumptions, weights, constraints, and sensitivity are explicit. Safety and regulatory requirements cannot be traded away for a higher numerical score. Digital technologies and AI can improve monitoring and prediction, but require validated models, secure data, human oversight, and clear governance. The final professional output is not simply a ranked table; it is an implementable recommendation supported by evidence, limitations, verification measures, and continuous improvement.

Key Terms

System boundary: Defined set of processes and flows included in an analysis.

Functional unit: Quantified basis used to compare alternatives providing the same service.

Design of experiments (DOE): Structured method for varying factors and estimating their effects.

Factor: Controlled input variable in an experiment.

Response: Measured output used to evaluate a process.

Interaction: Condition in which the effect of one factor depends on another factor.

Randomisation: Allocation of experimental runs to reduce bias from time-related effects.

Blocking: Grouping experimental runs to control a known nuisance source.

Replication: Repetition of independent experimental units to estimate variability.

Calibration: Comparison of an instrument with a traceable reference.

Uncertainty: Quantified doubt associated with a measurement result.

Specific energy: Energy consumed per defined functional output.

Life-cycle assessment: Evaluation of resource use and impacts across defined life-cycle stages.

Circularity: Retention of product and material value through reuse, repair, remanufacture, or recycling.

Multi-criteria decision-making: Structured comparison of alternatives using several criteria.

Normalisation: Transformation of criterion values to a common scale.

Reliability: Probability that an item performs its required function for a specified time and conditions.

Resilience: Ability of a system to withstand disruption, maintain function, and recover.

Functional constraint: Requirement that must be satisfied before alternatives are scored.

Data governance: Rules for data quality, access, security, ownership, and responsible use.

Review Questions

5 Short Answer Questions

1. Why must a functional unit be defined before comparing sustainable process alternatives?
2. Distinguish replication from repeated measurement readings.
3. What is an interaction in a DOE model?
4. Why should safety and regulatory requirements be treated as constraints?
5. State two limitations of a weighted-sum MCDM model.

5 Long Answer Questions

1. Explain how a team should plan and govern an integrated thermal or manufacturing laboratory project from charter definition to confirmation run.
2. Describe how calibration, uncertainty propagation, randomisation, blocking, and replication improve the credibility of experimental results.
3. Explain how energy, emissions, waste, quality, cost, reliability, and life-cycle impacts can be combined without hiding important trade-offs.
4. Develop a suitable DOE and analysis plan for reducing energy consumption in a heat-treatment process while maintaining hardness and dimensional quality.
5. Discuss how lean, circular, resilient, and digitally enabled operations can be incorporated into an industrial recommendation.

3 Analytical / Application Questions

1. A process uses 15 kWh to produce 25 parts, of which 3 are rejected. Calculate the specific energy per acceptable part and explain why this differs from energy per total part.
2. Construct a normalised weighted-sum decision model for three manufacturing alternatives using yield, cycle time, cost, and emissions. State the weights, justify them, and perform a sensitivity analysis.
3. A machine-learning model predicts tool failure from sensor data but was trained only during warm, stable operation. Evaluate the risks of deploying it during cold start-up and propose a validation and governance plan.

Multiple Choice Questions with Answers

Question 1

Which project statement is the most suitable industrial decision question?

A. Measure the furnace temperature.

B. Select a process condition that meets hardness requirements while reducing energy per acceptable part.

C. Collect as many readings as possible.

D. Choose the setting with the highest temperature.

Correct Answer: B. Select a process condition that meets hardness requirements while reducing energy per acceptable part.

Question 2

What is the primary purpose of randomising experimental runs?

A. To increase instrument resolution.

B. To eliminate the need for replication.

C. To reduce bias from time-related or uncontrolled effects.

D. To guarantee a normal response distribution.

Correct Answer: C. To reduce bias from time-related or uncontrolled effects.

Question 3

In a DOE model, an interaction indicates that:

A. Two instruments have identical calibration errors.

B. The effect of one factor depends on the level of another factor.

C. The response has no experimental variation.

D. All factors must be categorical.

Correct Answer: B. The effect of one factor depends on the level of another factor.

Question 4

A process consumes 12 kWh and produces 18 acceptable parts. What is its energy per acceptable part?

A. 0.50 kWh per acceptable part

B. 0.67 kWh per acceptable part

C. 1.50 kWh per acceptable part

D. 30 kWh per acceptable part

Correct Answer: B. 0.67 kWh per acceptable part

Question 5

In the expression m_CO₂e = E × EF, EF represents:

A. Equipment failure frequency

B. Energy efficiency

C. An emissions factor

D. Experimental fluctuation

Correct Answer: C. An emissions factor

Question 6

Before applying a weighted-sum MCDM score, which item should normally be screened as a mandatory constraint?

A. Aesthetic preference

B. Legal safety compliance

C. Stakeholder preference for a colour

D. A minor difference in cycle time

Correct Answer: B. Legal safety compliance

Question 7

For a non-beneficial criterion in min–max normalisation, the best alternative receives:

A. A score of 0

B. A score of 0.5 only

C. A score of 1

D. A negative score

Correct Answer: C. A score of 1

Question 8

Why is yield important when calculating energy per functional output?

A. Rejected parts consume no energy.

B. It accounts for energy associated with parts that do not provide acceptable output.

C. It always increases the measured cycle time.

D. It replaces the need for an energy meter.

Correct Answer: B. It accounts for energy associated with parts that do not provide acceptable output.

Question 9

The exponential reliability model R(t) = e^(−λt) assumes:

A. A constant failure rate over the period considered.

B. No possibility of failure.

C. A failure rate that always increases with time.

D. Only mechanical, not thermal, systems.

Correct Answer: A. A constant failure rate over the period considered.

Question 10

Which is the most appropriate use of AI in an industrial laboratory project?

A. Allowing an unvalidated model to override safety controls.

B. Using a validated defect-classification model with independent testing and human oversight.

C. Deleting data that disagree with model predictions.

D. Using training accuracy as proof of field reliability.

Correct Answer: B. Using a validated defect-classification model with independent testing and human oversight.