Digital Quality, AI and Future Developments

Introduction

Quality management is increasingly becoming a digital, connected and intelligent activity. Traditional quality systems relied heavily on paper records, periodic inspection, operator experience and retrospective analysis. Contemporary production systems can collect process data continuously, connect machines and people in real time, and use analytical models to identify quality risks before defects occur. This development is often described through terms such as Industry 4.0, smart manufacturing, cyber-physical production systems and digital quality management.

Digitalisation does not automatically create quality improvement. A poorly designed digital system can accelerate incorrect decisions, reproduce biased inspection practices, expose confidential production data or create additional administrative work. The central management challenge is therefore not simply to purchase software or install sensors. It is to design an integrated socio-technical quality system in which technology, processes, data, governance and employee capability reinforce one another.

This chapter evaluates quality management software, manufacturing execution systems (MES), industrial Internet of Things (IIoT) sensing, digital twins, machine vision, predictive quality, AI-assisted analytics and automation. It also considers cybersecurity, data governance, algorithmic bias, workforce capability and the emerging requirements of smart factories.

Learning Objectives

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

1. Explain the role of digital technologies in contemporary quality management.
2. Distinguish quality management software, MES, IIoT platforms and digital twins.
3. Evaluate machine-vision and AI-assisted inspection systems.
4. Explain predictive quality and the importance of data quality.
5. Assess cybersecurity, privacy, governance and algorithmic-bias risks.
6. Recommend implementation practices for a digitally enabled quality system.
7. Analyse the workforce and organisational capabilities required by smart factories.

1. From Digital Records to Digital Quality

Digital quality has developed through several overlapping stages.

Digitisation converts analogue information into digital form, such as scanning inspection sheets or storing calibration certificates electronically. Digitalisation uses digital information to improve a process, for example by automatically routing a nonconformance report to the responsible engineer. Digital transformation redesigns the operating model around connected data, automation, analytics and integrated decision-making.

A useful quality-information chain is:

Sense → connect → contextualise → analyse → decide → act → learn.

A temperature sensor may sense a process value. A network transmits it to a data platform. Contextualisation links the value to a machine, product, recipe, operator, time and lot. Analytics determines whether the value is normal or associated with future defects. A decision may adjust a process or stop production. The result is recorded and used to improve the model or process.

Digital quality supports the movement from:

• detection to prevention;
• sampling to continuous monitoring;
• isolated records to traceable product genealogy;
• expert-only interpretation to evidence-supported decisions;
• end-of-line inspection to closed-loop process control.

However, digital systems do not replace statistical thinking, process knowledge or customer requirements. A sophisticated dashboard cannot compensate for an unstable process, an invalid measurement system or a poorly defined specification.

2. The Digital Quality Architecture

A typical architecture contains several layers.

The architecture should be designed around business and quality decisions rather than around individual technologies. For example, the question “Can the process produce conforming product?” may require sensor data, measurement-system information, material genealogy, maintenance records and laboratory results. A single application rarely provides all of these reliably without integration.

Interoperability and system integration

Interoperability is the ability of systems, devices and people to exchange information and use it meaningfully. Syntactic interoperability concerns whether data can be exchanged in a common format. Semantic interoperability concerns whether the receiving system understands the meaning of the data. For instance, two systems may both use the field “temperature,” while one records degrees Celsius and the other degrees Fahrenheit.

Important integration principles include:

• use consistent equipment, product, defect and material identifiers;
• define units, time zones, precision and permissible values;
• preserve source data and transformation history;
• use standard interfaces where practical;
• document ownership of master data;
• test data exchange with realistic failure conditions.

3. Quality Management Software and MES

Quality management software

Quality management software (QMS) supports activities such as document control, training records, audits, corrective and preventive action (CAPA), nonconformance management, supplier quality and change control. Its principal value is controlled workflow and evidence. A QMS can ensure that an investigation has an assigned owner, required approval and a traceable closure record.

QMS selection should be based on process requirements rather than feature volume. Evaluation criteria include:

• support for the organisation’s quality processes;
• electronic signatures and audit trails;
• configuration flexibility without uncontrolled customisation;
• integration with ERP, MES, laboratory and maintenance systems;
• role-based access control;
• reporting and data export;
• validation and change-management support;
• availability, scalability and total cost of ownership.

A QMS is not necessarily a real-time production-control system. It may manage the investigation of a defect while the MES manages the execution and genealogy of the production order.

Manufacturing execution systems

An MES coordinates production activities between enterprise planning and shop-floor control. Typical functions include production dispatching, work instructions, labour and equipment tracking, electronic batch records, traceability, downtime recording, quality checks and performance reporting.

The quality value of an MES arises from contextualisation. Instead of storing an isolated measurement of 82, it can associate the measurement with a particular product, operation, machine, tool, operator, material lot and timestamp. This supports containment and root-cause analysis.

MES implementation risks include excessive customisation, weak master data, unclear ownership between production and quality, inadequate operator usability and unreliable interfaces. If operators regard the system as a bureaucratic burden, records may be delayed, copied or entered inaccurately.

QMS and MES comparison

The two systems should form a complementary control environment, not competing databases.

4. IIoT Sensing and Data Quality

The industrial Internet of Things connects physical assets, sensors, controllers, software and analytical services. Sensors can measure variables such as temperature, pressure, force, flow, current, vibration, humidity, position and acoustic emissions.

A sensor programme should begin with a decision. The organisation should ask: What quality risk is being monitored? What physical variable indicates that risk? What sampling rate, accuracy, range and environmental protection are needed? What action follows an abnormal value?

Measurement-system considerations

Digital data remains measurement data and is subject to the same principles as conventional inspection. Relevant characteristics include:

• accuracy and bias;
• repeatability and reproducibility;
• resolution;
• stability over time;
• calibration status;
• sampling frequency;
• sensor drift;
• missing and corrupted values.

A high-frequency sensor with poor calibration can produce a large quantity of misleading information. Measurement-system analysis and calibration control therefore remain essential.

Edge and cloud computing

Edge computing processes data close to the machine. It is useful where latency, bandwidth, availability or data confidentiality is important. Cloud computing provides scalable storage and analytical capacity, but may introduce connectivity, jurisdiction and cybersecurity concerns. A hybrid architecture often places immediate safety and control functions at the edge while sending selected, governed data to central platforms.

5. Digital Twins

A digital twin is a digital representation of a physical asset, process or system that is connected to the physical entity through data and can support analysis, prediction or decision-making. It differs from a static computer-aided design model because it is linked to operational reality and may be updated throughout the asset’s life.

Examples include:

• a twin of a machine spindle used to estimate tool wear;
• a process twin used to test recipe changes;
• a product twin containing configuration, inspection and service history;
• a factory twin used to examine capacity and material-flow alternatives.

A digital twin may be descriptive, diagnostic, predictive or prescriptive. A descriptive twin reports current status. A diagnostic twin helps explain a problem. A predictive twin estimates future behaviour. A prescriptive twin compares possible actions.

The usefulness of a twin depends on model validity, data quality, update frequency and decision integration. A detailed simulation that does not represent actual machine behaviour may create false confidence. Digital twins should therefore be verified against observed data and periodically revalidated when equipment, materials or operating conditions change.

6. Machine Vision and Automated Inspection

Machine vision uses cameras, lighting, optics, image-processing methods and decision rules to inspect products or processes. It is particularly valuable for high-speed, repetitive, contactless or hazardous inspection.

A typical system includes:

1. controlled lighting and product presentation;
2. image acquisition;
3. preprocessing, such as noise reduction or geometric correction;
4. feature extraction or learned representation;
5. classification, measurement or defect segmentation;
6. decision logic and disposition;
7. storage of images and inspection results.

Traditional vision systems use explicit rules, such as edge detection, dimensional thresholds or pattern matching. AI-based systems may learn defect features from labelled images. Neither approach is universally superior. Rule-based methods can be transparent and effective for stable geometries; machine learning can handle complex visual variation but requires representative training data.

Inspection performance should be evaluated using a confusion matrix:

The organisation must define the coding convention clearly because terms such as “positive” may refer either to defect detection or product acceptance. Important measures include precision, recall, false-accept rate and false-reject rate. In quality applications, a false acceptance may be more costly than a false rejection, but excessive false rejection also creates waste and undermines trust.

The inspection system must be validated across lighting changes, product variants, acceptable variation, realistic defect types and borderline conditions. A model trained only on clean laboratory images may fail in production.

7. Predictive Quality and AI-Assisted Analytics

Predictive quality uses historical and current process information to estimate the probability of a future quality event. Possible targets include defect occurrence, dimensional deviation, warranty failure, tool replacement need or laboratory-test failure.

A general predictive workflow is:

1. define the quality outcome and decision threshold;
2. collect and understand data;
3. clean, align and label records;
4. divide data into training, validation and test sets;
5. train a model;
6. evaluate it against appropriate performance and cost measures;
7. deploy with human oversight;
8. monitor drift, outcomes and unintended effects.

For a binary quality outcome, a logistic model may estimate the probability of a defect:

Here, Y = 1 denotes the defined event, such as a defect; p is its estimated probability; x₁, x₂, …, xₖ are predictor variables; β₀ is the intercept; βᵢ are model coefficients; and exp is the exponential function. The model assumes that the log-odds of the event are a linear combination of the predictors. It is appropriate only when the data structure, independence assumptions and relationship form are reasonably suitable; other models may be needed for time-dependent, hierarchical or nonlinear processes.

A model’s discrimination can be assessed using a confusion matrix, receiver operating characteristic analysis or precision–recall analysis. However, statistical performance is not the same as operational value. A model that predicts defects accurately but gives warnings too late may have little practical benefit.

Predictive versus causal analysis

Prediction asks, “What is likely to happen?” Causal analysis asks, “What intervention will change the outcome?” A variable may be highly predictive without being a useful cause. For example, a late-stage inspection result may predict final failure, but changing that result does not prevent failure. Quality engineers must combine analytics with process knowledge and designed experimentation before changing process settings.

Model drift

Model drift occurs when the relationship between inputs and outcomes changes. Causes include new materials, equipment ageing, product redesign, seasonal conditions, operator practices and changes in defect labelling. Monitoring should include data completeness, input distributions, prediction distributions, actual outcomes and performance by product family or production line.

8. Automation and Human Decision-Making

Automation can improve repeatability, speed, ergonomics, traceability and response time. Examples include automatic parameter adjustment, robotic handling, electronic approvals, automated quarantine and closed-loop control.

Automation should be classified by consequence and reversibility. A recommendation that an engineer can review is lower risk than an irreversible automatic release of safety-critical product. Appropriate controls include:

• defined authority levels;
• approval thresholds;
• safe-state behaviour;
• manual override;
• exception handling;
• event logging;
• periodic review;
• tested recovery procedures.

Human-in-the-loop systems retain human approval. Human-on-the-loop systems allow automation to act while a person supervises. Fully autonomous systems make decisions without routine human intervention. The appropriate arrangement depends on risk, regulation, process stability and the quality of available evidence.

9. Cybersecurity and Data Governance

Connected quality systems expand the attack surface. Threats include stolen credentials, ransomware, unauthorised parameter changes, manipulated sensor values, intellectual-property theft, denial of service and compromise of supplier connections.

A practical defence-in-depth approach includes:

• asset inventory and risk classification;
• network segmentation between enterprise, operations and control environments;
• least-privilege access;
• multifactor authentication where feasible;
• secure configuration and patch management;
• encryption in transit and at rest where appropriate;
• backups and tested recovery;
• monitoring and incident response;
• supplier cybersecurity requirements;
• secure software development and vulnerability management.

Safety and cybersecurity are connected. An attacker who changes a process setpoint may create both a security incident and a physical quality or safety hazard.

Data governance

Data governance assigns responsibility for data quality, access, retention, meaning and use. A governance framework should define:

• data owners and custodians;
• master-data standards;
• metadata and lineage;
• access and retention rules;
• correction and amendment procedures;
• auditability;
• privacy and confidentiality requirements;
• model and algorithm approval.

Data integrity requires that records are attributable, legible, contemporaneous, original or a verified copy, accurate, complete, consistent, enduring and available. Digital systems should preserve audit trails rather than allowing silent overwriting.

10. Algorithmic Bias, Explainability and Responsible AI

Algorithmic bias occurs when a system produces systematically poorer or unfairer outcomes for particular products, operators, suppliers, shifts, locations or defect categories. In manufacturing, bias may result from unrepresentative training images, inconsistent labels, class imbalance, sensor placement, historical inspection practices or proxy variables.

Risk controls include:

• representative sampling across products, shifts and operating conditions;
• clear labelling rules;
• independent test data;
• subgroup performance analysis;
• review of false positives and false negatives;
• explanation of important features or evidence;
• human escalation for uncertain cases;
• documented model limitations;
• periodic revalidation.

Explainability does not mean that every model must be simple. It means that users should understand the model’s purpose, inputs, limitations, confidence and appropriate use. No AI model should be treated as an unquestionable authority, especially where release, safety, employment or supplier decisions are affected.

11. Workforce Capability and Organisational Change

Digital quality requires blended capability. Operators need to understand digital work instructions, data-entry discipline, alarms and escalation. Quality professionals need competence in process analysis, measurement systems, statistics, data governance and model interpretation. Engineers need knowledge of automation, interfaces, cybersecurity and change control. Managers need to understand investment, risk, workforce development and ethical accountability.

Important capabilities include:

• data literacy and visualisation;
• statistical process control;
• measurement and calibration;
• problem-solving and root-cause analysis;
• system integration;
• cybersecurity awareness;
• AI and model-risk literacy;
• communication across technical and operational roles.

Training should be supported by redesigned jobs and usable interfaces. If a digital system increases data-entry effort without improving work, employees may develop workarounds. Participation by operators during design and pilot stages is therefore a quality-control mechanism, not merely a change-management courtesy.

12. Implementation Framework for Digital Quality

A disciplined implementation can proceed through the following stages.

Stage 1: Define value and risk

Identify customer-critical characteristics, major process risks, regulatory obligations and decisions that require better information. Establish measurable outcomes such as reduced escapes, shorter investigation time or improved traceability.

Stage 2: Assess the current state

Map processes, information flows, systems, manual workarounds, data defects and capability gaps. Determine whether the measurement system is adequate before automating collection.

Stage 3: Establish architecture and governance

Define system boundaries, interfaces, ownership, cybersecurity controls, retention rules and model-approval responsibilities.

Stage 4: Pilot a high-value use case

Select a manageable application with a clear baseline and feedback loop. Examples include predictive tool wear, automated inspection of a stable product or electronic nonconformance containment.

Stage 5: Validate and quantify performance

Compare the digital intervention with the existing method. Evaluate technical accuracy, process capability, false alarms, user acceptance, response time and economic impact.

Stage 6: Scale responsibly

Standardise successful practices, integrate systems, train personnel, monitor drift and review controls. Scaling should not mean copying an unvalidated model across different processes.

A simple financial evaluation may use return on investment:

Annual benefits may include avoided scrap, reduced inspection labour, lower warranty cost and reduced downtime. Annual costs include licences, sensors, integration, cybersecurity, training, maintenance and model monitoring. The calculation should state the time period, include relevant recurring costs and avoid claiming benefits that cannot be evidenced.

13. Illustrative Case: Predictive Quality in a Heat-Treatment Line

Background. An illustrative automotive-component manufacturer heat-treats steel parts in several furnaces. Final hardness testing identifies failures after substantial processing and handling have already occurred.

Problem. The plant experiences variable hardness results, repeated laboratory testing and occasional customer complaints. Operators record furnace temperatures manually at intervals, while the MES stores production orders separately from the laboratory system.

Analysis. A cross-functional team maps the process and discovers that temperature sensor calibration records, furnace recipes, material lots and hardness results use inconsistent identifiers. Before applying AI, the team standardises product and lot codes, verifies sensor calibration and links laboratory results to furnace cycles. The team also identifies that hardness failures are rare, making accuracy alone an unsuitable model-performance measure.

Approach and solution. Additional sensors record furnace-zone temperature and atmosphere conditions. The data is contextualised by furnace, recipe, material lot, cycle and time. A predictive model estimates the risk of hardness failure before unloading. A warning prompts the operator and engineer to review the cycle; automatic release remains prohibited. The model is tested on later production data rather than only on records used for development. Thresholds are selected with regard to the relative cost of an escaped defect and an unnecessary investigation.

The system includes an audit trail, role-based access, calibration alerts, manual override and a monthly review of false alarms and missed failures. Changes to steel grade or furnace configuration trigger model review.

Outcome. In this illustrative case, early warnings reduce late-stage testing and improve containment. The greatest benefit comes not from the algorithm alone but from data integration, measurement discipline and a defined response process.

Learning points.

1. Data preparation and process understanding precede model selection.
2. Predictive quality must be connected to a timely intervention.
3. Rare-event performance requires more than overall accuracy.
4. Human review is appropriate when the consequence of an incorrect release is high.
5. Governance, calibration and cybersecurity are part of quality performance.

14. Future Developments and Smart-Factory Requirements

Future smart factories are likely to combine interoperable equipment, autonomous material flow, advanced robotics, edge AI, digital twins, additive manufacturing, augmented-reality support and increasingly capable generative AI assistants. These developments may enable faster engineering changes, adaptive inspection and more personalised production.

Several requirements will become more important:

• trustworthy data spaces spanning internal and supply-chain partners;
• machine-readable product and process specifications;
• resilient and secure industrial networks;
• explainable and monitored AI;
• lifecycle traceability from design through service;
• energy and carbon data integrated with quality decisions;
• flexible automation for high-mix production;
• digital skills and continuous reskilling;
• standards-based interoperability;
• responsible treatment of worker and supplier data.

Generative AI may assist with searching procedures, summarising nonconformances, drafting investigation questions or querying maintenance knowledge. It should not be permitted to invent evidence, silently alter controlled documents or approve product without authorised verification. Outputs require source checking, access control and clear accountability.

Sustainability is also becoming part of digital quality. Energy, water, emissions and material-use data can be linked to product conformity and process capability. Nevertheless, sustainability metrics must be measured with the same discipline as quality metrics; an attractive dashboard is not evidence of improvement.

Challenges and Limitations

Digital quality programmes commonly face the following limitations:

• poor or inconsistent historical data;
• uncertain return on investment;
• legacy systems that cannot interoperate;
• cybersecurity vulnerabilities;
• vendor lock-in;
• model drift and changing processes;
• excessive alarms and dashboard fatigue;
• lack of explainability;
• insufficient employee capability;
• privacy and labour-relations concerns;
• dependence on connectivity and power;
• confusion between correlation and causation.

Technology should therefore be introduced incrementally, with explicit quality objectives, validated measurements and tested fallback procedures. The best digital system is not the one with the most functions; it is the one that improves decisions while preserving safety, traceability and professional accountability.

Best-Practice Checklist

Before approving a digital quality application, ask:

1. What quality decision or risk does it address?
2. Are the measurement method and data definitions valid?
3. Who owns the data and who is accountable for the decision?
4. Can the system interoperate with existing QMS, MES and control systems?
5. How are cybersecurity, access and recovery controlled?
6. What are the consequences of false alarms and missed defects?
7. How will model performance and drift be monitored?
8. What human review, override and escalation are required?
9. Are workers trained and involved in design?
10. What evidence demonstrates sustained improvement?

Chapter Summary

Digital quality links sensing, connectivity, contextualised data, analytics, decisions and learning. Quality management software governs documents, audits, CAPA and compliance, while MES coordinates real-time production execution and genealogy. IIoT sensors provide continuous information, but sensor accuracy, calibration and data governance remain essential. Digital twins represent physical assets or processes and can support diagnosis, prediction and scenario analysis when validated against reality.

Machine vision automates repetitive inspection, but systems must be evaluated for false acceptance, false rejection, robustness and representative conditions. Predictive quality estimates future quality outcomes and is valuable only when warnings lead to timely, controlled action. AI can support analytics but does not replace causal reasoning, process expertise or accountability.

Cybersecurity, data integrity, privacy, algorithmic bias and explainability are central quality concerns. Smart factories require interoperable architectures, resilient infrastructure, capable workers, lifecycle governance and continuous model monitoring. Successful implementation begins with a clear business problem, reliable measurements, a controlled pilot and evidence-based scaling.

Key Terms

Digital quality: Use of connected digital technologies to govern, monitor and improve quality activities.

Digitisation: Conversion of analogue information into digital form.

Digital transformation: Redesign of organisational processes and decisions around digital capabilities.

Quality management system (QMS): System for controlling quality processes, records, responsibilities and improvement activities.

Manufacturing execution system (MES): System that manages and records production execution between planning and shop-floor control.

Industrial Internet of Things (IIoT): Network of industrial devices, sensors, machines and software exchanging operational data.

Interoperability: Ability of systems to exchange and meaningfully use information.

Digital twin: Connected digital representation of a physical asset, process or system.

Machine vision: Automated analysis of images for inspection, measurement or process control.

Predictive quality: Use of data and models to estimate future quality outcomes.

Model drift: Change in the relationship between model inputs and outcomes over time.

Data governance: Assignment of rules, ownership and accountability for data quality, access, meaning and use.

Data lineage: Record of data origin, transformation and movement through systems.

Algorithmic bias: Systematic and undesirable differences in algorithmic outcomes across groups or conditions.

Explainability: Ability to communicate how an analytical model supports its output and its limitations.

Edge computing: Processing data near the physical source rather than exclusively in a central cloud.

Closed-loop control: Control in which measured output is used to adjust the process.

Human-in-the-loop: Arrangement in which a person reviews or authorises an automated recommendation or action.

Model validation: Evaluation that establishes whether a model is suitable for its intended use.

Cybersecurity: Protection of systems, data and operations from unauthorised access, disruption or manipulation.

Review Questions

Short Answer Questions

1. Distinguish digitisation, digitalisation and digital transformation.
2. What are the principal differences between a QMS and an MES?
3. Why is sensor calibration important in an IIoT quality system?
4. Define model drift and identify two possible causes.
5. State four elements of a defence-in-depth cybersecurity strategy.

Long Answer Questions

1. Explain how a digital quality architecture connects sensors, operations systems, analytics and human decisions.
2. Evaluate the advantages and limitations of machine vision compared with conventional inspection.
3. Discuss how predictive quality differs from causal analysis and explain why the distinction matters.
4. Describe a responsible implementation framework for an AI-assisted quality application.
5. Analyse the workforce capabilities and organisational changes required for a smart factory.

Analytical / Application Questions

1. A manufacturer has frequent final-inspection failures but limited process data. Design a staged plan for determining whether predictive quality is feasible.
2. An AI vision system has high overall accuracy but misses defects on one product variant. Identify likely causes and propose a validation and governance response.
3. A proposed system will automatically release safety-critical product based on sensor data. Evaluate the technical, cybersecurity, human-accountability and ethical controls that should be required before deployment.

Multiple Choice Questions with Answers

Question 1

Which system primarily manages audits, CAPA, document control and quality records?

A. Quality management system

B. Programmable logic controller

C. Warehouse management system

D. Computer-aided design system

Correct Answer: A. Quality management system

Question 2

What is the principal quality value of contextualising MES data?

A. It removes the need for calibration

B. It links measurements to products, operations, equipment and time

C. It guarantees that every product is conforming

D. It eliminates cybersecurity risks

Correct Answer: B. It links measurements to products, operations, equipment and time

Question 3

Which statement best describes a digital twin?

A. A static paper drawing of a machine

B. A duplicate physical production line

C. A connected digital representation of a physical asset or process

D. A database containing only financial data

Correct Answer: C. A connected digital representation of a physical asset or process

Question 4

Why can overall accuracy be misleading for a rare-defect AI model?

A. Accuracy cannot be calculated for inspection data

B. A model can appear accurate while missing most rare defects

C. Accuracy always equals recall

D. Rare defects automatically invalidate all models

Correct Answer: B. A model can appear accurate while missing most rare defects

Question 5

Which activity is most directly associated with preventing model drift?

A. Monitoring changes in inputs, outcomes and predictive performance

B. Deleting all historical data

C. Removing human review

D. Increasing screen brightness

Correct Answer: A. Monitoring changes in inputs, outcomes and predictive performance

Question 6

Which control is an example of defence in depth?

A. Using one shared administrator password

B. Disabling audit trails

C. Combining network segmentation, access control and recovery procedures

D. Connecting every machine directly to the public internet

Correct Answer: C. Combining network segmentation, access control and recovery procedures

Question 7

What is a major difference between prediction and causal analysis?

A. Prediction estimates likely outcomes, whereas causal analysis examines effects of interventions

B. Prediction never uses historical data

C. Causal analysis cannot involve experiments

D. Prediction always proves the root cause

Correct Answer: A. Prediction estimates likely outcomes, whereas causal analysis examines effects of interventions

Question 8

Which condition most increases the risk of algorithmic bias in machine vision?

A. Representative images from all product variants

B. Clearly defined defect labels

C. Training data that excludes important operating conditions

D. Independent testing on later production data

Correct Answer: C. Training data that excludes important operating conditions

Question 9

Calculate the ROI when annual benefits are $120,000 and annual costs are $80,000. Use ROI = (benefits − costs) ÷ costs × 100%.

A. 33%

B. 40%

C. 50%

D. 67%

Correct Answer: C. 50%

Question 10

Why should an automated release decision for safety-critical product normally include strong human accountability and override controls?

A. Automation is incapable of processing data

B. The consequences of an incorrect release may be severe and difficult to reverse

C. Human review always improves machine speed

D. Cybersecurity is irrelevant to release decisions

Correct Answer: B. The consequences of an incorrect release may be severe and difficult to reverse