Why Science? | Sectoral AI Centre of Excellence for Manufacturing

Manufacturing AI serves a defined transformation of materials, parts, information or energy. Manufacturing AI begins with a production question rather than an algorithm. Chapter 1 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Map inputs, operations, outputs, constraints and owners before selecting a model. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

An algorithm cannot compensate for an undefined process. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Start from the process”, Draw a process map. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 1 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Next chapter

Factory data combines sensors, controls, maintenance, quality, planning and operator records. Manufacturing AI begins with a production question rather than an algorithm. Chapter 2 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Align timestamps, units, asset identities and change histories. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

A large historian can contain gaps, drift and inconsistent meaning. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Build trustworthy factory data”, Draft a data dictionary. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 2 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Measurement-system analysis tests repeatability, reproducibility, bias, resolution and stability. Manufacturing AI begins with a production question rather than an algorithm. Chapter 3 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Calibrate sensors and compare variation with the process tolerance. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

AI cannot recover information a measurement system never captured. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Validate measurement systems”, Plan a gauge study concept. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 3 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Value links prediction to a feasible action, response time, cost and accountable decision. Manufacturing AI begins with a production question rather than an algorithm. Chapter 4 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Quantify baseline performance, error costs and intervention capacity. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

A metric gain cannot guarantee productivity or savings. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Define operational value”, Write a value hypothesis. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 4 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Computer vision can support consistent inspection when images, defects and lighting are controlled. Manufacturing AI begins with a production question rather than an algorithm. Chapter 5 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Define defect taxonomy, image protocol and part-level evaluation. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

A good image classifier does not prove product conformance alone. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Inspect quality with vision”, Design an inspection benchmark. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 5 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Predictive shopfloor models estimate faults or degradation from time-series and maintenance evidence. Manufacturing AI begins with a production question rather than an algorithm. Chapter 6 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Use chronological validation, event labels and lead-time metrics. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

Correlation with a fault does not identify its cause. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Predict shopfloor condition”, Build a failure timeline. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 6 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Optimisation selects plans under capacity, sequence, labour, material, energy and delivery constraints. Manufacturing AI begins with a production question rather than an algorithm. Chapter 7 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Test objectives, constraints and scenario sensitivity against current planning. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

An optimal mathematical plan may be infeasible on the floor. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Optimise operations”, Draft a scheduling model. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 7 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Sustainable manufacturing AI can target energy, scrap, water or emissions while preserving quality and throughput. Manufacturing AI begins with a production question rather than an algorithm. Chapter 8 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Use measured baselines, system boundaries and verified counterfactuals. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

Efficiency at one machine may shift impact elsewhere. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Connect sustainability evidence”, Create an impact boundary. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 8 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Generative design, simulation and surrogate models can explore alternatives faster. Manufacturing AI begins with a production question rather than an algorithm. Chapter 9 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Apply engineering constraints, manufacturability checks and physical validation. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

A generated design is not a qualified product or process. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Innovate products and processes”, Build a design funnel. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 9 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Industrial automation combines perception, planning, control, tooling and safeguarding. Manufacturing AI begins with a production question rather than an algorithm. Chapter 10 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Validate tasks, workspaces, failure modes and human interaction. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

Robot accuracy in a demo cannot establish production safety. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Automate with robots”, Map a guarded task. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 10 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

A sandbox lets teams evaluate tools, data and workflows in a controlled environment. Manufacturing AI begins with a production question rather than an algorithm. Chapter 11 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Isolate systems, use representative copies and predefine exit criteria. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

Sandbox performance cannot guarantee live-plant behaviour. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Use an AI sandbox”, Write a sandbox protocol. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 11 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Production AI must connect with controls, historians and enterprise systems without disrupting operations. Manufacturing AI begins with a production question rather than an algorithm. Chapter 12 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Define interfaces, latency, data ownership, fail-safe states and change control. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

A connected model can introduce new failure and attack paths. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Integrate operational technology”, Draw an integration boundary. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 12 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Industrial AI security covers data, models, software supply chains, identities and operational access. Manufacturing AI begins with a production question rather than an algorithm. Chapter 13 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Use least privilege, signing, logging, patching and incident response. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

A secure network perimeter cannot remove insider or model risks. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Secure the AI pipeline”, Create a threat register. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 13 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

AI changes decisions, workload, skills and accountability for operators and engineers. Manufacturing AI begins with a production question rather than an algorithm. Chapter 14 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Co-design interfaces, train users and test automation bias and override. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

A recommendation screen cannot transfer responsibility to a machine. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Design human-centred work”, Draft a human-in-the-loop plan. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 14 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

MLOps governs data, code, models, approvals, deployment, monitoring and rollback. Manufacturing AI begins with a production question rather than an algorithm. Chapter 15 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Use version control, traceability, drift tests and controlled releases. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

Frequent retraining is not the same as continuous improvement. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Operate manufacturing MLOps”, Build a release checklist. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 15 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

Scaling requires transferable definitions, adaptable models, local validation and sustained ownership. Manufacturing AI begins with a production question rather than an algorithm. Chapter 16 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Separate common components from site-specific configuration. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

Success on one line cannot establish value across sectors. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Scale across factories”, Create a scale-readiness matrix. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 16 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Next chapter

The receipt links process, data, measurements, model, safety, people, economics and decisions. Manufacturing AI begins with a production question rather than an algorithm. Chapter 17 connects machines, materials, people, process plans, sensors, quality and business decisions. AIMfg describes a national manufacturing centre jointly developed by A*STAR and MTI, with current focus areas covering operations optimisation and sustainability, predictive shopfloor work, quality assurance, product or process innovation and industrial automation.

Archive baselines, experiments, deviations, reviews and outcomes. A sound factory study records the product family, equipment state, process step, sensor identity, label source, sampling frequency, unit, shift, maintenance history, cost of error and intervention rule. Engineers build time-aware train and test splits, compare against current practice, examine drift and validate in a sandbox or controlled production trial before scaling.

Documentation cannot authorise unsafe production changes. A laboratory demonstration cannot promise factory uptime, savings, defect prevention or sustainability gains. Production data can be sparse, imbalanced, delayed or altered by maintenance and operator practice. AI must fit operational technology, cybersecurity, safety, quality systems and accountable human work; any change with product or safety consequences follows the organisation’s formal validation and change-control process.

For “Keep a manufacturing-AI receipt”, Assemble a one-page receipt. construct a fictional production table or use safe public data. Define the decision latency, false-alarm cost, missed-event cost, baseline, acceptance range and fallback. Plot performance by machine, product and time rather than quoting one average. Keep the activity away from live machinery and never bypass factory access, safety or information-security controls.

Did you know? The rare event may matter most in manufacturing. A dataset with 99 percent normal examples can reward a model that never detects the costly fault. Chapter 17 therefore uses precision, recall, lead time, calibration and operational response together, showing how a modest algorithm can create value only when labels, sensors and workflows are trustworthy.

Relevant skills span process engineering, manufacturing science, robotics, controls, computer vision, optimisation, data engineering, cybersecurity, maintenance, ergonomics and change management. Students can start with measurement systems, control charts, simple classifiers and simulation. Training providers and employers define current requirements; a strong portfolio explains data, limitations, validation and human decisions without promising a job or return.

Contents · Previous chapter · Continue in the Science Learning Hub