Digitalization, AI and Emerging Industrial Practice

Introduction and context

Industrial energy management and safety are increasingly shaped by connected measurement, automated control and data-driven decision-making. A modern factory may contain smart electricity meters, vibration sensors, programmable logic controllers (PLCs), supervisory control and data acquisition (SCADA) systems, cloud platforms, digital twins, machine-learning models, robotic systems and wearable devices. These technologies can reduce energy waste, detect abnormal conditions before failure, and improve worker protection. They also introduce new risks: poor-quality data, cyberattack, inappropriate automation, privacy violations and overconfidence in opaque algorithms.

Digitalization is therefore not simply the purchase of sensors or software. It is the coordinated design of physical equipment, communication networks, data governance, analytical models, human roles and management controls. The engineering objective is to improve performance while preserving safety, reliability, security and accountability.

Learning objectives

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

1. Explain the roles of smart meters, IoT sensors, SCADA and digital twins.
2. Evaluate predictive maintenance and AI-based energy and hazard applications.
3. Select appropriate computer-vision and wearable-safety applications.
4. Assess data quality, cybersecurity, privacy, bias and explainability.
5. Design human oversight and governance for automated industrial decisions.
6. Discuss the characteristics of future autonomous and low-carbon factories.

1. The industrial digital architecture

A useful architecture has five interacting layers:

The layers should not be treated as independent. For example, a predictive-maintenance model is useful only when a sensor measures the relevant failure precursor, the communication system preserves the time sequence, the model is validated, and maintenance personnel can act on the prediction.

Smart meters and IoT sensors

A smart meter records energy variables at defined intervals and can communicate readings automatically. Depending on the installation, it may measure active energy, reactive energy, demand, voltage, current, power factor and power-quality events. An IoT sensor is a connected measuring device that may monitor temperature, pressure, flow, vibration, humidity, acoustic emission, occupancy or gas concentration.

For electrical power, a simplified three-phase relationship is:

where P is active power in watts, V_L is line-to-line voltage in volts, I_L is line current in amperes and cosφ is power factor. This expression assumes a balanced three-phase system and approximately sinusoidal conditions. For distorted waveforms, true power should be obtained from appropriately sampled voltage and current waveforms rather than relying on the simplified equation.

Energy is the time integral of power:

For constant power, E = Pt. If power is recorded in kilowatts and time in hours, energy is obtained in kilowatt-hours. Meter interval, calibration, timestamp accuracy and communication reliability determine whether the data are suitable for billing, energy diagnosis or real-time control.

IoT deployments require attention to sensor placement, measurement range, accuracy, sampling frequency, battery life and environmental protection. A temperature sensor mounted near a heat source rather than in the process stream can produce precise but misleading readings. This illustrates the distinction between precision and validity.

SCADA and industrial control

SCADA systems collect data from field devices, present trends and alarms to operators, and may permit supervisory commands. PLCs and distributed control systems execute faster local control logic. A sound design keeps essential protective functions local and does not depend on an internet connection. SCADA connectivity should therefore support supervision and authorized control, not replace independent safety instrumented systems or physical emergency stops.

Typical SCADA functions include:

• real-time visualization and historical trending;
• alarm generation and acknowledgement;
• set-point supervision;
• event and audit logging; and
• reporting of production and energy indicators.

Alarm floods, poor prioritization and unreviewed set-point changes can reduce rather than improve safety. Alarm management should define consequence, priority, response time and responsible role for each alarm.

Digital twins

A digital twin is a maintained digital representation of a physical asset, process or facility that uses current data to support analysis or decisions. It is more than a static three-dimensional model. A useful twin has a defined purpose, a traceable relationship to the physical system, updating data and validated models.

A twin may be descriptive, diagnostic, predictive or prescriptive. For example, a descriptive twin shows current chiller performance; a predictive twin estimates the effect of fouling; a prescriptive twin compares cleaning or replacement actions. Model fidelity should match the decision. A highly detailed model is not automatically better if its parameters are uncertain or its computation is too slow for operational use.

2. Data quality and analytical foundations

Industrial AI is limited by the quality of its inputs. Important dimensions include:

• accuracy: closeness to the true value;
• completeness: absence of missing observations;
• timeliness: availability at the required time;
• consistency: agreement among systems and units;
• traceability: ability to identify source, transformation and custodian; and
• representativeness: coverage of normal and abnormal operating conditions.

A practical data-quality process includes sensor calibration, unit and tag-name standards, synchronized clocks, range checks, missing-value flags, change logs and documented treatment of outliers. Imputation may be acceptable for analysis but dangerous for protective control. A substituted temperature value must never silently appear equivalent to a measured value.

Useful energy indicators include specific energy consumption:

where SEC is specific energy consumption, E is energy consumed and Q is an appropriate production quantity, such as tonnes or units. Comparisons are meaningful only when product mix, operating hours, quality requirements and boundary conditions are comparable.

A baseline can be represented as:

where Q may represent production, T ambient temperature, H operating hours and M material or product mix. Energy savings should be estimated by comparing actual energy with the baseline under equivalent conditions, not by comparing two unadjusted monthly totals.

3. Predictive maintenance and AI applications

Preventive maintenance replaces or services equipment at planned intervals. Predictive maintenance uses condition data and models to estimate degradation, abnormality or remaining useful life. Prescriptive maintenance additionally recommends an action while considering risk, cost, production and safety.

Common analytical approaches include:

• threshold and rule-based detection;
• statistical process control;
• regression and classification;
• clustering and anomaly detection;
• time-series forecasting; and
• physics-informed or hybrid models.

A classification model can be evaluated using a confusion matrix. If a model predicts a bearing fault, a true positive identifies a real fault, a false positive raises an unnecessary inspection, a true negative correctly indicates normal operation, and a false negative misses a fault. In safety-critical applications, the cost of a false negative may be much greater than the cost of a false positive.

Remaining useful life predictions should be expressed with uncertainty rather than as an apparently exact date. Model performance may deteriorate when equipment type, load, environment or maintenance practice changes. This is known as distribution shift. Continuous monitoring, periodic validation and controlled model updates are essential.

AI-based energy optimization

AI can forecast demand, identify inefficient operating states, coordinate storage and adjust flexible loads. Optimization must respect constraints such as production requirements, equipment limits, comfort, emissions, electricity tariffs and safety interlocks. A simplified objective may be written as:

subject to equipment, process and safety constraints. The terms represent cost components and are not automatically comparable unless units or weighting factors are defined. An optimization that reduces electricity cost by violating temperature limits or increasing product defects is not a successful energy-management solution.

AI-based hazard prediction

Hazard models may combine permit-to-work information, equipment state, environmental measurements, near-miss reports and work schedules. Their output should support hazard identification and control selection, not declare an area safe. A low predicted probability is not evidence that a known dangerous condition can be ignored.

4. Computer vision and wearable safety devices

Computer vision can detect restricted-area entry, missing personal protective equipment, unsafe posture, vehicle–pedestrian proximity, smoke, flame or liquid leakage. Its reliability depends on lighting, camera position, occlusion, field of view and the diversity of people and equipment represented in training data. Vision systems should provide a clear escalation path and should not be the sole protection against catastrophic hazards.

Wearable devices may include location tags, inertial sensors, gas detectors, heart-rate monitors and fall detection. Benefits include lone-worker alerts, exposure monitoring and emergency localization. Limitations include battery failure, false alarms, discomfort, inaccurate positioning and worker concerns about surveillance. Workers should know what is collected, why it is collected, who can access it and how long it is retained.

5. Cybersecurity, privacy and responsible AI

Industrial systems have a distinctive cybersecurity requirement: safety and availability may be more important than confidentiality, although all three matter. Effective controls include asset inventories, network segmentation, least privilege, multifactor authentication where feasible, secure remote access, patch and vulnerability management, backups, logging, incident response and tested recovery. Legacy control equipment may not support modern security functions; compensating controls and isolation may therefore be necessary.

Safety engineering and cybersecurity must be integrated. A cyberattack that changes a set point can become a physical hazard. Conversely, an emergency response that disconnects a network may affect production monitoring. Cybersecurity exercises should include operations, maintenance, safety and management personnel.

Privacy issues arise when systems record faces, locations, physiological signals, productivity or behavioral patterns. Data collection should be necessary, proportionate, transparent and protected. Aggregated data may be suitable for trend analysis, whereas individual-level data require stronger justification and access control.

AI governance requires:

1. a defined purpose and decision boundary;
2. documented data sources and limitations;
3. validation against relevant operating conditions;
4. performance monitoring after deployment;
5. explainable outputs appropriate to the user;
6. human authority to challenge or override the system; and
7. records of decisions, changes and incidents.

Explainability does not mean that every algorithm must be mathematically simple. It means that users can understand the relevant factors, uncertainty, limitations and recommended response. Bias can arise from under-represented equipment, shifts, body types, lighting conditions or historical reporting practices. Testing should therefore examine performance across relevant groups and operating regimes.

6. Human oversight and safe automation

Automation should be designed around the question: what must the human know, decide and be able to do? Four broad arrangements are useful:

The appropriate arrangement depends on hazard severity, reversibility, uncertainty and response time. A human placed only as a nominal approver may suffer from automation bias and approve an incorrect recommendation without meaningful review. Interfaces should show confidence, evidence, constraints and escalation status rather than only a green or red label.

7. Illustrative case: connected compressed-air system

Background. An illustrative medium-sized plant operates several compressors supplying a manufacturing line. Electricity use is high, pressure fluctuates, and maintenance is largely calendar-based.

Problem. A baseline review finds simultaneous operation of lightly loaded compressors, night-time leakage and repeated high-temperature trips. Existing meters provide monthly totals but no compressor-level measurements.

Analysis. The plant installs calibrated power meters, pressure sensors, flow meters and temperature sensors. Timestamps are synchronized and data are stored in a historian. A dashboard calculates compressor-specific energy and compares demand with production. Maintenance staff inspect ultrasonic leak-survey results rather than relying solely on an anomaly score.

Approach. A rule-based control strategy first sequences compressors according to demand and minimum safe loading. A later predictive model identifies likely bearing degradation using vibration and temperature. The model recommends inspection; it cannot defeat a trip or authorize continued operation beyond safety limits. Remote access is segmented, accounts are individualized and model changes require review.

Outcome. The plant reduces unnecessary unloaded running, repairs leaks and detects one developing mechanical problem earlier than the calendar schedule would have done. The main benefit comes from measurement, operating discipline and human action; AI is a supporting component rather than a substitute for engineering judgment.

Learning points. Begin with a clearly bounded problem, establish data quality before advanced analytics, retain independent protective controls, and measure outcomes against a normalized baseline. A technically sophisticated model cannot compensate for missing meters, unclear ownership or poor maintenance response.

8. Future autonomous and low-carbon factories

Future factories are likely to combine autonomous mobile robots, collaborative robots, digital twins, flexible electrification, on-site renewable generation, batteries, demand response and industrial heat recovery. Carbon-aware control may shift flexible processes toward periods of lower grid emissions, provided production and safety constraints remain satisfied.

Autonomy will not eliminate engineering responsibility. New risks may include interacting robots, software-induced common-cause failures, opaque supply chains, dependence on communication networks and rapid changes in work organization. Safety cases should address both foreseeable misuse and algorithmic failure. Competence requirements will expand: mechanical engineers will need familiarity with data interpretation, control systems, cybersecurity, lifecycle assessment and human factors.

The transition should follow a staged approach: measure, stabilize, connect, analyze, pilot, validate, scale and continuously audit. Technology selection should consider total lifecycle cost, maintainability, interoperability, training, energy impact and end-of-life disposal—not merely purchase price or advertised accuracy.

Challenges and limitations

Digital systems can create excessive data without useful information, increase vendor dependence and impose hidden maintenance burdens. Wireless devices may suffer interference; cloud services may be unavailable; sensors may drift; models may degrade; and dashboards may encourage attention to what is measurable rather than what is important. Digitalization also has an embodied environmental cost through hardware manufacture, energy consumption and electronic waste.

Risk assessment should therefore compare the risk introduced by the digital system with the risk reduced by it. Independent verification, fail-safe states, manual fallback, competency training and periodic drills remain necessary.

Chapter Summary

Digital industrial practice links physical equipment, sensors, control systems, data platforms, analytics and human governance. Smart meters and IoT sensors provide the evidence needed for energy and condition management, while SCADA supports supervisory visibility and control. Digital twins connect current data with models for diagnostic and predictive decisions. Predictive maintenance and AI-based optimization can improve reliability and energy performance, but only when data are valid, models are monitored and recommendations respect process and safety constraints. Computer vision and wearable devices can strengthen hazard detection while raising concerns about false alarms, privacy and surveillance. Cybersecurity, explainability, bias control, human oversight and independent protective systems are essential. The future low-carbon factory will be increasingly autonomous and interconnected, but responsible engineering will remain central to safe and sustainable performance.

Key Terms

Smart meter: A communicating meter that records and transmits energy measurements.

IoT sensor: A connected device that measures a physical or environmental variable.

SCADA: Supervisory control and data acquisition system for monitoring and supervisory operation.

PLC: Programmable logic controller used for industrial control logic.

Digital twin: A maintained digital representation linked to a physical asset or process.

Data quality: Fitness of data for its intended use.

Specific energy consumption: Energy used per unit of production.

Baseline: Reference performance adjusted for relevant operating conditions.

Predictive maintenance: Maintenance informed by condition or failure predictions.

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

Remaining useful life: Estimated operating time before a defined failure or performance limit.

Computer vision: Analysis of images or video to identify objects, conditions or events.

Wearable safety device: Body-worn equipment that monitors hazards, location or worker condition.

Cybersecurity: Protection of digital systems, data and connected operations from harmful actions.

Distribution shift: Change between the conditions represented in model development and deployment.

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

Automation bias: Tendency to accept an automated recommendation without adequate verification.

Human oversight: Defined human authority to monitor, question, override or stop automated operation.

Low-carbon factory: Industrial facility designed to reduce greenhouse-gas emissions across its operations and lifecycle.

Review Questions

Short Answer Questions

1. Distinguish a smart meter from a general IoT sensor.
2. What are the five broad layers of an industrial digital architecture?
3. Why is data quality important in predictive maintenance?
4. State two cybersecurity controls suitable for an industrial network.
5. What is automation bias?

Long Answer Questions

1. Explain how SCADA, PLCs, sensors and human operators interact in a safe industrial control architecture.
2. Discuss the benefits and limitations of digital twins for energy management.
3. Compare preventive, predictive and prescriptive maintenance.
4. Evaluate privacy and ethical issues associated with wearable safety devices.
5. Explain why human oversight and independent protective systems remain necessary in autonomous factories.

Analytical / Application Questions

1. A plant has reliable monthly energy totals but no production-normalized data. Design a data-improvement plan before introducing AI-based energy optimization.
2. A computer-vision system detects missing helmets but performs poorly during night shifts. Identify likely causes and propose a validation and control strategy.
3. Develop a staged implementation plan for the compressed-air case, including measurement, cybersecurity, model validation, human response and performance evaluation.

Multiple Choice Questions with Answers

Question 1

Which system primarily provides supervisory monitoring, alarms and historical trends in an industrial plant?

A. SCADA

B. Bearing

C. Heat exchanger

D. Manual valve

Correct Answer: A. SCADA

Question 2

Under the stated balanced three-phase assumption, which expression gives active power?

A. P = V_L/I_L

B. P = √3 V_L I_L cosφ

C. P = V_L + I_L

D. P = I_L/(√3 V_L)

Correct Answer: B. P = √3 V_L I_L cosφ

Question 3

Which data-quality property concerns whether data cover the operating conditions relevant to deployment?

A. Precision

B. Representativeness

C. Encryption

D. Compression

Correct Answer: B. Representativeness

Question 4

What is the principal distinction between predictive and preventive maintenance?

A. Predictive maintenance uses condition information to guide timing

B. Predictive maintenance never requires sensors

C. Preventive maintenance is always controlled by AI

D. Preventive maintenance cannot be scheduled

Correct Answer: A. Predictive maintenance uses condition information to guide timing

Question 5

Why should an AI energy optimizer include process and safety constraints?

A. To guarantee that the algorithm uses no data

B. To prevent energy reduction from causing unsafe or unacceptable operation

C. To eliminate the need for operators

D. To make all equipment operate at maximum load

Correct Answer: B. To prevent energy reduction from causing unsafe or unacceptable operation

Question 6

Which is an example of a false negative in hazard prediction?

A. An alarm is raised when no hazard exists

B. A real hazard is correctly detected

C. A real hazard is missed by the system

D. No alarm is raised when no hazard exists

Correct Answer: C. A real hazard is missed by the system

Question 7

Which control most directly limits the damage caused by unauthorized remote access to an industrial network?

A. Network segmentation and controlled remote access

B. Increasing screen brightness

C. Removing all event logs

D. Disabling equipment labels

Correct Answer: A. Network segmentation and controlled remote access

Question 8

What does distribution shift mean in machine learning deployment?

A. The model is stored in two locations

B. Deployment conditions differ from those represented during model development

C. A sensor has two output wires

D. The production line changes ownership

Correct Answer: B. Deployment conditions differ from those represented during model development

Question 9

Which practice best supports responsible use of worker-location data from wearables?

A. Collecting all possible data indefinitely

B. Publishing individual locations publicly

C. Defining purpose, access rights and retention limits

D. Allowing vendors unrestricted reuse

Correct Answer: C. Defining purpose, access rights and retention limits

Question 10

In the illustrative compressed-air case, what should the predictive model recommendation do?

A. Replace all independent safety trips

B. Authorize operation beyond safety limits

C. Support inspection and maintenance decisions within defined controls

D. Disconnect every compressor automatically

Correct Answer: C. Support inspection and maintenance decisions within defined controls