14. AI, Cyber-Physical Control, and Industry 4.0

Introduction and context

Machine dynamics and control have traditionally been developed from mathematical models, experimentally identified parameters, and feedback controllers designed around known operating conditions. These methods remain fundamental. However, modern machines increasingly operate as connected cyber-physical systems (CPS): mechanical components, sensors, actuators, embedded computation, communication networks, cloud platforms, and human operators interact continuously. The machine is no longer an isolated plant; it is part of a data-rich production, transport, energy, or service ecosystem.

Industry 4.0 refers to the integration of cyber-physical systems, industrial communication, automation, data analytics, artificial intelligence (AI), and flexible manufacturing. In this setting, AI does not replace dynamics and control theory. Rather, it complements them by learning unknown relationships, identifying changing system parameters, detecting early faults, optimizing operation, and supporting human decisions. The central engineering problem is therefore not simply how to apply machine learning, but how to combine data-driven methods with physical knowledge, safety requirements, control theory, and responsible governance.

A machine-learning model may predict bearing degradation, estimate friction, or infer an unmeasured load. Yet its prediction can be unreliable when sensors drift, operating conditions change, or the training data omit rare failures. A reinforcement-learning controller may discover an efficient policy, but unrestricted exploration on a physical machine can cause damage. Cloud analytics may provide extensive computational power, while network delay or disconnection can make cloud-only control unsafe. These examples illustrate the defining challenge of AI-enabled machines: performance must be balanced with stability, safety, explainability, security, and maintainability.

Learning objectives

After studying this chapter, the reader should be able to:

1. Explain the architecture and significance of AI-enabled cyber-physical machines.

2. Distinguish supervised, unsupervised, self-supervised, and reinforcement-learning approaches used in machine systems.

3. Apply machine-learning methods to system identification, anomaly detection, and predictive maintenance.

4. Compare edge, cloud, and hybrid analytics for control and industrial decision-making.

5. Explain connected control, autonomous machines, and human–machine collaboration.

6. Evaluate data quality, model validity, explainability, cybersecurity, and safe deployment.

7. Design a staged validation and governance strategy for an AI-enabled control application.

8. Critically assess the limitations and professional responsibilities associated with intelligent machines.

From automated machines to cyber-physical systems

A conventional automated machine generally consists of a plant, sensors, a controller, actuators, and a human-defined control algorithm. A cyber-physical machine adds computational intelligence and communication. Its digital representation may include current states, historical operating data, maintenance records, simulated responses, and predicted future conditions. The physical and digital elements form a closed loop: measurements update the digital model, analytics generate decisions, and commands alter the physical process.

A useful abstraction is

x(k+1) = f[x(k), u(k), d(k), theta(k)] + w(k)

y(k) = h[x(k)] + v(k),

where x is the physical state, u is the control input, d represents disturbances, theta denotes possibly changing parameters, y is the measured output, and w and v represent process and measurement uncertainty. In a data-driven setting, an algorithm estimates f, h, theta, or an auxiliary quantity such as remaining useful life. The control decision may then be written as

u(k) = pi[x(k), y(k), z(k)],

where pi is a control policy and z may contain learned features, health indicators, or production context.

The cyber-physical architecture normally contains five interacting layers. The physical layer includes mechanisms, motors, drives, tools, and workpieces. The sensing and actuation layer acquires measurements and applies commands. The edge layer performs local filtering, feature extraction, fast inference, and safety functions. The supervisory or platform layer stores data, coordinates assets, and runs broader analytics. The human and enterprise layer connects operators, engineers, maintenance, production planning, and business systems. These layers are not rigid; their boundaries depend on latency, reliability, safety, and computational requirements.

Important concepts and terminology

A digital twin is a continuously updated computational representation of a physical asset or process. It may be physics-based, data-driven, or hybrid. A digital shadow is a related representation in which data flow primarily from the physical object to the digital object, without necessarily creating a closed control loop.

System identification is the estimation of a mathematical model from measured inputs and outputs. Machine-learning system identification extends this idea by using regression, neural networks, Gaussian processes, sparse models, or hybrid structures to represent nonlinear or time-varying behaviour.

An anomaly is an observation or pattern that differs significantly from expected behaviour. Fault diagnosis seeks to identify the cause or location of a fault, whereas prognostics estimates future degradation or remaining useful life (RUL). Concept drift occurs when the statistical relationship between inputs and outputs changes over time. Data drift refers to a change in the distribution of measured inputs, even if the underlying relationship remains similar.

A feature is a measured or calculated variable supplied to a learning algorithm. A health indicator is a feature or latent quantity that changes systematically with degradation. Inference is the computation of a model output after training. A policy is a rule that maps an observed state to an action, especially in reinforcement learning. A safe set is the set of states and actions that satisfy specified physical and operational constraints.

Machine learning for system identification

Classical identification methods estimate parameters of transfer functions, state-space models, frequency-response models, or differential equations. They are particularly effective when the physics is understood and the system operates near a known regime. Machine learning becomes valuable when the dynamics are strongly nonlinear, difficult to model analytically, affected by wear, or dependent on many contextual variables.

In supervised identification, a data set contains input histories and measured outputs. A model learns the mapping between them by minimizing a loss function such as mean squared prediction error. Linear regression, autoregressive models with exogenous inputs, support-vector regression, Gaussian processes, and neural networks are common choices. Recurrent networks and temporal convolutional models can represent sequential dependencies, while transformers may represent longer temporal relationships when sufficient data are available.

Purely black-box identification can fit training data accurately yet violate physical principles. A learned model may predict negative mass, impossible temperature, or an unstable response. Hybrid modelling addresses this problem by embedding physical structure in the learning process. Examples include:

• estimating unknown coefficients in a known dynamic equation;
• using a neural network only for an unmodelled residual force;
• imposing conservation laws or passivity constraints during training;
• combining a state observer with a learned disturbance model; and
• using a physics-based simulator to generate additional training data.

The quality of an identification experiment is as important as the algorithm. Inputs must sufficiently excite the relevant modes, while remaining safe for the machine. Sampling frequency must capture the fastest important dynamics, and sensor synchronization must be verified. Training, validation, and test data should be separated by operating condition or time, not merely by random row selection. Random splitting of highly correlated time-series samples can produce an unrealistically optimistic result because nearly identical observations appear in both training and test sets.

Model assessment should include one-step and multi-step prediction error, frequency-response agreement, stability, constraint violations, robustness to noise, and performance during transients. A model that predicts accurately for ten seconds may still be unsuitable for a controller if small errors accumulate over several minutes. Uncertainty estimates are particularly important. Gaussian processes, ensembles, Bayesian methods, and conformal prediction can indicate when a model is operating outside its evidence base.

Anomaly detection and intelligent maintenance

Anomaly detection is often the first AI application adopted in industrial machinery because it can provide value without directly commanding the plant. In a supervised approach, labelled examples of normal and faulty operation are available. Classification algorithms then distinguish classes such as imbalance, misalignment, lubrication failure, or sensor fault. In practice, labelled failure data are scarce, and failures may be too costly to generate. Consequently, unsupervised and semi-supervised methods are widely used.

A normality model is trained using data believed to represent healthy operation. An anomaly score may be based on reconstruction error, distance from a learned cluster, prediction residual, likelihood, or deviation from a digital-twin model. Autoencoders compress and reconstruct multivariate signals; principal-component models identify departures from dominant correlations; isolation-based methods detect observations that are readily separated from normal data. Time-frequency features from vibration, acoustic emission, motor current, temperature, and oil debris can reveal different fault mechanisms.

An anomaly score is not automatically a diagnosis. A high vibration level may result from imbalance, looseness, resonance, or an altered load. The system should therefore combine multiple signals, operating context, engineering thresholds, and maintenance history. Alarm design must address nuisance alarms, missed detections, alarm persistence, severity, and escalation. Precision, recall, false-alarm rate, detection delay, and cost-weighted consequences are more informative than accuracy alone when faults are rare.

Predictive maintenance uses health estimates to schedule inspection or replacement before functional failure. A useful maintenance decision considers RUL uncertainty, spare-part availability, production schedules, safety consequences, and the cost of planned versus unplanned downtime. AI should support, not obscure, the physical reasoning of maintenance personnel. A model prediction accompanied by the contributing signals, confidence range, and recommended inspection can be more useful than a numerical score alone.

Reinforcement learning and intelligent control

In reinforcement learning (RL), an agent observes a state, selects an action, receives a reward, and improves its policy through interaction. The objective is commonly expressed as maximizing expected discounted return:

J(pi) = E[sum gamma^t r(t)],

where r(t) is the reward and gamma is a discount factor. RL is attractive for problems involving sequential decisions, uncertain models, energy-management trade-offs, scheduling, and adaptive operation.

However, an industrial machine is not a simulation game. Exploration can create excessive loads, collisions, thermal stress, or unsafe human interaction. Direct trial-and-error learning on the plant is therefore generally unacceptable for safety-critical functions. Safer alternatives include offline RL using historical data, simulation-based training, model predictive safety filters, constrained RL, action shielding, and supervised pretraining. A learned policy should normally operate beneath a conventional safety layer that limits position, velocity, torque, temperature, pressure, and rate of change.

Reward design requires care. If a controller is rewarded only for production rate, it may increase vibration or consume excessive energy. A more appropriate objective may combine throughput, quality, energy, wear, constraint penalties, and safety margins. Reward hacking occurs when the agent exploits an unintended shortcut in the reward definition. Formal constraints and independent monitoring are needed to prevent this behaviour.

A practical architecture is hierarchical. A conventional low-level servo controller guarantees fast tracking and stability. A slower optimizer or RL agent selects set points, trajectories, gains, or operating modes. A safety supervisor verifies the proposed action before execution. This arrangement preserves established control functions while allowing learning at a level where uncertainty and validation are manageable.

Edge, cloud, and hybrid analytics

The location of computation is a control-design decision. Edge computing places inference close to the machine, often in a programmable logic controller, industrial computer, gateway, or embedded processor. It offers low latency, operation during network interruption, reduced bandwidth use, and improved data locality. It is suitable for vibration feature extraction, collision detection, local anomaly alarms, and fast adaptive compensation.

Cloud computing provides scalable storage, high computational capacity, fleet-level comparison, centralized model training, and integration with enterprise systems. It is appropriate for historical analysis, cross-site benchmarking, large-scale retraining, and strategic maintenance planning. Its disadvantages include communication delay, dependence on network availability, data-transfer cost, and increased exposure to remote attack.

A hybrid architecture assigns hard real-time and safety functions to the edge, while the cloud performs noncritical analysis and model development. A typical workflow is: sensors acquire data; the edge filters and validates signals; local models produce immediate decisions; selected summaries are transmitted to a platform; fleet-level models are trained; validated updates are signed and deployed through controlled release procedures. Raw data need not always leave the site.

Latency, jitter, determinism, bandwidth, synchronization, and failover must be specified quantitatively. A cloud recommendation may be acceptable for maintenance scheduling but unacceptable for a high-speed position loop. Time-sensitive networking, local buffering, distributed computing, and digital-twin simulation can improve reliability, but no communication technology removes the need for local safe-state behaviour.

Connected control and autonomous machines

Connected control links multiple machines, supervisory systems, robots, and production-planning tools. Coordination may reduce energy consumption, balance workloads, or prevent one machine from creating a bottleneck for another. Nevertheless, additional communication creates coupling and new failure modes. Delay, packet loss, inconsistent clocks, corrupted data, and unavailable services can destabilize a control loop or produce conflicting commands.

Distributed control should therefore define authority, timing, synchronization, and fallback behaviour. Safety-critical interlocks should not depend solely on a remote server. Commands should be authenticated, prioritized, bounded, and logged. A machine must know whether a received command is current, authorized, and physically feasible.

Autonomous machines extend connected control by sensing their environment, interpreting goals, planning actions, and adapting to changing conditions. Autonomy is not binary; it ranges from automatic execution of a fixed sequence to supervised autonomy and fully independent operation within a restricted domain. The operational design domain must specify the conditions under which the machine is expected to function. Outside that domain, it should slow down, request assistance, or enter a safe state rather than silently extrapolate.

Human–machine collaboration

Human–machine collaboration is strongest when responsibilities are allocated according to comparative capabilities. Machines are effective at high-frequency measurement, repetitive execution, rapid calculation, and consistent monitoring. Humans contribute contextual understanding, ethical judgment, improvisation, tacit knowledge, and responsibility for exceptional decisions.

An operator interface should communicate not only the recommended action but also its basis, confidence, urgency, and consequences. Explainable AI may use feature importance, local approximations, counterfactuals, saliency maps, or physically meaningful intermediate variables. Explanations must be evaluated for fidelity: a visually attractive explanation is not useful if it does not reflect the model’s actual reasoning.

Automation bias occurs when users accept an algorithmic recommendation without adequate scrutiny. The opposite problem, algorithm aversion, occurs when users reject useful assistance after observing one visible error. Training, calibrated confidence displays, clear escalation paths, and opportunities to override the system can improve trust. The override itself should be designed carefully: it must be accessible during emergencies but protected against casual or unauthorized use.

Data quality, explainability, and validation

AI performance is constrained by the quality and representativeness of its data. Important quality dimensions include accuracy, completeness, consistency, timeliness, traceability, resolution, synchronization, and relevance. Sensor calibration, missing-value handling, outlier treatment, unit consistency, timestamp integrity, and maintenance-event labelling deserve engineering control. A model trained on data collected only during normal production may fail during start-up, shutdown, overload, seasonal changes, or degraded sensor conditions.

Data governance should establish ownership, access rights, retention periods, metadata standards, version control, and audit trails. Data lineage should show how a raw measurement became a feature, model input, alarm, or control decision. Synthetic and simulated data can supplement scarce fault data, but differences between simulation and reality—often called the sim-to-real gap—must be measured and managed.

Validation should proceed in stages. First, verify the data pipeline and model independently of the plant. Second, conduct retrospective testing with time-separated records. Third, test in a high-fidelity simulator or hardware-in-the-loop environment. Fourth, deploy in shadow mode, where predictions are recorded but do not control the machine. Fifth, use a limited pilot with conservative bounds and human approval. Finally, expand operation only after predefined acceptance criteria are met.

Acceptance criteria should include not only average prediction error, but also worst-case or percentile performance, false alarms, missed faults, uncertainty calibration, latency, resource use, safety constraints, and performance across relevant operating regimes. After deployment, monitoring must detect drift, changes in alarm frequency, degrading accuracy, and unexpected action distributions. Model updates should be versioned, tested, approved, reversible, and traceable.

Cybersecurity and safe deployment

Connectivity enlarges the attack surface of a machine. Threats include unauthorized access, malware, ransomware, spoofed sensor data, replayed commands, denial of service, theft of proprietary models, and manipulation of training data. In a control system, integrity and availability are often as important as confidentiality. A false temperature value can be more dangerous than the disclosure of a temperature value.

A defence-in-depth strategy includes asset inventories, network segmentation, least-privilege access, strong authentication, secure boot, encryption, signed software and model updates, vulnerability management, logging, backup, intrusion detection, and incident-response procedures. Security should be considered during system architecture rather than added after deployment. An AI model itself can be attacked through adversarial inputs, data poisoning, model extraction, or malicious update packages.

Safety and cybersecurity are interdependent but not identical. A secure system can still be unsafe because of a flawed model; a physically safe fallback may limit the consequences of a cyberattack. Independent safety functions, watchdogs, redundant sensing, rate limiters, emergency stops, and fail-safe or fail-operational strategies should be selected according to the application. The principle of graceful degradation is important: when uncertainty, communication, or computation becomes unacceptable, the machine should transition to a known and manageable condition.

Professional deployment also requires compliance with applicable machinery-safety, functional-safety, industrial cybersecurity, data-protection, and quality-management requirements. The exact standards depend on sector and jurisdiction. Engineers should not treat an AI model as exempt from established verification, documentation, competence, and change-control obligations merely because its internal structure is statistical.

An illustrative case: AI-assisted monitoring of a high-speed spindle

Background

Consider a fictional precision-machining plant operating high-speed spindles. Each spindle has accelerometers, motor-current measurement, temperature sensors, and a drive controller. Unexpected bearing degradation causes dimensional errors and costly stoppages.

Situation and problem

The plant initially uses a fixed vibration threshold. The threshold generates frequent alarms during high-speed cutting but fails to identify some slowly developing defects at lower speeds. Maintenance personnel therefore distrust the alarm and often inspect components unnecessarily. The engineering team proposes an AI-based health-monitoring system.

Analysis

The team first separates operating regimes by speed, feed, tool type, material, and load. It discovers that vibration amplitude alone is not a reliable indicator: cutting conditions and tool wear strongly influence the signal. Time-synchronized vibration spectra, envelope features, motor current, temperature trends, and process context are therefore selected. Historical failure labels are incomplete, so a semi-supervised normality model is chosen, supported by a smaller set of confirmed fault records.

A physics-informed baseline describes expected rotational orders and thermal behaviour. The machine-learning model estimates a residual health score after accounting for speed and load. Confidence is reduced when the operating condition lies outside the training data. Maintenance records are linked to the data so that alarms can be evaluated against inspection findings rather than against assumptions.

Approach and solution

Feature extraction and first-level anomaly scoring are implemented at the edge to meet latency and availability requirements. The cloud platform stores summarized spectra, health indicators, and event windows for fleet analysis. The system initially operates in shadow mode for three months. Engineers compare its alerts with fixed thresholds and maintenance inspections. A two-stage alarm is then adopted: a persistent moderate score requests inspection at the next planned opportunity, while a high-confidence severe score triggers a controlled reduction in spindle speed and alerts the supervisor.

A conventional drive controller remains responsible for current and speed regulation. The AI system cannot directly bypass torque and speed limits. Model updates require data review, offline testing, signed deployment, and rollback capability. The interface displays the dominant frequency bands, relevant operating context, confidence level, and recommended inspection steps.

Outcome

In this illustrative case, the important outcome is not the replacement of the maintenance team by an algorithm. It is a better allocation of attention. Context-dependent monitoring reduces nuisance alarms, while the safety architecture prevents an uncertain prediction from becoming an unbounded control action. The system also reveals that sensor synchronization and maintenance labelling were initially greater barriers than model selection.

Learning points

The case demonstrates five principles: establish a reliable data foundation; distinguish anomaly detection from diagnosis; use edge computing for time-sensitive functions; retain conventional control and independent safety limits; and validate through shadow operation, human review, and controlled deployment.

Challenges, limitations, and common mistakes

The most common mistake is to begin with an attractive algorithm rather than a well-defined engineering decision. The question should be whether the system must detect a fault, estimate a parameter, optimize a set point, or authorize an action. Each purpose demands different data, metrics, latency, and acceptable risk.

Other limitations include insufficient failure examples, nonstationary operation, sensor degradation, hidden confounding variables, class imbalance, and weak labels. Correlation does not establish causation: a model may associate a particular production shift with faults without identifying the mechanical cause. High-dimensional models may also be difficult to maintain when components, materials, software, or operating policies change.

Overfitting, leakage, and distribution shift can make laboratory performance misleading. An algorithm that performs well on one machine may not transfer to another because of mounting conditions, manufacturing tolerances, or calibration differences. Excessive dependence on proprietary platforms can create vendor lock-in and obstruct auditability. Conversely, insisting on complete interpretability may prevent use of a model whose predictions are demonstrably safe and useful; the appropriate balance depends on consequence, uncertainty, and available safeguards.

Governance, ethics, and professional responsibility

AI-enabled machines raise questions of accountability. If an autonomous system causes damage, responsibility cannot be assigned vaguely to “the algorithm.” Roles must be defined across equipment manufacturers, integrators, software developers, plant owners, supervisors, and operators. Decisions, model versions, overrides, and alarms should be logged sufficiently to support investigation.

Worker monitoring and production analytics may expose personal or behavioural information. Data collection should be proportionate to the engineering purpose, access-controlled, transparent, and consistent with applicable privacy obligations. Workers should not be penalized solely on the basis of an unvalidated prediction. Human dignity, accessibility, and the effects of automation on skill and employment should be included in deployment planning.

Fairness is relevant when AI affects work allocation, performance evaluation, or access to training, even if the machine itself is the primary object of control. Engineers should communicate uncertainty honestly, document assumptions, and refuse deployment when evidence is inadequate for the claimed safety or performance level.

Current developments and best practices

Emerging systems increasingly use foundation models for engineering documentation, multimodal inspection, natural-language interfaces, and cross-asset reasoning. Federated learning can train across sites while retaining raw data locally, although it introduces challenges involving update security, statistical heterogeneity, and verification. TinyML and specialized accelerators allow inference on low-power sensors. Physics-informed neural networks, differentiable simulation, and operator-learning methods seek to combine generalization with physical structure.

These developments should be approached through disciplined engineering practice. A robust implementation should:

• define the operational purpose, constraints, and failure consequences;
• establish sensor, timestamp, calibration, and data-lineage requirements;
• choose the simplest model that satisfies the need;
• combine learned components with physical constraints and conventional safeguards;
• separate training, validation, testing, and deployment responsibilities;
• quantify uncertainty and specify out-of-distribution behaviour;
• keep safety-critical loops local and deterministic where necessary;
• secure the entire data, software, and model supply chain;
• use shadow mode, simulation, hardware-in-the-loop, and staged pilots;
• monitor drift, performance, alarms, and human overrides after deployment; and
• maintain version control, documentation, retraining criteria, and rollback plans.

The mature view of Industry 4.0 is therefore not the indiscriminate addition of AI. It is the deliberate integration of mechanical insight, control engineering, computation, communication, human expertise, and governance.

Chapter Summary

AI-enabled machines combine physical dynamics with sensing, computation, communication, learning, and human supervision. Machine learning can identify nonlinear or changing dynamics, detect anomalies, estimate health, and support maintenance and optimization. Supervised, unsupervised, hybrid, and reinforcement-learning methods have different data and safety requirements. Hybrid models and hierarchical control are especially valuable because they preserve physical constraints and established low-level regulation.

Edge computing supports low-latency and resilient operation, whereas cloud platforms support storage, fleet analytics, and model training; hybrid architectures commonly provide the best balance. Connected and autonomous machines require explicit authority, timing, fallback, and safe operating-domain definitions. Effective human–machine collaboration depends on meaningful explanations, calibrated confidence, training, and override procedures.

Data quality, uncertainty, validation, cybersecurity, and lifecycle governance determine whether an AI system is trustworthy. Safe deployment requires simulation, hardware-in-the-loop testing, shadow operation, staged pilots, independent safety limits, secure updates, monitoring, and rollback. The professional engineer must evaluate not only predictive performance but also consequences, accountability, privacy, security, maintainability, and ethical responsibility.

Key Terms

Cyber-physical system: An integrated system in which physical processes interact continuously with computation and communication.

Industry 4.0: The connected, data-driven integration of automation, CPS, analytics, and intelligent manufacturing technologies.

Digital twin: A computational representation of a physical asset that is updated using operational data.

System identification: Estimation of a dynamic model from measured inputs, outputs, and operating conditions.

Hybrid model: A model combining physical equations with data-driven components.

Anomaly detection: Identification of behaviour that deviates from an expected normal pattern.

Fault diagnosis: Determination of the likely source or type of a machine fault.

Prognostics: Prediction of future degradation or remaining useful life.

Concept drift: Change over time in the relationship between variables represented by a model.

Data drift: Change in the distribution of input data presented to a model.

Reinforcement learning: Learning a policy through interaction between an agent, environment, actions, and rewards.

Safe set: The set of states and actions satisfying defined physical and operational constraints.

Edge computing: Processing data near the machine or data source rather than relying on a remote platform.

Cloud analytics: Centralized, scalable computation and storage performed on network-accessible platforms.

Connected control: Coordination of control functions through communication among machines or systems.

Explainability: The degree to which a model’s output can be understood and justified.

Shadow mode: Deployment in which a model produces predictions but does not influence machine actions.

Concept drift: Change in the statistical relationship learned by a model during operation.

Defence in depth: Use of multiple independent protective layers against failures and cyber threats.

Graceful degradation: Controlled reduction of functionality when uncertainty, faults, or resources become unacceptable.

Review Questions

Short Answer Questions

1. What distinguishes a cyber-physical system from a conventional automated machine?

2. Why can random splitting of time-series data produce misleading validation results?

3. Define anomaly detection and explain why an anomaly score is not necessarily a diagnosis.

4. State two advantages of edge analytics for machine control.

5. What is meant by shadow-mode deployment?

Long Answer Questions

1. Explain how hybrid physics–machine-learning models can improve system identification for nonlinear machines.

2. Discuss the design and validation of an AI-based predictive-maintenance system.

3. Explain the risks of applying reinforcement learning directly to a physical machine and describe safer alternatives.

4. Compare edge, cloud, and hybrid architectures for industrial control applications.

5. Discuss how explainability, cybersecurity, and governance affect the professional deployment of autonomous machines.

Analytical / Application Questions

1. A robot controller performs well in simulation but becomes unstable after deployment because of communication delay and unmodelled friction. Diagnose the likely causes and propose a staged corrective strategy.

2. Design a validation plan for an AI system intended to reduce spindle speed when bearing degradation is suspected. Identify data, performance, safety, and rollback criteria.

3. An anomaly detector generates many alarms during product changeovers but misses a slow temperature-related failure. Analyze the data and modelling issues involved and recommend improvements.

22. Multiple Choice Questions with Answers

Question 1

Which feature most directly distinguishes a cyber-physical machine from a conventional automated machine?

A. It contains mechanical components.

B. It uses a feedback sensor.

C. It integrates physical operation with computation, communication, and digital models.

D. It requires an electric actuator.

Correct Answer: C. It integrates physical operation with computation, communication, and digital models.

Question 2

Why is a hybrid model often preferred to a purely black-box model for safety-relevant machine dynamics?

A. It eliminates the need for measured data.

B. It can incorporate physical constraints and known structure.

C. It guarantees zero prediction error.

D. It requires no validation.

Correct Answer: B. It can incorporate physical constraints and known structure.

Question 3

Which metric is especially informative when machine faults are rare and missed detections are costly?

A. Training-set accuracy alone

B. Number of model parameters

C. Precision, recall, and cost-weighted error measures

D. Cloud storage capacity

Correct Answer: C. Precision, recall, and cost-weighted error measures

Question 4

What is the principal safety concern with unrestricted reinforcement-learning exploration on a physical machine?

A. The algorithm cannot represent sequential decisions.

B. Exploration may produce damaging or unsafe actions.

C. Rewards cannot be defined mathematically.

D. The machine will always operate more slowly.

Correct Answer: B. Exploration may produce damaging or unsafe actions.

Question 5

Which function is generally most appropriate for edge computing?

A. Fleet-wide historical benchmarking

B. Long-term enterprise data warehousing

C. Fast local anomaly detection and safety-related inference

D. Cross-site strategic planning only

Correct Answer: C. Fast local anomaly detection and safety-related inference

Question 6

A model is deployed in shadow mode when it:

A. Has been trained without labels.

B. Runs on encrypted data only.

C. Produces predictions that are recorded but do not control the machine.

D. Is operated exclusively in the cloud.

Correct Answer: C. Produces predictions that are recorded but do not control the machine.

Question 7

Which situation is an example of concept drift?

A. A sensor’s timestamp is accidentally deleted.

B. The relationship between vibration and degradation changes after a redesign.

C. A data file is compressed.

D. The same normal operating data are copied twice.

Correct Answer: B. The relationship between vibration and degradation changes after a redesign.

Question 8

Which deployment practice best supports safe model updating?

A. Replacing the model immediately whenever new data arrive.

B. Using unsigned updates to reduce delay.

C. Versioning, testing, approval, signed deployment, and rollback capability.

D. Disabling all monitoring after commissioning.

Correct Answer: C. Versioning, testing, approval, signed deployment, and rollback capability.

Question 9

Why should a connected control system define local fallback behaviour?

A. Communication delay or loss can make remote commands unsafe or unavailable.

B. Cloud systems cannot store data.

C. Local controllers cannot measure physical variables.

D. Fallback behaviour removes the need for cybersecurity.

Correct Answer: A. Communication delay or loss can make remote commands unsafe or unavailable.

Question 10

Which human–machine interface practice most effectively reduces automation bias?

A. Showing only the final recommendation.

B. Hiding model uncertainty to avoid confusion.

C. Providing the recommendation, confidence, relevant evidence, and escalation options.

D. Preventing operators from overriding the system.

Correct Answer: C. Providing the recommendation, confidence, relevant evidence, and escalation options.