Artificial Intelligence and Machine Learning

Introduction and Context

Artificial intelligence (AI) is the field of computing concerned with designing systems that perform tasks commonly associated with human intelligence, such as perception, reasoning, learning, language use, planning, and decision-making. Machine learning (ML) is a major approach to AI in which a system improves its performance by extracting patterns from data rather than relying exclusively on manually written rules.

For mechatronics engineers, AI is not an isolated software topic. It connects sensors, actuators, embedded computers, industrial networks, control systems, mechanical design, and human-machine interaction. An autonomous mobile robot may use computer vision to identify obstacles, machine learning to classify terrain, and a conventional feedback controller to regulate motor speed. A production line may use an ML model to detect defective parts, while a predictive-maintenance system estimates whether a bearing is likely to fail.

AI systems should therefore be understood as engineering systems. Their performance depends not only on algorithms but also on sensor quality, data management, computational resources, physical models, safety constraints, and the operating environment.

Learning Objectives

After studying this chapter, you should be able to:

1. Distinguish AI, machine learning, deep learning, data science, and automation.
2. Explain intelligent agents and their interaction with environments.
3. Compare supervised, unsupervised, and reinforcement learning.
4. Describe the basic operation of neural networks and computer-vision systems.
5. Explain generative AI and distinguish it from predictive AI.
6. Select suitable evaluation metrics and interpret common errors.
7. Identify limitations, ethical risks, and safety requirements.
8. Propose AI applications in robotics, control, maintenance, and industrial automation.

1. Essential Terminology and Conceptual Boundaries

Artificial intelligence, machine learning, and deep learning

AI is the broadest term. It includes symbolic reasoning, search, planning, knowledge representation, optimization, robotics, and learning. An AI system may use explicit rules without learning from data.

Machine learning refers to algorithms that infer a model from examples. In supervised learning, examples include input-output pairs. In unsupervised learning, the algorithm searches for structure in inputs without specified output labels. Reinforcement learning uses feedback in the form of rewards or penalties to learn actions.

Deep learning is machine learning based on neural networks with multiple computational layers. The word “deep” refers to the number and organization of layers, not necessarily to superior performance in every application. Deep models are particularly effective when large datasets and substantial computation are available.

Automation executes a defined process with limited human intervention. It does not necessarily involve AI. A programmable logic controller running a fixed sequence is automated but may not be intelligent in the ML sense. AI becomes useful when the system must interpret uncertain information, adapt to variation, or make predictions.

A useful distinction is between a model and an algorithm. An algorithm is a procedure for learning or computing; a model is the resulting mathematical representation used to produce predictions or decisions. Training constructs or adjusts a model, whereas inference uses the trained model on new data.

Features, labels, parameters, and hyperparameters

A feature is an input variable used by a model. For a bearing-monitoring system, features may include vibration amplitude, temperature, rotational speed, and frequency-domain energy. A label is the target output, such as “healthy” or “faulty.” Parameters are learned from data, such as neural-network weights. Hyperparameters are selected by the designer, such as learning rate, tree depth, or number of hidden layers.

The dataset is commonly divided into training, validation, and test sets. The training set is used to fit parameters. The validation set supports model selection and hyperparameter tuning. The test set provides a final estimate of performance on unseen data. Test data must not influence model design; otherwise, the reported performance is optimistic.

2. Intelligent Agents

An intelligent agent perceives its environment through sensors and acts through actuators. In a robotic system, cameras, encoders, force sensors, and inertial sensors provide percepts. Motors, grippers, valves, and brakes produce actions.

A rational agent selects an action expected to maximize a performance measure, given its percept history, available knowledge, and uncertainty. Rationality does not mean that the agent is perfect. It means that the chosen action is appropriate according to the information and objectives available at the time.

A standard agent description identifies:

• Performance measure: what counts as success, such as safety, accuracy, energy efficiency, or cycle time.
• Environment: the physical or virtual setting in which the agent operates.
• Actuators: mechanisms by which actions affect the environment.
• Sensors: devices that provide observations.

This is often called a PEAS description. For an autonomous warehouse robot, the performance measure may combine collision avoidance, delivery time, battery consumption, and route reliability. The environment is a warehouse containing shelves, workers, and other robots. Actuators include wheel motors and brakes; sensors include lidar, cameras, wheel encoders, and inertial sensors.

Agent environments can be classified as fully or partially observable, deterministic or stochastic, episodic or sequential, static or dynamic, and discrete or continuous. Industrial environments are often partially observable, stochastic, sequential, dynamic, and continuous. These properties affect algorithm choice and safety design.

A practical AI architecture separates four functions:

1. Perception: converts raw sensor signals into useful information.
2. State estimation: combines observations with prior information to estimate the system state.
3. Decision or planning: selects a goal, path, or action.
4. Control and actuation: converts the decision into commands while respecting physical constraints.

AI should not automatically replace feedback control. A learned planner may generate a desired trajectory, while a proven controller tracks that trajectory and handles fast dynamics. This layered arrangement can improve safety and interpretability.

Figure: A perception–decision–action loop for an intelligent mechatronic agent.

3. Machine-Learning Paradigms

Supervised learning

Supervised learning estimates a mapping from inputs to known targets. For a dataset of n examples, a model produces prediction ŷᵢ from input xᵢ and parameters θ. Training minimizes a loss function:

Here, J is the average training loss, L is a loss function, fθ is the model, xᵢ is the input, yᵢ is the target, and n is the number of examples. The parameters θ are adjusted using an optimization method such as gradient descent. This formulation assumes that examples are sufficiently representative and that the loss reflects the engineering objective.

Regression predicts a numerical quantity, such as remaining useful life or motor temperature. Classification predicts a category, such as normal, overload, or bearing fault. Binary classification has two classes; multiclass classification has more than two.

A simple regression model is ŷ = wᵀx + b, where w is a vector of weights, x is the feature vector, and b is an intercept. For a linear model using mean squared error:

MSE has squared units of the target and is sensitive to large errors. The root mean squared error is RMSE = √MSE and has the same units as the target. These measures are applicable to numerical predictions but do not by themselves establish safety or usefulness.

Unsupervised learning

Unsupervised learning uses inputs without supplied labels. Clustering groups observations according to similarity. In k-means clustering, each observation is assigned to one of k clusters, and the centroids are adjusted to minimize within-cluster squared distance:

Here, Cⱼ is cluster j, μⱼ is its centroid, and ||·|| is Euclidean distance. K-means assumes that numerical features are meaningfully scaled and that cluster structure is reasonably represented by compact groups. The choice of k is a design decision, not a fact automatically discovered by the algorithm.

In maintenance, clustering can identify operating regimes or unusual patterns before failure labels are available. Dimensionality-reduction methods can provide compact representations for visualization, noise reduction, or downstream modeling. Unsupervised results require engineering interpretation; a mathematical cluster is not automatically a physically meaningful machine state.

Reinforcement learning

In reinforcement learning, an agent observes a state, selects an action, receives a reward, and transitions to another state. The objective is often to maximize expected discounted return:

Gₜ is the return at time t, rₜ₊ₖ₊₁ is a future reward, and γ, with 0 ≤ γ < 1, is the discount factor. A high γ values long-term outcomes; a low γ emphasizes immediate rewards. Reinforcement learning is useful for sequential decisions, but direct exploration on physical machinery can be unsafe. Simulation, constraints, offline data, and supervisory safeguards are important.

Overfitting and generalization

A model overfits when it learns accidental details of the training data and performs poorly on new data. Underfitting occurs when the model is too limited to represent relevant structure. Generalization is performance on previously unseen but relevant cases.

Regularization adds a penalty or constraint that discourages excessive complexity. Data augmentation, cross-validation, early stopping, simpler models, and careful feature design can also reduce overfitting. In industrial settings, splitting data randomly may be misleading if adjacent samples from the same machine appear in both training and test sets. A time-based or machine-based split may better represent deployment.

4. Neural Networks and Deep Learning

A basic artificial neuron computes a weighted sum followed by an activation function:

The input vector is x, weights are w, b is a bias, z is the pre-activation value, φ is an activation function, and a is the neuron output. Common activation functions include the rectified linear unit, ReLU(z) = max(0, z), and the sigmoid function, σ(z) = 1/(1 + e⁻ᶻ). A neural network connects many such units in layers.

During training, the network computes a prediction in a forward pass. The loss is propagated backward using the chain rule to calculate gradients. Gradient descent updates parameters according to:

Here, η > 0 is the learning rate and ∇θJ is the gradient of the loss with respect to parameters. The update assumes that the objective is differentiable or is handled with a suitable subgradient method. In practice, data are processed in mini-batches and several passes, called epochs, are performed.

Convolutional neural networks are effective for images because convolutional filters detect local patterns and share parameters across locations. Recurrent and attention-based architectures are used for sequences and multimodal data. Neural networks can approximate complex nonlinear relationships, but their high capacity creates risks of overfitting, distribution shift, and limited interpretability.

5. Computer Vision for Mechatronics

Computer vision converts images or video into information for inspection, localization, navigation, and interaction. A typical pipeline includes image acquisition, calibration, preprocessing, feature extraction or learned representation, inference, post-processing, and action.

Important tasks include:

• Image classification: assigns an image to a category.
• Object detection: identifies objects and their bounding boxes.
• Semantic segmentation: assigns a class to each pixel.
• Instance segmentation: separates individual objects of the same class.
• Optical flow and tracking: estimate motion across frames.
• Pose estimation: estimates the position and orientation of an object or body.

Camera calibration estimates parameters that relate image coordinates to physical geometry. Lens distortion, lighting, vibration, exposure, occlusion, and camera mounting affect reliability. A vision system should be tested under the range of conditions expected in operation, not only under laboratory lighting.

In robot manipulation, vision may locate a component, while force sensing verifies grasp quality. Sensor fusion can be more reliable than dependence on a single camera. The result of perception should include confidence and uncertainty where possible, allowing the controller or supervisor to respond conservatively.

6. Generative AI

Generative AI produces new content—such as text, images, audio, code, designs, or synthetic sensor data—based on patterns learned from training data. Predictive AI estimates an existing target, whereas generative AI samples or constructs outputs from a learned distribution.

Large language models generate sequences by estimating likely next tokens from context. Their outputs can be useful for documentation, code assistance, natural-language interfaces, and troubleshooting. However, fluent language does not guarantee factual accuracy, physical feasibility, or safe control commands. A language model should not be given unrestricted authority over safety-critical actuators.

Generative models can support engineering through synthetic data generation, design alternatives, simulation scenario creation, and operator assistance. Generated data must be validated because it may reproduce bias, omit rare faults, or contain unrealistic correlations. Industrial deployment should use access control, logging, human review, retrieval from approved technical documents, and strict separation between advice and executable control.

7. Model Evaluation and Deployment

Classification metrics

A confusion matrix records true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). Accuracy is:

Precision and recall are:

Precision answers: of the items predicted positive, how many were truly positive? Recall answers: of the truly positive items, how many were detected? In safety or maintenance, a false negative may be much more costly than a false positive, so recall may receive greater emphasis. The choice depends on consequences, inspection capacity, and operating policy.

The F1 score is the harmonic mean of precision and recall:

It is useful when both measures matter, but it does not represent financial or safety cost directly. Receiver operating characteristic and precision–recall curves show how performance changes with the decision threshold. For imbalanced fault data, accuracy can be highly misleading.

Regression and operational evaluation

Regression evaluation may use MAE = (1/n)∑|yᵢ − ŷᵢ|, RMSE, or a domain-specific cost. Mean absolute error has the same units as the target and is less sensitive to extreme errors than MSE. Evaluation should also include latency, memory use, energy consumption, calibration, robustness, maintainability, and failure behavior.

A model is not ready merely because it scores well on a test set. Deployment requires data-quality checks, monitoring for distribution shift, version control, cybersecurity, fallback behavior, and a process for retraining or retirement. The model card or equivalent documentation should record intended use, limitations, training data characteristics, evaluation conditions, and known failure modes.

Worked example: imbalanced fault detection

Suppose a classifier is evaluated on 1,000 machine cycles. It identifies 45 of 50 real faults and incorrectly labels 30 healthy cycles as faulty. Thus TP = 45, FN = 5, FP = 30, and TN = 920.

Accuracy = (45 + 920)/1,000 = 0.965, or 96.5%.

Precision = 45/(45 + 30) = 0.60, or 60%.

Recall = 45/(45 + 5) = 0.90, or 90%.

The high accuracy alone is not sufficient evidence of excellent fault detection. Recall is strong, but 40% of alarms are false positives. Engineers must decide whether the maintenance workload created by these alarms is acceptable and whether missing five faults is tolerable.

8. Applications in Mechatronics

Robotics

AI supports localization, mapping, object recognition, grasp planning, human activity recognition, and adaptive navigation. A robust robot combines learned components with geometric models, collision checking, and feedback control. Safety-rated sensing and emergency-stop systems should remain independent of an experimental ML model where required.

Control systems

ML can estimate unknown dynamics, tune controller parameters, predict disturbances, or provide high-level policies. A learned controller may be effective in simulation but fail when friction, load, temperature, or payload differs from training conditions. Stability, constraint satisfaction, and fail-safe operation must be analyzed rather than inferred from average test performance.

A common engineering pattern is hybrid control: a physics-based controller ensures basic regulation, while ML estimates uncertain terms or selects operating modes. This can reduce data requirements and preserve useful physical structure.

Predictive maintenance

Predictive maintenance uses sensor data to estimate degradation, detect anomalies, diagnose faults, or predict remaining useful life. The workflow includes defining failure modes, selecting sensing locations, synchronizing data, labeling events, handling censored observations, and connecting predictions to maintenance decisions. A technically accurate prediction has little value if no maintenance action follows it.

Industrial automation

Vision inspection, adaptive assembly, scheduling, energy optimization, robotic bin picking, and process-quality prediction are common applications. Integration requires communication with PLCs, robots, supervisory control and data acquisition systems, manufacturing execution systems, and databases. Timing guarantees and cybersecurity are as important as model accuracy.

9. Illustrative Case: Vision-Based Inspection of a Robotic Assembly Cell

Background

Consider a fictional but realistic illustrative case involving a robotic cell that inserts a gasket into a machined housing. Missing or misaligned gaskets cause leakage during final testing. Manual inspection is slow and inconsistent.

Problem

The manufacturer wants to detect three conditions: correct assembly, missing gasket, and visibly misaligned gasket. Parts vary slightly in color and surface finish, and lighting changes during maintenance.

Analysis

A camera is mounted above the fixture. Initial trials show that a simple brightness threshold works under one lighting condition but fails when reflections occur. The dataset contains many normal parts but relatively few defective parts. Randomly splitting adjacent images would overestimate generalization because nearly identical parts could appear in both training and test sets.

Approach and solution

The team installs controlled diffuse lighting, fixes the camera mount, and records images across shifts, batches, and acceptable variations. Images are divided by production batch so that test batches remain unseen during training. A convolutional classifier is compared with a rule-based geometric baseline. The system outputs class probabilities, but a high-risk decision rule sends uncertain cases to a human operator.

The vision result is not connected directly to the robot’s motion command. Instead, it is passed to the cell supervisor. If the result is “defective” or “uncertain,” the supervisor stops the next process step, illuminates an operator indicator, and records the image and batch identifier. A separate presence sensor provides a simple cross-check.

Outcome

The learned model detects most missing and misaligned gaskets, while the independent presence sensor catches some camera failures. Evaluation is performed on later production batches and includes false alarms per shift, missed defects, inference time, and recovery behavior. The system is accepted only after operators confirm that the additional inspections are manageable.

Learning points

1. Data collection and lighting design can matter as much as model selection.
2. Batch-based testing can reveal deployment performance more realistically than random image splitting.
3. Confidence thresholds should support escalation, not conceal uncertainty.
4. Redundant sensing and a safe supervisory layer reduce the consequences of model error.
5. The engineering objective is reliable quality control, not merely a high laboratory accuracy.

10. Limitations, Ethics, and Governance

AI systems inherit limitations from data, models, hardware, and deployment context. Important limitations include:

• Distribution shift: operating conditions differ from training conditions.
• Bias: data may underrepresent people, products, environments, or fault modes.
• Label noise: expert judgments and maintenance records may be inconsistent.
• Uncertainty: a model may be confidently wrong.
• Explainability limits: complex models can be difficult to audit.
• Resource demands: training and inference may require substantial computation and energy.
• Adversarial and cybersecurity risks: inputs or models may be manipulated.
• Automation bias: operators may trust an algorithm without sufficient verification.
• Privacy: cameras, worker data, and production records may contain personal or proprietary information.

Responsible engineering requires purpose limitation, data minimization, access control, traceability, human oversight, documented validation, and incident reporting. Safety-critical functions require hazard analysis, independent safeguards, verification, validation, and compliance with applicable laws and standards. The exact regulatory framework depends on jurisdiction and application; engineers must consult current requirements rather than assume that an AI model is exempt from machine-safety obligations.

AI should be evaluated across its life cycle: problem definition, data collection, development, testing, deployment, monitoring, updating, and retirement. A model that degrades silently can create greater risk than a model that fails visibly.

11. Contemporary Developments and Best Practices

Foundation models, multimodal systems, edge AI, digital twins, synthetic data, and physics-informed learning are influencing industrial applications. Edge AI places inference near sensors, reducing latency and dependence on cloud connectivity. Digital twins can provide simulation environments for testing policies and fault scenarios. Physics-informed approaches incorporate known relationships into learning, potentially improving data efficiency and extrapolation.

Best practices for a mechatronics project include:

1. Define the engineering decision and cost of errors before selecting an algorithm.
2. Establish a baseline using rules, statistics, or a simple physical model.
3. Collect representative data and document its provenance.
4. Separate training, validation, and test data according to deployment conditions.
5. Measure technical, operational, safety, and economic performance.
6. Use uncertainty thresholds, human escalation, and fallback modes.
7. Monitor data quality, drift, latency, and errors after deployment.
8. Secure models, interfaces, logs, and update procedures.
9. Keep the control authority of experimental AI bounded by verified constraints.
10. Communicate limitations clearly to operators and decision-makers.

Chapter Summary

AI is a broad field of computational intelligence, while machine learning learns patterns from data and deep learning uses multi-layer neural networks. Intelligent agents perceive environments, estimate state, decide, and act according to performance measures. Supervised learning predicts labeled outputs; unsupervised learning discovers structure; reinforcement learning optimizes sequential action using rewards.

Neural networks learn parameters by minimizing a loss, commonly through gradient-based optimization. Computer vision supports inspection, localization, navigation, and manipulation, but depends strongly on calibration, lighting, representative data, and uncertainty management. Generative AI creates new content and can assist engineering, but fluent outputs are not guaranteed to be correct or safe.

Evaluation must use appropriate metrics, realistic data splits, operational costs, and deployment tests. Accuracy alone is inadequate for imbalanced or safety-sensitive problems. In robotics, control, predictive maintenance, and automation, AI is most reliable when integrated with physical models, feedback control, redundancy, supervision, and documented safeguards. Ethical, cybersecurity, privacy, regulatory, and lifecycle considerations are essential parts of professional AI engineering.

Key Terms

Artificial intelligence: The field of building systems that perform tasks associated with intelligence.

Machine learning: Algorithms that learn patterns or mappings from data.

Deep learning: Machine learning using neural networks with multiple layers.

Intelligent agent: A system that perceives an environment and takes actions.

Percept: An observation received through sensors.

Feature: An input variable used by a model.

Label: The target output associated with a training example.

Parameter: A model quantity learned from data.

Hyperparameter: A design setting selected by the developer.

Supervised learning: Learning from input-output examples.

Unsupervised learning: Learning structure from unlabeled inputs.

Reinforcement learning: Learning actions from rewards and penalties.

Overfitting: Fitting training details that do not generalize.

Neural network: A layered computational model of weighted units and activations.

Computer vision: Computing methods for interpreting images or video.

Generative AI: AI that produces new content from learned patterns.

Precision: The proportion of positive predictions that are correct.

Recall: The proportion of actual positives that are detected.

Distribution shift: A change between training conditions and deployment conditions.

Model drift: Degradation caused by changing data or operating conditions.

Review Questions

5 Short Answer Questions

1. Distinguish artificial intelligence, machine learning, and deep learning.
2. What are the four components of a PEAS description?
3. Explain the difference between a feature, a label, a parameter, and a hyperparameter.
4. Why can accuracy be misleading for fault detection?
5. State two reasons why AI should not automatically replace a feedback controller.

5 Long Answer Questions

1. Explain the perception–decision–action architecture of an intelligent robotic agent and relate each stage to a warehouse robot.
2. Compare supervised, unsupervised, and reinforcement learning, including one mechatronics application of each.
3. Describe neural-network training using a loss function and gradient descent. Define the variables in the update equation.
4. Explain a computer-vision inspection pipeline and discuss the effects of calibration, lighting, occlusion, and data splitting.
5. Discuss ethical, safety, privacy, and cybersecurity issues in deploying AI in an industrial cell.

3 Analytical / Application Questions

1. A maintenance classifier produces TP = 72, FN = 8, FP = 18, and TN = 902. Calculate accuracy, precision, and recall. Which metric would you prioritize if missing a fault is much more costly than sending an unnecessary inspection?
2. Design a hybrid control architecture for a mobile robot in which machine learning assists but does not directly dominate safety-critical actuation. Identify sensors, learned functions, conventional control functions, and fallback behavior.
3. Propose a validation plan for a vision system used across three factories with different lighting and camera mounts. Explain how you would prevent data leakage, test distribution shift, monitor deployment, and decide when retraining is justified.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes machine learning from traditional rule-based automation?

A. Machine learning always requires a physical robot.

B. Machine learning infers patterns or mappings from data.

C. Machine learning cannot use numerical inputs.

D. Machine learning eliminates the need for sensors.

Correct Answer: B. Machine learning infers patterns or mappings from data.

Question 2

In an intelligent agent, which component converts observations into an estimate of the current situation?

A. Actuator

B. State-estimation process

C. Performance measure

D. Power supply

Correct Answer: B. State-estimation process

Question 3

Which learning paradigm uses labeled input-output examples?

A. Supervised learning

B. Unsupervised learning

C. Reinforcement learning

D. Random search

Correct Answer: A. Supervised learning

Question 4

A model has TP = 45, FN = 5, FP = 30, and TN = 920. What is its recall?

A. 60%

B. 90%

C. 96.5%

D. 98.0%

Correct Answer: B. 90%

Question 5

What is the principal risk of overfitting?

A. The model performs well on training data but poorly on new data.

B. The model always has too few parameters.

C. The model cannot calculate a loss.

D. The model requires no validation data.

Correct Answer: A. The model performs well on training data but poorly on new data.

Question 6

Which computer-vision task assigns a category to each pixel?

A. Image classification

B. Object detection

C. Semantic segmentation

D. Token generation

Correct Answer: C. Semantic segmentation

Question 7

In gradient descent, what does the learning rate control?

A. The number of classes in the dataset

B. The size of the parameter update

C. The camera resolution

D. The number of test examples

Correct Answer: B. The size of the parameter update

Question 8

Why is random splitting of adjacent industrial sensor samples potentially misleading?

A. It prevents any model from being trained.

B. It can place highly similar samples in both training and test sets.

C. It guarantees underfitting.

D. It removes all labels.

Correct Answer: B. It can place highly similar samples in both training and test sets.

Question 9

Which statement about generative AI is correct?

A. It only measures motor temperature.

B. It produces new content based on learned patterns.

C. It is identical to a fixed PLC sequence.

D. Its fluent output is guaranteed to be factually correct.

Correct Answer: B. It produces new content based on learned patterns.

Question 10

For a safety-sensitive fault detector, why might recall receive greater emphasis than accuracy?

A. Recall measures the number of model parameters.

B. Recall measures how many actual faults are detected.

C. Accuracy has no mathematical definition.

D. Recall eliminates false positives completely.

Correct Answer: B. Recall measures how many actual faults are detected.