Artificial Intelligence (AI) and Machine Learning

Artificial intelligence and machine learning

From programmed rules to learned patterns

Artificial intelligence (AI) covers methods for tasks such as interpreting images, processing language, planning and solving problems. Machine learning (ML) is the part of AI that builds models from examples. AI is the broader field; machine learning is one approach within it. In industrial automation, a useful application might inspect a product or flag a developing fault rather than hold a conversation.

Consider a PLC instruction: “Start the fan when the temperature exceeds the limit.” The engineer supplies the rule. An ML model instead learns relationships from historical measurements and uses them to assess new measurements. People still choose the data, objectives and evaluation criteria. Automation does not automatically mean AI.

Three terms that belong at different levels

How the concepts fit together
Term What it covers Example
AI Learning methods and other approaches, including rule-based expert systems. Reasoning through a diagnostic problem.
Machine learning Models whose relationships are learned from data. Estimating fault risk from measurements.
Deep learning ML using artificial neural networks with multiple layers. Detecting surface defects in camera images.

Generative AI produces content such as text or images. Computer vision works with visual information, while natural language processing works with human language. These describe capabilities, rather than alternative names for ML. A deep learning image classifier need not generate images. Likewise, AI and robotics overlap without being interchangeable: robotics involves physical mechanisms, while AI can operate entirely in software.

A practical example: Monitoring a motor

Suppose a plant records motor current, temperature, speed and vibration. Sensors provide measurements, a data logger keeps time-stamped records, and SCADA provides operating states and alarm history. Mechanical vibration changes with the operating conditions: a healthy motor under heavy load need not resemble the same motor running lightly loaded.

During training, historical measurements and verified maintenance records help a model learn useful relationships. During inference, the trained model evaluates new measurements. Its output might be “inspection recommended.” That is a reason to investigate, not proof of a failed bearing. Useful data analysis connects the output to physical evidence and the machine's operating conditions.

The working sequence is measurement → model → prediction → verification → action. Deployment does not imply continuous learning: retraining can be a separate, controlled process. IIoT supports connected data collection, and OPC UA supports information exchange. Neither is, by itself, a fault prediction model.

How models learn

Supervised learning uses examples paired with known outcomes. Deciding whether a sample is normal or faulty is classification; predicting a numerical energy consumption value is regression. Incorrect maintenance labels can teach the wrong relationship.

Unsupervised learning looks for structure in data without supplied outcome labels. It can group similar operating states. An unusual pattern may deserve attention, but unusual does not necessarily mean faulty: a new product or operating mode can also look unfamiliar.

Reinforcement learning develops a strategy from the rewards associated with actions and their outcomes. It has applications in robot learning, although trial and error on physical equipment requires carefully controlled conditions. Simulation and operating constraints matter.

Semi-supervised learning combines labelled and unlabelled examples. Self-supervised learning derives training targets from the data itself. Transfer learning adapts an existing model to another task; its performance in the new setting still needs testing.

Why an accuracy figure can be misleading

Illustrative example: A test contains 1,000 records: 990 normal and 10 faulty. A model that always answers “normal” achieves 99% accuracy while detecting none of the faults. Evaluate missed faults and false alarms as well as overall accuracy. The acceptable balance depends on the consequences of each error.

Keep training, validation and test data separate. Training learns the model; validation guides development choices; testing checks the selected result. Do not use the final test set to tune the model. For industrial time series, near-identical samples from the same fault event can leak information across a random split. Separating by time, machine or fault event should reflect the intended deployment.

Overfitting occurs when a model fits its training examples well but performs poorly on new ones. A replacement sensor, a different speed range or a new product can also shift the data the model sees. A successful initial test does not establish permanent reliability.

Where it belongs in an industrial system

Industry 4.0 applications can use ML for inspection, maintenance prioritisation and consumption forecasting. The physical behaviour of sensors and actuators remains an engineering responsibility. A model's recommendation does not define safe speed or temperature limits. Robot safety and other safety functions should not depend on an unqualified, general-purpose AI output.

A cloud connection is optional. A suitable model can run locally on an industrial computer. Assess latency, connectivity, privacy, processing requirements and maintenance costs when deciding where to run it. The Internet of Things connects devices; AI can help interpret the resulting information.

A generative system may produce a convincing technical explanation while recommending an incorrect connection or parameter. Fluent wording is not evidence of correctness. Verify equipment settings against the manufacturer's documentation. For industrial analytics and AI, judge value through reduced downtime, improved quality and the full cost of running the system.

Technical references: IBM — AI, ML and deep learning · Google — learning approaches · Google — training and test data · Google — evaluation metrics · NIST — trustworthiness and risk