|
Beyond generative AI: how Machine Learning enters industrial processes
Approaches and applications for quality control, fault detection and the regulation of machinery and processes.
When people talk about artificial intelligence, their first thought often turns to generative AI, its most visible and accessible form, capable of producing content and analyses from a simple prompt. The challenge changes, however, when we want to integrate Artificial Intelligence by understanding what actually happens in an industrial process and being able to intervene in a practical and automated way.
Problem, data, decision: what is Machine Learning?
One of the tools available to address this challenge is Machine Learning. It is a branch of artificial intelligence that allows a series of purpose-built algorithms to learn from data by recognising relationships and patterns without having to define a rule in advance for every possible situation.
In an industrial process, images, signals and production parameters are all data that can become useful information for identifying anomalies, predicting behaviour and supporting decisions. The possible applications are broad, ranging from quality control to process regulation.
However, there is no single approach that works for every application. The choice of the right framework to use depends on the problem to be solved, the available data and the role the algorithm will need to play within the process.
Supervised Learning: learning from previously classified examples
Supervised Learning is the most direct approach when examples are available that have already been assigned a classification: a conforming or defective product, an acceptable or unacceptable weld, an intact or damaged package. By training the algorithm with these cases, it learns to recognise the characteristics associated with each classification, acquiring the ability to assess new cases it has never encountered before.
Taking a machine vision system with images captured by a camera as an example, the model can learn to distinguish a conforming product from a defective one by observing differences in shape, colour or surface. The same logic can be applied to temperatures, pressures, vibrations and electrical signals to assess the quality of a processing operation or the condition of a machine.
For the system to be reliable, however, the examples must represent what actually happens during production, including normal variations in lighting, raw material, format or speed.
Examples of integration
- In the food sector, it can support product selection based on size, ripeness, integrity or the presence of visible contaminants.
- In packaging, it can check that a package is correctly sealed, the position of a label, the readability of a code or the presence of deformations.
- In the pharmaceutical sector, it can contribute to the inspection of tablets, capsules and containers, or to the classification of results from process sensors.
- In metal and material processing, it can help recognise surface defects, discontinuities, dimensional errors or operating conditions associated with a non-conforming result.
Unsupervised Learning: recognising deviations from normality
In many industrial processes, defects are rare. We may therefore have thousands of conforming examples, but very few samples representing possible anomalies. In these cases, instead of teaching the algorithm to recognise every individual defect, we can have it learn the characteristics of normal operation.
This approach is used in anomaly detection and is based on the Unsupervised Learning framework, which analyses data without an initial classification, or on models trained mainly with conforming examples. The system thus learns to recognise the expected behaviour of the technology on which it is installed and to flag deviations from it, even when the anomaly has never been observed before.
A difference, however, does not necessarily correspond to a defect. A new batch of raw material, a format change or an intentional adjustment can produce unusual data without indicating a problem. For this reason, particularly in the early stages, anomaly detection is used as a support tool: it flags cases that need further investigation and allows process experts to confirm or correct the assessment. This feedback can then be used to progressively improve the model.
Examples of integration
- On a packaging line, it can analyse images of conforming packaging and flag deformations, missing elements or configurations that have never been observed before.
- In a production plant, it can simultaneously monitor temperatures, pressures, flow rates and power consumption, detecting anomalous combinations even when each value, considered individually, still appears to be within the limits.
- In the chemical and pharmaceutical sectors, this approach can help identify process drift, unexpected variations between batches or behaviour that is inconsistent with production history.
- In thermal systems, it can recognise an anomalous temperature distribution, an unusual increase in consumption or a change in the material’s response to the applied energy.
- More generally, it is useful in predictive maintenance. By analysing vibrations, noise, current draw or other signals, the model can flag a gradual departure from usual operating conditions before the phenomenon becomes an evident fault.
Reinforcement Learning: learning to intervene in the process
While Supervised Learning and Unsupervised Learning help recognise or predict a condition, Reinforcement Learning allows the system to learn how to intervene. The algorithm, in this case referred to as an agent, observes the state of the process, changes one or more parameters and evaluates the result against an objective.
By repeating this cycle, it develops a strategy capable of adapting its decisions to the conditions detected, keeping the quality of what it must produce or control stable, reducing consumption or finding the best balance between efficiency and productivity.
In a thermal process, for example, the agent can learn to modulate the energy delivered based on the measured temperature, the product’s properties and the observed response, seeking to achieve the desired profile without exceeding safety or quality limits.
Why learning does not take place directly on the line
To identify the best strategy, the agent using Reinforcement Learning must experiment with different actions. Doing so directly on the line could produce waste or unsafe conditions. For this reason, learning is often developed through a process simulation or a digital twin.
The transition to the real system requires progressive checks, supervision and mechanisms that prevent the agent from leaving the permitted operating range. In an industrial setting, model performance is not the only criterion: stability, traceability, robustness and the ability to return the process to a safe condition also matter.
Examples of integration
- In food and pharmaceutical drying, where heat input must accompany the gradual loss of moisture, it could guide power regulation in microwave, radiofrequency or induction systems according to the material’s response.
- In packaging, it could instead adapt the sealing temperature to changes in line speed to help keep seal quality consistent.
- In chemical processes, it could regulate the feed to a reactor based on how the reaction evolves, seeking a balance between productivity and energy consumption within the defined operating constraints.
Which framework should you choose?
Choosing the right framework always starts with understanding the problem we want to solve and the role the system will need to play in the process.
Recognising known defects and identifying unusual behaviour may require different approaches, while intervening directly on the machine also requires evaluating the effects of decisions and defining the limits within which the algorithm can act.
The available data provide initial guidance, but even a large archive can prove of little use if it does not represent actual production conditions or if there is no in-depth understanding of the problem and what is to be achieved.
For this reason, choosing the most suitable model stems from a process of analysis and verification that combines Machine Learning expertise with knowledge of the physical phenomena of the process, making it possible to identify the quantities to measure and the acceptable margin of error in order to assess how well the algorithm’s capabilities meet actual production needs and design its integration into the plant accordingly.
Do you have a project in mind? Get in touch with our team.
We support companies developing innovative technologies throughout all the critical stages of the innovation process: from analysing and understanding the problem through to formulating and implementing technically sound and verifiable solutions.
Do you have a project in mind? Get in touch with our team.
We support companies developing innovative technologies throughout all the critical stages of the innovation process: from analysing and understanding the problem through to formulating and implementing technically sound and verifiable solutions.
Get the latest news
Subscribe to our newsletter and receive all the latest updates on innovations, stories and events.
Do you require any information?
We are at your disposal! Contact us and we will respond to your request as soon as possible.