Table of Contents
Introduction
Before addressing the integration of predictive maintenance and artificial intelligence, it is worth making a few brief introductory remarks. The thematic areas of predictive maintenance and artificial intelligence are increasingly interconnected in modern industrial processes.
Regardless of whether or not robotic machines are present in the company, those who work in production already know: unplanned downtime is never just a technical problem. And this is regardless of the level of industrial automation achieved.
A sudden breakdown generates a cascade of negative consequences: it is a missed delivery, a customer to call, a narrowing margin.
The problem is certainly not new; it is simply becoming increasingly expensive. The cause? More often than not, it is a lack of maintenance. It is a problem we have been dragging along for years. Due attention is not paid to maintenance, and when productivity is needed, we find ourselves chasing problems and improvising temporary solutions.
Advanced technologies, such as predictive maintenance and artificial intelligence, can help reduce downtime and optimize production processes.
A 2024 Siemens study estimates that large automotive companies can lose up to $695 million per year due to unplanned downtime, a cost that has grown by 150% compared to five years earlier. The 500 largest global companies lost an average of 11% of their annual turnover for the same reason.
For decades, there have been two main responses to solve the problem, variously designed and implemented in Italian, European, and global industrial realities.
Reactive maintenance: you wait for something to break, then you intervene. Investments have been made in huge spare parts warehouses and 24-hour on-site assistance services. This works fine, as long as you are lucky! And luck is not always on the entrepreneur’s side!
Preventive maintenance: scheduled interventions, regardless of what is actually happening on the system. Better, but you often end up replacing components that would have lasted for months. Other times, this approach means postponing an intervention on something that is about to fail. It works if you are good at planning interventions. With experience in machines and processes, preventive maintenance truly performs miracles. But competent people are needed at all levels; a single good maintenance technician is not enough!
Predictive maintenance and artificial intelligence: a new approach.
Combining the concepts of predictive maintenance and artificial intelligence allows for a change in the logic described above, enabling the introduction of new ways of working and thinking about maintenance. The introduction of artificial intelligence concepts into maintenance revolutionizes efficiency, and this translates—or rather, can translate—into staggering numbers. For now, there is no shortage of studies from major names and a few practical projects.
Predictive maintenance and artificial intelligence: some studies.
Some of the most interesting studies discussing predictive maintenance and artificial intelligence, with specific reference to potential benefits, provide very promising insights.
McKinsey estimates reductions in unplanned downtime of up to 50% and maintenance cost cuts between 10% and 40%.
IBM speaks of a reduction in failures of up to 70% in contexts where predictive maintenance and artificial intelligence are integrated at the production level.
Deloitte estimates an average ROI of 10:1 within two years of implementation.
In theory, predictive maintenance and artificial intelligence promise staggering numbers.
And we could further add:
According to Precedence Research and Grand View Research, investments in predictive maintenance and artificial intelligence are estimated to be around $14 billion in 2025 and will grow to $94 billion by 2035, with annual growth rates around 26-28%.
Suffice it to say that in 2024, over 38% of large global companies had already launched pilot projects in this area. It is therefore unwise to consider predictive maintenance and artificial intelligence as niche, futuristic concepts or, worse, as exclusively separate entities.
Predictive maintenance and artificial intelligence: how it works in practice.
The system combining predictive maintenance and artificial intelligence is data-driven. It is a system that requires data to function and “speaks” through data.
Data must come from production: sensors, PLCs, robots, vision systems—every element of the plant produces continuous information. Data is collected on temperatures, vibrations, consumption, processing parameters, and more.
Information that already existed, but was historically ignored or archived without being read. This data is then analyzed by machine learning algorithms, which look for anomalies, recurring patterns, and correlations between variables.
The difference compared to traditional monitoring is that the models are not static: they learn over time, adapt to the evolution of the plant, and refine their predictions.
The output is not a dashboard of alarms. It is a specific prediction: “this component will have problems within a certain number of working hours,” accompanied by operational instructions on what to do, where, how, and with what priority if there are relevant sequences to consider.
This is the step that matters: not knowing that something is wrong, but knowing what to do before something goes wrong, blocking production and undermining productivity. Naturally, these concepts apply not only to industrial robotics in the classic sense but also to the field of collaborative robots – cobots.
The reality is that predictive maintenance and artificial intelligence systems are changing the paradigm radically and definitively: it will be impossible, and unthinkable, to reverse course. We are moving, once and for all, from talking about predictive maintenance to prescriptive maintenance: not just “I predict the failure,” but “I tell you how to avoid it, in the right order considering the correct priorities.”
Edge computing is accelerating this transition, bringing analytical capabilities closer to the machine, reducing latency and dependence on cloud connectivity. This is because the combination of predictive maintenance and artificial intelligence requires distributed computing power, necessary to train AI algorithms and make them operate reactively.
Why many projects fail to take off
A significant portion of projects based on predictive maintenance and artificial intelligence remain stuck in the pilot phase. The reasons are quite recurring.
The most common problem is data quality.
Industrial data is often incomplete, unstructured, and distributed across systems that do not communicate with each other. A model combining the concepts of predictive maintenance and artificial intelligence works well when the data it is trained on is of high quality. If the information base is fragmented, the results will consequently be unreliable.
It is cliché to say, but in practice, this aspect is almost always underestimated. As the saying goes: Garbage IN, Garbage OUT! If I put garbage into my system, I get garbage out!
Then there is the so-called IT/OT gap, which is worth explaining. IT is the world of corporate information systems: ERP, databases, management software. OT is the world of machines: PLCs, SCADA, robots, industrial control systems, often with 15 or 20 years on their shoulders (or perhaps new, but based on technologies that are 15 or 20 years old).
These two worlds were designed in different eras, with different goals, for different purposes. And historically, they never needed to talk to each other. Integrating them is not a software problem: it is a problem of culture, priorities, and investment. IT is designed for maximum openness and data sharing (fast technology changes), while OT is designed for stability, reliability, and continuity over time (very slow technology changes).
Predictive maintenance and artificial intelligence: Trebi’s winning approach.
Even if the PLC is new, it is not necessarily easy to integrate. Choosing the correct PLC and the correct hardware and software architecture for a machine is a complex process that requires skills that manufacturers often lack. At Trebi, we understood this limitation a long time ago, and our machines have been gradually adapted over time; today, they are ready for the challenge.
Today, we make the IT and OT worlds coexist at their best: we use highly reliable and stable PLC systems and architectures, but we integrate them so that IT can collect data for greater efficiency.
These concepts are at the core of our work methodology and how we at TREBI conceive the role of artificial intelligence in modern industrial robotics.
Predictive maintenance and artificial intelligence: Trebi’s winning approach – part two.
At TREBI, the starting point was different from how many approach it: instead of starting from the algorithm, we started from the data.
As we have said, a fundamental point is the data and correct information.
Our systems are equipped with modern PLC systems that generate precise, structured, and continuous data on the actual production process: processing parameters, tool status, robot behavior, process variables. Not raw signals to be deciphered, but information that an AI system can work on reliably.
To solve the IT/OT gap problem, we have adopted standard communication protocols between machines and information systems. Modern protocols that allow every plant to expose its data in a uniform, secure, and interoperable manner. The practical result is that the customer’s IT system receives consistent data in real time, without the need for complex and expensive custom integrations.
The customer uses the data with their preferred analysis tools.
At TREBI, we do not sell analysis software: we provide what makes that software effective. By entering the world of TREBI, you gain access to a world of innovation, selection, and strategic design tailored to the customer.
Most importantly, we provide an industrial knowledge base built over the years, which explains what that data means, which patterns indicate a real anomaly, which variables truly matter, and which are background noise.
This is the part that is difficult to replicate. An algorithm can be bought. Knowing that a certain parameter variation on a specific type of processing anticipates a failure, while on another processing it is completely normal, is built with years of field work. At TREBI, this experience is structured and transferred, so that the customer’s predictive system works well from the start, without having to learn everything from scratch.
A further problem is the distance between “laboratory” AI and “production” AI. Many solutions work well in a controlled environment and stall as soon as they encounter the real variables of a factory: noise in the data, discontinuity in cycles, anomalous behaviors that the model never saw during the training phase.
Finally, there is the issue of skills. Let’s summarize the main points in this regard:
- Performing predictive maintenance with AI requires people who deeply understand both industrial processes and data (and therefore its structure, origin, and nature).
- A data scientist without factory experience is not enough to successfully integrate predictive maintenance and artificial intelligence.
- A maintenance technician who has never worked with models is not enough.
- It takes someone who understands artificial intelligence, maintenance issues, production, machine management, and real life in the factory.
Do such people exist? The reality is that this technology is in the development phase, and these people are currently rare to find. When they do exist, they are training to be capable of operating the system.
Because the reality is that AI systems are not magical; they only enhance people’s skills. If these skills are not present, then the technology does not perform as expected. It is like having a racing car without the right driver: you cannot expect to set the circuit speed record with the first driver who comes along!
A matter of method, not technology
Effectively combining predictive maintenance and artificial intelligence is no longer a bet on the future. It is a strategy that can be implemented today, but it is not a magic wand: you have to make the right choices and implement them correctly.
The most important change never concerns just the technology. It concerns the way of thinking about the factory: no longer as a system to be repaired when it breaks, but as a process to be governed continuously, with data as the primary tool.
Those who achieve the best results will not necessarily be those who adopt the most sophisticated AI. It will be those who manage to integrate it concretely into their processes. The secret to success will be: quality data, adequate skills, and a real understanding of what is happening on the plant.
Predictive maintenance and artificial intelligence: a combination that does not replace industrial experience, it amplifies it… provided that experience truly exists!


