Why we talk about predictive maintenance in the foundry today
In recent years, the topic of predictive maintenance in the foundry has become central to the debate on the manufacturing industry. Conferences, articles, supplier brochures: it seems that without an artificial intelligence system applied to maintenance, you are already out of the market.
The reality, as often happens, is much more multifaceted.
To understand it, let’s start with the fundamentals. There are three main approaches to industrial maintenance in the foundry, without even involving concepts related to artificial intelligence.
- Breakdown maintenance: intervention occurs when something breaks. Simple and inexpensive in daily management, but exposed to the risk of unplanned downtime. It is the shortcut taken to avoid planning. If you are lucky, you spend little and resolve it quickly, but studies show that in the foundry, this is a losing strategy by definition.
- Preventive maintenance: intervention occurs at regular intervals, regardless of the actual state of the component. It reduces sudden breakdowns but wastes resources on components still in good condition. Many times, parts that are not yet exhausted are replaced. This is currently the most widespread methodology.
- Predictive maintenance: intervention occurs when data indicates that a component is degrading, before the failure occurs. It is the theoretical ideal, but it requires solid preparation: data, analysis and, above all, in-depth knowledge of the process. It introduces a new way of working, a change in organization, and a change in mindset to the factory.
In the foundry, the difference between these three approaches is not just a matter of repair costs. Unplanned downtime on a deburring or finishing line can block the entire downstream production, compromise already confirmed deliveries, and generate scrap on already processed parts. Indirect costs often far exceed direct ones.
The real problem, therefore, is not “doing artificial intelligence” or “collecting data”: it is avoiding unplanned downtime and quality defects. Technology is a tool, not the end goal. This is fundamental to keep in mind when talking about predictive maintenance in the foundry.
An often underestimated element: modern plants already collect a huge amount of useful data today. PLCs, drives, industrial robots, vision systems — all these machines produce information continuously. The point is knowing how to read them and being able to decode the signals to perform predictive maintenance in the foundry.
At TREBI, we design robotic machines where data collection and its reorganization are integrated to have an AI-ready system. We promote preventive maintenance in the foundry (and beyond) and prepare our machines to support this and other implications related to artificial intelligence. Given these premises, the implementation of predictive maintenance systems in the foundry would seem to be just around the corner…
Predictive maintenance in the foundry: what are the peculiarities
Those who have never worked in a foundry struggle to fully understand how much the environment conditions every technical choice. The foundry is a highly variable process. Those who know it well know how quickly variables change and how complex it is to keep them under control.
This has direct consequences on the quality of the data that can be collected: the environment is difficult by definition, and obtaining clean data to process is often far from guaranteed.
To further complicate the picture, foundries often feature very heterogeneous production situations: small batches, substantially different productions, materials with characteristics that change from melt to melt. This makes it difficult to define a unique “normal behavior” for the machine, which is instead the basis on which any predictive system is built. This basic situation cannot be modified simply by planning the introduction of predictive maintenance in the foundry. There are constraints that must be observed. Ignoring this element means starting on the wrong foot… and getting nowhere!
Then there is the issue of coexistence between old and new machines. In many plants, a press purchased twenty years ago works alongside a latest-generation robotic machine. Connecting these worlds requires specific skills and necessarily tailor-made solutions.
Finally, when planning the introduction of predictive maintenance in the foundry, the experience of the operators must be considered. In the foundry, those who have worked on a plant for years “feel” when something is wrong: the noise changes, vibrations increase, the finished part looks different. This wealth of knowledge is precious and difficult to replace with an algorithm.
TREBI, having an extremely in-depth knowledge of the environment, ensures that data is comparable and reliable. The collected data comes from appropriately installed and protected sensors and is prepared for the acquisition system with processing entrusted to artificial intelligence.
Introducing or enhancing predictive maintenance in the foundry means facing a process made of technical premises, analysis of the specific situation, experience, and measurable effectiveness of the adopted solutions. In this article, we delve into some aspects related to the concepts of artificial intelligence and predictive maintenance.
Where it really makes sense to perform predictive maintenance in the foundry
Not all components of a plant deserve the same level of attention. This simple observation is very important, especially for companies approaching predictive maintenance in the foundry for the first time. A rational approach to predictive maintenance starts with identifying critical points: those components whose failure causes prolonged downtime, difficulties in sourcing spare parts, or degradation of product quality.
It is necessary to understand what is strategic and what is not. It is useless to monitor a marginal, easily replaceable process. It is necessary to monitor bottlenecks and equipment whose downtime has a high cost.
Focusing on predictive maintenance in the foundry, it will be essential to monitor the machines and devices that heavily influence productivity. Typically:
- Casting lines: if they do not work or generate scrap, the company’s profitability plummets.
- Production support systems: compressors, cooling systems. Everything that can immediately block production must be kept under control.
What data can be monitored to feed predictive maintenance processes in the foundry
Before talking about algorithms and artificial intelligence, it is useful to take an inventory of the data that a foundry plant can provide and that is therefore available to give life to predictive maintenance processes in the foundry. It is often surprising to discover how much information is already available without adding a single sensor.
- Cycle times: this is the easiest and most important parameter to monitor. A progressive slowdown in the cycle, even by a few seconds, can indicate degradation in progress. Variable cycle time is often a universal indicator of a problem generated by the machine or its use.
- Vibrations: variations in the vibration profile almost always precede a mechanical failure.
- Motor current draw: one of the most informative and least exploited parameters. A motor that draws more than normal is working against abnormal resistance — it could be a worn tool, a mechanical constraint, or a problem in the processed material.
- Pressures in hydraulic and pneumatic circuits: even small variations can indicate leaks, clogged filters, or valves starting to fail.
- Quality of the processed part: industrial vision systems can detect quality drifts even before the product actually becomes scrap. It is an indirect but powerful form of monitoring the plant’s status.
If you have read this far, we’ll tell you a secret: do you know which information is extremely precious but frequently ignored?
Temperatures! Whether they are from motors, drives, or hydraulic fluids, a progressive and unjustified increase is almost always a signal of something not functioning correctly.
In TREBI machines, much of this data is already ready to be read and analyzed for anomalies. With TREBI robotic machines, it is easy to introduce or expand predictive maintenance in the foundry!
Sensors, PLCs, and data collection: how predictive maintenance in the foundry is powered.
Once the data to be monitored has been identified, the next step is to organize the collection infrastructure. Here too, the guiding principle should be simplicity: the best solution is not the most sophisticated, but the one that works reliably in the real foundry environment.
In other words, implementing predictive maintenance solutions in the foundry does not mean introducing unnecessary complications that may be fragile and ineffective.
According to TREBI’s experience, the best method is to install sensors directly on the machines, have them pre-processed by the integrated PLC, and collect the pre-processed information through an artificial intelligence system installed at the customer’s site. This architecture guarantees data reliability and accuracy. Predictive maintenance in the foundry must work on qualitatively impeccable data.
One can also choose to connect external sensors directly to the AI system, but this introduces technological compatibility issues: sensors are designed for PLC systems and integrate with difficulty with PC-based systems like predictive maintenance AI platforms. Furthermore, raw data is normally of little significance: it must be reprocessed and filtered.
The role of artificial intelligence in predictive maintenance in the foundry
Artificial intelligence applied to predictive maintenance in the foundry is often presented as a monolithic solution. In reality, there are approaches with very different complexities, costs, and data requirements.
Too little is said about the fact that only by analyzing a company’s management data in an orderly and systematic way can concrete indications be obtained on efficiency and how to orient maintenance to improve it.
At the simplest level, we find analyses based on thresholds and statistics: if the vibration exceeds a certain value, an alarm is triggered; if the average temperature over the last seven days increases by 15%, a notification is generated. These are elementary tools, but surprisingly effective if the thresholds are correctly calibrated based on process knowledge.
At a higher level are trend analyses: you don’t just look at the instantaneous value, but at its evolution over time. A value that remains constantly below the alarm threshold but grows week after week is a signal that deserves attention.
Even higher are anomaly detection techniques and true machine learning: systems that learn the normal behavior of the plant and signal even subtle deviations that an analysis based on fixed thresholds would never detect. However, these AI-based predictive maintenance systems in the foundry require a sufficient amount of high-quality historical data and a training phase that cannot ignore real process knowledge. According to research by McKinsey, AI-based predictive maintenance can reduce maintenance costs by up to 10-25% and unplanned downtime by up to 50% — but only when the basic data is reliable and the process is well understood.
The real limits of predictive maintenance in the foundry
Those who sell predictive maintenance systems tend not to talk enough about specific problems, including, of course, failures. It is instead important to do so because AI projects, especially those oriented towards predictive maintenance in the foundry, are always difficult to make work and to justify initially. Often, miraculous results are expected, which do not arrive not because the system doesn’t work, but because the data is not of sufficient quality or the company is unable to analyze the results correctly.
False alarms are the most common problem. A system that is too sensitive or poorly calibrated continuously generates notifications that technicians quickly learn to ignore. If the system signals false anomalies, the maintenance technician stops using it carefully. This is how a predictive maintenance system in the foundry can be turned into a hazard.
Insufficient or poor-quality data makes any analysis unreliable. Poorly calibrated sensors, installed in non-representative positions, or in environments that are too aggressive, produce statistical noise, not information.
Costs can exceed benefits. Not all components justify the investment in an advanced monitoring system. For a gearbox costing a few hundred euros, replaced in half an hour, a five-thousand-euro monitoring system makes no economic sense. Costs must always be evaluated before performance. The right question to ask before starting to implement a predictive maintenance system in the foundry is: how much does downtime cost us? How much have we spent in past years on inefficiencies or non-conformities? The answer should indicate how much we are willing to invest in a predictive maintenance system.
Technical complexity is a real deterrent. A sophisticated system requires skills to be managed. If these skills do not exist internally, the system will soon be abandoned. A complex system, by definition, tends to be decommissioned over time.
At TREBI, we are convinced of one thing: in many cases, a correctly designed, robust, easily accessible machine with standard and well-documented components is worth more than an extremely sophisticated predictive maintenance system installed on a machine that is difficult to maintain. Design oriented towards reliability and maintainability is the first true tool for predictive maintenance in the foundry.
Before thinking about predictive maintenance in the foundry, therefore, a broad-spectrum reasoning on costs is necessary. If other methods guarantee efficiency with a smaller investment, predictive maintenance might simply not be worth the expense.
In the foundry (and beyond), human experience remains fundamental
There is an aspect that mathematical models still struggle to replicate: the ability of an experienced operator to perceive that something is wrong.
Let’s look at some concrete examples: a change in machine noise, a slightly different vibration underfoot, an unusual smell, a part that looks “not right” even if it is formally within tolerance. These are weak signals, difficult to quantify, but those who work every day on a plant learn to recognize them with surprising precision.
Ignoring this wealth of experience in the name of digitalization is a costly mistake. Monitoring systems work best when they are designed to dialogue with operators, not to replace them.
Technology must support experience, not erase it. A good predictive maintenance system in the foundry should be able to collect and structure operator observations as notes, reports, and perceived anomalies, integrating them with sensor data. In the foundry, this hybrid approach is often the most effective path. As highlighted by MIT Technology Review in an analysis of intelligent industrial systems, the best-performing predictive models are those that combine sensor data with the tacit knowledge of operators, redefining the human role from executor to active collaborator of the system.
The maintenance technician must be an active part of data collection, not just the recipient of alarms. They must be the first tool of predictive maintenance: artificial intelligence must amplify their capabilities, not replace their judgment.
Additional considerations on the relationship between humans, robots, and artificial intelligence can be found in our article on collaborative robots – cobots.
When it is NOT worth doing predictive maintenance
This is a topic that industrial marketing tends to avoid, but which an honest technical approach cannot ignore.
The determining factor is always economic: productivity and product quality must justify the investment.
Not all components and not all plants justify a predictive monitoring system. In particular:
- Inexpensive and easily replaceable components: if a component costs a few euros and can be replaced in ten minutes, continuous monitoring makes no economic sense. Preventive maintenance at regular intervals is more than sufficient.
- Simple plants with low production: the return on investment in a predictive system depends on the frequency of use and the cost of downtime. On a plant that works a few hours a day on non-critical processes, the numbers rarely add up.
- Components with unpredictable failure modes: some failures are genuinely random and have no detectable precursors. For these, the correct strategy is to invest in spare part availability, not monitoring.
The goal is not to have the most technologically advanced system. The goal is the real return on investment, measured in avoided downtime, improved quality, and reduced maintenance costs.
Cybersecurity and protection of know-how
A topic that is too often treated as secondary in the debate on predictive maintenance, while in the reality of manufacturing companies, it is increasingly central.
Process data from a foundry is not neutral data. It contains information on production rates, optimal processing parameters developed over years of experience, product quality, and raw material suppliers. It is, for all intents and purposes, company know-how.
It is necessary to structure the system that collects the data underlying the predictive maintenance system in the foundry so that it is closed and reliable. Putting data in the cloud or having it processed by an external AI system is by definition a risk: a risk of losing company know-how.
Contracts can formally protect the customer, but real protection is having the data physically under control. It is not paranoia: it is the responsible management of information assets that have real value. After all, AI systems that work in the cloud are only part of this technology: there are also systems that work locally, and they are still LLM (Large Language Model) systems.
Further insights on artificial intelligence in the foundry in this article.
The future of predictive maintenance in the foundry
The development directions consolidating in the sector are quite clear, although adoption times vary greatly from company to company.
On-machine artificial intelligence will become increasingly common: not centralized systems that collect data from many machines, but distributed intelligence directly in the machine controller, capable of reacting in real time without depending on external connections. AI integrated into vision systems and robots. Decentralizing skills directly into the equipment means that data never leaves the equipment itself, making it more reliable.
Assistants for operators. These are systems that guide the operator through setup, maintenance, and troubleshooting procedures, and will reduce dependence on individual experience and accelerate the training of new personnel.
Automatic diagnostics will evolve towards systems capable not only of detecting an anomaly but of identifying its cause and suggesting the correct intervention, reducing diagnosis times that today weigh heavily on maintenance costs.
Integration with MES and ERP will allow connecting the status of plants with production planning: a planned maintenance intervention can be automatically inserted into the production calendar, automatically optimizing resource utilization.
The digital twin: the virtual replica of the physical plant will allow simulating failure scenarios and testing maintenance strategies without interfering with real production.
At TREBI, these are not abstract future scenarios. The vision guiding the design of our plants is that of machines that are not only automatic but capable of guiding the operator, standardizing processes, and preserving the technical experience that every company accumulates over time. Instead of dispersing this information, we make it available for further future uses.
Conclusion
Predictive maintenance is not magic. It is not enough to install sensors and connect everything to an AI system to eliminate downtime.
Data alone is not enough. Process knowledge, the ability to interpret signals in their real context, and the discipline to act on the information provided by the system are needed.
The goal is not to collect data: it is to increase reliability, quality, and production continuity. These three results are achieved with a mix of well-chosen technology, design oriented towards maintainability, operator experience, and internal organization.
In the foundry, all this becomes even more complex, as the environment is difficult and the process less stable compared to other production contexts, such as mechanical machining.
Before talking about predictive maintenance, the costs and benefits of each strategy must be evaluated. And even before that, it is necessary to understand how much a sudden breakdown really costs and how much one is willing to do to avoid it.
