
A system rarely breaks down suddenly. Almost always, it sends signals for weeks: it vibrates a little more, draws a few extra amps, heats up where it previously did not, or runs for more hours to achieve the same result.
Predictive maintenance is a set of techniques that detect these signals and turn them into a scheduled intervention, rather than waiting for a breakdown or replacing components that are still in good condition according to a calendar.
However, there is an aspect that is rarely discussed, and it matters most to building managers: long before a machine stops working, a deteriorating machine starts consuming more energy. This excess consumption is paid for every day, and nobody notices it because it causes no service disruption.
Maintenance terminology is standardised: UNI EN 13306:2018 defines the general terms used in the sector. In practice, three categories are distinguished.
Corrective maintenance takes place after a failure. Preventive maintenance takes place beforehand, at set intervals, regardless of the component's actual condition. Predictive maintenance also takes place beforehand, but based on measured parameters that indicate ongoing deterioration.
The practical difference between the first two, with examples and advantages of each, is explained on the page about the difference between preventive and corrective maintenance. Here, the focus remains on predictive maintenance.
One clarification is worth adding because it is often lost in commercial language. Predictive does not mean that software guesses the date of a breakdown. It means monitoring a physical quantity correlated with a component's health and intervening when that quantity deviates from normal behaviour.
The cycle has four steps, and the fourth is the one most often skipped.
Measure. Vibration, temperature, electrical current draw, differential pressure, operating hours and flow rate. The choice of measurements depends on the machine: vibration and current draw matter for a fan, while temperatures and pressure drop matter for a heat exchanger.
Establish what is normal. Without a baseline built over several weeks of operation, any threshold is arbitrary. This phase requires the most patience and is the one most often shortened.
Detect deviations. A value outside the threshold, a gradual drift, or behaviour that does not match external conditions or the load.
Generate the intervention. The anomaly becomes a work order assigned to someone, with a deadline and a history. An alert that arrives by email and simply sits there is not predictive maintenance; it is just another alarm to ignore.
There is a great deal of confusion around this topic, so precision matters.
Machine learning algorithms help identify patterns that a fixed threshold cannot detect: combinations of multiple parameters, behaviour that changes according to load or season, and drift that appears only under certain operating conditions.
They need historical data to work, particularly records of failures. A model trained on equipment that has never broken down has little to learn from. This is the most common practical limitation in buildings, where collected data is scarce and often discontinuous.
The operational implication is clear. In most cases, it makes sense to start with well-designed rules and thresholds based on a few measurements, collect data for one or two years, and introduce models when there is enough material to train them. Doing the opposite—buying the algorithm before having the data—is the most common and most expensive mistake.
The next step is prescriptive maintenance, where the system not only flags the anomaly but also recommends corrective action and estimates its impact. This is a real direction of development, but it presupposes everything that comes before it.
This is the part that directly concerns building managers and is often overlooked in industrial discussions of predictive maintenance.
A dirty air filter increases pressure drop, so the fan draws more power to move the same amount of air. A fouled heat-exchange coil transfers less heat, and the system compensates by extending its operating hours.
A refrigeration circuit with a reduced refrigerant charge delivers less performance for the same amount of work. A miscalibrated sensor causes the system to regulate itself against an incorrect value for months.
None of these cases causes a shutdown. Instead, they gradually produce higher energy bills and remain invisible until someone compares consumption with the equivalent period of the previous year.
This leads to the most useful operational insight in the entire discussion: electrical current draw is both an indicator of consumption and an indicator of equipment condition. The same measurement used for energy consumption monitoring can signal component deterioration without installing diagnostic-specific sensors.
This is why energy management and maintenance make sense on the same platform, rather than in two systems that communicate with each other once a month.
This is why energy management and maintenance make sense on the same platform, rather than in two systems that communicate with each other once a month.
Deloitte's 2017 position paper indicates a reduction in overall maintenance costs of between 5% and 10%, an increase in equipment availability of between 10% and 20%, and a reduction in planning time of between 20% and 50%.
The same document also reports much higher averages, such as 70% fewer failures, without specifying the underlying basis.
The 2018 PwC-Mainnovation study, conducted among 268 companies in the Netherlands, Germany and Belgium, is more cautious: 60% of companies that adopted predictive approaches achieved an average uptime increase of 9%.
Two practical points follow. First, the highest figures almost always come from industrial environments with critical machinery operating continuously, conditions that do not apply to commercial buildings.
Second, the only figure that matters for an investment decision is the one derived from your own records of downtime and maintenance interventions, not an industry benchmark.
Four things are needed: sensors measuring significant parameters, continuous data collection, a maintenance management system that receives anomalies and turns them into tasks, and someone who regularly reviews the data.
There is no need to start with the entire system. The sensible approach is to focus on components where failures are most expensive, either because they interrupt services or because emergency interventions cost more: chillers, heating plants, elevators, pumps and main electrical panels.
Predictive maintenance does not pay for itself on low-value components that are easy to replace. An inexpensive lamp or valve is better managed on a schedule or replaced when it fails, and stating this openly is part of an honest project.
In air handling units, differential pressure across filters, fan current draw, and supply and return air temperatures are monitored. Together, these three parameters indicate filter fouling, loose belts and fouled coils.
For chillers, the key parameters are pressures, condensing and evaporating temperatures, and current draw relative to load, because efficiency drops well before the chiller stops working.
For pumps, the combination of vibration and current draw provides a signature that can anticipate deteriorating bearings and cavitating impellers.
For electrical panels, periodic thermography detects loose connections, which are a frequent cause of failure and a fire risk.
For elevators, the number of journeys, operating times and restoration interventions reveal a great deal about system condition, and these data are already generated by the control panel.
UTwin's maintenance software is built within a building digital twin: each component is located on the floor plan and has its own record, history and documents. Interventions can be assigned and tracked, and remain accessible even when the maintenance provider changes.
What distinguishes this approach is its integration with energy management software. Data from smart meters and IoT sensors simultaneously support consumption analysis and equipment-condition monitoring. As a result, a change in current draw can trigger a work order without manual transfers between two separate systems.
Where an automation system already exists, the platform works alongside it and reads its data. The relationship between building automation and digital twins is explained on the dedicated page.
Preventive maintenance takes place at predetermined intervals, regardless of the component's condition, and can therefore replace parts that are still efficient or intervene too late. Predictive maintenance is based on measured parameters indicating ongoing deterioration, so its frequency adapts to the machine's actual condition.
No. Algorithms help when sufficient historical data is available and when deterioration signals depend on a combination of multiple variables. In many building systems, a well-configured threshold based on two or three significant measurements already delivers most of the benefits and allows you to collect the data needed later.
Published estimates vary considerably. Deloitte indicated a reduction in maintenance costs of between 5% and 10% in 2017, while the 2018 PwC-Mainnovation study found an average uptime increase of 9% among 60% of the companies involved.
The much higher figures often quoted generally concern critical industrial systems, so a reliable estimate should be derived from your own downtime and maintenance records.
Start with systems where failures are most expensive, either because they interrupt services or because urgent interventions cost more: chillers, heating plants, pumps, elevators and main electrical panels. There is no return on investment for inexpensive components that are quick to replace.
The relationship is closer than it might seem. Dirty filters, fouled coils, insufficient refrigerant charge and miscalibrated sensors increase energy consumption long before causing a shutdown. Because electrical current draw is both a consumption measurement and an indicator of equipment condition, monitoring it serves both purposes with the same installation.
Informative Material