The prediction is not the maintenance strategy
Twelve days to failure sounds precise. It is not the same as knowing what to do. The plant must still judge how critical the asset is, whether production can release it, whether the right parts and skills are available, and whether intervening too soon would sacrifice useful life that could have been retained without increasing risk.
That distinction matters more as condition monitoring and AI make predictions easier to generate. Manufacturers can instrument more assets and identify patterns that once depended on an experienced analyst. Yet a prediction can create cost or false confidence when separated from asset criticality, operating context and the workflow required to respond.
Not every asset deserves the same answer
Maintenance strategy begins with consequence rather than technology. A compressor that can stop an entire process requires a different level of protection from a small motor that can be replaced from stores in ten minutes. The monitoring frequency, analytical depth and response plan should reflect what failure would mean to safety, production, quality and cost.
That hierarchy is central to the way Emerson approaches asset health. Drew Mackley, Director of Sales Enablement for Reliability Solutions, argues that run-to-failure remains a legitimate choice for the right equipment. “Customers have critical assets, essential assets, balance-of-plant assets and equipment where run-to-failure is the maintenance strategy. If an asset has little effect on the process, there is a replacement on the shelf and it takes ten minutes to change, investing heavily in monitoring may not make sense. There is no single technology or frequency that fits every asset, because the decision should follow criticality and consequence.”
The economics of monitoring have changed. Wireless sensors and edge analytics can cover equipment that once sat outside formal programs. Critical machines may justify readings every second and rapid protective action, while lower-consequence assets can be checked hourly. The value lies in a graduated strategy rather than placing the same sensor, threshold and response on everything that rotates.
Continuous collection also removes the blind spots created by monthly routes. A machine that is unavailable when the analyst arrives may go 60 days without a reading, while a fast-developing fault can move from harmless to critical between inspections. Automated collection gives the team more time to react, but it does not remove the need to understand what the signal represents.
Better prediction begins with better context
AI is often presented as the missing intelligence that will convert sensor data into reliable maintenance decisions. Mackley sees it as an important efficiency tool but warns that pattern recognition cannot compensate for information that was never captured. A sensor with an inadequate frequency range cannot reveal a gearbox fault outside that range, regardless of the sophistication of the model applied afterward.
“AI only knows what it has been taught and what the sensing technology allows it to see,” he says. “It can recognize that a pattern is different, but it cannot always explain severity, remaining useful life or the directional evidence, a trained vibration analyst would consider. It is a strong starting point for someone who is not a specialist, but the underlying data still must be fit for the fault you are trying to detect, and the recommendation still needs a second opinion where the consequence is high.”
The same distinction appears in ABB Motion Services’ work with drives and motors. For Sachin Mathur, Global Head of Digital, a real-time reading becomes more useful when it is connected with service history, prior repairs, operating load and the production schedule. The objective is not simply to know that behavior has changed, but to understand why it changed and what the business can do next.
“A maintenance record might tell you when a bearing or belt was replaced, while real-time monitoring shows how the asset is behaving today,” Mathur explains. “The value comes when those dimensions are brought together with information from systems such as ERP or computerized maintenance management. You can then move from detecting and diagnosing a problem to predicting how it may develop and making a recommendation that reflects the asset’s history and its role in the operation.”
That context becomes harder in mixed fleets spanning several generations and suppliers. A useful strategy cannot depend on replacing every installed asset or limiting analysis to one vendor’s equipment. It must combine information while retaining the engineering knowledge needed to interpret different machines correctly.
Production context introduces another layer. The same event can influence OEE, scrap, quality and maintenance, but separate applications may present it as several unrelated problems. Leonor Marques, Architecture & Advocacy Director at Critical Manufacturing, says the relationship between those outcomes must survive the journey through different systems.
“If you do not know that the same shop-floor event is behind them, you can search for different root causes and try to solve each symptom independently,” she says. “The architecture has to preserve the relationship between the asset condition, the product being made, the machine configuration and the production event, otherwise the insight may be accurate in isolation and still lead to the wrong action.”
The best intervention is not always the earliest
Predictive maintenance is frequently described as finding faults sooner, but the real objective is choosing the best moment to intervene. Acting at the first sign of degradation can waste remaining economic life, while waiting too long can increase energy use, damage connected equipment or extend the eventual outage. A maintenance strategy therefore must balance risk, production commitments and the changing cost of continued operation.
Enel, a global energy company and major renewable power operator, shows how predictive analytics can support more precise maintenance decisions across a large generation fleet. Its remote diagnostic center monitors 1,285 assets using thousands of models and plant data tags. The system does not merely announce that a turbine filter is degrading. It compares pressure change, expected remaining life, operating performance and the economics of replacement to determine when a change creates the greatest value.
In one instance, Enel found that a filter had previously been replaced while significant economic life remained. In another, replacement was delayed long enough for excess gas consumption to cost approximately €18,000 a week. Since 2020, the company reports that predictive analytics has helped identify 461 failures and avoid almost €47 million in estimated losses. The strategy is not “replace when predicted.” It is to use the prediction to select an intervention point that protects availability without over-maintaining the asset.
The same program connected maintenance decisions with emissions. Earlier detection reduced the need for less efficient replacement power when thermal or geothermal assets were unavailable, producing substantial estimated carbon reductions. That wider outcome matters because reliability, energy efficiency and sustainability often arise from the same operational decision rather than separate projects.
This convergence is central to Mounir Boemond’s work as Global Director of Sustainability Value at AVEVA. “The useful KPIs are the ones that both the plant floor and senior management can act on,” he says. “Energy consumption can support emissions reporting, but it can also improve equipment effectiveness and profitability. Once operational data is contextualized, predictive analytics can help the organization move from reacting to failure toward preventing the cost, waste and environmental impact that develop before the asset finally stops.”
A recommendation needs somewhere to go
The most sophisticated prediction still fails if it ends in a dashboard that nobody owns. Maintenance teams need an agreed route from detection to verification, work-order creation, planning, parts availability and execution. Operations must know which alerts require an immediate production decision, and which belong within the normal maintenance process.
Mathur describes the progression as insights, recommendations and actions. Many manufacturers can already generate the first layer. AI is accelerating analysis and making recommendations easier to explain, including through agents that allow engineers to question a system in natural language. The harder step is embedding the result in the operating model.
Scaling exposes that weakness. A pilot covering 20 assets can be managed through a committed project team, but a deployment covering thousands requires sustained competence and clear ownership. Some plants expect users to monitor dashboards themselves, even though the workforce shortage that justified the technology also limits the time available to interpret it.
“The benefit does not come from giving the customer another screen with red and green lights,” Mathur says. “It comes when somebody takes responsibility for monitoring the fleet, explains what may go wrong and helps the plant move quickly toward the corrective action. Technology is advancing faster than operating models, budgets and accountability, so the obstacle is often not whether the prediction can be made, but who owns the response at scale.”
Closed-loop maintenance will become more realistic as automated collection, analytics and work-order generation mature. Mackley expects those elements to connect from detection through correction, but autonomy should be earned through evidence and bounded by consequence. Low-risk decisions can tolerate more automation than an intervention affecting safety, quality or a critical process.
The maintenance leaders will not be those producing the greatest number of predictions. They will know which assets merit attention, what evidence is sufficient, how long intervention can safely wait and who has authority to act. Prediction is one input into that system. The strategy is the set of decisions that turns it into reliable production.
