Skip to content
DT Nova

Manufacturing predictive analytics: use cases, data and limits

Manufacturing predictive analytics uses the data a plant already produces, from machines, production orders, quality checks and sales, to estimate what is likely to happen next: a bearing that will fail, a batch that will be scrapped, a spike in orders or a line that will overrun its energy budget. This guide explains the use cases that pay back, the data each one needs, the limits nobody mentions in a sales demo, the EU rules that apply, and how to start without a large programme.

It is written for production, maintenance, quality and operations managers in small and midsize manufacturers. For the four types of analytics and where prediction fits among them, see our guide to types of analytics.

The short version

  • Prediction is worth something only if someone acts on it earlier than they could have otherwise. Start from the decision, not the data.
  • The best first use cases are usually scrap and quality, demand and material planning, and maintenance on a few critical assets.
  • You need history of the event you want to predict. Rare failures make specific failure prediction hard; anomaly detection is often the realistic first step.
  • Most of the work is data: linking machine data to orders, batches and maintenance records with the right timestamps.
  • Most factory analytics is not high-risk under the EU AI Act; AI that is a safety component of machinery is a different matter.

Where manufacturing predictive analytics pays back

Use caseWhat is predictedData it needsWho acts on it
Predictive maintenanceLikely failure or abnormal behaviour of a critical assetCondition and process data, alarms, maintenance and failure historyMaintenance planners
Quality and scrapBatches or parts likely to fail inspectionProcess parameters per batch, material lots, inspection resultsQuality and shift leaders
Demand forecastingOrders per product and periodOrder history, customer forecasts, seasonality, price changesSales and production planning
Material and lead timesLate supplier deliveries, shortagesPurchase orders, confirmations, receipts per supplierPurchasing
Throughput and lead timeOrders likely to miss their dateRoutings, actual times, queues, machine statusProduction planners
Energy useConsumption per line, shift or productMeter data, production volumes, weatherOperations and facilities

Demand and quality are often better first projects than maintenance, because the data already sits in the ERP and quality system and the outcome is easy to check a few weeks later. For retail-side forecasting, see our guide to predictive analytics for retail.

Predictive maintenance, honestly

Predictive maintenance is the use case everyone mentions first. The idea is sound: watch vibration, temperature, current, pressure or cycle times, learn what normal looks like, and act before a failure. The catch is that a model learns from examples, and a well-run plant has few failures of any one kind. A model that has seen three bearing failures cannot reliably predict the fourth.

Three steps usually work better than a leap to prediction:

  1. A working preventive plan. Time and usage-based maintenance, done on schedule and recorded, so there is a clean history. See preventive maintenance scheduling software.
  2. Condition limits. Simple thresholds on measured values, set from the manufacturer's data and experience, that raise a work order when crossed.
  3. Anomaly detection, then prediction. A model that flags behaviour unlike the asset's own normal, reviewed by a technician; with time and labelled outcomes, specific failure prediction for the assets where it pays.

Every alert has a cost: a technician's time, a stopped line, lost trust when it is wrong. Measure false alarms as carefully as caught failures.

The data you need, and where it lives

Factory data sits in layers. Controllers and sensors on the machines hold process values, counts, states and alarms. Manufacturing execution and SCADA systems hold what ran when. The ERP holds orders, materials, routings and costs. Quality systems hold inspection results, and maintenance systems hold work orders and failures. Prediction needs these linked: the process values for a batch, the material lot it used, its inspection result and the machine's maintenance state at the time.

For the machine layer, OPC UA is the common standard; the OPC Foundation describes it as a platform independent service-oriented architecture, released in 2008, with encryption, authentication and auditing (read on 11 November 2026). Many machines can provide data through it directly or through a gateway. For the rest, interfaces to the ERP and quality systems, or regular exports, are enough to start. For the link between connected equipment and your systems, see IoT software development; for planning and orders, see manufacturing ERP.

Two practical points decide most projects. Clocks must agree, because a model linking a temperature spike to a defect is useless if the two systems are minutes apart. And the event you want to predict must be recorded consistently: a failure noted as "fixed" in a free text field cannot train anything.

Access to machine data under the Data Act

The EU Data Act gives equipment owners a stronger hand. Connected products shall be designed and manufactured so that product data and related service data, including the metadata needed to interpret and use them, are by default easily, securely and free of charge accessible to the user, in a structured, commonly used and machine-readable format, and where relevant and technically feasible directly. This applies to connected products placed on the market after 12 September 2026 (Regulation (EU) 2023/2854, Articles 3(1) and 50, read on 11 November 2026). When buying new machines, ask the supplier how product data will be made available and in which format, and write it into the purchase.

How many manufacturers use AI today

Eurostat's 2025 survey of EU enterprises with at least 10 employees found that 19.95% used at least one AI technology, against 55.03% of large enterprises. In the manufacturing sector, AI software or systems were used mostly for marketing and sales (30.41%) and for the organisation of business administration processes or management (27.05%). Among enterprises that had considered AI but did not use it, the most common reason was lack of relevant expertise (70.89%), followed by lack of clarity about the legal consequences (52.52%) (Eurostat, Use of artificial intelligence in enterprises, read on 11 November 2026). In other words, production analytics is still early for most plants, and the barrier is skills and clarity more than technology.

The EU AI Act and production analytics

Most manufacturing predictive analytics, such as demand forecasts, scrap prediction and maintenance alerts reviewed by people, does not fall into the AI Act's high-risk categories. The high-risk rules for products apply where an AI system is a safety component of a product covered by the EU product legislation listed in Annex I of the Act, which includes machinery, or is such a product itself. Under the Act as amended by the Digital Omnibus on AI, those rules apply from 2 August 2028, and the high-risk rules for the uses listed in Annex III apply from 2 December 2027 (AI Act Service Desk, Article 113, consolidated text as at 27 July 2026, read on 11 November 2026).

The practical line: an analytics model that advises people is ordinary software with good practice around it; a model that directly controls a safety function of a machine is a product safety question for the machine's manufacturer. Analytics about workers, such as monitoring individual performance, raises separate data protection and AI Act questions and should be scoped with care.

Common limits and how to handle them

  • Too few examples. Group similar assets, start with anomaly detection, or predict a more frequent event such as a quality deviation.
  • Drift. A new material, tool or product changes the data; models are monitored and retrained, and the system shows when a model was last checked.
  • Black boxes. Planners trust predictions they can understand; show the main reasons behind each alert and keep simple models where they perform as well.
  • No owner for the action. A prediction with nobody responsible for acting on it is a dashboard nobody opens. Agree who acts and how fast before building.
  • Measuring the wrong thing. Model accuracy matters less than the business result: downtime avoided, scrap reduced, stock lowered.

How to start

  1. Pick one decision that is made too late today, with a clear cost, such as a scrapped batch, an emergency repair or a stock-out.
  2. Check that the history exists: at least a year of the outcome and the data before it, with timestamps that agree.
  3. Record the baseline: how often the event happens and what it costs.
  4. Build the simplest model that could work, and run it next to the current process for a few weeks without acting on it.
  5. Act on it with one team, measure against the baseline, and only then extend to more lines, plants or use cases.

How DT Nova helps

We start with the decision and the data you have, link the systems that hold it, and build the forecast or alert into the tools people already use: the maintenance system, the planning screen or a dashboard. Data stays in your account in an EU data region, models and code are yours, and every prediction can be traced back to the data it came from. For the analytics strategy around it, see data analytics consulting; for reports and screens, see dashboard development. For AI uses beyond prediction, such as visual inspection, knowledge assistants and AI inside machines under the Machinery Regulation, see manufacturing and artificial intelligence. To pick the first use worth piloting, see our AI consulting services.

Questions people ask

What is predictive analytics in manufacturing?

It is the use of historical and live data from machines, production, quality and supply to estimate what is likely to happen next: which machine is likely to fail, which batch is likely to be scrapped, how much a customer will order or how much energy a line will use. The estimate is useful when someone acts on it earlier than they could have otherwise.

Is predictive maintenance the same as preventive maintenance?

No. Preventive maintenance follows a plan: by calendar, by running hours or when a measured value crosses a limit. Predictive maintenance uses a model to estimate when a failure is likely and schedules the work just before it. Most plants need a working preventive plan first; prediction is added for the critical assets where it pays.

How much data do we need?

More history of the event you want to predict than most plants expect. For failures, that means a record of past failures with dates and causes, linked to machine data from the weeks before them. For quality, it means inspection results linked to process data per batch. Where failures are rare, a model that flags unusual behaviour is often more realistic than one that predicts a specific failure.

Does the EU AI Act apply to predictive analytics in a factory?

Most production analytics, such as forecasts, scrap prediction and maintenance alerts reviewed by people, is not a high-risk AI system. The Act's high-risk rules for products apply where an AI system is a safety component of a product covered by EU product law listed in its Annex I, such as machinery, or is such a product itself. Under the amended Act those rules apply from 2 August 2028.

Can we start without new sensors?

Often, yes. Many machines already record running hours, counts, alarms and process values in their controllers, and the ERP, MES and quality systems hold years of history. A first project usually uses data that already exists and adds sensors only where a specific gap shows up.

Sources