Predictive maintenance uses sensor data and machine learning to estimate when a machine or vehicle is likely to fail, so you fix it just before it breaks rather than on a fixed calendar or after a breakdown. IoT devices collect vibration, temperature, current, pressure and fault-code data; edge and cloud models turn that data into alerts and work orders. In 2026 the main changes are cheaper on-device inference, generative AI assistants that help technicians act on alerts, and new EU rules on who can access machine and vehicle data.
Updated October 2026
What changed in 2026
- Downtime is more expensive. Siemens' 2024 study estimates the 500 largest companies lose about $1.4 trillion a year to unplanned downtime, around 11% of revenue. An hour of lost output at a large automotive plant can cost up to $2.3 million (AEMT summary of Siemens TCOD 2024).
- More firms run dedicated teams. The same study found nearly half of surveyed firms now have a dedicated predictive maintenance team, double the 2019 share.
- Assistants sit on top of the alerts. Siemens added a generative AI interface to Senseye Predictive Maintenance in 2024 to search past cases and suggest fixes in several languages (Drives & Controls). CMMS and fleet vendors now ship similar assistants.
- Data access is regulated in the EU. From 12 September 2026, new connected products sold in the EU, including industrial machinery and vehicles, must be designed so users can access their product data by default, where relevant and technically feasible (Matheson).
- AI Act deadlines moved. High-risk AI duties for AI used as a safety component in regulated products (Annex I, such as machinery) now apply from 2 August 2028 rather than 2026 (Orrick).
Reactive, preventive, condition-based, predictive: what is the difference?
| Approach | When you act | Data needed | Typical problem |
|---|---|---|---|
| Reactive (run to failure) | After it breaks | None | Unplanned stops, secondary damage |
| Preventive / time-based (TBM) | On a calendar or mileage interval | Hours, km, dates | Replaces healthy parts; misses random failures |
| Condition-based (CBM) | When a reading crosses a threshold | Live sensor values | Thresholds are set by hand and fire late |
| Predictive (PdM) | When a model forecasts failure inside a set window | Sensor history + failure and work-order history | Needs labelled failures and good data plumbing |
| Prescriptive | Model also recommends the action | All of the above + repair knowledge | Hardest to trust; needs a human in the loop |
The reason fixed schedules fall short is that most failures are not tied to age. ARC Advisory Group reports that only about 18% of asset failures are age-related, while 82% follow a random pattern (ARC Advisory Group). A calendar cannot catch a random failure; a sensor trend can.
How IoT predictive maintenance works
- Sense. Fit or read sensors: vibration (accelerometers), temperature, motor current, oil quality, acoustic, pressure. Vehicles already produce engine fault codes and telematics data through the CAN bus.
- Collect at the edge. A gateway or onboard unit samples data, filters noise and computes features such as RMS vibration or spectral peaks. Raw high-frequency vibration data is too large to stream from every asset.
- Score. A model flags anomalies (something looks different) or predicts a failure class or remaining useful life (RUL). Simple models often run on the gateway; heavier models run in the cloud.
- Decide. Alerts are ranked by risk and cost and checked against spares, crews and production or route plans.
- Act. A work order is raised in the CMMS or fleet system. Increasingly an assistant drafts it, with the likely cause and the parts list, for a planner to approve.
- Learn. The technician's finding (true fault, false alarm, root cause) is fed back as a label. Without this step models drift and alert fatigue sets in.
Models that do the predicting
- Anomaly detection (autoencoders, isolation forests, statistical baselines): useful when you have few recorded failures, which is most sites.
- Classification (gradient-boosted trees, CNNs on spectrograms): "which fault is this?" once you have labelled examples.
- Remaining useful life (LSTMs, survival models): "how many hours or km are left?" Needs run-to-failure histories.
- Physics-informed and hybrid models: combine engineering rules with ML, which helps when data is thin.
- Foundation models for time series and language: in 2026 these are used mainly to summarise alerts, search manuals and past work orders, and draft reports. They support the predictive model; they do not replace it.
Predictive maintenance for industrial assets
Rotating equipment is where IoT predictive maintenance pays back first: motors, pumps, fans, compressors, gearboxes and conveyors. Bearings and misalignment show up in vibration well before failure. Thermal cameras and machine vision add coverage for electrical panels, belts and leaks (see our guide to computer vision solutions).
What the evidence says about results:
| Source | Finding |
|---|---|
| McKinsey (2017) | Predictive maintenance typically reduces machine downtime by 30–50% and increases machine life by 20–40% |
| Deloitte | Maintenance cost down 5–10%, equipment uptime up 10–20%, planning time down 20–50% |
| US DOE FEMP | Predictive saves 8–12% over preventive and 30–40% over reactive maintenance |
| McKinsey (2021) | Caution: false positives can erase the savings, so precision matters as much as recall |
Treat these as ranges from studies, not promises. Your result depends on how costly your downtime is and how good your failure history is. A digital twin of the line helps test "what if we run this pump to the next shutdown?" before you decide (see our guide to digital twin technology).
Predictive maintenance for vehicle fleets
Fleets have one advantage over plants: the data already exists. Trucks, buses and vans log diagnostic trouble codes (DTCs), engine hours, oil and coolant temperatures, battery voltage, brake and tyre data. Telematics devices or OEM data feeds send it to a fleet platform.
Common fleet use cases:
- Battery and starter failure forecast from cranking voltage trends.
- Aftertreatment problems (DPF, DEF/SCR) caught from regeneration patterns before the vehicle derates.
- Brake and tyre wear from pressure, temperature and driving behaviour.
- EV fleets: battery state-of-health, thermal behaviour and charger faults.
The shift since 2024 is from "alert on a fault code" to "predict the fault code". Fleet platforms such as Geotab, Uptake and MaintainX now advertise ML-based predictive alerts and assistants that learn from manuals and work-order history (Geotab, Aug 2026). Edge AI on the vehicle matters where connectivity is poor: the model scores data onboard and sends only the alert.
The operational payoff is scheduling: a predicted failure becomes a planned stop at a depot, not a roadside breakdown and a tow. Route planning can then work around the vehicle.
Plants vs. fleets at a glance
| Industrial assets | Vehicle fleets | |
|---|---|---|
| Main data | Vibration, temperature, current, acoustic | DTCs, CAN bus signals, telematics, GPS |
| Sensors | Usually retrofitted | Mostly built in |
| Connectivity | Plant network, often stable | Cellular, intermittent |
| Where models run | Gateway + cloud | Onboard unit + cloud |
| Cost of a miss | Line stoppage, scrap, safety | Roadside breakdown, missed delivery, safety |
| Regulation to watch | EU Data Act, machinery rules | EU Data Act, vehicle type-approval and data access |
How to start: a 90-day plan
- Pick 10–30 critical assets or one vehicle class with costly, repeated failures.
- Pull 12+ months of work orders and match them to sensor or telematics history. This is your label set.
- Fill sensor gaps only where data is missing; fleets often need none.
- Start with anomaly detection and a clear alert threshold; add failure classification once labels grow.
- Wire alerts into the CMMS or fleet system, not a separate dashboard nobody opens.
- Measure false-alarm rate, lead time before failure, and avoided downtime hours; review monthly with technicians.
Challenges to plan for
- Too few failures to learn from. Well-run assets rarely fail; use anomaly detection and physics rules first.
- False alarms. They destroy trust faster than misses. Track precision from day one.
- Messy work orders. Free-text notes and missing failure codes are the biggest data problem. Language models now help clean them.
- Data ownership. OEMs, telematics vendors and operators all hold parts of the data; the EU Data Act changes who can request it.
- Security. Every gateway is a network entry point; patch and segment it.
Where it's heading: 2026 → 2030
| Area | Around 2020 | 2026 (now) | By 2028 (expected) | By 2030 (expected) |
|---|---|---|---|---|
| Adoption | Pilots; dedicated PdM teams in a minority of firms | Nearly half of surveyed firms have PdM teams (Siemens TCOD 2024) | PdM standard on critical assets in large plants and fleets | PdM as default for new connected equipment |
| Technician tools | Dashboards and threshold alerts | Generative AI assistants search past cases and draft work orders (Siemens Senseye) | Agents propose and schedule work orders, with planner approval | Closer links between agents, spares and production or route planning |
| Where models run | Mostly cloud | Edge gateways and onboard units for scoring; cloud for training | More on-device inference on low-power chips | Edge-first for most scoring |
| Data access | Held by OEM or vendor | EU Data Act access-by-design for new connected products from 12 Sep 2026 (Matheson) | Third-party maintenance providers use user-shared data | — |
| AI regulation | None specific | EU AI Act in force; high-risk timelines deferred (Orrick) | High-risk duties for AI safety components in regulated products from 2 Aug 2028 | — |
Columns marked "expected" are editorial forecasts based on the trends above, not sourced predictions.
How NexaStack helps
NexaStack's Edge AI platform runs and manages AI models on edge gateways and onboard units, so vibration, telematics and vision models score data next to the asset and send only the alerts that matter. The same platform handles model updates, monitoring and rollback across plants and fleets, and connects alerts to your existing maintenance systems.
Frequently asked questions
What is the difference between TBM and CBM in TPM?
Time-based maintenance (TBM) services equipment on a fixed schedule, such as every 500 hours. Condition-based maintenance (CBM) acts when a measured condition, such as vibration, crosses a limit. Predictive maintenance goes one step further by forecasting when that limit will be reached.
Is IoT still a thing in 2026?
Yes, though the label is used less. The sensors and gateways are now part of edge computing and physical AI platforms, and predictive maintenance is one of the most common reasons companies deploy them.
What data do you need for IoT predictive maintenance?
Sensor or telematics history plus a record of past failures and repairs. The repair history is usually the harder part, because it provides the labels the model learns from.
How does predictive maintenance work for fleets?
Vehicle fault codes, engine data and telematics are sent to a model that spots patterns that come before known failures. The fleet manager gets an alert and books the vehicle into a depot before it breaks down on the road.
What are the four types of IoT?
They are commonly grouped as consumer, commercial, industrial (IIoT) and infrastructure IoT. Predictive maintenance belongs mainly to industrial and infrastructure IoT, plus connected vehicles.
Start predictive maintenance on your critical assets
Pick one line or one vehicle class and see how NexaStack scores sensor and telematics data at the edge and turns it into work orders your team trusts. Book a predictive maintenance pilot scoping session →
