Predictive maintenance in manufacturing and equipment fleets: what works today
Updated · Reviewed by Israel Margulies, founder and CEO of High Point GPS
Predictive maintenance means servicing a machine when measurements of its condition, such as vibration, temperature, oil quality or fault codes, show a failure starting, instead of on a fixed schedule. Manufacturing plants have used it for decades on fixed rotating equipment. Equipment fleets can use a practical version built on telematics readings, engine fault codes and oil analysis, with clear thresholds doing most of the work and AI models as an optional layer on top.
What predictive maintenance is
The US Department of Energy’s O&M Best Practices Guide defines predictive maintenance as measurements that detect the onset of degradation, so the cause can be removed or controlled before the component deteriorates in any significant way. NIST’s survey of US manufacturers puts reliability-centered and condition-based maintenance under the same umbrella: you measure condition or reliability, often with sensors and history, and plan work from it.
The difference from preventive maintenance is the trigger. Preventive work happens because the calendar or hour meter says so. Predictive work happens because a reading says so. NASA’s reliability-centered maintenance guide explains why the trigger matters: in the failure studies it cites, random failures made up 77 to 92 percent of the total and age-related failures only 8 to 23 percent, and the guide notes that picking a fixed interval for a component that fails at random is difficult. Measurement helps with the subset of those failures that give a detectable warning, such as rising temperature or wear metals in the oil; some failures give none.
Predictive maintenance in manufacturing: what the data says
NIST’s 2020 study of US discrete manufacturing (NIST AMS 100-34) is the most useful public data set. Across the plants surveyed, maintenance was on average 45.7 percent reactive, 31.8 percent preventive and 17.3 percent predictive. The study put losses from preventable maintenance problems at $119.1 billion in 2016, most of it lost sales from delays and defects.
Among plants that were under 50 percent reactive, the half that leaned most on predictive maintenance reported 15 percent less downtime and an 87 percent lower defect rate than the other half. NIST calls these results anecdotal and notes that some are not statistically significant because of the small number of respondents, so read them as a direction and not as a savings forecast.
- Vibration analysis on motors, pumps, fans, gearboxes and spindles
- Infrared thermography on electrical panels, bearings and steam systems
- Lubricant and wear-particle analysis on gearboxes, hydraulics and engines
- Ultrasound for leaks and early bearing wear
- Motor current and controller data from PLCs and CNC machines
What transfers to mobile equipment and trucks
A plant’s machines sit still, run at steady loads and plug into the building’s network. A fleet’s machines move between job sites, idle for days, then work flat out, and a wired vibration sensor on every bearing isn’t realistic. The condition data a fleet can get at reasonable effort comes from three places.
First, the engine’s own diagnostics. Heavy-duty engines report through SAE J1939, whose diagnostics layer (J1939-73) defines messages that periodically broadcast active diagnostic trouble codes and the status of the operator’s warning lamps. A telematics device can read those codes and send them to you the moment they appear. Second, live readings such as coolant temperature, battery voltage, fuel level and engine hours. Third, periodic samples: oil analysis on engines, hydraulics and final drives, and measurements technicians take at inspections, such as undercarriage wear or tire tread depth.
AI and IoT predictive maintenance: where each fits
IoT predictive maintenance means sensors send readings over a network so someone, or something, can watch them continuously. AI for predictive maintenance usually means a model trained on those readings and on past failures that flags unusual patterns or estimates remaining life. Both work best where you have many identical machines, steady duty cycles and a long, labeled failure history. A plant with forty identical pumps fits. A fleet of twelve excavators from three brands, each doing different work, gives a model very little to learn from.
Before you pay for prediction, check that your basic data is clean: correct meter readings, a problem and cause code on every repair, and a record of which alerts turned out to be real. Without that, no model has ground truth to learn from, and you can’t tell whether its warnings are any good. The DOE guide also lists the downsides: investment in diagnostic equipment, staff training, and savings that management doesn’t see right away.
A practical predictive program for a fleet
Start with thresholds you can explain. A hypothetical example: a dozer’s coolant temperature runs in a steady band all season. You set a rule that any reading above the top of that band opens an inspection work order and alerts the shop lead. Weeks later the rule fires, a technician finds a radiator core packed with debris, and the machine is cleaned on a Friday afternoon instead of overheating on a Tuesday morning.
- Pick the machines where an unplanned failure costs the most, using your criticality ranking
- List the two or three failure modes that hurt each one most and the reading that shows each early
- Set thresholds from the manufacturer’s limits and your own normal operating range
- Decide what each alert does: open a work order, take the unit out of service, or notify someone
- Review every alert after the fact and tighten or loosen the threshold
- Keep the preventive schedule running underneath; prediction adds to it and doesn’t replace it
Choosing predictive maintenance software
Predictive maintenance solutions range from sensor-and-analytics platforms built for plants to maintenance systems that act on telematics data. For a fleet, the questions that matter are practical ones. Can it read the telematics you already have, or does it need its own hardware? Does an alert become a work order with the reading attached, or only an email? Can you set different limits per machine or category? Does it keep the history so you can check which alerts were right?
If a vendor claims AI prediction, ask what data the model was trained on, how many failures of your machine type it has seen, and what its false-alarm rate is. If the vendor can't answer those three questions for your machine type, judge the product on its alerts and work orders alone.
Where readings and fault codes fit in FleetZap
FleetZap does not run AI failure prediction. It gives you the condition-based layer a fleet can use today. With optional GPS trackers connected (FleetZap’s team can supply, install and connect them), engine hours and odometer arrive automatically, along with fuel level, coolant temperature, battery voltage and diagnostic trouble codes where the machine provides them. Fault codes open repair work orders and alert the right people.
Threshold automations watch coolant temperature, fuel, battery, speed or meter readings, above or below a limit you set per asset or per category, and can open a work order, take the unit down, send an alert or send an email. Hour-based PM schedules keep running on the real hours. Each machine also carries a health score from 0 to 100 that drops for overdue PM, repeat problem codes and recent repairs, with the reasons shown, and the dashboard lists the machines at risk.
Key takeaways
- Start with the five machines whose breakdown stops the most work, and give each one rule per failure mode before you add any more readings.
- Run oil analysis on the components with the highest repair bills; it is the cheapest condition check a mixed fleet can add.
- Log whether each alert was real; after a season that record tells you which thresholds to tighten and gives any future model something to learn from.
- If more than half your alerts turn out to be false, widen the threshold before people start ignoring them.
Sources
Frequently asked questions
What is predictive maintenance software?
Predictive maintenance software collects condition data from machines, such as sensor readings, engine fault codes or oil analysis results, compares it with normal limits or learned patterns, and raises an alert or a work order when a problem starts. Some products include their own sensors and AI models; others act on data from telematics you already run.
How is AI used for predictive maintenance?
AI models learn what normal looks like from sensor history and past failures, then flag readings that drift from it or estimate how long a part has left. They work best on many identical machines with long failure records. With small or mixed fleets, simple thresholds on the right readings often catch the same problems with less setup.
What is IoT predictive maintenance?
IoT predictive maintenance uses connected sensors to send equipment readings over a network continuously, instead of a technician taking them by hand. For vehicles and mobile equipment, the telematics device is the IoT link: it reads engine hours, temperatures, voltages and fault codes from the machine and sends them to your maintenance system.
Is predictive maintenance worth it for a small fleet?
A full sensor-and-analytics program usually isn’t, because the cost of hardware, training and analysis is spread over too few machines. A light version often is: read fault codes and a few key readings through telematics on your most critical units, run oil analysis on big components, and act on clear thresholds while keeping your preventive schedule.
What equipment benefits most from predictive maintenance in manufacturing?
Rotating equipment that runs at steady loads and costs a lot when it stops: motors, pumps, fans, compressors, gearboxes and machine-tool spindles. These show measurable early signs such as rising vibration, heat or wear particles in the oil, which gives the maintenance team time to plan the repair before the machine fails.