Predictive Maintenance with IoT and AI: A Practical Guide for Mid-Sized Factories

In many mid-sized factories, machine maintenance still works in one of two ways: fix it when it breaks, or replace parts on a fixed schedule without looking at their actual condition. The first approach causes sudden production stops and late orders. The second is safer, but it often throws away parts that are still good and still fails to prevent breakdowns that occur between scheduled services.

Predictive maintenance offers a middle path: monitor machine condition continuously through sensors, then use data analysis and AI to estimate when a component is likely to fail. This used to be associated with large companies that had their own data teams. Today wireless sensors are cheaper, IoT gateways are easier to install, and processing can happen on edge devices or in the cloud. Industry reports from 2026 show predictive maintenance moving from pilots into routine deployment, including at mid-sized plants. Another driver is the loss of maintenance know-how as senior technicians retire. This article explains how to start realistically without a large, risky project.

Summary

• Start with one or two machines that cost the most when they stop, not the entire factory
• Vibration, temperature, current, and pressure sensors are enough for many common failure types
• AI models need breakdown and maintenance history, so work orders must be recorded properly from the start
• System alerts must become work orders in the ERP or CMMS, not just notifications
• Calculate benefits from downtime avoided, spare parts saved, and overtime reduced
• Keep experienced technicians involved because their knowledge makes the model accurate faster

Three approaches to maintenance

Reactive maintenance means repairing a machine after it breaks. It looks cheap because there is no routine activity, but the hidden losses are large: production stops, materials are wasted, deliveries are late, and spare parts have to be bought urgently at higher prices.

Preventive or scheduled maintenance means changing oil, bearings, or belts after a certain number of hours or days. This is far better and is a baseline every factory should have. Its weakness is that schedules are based on averages, while the condition of each machine differs with workload, environment, and operator quality.

Predictive maintenance looks at actual condition. If bearing vibration starts rising outside its normal pattern, the system warns days or weeks before failure. Technicians can schedule the replacement during a production break. This approach does not replace preventive maintenance; it complements it for critical machines. The basics of asset management and scheduled maintenance are covered in asset management and equipment maintenance with ERP.

Choosing the first machine

A common mistake is wanting to put sensors on every machine at once. The result is high cost, piles of data, and an overwhelmed team. It is better to pick one or two machines that meet these criteria:

• If this machine stops, the whole production line stops
• It has failed before and the cost was felt, in both downtime and repairs
• Its spare parts have long lead times or are expensive
• Its failure modes can be detected from physical condition, such as vibration, heat, or current
• There is a technician who understands it and is willing to be involved

Examples include the main air compressor, circulation pumps, conveyor drive motors, plastic injection machines, or extruders. Machines like these usually have rotating parts with well-understood failure patterns.

The sensors and data you need

For many cases, a few sensor types are enough:

• Vibration sensors to detect worn bearings, misaligned shafts, or loose components
• Temperature sensors to detect excess friction, poor lubrication, or overload
• Current sensors to see changes in motor load
• Pressure and flow sensors for pumps and hydraulic systems
• Operating hour and cycle counters to know actual usage

Battery-powered wireless sensors now make installation possible without pulling new cables. Data goes to a gateway and then to a local server or the cloud. For factories with unstable internet connections, initial processing can run on an edge device so alerts still appear when the connection drops. Practical considerations for choosing devices are in realistic IoT and sensors for business.

Data from operational systems

Sensor data alone is not enough. The model needs context: what product is being made, how fast the machine is running, when a component was last replaced, and what failures have occurred. This data usually lives in the ERP, the production system, or technicians' notebooks. If maintenance history is still on paper, start recording it digitally with failure types and replaced parts. Without failure labels, AI can only detect anomalies; it cannot yet name the problem.

How AI works in predictive maintenance

There are several levels of analysis, from simplest to most advanced.

Thresholds

The first level is thresholds: if bearing temperature exceeds 80 degrees or vibration passes a certain value, send an alert. This is easy, transparent, and often already delivers large benefits. Industry standards for machine vibration levels can serve as an initial reference.

Anomaly detection

The second level is anomaly detection. The model learns the machine's normal behaviour across operating conditions, then flags deviations. This is useful because it does not need many failure examples, only data from when the machine is healthy.

Remaining useful life

The third level is estimating remaining useful life. The model estimates how many days remain before a component needs replacing. This requires enough failure history, so it usually becomes realistic only after the system has been running for some time.

Mid-sized factories should start with thresholds and anomaly detection. Both deliver results quickly and build the team's trust. General data readiness before AI projects is discussed in company data readiness before starting an AI project.

From alert to action

Many predictive maintenance projects fail not because the model is wrong but because nobody acts on the alerts. A dashboard notification that no one opens is worthless. A good flow looks like this:

• The system detects an abnormal condition and rates its severity
• For medium severity, the system automatically creates a maintenance work order in the ERP or CMMS
• The maintenance supervisor schedules the job in the nearest production break
• The system checks spare part stock and raises a purchase request if needed
• The technician records the actual finding, for example the bearing really was worn or a sensor was loose
• That record becomes feedback to improve the model

Technician feedback is essential. Every false alarm and every missed failure should be recorded so accuracy improves over time.

An example from a plastic packaging plant

A packaging plant in East Java installed vibration and temperature sensors on its main extruder motor. In the second month, the system detected vibration slowly rising at a specific frequency. A work order was created, a technician checked during a shift change, and found a bearing beginning to wear. It was replaced in two scheduled hours instead of failing in the middle of a large order. A single event like this often pays for the sensors on that machine.

Calculating the benefit honestly

Many vendors promise savings with large percentages. Treat those figures as marketing claims until they are proven in your own plant. A more honest approach is to calculate from your own data:

• Average unplanned downtime hours per year on the chosen machine, multiplied by production value per hour
• Emergency repair costs, including overtime and express spare part shipping
• Components replaced too early because of preventive schedules
• System costs: sensors, gateways, software, integration, and team time

Compare the figures before and after over six to twelve months. If the benefit is clear, expand to the next machine using the same priorities.

Preparing the maintenance team and process

Technology is only half the job. The other half is people and process. Decide who receives alerts, who decides on action, and what the response deadline is for each severity level. Train technicians to read simple vibration and temperature charts so they understand the reasoning behind an alert rather than just taking orders from the system. Schedule a short weekly meeting to review alerts raised, actions taken, and findings on the floor. Also capture senior technicians' knowledge, such as sounds or symptoms that usually precede a failure, in digital maintenance records. That knowledge is valuable for training new technicians and for improving analysis rules and models.

Common mistakes to avoid

• Installing sensors on every machine before proving value on one
• Not recording failure history, so the model cannot learn
• Alerts that only appear on a dashboard and never become work orders
• Too many false alarms, so technicians stop trusting the system
• Neglecting basic preventive maintenance because a predictive system exists
• Not involving technicians from the start, even though they understand machine symptoms best
• Choosing a closed platform whose data is hard to export to other systems

Frequently asked questions

How much does it cost to start?

It depends on the number of machines, sensor types, and integration. A pilot on one or two machines is usually far smaller than the loss from a single production stop on a critical machine. Calculate your loss per hour of downtime first to set a sensible investment ceiling.

Can old machines be monitored?

Yes. Vibration and temperature sensors are mounted on the outside of the machine and do not need a modern control system. Older machines often benefit the most because they are more prone to failure.

Do we need a data scientist?

Not in the early stages. Many platforms provide built-in analysis. What you need is someone who understands the machines and processes, and a partner who can integrate the data with your maintenance system.

What about network security?

Separate the sensor network from the office network, restrict device access, and make sure gateways receive security updates regularly.

Conclusion

Predictive maintenance is no longer a technology suited only to giant factories. By starting with the machine that costs the most when it stops, using simple sensors, recording maintenance history properly, and connecting alerts to work orders, mid-sized factories can reduce production stops in a real way. The key to success is not the most advanced algorithm but data discipline, technician involvement, and a clear follow-up flow.