Predictive Maintenance Without a Data Science Team
Predictive maintenance AI is within reach of a plant with no data scientists. Here is what to start with, how a first pilot runs, and where it goes wrong.
Predictive maintenance AI uses sensor data and maintenance records to warn you that a machine is drifting toward failure, early enough to plan the repair. You do not need an in-house data science team to start. You need a few well-chosen machines, clean maintenance logs, an engineer who knows the failure modes, and a partner who builds and hands over the model.
This guide is for plant managers, maintenance heads and reliability engineers at Thai factories. It explains what the technology does, which approach fits which data, what a small first project looks like, and where it goes wrong. For the wider picture, see our page on AI for manufacturing and automotive.
What predictive maintenance actually predicts
A predictive maintenance model does not see the future. It compares what a machine is doing now with what it normally does, and with what it did before past failures. When the pattern moves in a way that has preceded trouble, it raises an alert.
Three kinds of output are common, and they need different amounts of history:
| Output | Question it answers | History needed |
|---|---|---|
| Anomaly alert | "Is this machine behaving differently from normal?" | Weeks or months of normal running data |
| Failure classification | "Which fault pattern does this look like?" | Labeled examples of each fault |
| Remaining useful life | "How long until it needs work?" | Many run-to-failure histories |
Most plants should start with the first row. Anomaly detection needs normal data, and you already have plenty of that. Failure examples are rare because you work hard to prevent failures, and that scarcity is the central difficulty of the whole field.
Why you do not need a data science team to begin
The heavy parts of the work are repeatable. Reading sensor data, cleaning it, building a baseline for each machine and tuning alerts follow a known pattern. What cannot be outsourced is knowledge of your machines.
Your maintenance engineers know that press 12 runs hotter in the afternoon, that a certain bearing is replaced early on one line, and that a "failure" in the log sometimes means a missed adjustment. A model built without that knowledge produces alerts nobody trusts. So the roles split like this:
- Your side: pick the machines, explain failure modes, review alerts, and record what you find when you open the machine.
- Partner side: connect the data, build and tune the model, set up the alert flow, and train your team to run it.
- Shared: agree on what counts as a useful alert before you start.
That last point matters most. If you cannot say what a good alert looks like, no model can be tuned to it.
Start with the data you already have
Many plants already collect more than they use. Typical sources are:
- Vibration, temperature and current readings from sensors or drives, often stored in a historian.
- PLC signals such as cycle time, pressure, speed and motor load.
- Maintenance records from a CMMS or even spreadsheets: dates, machine, what was found, what was replaced.
- Operating context such as product type, shift and ambient conditions, which explain why readings change.
Maintenance records are usually the weak link. Free-text notes like "fixed noise" or entries made days after the work make it hard to match a failure to the sensor pattern before it. Spend your early effort cleaning these logs. A small, accurate set of failure records beats a large, vague one.
If a machine has no sensors, adding a few is often less effort than building a complex model on poor data. Vibration and temperature on rotating equipment such as motors, pumps, fans and gearboxes are the usual first step. Standards such as ISO 20816 give accepted ways to judge machine vibration, which helps you set sensible alert limits before any AI is involved.
Choosing the right approach for each machine
Do not apply one method everywhere. Match the method to the data.
Rules and limits first. For a simple case, such as a motor that must not exceed a temperature, a threshold alert is enough. AI adds nothing here, and a good partner will say so.
Statistical or machine learning baselines. When normal behavior depends on load, speed or product, a model learns the normal pattern for each condition and flags departures. This suits most rotating equipment and presses.
Failure-pattern models. If you have enough labeled failures of the same type across similar machines, a model can name the likely fault. This takes more history and more review, so treat it as a second phase.
Language assistance for the technician. A maintenance copilot can turn an alert into plain Thai or English, show the related manual page and similar past repairs, and draft a work order. This helps the people acting on the alert, though it is not a prediction. Our plant knowledge work, described under enterprise knowledge AI, covers answering from SOPs and manuals with citations.
A realistic first project
A sensible first pilot is narrow. For example, pick three to five machines of the same type, such as motors on one line, where past failures are well documented and downtime hurts. This is an illustrative scope, not a benchmark.
A typical sequence looks like this:
- Choose machines and define failure. Agree what counts as a failure, a near miss or a planned stop.
- Check the data. Confirm sensors, sampling rates and maintenance records are usable. Fix gaps now.
- Build the baseline. Train on normal running, then test against past events you know about.
- Run in shadow mode. The model raises alerts, but maintenance does not act on them yet. You compare alerts with what engineers find.
- Go live with review. Alerts reach a named person who confirms or rejects each one, and the model learns from that feedback.
- Decide on scale. Expand only if the alerts were useful enough to change how you plan work.
Shadow mode is the step teams skip, and it is where trust is built. It also shows you the false alarm rate before anyone is woken at night.
Where it runs and who can see the data
Production data often reveals output volumes and process settings, so many plants prefer to keep it in their own network. Models can run on edge devices near the machines, or on in-country cloud infrastructure, with decisions written back to your maintenance system. Work with your IT and OT security teams on this early, and check current PDPA guidance if operator or personal data is involved. Our AI infrastructure practice covers edge and in-country hosting.
For a view of how time-series prediction can be presented to a decision maker, our power demand forecasting demo shows a forecast with a likely range, checks against simple guesses and a person making the final call. It is a working demo with sample data about grid load, not maintenance, but the habits are the same: show uncertainty, compare against a simple baseline and keep a person in charge.
What can go wrong
Too many false alarms. If technicians check ten alerts and find nothing, they stop reading them. Tune alert limits against how many checks your team can handle.
Too few failures to learn from. Do not promise remaining-life estimates without run-to-failure histories. Start with anomaly alerts.
Changed conditions. A new product, a replaced part or a season change shifts the baseline. Plan to retrain and monitor drift.
Sensor faults mistaken for machine faults. A loose cable can look like a bearing problem. Include sensor health checks.
No action path. An alert that creates no work order is wasted. Connect it to the way your team already plans repairs.
When not to do this: if a machine is easy to replace, rarely fails, or has no stable operating pattern, a run-to-failure or time-based plan may be the better choice. Predictive methods earn their place where unplanned stops disrupt delivery or safety.
How to start
Make a list of the five machines whose unplanned stops hurt most. For each, note the sensors available, how failures are recorded and who knows the failure modes. That list is the brief for a first conversation.
If you are also looking at the line itself, read our companion post on AI visual inspection for EV parts. To review your machine list with a consultant, book a consultation.
Frequently asked questions
What data do I need to start predictive maintenance?
Start with sensor readings such as vibration, temperature and motor current, plus maintenance records that say what was found and when. Operating context like product type and load also helps. Accurate failure records matter more than volume, so clean the logs first.
How many failures do I need before the model works?
Anomaly detection needs mainly normal running data, so you can start without many failures. Failure classification and remaining-life estimates need many labeled failures or run-to-failure histories. Most plants should begin with anomaly alerts.
Can predictive maintenance run without sending data to the cloud?
Yes. Models can run on edge devices near the machines or on in-country or on-premises infrastructure, with alerts written to your maintenance system. Agree the design with your IT and OT security teams.
Which machines should we try first?
Pick a few machines of the same type, such as motors or pumps on one line, where unplanned stops hurt delivery and past failures are documented. Narrow scope makes the results easier to judge.
Ready to put this to work?
Turn this playbook into a working system
From a first agent to a national platform, the same team plans it, builds it and runs it with you.
