Predictive Maintenance In Manufacturing, Step By Step

Key Highlights

• A first phase on one critical asset needs a handheld vibration meter, technician time, and a structured breakdown log. Fixed sensors and AI can wait for its results.
• Eight weeks of warning on a bearing with a six-week parts lead time leaves two weeks to act.
• Poor fitting and contamination damage bearings before readings can show it, so mounting practice matters too.

How Mid-Size Plants Can Start Predictive Maintenance In Manufacturing

A vibration reading can show a failing bearing weeks or months before it stops a production line. That warning prevents downtime only if it comes earlier than the time it takes to get the part and a crew. It also needs one person who is responsible for acting on it. Those two conditions decide whether predictive maintenance reduces unplanned stops at your manufacturing plant.

For a mid-size plant, the first phase can be small and inexpensive. One critical asset, a structured breakdown log, monthly handheld readings, and alert levels from a published standard are enough to test the approach. The phase should end with a funding decision, based on your own data, about fixed sensors and machine learning.

What Is Predictive Maintenance In Manufacturing?

Predictive maintenance in manufacturing is a way of scheduling repairs based on the measured condition of each piece of equipment. Sensors or handheld readings track signs of wear, such as vibration and temperature. When a part starts to wear out, your team replaces it during a planned stop, before it breaks and halts production.

Predictive maintenance suits parts that wear gradually, such as bearings in motors, gearboxes, pumps, and fans. At larger scale, machine learning models analyze continuous sensor data to flag changes across many assets at once. Most plants run predictive maintenance alongside the two other common approaches.

What The First Phase Costs

The business case starts with the cost of unplanned downtime on the asset you choose. Your existing maintenance and production records should hold the figures for one recent breakdown.

• Lost output, meaning the hours of downtime times the value of an hour of production
• Scrapped material
• Overtime for the repair crew
• The premium paid for rush parts

For example, six hours of downtime at $5,000 an hour is $30,000 in lost output. Add $2,000 of scrap, $1,500 of overtime, and a $1,000 rush premium, and one breakdown costs $34,500. A handheld vibration meter lists at about $2,800 from a US test equipment distributor in 2026, so preventing one such breakdown would repay it more than ten times. Replace these illustrative figures with your own.

Plan for the baseline to take about three monthly readings. The decision on fixed sensors takes longer, because it needs at least one rise in readings ahead of a repair. On a reliable asset, that can take most of a year, so agree on the timeline with your executive team before you start.

How Predictive Maintenance Works On One Production Line

Consider a conveyor drive: the motor and gearbox that move product between stations and stop the line when they fail. The objective is to detect bearing wear early enough to schedule the replacement during a planned stop. Each stage below has a named owner, because an alert without one tends to go unanswered.

Choose The Asset That Stops Production

Start with the asset whose failure halts the whole line. A conveyor drive qualifies, because every station after it stops too. Assets with a standby unit can fail without halting output, so they can wait. Your maintenance lead and production manager should agree on the choice, since they know which failures have cost the most. Keep the first phase to that one asset until it has completed a full cycle, including a repair. Each extra asset adds readings, alerts, and log entries for a small team to follow.

Standardize The Breakdown Log

The breakdown log is the first data asset the program creates. Technicians record every breakdown in the same six fields, in your maintenance software or a shared spreadsheet. If past entries are thin, start the fields now and treat the next breakdown as the first entry.

• Date Of The Breakdown
• Part That Failed
• First Warning Sign (Noise, Heat, Or Vibration)
• Cause Found During The Repair
• Repair Made
• Hours Of Lost Production

Each field needs a list of fixed choices. A note such as “fixed motor” cannot be counted or compared later. We saw this while working with SKF, which Accedia’s team enhanced and maintained for three years. The tool logged why each machine part came out of service in the same categories, such as vibration, noise, or temperature. That consistency supported SKF’s decisions to reuse, remanufacture, or scrap each bearing.

Measure Vibration With A Handheld Meter

Vibration is the most practical first measurement for a motor and gearbox. It shows bearing wear, shafts that have drifted out of line, and parts that spin unevenly. A 2026 guide to warning times puts its typical warning time at one to nine months, while thermal imaging gives a much shorter warning.

A technician takes each reading at the same point on the housing, at the same speed and load, so every result compares with the last. Until your own history shows the warning time for this asset, read monthly. Once it does, read at least twice within that warning time.

Readings will not catch every part failure. SKF’s 2026 bearing maintenance page attributes 16% of premature bearing failures to poor fitting and 14% to contamination, and both begin before wear is measurable. The rest, poor lubrication and fatigue, build gradually and give readings time to warn you.

Set Alert Levels And Assign An Owner

The first alert levels come from a published standard. In ISO 20816-3, the International Organization for Standardization (ISO) sets guideline vibration levels for industrial equipment above 15 kilowatts (about 20 horsepower). That includes conveyors and electric motors with a flexible coupling. Severity-grading meters show where each reading falls on a similar scale, from good to unacceptable.

• A limit alert, set where readings leave the satisfactory band. For motors under 15 kilowatts, the motor maker can supply the value.
• A rise alert, triggered by a sustained increase over the asset’s baseline. Your maintenance lead sets the size of that increase.

The baseline is the average of readings on a healthy machine part once they stop changing from one visit to the next. Your team adds both alert levels to the reading history, kept with the breakdown log, so an out-of-range reading is flagged automatically.

Every alert needs one named owner, usually the maintenance lead, and a backup for nights and weekends. If alerts trigger during normal operation, apply these methods for reducing false predictive maintenance alerts early, before the team stops trusting them.

Turn Each Alert Into A Planned Repair

1. Your maintenance lead confirms the alert with a second handheld reading, to rule out a bad measurement.
2. The planner opens a work order and checks stock for the bearing.
3. With the part reserved, the planner books the replacement into the next planned stop with the production manager.

Warning time has value only when it exceeds the time needed to get the part and a crew. Eight weeks of warning on a bearing with a six-week parts lead time leaves two weeks to act. Keep a spare on site for any part your supplier cannot deliver within the expected warning time.

When readings climb faster with each visit, less warning time remains, so bring the repair forward. If the next planned stop is too far away, compare the cost of a short extra stop with the cost of a breakdown.

Learn From Every Replaced Bearing

Fit the new bearing with the correct tools, since poor fitting is one of the failures readings cannot detect in time. Bearing Assist, the mounting app our team develops with SKF, guides technicians through each installation step by step for exactly this reason.

Each repair should also improve the data. Close the work order only once the record is complete.

• Photograph the removed bearing.
• Enter the cause found during the repair in the breakdown log.
• Compare the damage with the alert. If the wear was more advanced than the alert suggested, lower the alert level for the next bearing.

The cause field is the hardest to keep consistent, because two technicians can describe the same damage differently. We built the Accedia AI Defect Detector for this step. It identifies the component and each defect from a photo, grades the part from Good to Critical, and stores every analysis so similar failures can be searched later. Cases it is unsure about go to an engineer.

When To Fund Fixed Sensors And AI

Fixed sensors are worth funding once handheld readings have shown a clear rise ahead of at least one repair. Wireless vibration sensors bought outright list at roughly $350 to $580 each, according to a 2026 sensor maker comparison. The same comparison cites third-party estimates of about $35 per sensor per month for subscription models, so one year of rental costs roughly the purchase price.

Controls engineers install the sensors, and the rest of the work sits with your team.

• A network for sensor data kept separate from the machine controls, to limit the damage of a cyberattack
• Access rules that decide who can open the readings
• Internet of Things software that collects the data and links confirmed alerts to your maintenance software, so work orders open without retyping

Older equipment often has no network connection, and getting live data off legacy equipment adds wiring and hardware to the budget. Plan for that before ordering sensors for an asset that predates your network.

Getting Data Off Older Equipment

Before ordering sensors, check whether the drive’s controller already records usable data. Some controllers do, some speak protocols their makers will not document, and some output nothing until a retrofit sensor is added. Our article on getting live data off legacy equipment covers the gateways and translation work involved.

Controls engineers install the sensors, and the rest of the work sits with your team.

• A network for sensor data kept separate from the machine controls, to limit the damage of a cyberattack
• Access rules that decide who can open the readings
• Internet of Things software that links confirmed alerts to your maintenance software, so work orders open without retyping

Connecting older equipment also opens new security risks. We check the security of connected devices and their hardware protocols against standards from the Open Worldwide Application Security Project (OWASP) and frameworks such as TISAX, the automotive industry’s information security assessment.

When Machine Learning Adds Value

Machine learning models need continuous data. Once fixed sensors report around the clock, a model can track vibration, temperature, and load together and learn each asset’s normal pattern. A small team can then watch dozens of assets, which monthly handheld readings cannot support. The breakdown log from the first phase becomes the labeled history that model alerts are checked against.

What The First Phase Should Decide

A mid-size plant can start predictive maintenance with one critical asset and a consistent breakdown log. Every alert also needs an owner with the authority to act in time. That first phase costs one handheld meter and your team’s time, and it shows whether the approach works on your equipment.

Its results give your executive team the basis for the next decision. If the readings warned early enough and each alert led to a planned repair, fixed sensors and machine learning have a business case built on your own data. Otherwise, the first phase has shown what to fix before you spend on sensors.

Our Manufacturing team builds the software side of programs like this, from data collection on older equipment to links between alerts and maintenance systems. Contact our manufacturing team to discuss where to start on your line.