Start your predictive maintenance implementation with one machine whose failure truly costs the business. First, choose an asset that stops the line, creates quality losses, requires expensive service, or disrupts the production schedule. Then calculate the cost of failure, check the data, and define what the team should do after an alert.
This article is for you if you are responsible for maintenance, production, automation, or plant development and want to move from a test to a solution that actually works on the shop floor. Let’s start from the beginning.
Predictive maintenance is a maintenance strategy that uses machine condition data and failure risk to plan service activities.
Instead of replacing a part every six months or reacting only after the line stops, you monitor signals that may indicate a machine’s condition is getting worse. These may include:
Predictive maintenance can also be described as maintenance based on forecasting failures from observed data, such as temperature, noise, or vibration. ISO 17359, the standard for machine condition monitoring, also indicates that a good program should begin with an analysis of costs, the importance of machines to production, and failure modes.
A useful rule of thumb is this: prediction does not start with an algorithm. It starts with identifying which failure hurts your plant the most and whether you can see it earlier in the data.
Most of the time, it is not because the technology does not work. The reason is usually simpler. The pilot looks good in a presentation, but it does not change the day-to-day work of the maintenance team because:
Manufacturing companies often struggle with data quality, failure descriptions, and turning analysis into action. Research reviews on predictive maintenance also confirm common barriers: noise in industrial data, incorrect readings, different machine operating conditions, and models that work only for one specific asset.
That is why the goal of the first implementation should be an earlier maintenance decision that the team understands and can act on.
Get 5 chapters of the book for free!
Join the newsletter and gain access to 40% of the book
”15 Steps to Buying an Information System”
At the start, choose one machine, not an entire line. This will make it easier to monitor the outcome.
The best first candidate is a piece of equipment that:
In its report The True Cost of Downtime 2024, Siemens states that in a large automotive plant, one hour of unplanned downtime can cost $2.3 million. This number should not be applied directly to every plant, but it clearly shows why selecting the first machine matters more than selecting the technology.
A simple table can help.
| Question | Good Signal | Risk |
| Does the failure stop the line? | Yes, downtime is visible in OEE | Low impact makes the project harder to justify |
| Has the failure already repeated? | There are maintenance work orders and shift reports | No history makes evaluation difficult |
| Are there symptoms before failure? | Vibration, temperature, current, pressure | Sudden failure with no symptoms is a poor starting point |
| Does maintenance know what to check? | There is an inspection procedure | An alert without action loses its value |
| Does the machine run often? | Data is generated daily | Limited operation means limited data |
You do not need a perfect financial model right away. You need a number that helps you make a decision. Include:
A pump failure stops the line for five hours. Each hour of downtime means $18,000 in lost margin. Service, parts, and defects cost another $12,000.
Cost of one failure:
5 × $18,000 + $12,000 = $102,000
If this failure happens four times a year, the annual cost is approximately $408,000.
This makes it easier to decide whether monitoring, data integration, and team involvement make financial sense.
McKinsey reports that predictive maintenance can reduce machine downtime by 30 to 50% and extend machine life by 20 to 40%. The U.S. Federal Energy Management Program also indicates that PdM programs can deliver 8 to 12% savings compared with preventive maintenance alone.
Naturally, treat these numbers as a reference point, not as a guaranteed result in every plant.
Saying that you want to predict press failures is too broad. A better goal might be:
This will help you choose the right data, response threshold, and person responsible for checking the machine.
If you cannot identify a failure mode, start with a workshop involving maintenance, automation, and production. The team can usually point to three to five recurring problems that create the most stress and cost.
You usually do not need to start with new sensors. Some of the data already exists; it is just scattered across different systems.
Check:
A vibration chart without context can be misleading. A machine behaves differently during startup, product changeover, and stable operation.
McKinsey points out that data quality is one of the barriers to AI applications in manufacturing. That is why, before building a model, you need to organize the basics: machine names, event times, failure causes, and the way maintenance responses are recorded.
Not every predictive maintenance initiative requires an advanced model from day one.
Sometimes it is enough to detect a trend:
This can give the team several days or hours to respond. Often, that is all you need to order a part, plan downtime, and avoid a repair at the worst possible time.
A predictive model makes sense when you have historical data, labeled failures, and a repeatable pattern. Without that, the system may look impressive but fail to help the people working on the shift.
This stage often determines whether the implementation succeeds.
An alert should clearly tell you:
A sample process may look like this:
People are more likely to use the system when they see that an alert leads to a specific decision, not just another notification.
Below is a simple plan that helps keep the pilot from dragging on indefinitely.
| Stage | What to Do | Outcome |
| Days 1 to 15 | Choose one machine and calculate the cost of failure | Clear business case |
| Days 16 to 30 | Define one failure mode and check the data | You know what to look for |
| Days 31 to 50 | Collect data from normal machine operation | Baseline established |
| Days 51 to 70 | Set up an alert and response procedure | The team knows what to do |
| Days 71 to 90 | Evaluate the results and decide on the next step | Decision: improve, scale, or stop |
After 90 days, check whether:
Predictive maintenance does not solve every maintenance problem.
It is better to start with something else when:
In that case, a better first step may be organizing maintenance work orders, analyzing root causes, monitoring basic parameters, or changing the preventive maintenance schedule.
If you want to start by organizing data from machines, production, and maintenance, explitia can help you take the first step.
This is especially useful when you already have data in PLCs, SCADA, MES, or a maintenance system, but the team does not have one shared view that supports fast decision-making.
A predictive maintenance implementation often starts with the need to connect data so that maintenance and production can see the same problem at the same time.
Run a small audit of one machine. Choose the asset whose failure hurts production the most. Calculate the cost of one hour of downtime and check what data you already have. Then describe one failure mode and the team’s response after an alert.
If this exercise reveals a real cost, available data, and a clear maintenance decision, you have a strong starting point. In that case, predictive maintenance can become a calmer, more reliable way of working for both maintenance and production.
Predictive maintenance is maintenance based on machine condition data and predicting the risk of failure. Unlike preventive maintenance, it is not based only on time or the number of cycles, but on signs that a machine’s condition is getting worse.
The best way to start is with one machine that has a high cost of failure. Then calculate downtime cost, choose one failure mode, check the available data, set up monitoring or alerts, and define how the maintenance team should respond.
Not always. The first stage can be based on condition monitoring, trends, and alarm thresholds. AI makes sense when you have good historical data, documented failures, and repeatable patterns.
A meaningful first stage can be planned for about 90 days. That is enough time to choose a machine, check the data, set up alerts, and evaluate whether the project creates operational value.
The most common issues are choosing the wrong machine, lack of failure data, no alert response procedure, too broad a scope, and measuring success by technology instead of its impact on downtime, costs, and the maintenance team’s work.
