The PM Report Came Out That Day, And The Team Leader's Face Turned Livid
That afternoon, the production line's PM (Preventive Maintenance) report had just been released. The team leader's face instantly changed as he looked at the data. He pointed to one of the machines' MTBF (Mean Time Between Failures) and said to me, "Ah Han, look at this machine. It broke down again not long after the last PM. Now its MTBF has dropped to 1.08, and its CPK is only 0.85. If this continues, how will we deliver next week's orders?" I walked over to look. Sure enough, that old etching machine, despite undergoing regular preventive maintenance, still frequently crashed recently. Each downtime lasted several hours, causing everyone to be in a frantic mess. You tell me, should this PM be done or not?
Where Is The Problem? It's Not That Maintenance Is Useless, But That The Method Is Wrong!
In fact, this is a situation we often encounter. Traditional "Preventive Maintenance" (PM) involves periodically and quantitatively replacing parts and cleaning machines. It's like your car or motorcycle, where you change the engine oil and check the tires every certain mileage. The advantage is that it's simple, clear, and has a fixed schedule. But the drawbacks are also obvious: the parts you replace, to be honest, might still be usable for a long time—that's waste. Conversely, some parts might fail prematurely before the PM time arrives, catching you off guard, just like our etching machine which had issues again not long after PM, with its DPMO directly soaring to 6210.
In other words, preventive maintenance is like "treating illnesses when they arise, and taking tonics even when healthy"—it reduces risk but is not highly efficient. "Predictive Maintenance" (PdM), however, is different. It's more like "prescribing precise medication based on bodily conditions." It involves collecting real-time data from machines through sensors, such as vibration, temperature, current, and then using data analysis to predict when a machine might encounter problems.
How Is It Actually Done? By Looking At Data, Not Calendars
The core of predictive maintenance is shifting from "period-based" to "condition-based." Here are some common practices:
- Vibration Analysis: Many machine bearings and motors exhibit abnormal vibration frequencies before they fail. We install sensors to monitor these subtle changes. If a certain vibration frequency starts to deviate from the baseline value—for example, gradually rising from a normal 0.1mm/s to 0.5mm/s, or even reaching 1.0mm/s—then you know this bearing might be nearing its end, and you can schedule its replacement in advance.
- Infrared Thermal Imager: Overheating of some internal machine parts is often a precursor to failure, especially for high-power components like electrical panels and motors. Regularly scan with a thermal imager. If a spot's temperature is abnormally high—for example, suddenly jumping from a normal 40°C to 70°C, or even 80°C—it likely indicates aging wiring or poor contact, which can be addressed proactively.
- Oil Analysis: For machines that use lubricating oil, regularly draw oil samples for testing, analyzing metal particles, moisture content, and other factors. If copper and iron filings are found to be abnormally increasing—for example, suddenly rising from 5ppm to 50ppm—it indicates severe internal wear of the machine, and maintenance can be scheduled in advance.
So the key point is that we no longer replace parts simply because "time is up," but rather "take action" when the data tells us it's "about to fail." This way, we avoid wasting still-usable parts and prevent the disaster of unexpected downtime.
The Most Common Pitfall: Too Much Data, No One Understands It!
To be honest, the most common problems encountered when implementing predictive maintenance are not technical feasibility, but rather "data explosion" and "insufficient manpower." You install a bunch of sensors, and data pours in every second. What's the result? Everyone is busy firefighting and has no time to look at those charts or analyze those trends. Often, a machine has already failed, and only upon reviewing the data later do you realize, "Oh, it had already issued warnings!"
I remember one time, to track the lifespan of a certain pump, we installed a bunch of pressure sensors. As a result, the system issued an alarm every five minutes, but everyone was already overwhelmed dealing with production line issues. Who had time to pay attention to those "not-yet-failed" alarms? Eventually, there were too many alarms, everyone became desensitized, and legitimate alarms were ignored. Frankly, it boils down to too many "false alarms," causing people to lose confidence. Honestly, at this point, someone is needed to "translate" the data into actionable information, rather than just throwing out a pile of numbers.
One Thing You Can Do Today
Start with a "pain point" machine and implement the simplest temperature or vibration monitoring.