Predictive maintenance on a case erector has a reputation problem, and it is deserved. Most installations fail not because the sensors are wrong but because nothing was ever set up to act on the data. A machine instrumented with six sensors and no defined threshold produces a dashboard nobody reads, and the plant has spent money to move a breakdown from an unplanned stop to a planned one that still nobody schedules.

This article covers the six signals that genuinely predict failure on a case erector, the thresholds that separate normal variation from a developing fault, what the hardware costs, and the three failure modes where predictive maintenance actually returns the investment.

The six signals worth instrumenting

A case erector does not need comprehensive condition monitoring. It needs the small number of measurements that change before a failure, and it needs them at a resolution fine enough to see a trend rather than a snapshot.

Vacuum level and vacuum stability. Vacuum is the single most informative signal on the machine because vacuum loss is the mechanism behind most pick failures, most double feeds and a large share of jams. Normal operation sits at 0.4 to 0.6 bar at the cup manifold. The absolute value matters less than the stability — a station holding 0.55 bar with no ripple is healthier than one holding 0.50 bar that swings by 0.08 bar on every cycle. Ripple means a leak that opens and closes, which is usually a worn cup lip, a cracked hose or a partially blocked filter.

Servo current or torque. Rising servo current on a forming axis indicates increasing mechanical resistance: a dry or worn gearbox, a foreign object in the fold path, or a rail that has moved. A trend increase of 15 per cent above the commissioning baseline is worth investigating before it becomes a fault. Sudden spikes are different and usually indicate a jam rather than wear.

Cycle time and cycle time deviation. Cycle time drifting upward by a few milliseconds per cycle is the classic signature of a machine getting slower before it fails. The more useful figure is the deviation between cycles: normal variance on a healthy machine is below 3 per cent; above 5 per cent the machine is being slowed by something intermittent, whether friction, vacuum loss, board variation or a sensor that is beginning to fail.

Adhesive temperature and heater power. Where the machine seals with hot melt, the temperature reading alone is insufficient — a heater element losing capacity holds temperature at idle and falls behind under duty. The signal that predicts failure is the heater power required to hold setpoint, trending upward. A deviation of 5 degrees Celsius beyond the setpoint tolerance is the alarm threshold; a 15 per cent rise in heating duty is the predictive one.

Compressed air pressure and flow. Pressure verifies supply adequacy; flow reveals leakage and demand change. A plant that adds a machine elsewhere in the hall without uprating the compressor sees the effect here first, which is useful because the resulting intermittent pick failures on the erector look like a machine fault.

Vibration, root mean square acceleration. Broadband vibration on the frame and on the major rotating assemblies gives the earliest indication of bearing degradation and of mechanical looseness. The absolute figure is machine-specific; the trend is what matters, and a 20 per cent rise over a rolling four-week baseline justifies inspection.

Data architecture: keep acquisition away from control

The most common architectural mistake is to try to extract data by modifying the PLC program that runs the machine. This creates a change to a validated control system to satisfy a reporting requirement, and it is the wrong trade in every case.

The correct arrangement leaves control entirely alone. The PLC continues to run the machine and continues to expose its existing data through OPC UA or Modbus TCP. An edge gateway reads those registers, plus the standalone sensors wired to its own inputs, and publishes over MQTT to a broker, a local historian or a dashboard. Nothing in this path can affect machine operation, and the segregation is what makes the installation acceptable to a plant’s engineering and validation functions.

Two practical points follow. Configure the gateway for store-and-forward, so that a network interruption does not lose the trend data that makes the whole exercise worthwhile. And keep the sensor wiring physically separate from the control cabinet where possible, because a sensor fault that can trip a safety circuit is a worse outcome than the fault it was installed to predict.

If the machine is in a regulated production environment where the control system is validated, this architecture is the only one that avoids triggering a revalidation. The same principle governs change control in cleanroom-compliant installations, where the validated configuration is the baseline and reporting must be added alongside it rather than inside it.

The three failure modes where the investment returns

Predictive maintenance pays on unplanned stops that are both frequent enough to matter and predictable enough to be seen coming. Three failures on a case erector meet both conditions.

Cup and vacuum generator replacement. Worn cups and degraded ejectors cause pick failures that stop the machine and take 40 to 90 minutes to diagnose on a line where the cause is not obvious. Vacuum ripple and a slow decline in pickup success give three to five days of warning, which is enough to schedule a cup change during a planned break. This is the single highest-return instrumentation on the machine.

Gearbox and forming axis wear. Servo current trending upward gives weeks of notice. The maintenance action is lubrication or replacement during a planned shutdown; the alternative is a seized axis, which is a multi-shift repair and frequently damages the tooling as well.

Adhesive system heater degradation. A heater element losing capacity produces progressively worse bonds, which appear as intermittent adhesion failures before any alarm is raised. The mechanism is worth connecting to the larger adhesion picture covered in the tape and adhesive troubleshooting material, because a heater that is 15 per cent short of duty produces the same symptom as a badly selected adhesive.

Payback, calculated

Take a line running at 40 CPM producing cases with a realistic contribution margin. Unplanned stoppages of 45 minutes typically cost 1,200 to 4,000 USD each once scrap, lost output, operator time and restart losses are counted. The range is wide because it depends entirely on whether the line can be buffered and whether the stop lands in a peak window.

A monitoring package covering vacuum, servo current, cycle time and adhesive system duty costs roughly 2,000 to 3,500 USD per machine installed, including the gateway. If the package prevents 1.5 unplanned stops per year, the payback interval is 12 to 24 months. Beyond that interval the instrumentation is delivering free reliability for the remainder of the machine’s life.

The calculation is sensitive to one assumption: the cost of a stop. Rather than debate it, measure it. Record the duration and the recovered output of the next three unplanned stops, and use the average. An assumption measured on your own line beats any benchmark figure, including the one above.

What not to do

Three failure patterns account for most disappointing installations.

Instrumenting signals with no action attached. If there is no defined maintenance action for a threshold being crossed, the signal is noise. Before installing a sensor, write down what the maintenance team will do when it alarms. If the answer is “investigate”, the sensor is not ready to be installed.

Collecting at too coarse a resolution. A vacuum reading averaged over an hour hides the ripple that indicates a failing cup. Where the signal of interest is per-cycle, acquisition has to be per-cycle, and a PLC scanning at its normal rate already provides it.

Alarming on absolute values instead of trends. Absolute thresholds generate false alarms during normal variation and product changes. Trend comparison against a recorded baseline is what makes the signal useful, which is why establishing the baseline comes first and why commissioning records matter.

Implementation in three stages

Stage one: manual baseline, four weeks. Record vacuum, cycle time, adhesive duty and any stop events by hand on a shift sheet. No instrumentation, no capital. Four weeks of manual records establish the range of normal operation for this machine on this product mix, and they establish whether the stop rate even justifies instrumentation.

Stage two: instrument and validate. Install sensors and the gateway, and run them alongside the manual record for two weeks. The purpose is to confirm the instrumented readings agree with the observed baseline. Where the two disagree, the sensor placement or the scaling is wrong, and it is better to find that before anyone trusts an alarm.

Stage three: connect alarms to the spares process. Configure thresholds against the validated baseline, and connect each threshold to a specific spare part and a specific maintenance action with an owner. Set the trend report to be reviewed at a fixed interval — monthly is usually right — because a trend only predicts anything if somebody looks at it.

FAQ

How many sensors does a case erector actually need?

Four signals cover most of the value: vacuum level and ripple, servo current on the forming axes, cycle time deviation, and adhesive system duty. Vibration and air flow are useful additions on machines with known gearbox or air supply issues. Beyond that the marginal signal usually has no action attached.

Can I read data from the machine without touching its PLC program?

Yes, and this is the correct approach. Use the PLC’s existing OPC UA or Modbus TCP interface and read the registers through a separate gateway. Never modify the control logic to obtain reporting data.

Will this work on an older machine?

Older machines often make better candidates, because their mechanical condition is more variable and their failure modes are more predictable. The constraint is whether the existing controller exposes data at all; where it does not, standalone sensors on the gateway’s own inputs bypass the question entirely.

What is the smallest useful installation?

Vacuum plus cycle time deviation, on one machine. Both signals come from hardware that is inexpensive and easy to mount, and between them they catch the two most common causes of unplanned stops in the list above.

How long before the data is useful?

Roughly six weeks: four weeks of manual baseline and two weeks of validation. Trend-based prediction needs a baseline more than it needs history, and a machine without a recorded baseline cannot produce a meaningful trend no matter how much data is collected.

Does this replace scheduled maintenance?

No. It supplements it by moving certain interventions from a calendar to a condition. Calendar maintenance still covers lubrication, cleaning and inspection items that produce no useful signal.

Start with the baseline, not the sensors. If you can supply the current stop pattern, cycle time and vacuum readings for a machine, we can identify which of the six signals will produce a return on your line and which will not. Send the figures to [email protected] or contact us on WhatsApp at +86 13681839278. We will confirm the sensor set, the thresholds to configure against your machine’s own baseline, and the maintenance actions that each alarm should trigger, so the installation produces decisions rather than dashboards.