Hour 61 of the Lyo Cycle Is the Worst Time to Lose Vacuum. The Leak-Rate Trend Called It Nine Days Earlier.

64%, and the number your maintenance planner already writes down every cycle

Predictive maintenance is now the most adopted use of AI in manufacturing — around 64% of manufacturers, ahead of quality control at 58% and supply chain at 52%. That gets quoted in rooms where nobody can name a machine. So name one. In a sterile fill-finish contract manufacturer the machine is the freeze dryer, and the number is the chamber leak rate your maintenance tech writes at the end of every cycle on a log sheet in a binder in the mechanical room.

That binder is the whole story. The value in it is not any single reading — every reading passes — it is the slope across the last ten cycles, and nobody has time to draw the slope. What changed in 2026 is not new sensors. It is that the readings you already take, together with the photographs and thermal images from rounds you already walk, are watched continuously instead of being checked against a service interval printed in 2019.

The definition, plainly: predictive maintenance means the machine is serviced when its own trend says it is drifting, not when the calendar says twelve months have passed. In a plant where a single unplanned failure can cost more than a year of maintenance labour, that distinction is the whole argument.

Hour 61 of a 68-hour cycle, and what it takes with it

Picture the worst version, because most sterile CDMOs have either lived it or been two hours from it. An 18,000-vial lot of a client's clinical drug substance goes into the lyophilizer on a Saturday. Freezing down to −45 °C on the shelves, then primary drying under about 100 mTorr for the better part of two days, then secondary drying, then stoppering under nitrogen backfill. Somewhere around hour 61 the chamber pressure starts climbing off setpoint. The condenser cannot hold temperature. The vacuum pump is at its limit and the product temperature crosses the collapse point on half the shelves.

What you lose is not one batch. You lose the client's drug substance, which for a biologic took eighteen months to make. You lose the suite for two to three weeks while you clean, investigate and requalify. If the repair means opening the chamber and touching the sterile boundary, you are into a change control and quite possibly a fresh aseptic process simulation before you fill again — and your fill schedule is booked eleven weeks deep, so every program behind this one slips. You write a deviation under 21 CFR 211.192 and, because the leak test has been drifting, you run a look-back across every lot since the last unambiguously good result. Your client's clinical site loses a dosing window.

All of that from a vacuum pump that had been telling you for nine days.

The signals that show up before the alarm does

None of these are exotic. Every one of them is already being measured or already being written on a log sheet in a sterile facility that runs a lyophilizer.

• Post-cycle leak rate — the value itself stays inside the acceptance limit, but the slope across the last eight or ten cycles points one way.
• Condenser pull-down time — the minutes to reach −55 °C creeping from twenty-two to thirty-four across a campaign is a refrigeration circuit losing capacity.
• Vacuum pump ultimate pressure and oil temperature — the pump's own quiet decline, usually visible a full campaign before it stops holding under load.
• Comparative pressure measurement — the gap between the capacitance manometer and the Pirani gauge at end of primary drying moving cycle over cycle.
• Silicone oil reservoir level — a couple of millimetres a cycle is a heat transfer fluid leak in the making, which is the failure that contaminates product rather than merely stopping it.
• Compressor amp draw and discharge temperature, plus the thermal image your facilities tech takes on rounds and drops into the folder nobody opens.

The 2026 difference is that these live in different places — the batch report out of your DeltaV or SIMATIC control system, the historian, a log sheet, a phone camera roll, the work order history in Blue Mountain RAM — and the machine reading them holds all of it at once and says a plain sentence: leak rate has risen 40% over eight cycles while condenser pull-down has lengthened by nine minutes; these two together usually mean a seal, not a gauge. That sentence is what your facilities engineer would have written if he had had a free Thursday, which he has not had since March.

flowchart TD
  A["Lyo cycle ends, leak-rate test run"] --> B["Readings land in the historian and the batch report"]
  B --> C["Agent compares against the last ten cycles"]
  C --> D{"Leak rate or pull-down time drifting?"}
  D -->|No| A
  D -->|Yes| E["Work notification raised in Blue Mountain RAM"]
  E --> F["Planner books it into shutdown week"]
  F --> G["Change control and requalification by validation"]
  G --> A

You get one week a year to act on this, and it is usually in July

This is a scheduling problem more than an engineering one. You cannot open a qualified freeze dryer whenever you like. The sterile suite is booked for months, every intervention drags a change control and requalification behind it, and the only realistic window to break into that chamber, change pump seals, service the refrigeration circuit and re-run the leak test is the annual shutdown — the week around 4 July at most US sites, or the week between Christmas and New Year.

Which means the decision is not "should we fix this." It is "do we know about it by mid-May." A drift detected in June goes into the shutdown scope, gets planned, gets parts ordered, gets a validation engineer assigned and a protocol written. The same drift detected in September gets carried for ten more months, and everyone in the room knows they are gambling. Watching the trend continuously moves the discovery date forward by weeks, and weeks is exactly the currency this decision is priced in.

What a lost cycle costs against what watching costs

Illustrative numbers for a mid-size sterile CDMO — use your own batch value and slot rate.

If watching the trends moves one event from unplanned to planned once every five years, that is $2.04M against roughly $228,000 of planned work and watching over the same five years. The tool does not need to be clever. It needs to be right once.

What the trend cannot tell you

A rising leak rate does not tell you where the leak is. It narrows the search — door seal, isolation valve seat, a shelf fluid connection, an instrument port — and your facilities engineer and the equipment vendor's service team still find it, still decide whether the fix is inside the sterile boundary, and still own the call on whether the unit runs the next campaign. That call is a judgment about patient risk, not a reading.

Two more honest limits. First, this only works on units with enough cycle history to have a shape; a freeze dryer that runs six campaigns a year takes a long time to build a meaningful trend, and on those units your best signal is still the tech who has listened to that vacuum pump for eleven years. Do not throw him away. Give him the chart. Second, every one of these decisions carries a paperwork tail — an unscheduled intervention on a qualified unit means change control, and possibly requalification, and possibly an aseptic process simulation before you fill product again. A trend alert that skips that tail creates a worse problem than the one it solved.

Start small on Monday: pull the last twenty-four cycles of leak-rate results and condenser pull-down times off the log sheets and the historian into one sheet, and plot them. If they are flat, you have learned something for the cost of an afternoon. If they are not, you have just found your shutdown scope while there is still time to order the parts.

Frequently asked questions

Does adding this count as a change to a qualified system?

Reading data that already exists and drawing a chart does not change your equipment or its control system, and generally sits outside change control as a monitoring aid. The moment anything writes back into the control system or triggers an automated action, you are changing a qualified system and it goes through change control. Keep it read-only and this stays simple.

Our lyo is from 2011 and barely talks to anything. Is it out?

No. The two most useful signals — post-cycle leak rate and condenser pull-down time — are already recorded by hand at nearly every site. A trend built from typed-up log sheets is worth more than no trend, and older units drift more predictably than new ones. The camera on your tech's phone covers the visual and thermal side.

Should we start with the freeze dryer or the water system?

Start with whatever costs you most per hour when it stops mid-run. For most sterile CDMOs that is the lyophilizer, because a failure there destroys product already committed. If you are oral solid dose, the same logic points at your fluid bed or your tablet press tooling instead — the machine that fails with a batch already inside it.

Our client asked whether AI is making our maintenance decisions. What do we say?

That it is not. The trend raises a work notification; a named planner schedules it; a named engineer performs it; validation approves any requalification. Show the audit trail on one work order and the question usually ends there.

One footnote from the operations side. Every unplanned event of this kind generates a wave of phone calls in the same forty-eight hours — the client's program manager, the equipment service company, three parts suppliers, and the business development lead of the customer whose slot just moved. CallSphere builds AI voice and chat agents that keep your main line answered and log who called about what while your engineering and quality people are standing in front of the machine. It does not read your vibration data; it makes sure the phone stops being one more thing on fire.