AI thermography on the aerospace CNC spindle — the bearing is changed on schedule, not on breakdown.

The problem

By the time the spindle warns you, the structural part is already lost.

The spindle of an aerospace aluminum CNC is the most expensive item in the cell and the one you can least afford to change at the wrong moment. The failure always follows the same script:

1. The spindle bearing goes into fatigue — micro-seizures, coolant contamination, a peak axial load during a finishing pass.
2. It starts running hotter than normal, long before it moves a single micron of vibration.
3. The PLC control fires the pre-alarm when the thermal pattern is already off, too late to get the structural part in progress out safely.
4. The spindle is changed on breakdown, not on schedule. The cell stops for days. The aerospace part in the queue —a wing rib, a structural bed— is rescheduled or reworked, depending on what was left inside.

The maintenance manager knows what is happening. His people know what is happening. What he does not have is a learned thermal pattern of a healthy spindle running its normal recipe to compare against what the operator is seeing on the screen. That pattern is built —if it is built at all— in a spreadsheet only known to the person who retires next year.

How it fits the IRIS system

iLEAN Edge does not add another sensor — it learns the thermal pattern of the healthy spindle and watches the drift.

The spindle problem is not a shortage of data: the CNC's PLC already has spindle temperature, vibration and headstock power draw. What is missing is a layer that learns the normal heating pattern per recipe and per tool, compares it in real time against what the thermal camera is seeing, and cross-checks the drift with the MES production plan and the real spindle hours. That layer is iLEAN — the filler that covers what ERP, MES, SCADA and CMMS left exposed.

Edge sees the bearing's thermal pattern. Connect captures the real spindle hours and the batch recipe. The agent cross-checks against the aerospace plan and proposes the change window. The person signs — the spindle does not change itself.

The three iLEAN pieces applied to aerospace CNC spindle thermography:

• Edge — a physical terminal in the cell with a fixed thermal camera aimed at the spindle housing. A CNN trained on the normal heating pattern per machining recipe (roughing vs. finishing, tool used, speed). It detects the drift against the pattern, not a fixed threshold. It raises the alert in the CMMS and a status light in the cell. It works without the network: if the plant loses WiFi, Edge keeps watching and recording the pattern.
• Connect — captures the real spindle hours from SCADA, the recipe from MES, the change history from the CMMS, and the maintenance manager's spreadsheet where he tracks the bearings replaced over the last few years. And it captures from outside the delivery note of the spare bearing, the supplier's lead-time notice, the aerospace customer's message about urgent parts.
• Agent — cross-checks the thermal drift from Edge with the aerospace production plan, the real hours and the bearing stock. It proposes a change window that does not cut any structural part in half. If the drift accelerates, it raises the priority of the alert. The maintenance manager signs the work order and the change goes into the schedule.

See the full IRIS architecture →

Before and after

Spindle by ear vs. spindle with a learned thermal pattern

Aspect	Classic maintenance	With iLEAN Edge thermography
Normal spindle pattern	In the maintenance manager's head	Learned per recipe and tool, checked in real time
Drift detection	When the PLC fires the pre-alarm	Much earlier — at the subtle change in thermal pattern
Bearing change	On breakdown, cell stopped	On schedule, without cutting an aerospace part in progress
Structural part in the queue	Rework or scrap whatever was left inside	Cleared before the change, no rework
Operation without network	n/a	Edge keeps watching on cabinet power
The veteran's knowledge	Retires with him	Captured as a learned pattern in Central

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Tier 2 aerospace plant with 6 to 12 aluminum CNC cells, structural parts with demanding tolerances, roughing + finishing recipes on the same machine.
• Edge pilot on one critical cell (thermal camera + integration with MES and CMMS). First expected value within a few weeks: learned thermal pattern and first drift detected.
• Indicative payback between 4 and 9 months, depending on the historical frequency of spindle failures and the average cost of an unscheduled change.
• Expected reduction of unscheduled spindle failures ≥ 30% over the baseline, and practically zero structural parts left half-machined by a sudden failure.
• Hard lever: a single spindle failure avoided + zero aerospace parts scrapped mid-cycle. One avoided failure pays for the pilot.

And the IT director's reasonable doubt

«What if the AI gets it wrong and stops my spindle when it was fine?» — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI does nothing but compare a thermal pattern against a learned pattern, the best models brought the error below 1.5% ^[1]. And even so, critical decisions are not made alone: iLEAN proposes the change window and the person signs. The three safety rings are there precisely for this.

^[1] OpenAI paper «Why Language Models Hallucinate», 2025 — on the reliability of AI in anchored tasks.