Powder metallurgy compaction — parts in tolerance shift after shift, not just in the SPC sample.

The problem

The green part falls out of tolerance by a few micrometres nobody looked at together.

In powder metallurgy compaction, the green height and final density of the part are the result of a fine balance between dosed powder weight, closing hydraulic pressure and tooling temperature. SPC looks at one sample every so many cycles. By the time that sample confirms a deviation, hundreds of parts have already been made with the same drift — some pass, some do not, and all of them go into the furnace. The problem is not SPC; it is what happens between one sample and the next:

1. Powder moisture changes with the shift (hopper openings, warehouse conditions) and alters the real compressibility of the powder in the die.
2. The supplier batch arrives with a compressibility slightly different from nominal and nobody retunes the closing pressure.
3. The die heats up as the shift goes on, and that shifts green height by a few micrometres.
4. The hydraulic oil filter starts to clog and the real pressure at the cylinder drops without the setpoint changing.

The classic system (fixed setpoint + sampling SPC + an experienced shift supervisor) works nearly always. When it does not, the bill is paid in post-sintering grinding, in scrapped critical parts (hydraulic pumps, synchros) or by an automotive customer that detects the dimensional deviation at goods-in. It is not a lack of process; it is information living in silos at the critical moment.

How it fits the IRIS system

iLEAN Edge does not add another SPC — it seals the crack between pressure, powder weight, batch and die.

The piece that solves the pain is iLEAN Edge: a physical terminal on the shop floor, next to the press cabinet, that watches hydraulic pressure the way a veteran press supervisor would if the powder scale, the die thermocouple and the circuit filter were all on the same screen at the same time.

Edge sees the hydraulic pressure curve of every cycle. Connect reads the batch, the powder supplier ticket and the ambient humidity wherever they live. The agent cross-checks pressure + batch + die and proposes a setpoint adjustment or a service to the shift supervisor. The person signs — the part stays in tolerance.

The three iLEAN pieces applied to hydraulic pressure in PM compaction:

• Edge — a terminal next to the press cabinet, reading main cylinder pressure, cushion pressure, ram position and, where it exists, ejection force. The CNN learns the normal signature of the pressure curve per SKU and per tool — not the setpoint, but the shape of the cycle. It cross-checks against the powder weight dosed at fill and the tooling temperature. It works without a network: if connectivity to the MES drops, Edge keeps watching cycle by cycle, because what is critical cannot depend on the WiFi.
• Connect — captures the powder supplier batch (scale ticket, compressibility certificate, packing date), warehouse ambient humidity, the die plan and SKU changeovers. And it also captures what arrives from outside — a supplier email warning that the next batch comes with a different baseline, a lab observation about the powder grade — at second zero.
• Agent — cross-checks the pressure-density signature cycle by cycle against moisture, batch and die temperature. When the drift exceeds tolerance, it proposes to the shift supervisor: adjust the pressure setpoint by 2 bar, schedule hydraulic filter cleaning at the end of the shift, or hold the current batch for sampling. The supervisor signs. It does not recalibrate the press on its own — except where the management team configures autonomy.

See the full IRIS architecture →

Before and after

PM compaction with classic SPC vs. compaction with iLEAN Edge on hydraulic pressure

Aspect	Fixed setpoint + sampling SPC	With iLEAN Edge on pressure + batch + die
Control frequency	One sample every N cycles	Every cycle
Detection of inter-shift drift	By the time SPC confirms it, there is scrap	Before the next sample fails
Cross-check with moisture / batch / die	Each in its own system and its own person	Cross-checked by the agent on the shop floor
Hydraulic filter servicing	Fixed calendar	Proposed when the real drift justifies it
Operation without a network	Setpoint yes, context no	Edge keeps watching cycle by cycle on the cabinet light
Critical part dossier	Sampled curve, reconstructed	Per part — pressure curve, batch, human signature

Impact estimate

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

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

• Powder metallurgy plant making automotive parts (synchros, hydraulic pumps, gears, structural parts) with a hydraulic compaction press and a sintering furnace, with documented dimensional scrap between shifts.
• Edge pilot on the press (pressure transducers + integration with the fill scale, MES and SPC). First expected value within a few weeks: detecting the drift cycle by cycle before SPC confirms it.
• Indicative payback between 4 and 9 months, depending on the frequency of documented incidents and the average cost of scrap + post-sintering grinding.
• Expected reduction of dimensional scrap of ≥ 30% in the first year on critical parts. A reminder of the automotive quality standard: in the order of 25 PPM ^[1] — inter-shift drift is hard to sustain at that level without Edge watching cycle by cycle.

And the fair doubt of the quality manager

«What if the AI changes the setpoint on its own and knocks my press out of calibration?» — the proposal does not execute itself. iLEAN works on an anchored task: it reads the pressure curve the transducer already provides and compares it with the learned signature. On anchored tasks, the best models brought error below 1.5% ^[2]. And even so, the shift supervisor sees the curve, sees the proposal, and signs. The setpoint does not change on its own — except where the management team configures autonomy.

^[1] Automotive quality standard ~25 PPM — source Symestic.

^[2] OpenAI paper «Why Language Models Hallucinate», 2025 — on the reliability of AI in anchored tasks.