Car hydraulic lines with AI — a line that passes the leak test can be a leak six months from now.

The problem

Three critical steps and a final test that detects too late.

Anyone manufacturing hydraulic lines for brakes, power steering, transmission or cooling lives with three pains:

1. Bending — ovalization at the bend, internal wrinkles, a microscopic crack in the outer fiber. The part passes visually and the defect stays hidden.
2. The connector weld — a bead with lack of penetration holds the pressure of the test bench and fails under real vibration in the car.
3. The flare — geometry out of tolerance that the operator forces into place and that ends up as a slow leak. The customer finds it months later.

The leak test detects the gross leak, but not the root cause nor the defects that will show up with mileage. When the customer returns a batch, you contain a thousand units because classic traceability is by the hour, not by the part. The classic system runs on fear: contain just in case, contain everything, contain late.

And the OEM customer is demanding PPM levels on the order of 25 parts per million ^[1]. At that figure, there is no room for "it passed the test, we're done".

How it fits the IRIS system

iLEAN does not replace your leak test bench — it puts an eye on every step and links the result to the part.

The problem with hydraulic lines is not a lack of a final test, it is blind spots at every intermediate step that nobody watches part by part. iLEAN acts as the putty that fills those blind spots — Edge on bending, welding and flaring; Connect reads the bender's PLC and the test bench; the agent cross-references with the batch and assembles the file.

Edge sees every step. Connect reads the actual program, not the setpoint. The agent cross-references batch, test and operator; when there is a return, it contains by the part, not by the batch. The person signs — the file comes out on its own, the critical decision does not.

The three iLEAN pieces applied to hydraulic line control:

• Edge — machine vision terminals (CNN) on bending (ovalization, external wrinkles), welding (bead, apparent penetration) and flaring (geometry, symmetry). It triggers the actuator before the next step — the suspect part never reaches the final test, and the test is not the first time anyone looks at it. It works with no network.
• Connect — captures the bending program actually executed (real angle, radius, PLC parameters), the welding parameters, the leak test bench result, the incoming tube batch. If the machine is old and isolated, it reads its screen by vision; it does not force you to replace it.
• Agent — cross-references images, parameters, test and batch. It spots that bender 3 is drifting on ovalization shift after shift, before rejects pile up. It assembles the PPAP / IATF 16949 file per part. When the customer raises an 8D, containment is by the part, not by the whole batch.

See the full IRIS architecture →

Before and after

Classic control vs. cross-referenced control with iLEAN

Aspect	Leak test + self-inspection	With iLEAN Edge + Connect + Agent
Bending defect	Passes the leak test, fails months later	Edge catches it as it leaves the bender
Lack of penetration in the weld	Holds on the bench, not under vibration	Edge sees the abnormal bead, pulls the part
Traceability facing an 8D	Containment by the hour — a thousand parts	By the part — only the affected ones
Machine drift	Discovered with a rejected batch	Pattern tracked by the agent, early warning
PPAP / IATF file	Manual reconstruction weeks later	Ready on the fly, per part
Operation with no network	n/a	Edge keeps inspecting on the light from the panel

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.

• Automotive hydraulic line plant with several benders, several welding stations and a shared leak test bench.
• Edge pilot on one critical bender + the flaring station + integration with the MES and the test bench. First value expected within a few weeks.
• Expected reduction of customer PPM of ≥ 30% within the pilot scope.
• Indicative payback between 4 and 9 months, dominated by finer containment when an 8D hits + fewer field returns + the PPAP file being ready.

And the CAIO's reasonable doubt

“Is AI reliable enough for a safety part?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI simply classifies an image against a known pattern or reads a value from one system and writes it into another, the best models brought error below 1.5% ^[2]. And even so, what is critical is never decided alone: iLEAN's three safety rings guarantee that a held part is validated by a person, and that the file going out to the customer carries a human signature.

^[1] Automotive quality standard on the order of 25 PPM. Symestic.

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