Skip to content
Zester
Sparks flying from a grinder in a metal workshop
All applications

Forging A closed-die forging unit

Reading the forge the way the master does

Thermal vision and acoustics at the press, trained on the forge master's own calls, so an off-temperature billet is flagged before it is struck.

Stage
Reference design
Sector
Forging
Capabilities
Computer VisionEdge AIIndustrial IoTVoice Capture
Deployment
Edge: Jetson Orin, thermal + RGB camera, contact mic. Works offline.
Integrations
Furnace PLC (Modbus), WhatsApp alerts, shift report PDF
FURNACETHERMALDIEACOUSTIC
Signal path. Illustration, not a site photograph.

The knowledge at risk

It lives in the head of the forge master.

He reads steel by colour, by the ring of the first blow, and by how the flash spreads. He rejects a billet in under a second. He cannot write down why, and nobody has asked him to.

Shop-floor constraints

  • Radiant heat, scale and vibration at the press.
  • Cycle time of a few seconds per part. No time for a screen.
  • Operators do not read English. Alerts must be a light, a sound, or Gujarati voice.
  • Unreliable internet on the shop floor.

What we build

  1. 01

    Instrument the press

    A thermal camera watches the billet leave the furnace. A contact microphone listens to each blow.

  2. 02

    Record the master's calls

    Every accept, reject and re-heat he makes is logged with one button and an optional Gujarati voice note.

  3. 03

    Learn his threshold

    An edge model links the thermal profile and blow acoustics to his decisions, per die and per grade.

  4. 04

    Flag before the strike

    Off-temperature billets trigger a stack light at the press. The master overrides; the override becomes a label.

Today: what it improves

  • Fewer under-heated strikes, so less die wear and fewer laps and cracks.
  • Newer operators get the master's threshold as a light, not a lecture.
  • A shift report of temperature spread per die, without a clipboard.

Over time: how it learns

A record of what the master saw, heard and decided, blow by blow, so the system keeps learning his calls.

  • D1Thermal video of each billet from furnace door to die
  • D2Hammer-blow acoustics, time-aligned to the frame
  • D3The master's accept / reject / re-heat label per billet
  • D4Gujarati voice notes transcribed against the part
  • D5Die, grade and furnace setpoint context from the PLC

Recorded on site and owned by the plant. The system uses it to keep learning your floor.

Architecture

  1. 01

    Capture

    Thermal + RGB camera, contact microphone, one-button label box

  2. 02

    Edge

    Jetson Orin: frame sync, temperature estimate, acoustic features

  3. 03

    Model

    Per-die classifier trained on the master's labelled decisions

  4. 04

    Record

    Local store, synced when online: video, audio, labels, PLC context

  5. 05

    Surface

    Stack light at the press, WhatsApp alert, daily shift report

What we will measure

  • Agreement between the model and the master's calls.

    PILOT MEASUREMENT REQUIRED: agreement rate over four weeks of labelled billets.

  • Fewer defects traced to billet temperature.

    PILOT MEASUREMENT REQUIRED: lap and crack rejection rate before and after.

  • The model keeps learning the master's calls.

    PILOT MEASUREMENT REQUIRED: labelled billets per month and model agreement over time.

Run a plant with an expert like this?

Tell us whose judgement your floor cannot run without. That is where a pilot starts.

Solve a Real Problem[email protected]