Every record cites a clause
Bearings map to ISO 15243 damage classes and gears to ISO 10825, 13 modes carrying the standard's own numbering, so a finding traces back to the document.
ISO 15243 · ISO 10825 · multi-tenant · exportable
A photograph of a returned gearbox part becomes a standards-linked failure record: the damage mode, the severity, the mechanism that caused it, and the batch it came from. The model suggests. The inspector decides. Every confirmation becomes training data.
ISO failure modes
Cause mechanisms
Severity grades
Rolling-bearing damage classes
Core objectives
Witness is not a general defect detector. It is scoped to rotating equipment, to two published standards, and to the decision an inspector actually has to make.
A returned part usually arrives with a photograph and nothing else. Witness turns that photograph into a record that cites a clause of a published standard, so the finding survives a warranty dispute.
The damage mode says what the surface shows. The attribution says why it got there. The same pit can mean end of life or a contaminant dent that seeded it early, and the batch-level action differs.
Every confirmed or corrected record is a (model, human) pair. The workspace accumulates a labelled set that belongs to the plant, and exports on demand as JSON.
Workflow
Each stage does one job and hands off. Nothing decides for itself past the confidence gate. That gate is the single place the system chooses whether a person is needed.
Score above the threshold files itself. Below it queues for a person.
Upload known-good reference photographs for each part family. This set defines what normal looks like for this workspace, not for the industry.
The photograph is read for EXIF, scored for provenance, and hashed. A perceptual hash catches the same part submitted twice.
Stage one asks one question: is this part abnormal against the reference set? A PatchCore-style memory bank answers it, and returns a heat map.
Stage two asks a second question: which ISO mode is it? The choice is forced. The model must pick from the standard's own list, never invent a label.
A confidence gate splits the traffic. High confidence files itself. Low confidence queues for an inspector, whose call is the authoritative one.
The loop does not stop at the finding. A confirmed record links back to the batch that produced the part, so a repeated mode across one supplier becomes a fleet signal rather than a pile of separate complaints.
What it knows
Bearings and gears fail in different ways, and each has its own catalogue. Witness carries the standard's own numbering, so a finding traces back to the document rather than to a label somebody invented.
Six primary damage classes, from rolling-contact fatigue to fracture and cracking. 6 modes carry clause numbers such as 5.1 and 5.6.
7 wear and damage classes covering the flank, the root and the mesh: wear, scuffing, contact fatigue, cracks and tooth fracture.
A mode describes the surface. An attribution names the mechanism that put it there. Witness records both, because the fleet action depends on the second one.
Reference geometry
Severity · ISA-101 discipline
Severity never rides on colour alone. Every grade pairs a hue with a filling-disc glyph and a numeral, so the signal survives colour-vision deficiency, a greyscale print, and a washed-out bench display.
Return to service
Re-inspect at next interval
Schedule rework
Pull the part
Quarantine batch, escalate
Features
The first finding is useful. The hundredth is where the system earns its place: patterns across suppliers, batches and duty cycles become visible.
Bearings map to ISO 15243 damage classes and gears to ISO 10825, 13 modes carrying the standard's own numbering, so a finding traces back to the document.
Postgres row-level security scopes every row to the signed-in tenant. Reference sets, findings and batches belong to the workspace that made them.
Confidence is shown, not hidden. A record that the model was unsure about is marked unconfirmed until a person confirms or corrects it.
A part carries a batch code and a supplier. Link enough findings and a repeated mode stops being an anecdote and becomes a warranty case.
Insights aggregates on three axes at once: ISO mode against severity against attribution, so a lubrication problem separates from a design problem.
Confirmed records export as JSON, with the model's suggestion and the human call side by side. That pairing is what trains the next model.
Open the console and walk the loop end to end: enrol a reference set, score a part, accept or correct the ISO mode, then watch the finding appear in the fleet cube against its batch.