Home Appliances: Ring Outer-Edge Chip Inspection — Feasibility Study

01/10Problem

A local notch that can appear anywhere on 360°

The morphology is simple: handling or clamping knocks the outer edge, a small piece breaks away, and a notch is left behind. The difficulty is not the shape but three conditions around it. First, it appears at an arbitrary angle—there is no fixed inspection window, and all 108 mm of rim has to be covered. Second, the part is an alloy casting with oil film and directional glare on the face; under ordinary lighting both a specular patch and normal turning marks can be mistaken for an anomaly. Third, both faces must be inspected, and one exposure only ever sees one of them.

The harder gap is the missing standard. How small a chip still counts as NG has not been defined by the customer. Classic AOI practice is to assume a grade, collect samples, train, tune thresholds—and redo the whole loop once the real standard lands. That second loop usually costs more than the first. This plan therefore treats the minimum detectable grade as a parameter still to be filled in rather than as a given.

02/10Measurement

Establishing what “how big” means

We have a batch of result images produced by the project's own model, with each notch already boxed (the untouched original output is shown at the lower right of Fig. 01). A box says “something is here”, not “how big it is”. Since the 34.46 mm outside diameter is known, the scale can be recovered: fit the outer circle to the bright-field silhouette, take the radius in pixels, divide into the true diameter, and the mm/pixel of the batch falls out. The innermost 12 % of contour points are dropped from the fit so the notch itself cannot pull the datum circle inward.

Once scaled, each notch converts into two physical quantities: the maximum radial intrusion relative to the fitted circle, and the angular span where the curve dips below half that depth, expressed as arc length. Across seven chips these land in 0.046 – 0.244 mm and 0.23 – 0.75 mm respectively.

Two comparisons are worth noting. Samples C and D are almost identical in depth but differ by more than three times in arc length: depth and arc length are not proportional, so reporting only an area or a box size merges “long and shallow” with “narrow and deep”—two cases that usually carry different process verdicts. Sample G is the shallowest on file at about twice the baseline jitter of its own frame, close to the measurement floor of this batch.

One caveat: these result images are 1024 pixels on the short side, about 0.041 mm/pixel, roughly half the sampling of the original capture. The table therefore reports an order of magnitude measured on downsampled images; at full resolution the real margin is wider.

03/10Sampling

0.020 mm/pixel, and the whole rim in a single frame

The existing setup is a five-megapixel monochrome camera, a 12 – 36 mm zoom industrial lens and a dome light, at 80 – 120 mm working distance with the part held still. Taking 2448 × 2048, covering Φ34.46 mm with clamping margin puts the frame at roughly 42 mm across, which over the 2048-pixel short side is 0.020 mm/pixel; the focal length implied by that field falls between 12 and 18 mm, comfortably inside the existing zoom range.

Two numbers follow. First, the circumference π × 34.46 ≈ 108.3 mm spreads over about 5300 pixels at 0.020 mm/pixel—the entire inspected edge fits in one frame, with no segmented capture and no rotary stage. That is a structural difference from large-format inspection: no segments means no seams between them, no repeat-positioning error and no dozens of exposures in the cycle. Second, a notch 0.45 mm along the arc and 0.23 mm deep occupies roughly 22 × 11 pixels at full resolution—enough for segmentation to track the contour, but not generous, which is why this sampling line is a hard constraint rather than headroom.

The dome light deserves its own note. Wide-angle diffuse illumination drives the alloy face to uniform white, flattening oil film, turning marks and directional glare together, while the background stays dark—so the outer edge becomes a high-contrast black-and-white boundary. This changes the nature of the problem: a chip is no longer a texture anomaly on a surface but a missing piece of a silhouette. Geometric criteria are far more stable than textural ones, and they map directly onto drawing tolerance.

The zoom lens is the one risk in this configuration. A 12 – 36 mm zoom is convenient during setup, but if focal length or aperture is knocked, the scale changes and every millimetre criterion silently goes with it. It must be locked before production, or replaced by a fixed focal length in the same range. The exact lens model is to be confirmed from the nameplate.

04/10Datum & Fixture

Three-centre alignment governs lighting, not centring

“Lens axis = dome centre = part centre” is the basic requirement of this optical path. Taken apart, however, the three deviations behave very differently, and the fixture only has to control one of them.

Lateral offset is absorbed in software. When the part sits off-axis, a circle parallel to the sensor still projects as a circle; only its position moves. The datum circle is fitted from the frame itself, so its centre travels with the part and the radius curve is unaffected. In several of the frames behind Fig. 01 and Fig. 03 the ring is in fact not centred, and the measurement still holds. What offset does consume is field margin—shift too far and the rim is clipped by the frame.

Height variation is absorbed the same way. A change in face height changes magnification: at 100 mm working distance, 1 mm of height is about 1 % of scale, which on the radius is of the order of 0.17 mm—larger than any chip in the table. But as long as the criterion is defined as radial intrusion relative to this frame's fitted circle, a uniform scale change cancels out. Define it instead as an absolute diameter and that 0.17 mm lands straight on the result. The form of the criterion, not the hardware, is what makes the difference.

Only face tilt has to be held mechanically. With the face tilted by α relative to the optical axis, the outer circle projects as an ellipse whose major and minor axes differ by roughly D(1 − cos α). With D = 34.46 mm:

The fixture's key specification is therefore face tilt rather than concentricity, and holding it within 3° is a safe target. If that proves mechanically hard there is a fallback: ellipticity is a slowly varying second harmonic around the rim while a chip is a local spike, so the two look nothing alike on the unwrapped curve and an ellipse fit—or simply removing the low-order harmonics—can suppress it. That is margin, though, and should not be the first line of defence.

So what does three-centre alignment actually protect? The symmetry of the illumination. A dome's brightness distribution is symmetric about its own axis, so once the part is off-centre the incidence conditions differ around the rim, the width of the bright-to-dark transition drifts with azimuth, and the effective binarisation threshold drifts with it. That error cannot be corrected downstream. The existing material supports this: the frame with the most uneven rim brightness, where turning marks show through at the lower edge (sample F), also carries the highest baseline jitter in the set at 0.027 mm—about twice that of the cleanest frame (sample E, 0.013 mm).

05/10Unwrapped Rim

Unroll the circle into a line

That a chip can sit anywhere on the rim looks like a nuisance, but it has a clean answer: fit the outer circle to establish centre and datum radius, sample the contour radius angle by angle, and unroll the whole rim into a single r(θ) curve. Once unwrapped, positional randomness disappears—the entire circumference shares one criterion, and a chip shows up as a dip in the curve.

The curve settles three things at once. First, the verdict becomes a physical quantity: so many millimetres of radial intrusion over so much arc, both of which map straight onto drawing tolerance, with no need to explain what a model confidence of 0.87 means. Second, signal-to-noise becomes computable—the P95 residual of the intact rim is the noise floor of this optical path, about 0.015 mm here against a 0.23 mm chip, a ratio of roughly fifteen; even the shallowest chip at 0.05 mm keeps about twice the margin. “Can it detect 0.1 mm” stops being a matter of impression and becomes one division. Third, the grade is adjustable: fixing the minimum detectable size simply moves the orange line up or down, with the model untouched.

None of this replaces the segmentation model. The unwrap supplies a stable measurement frame; the model decides which stretch of the rim actually belongs to a chip against a busy background—separating burrs, clamping marks and fixture occlusion from genuine knocks. Criterion and perception divide the work; they are not substitutes.

06/10Both Faces

One camera with a flip, or two cameras in parallel

One exposure sees one face, so two exposures per part are unavoidable. With cycle time and budget both open, two configurations stay on the table: option A uses one camera, lens and dome to shoot the faces in turn with a flip in between; option B uses two optical paths and images both faces at once, eliminating the flip.

The point worth making is that the two configurations are the same system in software. One camera or two, what reaches the state machine is a front-face verdict and a back-face verdict, merged under one part ID, with a chip on either face condemning the whole part. Option A adds a “flipped” state; option B adds a parallel capture branch. Changing configuration changes the orchestration, not the inspection criteria.

That opens a practical path: start on option A, settle the minimum detectable grade, the sample library and the thresholds, then scale to option B once cycle time is fixed. Samples, labels and thresholds all carry over; only the second optical path needs new calibration.

07/10Tensor Flow

Two branches into one workflow

The vision agent runs at the edge: configuration, inference, display and output all live in one piece of software, orchestrated by dragging nodes out of an operator library and wiring them along the data flow. This project's graph is short—one branch per face: capture, pre-processing (outer-circle location and polar unwrap), semantic segmentation, event detection; the two event streams merge into one state machine, followed by image archiving and a serial device.

Double-clicking a node opens its sub-editor: semantic segmentation to the annotation tool, event detection to the GPU event orchestrator, state machine to the state diagram editor. The structure of this graph is independent of part geometry; moving to another ring means new calibration values, samples and thresholds, while the graph itself is reused as-is on comparable stations.

08/10Segmentation

Pixel masks, with 1–10 samples per new morphology

Chip boundaries are irregular—a sharp V in one part, a long shallow scallop in another—so box regression fits them poorly, while the downstream depth and arc conversion depends entirely on contour accuracy. The plan therefore uses pixel-level semantic segmentation: a single class, “outer-edge chip”, defined in the annotation tool from a handful of field images, with the product side starting from 1–10 sample images and minutes of setup and the mask converted to millimetres afterwards.

Having only one class actually clarifies the sampling strategy: what needs collecting is not more defect types but a sample spread that covers the detection boundary—obvious chips, borderline chips, sound parts, and the confusers that invite false calls (clamping marks, burrs, oil residue, fixture occlusion). Borderline samples matter most, since they decide where the minimum detectable grade can sit without either missing parts or crying wolf. Training and inference both run locally on the edge GPU, so station imagery stays on site.

09/10Events & FSM

From mask to “deeper than 0.10 mm”, then to one verdict per part

Segmentation only sees pixels; the line needs decidable events. Events are assembled in the GPU event orchestrator: an “object present” primitive confirms the mask lies in the rim band, a “width” primitive takes the radial intrusion against the fitted circle and compares it with a millimetre threshold, and the two meet at an AND. Arc length is a second quantity read off the same mask and becomes its own event; chip count becomes a third.

Keeping thresholds at the event layer rather than inside the model is the pivot of this structure, and it matters more than usual here because the minimum detectable grade has not been set yet. The system can go live on an initial grade and start producing data; when the customer's standard arrives, what changes is a millimetre value in the recipe, with samples, labels and model untouched. The same holds for the allowed arc length and the permitted chip count per part.

The state machine then assembles these events into the station's flow. A part arrives and enters front-face inspection; only a passing front face triggers the flip and back-face inspection; only two passing faces release the part and increment the count. A hit on either face turns NG, and the archive records angular position, radial depth and arc length together. One cycle-time detail is worth flagging: a part already condemned on the front face is not flipped—it is settled immediately, and flipping only ever happens on parts that passed.

10Deployment

Edge-local inference: collect the data, then set the standard

The system sits on five layers: camera and dome light capture; gigabit industrial Ethernet moves frames inside the plant network only; an edge GPU runs segmentation, size conversion, event decisions and state transitions; the front end offers live view, NG alarms, manual review and retrieval by time, part ID or angle; and a recipe layer holds camera parameters, calibration values, model versions and thresholds together. Raw frames, masks, result images and verdicts are archived per part—both as traceability material and as the sample source for the next round of tuning.

For a project whose standard is still open, that archive carries extra value: once the system runs, every part's depth and arc length is recorded, and within a few weeks there is a real size distribution. The minimum detectable grade can then be set on data rather than on impression.

System hardware

The table fixes equipment categories and configuration direction only; exact models are frozen after physical verification and once cycle time is known.

Rollout sequence

One step departs from the usual order: standard alignment comes before sample collection. Labelling against an undefined criterion means the labelling convention is reworked along with the criterion.

The boundaries are worth stating. This plan addresses contour-type defects—knocks that remove material from the outer silhouette. If acceptance also covers texture-type defects such as face scratches, pitting or discolouration, blowing the face to white with a dome no longer works and a second recipe with low-angle or ring lighting is required, which is a different technical route. Part height is still unconfirmed and affects both depth of field and fixture design; the lens model and outer diameter must be verified physically, with the dome aperture following from that. All measurements quoted here come from downsampled result images and are used to establish orders of magnitude; the minimum detectable grade and the OK/NG boundary follow full-resolution imagery and the customer's acceptance standard.