Battery Tab Appearance Inspection — Field Plan

01/10Challenge

Too many defect names, too few ways to image them

The list holds 30 entries: 14 on aluminium tabs, 14 on copper tabs, 2 on embossing. The names differ widely—“adhesive pile-up” and “double adhesive (mixed sheets)” are two process causes; “inclusion 0.3 mm” and “inclusion 0.5 mm” are two quality grades. Treated the classic AOI way, 30 names mean 30 sample sets, 30 threshold groups and 30 rounds of onsite tuning, repeated whenever the material or the incoming batch changes. Training engineers stay resident, and replication never starts.

Look at it from the sensor side instead. A camera cannot see a process cause; it sees grey levels. Pile-up and double adhesive are both a raised band outside the tape boundary; 0.3 mm and 0.5 mm inclusions are two grades read off the same mask. What actually needs distinguishing is which light makes the defect appear at all. So the first step is not choosing a model—it is pushing 30 names back onto the imaging layer and counting how many genuinely different ways of seeing remain.

02/10Families

Thirty into seven: shared light inside a family, new recipe across

Re-sorted by “what it looks like and what light it needs”, the list collapses into seven families. The immediate payoff: one optical path and one segmentation class set per family, with a recipe change only across families—which is how 30 rounds of tuning become the 4 lighting experiments of Section 04.

Two notes. First, aluminium and copper are different substrates, but a family images the same way on both—copper's higher reflectance shows up as a brightness offset inside the same recipe, not as another family. Second, the 0.3 mm and 0.5 mm inclusions are not two classes but two grades on one mask: segmentation gives the contour, the event layer compares its max bounding size against a threshold. Changing the acceptance grade edits one millimetre value; the model is untouched. On lines where tolerances move often, that separation is what keeps maintenance cost flat.

03/10Sampling

0.017 mm/pixel: draw the sampling line before discussing accuracy

The smallest criterion on the list is a 0.3 mm inclusion. For it to form reliably in the image, sampling resolution has to be settled first. The imaging trial uses 0.017 mm/pixel: a 0.3 mm target spans about 18 pixels and a 0.5 mm target about 29—enough for a pixel-level mask to hug the contour, and enough to separate the two grades. Relax to 0.05 mm/pixel and 0.3 mm is down to six pixels, at which point even a strong model is only guessing.

From 0.017 mm/pixel and 4096 × 3000, the single-frame field follows: 68 × 52 mm. That number then decides two things—a 25 mm fixed lens at roughly 120 mm working distance, and the fact that a 1410 mm edge has to be captured in segments, which is the station-layout problem of Section 05. The light sits closer than the camera (LD2 ≈ 55 mm) to widen the angular difference between zones while leaving mechanical room for loading.

Camera, lens and light reflect the imaging-validation stage. Exposure, working distance and station layout freeze after official samples, mechanical space and real cycle time have been checked.

04/10Zoned Light

One light: lighting a different zone is changing the incidence angle

The seven families do not ask for the same light. A scratch is a microscopic groove in the foil with no height of its own—only near-coaxial light, returning along its own path, makes the groove wall fall off the bright field and read as a dark line. A bubble is a dome under the adhesive—only grazing light leaves the bright-one-side, dark-the-other signature on a frontal view. Adhesive scratches sit between the two and want a mid-high angle from the inner rim of the sectors. Dedicating a lamp to each morphology would put four light assemblies on one station, which neither the budget nor the mechanical space supports.

A 36-zone white light removes that need. The zones already sit at different radii and azimuths, so choosing which ring and which azimuths to light is equivalent to changing the incidence angle—no camera move, no lens change, no added lamp; the switch is one recipe command. The four morphologies therefore collapse into four recipes along a single axis: the closer a defect is to pure texture, the further inward the light moves; the more three-dimensional the relief, the further outward.

05/10Segmented Capture

A 68 mm field has to cover a 1410 mm edge

One frame covers about a twentieth of the long edge. 1410 ÷ 68 ≈ 20.7, so 21 segments without overlap and about 23 with a 10 % stitch margin. Each segment still needs one exposure per lighting recipe to cover all seven families, which puts a part at 84–92 exposures. That is where cycle time is decided—not in inference latency.

Exposures in the hundreds leave three routes: one camera scanning the edge, trading motion for time; several cameras in a row splitting the segments, trading hardware for time; or narrowing the mandatory inspection area by process agreement so the segment count drops. Each route has its own cost, and the choice has to be settled during PoC with official samples and real takt—not on paper. The segment and exposure figures here are magnitudes derived from the parameters above; the field calibration is what counts.

06/10Tensor Flow

Four recipes, four branches, one workflow

The vision agent runs on the edge: configuration, inference, HMI and outputs share one software stack, wired zero-code on a canvas—drag operators from the library, connect them by data flow. This project is not a single serial chain but four branches split by lighting recipe: each lights its zones, captures, segments and detects events; all four merge into one state machine, then save images and talk to the PLC over serial.

Double-click a node for its sub-editor: segmentation → annotate tool, event detection → GPU event composer, state machine → state-graph editor. A product change is mostly recipes and samples; moving the same flow to another size or station means editing zone combinations and thresholds, not rewriting line software.

07/10Segmentation

Pixel masks: lock a new defect with 1–10 samples

Tab defects rarely have regular boundaries: a scratch is a fine broken line, an inclusion is an arbitrary blob, an adhesive bump is a gradient bright patch. Boxes fit these poorly, and the grade decision downstream depends on contour size. The plan therefore uses pixel-level semantic segmentation: each class is defined in the annotate tool from a small set of line images—product guidance is 1–10 sample images and minutes of setup—and the resulting mask drives area, length or max bounding size.

This is what makes “30 defect classes” affordable. A few-shot base means adding a class does not restart a collection-and-long-training cycle: between aluminium and copper, or between batches, similar morphologies often transfer with a handful of extra labels. Training and inference both run locally on the edge GPU, so station frames never leave the plant—good for latency and for data exposure alike.

08/10Events

Turning a mask into “inclusion over 0.5 mm”

Segmentation only sees pixels; the line needs decidable events. Events are composed in the GPU event composer: primitives such as object presence, width measurement and count are given a target class and a region, then combined with AND / OR / NOT. The inclusion grade lands here—object presence confirms the mask belongs to the inclusion family, and width measurement takes the mask's max bounding size against a millimetre threshold.

Keeping thresholds in the event layer rather than inside the model is the structural point. The 0.3 mm and 0.5 mm grades share one mask and one model, so changing the acceptance grade is a single millimetre edit. The same holds for adhesive-overrun tolerance and scratch-length limits: they live in the recipe, and perception stays untouched.

09/10State Machine

Scan segment by segment; release only when every segment and family is clear

Segmented capture creates a problem manual inspection never had: one part yields twenty-odd segments and a hundred-odd frames, and something has to fold them into a single verdict. The state machine encodes the floor process: once the part is in place it enters segmented scanning, self-loops after each segment finishes its four recipes and advances to the next. Only when every segment and every family is clear does it release and count; a hit in any family at any segment turns it NG, and the saved image carries the segment index and the class. An NG part leaves for review and, after re-judgement, re-enters the cycle.

10Deployment

Edge-local inference, with samples and verdicts accumulating together

The system sits on five layers. Cameras and the zoned light capture; a gigabit industrial network keeps traffic inside the plant, so images never leave the site; the edge GPU runs segmentation, event decisions and the state machine; the front end provides live view, NG alarms, human review and retrieval by time, part ID or defect; recipes centrally manage camera parameters, zone combinations, model versions and thresholds. Raw frames, masks, result images and verdicts are archived per part—material for traceability, and the sample source for the next round of model tuning. The longer the line runs, the more usable samples exist, which is exactly the input continuous iteration needs.

Equipment list

The table fixes categories and configuration direction only; specific models freeze after official samples, mechanical space and cycle time are checked.

Rollout

Staged delivery, each stage with an acceptable output:

The boundary is explicit: this plan targets metal–adhesive composite surfaces with many similar-looking defect types and frequently moving tolerance grades. The cycle-time ceiling of segmented scanning is set by the mechanical design; where takt is tighter, it has to be balanced with parallel cameras or a narrower mandatory area. Defects inside the adhesive layer, and curled or badly warped parts that fall outside depth of field, need process support or an added orthogonal view. All parameters and segment magnitudes above come from the imaging validation and derivation; recognition rates and decision thresholds are confirmed on official samples in a fixed fixture.