Beyond Pass or Fail: Rethinking Visual Inspection in Battery Manufacturing

When a scratch appears on the housing of a consumer gadget, the customer notices and the manufacturer issues a refund. When a metallic particle is sealed inside a battery cell, nobody notices anything. It can pass formation and end-of-line testing and ship. Months later it works through a separator, and occasionally something goes badly wrong. That asymmetry is what makes battery quality so different, and why the pressure on inspection keeps rising as electric vehicles and energy storage push volumes higher. 

I have spent more than twenty years in technology, the last decade putting AI and computer vision to work on factory floors. One conclusion keeps coming back to me: the real opportunity in battery quality lies in changing what we ask inspection to do, well beyond catching a few more defects. 

Inspect before the cell closes 

Once a cell is sealed, a camera can no longer see what matters most. X-ray and CT can check electrode alignment and anode overhang in finished cells, but they are slower and costlier, and they find problems after most of the value has been added. The cheapest place to stop an internal defect is upstream, while electrodes and separators are still open to view. 

That means inspecting at coating, where pinholes, agglomerates and streaks first appear; at calendering and slitting, where burrs and edge cracks are born; at stacking or winding, where a folded separator or misaligned electrode becomes permanent; and at tab welding, where spatter and incomplete fusion decide long-term reliability. A defect caught at the coater costs a few meters of foil. The same defect found after formation and aging costs a finished cell. 

Where traditional inspection runs out of road 

Most plants start with people looking at parts, then add rule-based machine vision. Rules work when the world holds still: the same part, the same lighting, a defect with a clean geometric signature. Battery lines rarely offer that. 

Coatings change in texture and reflectivity from lot to lot. Separators are thin and translucent. Pouch films are glossy and deformable; prismatic cans are highly reflective. Defects do not follow a template, and they shift whenever a material, supplier or process setting changes. Learning-based inspection helps because it models normal variation and flags whatever falls outside it, including defects nobody wrote a rule for. 

That does not make rules obsolete. I do not see deterministic and learning-based vision as opposing choices. A coating-width measurement or tab-position check is often best handled by a gauge or classic vision tool, and some problems are better solved by an upstream process change than by any camera. Technology selection should follow the quality problem. 

The pilot is rarely the hard part 

Across automotive, aerospace, electronics and industrial deployments, the technology has seldom held a project back. Trouble starts when a pilot has to become a production system. A pilot gets a curated dataset and a team watching closely. Production gets drifting light, a camera someone knocked on Tuesday, a new foil supplier and a coater that does not stop to wait. 

False calls deserve more attention than they get. Every good electrode rejected is unnecessary scrap, and every false alarm teaches operators to doubt the system. A system that cries wolf gets bypassed, and a bypassed system catches nothing. Escapes and false rejects must be managed together, with thresholds agreed by quality and production. 

A few other lessons hold. Image quality and labeling discipline matter more than model architecture. The most worrying defects have the fewest examples, so systems must learn from scarce data and keep learning after launch. Decisions belong at the edge, close to the line. And none of it counts unless results reach the MES, the PLCs and the quality workflow. 

Treating inspection as a source of insight 

For decades, inspection has been a gate: a part passes or it does not. That framing throws away most of what the camera knows. Every image says something about the process that produced it. A gradual rise in coating edge irregularities may point to a drifting slot-die setting. Particle detections clustered around one station may lead to a contamination source upstream. A small shift in weld appearance can show up before yield falls. Tied to process data, a defect stops being an isolated event and becomes evidence. The question changes from “did this cell pass?” to “what is the process telling us, and what should we adjust before the next thousand cells?” 

Evidence is only useful if it is traceable. I think of it as a simple chain recorded for every cell: which part, what we saw, what we did, and what changed. With those four links in place, root-cause work starts from images captured at the moment of production, and the loop between defect and cause can shrink from weeks to hours. 

What this asks of leaders 

Much of the work is organizational. Quality, process, automation and IT teams must agree on what counts as a defect, how much risk is acceptable and who acts on the data. That conversation is often harder than the engineering. 

It also asks leaders to rethink the role of people. The scarcest resource in a battery plant is rarely labor; it is judgment. The engineer who can tell a harmless coating mark from a real defect at a glance took years to develop, and new gigafactories are opening faster than that judgment can be grown. The goal is not to remove experts from the loop but to capture their calls and apply them consistently across every shift and plant. 

A few habits have served me well. Start from a business outcome such as yield, scrap or escape rate, and work backward to the model. Earn the trust of operators and process engineers. Plan for the second year as carefully as the first month, with monitoring, retraining and drift management built in. And judge success by fewer escapes, less waste and faster learning, not by accuracy on a test set. 

Looking ahead 

OEMs, regulators and end customers are asking for deeper traceability and firmer proof that safety-critical processes are under control. Manufacturers that treat visual inspection as core infrastructure rather than a cost to squeeze will be better placed to scale. 

Cameras and algorithms will not write the next chapter on their own. It will belong to organizations that turn what their lines see into what their teams know, and act on it quickly. The value of AI in manufacturing is that it gives good people better information at the speed production demands. 

About the author: Rajesh Iyengar is the Founder and CEO of Lincode Labs Inc., an AI visual inspection company serving manufacturers in automotive, aerospace, electronics and industrial sectors. He is the author of the Amazon bestseller Teach the Factory to See and a member of the Forbes Technology Council. With more than 20 years in technology, including three successful exits and CTO roles at Si3D Systems and PurpleData, he speaks at industry events across the U.S., U.K. and Germany.