Almost nothing on a battery line photographs well. Electrode coating is low-contrast and uneven, and its dark spots and exposed foil are micron-scale. Cylindrical cells have mirror-finish metal cans, where reflections are read as defects while real scratches and dents disappear into the glare. Tab-welding burrs, misalignment and cold joints hide between layers, which 2D radiography struggles to separate.
What makes it harder is the asymmetry of cost: a missed cold joint is a thermal-runaway safety risk, so lines tighten the criterion — and pay for it in false calls and stoppages. The goal here is never "catch a bit more". It is to drive escapes toward zero without letting false calls rise.
2. How the solution is put together
Imaging first: electrode coating uses 3D ACI, fusing point cloud with deep learning to read height differences in a low-contrast layer; cylindrical cells use multi-light computational imaging to suppress specular reflection before any judgement is made.
Model handles sample scarcity: APDT positive-sample learning trains on good parts only — 1–20 good samples to go live, so a new line does not have to collect a defect library first.
Quality gate on the existing line: critical steps such as tab welding run a 100% automatic gate that identifies weld defects layer by layer, with no added cycle-time penalty.
Line integration: results, defect images and coordinates return to MES / SPC and stay traceable by cell ID. Production data stays inside the plant.
3. Choosing the configuration
Process step
Recommended
What it addresses
Electrode coating
3D ACI (point cloud + deep learning)
Micron-scale dark spots, scratches, exposed foil, material drop-out
Tab welding
ACI quality gate, retrofitted
Burrs, misalignment and cold joints, identified layer by layer
Cell appearance
2D ACI with multi-light imaging
Wrinkles, leakage, bulging and scratches on mirror-finish cans
DaoAI's deployment runs in four steps. ① Assessment — set the decision boundary separately for safety-related defects (cold joints, burrs) and cosmetic ones; their tolerances differ. ② Configuration — match systems and lighting to each process step, with a quote and a delivery plan. ③ On-site deployment — one good part programs the system automatically in 30 seconds to 5 minutes; operators are up to speed in about 30 minutes. ④ Feedback — re-inspection results flow back and the model updates in minutes.
5. What to measure afterwards
Track two groups separately: escape and detection rates on safety-related steps, accuracy and false-call rate on cosmetic ones. Across DaoAI's deployed battery lines the typical picture is 50 µm smallest detectable feature on electrode coating at 97%+ accuracy with 100% inline inspection and no speed loss; >99% detection on tab welding with cold-joint escapes close to zero and no cycle-time penalty; and 98%+ judgement accuracy with 99%+ defect detection on cylindrical cell appearance.
Figures are the typical range our solution reaches in anonymised industry scenarios. Actual values vary with process step, material and cadence; on-site measurement governs.
Which key battery production steps does AI vision inspect?
It covers inline electrode coating, separator pinholes, tab-welding quality, X-ray / CT electrode alignment, and reflective cylindrical-can inspection, protecting consistency and safety gates.
Coating defects are tiny and lines run fast, can AI keep up?
Yes. DaoAI models are optimized for high-speed coating, detecting exposed foil, dark spots and thickness variation in real time to match high-throughput continuous production.
How are safety-critical items like tab welding and electrode alignment gated?
DaoAI inspects tab-weld morphology with 2D vision and checks internal electrode alignment via X-ray / CT, serving as a safety gate that lowers thermal-runaway risk.
Cylindrical cans are highly reflective and prone to false judgments, what helps?
DaoAI suppresses metallic glare with feature-cognition models and optimized lighting, reliably finding scratches, dents and stains on curved cylindrical cells while reducing false calls.
