Discover how EA Series with NVIDIA Jetson Orin NX and 67 TOPS enables flexible on-device AI visual inspection for lithium battery structural parts, with 0.02 mm defect detection.
Case Details
EA Series Brings Flexible On-Device AI Visual Inspection to Lithium Battery Structural Parts
Background
A domestic manufacturer of precision lithium battery structural parts—battery cases, covers, and explosion-proof sheets—supplies leading power battery makers. Its lines run high-mix, low-volume production, switching more than 40 product models per month. The existing “manual inspection + simple machine vision” system could no longer meet rising quality standards or fast delivery requirements.
Challenges
- Slow changeover: Each new model required 3–5 days of on-site parameter tuning, incompatible with “order today, produce tomorrow” demands.
- Limited defect coverage: Traditional algorithms missed subtle and irregular defects; over 80% of customer complaints involved non-standard defects.
- Small-batch cost inversion: Dedicated inspection stations were uneconomical for orders of a few thousand pieces; manual inspection cost more than 3× large-batch rates.
- Space constraints: Split “industrial PC + vision controller” setups were too bulky for existing stations; layout changes were costly and slow.
Solution: EA230 Edge AI Computer
Beilai Technology deployed the EA230 industrial edge AI computer as a compact, flexible inspection platform.
- Compute: NVIDIA Jetson Orin NX 8GB, 67 TOPS INT8, about one-third the size of a traditional industrial PC.
- Vision: 2 GigE Vision industrial cameras with ring lights for top and side inspection; hardware-triggered synchronized acquisition; PoE power to reduce wiring.
- Line integration: Native Modbus TCP/IP and PROFINET connect directly to the PLC and barcode scanner. Product model information syncs automatically, and the system calls the corresponding AI model.
- Industrial design: Fully sealed fanless enclosure, -20°C to 60°C operation, wide-voltage input, and isolated IO for lithium battery workshop conditions.
- Deployment: DIN-rail mounting inside the existing electrical cabinet—no extra floor space.
Software and Model Workflow
- Customer quality staff label only 50–100 defect samples using a cloud tool; the system handles training and quantization automatically—no algorithm engineer required.
- EA230 stores up to 50 product models locally. After scanning, the correct model switches within 0.5 seconds, enabling “changeover means production.”
- Defects are graded automatically: critical defects trigger rejection; suspected defects go to manual review, reducing review workload.
Rollout
A pilot line went live in 3 days and reached stable mass production within a week. Using a standardized template, the remaining 8 lines were upgraded within half a month; average deployment time per station dropped to 2 hours.
Results After Three Months
- Changeover: Single-model inspection deployment cut from 3–5 days to under 4 hours; new models supported per month rose from 10 to 45.
- Mixed-model production: Up to 12 models on one line with zero-pause model switching; changeover downtime reduced by over 90%.
- Small orders: Per-piece inspection cost down 65%, improving small-order profitability.
- Quality: 23 defect types identified; minimum defect 0.02 mm; detection rate 99.6%; false alarm rate under 3%; customer appearance complaints down 71%.
- Efficiency: 120 ms per piece; over 70,000 pieces/day per station; inspection staffing reduced 70%.
- Traceability: All images and defect data stored locally and linked to barcodes for full lifecycle traceability.
- Cost: Hardware investment per station down 35%; annual O&M labor cost down 60%; payback about 7 months.
Conclusion
EA series delivers inclusive on-device AI compute, native vision integration, and frontline model iteration. For high-mix, low-volume manufacturing, it turns AI inspection from a fixed, costly system into a flexible production tool—deployable quickly, adaptable to changeovers, and affordable for small batches.
