Engineering Path for AI Defect Detection in PV Inspection: From Training Set to the Production Line

Almost every PV inspection vendor's pitch deck mentions "AI defect detection." The reality on the shop floor is less romantic: great demos, very few stably-running production AI systems. The bottleneck is not the algorithm — ResNet, EfficientNet, Transformer architectures all handle image classification just fine. The bottleneck is the engineering path from algorithm to line.

This article distills MVCreate's experience building production AI into SC-EPL, SC-PLEL-PS, SC-MC-W and SC-Seed, and lays out the full path.

1. Why "Just Train a Net" Fails in PV

Versus generic vision tasks (cat/dog classification, face recognition), PV defect detection is structurally different:

1.1 Many defect classes, extreme class imbalance

Common PV defects run to 30+ types: hidden crack, broken finger, dark corner, dark spot, dark edge, concentric ring, snake line, cold solder, over-etch, mismatch, PID, hotspot precursor, moon-crescent mark, dislocation line, fiber anomaly, and more. All need to be recognized.

But distribution is brutally imbalanced — tens of thousands of hidden-crack examples per year versus hundreds of snake-line examples. Vanilla classification loss severely overfits to majority classes under this long tail.

1.2 Highly subjective defect boundaries

What counts as a "significant hidden crack"? What is a "mild crack, pass"? Even experienced inspectors agree only ~80% of the time on the same image. This is a natural source of label noise.

1.3 Zero-tolerance on the line

An AI grader that misses a severely cracked module and lets it ship to the downstream customer incurs rework costs on the order of 100× the inspection cost. This zero-tolerance regime puts extreme pressure on recall — especially recall on severe defects.

2. Six Stages of the Engineering Path

Stage 1 — Data collection and labeling

The most expensive, most tedious stage. It also sets the ceiling for everything that follows.

Diversity matters. The training set must span:

• Multiple production lines (different cameras, different lighting)
• Multiple cell technologies (mono, multi, TOPCon, HJT, perovskite)
• Multiple process regimes (normal, edge-of-spec, fault batches)
• Multiple time points (fresh vs aged cells)

MVCreate has accumulated 2.5M+ labeled samples across 30+ defect classes over the past five years.

Double-blind labeling with arbitration. Each image is labeled by two independent inspectors; disagreements are resolved by a third-party arbitrator. Final label consistency holds above 95%.

Active learning to cut cost. Not all samples need human labeling. The current model predicts on new data; only the lowest-confidence samples go to human review. Active learning cuts labeling cost to roughly one-tenth of the naive approach.

Stage 2 — Model selection and training

For fine-grained PV vision, the standard structure is two-stage:

1. Segmentation — U-Net / DeepLab localizes the defect region
2. Classification — ResNet / EfficientNet classifies the region

MVCreate's production model uses a customized two-stream architecture:

• One stream on the raw EL image
• A second stream on the difference image (vs. local background)
• Features fuse at a mid-layer

This design measurably improves recognition of "defects only visible against context" — mild dark spots, subtle edge degradation.

Training strategies:

• Long-tail distribution handled with Focal Loss + class-balanced sampling
• Severe-defect recall monitored as a separate target, >99.5%
• 5-fold cross-validation on every training run to avoid line-specific overfit

Stage 3 — Inference acceleration

Line cycle time is a hard constraint. SC-EPL runs at 0.5–2 s per cell; AI inference must fit inside that envelope.

Acceleration stack:

• Model compression — knowledge distillation from a ResNet-50-class teacher to a ResNet-18-class student. Accuracy loss <1%, speed up ~3×
• Mixed precision — FP32 → FP16 on NVIDIA GPUs, another 2×
• Engine compilation — TensorRT / ONNX Runtime converts PyTorch graphs into optimized GPU kernels
• Batch optimization — batching across multiple cells exploits GPU parallelism

SC-EPL's AI inference latency stabilizes at 200–400 ms, comfortably under the 2-second cycle ceiling.

Stage 4 — Deployment

Two deployment patterns:

Pattern A — local inference. Each station has a GPU; model runs locally. Low latency, data never leaves the facility. Higher hardware cost and harder multi-station updates.

Pattern B — edge server, centralized inference. One GPU server per line, all stations stream images to it. Higher hardware utilization, one-shot updates. Network latency, single point of failure.

MVCreate recommends Pattern B for large lines (>10 stations) and Pattern A for R&D and small lines.

Stage 5 — Drift monitoring

AI models are not trained once and forgotten. Two drift modes hit production:

1. Data drift — process changes (e.g., P-type → N-type) alter EL image statistics, invalidating the old model
2. Concept drift — customer definitions of "defect" evolve (e.g., from "only severe cracks" to "mild cracks too"), making the old decision boundary stale

Monitoring:

• Daily automatic tracking of model confidence distribution
• Weekly human review of ~200 newly-inferenced samples, comparing AI verdict vs. human verdict
• Any drop >3% in agreement triggers a retraining cycle

Stage 6 — Closed-loop iteration

The AI system has to keep seeing and learning from new data. MVCreate's loop:

1. AI inference on the line auto-collects low-confidence samples (typically 2–5% of all samples)
2. Monthly human review of these low-confidence samples produces new training data
3. Quarterly full retraining cycle
4. New model validated on a held-out set is canary-released (5% → 20% → 100%) to the line
5. Monitoring during rollout — any anomaly triggers instant rollback

3. Three Products Built on This Stack

SC-EPL (production AI grader)

• 0.5–2 s per cell
• 30+ defect classes for crystalline silicon
• On-line fine-tuning capability for per-line adaptation
• Deployed across major Chinese TOPCon/HJT lines

SC-PLEL-PS (R&D AI analysis)

• PL + EL dual-mode
• Crystalline silicon + perovskite + tandem support
• AI model outputs quantitative parameters (minority-carrier lifetime, series resistance), not just classification
• Widely used in R&D labs

SC-DEL family (field AI inspection)

• Portable and UAV form factors
• AI model hardened for field conditions — light variability, angle variability, distance variability
• Edge inference, no cloud dependency

4. Three Engineering Lessons

After many years of this work, three lessons stand out:

Lesson 1: Data matters 10× more than model architecture

Effort spent on labeling and data hygiene caps everything else. Months cleaning labels often beats months tuning architecture.

Lesson 2: Recall on severe defects is the only hard target

Over-kill can be mopped up by downstream human review. A miss cannot. Every optimization decision must respect severe-defect recall ≥99.5%.

Lesson 3: Explainability wins customer trust

An AI system on a production line cannot be a black box. MVCreate's AI outputs both a verdict and a saliency heatmap — which region, why. Letting inspectors "see what the model is thinking" earns more trust than another 0.5% accuracy would.

5. Common Misconceptions

Misconception 1 — AI replaces human inspectors. It doesn't. AI handles 90–95% of routine samples; the remaining 5–10% boundary cases still need humans.

Misconception 2 — bigger models are better. On a production line, inference speed is a hard constraint. Real deployed models are typically 5–10× smaller than SOTA.

Misconception 3 — train once, run forever. False. Retrain every 3–6 months or drift silently degrades the model.

Closing

Getting AI defect detection from a slide bullet to a stably-running production system means five engineering stages done right: data, algorithms, deployment, monitoring, iteration. MVCreate's AI product line isn't built on a one-shot "big model" — it's built on a full engineering stack spanning collection, labeling, training, deployment, monitoring, and iteration.

To discuss AI-based PV inspection or arrange a demo, reach the MVCreate technical team (+86 159-5048-9233).

Website: www.mvcreate.com

Originally published by Vision Potential / MVCreate.