This project aims to develop ultra-lightweight, deployment-ready CNN architectures
for defect detection (Curling) in industrial PBF-LB/P (Selective Laser Sintering) processes.

The central research question of Phase‑1:

Can we design a computationally efficient CNN that remains reliable under extreme class imbalance (~0.9%) while preserving strict anti-leakage guarantees?

Four progressively designed lightweight CNN models:

Model Purpose Design Philosophy

PicoCNN Minimal baseline Ultra-small reference model

NanoCNN Compact CNN Standard lightweight convolution

NanoLightCNN Efficiency-optimized Reduced capacity, faster inference

All models follow identical training and evaluation protocols to ensure fair comparison.

Phase‑1 enforces a cryptographically verifiable protocol:

• Canonical configuration (fixed_config.json)
• SHA256 checksum validation at runtime
• Immutable configuration mapping
• Deterministic execution (fixed seeds, disabled cuDNN benchmarking)
• Strict 4-fold group-aware cross-validation
• 20% holdout test isolation
• No temporal or segment leakage
• SHA1 duplicate-content verification across splits

This ensures:

✔ No configuration drift
✔ No silent hyperparameter modification
✔ No cross-fold contamination
✔ Full audit readiness

Industrial defect ratio ≈ 0.9% positives

Mitigation strategy:

• Cost-sensitive cross-entropy (frequency-based weighting)
• Minority-only augmentation (train split only)
• Validation-based threshold calibration (τ* maximizing F1)
• AUPRC reporting for imbalance-aware evaluation

No oversampling.
No distribution distortion.
No heuristic batch rebalancing.

Per fold: - F1@τ* - AUPRC - Accuracy - Optimal threshold (τ*)

Holdout: - Final unbiased evaluation - No threshold tuning on holdout

Capacity Reporting: - Trainable parameters - MACs - FLOPs (≈ 2 × MACs) - Single-thread CPU latency

cnn_models/
    PicoCNN/
    NanoCNN/
    NanoLightCNN/
    MicroLiteCNN/

configs/
splits/
stats/
reports/audit/
scripts/preprocess/

Each model directory contains:

• artifacts/
• calib/
• capacity/
• configs/
• logs/
• meta/
• weights/
• model_*.py

All folds (1--4) are complete and verified.

✔ Architectures finalized
✔ Anti-leakage validated
✔ Cross-validation locked
✔ Holdout metrics recorded
✔ Capacity benchmarks measured
✔ Configuration cryptographically frozen

Phase‑1 is officially closed and ready for temporal modeling extension (Phase‑2: LSTM-based sequence modeling).

Ali Vaezi
University of Applied Science Vienna