Skip to content

progress.md

Latest commit

 

History

History
30 lines (26 loc) · 2.83 KB

progress.md

File metadata and controls

30 lines (26 loc) · 2.83 KB

Progress: Normal-First Defect Learning

2026-04-21

  • Created planning scaffold for a reusable industrial defect detection方案.
  • Confirmed /Users/jilanfang/yolo is not currently a git repository.
  • Confirmed the workspace had no existing markdown/docs structure before this task.
  • Created document package under docs/normal-first-defect-learning/.
  • Wrote the general solution, technical architecture, customer pilot playbook, and research notes.
  • Verified the final file structure and line counts.
  • Added a 2026 literature scan covering recent PatchCore/pseudo-label/YOLO/active-learning/real-world industrial anomaly detection papers.
  • Added a staged research roadmap that starts with a no-large-model baseline and progressively introduces ROI classifiers, vector retrieval, 1B-3B assistants, 3B-7B reviewers, and 14B-30B offline analysis.
  • Added a detailed key-paper review and breakthrough analysis covering the closest PatchCore-pseudo-label-YOLO work, anomaly-to-classification, MVTec AD 2, ISP-AD, and active learning papers.
  • Added an industrial implementation blueprint covering on-site survey, mechanical fixture, camera/lens/lighting, edge compute, software services, data loop, PLC/MES integration, deployment, operations, and acceptance metrics.
  • Added a human-accelerated inspection strategy for the realistic case where OK samples are available, NG samples are only 20-30, and the machine's first role is to speed up human inspection rather than fully automate final judgment.
  • Saved a local hardware and customer deployment plan covering MacBook Air, XPS 15 9560, E5-2696v4 desktop, GPU choices, and customer-side machine tiers. Marked as saved only, not implemented.
  • Implemented normal-first-defect/ as a runnable Python scaffold with configs, scripts, core modules, tests, notebooks, and docs.
  • Followed a red-green cycle for the core package: initial tests failed because nfdi did not exist, then implementation made the suite pass.
  • Verified a CLI smoke path using temporary data: manifest preparation, heatmap-to-ROI generation, candidate scoring, review queue generation, metrics evaluation, and static HTML review rendering.

Verification

  • find . -maxdepth 3 -type f shows all expected deliverables.
  • wc -l reports 1027 total lines across the 8 markdown files.
  • Pending after literature scan: rerun file structure and line-count verification.
  • Pending after research roadmap: rerun file structure and line-count verification.
  • Pending after key-paper review: rerun file structure and line-count verification.
  • Pending after industrial implementation blueprint: rerun file structure and line-count verification.
  • Pending after human-accelerated strategy: rerun file structure and line-count verification.
  • Scaffold verification completed with python3 -B -m unittest discover -s normal-first-defect/tests and CLI smoke scripts.