Diagnostic-Equipment QC Automation & Performance Grading

Replacing manual Excel QC with a two-stage LSTM and 10 graded metrics that cut QC time ~93% and won an R&D President's Award.

English한국어

Role: Technical Lead / Data Scientist — 11-person cross-functional team (Data Scientist 1, Full-stack 3, mechanical engineers 4, patent 3)  ·  Period: 2020.12 – 2021.09  ·  Stack: Python, R, PyTorch (LSTM), PCA/t-SNE/DBSCAN, Isolation Forest, R Shiny

When PCR reagents are deployed onto third-party instruments, a two-stage QC process gauges device performance. I led its automation and re-graded devices on a differentiated A+/A/B/F scale.

  • Legacy process — manual Excel inspection, only two metrics, ~100-sample batch pass/fail where a single failing signal failed the whole device.
  • Failure modes — fragile to human error, and unable to separate machine defects from operator mistakes.
Diagnostic-Equipment QC Automation & Performance Grading — project poster

Highlights

  • QC time cut ~93% (400h → 28h per 100 devices) and annual operating cost reduced ~13× (≈₩800M), replacing manual tracking with an automated platform.
  • Two-stage LSTM predicts Step-2 results from Step-1 data, so only likely-fail devices proceed to Step 2 — trained on 2,201 devices / 2,552 runs / 61,248 signals: pass/fail classification 94.5%, grade classification 82.7%.
  • 10 new performance metrics — amplification efficiency, SNR, baseline stability, optical/thermal uniformity, negative/positive signal noise, time-series decomposition (trend/seasonal/remainder) residual variance, outlier labeling, RSS — driving an A+/A/B/F grade (full fleet: A+ 7.01%, A 12.91%, B 75.72%, F 4.36%).
  • Anomaly detection via PCA, t-SNE, DBSCAN, Isolation Forest, and the 3-sigma rule, with an R Shiny real-time dashboard for upload → auto-analysis → visualization.
  • R&D President’s Award and two first-inventor patents; grading enables resource optimization (A-grade devices to critical use; B-grade for training) instead of scrapping.

Approach

An ETL pass unifies scattered Excel QC data; a two-stage LSTM predicts Step-2 outcomes from Step-1 signals, and 10 metrics drive a differentiated grade surfaced through a dashboard.

flowchart TB
    M[Manual Excel QC] --> ETL[ETL to unified dataset<br/>2,201 devices, 61,248 signals]
    ETL --> LSTM[Two-stage LSTM<br/>Step 1 predicts Step 2]
    ETL --> MET[10 performance metrics<br/>SNR, uniformity, RSS, ...]
    LSTM --> GR[A+/A/B/F grading]
    MET --> GR
    GR --> DASH[R Shiny dashboard]
    GR --> T[QC time ~93% lower<br/>~13x cost reduction]