Skip to content

ch09 — Industrial Practice: a Semiconductor FDC/ATE Perspective

Companion notebook: notebooks/ch09_industrial.ipynb — the BYOD workflow

The gap between benchmarks and the fab

Benchmark assumption Semiconductor FDC/ATE reality
a normal train span exists "normal" changes with every recipe and PM cycle
uniform sampling per-step sampling rates differ; event-driven logging
one long continuous series many short fragments per wafer/lot
point labels weak labels at the level of "this lot failed"

Regime vs fault — the most common failure mode

Recipe switches, post-PM behavior and seasonal shifts are operating-regime changes: statistically anomalous, but not faults. Mitigations:

  1. split data by regime metadata (recipe ID, …) and train per-regime models
  2. drift-adaptive thresholds like DSPOT (ch08)
  3. separate changepoint detection from anomaly detection — SKAB labels both

Sparse labels, irregular sampling, cold start

  • Sparse labels: select models with threshold-free metrics (VUS), guarantee the false-alarm rate with conformal thresholds
  • Irregular sampling: breaks the SSM Δ-discretization assumption (ch06) — resample, or add the interval as a feature. Matrix Profile also assumes uniform sampling.
  • Cold start: no train data for a new tool or recipe — foundation-model zero-shot (ch06) or transfer from similar equipment. The SECOM card (docs/datasets/secom.md) illustrates the difference between a time-series problem and static FDC summary statistics.

The BYOD workflow

# unlabeled CSV: scores + threshold decisions only
tsad-forge run --model sub_pca --data my_sensor.csv

# with a label column, the full metric suite is computed
tsad-forge run --model iforest --data my_labeled.csv --seed 0

CSV conventions: numeric columns = channels, optional timestamp column = ordering, optional label column = 0/1. Missing values are linearly interpolated. The span before the first anomaly becomes train (preserving the unsupervised assumption).

Recommended starting combo: sub_pca + conformal — simple, fast, with a false-alarm guarantee. Only then consider deep models (the leaderboard is the evidence for this ordering).