Smart India Hackathon 2026 · PS 26170 · ISRO

Catch defects thedatasheet limit can't see.

AI-driven lot-relative anomaly detection and physics-informed drift forecasting for ESS burn-in screening — explains every flag to QA engineers with per-feature attribution and Arrhenius trajectory overlays.

0
Parts screened
0
Flagged for review
0.0%
Recall (synthetic eval)
±0.0µA
168h drift MAE
< 2s
Detection latency / lot
0 parts
5 lots screened
±0.0µA
Drift prediction accuracy
~40%
Screening time saved

Two modules. One explainable risk score.

Every part gets a fused, lot-relative score — not an absolute-limit pass/fail.

Module A — Fused Outlier Score

Isolation Forest catches multivariate pattern drift; ECOD catches per-parameter tail anomalies — parameter-free, O(n·d log n).

Physics Baseline

Arrhenius / log-linear extrapolation grounds every drift forecast in reliability-engineering theory.

Per-Lot Normalization

MAD-based robust z-scoring per lot per timestamp — the mechanism that makes detection dynamic, not static.

Module B — Drift Predictor

XGBoost residual correction on top of the physics baseline forecasts val_168h from val_0h + val_24h, MAE-optimized.

SHAP Explainability

Every flag decomposes into per-feature contribution — never a black-box score.

Recall-First Threshold

Tuned via PR curve — false negatives (missed defects) cost far more than false positives in mission-critical hardware.

Synthetic Injection Validation

Threshold calibration on synthetically injected drift patterns — detector itself stays unsupervised.

QA Dashboard

Flagged parts, physics-curve overlay, lot percentile, accept/reject override log — all in one interactive console.

End-to-end pipeline

1

Ingest & normalize

Raw ESS logs → per-lot MAD z-score + log-transform

2

Feature decorrelation

val_0h, early/mid/late slopes — not raw correlated timestamps

3

Module A ‖ Module B

IF + ECOD fused score, parallel to physics + XGBoost forecast

4

Risk fusion

Weighted combine → single explainable risk flag

5

SHAP + physics overlay

QA-facing justification, not a black-box number

O(n·d log n) at fab-lot scale

ECOD and Isolation Forest both scale linearly-ish in features and log-linearly in lot size — unlike LOF's O(n²), this holds up on large ESS datasets.

O(x) = Σⱼ −log(min(F̂ₗ(xⱼ), F̂ᵣ(xⱼ)))
// ECOD: sum of per-feature tail-probabilities
O_IF(x) = 2^(−E[h(x)] / c(n))
// Isolation Forest: path-length anomaly score
flag if 0.45·norm(O_IF) + 0.55·norm(O_ECOD) ≥ τ

See it flag a real part

Live data from the screening pipeline. Click a module to see its detection output.

Part #0401-009·L-0401
Anomaly
O_IF
0.685
O_ECOD
25.27
Fused Score
0.858
Lot Percentile
93.8th

SHAP / ECOD Feature Attribution

Driver: tpd_96h contributed 4.6 of 25.3 total ECOD score — this part sits at the 93.8th percentile within its lot's distribution.

Tech Stack

Python
Core pipeline
PyOD
ECOD + Isolation Forest
XGBoost
Residual regression
SHAP
Explainability
scikit-learn
Preprocessing
NumPy / Pandas
Feature engine
Next.js
QA Dashboard
Recharts
Visualizations

Recently Flagged

Top flagged parts across 5 lots — click to drill down

View all
Part IDLotIF ScoreECOD ScoreFusedStatus
#0401-001L-04010.37617.630.654Warning
#0401-004L-04010.36416.310.612Warning
#0401-006L-04010.21114.840.503Warning
#0401-009L-04010.68525.270.858Anomaly
#0401-014L-04010.14314.250.456Warning
#0401-017L-04010.36512.250.501Warning

Ready for mission assurance.

Built for ISRO's ESS pipelines — extensible to defense, aviation, and medical device screening.

Launch QA Console