AI-Driven Anomaly Detection in Component Burn-In & Screening
Indian Space Research Organisation(ISRO) · Smart Automation · Software
Exceptionally clear and light to build, but the entire demo rests on synthetic burn-in data — invest in making the drift and latent defects realistic and non-trivial, because a generator that makes the anomalies obvious proves nothing to a reliability judge.
What it actually is
Space electronics are burned in at high temperature to weed out weak parts, but today's pass/fail limits miss 'latent defects' — parts that stay inside the absolute limit yet drift suspiciously over time. The ask is an ML system that catches these two ways: flagging parts that are outliers relative to their lot, and predicting from early readings which parts will drift too far by the end of the test.
What to build
Two modules on time-series parametric burn-in data measured at intervals like 0, 24, 96 and 168 hours: a dynamic outlier detector that flags a part as anomalous relative to its lot's distribution even when it passes the absolute datasheet limit — the part reading 45 microamps in a lot averaging 10 is anomalous though the limit is 50 — and a drift predictor, a regression model taking the 0-hour and 24-hour values and forecasting the 168-hour value, flagging for early rejection any part whose predicted drift rate exceeds a safety slope, with explanations a QA inspector can act on, evaluated with heavy penalty for false negatives.
Smallest thing that wins the room
Run a lot through both modules and show a part passing the absolute limit but flagged as a lot-relative outlier, and another flagged early because its 0-hour and 24-hour readings predict a 168-hour value beyond the safety slope, each with a plain explanation of why.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 62% of the 226 · #87 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: defence, intelligence and space bodies drew small fields.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
3/5Exceptionally clear and genuinely useful for high-reliability screening, with the asymmetric false-negative metric matching the real cost structure, but the whole demonstration rests on synthetic burn-in data you construct, so the models can only catch the latent defects you built into that data — realism of the generator is the crux an ISRO reliability judge will probe.
Feasibility
3/5The methods — distribution-based outlier detection and early-to-late regression — are standard and light, but no real burn-in dataset is provided, so you must synthesise realistic parametric drift data with embedded latent defects, and the realism of that synthetic data determines whether the models demonstrate anything meaningful.
Innovation scope
3/5The two modules and their logic are specified in detail, including the worked outlier example and the drift-prediction inputs, so the approach is largely given and your room is in the modelling quality and the explainability.
Clarity
5/5The description defines both modules precisely, gives a concrete numeric example, specifies the regression inputs and outputs, and states the evaluation metrics including the asymmetric false-negative penalty, making it exceptionally clear.
Effort
MediumTwo focused models on tabular time-series data plus explainability and a synthetic data generator is a contained, well-bounded build.
Demo-ability
MediumThe lot-relative outlier and the early-drift prediction are clear, meaningful results, but they run on synthetic data you generated, so the demo shows the method rather than a real defect caught.
In its favour
- Green flag: The description gives a concrete numeric example and exact module definitions, so there is zero ambiguity about what to build
- Green flag: The asymmetric metric penalising missed defects matches the real cost structure and pushes you toward the right trade-off
- Green flag: The methods are light and well-understood, so you can reach a polished, explainable result rather than a half-working one
- Green flag: The high-reliability screening framing is genuinely valuable and the niche thins the field
Against it
- Red flag: No real burn-in dataset is provided, so you synthesise it, and the models can only catch the latent defects you deliberately embedded
- Red flag: If your synthetic drift is too clean or too obvious, the outlier and drift detection become trivial and prove nothing
- Red flag: Predicting a 168-hour value from just the 0-hour and 24-hour readings is genuinely hard if the synthetic drift is realistically noisy, so honest accuracy may be modest
- Red flag: An ISRO reliability judge knows real burn-in behaviour and will question whether your synthetic data reflects it
What you will be writing
- Distribution-based outlier detection (robust z-score / Mahalanobis)
- Early-to-late drift regression
- Synthetic burn-in time-series generation
- False-negative-weighted evaluation
- SHAP explainability for QA
- Safety-slope thresholding
- Reliability engineering
- Anomaly detection
- Electronics screening
Prior art to read before you start
lot-relative outlier detection · parametric drift prediction · latent defect screening
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.