DecodeX
Home/Problem Statements/SIH26165
All 226 Statements
SIH26165SoftwareSmart Automation
₹1,50,000 Prize (₹75k Joint)
•
Deadline: 20 September 2026
•
Submitted Ideas: 0/500

AI/NLP Engine to Detect Serious Injury & Fatality (SIF) Precursors in OIL's Unsafe-Act/Unsafe-Condition and Near-Miss Reports

Ministry / Organization: Oil India Limited
SIH Official Portal
Technical Complexity
7/10
Dataset Accessibility
6/10
36h Hackathon Viability
9/10
Step 0 • Zero-Jargon Layman Summary

“Oil India Limited collects thousands of safety reports about unsafe acts and near-misses, but currently reviews them manually after long delays. This project requires an AI tool that reads these text reports, identifies which ones have the potential to cause fatal injuries, categorizes them according to global safety rules, and highlights high-risk areas on a dashboard.”

The DecodeX PS Decoder™

Deconstructed Problem Blueprint

Step 1

WHAT TO BUILD?

An AI-powered Natural Language Processing (NLP) web platform that ingests unstructured safety text reports, classifies them for Serious Injury & Fatality (SIF) potential, tags them to IOGP Life-Saving Rules, and visualizes precursor risk patterns on an interactive dashboard.

Step 2

WHO IS IT FOR?

The primary beneficiary is Oil India Limited (OIL), specifically their Health, Safety, Security, and Environment (HSSE) departments, plant managers, and field safety officers who need to allocate safety interventions effectively.

Step 3

WHY DOES IT EXIST?

Low-severity incidents and fatal incidents have different underlying causes. Manual triage is slow, prone to human error, and fails to separate the 20-25% of reports carrying genuine fatal potential, leading to repeated avoidable accidents.

Step 4

INPUT → PROCESS → OUTPUT ARCHITECTURE PIPELINE

01 • Inputs & Telemetry

Free-text Unsafe-Act / Unsafe-Condition (UA/UC) observation records, near-miss descriptions, and incident logs containing unstructured text, timestamps, locations, and activity types.

02 • Processing & AI Core

Text preprocessing, embedding generation, fine-tuned transformer model (or zero-shot LLM with prompt engineering) for binary SIF-potential classification, multi-label text classification for IOGP Life-Saving Rule mapping, and spatial-temporal aggregation for precursor pattern mining.

03 • Actionable Deliverables

An interactive executive dashboard showing SIF-precursor heatmaps, risk rankings by site/activity, auto-tagged reports matching IOGP rules, and automated safety alert notifications for high-risk clusters.

Step 5

CONSTRAINTS

  • Must handle domain-specific oilfield and industrial jargon, abbreviations, and poorly structured or typo-ridden free-text logs.
  • Must ensure data privacy and compliance with internal corporate data handling policies.
  • Needs to run efficiently with low inference latency to handle bulk report uploads smoothly.
Step 6

JURY METRICS

  • Classification accuracy, precision, and recall specifically for identifying SIF-potential reports (minimizing false negatives on fatal risks).
  • F1-score for accurate mapping to IOGP Life-Saving Rules.
  • Dashboard usability, responsiveness, and clarity of risk visualization during live demo.
Step 7

CORE BOTTLENECK

Handling ambiguous, highly abbreviated, and noisy field-operator notes where critical safety context is implied rather than explicitly stated.

Why You Should Choose This PS (Pros)

  • Extremely clear problem definition directly mapped to an established industry safety framework (DEKRA / IOGP).
  • High social and operational impact by potentially saving human lives in hazardous industrial environments.
  • Strong scope for a visually striking dashboard ('wow factor' for hackathon judges).

Challenges & Risks to Watch (Cons)

  • Lack of a real pre-provided dataset in the problem statement requires teams to synthesize realistic dummy training/testing datasets.
  • Text classification for high-stakes safety requires high precision; false negatives on fatal precursors can be critical.
  • Crowded domain if teams rely on generic wrapper APIs without adding proprietary domain logic.
Hackathon Blueprint

Recommended Tech Stack & Grand Finale Features

Frontend & Mobile
React.jsTailwind CSSRecharts or Chart.js for dashboard visualizationsLeaflet.js for facility map risk overlays
Backend & APIs
FastAPI (Python)Pydantic for data validationCelery for background batch processing of large report files
AI / ML & Models
Sentence-Transformers for text embeddingFine-tuned lightweight LLM or Hugging Face classifier (e.g., DeBERTa / RoBERTa) for SIF and IOGP taggingLangChain or LlamaIndex for structured parsing
Database & Cloud
PostgreSQL with pgvector for semantic search and log storageRedis for caching dashboard stats

Killer Features to Impress the SIH Jury

An 'Auto-Triage & Risk Scoring' module that color-codes incoming raw logs in real-time.
An interactive Root Cause Analyzer that clusters recurring barrier failures (e.g., procedural vs equipment failure).
An NLP-powered Query Assistant ('Ask HSE') where safety officers can type natural language questions like 'Show me all confined space near-misses at Rig 4 last month'.
View Official SIH 2026 Portal Raw Text
▼
•Background OIL collects large volumes of UA/UC observations, near-miss and incident reports through its HSSE platform but these are triaged manually after certain time intervals such as monthly, quarterly etc.However, Global best practice (DEKRA Martin & Black 2015; EEI SIF Precursor model; VelocityEHS 2024 PSIF classifier) has established that low-severity incidents do not share the same causes as fatalities — non-fatal US accidents fell 51% over 15 years while fatalities fell only 25.5%.Leading operators therefore separately flag the ~20–25% of reports carrying genuine fatal potential.

Problem Description Build a prototype that ingests OIL's free-text safety reports and automatically a) Classifies each as SIF-potential vs non-SIF-potential b) Tags it to the relevant IOGP Life-Saving Rule (e.g., Energy Isolation, Hot Work,Confined Space, Line of Fire)

c) Surfaces recurring precursor patterns (activity, location, barrier failure) via a dashboard.

Expected Outcome/Solution A working AI/NLP with an interactive dashboard that ranks sites/activities by SIF-precursor density and auto-maps to Life-Saving Rules, enabling HSE to focus interventions where fatal potential is highest.

Relevant Data Availability (if any)

OIL's UA/UC observations, near-miss and incident reports.