✈️ Aviation Maintenance NLP: Automated Fault Classification Pipeline

Classifies industrial fault reports from free-text maintenance logs using Azure ML + NLP

---

📋 Table of Contents

• Problem Statement
• Solution Architecture
• Dataset & Preprocessing
• Model Performance
• Real Examples: Input → Output
• How to Run
• Business Impact

---

1. Problem Statement

In the aviation and manufacturing sectors, mechanics and technicians log maintenance issues using unstructured, free-text descriptions. This creates a massive bottleneck: human engineers must manually read thousands of logs to assign standardized JASC (Joint Aircraft System/Component) codes before work orders can be routed.

This pipeline automates the classification of unstructured discrepancy logs from the FAA Service Difficulty Reports (SDRS), instantly mapping messy text to standard fault codes with attached confidence scores to maintain a "Human-in-the-Loop" safety standard.

---

2. Solution Architecture

This project implements a complete, end-to-end Connected Industry 4.0 data pipeline:

flowchart TD
    A["🖥️ EDGE DEVICE SIMULATION\nDocker + Python\nproducer.py — streams JSON batches every 10s\nfrom FAA SDRS dataset"]
    B["🗄️ DATA LAKE\nAzure Blob Storage\nraw-logs container\nCold storage landing zone"]
    C["⚙️ CLOUD ETL\nAzure Data Factory\nDate normalization · Null dropping\nSchema standardization"]
    D["🏢 DATA WAREHOUSE\nAzure SQL Database\nCleanAviationMaintenanceData view\n119 JASC fault classes"]
    E["🤖 AI ENGINE\nAzure Machine Learning\nTF-IDF Bigrams → Logistic Regression\nConfidence Score 0–100%"]
    F{"Confidence\n≥ 60%?"}
    G["✅ AUTO-ACCEPTED\nLog cleared automatically\nWritten to AIPredictions table"]
    H["🔴 HUMAN REVIEW\nFlagged in dashboard\nEngineer manually verifies"]
    I["📊 EXECUTIVE DASHBOARD\nPower BI\nFleet health trends · JASC frequency\nRed-flag low-confidence predictions"]

    A -->|"JSON payload stream"| B
    B -->|"Blob ingestion trigger"| C
    C -->|"Structured data load"| D
    D -->|"SQL data pull"| E
    E --> F
    F -->|"Yes"| G
    F -->|"No"| H
    G -->|"Predictions push"| D
    H -->|"Predictions push"| D
    D -->|"Live SQL connection"| I

    style A fill:#0d1f2d,stroke:#00d4ff,color:#c8d8e8
    style B fill:#0d1f2d,stroke:#7fff6b,color:#c8d8e8
    style C fill:#0d1f2d,stroke:#ffd166,color:#c8d8e8
    style D fill:#0d1f2d,stroke:#c77dff,color:#c8d8e8
    style E fill:#0d1f2d,stroke:#ff6b35,color:#c8d8e8
    style F fill:#0d1520,stroke:#ff6b35,color:#ffd166
    style G fill:#0d2010,stroke:#7fff6b,color:#7fff6b
    style H fill:#2d0d0d,stroke:#ff6b35,color:#ff6b35
    style I fill:#0d1f2d,stroke:#00d4ff,color:#c8d8e8

Loading

Layer Summary

Layer	Component	Description
🖥️ Edge	Docker + Python	producer.py simulates live edge telemetry by sampling raw FAA SDRS data and streaming JSON batches every 10 seconds
🗄️ Data Lake	Azure Blob Storage	Landing zone catching streaming JSON payloads in a raw-logs container
⚙️ ETL	Azure Data Factory	Ingests raw blobs, standardizes date formats, drops null records, and loads structured data into the database
🏢 Data Warehouse	Azure SQL Database	Hosts raw telemetry and serves a clean curated view (CleanAviationMaintenanceData) optimized for ML
🤖 AI Engine	Azure Machine Learning	Pulls SQL data, vectorizes text via TF-IDF, classifies logs with Logistic Regression, and pushes predictions + confidence scores to AIPredictions table
📊 Dashboard	Power BI	Connects to Azure SQL to visualize fleet health trends and flags low-confidence predictions for human review

---

3. Dataset & Preprocessing

The model is trained on the real-world FAA Service Difficulty Reports (SDRS), which contains historical aircraft malfunction records.

Dataset Split:

• 🟢 Training: 580 rows
• 🔵 Testing: 146 rows

Features:

Role	Field	Description
Input	Discrepancy	Unstructured mechanic notes including abbreviations, part numbers, and misspellings
Target	JASCCode	Standardized 4-digit system component code spanning 119 unique classes

NLP Preprocessing — TF-IDF Vectorizer Configuration:

TfidfVectorizer(
    stop_words='english',   # Strip standard English stop words
    ngram_range=(1, 2),     # Capture bigrams
    max_features=2500,      # Expand vocabulary
    min_df=2                # Ignore ultra-rare typos
)

---

4. Model Performance

Logistic Regression (max_iter=2000, C=10) outputs both a predicted fault class and a Confidence Score (0–100%).

⚠️ Logs scoring below 60% are automatically flagged for manual review.

Metric	Score
Baseline AI Accuracy	45.89%
Upgraded AI Accuracy	67.12%
Average Fleet AI Health	69.88%

---

5. Real Examples: Input → Output

✅ Example 1: The Confident Automation

Raw Input:

AFTER LANDING FA REPORTED STRONG BURNING SMELL IN THE AFT GALLEY AND SAID THAT 
THE PAX IN THE LAST TWO ROWS WERE COUGHING... R/R BOTH RECIRC FILTERS PER A220 AMP 21-22.

Field	Value
Predicted JASC Code	2120
Confidence Score	67.91%
Action	✅ Auto-accepted — Confidence > 60%, log cleared without manual intervention

---

⚠️ Example 2: The "Human-in-the-Loop" Edge Case

Raw Input:

FOUND MAJOR CORROSION AND HUGE CRACKS ON THE LEFT LANDING GEAR DOOR IN THE FWD CARGO AREA.

Field	Value
Predicted JASC Code	5320
Confidence Score	12.18%
Action	🔴 Flagged for review — Dashboard highlights row in red, alerting an engineer to manually verify

---

🛠️ Tech Stack

• Cloud: Azure Blob Storage, Azure Data Factory, Azure SQL Database, Azure Machine Learning
• ML: Logistic Regression, TF-IDF Vectorization (scikit-learn)
• Edge Simulation: Python, Docker
• Visualization: Power BI
• Dataset: FAA Service Difficulty Reports (SDRS)

---

6. How to Run

1. Start the Edge Data Stream (Docker)

Ensure Docker is installed, then build and run the data producer:

docker build -t aviation-edge-node .
docker run -e AZURE_CONNECTION_STRING="DefaultEndpointsProtocol=https;AccountName=YOUR_ACCOUNT_NAME_AND_KEYS_HERE" aviation-edge-node

2. Azure Setup

• Trigger your Azure Data Factory pipeline to move the generated JSON blobs from the raw-logs container into your Azure SQL database.
• Execute the SQL script provided to generate the CleanAviationMaintenanceData view.

3. Train & Score the Model

Open the Azure ML workspace, configure your database credentials, and run nlp_model.ipynb to generate predictions and push them back to the database.

4. View the Dashboard

Open Aviation Fault Classifier.pbix in Power BI Desktop and click Refresh to pull the live AI predictions from your Azure SQL instance.

---

7. Business Impact

In the context of Connected Industry 4.0, shifting from reactive repairs to predictive maintenance requires massive amounts of structured historical data. Unstructured text logs are a dark data asset — they contain the ground truth of machine failure but cannot be read by digital twins or forecasting algorithms.

By deploying a containerized NLP pipeline to standardize this text into machine-readable fault codes at scale, organizations can:

• Unlock years of legacy data, making it available for predictive model training
• Reduce Aircraft On Ground (AOG) downtime through faster, automated fault triage
• Optimize the spare parts supply chain with structured, queryable failure history

This directly accelerates the transition from legacy maintenance logs to a fully data-driven, predictive fleet management system.

---

Dataset is intentionally constrained to 726 records as a proof-of-concept pipeline demonstration. The architecture is designed to scale: replacing the FAA SDRS sample with a full enterprise maintenance database would significantly improve model performance.

---

👨‍💻 Author

Alireza Sorousheh

Maintained for aviation safety research and industrial NLP classification.