Aviation Safety Analysis — Proposal, KPIs, Stakeholders, Datasets

Overview

This project analyzes ASRS aviation incident reports to (1) describe historical trends, (2) compare recent years (2024–2025) to prior years, and (3) explore factors associated with change (e.g., FAR part, airspace, time of day). We combine exploratory analysis, data cleaning, and modeling (OLS and SARIMA) to inform potential safety interventions.

---

Datasets

• Aviation Safety Reporting System (ASRS) online database
  https://asrs.arc.nasa.gov/search/database.html

---

Stakeholders

• Regulators: FAA, NTSB
• Operators: Airlines, Part 135 operators, General Aviation community
• Air Traffic Services: ATC facilities and management
• Industry & Unions: OEMs, pilot associations, safety teams
• Public & Insurers: Travelers, risk analysts

---

KPIs

• Trend indicators: Year-over-year and month-over-month incident counts by FAR Part (91/121/135) and airspace class (A–E).
• Category-specific trends: “Human Factors,” “Aircraft,” “Company Policy,” “Weather,” etc.
• Recent-year comparison: Are 2024–2025 incident patterns different from prior years?
• Seasonality: Magnitude and timing of monthly seasonal effects.
• Forecast stability: SARIMA projections for 121 and 91 (short- and 5-year horizons).

---

Data Cleaning (high level)

• Header fix: First row contained true headers; promoted to column names.
• De-duplication: Where multiple reports referenced the same incident, retained a single record per incident/field to avoid double counting.
• Date hygiene: Removed obviously invalid dates (e.g., “0 BC”).
• Normalization:
  • FAR Parts like 121; 121 → 121.
  • Light/Day labels standardized (e.g., Day, day, Daylight → Day).
• Airspace anonymization: Standardized airport labels to anonymized zzz to ensure consistency across years after ASRS policy changes.
• FAR ambiguity: Excluded small cases with conflicting FAR parts to maintain clean grouping.

---

Exploratory Data Analysis (can be seen in repo)

• Accidents over time: Overall decrease over time, though not entirely monotonic.
  

• Accidents per state: Top three are California, Texas, Florida.
  

• Top 3 states over time: Declines since ~2015 are most apparent.
  

• Accidents per country: U.S. highest (dataset focuses on U.S.-origin flights), followed by the Philippines.
  

• FAR 121 × Airspace class (A–E):
  
  

• Time of day: Distribution across daylight vs night operations.
  

• By FAR Part:
  
  
  

• Top 10 aircraft models: Commercial fixed-wing dominates counts (usage effects likely).
  

---

Modeling & Results

OLS (trends)

• All incidents (2006–2024): Yearly incidents decrease over time (p = 0.023).
  R² = 0.337, adj. R² = 0.286.
  

• FAR 121 (commercial): Clearer decreasing trend (p = 0.000).
  R² = 0.696.
  

• Within FAR 121 by Primary Problem:

  • Human Factors: decreasing (p = 0.024; ~1,000/yr; coef ≈ −22).
    
  • Aircraft: decreasing (p = 0.014; ~1,000/yr; coef ≈ −24.78).
    
  • Company Policy: decreasing (p = 0.000; coef ≈ −17.23); approaches near zero by 2023; R² ≈ 0.710.
    
  • Procedure, Weather, Environment (non-weather), ATC Equip/Nav, Charts, Airspace Structure, Software/Automation: no significant trend (given current data coverage).

• Airspace subset:

  • Clear decreases in A, B, E, and unspecified.
  • Class C: no significant trend.
  • Class D: marginal increase (p = 0.054; ~50/yr). Restricting to Human Factors clarifies a significant increase (p = 0.001).
    
    

• FAR 91 (general aviation): Increasing over time (p = 0.016).
  

• FAR 135 (charter/air taxi): No significant trend (less data overall).

• Seasonality (FAR 121, monthly):

  • One-hot months (Jan baseline): February lower than January (short month effect); others mixed.
  • Fourier (cyclical) encoding reveals seasonal structure (γ₂ ≠ 0, p = 0.009).
  • Fit quality modest: adj. R² ~ 0.315–0.317.
  

SARIMA (forecasts)

• FAR 121: Early years (2005–2009) higher/volatile; forecasts suggest lower, more stable rates going forward.
  
  
  
  

• FAR 91: Similar story—historically higher/volatile; forecasts remain relatively low/stable.
  
  

---

Key Findings

• Overall incidents show a downward trend; strongest for FAR 121.
• Human Factors, Aircraft, Company Policy categories under FAR 121 are decreasing; company-policy incidents approach zero by 2023.
• Class D airspace shows a marginal increase, significant when focusing on Human Factors—but counts are small; interpret with caution.
• FAR 91 incidents increase, potentially reflecting reporting or exposure changes.
• Seasonality exists (notably February lows); month effects explain limited variance.
• Forecasts (SARIMA) indicate stable/lower incident levels in near term for 121 and 91.

---

Limitations

• ASRS is voluntary; reporting practices vary by part/category and over time.
• Airspace labels and airport anonymization shift across years.
• Part 135 sample smaller → less statistical power.

---

Next Steps

• Extend category-specific analyses beyond 121 (where feasible).

---