A Data-Driven Early Warning System for Mining Accidents

Background

Often, safety condition of a mine deteriorates before an accident happens. For example, before the 2010 Upper Big Branch Mine disaster, one of the largest in the U.S. history, the mine displayed an alarming rising trend of safety violations. Similar building-up behaviors apply to other major mining accidents in the U.S. According to the Mine Safety and Health Administration (MSHA), on average, there are more than a hundred thousand citations and orders issued yearly. The MSHA accident and inspection databases are an untapped resource for safety analysis. Our model analyzes semi-structured data from MSHA, attempting to uncover a company's safety culture from its public regulatory records. This work has implications for developing a data-driven early warning system. In the future, we plan to extend this research to inspection data from the Occupational Safety and Health Administration (OSHA) and the Environmental Protection Agency (EPA).

Goal Statement

Create a statistical model as follows:

• Training data: historical data of mines, mine accidents, and MSHA violations
• Input: mine type (or mine identifier), past data of this mine, i.e., accidents and violations
• Output: chance of this mine to have a severe (resulting fatality or disability) accident in the future

Data Source

• Current data set used for this project: https://arlweb.msha.gov/OpenGovernmentData/OGIMSHA.asp
• Other data set: https://enforcedata.dol.gov/views/data_catalogs.php
• IMPORTANT: arlweb.msha.gov no longer includes proper headers in their data dumps (i.e., the first row contains actual values instead of column names). To download the working RData files, follow these instructions:
  • Old data sets are uploaded to GitHub via Git Large File Storage (LFS). Install Git LFS to work with large files.
  • Run git lfs install to install the Git command line extension (you only need to install once).
  • Clone this repo and the full RData files will be downloaded to your local repo.
    • If you've already cloned this repo before installing LFS, remove it and clone again so that the RData files can be properly downloaded.

Repository Hierarchy

Description	Working directory	Sub-directory	File
Main script	./Houston/	src/	main.R
Data consolidation			consolidate_data.R
Statistical model			conditional_logistic.R
Function 1			prepare_violation.R
Function 2			roll_over.R
Consolidated data		output/	Consolidated.RData
Result			Result_clogit.RData
Mine data		data/	Mines.RData
Accident data			Accidents.RData
Violation data			AssessedViolations.RData

• Required R packages: caret, dplyr, e1071, RcppRoll, plm, survival

Manuscript

• Work in progress. Contact Yu.

Updates

2017-10-01

• The MSHA data website we used before (link) no longer offers proper headers for their txt data dumps. We will not update our processed data (.RData) any more.
• Accidents, violations, and mines were updated on 2017-04-08.
• They are uploaded to GitHub via Git Large File Storage (LFS).

2017-06-13

• Updated README.md

2017-06-12

• Minor edits on Catherine's codes
  • Added /Houston/src/main.R
  • Replaced load() and .RData with source() and .R to import functions
  • Fixed a problem where extracted .txt files are not deleted in /Houston/src/import_msha_txt.R
  • Added out_sample_model in /Houston/src/conditional_logistic.R