Maritime accidents: root causes and saving lives

This project was undertaken by Mr Felipe Dutra and Mr Mark Montanana.

• We investigated the current reporting methods and documentation used by CHIRP, particularly in light of the improvements made to their reporting system last year.
• By means of statistical sampling techniques we were able to select a representative sample for text analysis. The text analysis revealed themes in line with previous research that identifies a ‘deadly dozen factors.’
• The analysis outlines a series of recommendations for improving data quality going forward.
• Topic analysis based on Latent Dirichlet Allocation (probabilistic modelling) revealed different themes according to the severity of the accident.
• A proof-of-concept predictive model was developed to impute the field of incident severity. To build the model, we used feature engineering techniques that transformed verbatim data into vectors, so the data is on a format amenable for machine learning models.

Confidential Human Factors Incident Reporting Programme (CHIRP) work to help improve aviation and maritime safety and build a Just Culture by managing an independent and influential programme for the confidential reporting of human factors–related safety issues.

CHIRP has a database of incident reports going back to 2002, but the reporting system was improved last year. So far just over 100 reports are available in the new format. The results from these systems are used by the Maritime & Coastguard Agency to communicate the major human contributory factors that result in maritime accidents, incidents and errors.  

These are communicated in terms of the ‘deadly dozen’ factors. These dozen were chosen from a full list of 65 human behaviour factors which affect overall safety performance.  

Both CHIRP’s Maritime Director and the MCA consider that these dozen factors were not selected scientifically, and that reporting could be improved.  

They were keen to leverage this resource to draw insights from the data to improve their reporting system going forward.

• Text analysis of key themes around report severity.
• Recommendation to streamline and improve quality of reporting going forward.
• Predictive model to impute missing field of report severity.

• Suggestions to improve the reporting system.
• Identifying primary causes of accidents and pattern recognition.
• Key insights about drivers of accident severity.
• Missing data imputation methodology that can be developed further.