Prediction of Safety Incidents, Year 2

This project seeks to define, develop and test an algorithm that will use hazard identification data as input to predict when and where there is a high probability of a safety incident occurring. In comparison to safety incidents, collection of inspection data or other similar data is cheap. However, the comprehensive value of this data at NASA is still untapped as a reliable leading indicator for safety incidents. Even when data is combined, it is viewed as a current status with no predictive information as shown in the graph below which trends numbers of different types of safety reports over time. This project was separated into two phases: Data Integration and Predictive Capability. During year one of this project most of the effort was directed to Data Integration. We have been able to implement a new tool using an existing platform that integrates several data sources with a search and export capability. Seven databases owned by SSC, MSFC and the NASA Safety Center were identified for integration and six have been successfully incorporated. The tool will be ready for testing by mid-August with the majority of features incorporated. At the end of FY19, this tool will be in use by SMA Directorate personnel. During the end of year one, the effort toward the predictive capability phase of the project has progressed. Three data sets that house mishaps, close calls and safety and quality inspections have been combined into one numerical dataset. Machine learning experiments have been run on the data to answer some basic questions about machine learning and our data set. Firstly, the question is: Can Machine Learning be used to gain an accurate enough solution for this problem? If the answer is yes then, we need to know the limitations of machine learning in relation to this problem and what additional data may be useful. Ultimately we expect to be able to write requirements for further algorithm development. These requirements will be used to solicit solutions from one or more of the following sources: Louisiana Research Consortium, Mississippi Research Consortium, NASA Crowdsourcing. Safety at NASA is ripe for this innovation as we continue to drive safety incidents to low levels using traditional root cause techniques after an incident occurs. It is also increasingly difficult to predict future incidents through experience and instincts of safety personnel, as personnel retirement and attrition erodes the agency expertise. We seek to take advantage of advances in the areas of machine learning, data analytics and big data to save time, money and lives.