Machine Learning Based Fire-Ignition Video Analysis Tool

Accurate modeling of threat flash/function and fire ignition in aircraft dry bays is dependent on an understanding of the phenomena involved in the impact of ballistic threats with flammable fluid containers such as a fuel tank. The live fire tests for these events are riddled with obstacles hindering the collection of complete and consistent data, impeding a full understanding of the underlying phenomena. Engineering models and simulations are built and/or validated with measured test data, so any measurement errors or sampling gaps may be inadvertently incorporated into the models. To solve this problem, a set of analysis tools is needed to consistently provide automatic data extraction from sensor data during ballistic test events. A successful development would provide an accurate understanding of the events leading up to a dry bay fire.Manually annotating High Speed Video (HSV) of ballistic events is very time consuming and error prone. This is primarily due to challenges associated with video calibration, the unique characteristics of each flash or function cloud, and the current approach to camera position, alignment and configuration. In addition, the flash and function definition often vary by analyst or by test, especially related to the end of a test event. This limits the usability of the data for model or simulation creation. Unfortunately, despite efforts to avoid this, a significant amount of ballistic test data has been gathered inconsistently over time. This has occurred for several reasons, including data collected at multiple locations over many years, subjective human measurement, and unperceivable data. Documentation of tests and the resulting data has frequently been focused on the current project, leaving insufficient details to allow the results to be incorporated into future models. Skyward's development of a fire ignition video analysis tool will provide essential consistent data collection and annotation across current and future efforts.Significant progress towards the goal of automated video annotation of fragment tests was made by Skyward, Ltd. (Skyward) under the Joint Aircraft Survivability Program (JASP) Next Generation Fire Model (NGFM) effort. That effort showcased how computer vision (CV) and machine learning (ML) techniques can be used to automatically correct video, determine flash duration, and localize the flash event for the front face flash of a fragment impacting a metal panel. The proposed effort will build upon existing tools developed by Skyward in the NGFM effort These tools are already in use for analyzing test events in ongoing NGFM efforts. Risk reduction research in the Phase I effort will focus on expanding capabilities to further automate data analysis and evaluate the existing tools on additional tests.