Explaining Aviation Safety Incidents Using Deep Learned Precursors

Although aviation accidents are rare, safety incidents occur more frequently and require careful analysis for providing actionable recommendations to improve safety. Automatically analyzing safety incidents using flight data is challenging because of the absence of labels on timestep-wise events in a flight, complexity of multi-dimensional data, and lack of scalable tools to perform analysis over large number of events. In this work, we propose a precursor mining algorithm that identifies correlated patterns in multidimensional time series to explain an adverse event. Precursors are valuable to systems health and safety monitoring in explaining and forecasting anomalies. Current precursor mining methods suffer from poor scalability to high dimensional time series data and in capturing long-term memory. We propose an approach by combining multiple-instance learning (MIL) and deep recurrent neural networks (DRNN) to take advantage of MIL's ability to model weakly-supervised data and DRNN's ability to model long term memory processes, to scale well to high dimensional data and to large volumes of data using GPU parallelism. We apply the proposed method to find precursors and offer explanations to high speed exceedance safety incidents using commercial flight data.