Real-Time Anomaly Detection for Flight Testing Using AutoEncoder and LSTM

Abstract

Flight testing is crucial in validating the functionality and safety in new commercial aircraft design before mass production. The challenge is to support real-time analysis of high-dimensional time series data generated from tens of thousands of sensors around the aircraft during test flights. We propose a novel 2-stage approach, using a fine-tuned autoencoder to extract the generic underlying features of high-dimensional data, followed by a stacked LSTM using the learned features to predict aircraft time series and to detect anomalies in real-time for flight testing. A novel Timestep(TS)-buffer is introduced to avoid redundant calculations of LSTM gate operations to reduce system latency. Compared with a software implementation of the AutoEncoder-LSTM on CPU and GPU, our FPGA design is respectively 36.3 and 23.9 times faster and consumes 247 and 499 times less energy.