1. The parameters considered for the fault detection study included body-axes commanded rates, actual aircraft body-axes rates, and corresponding neural network outputs. The training data were collected by combining data from non-failed axes using several single axis failure cases. For the single axis failure cases, the failures happened 2-3 seconds after the beginning of data collection. No experimental data were collected for cases without a failure. The test data were generated by windowing the data 1.5 seconds before and 1.5 seconds after the failure. As a result, the "normal" part of the test data looks similar to the training data for each case. It is to be noted that some of the entries are zero at the beginning of the data set, because the pilot has likely not entered a command to move by then, and the aircraft would still beflying straightand level. 5.2 Experiments
2. Figures 8-10 display results for other failures. There are two types of results reported for different fault detection scenarios. One is the detection result (horizontal bar on the top), which displays raw detection result for each data point, and the other one is isolation result (horizontal bar at the bottom), which displays detection signal based on the buffering and sensitivity parameters defined by the user. All types of faults occur right after the 1000thtime stamp, butbased on the buffering and sensitivity, the anomaly alarmis triggered atvarious times as reflected in each of the detection result figures. For example, for the tail failure detection result, the actual fault occurs at the 1000thtime stamp, the detection occurs at the 1000thstamp, the isolation result gives warning at the 1000thtime stamp as well. However, the actual fault alarmgivenby the isolationresult is at the1200thtimestamp.