Performance of a neural network for recognizing AC current demand signatures in the space shuttle telemetry data

We describe performance of an analog neural network trained to identify signatures from the AC electrical power system on the Space Shuttle Orbiter. The network is based on the Intel ETANN analog neural network and identifies various electrical equipment from their transient signals. These signals are sampled during their first 6 seconds through a tapped analog delay line and presented to 60 input neurons. At a time resolution of 1/10 sec, the network will recognize the electrical "fingerprints" by producing a "true/false" pattern on the output layer. Results are discussed in terms of two learning paradigms (BP and MRIII) as well as details on the training and on the errors obtained. The network has also been extended to include "overlapping" signals.