In our diagnostics studies, we want to develop solutions to dynamical estimation problems on real systems whose behavior is represented by discrete time samples of sensor data captured and stored in a very large database. We have developed methods to utilize the generalized extended Kalman filter (GEKF) for training time lagged recurrent neural networks (TLRNN's). Our results indicate that the GEKF algorithm can be effectively used to train large TLRNN's on extensive time series of noisy data, without the need to add ad hoc noise in the training process to sustain performance improvement. These results have been achieved not only on the training data but in tests of generalization performance, measured by analysis of out of sample data drawn from a similar, but different system.
展开▼
