Using Multiple Nets to Address Noisy Data

P.L. Juell, S. Marla, and L.J. Francl (USA)


multiple neural nets, noise, training, backpropagation, model development


We present a way to improve the answer quality of neural networks faced with noisy input data. We have input values that are often contaminated by sensor errors. We trained a network to estimate our noisy variable, in our case the leaf wetness variable. This predicted output was included with the real sensor input data in training the neural network. The algorithm uses self-organizing data filters to filter noise present in the input data and collects underlying desirable features efficiently in neural net training. The multiple network system improved answer quality over other attempts to address the problem of noise in the data. The developed network design was successfully employed with noisy input weather data for development of a model of wheat crops including a noisy variable for leaf wetness.

Important Links:

Go Back