Process modeling and identification, Fault detection and diagnosis, Adaptive autoencoder, Neural networks
Abstract
This work proposes a novel neural network architecture based on Parallel Layer Perceptron. This new architecture aims to overcome some limitations of the aforementioned model. Thus, a neural network with multiple adjustable parallel layers capable of solving tasks with multiple outputs is achieved. The proposed algorithm is applied to monitor industrial processes in its adaptive autoencoder form, capable of iteratively updating its parameters from a data stream. The model is validated in two examples that show its effectiveness in input reconstruction and in providing a latent layer with significative features for event detection.
