Linear pruning techniques for neural networks-based on projection latent structure

Based on a 5-layer neural network, nonlinear principal component analysis (NLPCA) has been widely applied in many problems. However, when applying NLPCA in modeling and monitoring chemical process, it´s hard to determine the number of hidden layer´s nodes. This results in a tendency to use networks much larger than required. A brief review of presented linear neural network pruning techniques is given, and an improved linear pruning method based on Projection Latent Structure (PLS) is presented. The advantages of propose approaches are discussed and illustrated via modeling the famous Tennessee Eastman chemical process.