A Deep Nonconnectionist Learning Framework for Industrial Data Modeling

Despite the extensive applications of deep neural networks in data modeling filed, there are still some obvious deficiencies for the implementation in modern industrial cases. There are mainly reflected in the following aspects: first, the architectures are difficult to configure; second, the modeling process is time-consuming; third, the training procedure easily falls into the local optimum situation. To overcome these problems, exploring the nonconnectionist learning model has become a popular topic recently. This article proposes a deep nonconnectionist learning model based on kernel principal component regression (KPCR), which is referred to as stacked KPCR (SKPCR). By stacking multiple KPCR modules, a multilayer learning model is constructed by adopting hierarchical feature extraction. In SKPCR, the model structure is determined incrementally and there is only one parameter needed to be configured for each layer. Furthermore, an enhanced learning strategy is designed for alleviating the information loss problem in the training process. An actual industrial case is used to validate the effectiveness, including the prediction performance and modeling efficiency, of our proposed method.