An adaptive mode convolutional neural network based on bar-shaped structures and its operation modeling to complex industrial processes

Abstract

Optimal operation modeling plays an important role in complex industrial processes; however, with the increasing complexity and high nonlinearity in industrial processes, it becomes more and more difficult to establish an accurate operation modeling using first-principles methods. In this paper, an adaptive mode convolutional neural network framework based on bar-shaped structures (BS-AMCNN) is proposed, which is a data-driven model. First, a bar-shaped structure is designed to deal with the industrial process data specifically. The bar-shaped structure can transfer the advantages of CNN on processing image data to processing industrial process data. Meanwhile, the convolution windows and pooling windows in the proposed BS-AMCNN algorithm is replaced by translation-only sliding bar-shaped windows. Therefore, the algorithm can adjust the CNN structure adaptively among three different modes depending on different process statuses. the optimal operation model can be obtained with the proposed BS-AMCNN method accordingly. An experiment on real complex industrial process, methanol production process, is carried out, which validates the effectiveness of the proposed method. The proposed method is further compared with the traditional CNN method, and the back propagation (BP) method. The results demonstrate the effectiveness of the proposed method.