Adaptive deep fusion neural network based soft sensor for industrial process

Deep neural networks have become an important tool for soft sensor modeling. However, common deep autoencoder networks are limited to mining the effective information of each input layer during hierarchical training, ignoring the loss of effective information in the original input data and accumulating it layer-by-layer, resulting in incomplete feature representation of the original input. At the same time, there is a lack of mining for temporal correlation between process samples and an adaptive mechanism to strengthen temporal related features, resulting in insufficient process information mining. In addition, deep neural networks generally have overfitting problems. To this end, an adaptive deep fusion neural network (ADFNN) method is proposed. This method reconstructs the original input data at each layer of the feature extraction network. By using the reconstructed original input error in pre-training loss, it reduces the loss of effective information from the original input. Simultaneously, incorporating sliding windows and self-attention mechanisms to select and calculate the contribution of historical samples to the current sample, integrating temporal related information, and overcoming dependence on high-dimensional local features by minimizing Kullback-Leibier (KL) divergence penalty terms. Finally, the temporal features are adaptively weighted and connected to a fully connected network to achieve quality prediction. Simulation experiments were conducted in cases of debutanizer and industrial polyethylene production to verify the effectiveness of the proposed method. The experimental results show that compared to the stacked autoencoder (SAE), target dependent stacked autoencoder (TSAE), and stacked isomorphic autoencoder (SIAE) models, the proposed method ADFNN has improved prediction accuracy by 2.4%, 1.7%, and 0.5% in the case of a debutanizer, respectively. In the industrial polyethylene production case, it has increased by 3.6%, 3.3%, and 1.8%, respectively. By using the reconstructed original input error in pre-training loss, it reduces the loss of effective information from the original input; Simultaneously incorporating sliding windows and self-attention mechanisms to select and calculate the contribution of historical samples to the current sample, integrating temporal related information, and overcoming dependence on high-dimensional local features by minimizing KL divergence penalty terms. Finally, the temporal features are adaptively weighted and connected to a fully connected network to achieve quality prediction.