A Soft-Sensor Modeling Method Based on Gated Stacked Isomorphic Autoencoder

Soft sensing mainly studies the real-time prediction of some key performance indicators or quality variables in the actual production process, which has the role of guiding production in the actual production process. The stacked autoencoder is a multi-layer autoencoder structure. The input variables will be encoded and decoded through each layer of autoencoders, and the obtained hidden features will be retained as the input of the next layer. In this way, high-level data features can be successfully learned from the input layer to the intermediate layer. Stacked isomorphic autoencoder reconstruct an identical data input layer, and the reconstruction is performed by minimizing the error of the decoded data from the original input data. However, for a soft measurement model, some data information may also reduce the accuracy or generalization of the model. In order to overcome the above shortcomings, this paper proposes a gated stacked isomorphic autoencoder, which evaluates the contribution of each hidden layer feature through the gating unit, and then integrates the information of different hidden layers to complete the prediction and estimation of related main variables. Finally, the effectiveness and feasibility of the method are verified in practical industrial cases.

Automated summary

Soft sensing focuses on real-time prediction of key performance indicators or quality variables in production processes. Stacked autoencoders help learn high-level data features, while gated stacked isomorphic autoencoders integrate information from different hidden layers for prediction and estimation.