Development of Neighborhood Double Preserving Autoencoder Models for Soft Sensors in Industrial Processes

In recent years, soft sensors have been widely applied in the process industry. Autoencoder (AE) models possess the capability to extract deep features, which would be important for constructing regression models for predicting quality variables. It is reported that preserving the neighborhood structure of the original data is also important. However, existing AE models rarely consider this issue, and we would like to make the models more capable of neighborhood structure preservation based on manifold learning concepts. To solve this problem, a novel neighborhood double preserving autoencoder (NDPAE) is proposed in this paper. Different from the traditional AE, a novel loss function is established by introducing a double constraint term. This model can extract the neighborhood structure of data while maximizing the retention of its information, and more accurately predict the quality variables. The experimental results on the industrial sulfur recovery unit (SRU) process demonstrate that the proposed NDPAE model is more effective compared with other common models.