부산대학교 도서관

Data-driven soft sensor methods have been widely used in municipal wastewater treatment processes to achieve efficient monitoring of effluent indicators. However, the complex biochemical reaction mechanisms in the wastewater treatment process lead to process data with strong nonlinear and time correlation characteristics, which causes the performance of the current state-of-the-art soft sensor techniques to be limited. Therefore, in this article, a novel Transformer network is introduced to construct a soft sensor model. The model structure utilizes a positional encoding mechanism combined with a multihead attention mechanism for the parallel processing of data, which can establish global interdependencies in the time series to fully extract the long-term time correlation of the time series data. Subsequently, the model is introduced with a residual connection module to successfully ensure the extraction capability of the model for nonlinear characteristics while also avoiding the problem of gradient disappearance and ensuring the performance of the model. Finally, the effectiveness and feasibility of the proposed method were verified on the benchmark simulation model.