A Periodic Recursive Network Based on Temporal Attention Convolution for Soft Sensing

Quality prediction ensures the safety of industrial production. However, the nonstationary period characteristic produced by repeated motions makes soft sensing difficult. In this paper, a period recursive network (PRN) is proposed to achieve nonstationary quality prediction. At first, period length of original data is calculated in the frequency domain space using fast Fourier transform (FFT). Secondly, original data is divided by period length and a temporal attention convolution network (TACN) is designed to extract the short-term dynamic features of each subperiod data. Then, a recursive structure is constructed to facilitate information transfer within individual subperiods and the features of multiple subperiods are concatenated recursively in PRN. The concatenated features have long-term memory across periods and will be used for soft sensing of quality variables. The quality prediction effect of PRN has been verified in a real thermal power generation process.