Flexible Clockwork Recurrent Neural Network for multirate industrial soft sensor

Abstract

Data-driven-based soft sensors play significant roles in predicting key quality and optimizing the production process. Considering the difficulty and cost of variable acquisition, these variables are generally collected at different sampling rates. However, traditional soft sensor methods assume that data are uniformly sampled, which cannot be directly applied to multirate industrial scenarios. In this paper, a Flexible Clockwork Recurrent Neural Network (FCW-RNN) is proposed for multirate industrial soft sensors. First, the multirate data are divided into different variable groups according to their sampling rates. Then, an FCW-RNN is developed to map these variable groups into a common hidden space separately. A flexible clockwork mechanism is designed to incrementally update the hidden space at each moment based on the specific variable groups that have been sampled. Considering different sampling rates, the hidden space will be updated with time until all variable groups are sampled. In this way, we integrate the information of multirate data into the uniform hidden space step by step. Finally, a prediction module is established to calculate the hard-to-measure variables based on the hidden space. The effectiveness of FCW-RNN is demonstrated in a real coal mill case.