Model-level Weight Update Domain Adaptive Dynamic CNN Soft Sensor for Free Calcium Ion Concentration in Cement Clinker

Abstract

The paramount determinant of clinker quality within the cement clinker production process resides in the precise prediction of F-CAO (free calcium ion) content. Given the intricate nature of cement clinker production, a profound coupling and temporal progression are inherently intertwined between the process data, as working conditions undergo gradual transformations over time. This research suggests a Model-level Weight Update Domain Adaptive Dynamic CNN Soft Sensor for Free Calcium Ion Concentration in Cement Clinker to overcome the aforementioned issues. We introduce a graph-structured two-dimensional methodology aimed at transforming both labeled and unlabeled data within the original process dataset into two-dimensional process data blocks. This innovative approach serves to enhance the spatial and temporal attributes of the labeled samples within the comprehensive sample set. Importantly, this methodology effectively addresses the intricate challenges posed by robust data coupling and the substantial presence of unlabeled samples. In response to the challenge posed by the frequent temporal fluctuations in working conditions, we introduce the Model-level Weight Update Domain Adaptive (MWUDA) approach. MWUDA leverages the hidden layer outputs from both the source and target domains to optimize the weight updates of the fully-connected layer within the CNN model. This strategic weight optimization is aimed at ameliorating the performance degradation observed in static models when confronted with shifting patterns in sample distributions. Subsequently, a limited subset of source domain samples exhibiting analogous regularities to those in the test samples is identified, enabling the acquisition of refined weights through the MWUDA method. Employing this strategy safeguards the precision of model updates throughout the MWUDA computation procedure, ensuring a robust and accurate adaptation process. The effectiveness of the proposed method is verified in an experimental study with real cement production data.