Deep matrix factorization considering dynamic constraints to complete missing data of complex industrial processes

Abstract

In the complex industrial processes, data loss is an unavoidable issue. Due to the lengthy process flow and complex reaction mechanisms, traditional data completion methods fail to deliver satisfactory results when data loss occurs. To address this challenge, this paper proposes deep matrix factorization considering dynamic constraints (DMFDC). This algorithm combines traditional matrix factorization with artificial neural networks, leveraging the strengths of neural networks to approximate nonlinear mappings in latent variable models and utilizing all available information to minimize discrepancies between raw and generated data. Additionally, DMFDC accounts for the dynamic characteristics of the complex industrial system, employing differential operations to transform irregularly changing industrial data into a more stable sequence, thereby enabling the model to better capture data evolution patterns. This approach allows DMFDC to intelligently address the issue of missing dynamic data in the complex industrial process and to predict missing values more accurately. To evaluate its effectiveness, we conducted case studies under various missing data conditions based on a digestion dataset collected from actual alumina production sites. The results indicate that DMFDC achieves higher data completion accuracy than other methods, confirming the applicability of our approach in diverse situations involving missing data.