Industrial Data Imputation Based on Multiscale Spatiotemporal Information Embedding With Asymmetrical Transformer.

Area of Science:

• Process Engineering
• Data Science
• Artificial Intelligence

Background:

• Missing data in process industries hinders data-driven monitoring and soft sensor modeling.
• Nonlinear process data with spatiotemporal coupling and distribution shifts challenge traditional imputation methods.

Purpose of the Study:

• To develop a novel data imputation framework, MSST-Former, addressing limitations of existing techniques for process data.
• To integrate global and local perspectives for enhanced imputation of time-series and multivariate process data.

Main Methods:

• The MSST-Former framework utilizes a hybrid 1-D CNN for local correlations and an encoder-decoder with iTransformer and Transformer blocks for long-term dependencies.
• A multilayer residual network embeds multi-scale features for robust data imputation.

Main Results:

• Experiments on real-world industrial datasets demonstrate the superiority and robustness of MSST-Former compared to baseline and state-of-the-art models.
• The framework effectively captures complex spatiotemporal correlations and handles distribution shifts in process data.

Conclusions:

• MSST-Former offers a powerful solution for missing data imputation in process industries.
• The proposed method enhances the efficacy of data-driven process monitoring and soft sensor modeling.