A soft sensor net based on the symplectic decomposition-global attention reconstruction architecture for biopharmaceutical industry

Non-linearity, time-varying properties, and high noise levels in biopharmaceutical process data have been recognized as critical factors affecting the accuracy of data-driven soft sensors. To address these issues and enhance prediction precision, we introduce BPSN, an innovative soft sensor framework grounded in the symplectic decomposition-global attention reconstruction architecture. Symplectic geometry mode decomposition effectively adapts to data complexity and reduces noise. A reconstruction module combines global attention mechanism and reversible instance normalization to enhance sharp signal features via Manhattan distance while addressing internal drift. Experiments show that the proposed soft sensor model outperforms state-of-the-art models in predicting key indicators: bacterial concentration, viscosity, and reducing sugar content in the erythromycin fermentation process. This illustrates its practical applicability and exceptional performance in biopharmaceutical industry. The source code is available at: https://github.com/Joss0623/BioPharmaSoftNet.git.