PLSTM-MTGF: A deep learning fusion model enabling real-time multi-target monitoring of penicillin fermentation via near-infrared spectroscopy.

Real-time, multi-parameter monitoring is essential for optimizing industrial penicillin fermentation, yet conventional assays are slow and prone to contamination. This study introduces a lightweight dual-stream fusion model-Partial Least Squares Regression-Long Short-Term Memory with Multi-Task Gated Fusion (PLSTM-MTGF)-for online monitoring of four key indicators via near-infrared spectroscopy: residual sugar, amino nitrogen, cell density, and potency. The architecture integrates partial least squares regression for spectral compression with a long short-term memory-self-attention network to capture process dynamics, and employs a multi-task gated fusion mechanism to balance task-specific representations. Evaluated on 24 industrial batches under a strict batch-wise scheme, the model achieved high accuracy across all targets (R2 = 0.89-0.98; RPD = 3.0-7.1) and outperformed representative linear and deep-learning benchmarks. Ablation studies confirmed the indispensability of both spectral and temporal streams and the gating strategy. With a compact size of 0.97 M parameters and real-time inference on a standard CPU, the model is readily deployable for routine bioprocess monitoring. These results validate that fusing chemometrics with deep learning provides a robust, interpretable strategy for multi-target, online bioprocess analytics.