Abstract
Batch manufacturing is central to fine chemicals, pharmaceuticals, and bioprocessing. Its operation evolves across phases and recipes, which yields high-dimensional trajectories and strong batch-to-batch variability. Meanwhile, key quality-indicative variables are often measured offline and cannot be used as online model inputs. This work presents an integrated deep learning framework that unifies soft sensing and process monitoring in a single module using only process variables as inputs. A multi-scale Temporal Convolutional Network with multiple kernel sizes extracts complementary dynamic features from sliding windows. These features are concatenated and pooled into a compact representation that feeds two task branches. A variational autoencoder branch reconstructs the input window and provides fault monitoring signals via reconstruction deviation while regularizing the latent space through KL divergence. In parallel, a prediction branch estimates the quality-indicative variable directly from the pooled temporal features without using the variational latent sample. This separation preserves a stable quality mapping while retaining a probabilistic reconstruction model for monitoring. During inference, reconstruction error and prediction error are fused into a joint state score that more comprehensively reflects system state changes than either deviation alone. Diagnostic heatmaps are produced from residual maps and optional SHAP attributions to highlight contributing variables and time segments. The framework is validated on the Industrial Penicillin Simulation, an industrial-scale penicillin fermentation benchmark. Results show stable convergence of reconstruction, prediction, and KL terms, clear fault-set monitoring trajectories, and interpretable heatmaps that support actionable diagnosis.