Local Neuron Analysis for Monitoring Dynamic Nonlinear Batch Processes

Dynamic characteristics and nonlinearity may exist in both the time and batch directions of batch processes, so monitoring the batch process is a challenge. The neural component analysis (NCA), which uses neural networks to address nonlinearity, is effective for continuous processes. However, it cannot handle dynamic features in batch processes, as it assumes time-independent variables and is designed for 1-D data, making it unsuitable for batch process monitoring. Herein, this article proposed a novel batch process monitoring method called local neuron analysis (LNA). First, in contrast to existing NCA-related methods, LNA innovatively integrates the dual advantages of local receptive field (LRF) mechanisms and autoregression (AR) modeling. The convolutional and deconvolutional modules effectively extract cross-batch correlation features, while the AR model precisely captures intrabatch temporal dynamic characteristics. Furthermore, unlike black-box gradient estimation, we develop a theoretical LNA optimization framework with derived analytical gradients for interpretable and efficient learning. Third, the proposed statistical index achieves theoretically balanced fault sensitivity and diagnostic robustness, creating an interpretable paradigm for nonlinear dynamic monitoring while preserving real-time computational efficiency. Finally, the simulation results of the proposed second-order nonlinear autocorrelation model demonstrate that LNA overcomes the challenges NCA faced in detecting faults in earlier stages and is particularly sensitive to faults resulting from recursive accumulation in nonlinear processes. Moreover, when applied to the fed-batch penicillin production (FBPP) process, LNA achieves high detection accuracy, low false alarms, and minimal delay in detecting minor, incremental faults in both aeration rate and agitator power.