Uncertainty and Reliability Analysis of Neural Networks Based Soft-sensors With Applications in Wastewater Treatment*

To overcome the transparency and interpretability issues of neural network-based soft sensing, a data-driven framework by combining neural network prediction models and sensitivity analysis methods is constructed in this paper, with application in wastewater treatment plant analysis. The results show that readily biodegradable substrate effluent (SS-e) and ${\text{N}}{{\text{H}}_4}^ + + {\text{N}}{{\text{H}}_3}$ nitrogen effluent (SNH-e) are both important input factors affecting the difficult-to-measured effluent quality, chemical oxygen demand (COD), by resorting to PAWN sensitivity analysis and four neural network models. In addition, four neural network models worked together with bilateral sensitivity analysis for the input perturbation analysis, which shows that the addition of Gaussian perturbation to SS-e and flow rate influent (Q-in) has a greater effect on the prediction which results large fluctuations of output results. Finally, the effectiveness of this framework is confirmed based on above results.