<p>Data-driven models have been an effective way for the early-warning of membrane fouling (MF) in wastewater treatment processes. However, the collected samples in this process always suffer from disjunction due to changes in working conditions or serious perturbations, which can degrade the performance of continuous early-warning with a data-driven model. To solve this issue, a robust intelligent early-warning method is proposed to discriminate the incidents of MF. Primarily, the collected samples are divided into several sequences, followed by the observation of working conditions based on the spatial cross-entropy loss. This approach captures sample fluctuations under various working conditions, which facilitates the identification of underlying patterns. Then, a robust predictive model is built to identify the state of MF with a compact recurrent fuzzy neural network (C-RFNN). The C-RFNN adapts its structure by adding or pruning fuzzy rules with temporal information entropy, which enhances model adaptability to dynamic environments. Third, an adaptive parameter update strategy is employed to optimize the model parameters in batch processing. With this strategy, the model achieves efficient parameter updates and enhances its robustness against perturbations over time. Finally, the proposed approach is validated with practical examples. The experimental outcomes demonstrate that it surpasses baseline models in both predictive accuracy and robustness.</p>