Robust incremental random RNN for process industries soft sensing

Recurrent neural networks (RNNs) serve as critical tools for soft-sensing modeling in process industries. However, their network architectures typically require manual tuning, and training data contaminated by abnormal samples can significantly hinder model generalization. To address these limitations, we present an Incremental Random Recurrent Neural Network (IRRNN) model that enables automated architecture expansion and random parameter learning. A tailored Robust Incremental Random (RIR) learning algorithm is subsequently developed for this framework. The algorithm employs derivative-based abnormal samples detection and state resampling techniques to transform contaminated datasets into semi-supervised learning scenarios. This is followed by a constrained optimization approach incorporating quadratic inequality constraints to determine output weights, further fortifying model robustness. Benchmark evaluations and industrial case studies on hematite grinding processes soft sensing demonstrate the proposed method’s superior resilience to abnormal samples and prediction accuracy compared to conventional approaches.