Novel Input Space-Relevant Distributed Extreme Learning Machine Integrated With LSTM for Industrial Soft Sensing

Soft sensor technologies have been extensively applied in industrial processes to utilize directly measurable process variables for inferring difficult-to-measure quality indicators. However, the effectiveness of existing soft sensors based on data-driven methods is often constrained by the inherent properties of process variables, including strong nonlinearity, intricate dynamics, and spatial coupling. In response to the aforementioned challenges, a novel model is developed in this study, namely the input spatially relevant distributed extreme learning machine integrated with LSTM (NISRDELM-LSTM), for predicting key quality variables. In the proposed NISRDELM-LSTM model, a clustering approach based on the k-shape method is employed to partition the input feature space, enabling effective categorization of sample data. The Distributed Extreme Learning Machine (DELM) subsequently maps these clustered results into a high-dimensional feature space through nonlinear transformations, enhancing feature representation. These enriched feature representations serve as discriminative inputs for the LSTM network, which captures temporal dependencies and refines prediction accuracy. By integrating DELM’s feature extraction with LSTM’s sequential modeling capabilities, the proposed NISRDELM-LSTM model enables accurate estimation of critical quality indicators within complex industrial operations. To demonstrate the predictive capabilities and practical utility of the developed model, extensive experimental evaluations are conducted on two representative industrial benchmark processes.