Fuzzy-Constrained Incremental Random Weight Network for Industrial Process Soft Sensing

Data-driven modeling has emerged as a powerful approach for soft sensing of key performance indices in industrial processes. This approach can leverage the collected process data to establish precise soft sensors without requiring explicit physical knowledge. To cope with uncertainties and nonlinearities present in industrial process data, a novel fuzzy randomized incremental model named fuzzy compact constraint-based incremental random weight network (F-CCIRWN) is proposed for industrial process soft sensing by integrating the Takagi-Sugeno (T-S) fuzzy system into CCIRWN. Specifically, all inputs from the input layer are sent to a set of T-S fuzzy subsystems to enhance fuzzy reasoning capability. The defuzzification outputs produced by the fuzzy subsystem layer are directly linked to the output layer. Additionally, the fully connected layer is introduced between the fuzzy subsystem and output layers to perform nonlinear transformation. Then, the fuzzy c-means approach is adopted to estimate the clustering centers of Gaussian membership functions and determine the number of the fuzzy subsystem rules associated with the fuzzy subsystem layer. Subsequently, the Greville’s method is employed to design a compact constraint that is capable of effectively choosing the well-performed random parameters in the fully connected layer. Finally, the comprehensive performance evaluation of the proposed F-CCIRWN is conducted on nonlinear dynamic system identification, chaotic time sequence prediction, and data-driven modeling based on real-world benchmark datasets and an industrial dataset. The results indicate that the proposed F-CCIRWN is more suitable for industrial process soft sensing compared to some other state-of-the-art modeling methods.