Abstract
This paper proposes a Fast Geometric Constructive Neural Network (FastGCNet) that aims to provide a fast learning paradigm for data-driven modelling in resource-constrained Industrial Internet of Things (IIoT) devices. First, by integrating the concept of simulated annealing with the block incremental method, we have developed a dynamic block incremental approach and incorporated it into a compact angle control strategy, resulting in a compact angle control strategy with variable block incremental. This strategy allows multiple new nodes (i.e., node block) to be added to the hidden layer simultaneously, which in turn greatly accelerates the convergence efficiency of the network. Second, by combining the dynamic block increment method with the Grenville iterative method, we propose a fast iterative method that reduces the computational consumption of evaluating the output weights each time a new node block is added. Specifically, when adding a new node block, the method is able to obtain the output weights of the whole network after adding a new node block on the basis of the old output weights by iterative updating, which avoids the recalculation of the output weights of the whole network, and thus greatly reduces the computational consumption in the process of network construction. Thirdly, this paper demonstrates that FastGCNet has universal approximation property through theoretical analyses. The results from four benchmark datasets and a real ore grinding process show that FastGCNet can effectively reduce the computational consumption in the modelling process while maintaining considerable accuracy.