Abstract The kernel extreme learning machine (KELM) has attracted attention for failure diagnosis of turbofan engines, but its application for time-sensitive scenarios is inherently limited by its lack of sparseness.… Click to show full abstract
Abstract The kernel extreme learning machine (KELM) has attracted attention for failure diagnosis of turbofan engines, but its application for time-sensitive scenarios is inherently limited by its lack of sparseness. The original KELM constructs the hidden layer using all the training samples; thus, the real-time performance may be seriously degraded for large datasets. To address this limitation, a novel iterative picking scheme for improving the sparseness of KELM is proposed in this study. It has two noteworthy features, a compact structure and a sparse solution, which gives better real-time performance. The proposed scheme improves the sparseness of the KELM algorithm by two alternating components, an insertion strategy to expand the network and an elimination strategy to reduce the scale of the network. Validation on regression and classification benchmark datasets demonstrates that the proposed algorithm can produce a sparse network structure with fewer hidden nodes without sacrificing the model's accuracy. Finally, simulations of failure diagnosis of a turbofan engine show that the proposed algorithm can perform at an accuracy comparable to that of KELM with a much faster testing speed, which is crucial in real-time and onboard applications.
               
Click one of the above tabs to view related content.