Aiming at the structure optimization of the extreme learning machine(ELM), an improved sensitivity-analysis based pruning ELM(ImSAP-ELM) algorithm is proposed. The ??₂-regularization factor is introduced into the SAP-ELM by using the ImSAP-ELM. The leave-one-out(LOO) criterion is utilized for selecting an appropriate number of hidden neurons. In addition, the computing expression of output weights based on singular value decomposition(SVD) is deduced, which overcomes the problem that computing result is invalid when the matrix is singular. The proposed ImSAP-ELM is applied to fault prediction. Associated with some groups of known fault data under the same fault type, a number of ImSAP-ELM based models are built. All the prediction values from different ImSAP-ELMs are fused with weighted sum. The case study on a certain unmanned aerial vehicle transmitter shows that, comparing with the ELM, the optimally pruned ELM(OP-ELM) and the SAP-ELM, though ImSAP-ELM time consuming is the highest, the prediction error of the ImSAP-ELM is lower than other 3 algorithms.