Aiming at the problems of the shortage of fault samples for active new equipment and the low accuracy of existing algorithms for fault detection, the multiple kernel learning(MKL) and the one-class extreme learning machine (OC-ELM) are combined, and the $l_p$-norm constrainted multiple kernel learning one-class ELM($l_p$-MKOCELM) is proposed. Under the $l_p$-norm constraint, a mathematical optimization form combining the MKL and the OC-ELM is defined, and the update method of combination weights of the base kernel and Lagrange multipliers are derived.To facilitate the implementation of fault detection, the test statistic and detection threshold based on the $l_p$-MKOCELM are defined. The approximate equivalence of different norm constraints is confirmed through experiments. The proposed method is applied to the commonly used UCI data set and test data of an equipment. The experimental results show that, compared with the traditional SVDD, PCA, OC-SVM, and OC-KELM, the proposed method can significantly improve the detection accuracy while balancing missing alarm and false alarm.