Abstract: As the service time of military equipment increases, equipment failure data is continuously accumulated during events such as routine maintenance, training, and combat readiness exercises, and the data presented is often imbalanced to varying degrees and consists of small samples. In addition, due to fault tolerances of various electrical component parameters in the equipment and widespread nonlinearity and feedback loops of the circuit, it is often difficult to accurately express the fault mechanism using mathematical models. This poses new challenges for the fault diagnosis of equipment. To address the aforementioned problems, machine learning methods are widely used for fault diagnosis. The essence of such methods is that they transform a fault diagnosis problem into a pattern recognition problem. By learning the characteristic data of normal modes and various failure modes, a diagnosis model is constructed and, ultimately, a diagnosis strategy is formed. Aiming at the problems of the unbalanced distribution of various fault samples from equipment and low fault diagnosis accuracy of existing algorithms, in this paper, we define a regularized weighted multiple kernel ensemble under a p-norm constraint by introducing a p-norm constraint weighted multicore extreme learning machine and an ensemble learning strategy based on the AdaBoost fault diagnosis model of extreme learning machine. Under the p-norm constraint, the model performed two types of adaptive sample weight distribution based on the size of various fault samples; simultaneously, the model combines the multisource data fusion and extreme learning abilities of the multiple kernel learning machine with high efficiency. The weight of a sample, W , is integrated into the optimization objective function of the multiple kernel extreme learning machine. Through the Adaboost integration strategy, the information-rich sample in the model is adaptively improved. Thus, the weight of a sample significantly improves the accuracy of fault diagnosis. Taking 6 UCI public data sets and 1 actual installation case as examples, a fault diagnosis experiment was conducted. The results of the experiment show that the model constructed in this study has significantly improved diagnostic accuracy compared with other models such as kernel extreme learning machine, weighted kernel extreme learning machine ( \boldsymbolW^\left( 1 \right) and \boldsymbolW^\left( 2 \right) weighting method), and weighted multiple kernel extreme learning machine under 1-norm constraint, and the model’s diagnostic performance impact is limited.