Abstract: A fault diagnosis method for conventional circuit breakers based on a multi-task learning combination model is proposed to address issues such as individual sample differences, noise interference, and difficulty in determining classifier parameters in vibration signals. Firstly, multivariate variational mode decomposition (MVMD) is used to decompose the vibration signal and obtain intrinsic mode functions (IMFs) that meet the threshold requirements. The time-domain and frequency-domain features are accurately extracted to reduce the impact of noise interference and signal differences; furthermore, the kernel principal component analysis (KPCA) algorithm is adopted to reduce the dimensionality of feature datasets. Different feature extraction methods are compared and the effectiveness and advantages of MVMD-KPCA are verified. Meanwhile, the parameters of the kernel extreme learning machine (KELM) are optimized by using the improved northern goshawk optimization (INGO) algorithm to enhance its classification performance. Finally, the reduced dimensional feature dataset is input into models such as INGO-KELM for comparison. The results show that the MVMD-KPCA method performs well in processing complex and nonlinear datasets. Compared with other comparative models, MVMD-KPCA and INGO-KELM have an average diagnostic accuracy of 99.83% for conventional circuit breakers, demonstrating stronger predictive ability and stability.