<p>Rotor–stator rub-impact faults, as a typical secondary failure in large rotating machinery, exhibit weak characteristic and diagnostic complexities. In order to effectively extract fault information, an integrated methodology, using improved multivariate variational mode decomposition for rub-impact fault detection and location is proposed. Firstly, in order to achieve accurate signal decomposition, the key parameters of multivariate variational mode decomposition, modal number and penalty factor, are adaptively determined by Bayesian optimization algorithm. Taking the negative signal-to-noise ratio as the objective function of Bayesian optimization minimizes the impact of noise on the signal and maximizes the intensity of the useful signal after decomposition. Secondly, signals are adaptively decomposed based on the determined modal number and penalty factor to obtain the corresponding component signals. Thirdly, from the perspective of signal complexity change before and after a rub-impact fault, and taking advantage of the complexity parameter in Hjorth parameters, which is robust to noise, the feature vector is constructed of each component signal. Finally, the rub-impact fault and rubbing positions are identified based on the constructed feature vectors with 1D convolutional neural networks. In ten independent tests, the recognition rate of the proposed method was all above 98.5%. Compared with the comparison methods, the average recognition rates increased by 9.5%, 15.4%, 22.3%, 29.0% and 6.2%, respectively.</p>