A feature extraction method based on adaptive multivariate variational mode decomposition (MVMD) combined with multi-variate multiscale dispersion (mvMDE) was proposed to address the problem of complex equipment fault signals and difficult fault feature extraction in early or composite faults. Firstly, a method for determining the decomposition layers of MVMD was proposed by defining the channel energy difference rate. Based on the signal itself, the decomposition layers are adaptively determined, which solves the problem of strong subjectivity in decomposition layers and was verified through simulation analysis. Secondly, a feature vector was constructed using the mvMDE values of intrinsic mode functions (IMFs) after adaptive MVMD decomposition, and input into a support vector machine for fault diagnosis and recognition. The effectiveness of the algorithm was verified using multi-channel vibration signals from the diesel engine. Finally, the influence of operating conditions, the number and location of measuring points on fault recognition rate was studied. The results indicate that the diesel engine exhibits the highest recognition rate when operating at half load and 1 500 r/min. The number of measurement points is beneficial for identifying misfire faults, and measurement points near the faulty cylinder contain more fault information.