Abstract: Aiming at the problems of time-consuming pre-training and poor diagnostic accuracy in the process of unsupervised training of traditional Deep Belief Network (DBN), In this paper, an Improved Multi-Objective Dragonfly Optimization Adaptive Deep Belief Network (IMOD-ADBN) is proposed for analog circuit fault diagnosis. Firstly, an adaptive learning rate is proposed according to the similarities and differences of parameter update directions to improve the convergence speed of the network. Secondly, traditional DBN uses Back Propagation (BP) algorithm in the supervised tuning process. However, BP algorithm has the problem that it is easy to fall into local optimum. In order to improve the problem, IMOD algorithm is used to replace BP algorithm to improve the accuracy of network classification. In the improved MODA algorithm, Logistic chaotic imprinting and oppositional jumping are added to obtain the Pareto optimal solution, which increases the diversity of the algorithm and improves its performance of the algorithm. The proposed algorithm is tested on eight multi-objective mathematical benchmark problems and compared with three meta-heuristic optimization algorithms (MODA, MOPSO, and NSGA-II), and the stability of IMOD-ADBN network model is proved. Finally, IMOD-ADBN is applied to the diagnosis experiment of a two-stage four-op-amplifier double-second-order low-pass filter. The experimental results show that the proposed IMOD-ADBN can ensure classification accuracy on the basis of fast convergence, and IMOD-ADBN has higher diagnosis rate than other methods mentioned in this paper, which can realize the classification and location of high-difficulty faults.