Owing to safety limitations and data collection costs, scenarios with imbalanced data usually arise, posing a great challenge for precise fault diagnosis. Targeting imbalanced fault diagnosis and the high computational cost of mainstream ensemble learning methods currently used, this article proposes a lightweight and accurate scheme based on a progressive joint-transfer ensemble network (PJTEN) and a Markov-lightweight strategy (MLS). Specifically, a PJTEN is developed, incorporating a multiple excitation-channel attention basic estimator and progressive joint-transfer strategy (PJTS) to maintain diversity of basic estimators better and focus more on key information from minority classes. Besides, the MLS guided by Markov transition probabilities is for the first time constructed for ensemble learning to reduce the network redundancy by alternating optimization. Using a standard dataset and a brand-new dataset of a real ship propulsion system, the proposed method achieves leading results in Accuracy, F1 score and MCC, compared with eight cutting-edge methods, thereby validating its substantial value. In terms of lightweight operation, such as temporal complexity (TC), spatial complexity (SC), and time efficiency, it is also ahead of the latest ensemble-based methods.