Motivated by decoupling and diagnosing of the multiple faults of diesel engine lubrication system, a fault diagnosis method based on Bayesian networks is proposed. The establishment of a diagnosis Bayesian networks model for the lubrication system includes two processes: representing the coupling relationship between multiple faults through directed acyclic graph, and setting up the quantitative parameters in the form of probability. Before the implementation of the proposed method, the diagnostic task of the lubrication system is decomposed according to the types of failure. A fault tree model is utilized in every single diagnostic subtask to analyze the coupling between failures and symptoms and between different failures. The fault tree can be naturally converted into the topology of the lubrication system diagnosis Bayesian network. In the aspect of quantitative parameters, the noisy-OR/AND model is adopted to analyze the strength of causal relation between failures and symptoms. The historical health information can be represented through the prior probability of faults. The methodology proposed is illustrated by taking the failure‘low oil pressure’as an example. Testing results from two actual fault cases verify the effectiveness of the proposed methodology.