Addressing the limitation embedded in traditional reliability analysis and prediction approaches when dealing with a multi-level & imbalanced information system, a Bayesian-based information extraction and aggregation approach is proposed. By introducing the direct prior, induced prior and combined prior, the Bayesian inference process is reformulated. Novelties include an adaptive Bayesian melding algorithm, which integrates the lower level information to compensate for higher level information inadequacy for a better reliability analysis and prediction results. A multi-level electromechanical system with imbalanced information is demonstrated for validation and benefit illustration, the results show an improved accuracy compared with traditional approach.