The knowledge about aluminum electrolysis production has properties of interdisciplinarity, uncertainty, multi-source and heterogeneity. Using the traditional knowledge representation may give rise to the problems like combination explosion, ambiguity, knowledge selection difficulty, poor understandability and so on, all of which will thereby reduce the efficiency and accuracy of the condition diagnosis of aluminum electrolytic cells. To solve the above problems, a new knowledge representation method that combines Bayesian conditional probability and traditional semantic net is proposed in the paper, and it is named the knowledge representation model based on Bayesian probability semantic network. In this method, the knowledge element and the probability are used for correlation and multiplication respectively, which can solve effectively the problems of ambiguity, knowledge selection difficulty and poor understandability in the process of knowledge reasoning. Meanwhile, knowledge reduction method based on combination elimination is proposed, aiming to solve the storage and calculation problems caused by redundantly repeated knowledge factors and high matrix dimension in the relational matrix. The rationality, feasibility and effectiveness of the Bayesian probability semantic network model are verified by a case study of the aluminum electrolytic cells condition diagnosis.