To solve the problems of poor efficiency of production rule in big data and hard weight distribution in the case of heterogeneous information fusion, a knowledge representation method of production rule based on echelon inference and benefit principle was designed taking the data and voice as the information source for fusion. A grouping & ranking inference strategy was proposed on the basis of different sensitivity between mapped knowledge elements and unmapped knowledge elements. The weight of knowledge element was specified by the benefit value obtained from mutual information and information gain. We simulated the engine fault by using the flight data and the synchronous cockpit voice as the information source for making failure prediction. The result shows that our knowledge representation method can provide higher efficiency, stronger intensity and more accurate diagnosis result, which can be used in practical fault diagnosis and fault prediction.