The self-excited vibration of rolling mills seriously affects the product quality during the rolling process of the strip steel. When the rolling mills operate at variable speeds or working conditions, the self-excited vibration has many influence factors and the milling housing possesses diverse modal information. A single sensitive characteristic index cannot illustrate the trends and states in self excited vibration of rolling mills. For solving this situation of rolling mills, first, various signal processing technology is used to analyze the characteristic of vibration signal of rolling mills. Multiple features extraction can ensure complete and effective information for condition assessment of self-excited vibration of rolling mills. Then, a feature selection method based on maximum minimum redundancy is devised to remove unrelated or redundancy information and reduce the features’ dimensionality. The optimal feature subset is selected from multiple features. Last, SOM neural network fusion method is used to construct a characteristic index which illustrates the trends of self-excited vibration of rolling mills. A method of setting alarm threshold is proposed based on 6σprinciple. The vibration data of self excited rolling mills is used to verify the affection of alarm threshold. The result shows that the constructed index could forecast the trends of self excited occurred, which enable the operators to slow down the rolling mills in time and reduced the fraction defective and cost.