The flavor of beer is an important means of evaluating its quality. Beer flavor is the integrated embodiment of beer smell and taste information. The aroma and taste of beer were detected by electronic nose and tongue fusion system. Principal component analysis (PCA) was respectively used for reducing the dimension of detected information, and the principal component of test data by electronic nose and tongue were extracted to fuse as the characteristic data. The classification of beer was achieved by smell and taste comprehensive information. Due to the difference in data of sensor array, the traditional K-means algorithm clustering results were depended on the selection of initial value, and it was easy to fall into local optimum. A modified K-means algorithm based on particle swarm optimization was proposed, which was based on the characteristics of fusion data. The weight coefficient was optimized in the course of operation. With the increase of iteration number, the convergence speed was adjusted, the particle search tended to be more balanced. Meanwhile, the compression factor was introduced to balance the global and local conflicts. Compared with K-means algorithm, the modified algorithm had better global convergence in experiments. It also can overcome the disadvantage which was easy to fall into the local optimum, and converge to the optimal solution. The experimental results showed that the clustering effect in five kinds of beer was obvious, and the correct rate was stable at 93.3%.