Recently, clustering illustrates its importance in knowledge discovery. However, most of the considered algorithms are efficient only on those linear separable datasets. Although the kernel based methods perform better with non-linear separable ones, at the same time, they are suffered from the missing of priority knowledge of a suitable kernel. Two latest methods which approximate date to infinity dimensional kernel feature spaces are considered in this paper. Furthermore, new attribute multi kernel weighted fuzzy c-means are proposed by combining the randomly projected features. From the experiments shown in the simulation section, we conclude that we develop an improved version of attribute multi-kernel weighted fuzzy c-means, which achieves better accuracy.