There have been some studies of exoskeletal robots to support upper limb motion (especially for the physically weak). However, most robots don't completely support internal/external rotation, one of 3 DOF in the shoulder joint. This is because the internal/external rotation is performed by the activities of the rotator cuff, which consists of deep layer muscles. In other words, it is difficult to recognize the surface electromyogram (SEMG) signals, which are generated at muscles far from the skin. In this paper, our aim is to quantify the differences in the SEMG, between the surface layer and deep layer muscles and apply them to the discrimination of the external rotation under the experimental evaluation with twelve young subjects. Four kinds of parameters, such as Zero Crossing (ZC), were selected to express the characteristic of high frequency component in the EMG signal. Specifically, the classification was shown to be 97% successful by applying two features to the learning machine. Hence, it is almost possible to assume that either the deep or the surface muscle is active and discriminate the motions which their muscles involve.