In this paper, a multivariate statistical method called Principal Components Analysis, PCA, is utilized for detection faults in a 3-PSP parallel manipulator. This statistical method transfers original correlated variables into a new set of uncorrelated variables. PCA method can be used to determine the thresholds of statistics and calculate square prediction errors of new observations for checking the system when a fault occurs in the robot. To investigate on the ability of the PCA method for faults detection of the robot, a nonlinear model-based controller called Computed Torque Control, CTC, is designed. In this control scheme, rigid-body inverse dynamics model of the robot is utilized to linearize and to cancel the nonlinearity in the controlled system. Also, instead of using the robot prototype model, direct dynamics of the robot is used in the robot-control system. In this paper, two faults are artificially applied to the robot-control system. These two faults consist of faults in servo drive or servo motors and faults in joints clearances or position sensors. Finally, these faults are applied on the robot throughout a desired end-effector trajectory and the resultant outputs are obtained for both with and without faults in the manipulator. Consequently, the desired and faulty outputs are compared and faults detection using PCA method for the robot is performed.