Fault diagnosis of robot is difficult since it consists of several interconnected components. To solve this, system-level diagnosis framework is developed using the motor control signal, which can estimate the current health of two components: harmonic drive and timing belt, as well as the system performance. Features are extracted using wavelet packet decomposition and selected by trendability and separability. Artificial neural network and Gaussian process regression are used to evaluate components’ health and system performance, respectively. Finally, optimum maintenance is planned by simulating how well the system is restored when each component is replaced.