It is a linchpin for ensuring the trajectory accuracy, positional accuracy and kinematic stability of the end-effector to obtain the parameters accurately of the nonlinear friction torque model of robot joints. robot However, it is difficult to obtain these parameters accurately torque by experiments. This work develops a data-driven inverse method for key parameters of nonlinear friction torque model. Firstly, a kinematic model of three-joint robot is established, using which the angular displacement, velocity and acceleration of joints are obtained for any planned trajectory. The Newton-Euler inverse dynamics analysis is then carried out to calculate the torques of the joints considering different parameter values of stribeck nonlinear frictions, for a given trajectory at the end-effector. Next, sensitivity analysis of the friction parameters to the torques is carried out to extract the sensitive friction parameters. Finally, an inverse problem neural network is established. Sampling in the sensitive parameter space is carried out by random sampling method. And the sample points needed for training the inverse problem neural network are calculated to train the inverse problem neural network. The results show that the average error of friction parameters is less than 2.5%.