The identification process based on time-domain nonlinear subspace is cumbersome under different excitation level tests due to the difficulty in selecting the nonlinear description functions. Nonlinear restoring force data reconstructed by time domain nonlinear subspace method and nonlinear response data measured at the nonlinear position under different excitation levels are used to train the neural network model， which is then used to equivalently represent the response-nonlinear restoring force mapping relationship. As a result， the identification process is no longer dependent on the system models. In other words， the nonlinear restoring force can be obtained as long as the response at the nonlinear position is measured， which significantly improves the computational efficiency. Furthermore， a force identification method based on neural network and subspace method is proposed to deal with the problem that the external force of nonlinear system is difficult to measure. The effectiveness and feasibility of the proposed force identification method are verified via numerical and experimental studies of systems with clearance nonlinearity.