The attitude quaternion and angular velocity of the spacecraft need to be determined for missions such as in-orbit repair and removal. Failed satellites are often in a free tumbling state， usually with flexible sails， and their motion patterns are more complex than those of rigid sails. On the one hand， the attitude determination of space invalid satellites often uses LIDAR and binocular camera as measurement equipment. Its measurement accuracy is often affected by light and magnetic field， which will greatly interfere with the recognition accuracy. On the other hand， the mass characteristics of flexible spacecraft are easy to change， which makes it difficult to accurately describe its dynamic model. Aiming at the problem that it is difficult to obtain the state of the flexible free-tumbling target， an attitude estimation method based on traceless Kalman filter is proposed， and neural network is used to compensate the model error of flexible spacecraft. The simulation results show that the prediction error of attitude quaternion of the flexible spacecraft by unscented Kalman filter（UKF） algorithm is within the range of 10-3， and the angular velocity error is within the maximum of 0.08 rad/s， and the prediction error of quaternion is stable within the range of 9×10^-4 and the angular velocity error is stable within the range of 1.5×10^-3 after using neural network to compensate the dynamics model. The results show that after using neural networks to compensate for the uncertainty terms of the flexible spacecraft dynamics model， the attitude estimation accuracy of the UKF for flexible free-tumbling targets is sufficiently high.