The invention relates to a neural network-based ultra-high-speed permanent magnet synchronous motor speed observation method. According to the method of the invention, a speed estimation equation is discretized; the discrete linear equation is substituted into a three-layer dynamic recurrent neural network to calculate a rotation speed estimated value and a rotor angle estimated value; the rotor angle estimated value is corrected through current estimation error; and the rotor angle estimated value is further substituted into the dynamic recurrent neural network for operation; and the rotationspeed of an ultra-high-speed permanent magnet synchronous motor can be estimated in real time in the dynamic operation process of the ultra-high-speed permanent magnet synchronous motor. With the method of the invention adopted, sensorless rotation speed observation is realized, and the rotation speed control precision of the motor is improved.