The paper aims to design a compound observer based on high frequency square wave injection and improved sliding mode observer to solve the problem that the back EMF is not easy to detect when the motor is running at low speed and the high speed position accuracy is low, to improve the product efficiency and quality of automatic packaging production line and improve the performance of speed sensorless vector control system of permanent magnet synchronous motor (PMSM). The high frequency square wave signal injection method was used to estimate the rotor position information at low speed, and the improved sliding mode observer (ISMO) was used to estimate the rotor position at medium and high speed. Secondly, the double boundary layer switching function and variable sliding mode gain were introduced to speed up convergence and reduce chattering. At the same time, PPL based on the single neuron PI was designed to analyze the rotor information, to improve the speed and position estimation accuracy. Compared with the single algorithm, the position estimation accuracy of the composite observer was improved. Matlab/Simulink simulation experiments and physical verification experiments show that compared with the traditional algorithm, the algorithm has the advantages of high position estimation accuracy and good robustness.