Abstract: Levelling and stability of agricultural machinery can be required in hilly and mountainous areas. In this study, an electromechanical machine was designed to combine with the omnidirectional levelling control system using BP neural network PID. Firstly, the structure scheme of the electromechanical machine was provided, including an upper working platform, a middle omnidirectional leveling, and a lower tracked driving. The electromechanical leveling system was achieved in the maximum leveling angles of 20° and 25° in the lateral and longitudinal directions, respectively. Secondly, the angle error leveling was used for the omnidirectional leveling system in the control principle. A mathematical model of the servo-electric cylinder was established for the servo motor and transmission. Then, the BP neural network PID algorithm was designed for the leveling control, in which the weight coefficients of the BP neural network were dynamically adjusted, according to the error of body pitch angle. The PID parameters were optimized to improve the leveling speed, accuracy, and stability of the electromechanical machine. Furthermore, a servo-electric cylinder model was established for the BP neural network PID controller in Matlab/Simulink. An integrated mechanical model of the entire machine was established in Adams software. The simulation and experiment were conducted to compare the leveling control performance of improved and classical algorithms using the Matlab/Simulink-Adams co-simulation platform. A classical PID control algorithm was used as a reference. The simulation results show that the better leveling control performance of the BP neural network PID control algorithm was achieved in the lateral leveling control with a tilt angle of 20° and the longitudinal leveling control with a tilt angle of 25°, compared with the classical one. The average reduction in the lateral and longitudinal leveling rise time was 20.09%, and the average reduction in the leveling stabilization time was 18.99%. The steady-state error was controlled within 0.30°. Finally, the static and dynamic tests were conducted on the physical prototype. The static test results show that the BP neural network PID control shared a shorter time of levelling stabilization, smaller maximum inclination overshoot, and lower steady-state error in both lateral 20° and longitudinal 25° levelling control, compared with the PID control, the differences in the leveling stabilization time were 0.5, and 0.4 s, respectively, and the differences in the maximum inclination overshoot were 1.0°, and 0.5°, respectively. The differences in the steady-state error were 0.20°, and 0.15°, respectively, indicating the errors within a small range. The dynamic test results show that the electromechanical working machine was efficiently operated at a speed of 3 km/h on hilly terrain with uneven surfaces when the inclination angle of the fuselage changed significantly. Compared with the PID control, the BP neural network PID control reduced the average stabilization time by 0.5 s for both longitudinal and lateral leveling, resulting in a higher leveling speed. Additionally, the errors in the longitudinal and lateral leveling angles were within ±1.5°, indicating the higher leveling accuracy. The effectiveness of the leveling system and the BP neural network PID control algorithm was verified through experiments. The automatic leveling control of the working platform was realized to fully meet the requirements of levelling performance for the hilly and mountainous operations. The automatic leveling technology of agricultural machinery was avoided the work quality that caused by uneven road surfaces, and large slopes. The finding can also provide a practical reference to explore the structure design and leveling control of the omnidirectional leveling system in the electromechanical working machine, particularly for agricultural machinery.