To improve the operational efficiency and measurement accuracy of the dynamic check weigher, the interference of mechanical vibration to the measurement and the generating mechanism of the sensor's nonlinear characteristics are deeply analyzed. A multi-layer BP neural network based on ADAM optimizer is proposed to realize the nonlinear correction of weighing sensor and estimates the dynamic weighing results accurately. The classical gradient descent algorithm, gradient descent algorithm with momentum and root-mean-square propagation algorithm are compared with the ADAM algorithm through experiment. According to the results, the ADAM algorithm had faster convergence speed and more accurate prediction results as it comprehensively considered the first and second sample moment of parameter's gradient. The high speed dynamic check weigher with full range of 400 g and maximum running speed of 2 m/s is manufactured, The type test results showed that all of its indicators are meet the requirements of national standard GB/T 27739-2011 Automatic Divider for XIII check weigher.