Due to the complexity of signal components and weak fault features in the early fault of ammunition feeding and conveying system, a fault diagnosis method based on adaptive noise complete empirical mode decomposition (CEEMDAN) and least squares support vector machine (LSSVM) optimized by von Neumann topology (VN) improved whale algorithm (WOA) was proposed. After preprocessing the measured signal, i.e. removing the trend term and zero drift, the feed signal was decomposed by CEEMDAN to obtain the modal component (IMF).Then, according to the correlation coefficient and kurtosis criterion, the IMF components that met the criteria were selected, and the distribution entropy(DE) of these components was extracted as the feature. Finally, the fault diagnosis model of VNWOA-LSSVM was used to diagnose the vibration signal of the ammunition feeding system under three different working conditions, and the fault recognition rates of LSSVM, PSO-LSSVM, GA-LSSVM and WOA-LSSVM were compared. The experimental results show that the recognition rate of VNWOA-LSSVM is the highest, and it is up to 94.03%.