The present paper used synthesized data from the experiment samples to replace partial basic experiments, and increased the training samples amount from 14 to 27. In principal component analysis (PCA), the dimensionality of multivariate data was reduced to n principal components and almost all data information was kept. The PCA reduced the network’s input nodes from 60 to 3 to simplify the neural network’s structure. Finally, back-propagation neural network was used to train and predict these samples. It had 27 training samples, the input layer had three nodes, the hidden layer had two nodes, and the output layer had two nodes. Its excitation function is variable learning rate method. The results show that the coefficient of recovery can reach 89.6-109.0. It has reached the expected purpose.