At present,the wood grades of the national musical instrument Chinese zither panel mainly rely on the personal experience of the musical instrument technician.This method relies on experienced technicians and is susceptible to subjective judgment. In view of this situation, we use the Paulownia wood used to make the Chinese zither panel as an experimental sample. We propose a method of using near-infrared spectroscopy and the improved BP neural network to rapidly identify different grades of Chinese zither panels .Because Near-infrared spectroscopy can characterize a number of material structure and composition information, with the low cost of measuring instruments and many measuring accessories,we conduct an experimental analysis of the near-infrared spectral data of Paulownia panel. In the experiment, spectral denoising is first performed to eliminate system errors and improve spectral resolution, regarding the root mean square error and the square sum of signals as the evaluation criterions of various pretreatment methods.Therefore ,the first derivative is selected as the final pretreatment method, and 15 is the appropriate filter denoising window size.The principal component analysis is then used to compress the data and the Mahalanobis distance is used to eliminate the modeling set’s abnormal samples to create a more representative modeling set. Then, an unsupervised clustering is used to analyze the Paulownia panel grades , which proves the feasibility of grade classification.Since H2O has a large absorption in the near-infrared spectral region, according to experimental spectral analysis results,we do not consider the fundamental frequency vibration band (5 396 to 4 978 cm-1) and the first overtone vibration band (6 800 to 7 000 cm-1),but consider only the remaining near-infrared spectral band. Different spectral bands are combined, and seven bands are used as input to the neural network to carry out the panel grade recognition.We also improve the traditional BP neural network model. The learning rate of BP neural network is set by an adaptive optimization strategy to speed up the traditional neural network’s training rate.At the same time, the cross entropy function is used as the cost function to speed up the updating of the weight.The Relu function is selected as the transfer function between the input layer and the hidden layer, which improves the training speed of the model and effectively prevents over-fitting.The Softmax function is chosen as the transfer function of the last layer to reduce complex calculations. By this way, the final BP neural network is constructed.The amount of spectral information that can be extracted by different principal component variables is different.We adjust the input of the BP neural network model by increasing the number of principal components and adjusting the spectral band interval.When the number of principal components is 11 and the spectral intervals are 10 000 to 7 000 cm-1 and 4 976 to 4 000 cm-1, the unknown sample’s recognition rate reaches 99.7% , and the selected spectral range covers all the characteristis of C—H bond and other bond information.The results show that near-infrared spectroscopy combined with BP neural network can effectively identify different grades of Paulownia panel, thereby reducing manual detection errors, shortening the processing time, and better meeting the needs of the instrument market.