The authenticity of documents is an important work in the current stage of litigation review. In judicial cases, erasablepens are often used to forge documents, contracts and other criminal acts. The identification of ink composition and handwriting modification is the key research in the field of document inspection. Special thermal color pigment is the main component of erasable ink; its color principle is that temperature change will produce the disappearance and recurrence of handwriting, color fades above 65℃, and color recurrence below -18 ℃. The identification of its species can identify the authenticity of the case evidence and provide support for the litigation process of the case. The ultra-high spectral resolution of hyperspectrum has good feature selectivity for polymer materials, which can effectively collect data for common ink components. In this experiment, a total of 45 erasable pen ink samples from 22 brands were collected, which can be divided into four types: tungsten carbide pen beads, bullet pen beads, full needle tube and half needle tube, and the hyperspectral information of 450~950 nm band was collected uniformly. As for the redundancy of background noise in spectral data, the principal component analysis (PCA) was used to reduce the dimensionality of the data and extract the feature variables. Based on the dimensionality reduction data, different Hilbert transform (HT) types were used for signal filtering, and effective signals were further selected to improve the modeling effect. Two artificial neural network models, Multilayer Perceptron (MLP) and radial basis function neural network (RBFNN), were selected for sample recognition. The feature variable class modeling accuracy based on 23-dimensional principal component extraction is 81% and 84%, respectively. After the Hilbert high-pass filtering processing, the classification accuracy can be increased to 88. 9% and 92%, effectively improving recognition accuracy. In order to further distinguish the types of different samples, Fisher discriminant analysis method was selected for modeling. The identification accuracy of the original data of each sample in the FDA model was 44%, and the FDA modeling accuracy of the optimal PCA-HT treatment was 93. 3%, which could distinguish different types of erasable ink. The results show that PCA can reduce the dimension based on retaining the effective spectral information, improving the model accuracy and shortening the running time. Compared with the original spectral data, the modeling effect is good, and the spectral data after the Hilbert transform can further improve the effective spectral information to further improve the modeling accuracy. This experiment determined the optimal PCA-HT-FDA model and the best erasable ink hyperspectral identification model, which can provide a certain reference for forensic experts.