More than 26 million tons of waste textiles are produced in China every year, and with the development of the economy, it is showing a trend of sharp increase year by year, but its recycling rate is less than 10%. The diversity of waste textile components and the complexity of the structure are the biggest obstacles that affect their accurate classification, rapid recycling and high value-added reuse. Manual identification and sorting are time-consuming, labor-intensive and inaccurate, while near infrared spectroscopy analysis technology can quickly, non-destructively and efficiently identify and sort waste textiles. According to the optimized test conditions explored in the early study, the online raw near-infrared spectra for polyester, cotton, wool, nylon, silk, viscose, acrylic, polyester/wool, polyester/cotton, polyester/nylon, silk/cotton blended fabrics and “special type”, a total of 1060 waste textiles samples of 12 types were collected with the self-developed “online near-infrared high-efficiency identification and automatic sorting device for the fiber products”. Based on the online original NIR spectra of the collected samples, the convolutional neural network method was used to conduct network training according to the input sample spectral data and corresponding classification labels, and an online NIR qualitative identification model for 12 types of waste textiles was established. The two-dimensional model was better compared to the one-dimensional and two-dimensional convolutional neural network models-. The one-dimensional array of 901~2 500 nm was normalized and converted into a two-dimensional grayscale image of 40×40 pixels and then alternately performed multiple convolutions and pooling for spectral feature extraction, compression and data dimensionality reduction. The class probability value of each kind of waste textile sample was obtained through model calculation, and its maximum value was taken as the final classification of this kind of fabric. In the model’s training process, epoch number was set to 500, batch size was set to 32 and the learning rate was 0.001. After training, the preset 12 types of fabric labels were output, and the internal training accuracy of the model can reach 96.2%. To verify the applicability of the model, the prediction test of the model was carried out with 232 fabric samples that did not participate in the modeling, and the recognition accuracy was 96.6%. After the model was imported into the “Textile Online Master Control Program”, the 12 types of fabrics included in the modeling samples could be identified and automatically sorted. The identification and sorting time of each sample is less than 2 s. The establishment of the model and the application of the device provide a new sorting technology and equipment for the recycling of waste textiles in China.