In the detection and maintenance of winder, a large number of text maintenance records containing rich fault characteristics will be saved, but at present, the fault diagnosis of textile winder still mainly relies on manual investigation, failing to realize intelligent diagnosis, and the text maintenance records can not be fully utilized. Therefore, aiming at the problem of textile winder fault text classification, a fault classification method for textile winder named ERNIE-TEXTCNN-LightGBM (ETL) is proposed based on short text and sparse feature data. Through mask language model training of ERNIE model, The contextual semantic data and entity knowledge data are obtained respectively, and the semantic characteristics of the fault article are fully grasped. Then TEXTCNN model is used to extract local features from convolution layer and pooling layer, and learn the semantic association between adjacent words to overcome the loss of local features. Finally, the LightGBM model is integrated to reduce the parameters of the model, to realize the fault classification of the maintenance text, to assist the fault maintenance of the winder, and to improve the production efficiency of the factory. In order to verify the effectiveness of the algorithm, ETL is compared with other common text classification models, and the experiments show that the proposed method is effective and superior in realizing the maintenance text fault diagnosis of winder.