The work aims to propose a convolutional neural network classification and detection method which can meet the factory requirements, for the purpose of solving the problem of high manual dependence and low efficiency in traditional non-woven fabric defect classification and detection. Firstly, a total of 70 000 images of five types of non-woven fabrics including dirty spot, fold, fracture, missing yarn and no defect were established. Secondly, a neural network of the convolutional layer and pooling layer with a number of different neurons was constructed. Then, the back propagation algorithm was used to update weights layer by layer, and the loss function was minimized by gradient descent method. Finally, Softmax classifier was used to realize defect classification and detection of non-woven fabric. A 12-layer convolutional neural network was constructed and tested with 20 000 samples. The accuracy of samples without defects could reach 100%, the classification accuracy of defect samples was above 95%, and the detection time was within 35 ms. The proposed method can meet the requirement of real-time classification and detection of non-woven fabric defects in industrial production lines.