摘要
Abstract
In view of the serious overlapping of infrared spectral data features in the qualitative analysis of the composition of waste textiles and the insufficient classification accuracy of traditional models,proposes a deep learning model based on the SDP(symmetrized dot patterns)and SE(squeeze excitation)attention mechanisms.The infrared spectral data of 20 types of single and mixed component textile fibers such as cellulose,wool,and polyester were collected(400 training sets and 200 test sets for each type).The spectral data was mapped to po-lar coordinate images using SDP conversion to effectively enhance feature separability,and the optimal parame-ters were determined through experiments(gain angle g=30°,interval factor b=29).On this basis,the SE-DCNN model was constructed by combining the hole convolution to expand the receptive field and introducing the SE attention mechanism to optimize the feature channel weight allocation.Ablation experiments showed that after fusing the SE module with the hole convolution,the model's determination coefficient(R²)on the test set in-creased to 0.992,and the mean absolute error(MAE)and root mean square error(RMSE)were reduced to 0.64 and 2.24 respectively.Compared with traditional methods(such as SVM,PCA+KNN)and 1D-CNN,SE-DCNN performs best in the mixed component prediction task(test set R²=0.947,MAE=1.05).关键词
废旧纺织物/红外光谱/对称点模式/注意力机制/空洞卷积Key words
waste textiles/infrared spectrum/symmetrized dot patterns/attention mechanism/void convolution分类
轻工纺织
