Abstract:

Objective Textiles with similar contents, such as cashmere and wool, are known to be tough to categorize. When their amounts in waste textiles are close, conventional detection models have difficulties to distinguish them. This often results in low accuracy in identifying cashmere and wool. This study aims to build a hybrid CNN-LSTM model combined with local feature extraction with temporal feature extraction, so as to detect cashmere and wool content accurately with near-infrared spectroscopy(NIR). This method is expected to extract the variations of absorption peaks and the relationship between absorption peaks, and then obtain the complex spectral variations.

Method Cashmere and wool samples with similar colour radii are collected for preparing mixed samples of cashmere and wool in various ratios using the potassium bromide pressing method. Spectra using a near-infrared spectrometer are obtained to create a database, and the NIR spectral data are nomalized with the Z-Score algorithm, which helps remove differences in magnitude. Gaussian noise data enhancement follows to boost data diversity and the dataset is split using 10-fold cross-validation which is used as input to a hybrid CNN-LSTM model. MLP regression prediction is performed through the convolutional layer and gated sequence.

Result In the experiments, the improved CNN-LSTM-MLP model predicts cashmere and wool contents well. The model efficiently pulls out key spatial features from near-infrared spectral data which cuts down on processing time and boosts feature extraction efficiency. It uses depth-separable convolution to first identify peaks from chemical bond vibrations in the near-infrared spectra, and combines these features using stationary point convolution. The LSTM (long short-term memory) layer effectively captures the time-based relationships in spectral data to enhance the model's sensitivity to changes in spectral features. It also preserves contextual information across bands using forgetting gates and memory units. As a result, the model can focus on intensity changes in individual absorption peaks. It also learns patterns among several key bands, such as the connections between absorption peaks from 2 028 to 2 470 nm. By modelling the spectral sequence, LSTM reveals how different absorption peaks relate over time. It boosts the weight of continuous absorption peak bands, and allows it to better spot changes in mixing ratios and peak intensity, even with minor differences. The results show that the improved model predicts more accurately than the conventional methods. In the 10-fold cross-validation, most R² values on the test set are above 0.95 and the MSE is below 0.01, indicating that the improved CNN-LSTM-MLP model is accurate and stable in analysing near-infrared spectroscopic data. It also offers strong support for quick and accurate detection of cashmere and wool content.

Conclusion The proposed new prediction model for cashmere and wool uses an enhanced CNN-LSTM-MLP architecture in combination with deep separable convolution (DSC) with conventional convolutional neural networks (CNN). This setup effectively extracts spatial features from near-infrared spectral data, and also models temporal features with LSTM layers. Quantitative predictions are achieved using a multilayer perceptron (MLP), and deep separable convolution cuts down the model's computational complexity, keeping rich channel information in the high-dimensional spectral data. The proposed model greatly improves prediction accuracy, and the efficient algorithm results in the need of little computing power to predict cashmere and wool contents. The model performs well in analysing cashmere and wool and it is practically useful for spectral analysis in real production.

Key words: deep learning, near-infrared spectroscopy, convolutional neural network, long short-term memory, quantitative detection, cashmere, wool