Rapid identification of silk products by attenuated total reflection–mid-infrared spectroscopy and extreme learning machines

Silk is a high-quality fiber of natural protein and its amino acid composition is close to that of the human skin. The texture of pure silk is elegant, soft and comfortable to wear, with good breathability and hygroscopicity, as well as excel dirt absorption and sterilization. It is essential to identify natural pure silk products and various non-silk and adulterated silk products with other fibers. The work focuses on exploring the feasibility of combining attenuated total reflection–mid-infrared (ATR-MIR) spectroscopy feature selection, extreme learning machines (ELM), and ensemble and for rapid identification of silk products. A dataset consisting of 70 silk samples and 83 other fiber samples was prepared. The dataset is split evenly into training and test sets for building and validating predictive models. Simulated annealing (SA) was used for pre feature selection. A total four kinds of ELM models including full-spectrum, subspace and ensemble models were developed and compared. The optimized local model and the ensemble model achieve the best and most stable performance, i.e., with 100% accuracy. It seems that SA, ELM combined with ATR-MIR technique may be a potential alternative to traditional analytical methods of silk identification, which can avoid time-consuming and labor-intensive chemical analysis.