Rapid multi-indicator detection of high final temperature biochar based on near-infrared spectroscopy and convolutional neural network

文献类型： 外文期刊

第一作者： Pan, Yuxuan

作者： Pan, Yuxuan;Jiang, Hanlu;Li, Fupeng;Wang, Feiyun;Lv, Chengxu;Jia, Xiaofeng;Shang, Xuechuang;Zhou, Haiyan;Wang, Yixian

作者机构：

关键词： Biochar; NIRS; CNN; Data augmentation

期刊名称：INFRARED PHYSICS & TECHNOLOGY （ 影响因子：3.4； 五年影响因子：3.4 ）

ISSN： 1350-4495

年卷期： 2025 年 149 卷

页码：

收录情况： SCI

摘要： The fixed carbon, volatile matter, and ash content are pivotal indicators for determining the heating value of biochar and evaluating its quality. Rapid detection of these indicators holds significant importance in the biochar industry. However, traditional regression models based on near-infrared spectra struggle to achieve precise quantification due to the weak response of biochar samples produced under high final temperature conditions in the near-infrared band. To address this issue, the laboratory production of biochar was conducted, simulating process parameters such as feedstock size, heating temperature and insulation time in the carbonization production line. Through data augmentation techniques, a sufficient large spectral dataset was obtained for model training. A one-dimensional convolutional neural network (1D-CNN) architecture was proposed, encapsulating convolutional layers, batch normalization (BN) layers, activation functions and pooling layers into convolutional blocks. Hierarchical fast perception and nonlinear fitting of spectral features were achieved through multiple convolutional blocks, decreasing convolutional kernels and multiple fully connected layers. The augmented dataset was subjected to noise and error correction using three preprocessing methods: Savitzky-Golay (SG), Multiplicative Scatter Correction (MSC), and Standard Normal Variate (SNV). The results indicated that the predictive performance of the 1D-CNN model optimized with SG preprocessing was the best. For fixed carbon, volatile matter, and ash content, the models achieved R2 greater than 0.94, RPD greater than 4 and RMSE (%) for 1.70, 1.37 and 2.20, respectively. The RMSE (%) for the three components were reduced by 61.63 %, 71.40 % and 46.60 % compared to the optimal results of the PLS and SVM models. The predictive performance of the models was excellent and suitable for precise prediction. The 1D-CNN models developed in this study achieved accurate predictions of fixed carbon, volatile matter and ash content in biochar, providing a reliable reference for predicting the calorific value and monitoring the quality of high final temperature biochar.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序