基于极限学习机自编码算法的近红外光谱模型传递的研究

Study of near infrared spectral model transfer based on an extreme learning machine auto-encoder algorithm

中文关键词:  近红外光谱技术  极限学习机  模型传递  酸值  过氧化值

英文关键词:near infrared spectroscopy  extreme learning machine  model transfer  acid value  peroxide value

中文摘要:

      目的 针对食用油的酸值和过氧化值进行分析, 探究极限学习机自编码算法(transfer via extreme learning machine auto-encoder algorithm, TEAM)在近红外光谱上的模型传递。方法 使用MATRIX-F和VERTEX-70两种红外光谱仪采集食用油近红外光谱数据, 利用多元散射校正方法对光谱数据进行预处理。然后基于TEAM建立传递模型, 并与直接标准化、分段直接标准化和斜率偏差校正算法的建模效果进行了对比。结果 经TEAM算法模型传递后提高了模型的精确度, 食用油酸值模型中, 决定系数(R2)从?1.3984升高到0.8553, 预测集均方根误差从0.6130 mg/g降低到0.2578 mg/g, 食用油过氧化值模型中, R2从0.6170升高到0.8987, 预测集均方根误差从16.1530 mmol/kg降低到10.4150 mmol/kg。结论 极限学习机自编码算法使从机数据更好适应主机模型, 提高了模型的稳定性和准确性。

英文摘要:

      Objective To investigate the model transfer of the transfer via extreme learning machine auto-encoder algorithm (TEAM) on near infrared spectrum by analyzing the acid value and peroxide value of edible oil. Methods The near infrared spectral data of edible oils were collected by MATRIX-F and VERTEX-70 infrared spectrometers, and the spectral data were preprocessed by multiple scattering correction method. Then, a transfer model was established based on TEAM, and the modeling effects were compared with those of direct standardization, piecewise direct standardization and slope deviation correction algorithms. Results After TEAM algorithm model transmission, the accuracy of the model was improved. In the edible oil acid value model, coefficient of determination (R2) increased from ?1.3984 to 0.8553, the root mean square error of prediction set decreased from 0.613 mg/g to 0.2578 mg/g, and in the edible oil peroxide value model, R2 increased from 0.6170 to 0.8987. The root mean square error of prediction set decreased from 16.153 mmol/kg to 10.4150 mmol/kg. Conclusion The extreme learning machine self-coding algorithm makes the slave data better adapted to the host model and improves the stability and accuracy of the model.