深度学习与可解释人工智能驱动可见-近红外光谱无损检测牛肉硫代巴比妥酸含量

Non-destructive detection of the content of thiobarbituric acid in beef by visible-near infrared spectroscopy driven by deep learning and explainable artificial intelligence

中文关键词:  硫代巴比妥酸反应物  深度学习  可解释人工智能  长短期记忆网络  无损检测  可见近红外光谱

英文关键词:thiobarbituric acid reactive substances  deep learning  explainable artificial intelligence  long short-term memory  non-destructive detection  visible-near infrared spectroscopy

中文摘要:

      目的 利用深度学习与可解释人工智能驱动的可见-近红外(visible near infrared, Vis-NIR)光谱技术实现牛肉中硫代巴比妥酸反应物(thiobarbituric acid reactive substances, TBARS)快速无损检测。方法 通过对420份牛肉样品进行Vis-NIR光谱采集和TBARS测定, 并比较不同的预处理方法、特征选择方法和建模方法, 实现牛肉TBARS的无损快速检测。结果 标准正态变换-逐步投影算法-长短期记忆网络模型表现最优[测试集决定系数(R-square of prediction set, R2P)=0.8378, 测试均方根误差(root mean square error of prediction set, RMSEP)=0.0679 mg/kg, 相对预测偏差(residual prediction deviation, RPD)=2.4833], 优于偏最小二乘回归(partial least squares regression, PLSR)和极端梯度提升(extreme gradient boosting, XGBoost)。进一步通过沙普利可加性解释(Shapley additive explanations, SHAP)可解释性分析, 揭示了1080 nm、956 nm等关键波长对TBARS预测的贡献机制。结论 本研究提出的方法为牛肉TBARS的快速、无损、精准且可解释检测提供了创新策略。

英文摘要:

      Objective To achieve rapid and non-destructive detection of thiobarbituric acid reactive substances (TBARS) content in beef by leveraging deep learning and explainable artificial intelligence-driven visible-near infrared (Vis-NIR) spectroscopy. Methods Vis-NIR spectra were collected from 420 beef samples, and the TBARS values were determined experimentally. Different spectral pretreatment methods, feature selection methods and modeling algorithms were compared to realize the non-destructive and rapid detection of TBARS in beef. Results It demonstrated that the standard normal variate preprocessing-successive projections algorithm for feature selection-long short-term memory network model achieved the best performance [R-square of prediction set (R2P)=0.8378, root mean square error of prediction set (RMSEP)=0.0679 mg/kg, residual prediction deviation (RPD)=2.4833], outperforming partial least squares regression (PLSR) and extreme gradient boosting (XGBoost). Furthermore, through the interpretability analysis using Shapley additive explanations (SHAP) revealed the contribution mechanisms of critical wavelengths, such as 1080 nm and 956 nm, to TBARS prediction. Conclusion The proposed method offers an innovative strategy for rapid, non-destructive, accurate and explainable detection of TBARS in beef.