In order to predict the quality of Japanese fermented soy sauces, this study focuses on selecting relevant variables for developing a flexible and objective model. There were 74 parameters with the potential to influence the overall acceptability of soy sauce being measured and regarded as potential variables for predicting the sensory scores of soy sauce samples. The variable selection approach was inspired by Compressed Sensing (CS) theory and has been used for the first time on the calibration set (soy sauce samples were collected directly from the Akita Prefectural Soy Sauce Competitions in 2016 and 2017) to evaluate the contribution of each predictive variable to the sensory score. Consequently, 30 predictive variables which make a great contribution to the quality for predicting soy sauce were successfully selected by CS-based method. The selected variables covered the important variables of sensory evaluation such as color, taste, and fragrance. Subsequently, the model for predicting soy sauce quality was established using partial least squares regression, based on the selected variables. The validity of the model was evaluated using soy sauce samples produced in 2018 leading to values of r² and RMSEP for the validation samples of 0.80 and 11.47, respectively. Therefore, the model was considered to be suitable for predicting the sensory quality of soy sauce. The results also confirmed that the CS-based method provided a new approach to selecting variables of practical importance for developing a predictive model.

为了预测日本发酵酱油的质量，本研究着重于选择相关变量以建立灵活而客观的模型。有74个参数可能会影响被测量的酱油的总体可接受性，并被视为预测酱油样品感官评分的潜在变量。变量选择方法是受压缩感知（CS）理论启发而来的，并且首次用于校准集（酱油样品直接从2016年和2017年的秋田县酱油大赛中收集）来评估每个变量的贡献感官评分的预测变量。因此，通过基于CS的方法成功地选择了30个对酱油质量预测有重要贡献的预测变量。选择的变量涵盖了感官评估的重要变量，例如颜色，味道和香味。随后，基于所选变量，使用偏最小二乘回归建立了预测酱油质量的模型。使用2018年生产的酱油样品评估模型的有效性，得出的价值为验证样本的r ²和RMSEP分别为0.80和11.47。因此，该模型被认为适合预测酱油的感官品质。结果还证实，基于CS的方法为选择对开发预测模型具有实际重要性的变量提供了一种新方法。