Machine learning combined with sensory evaluation and multi-sensor technology to evaluate the overall quality of commercial soybean paste in China

• Sensory evaluation combined with three sensors was used as the evaluation criteria for commercial soybean paste. • Using modern machine learning technology to predict the quality of soybean paste. • SVR model had better prediction performance than other models. Soybean paste has been a prominent condiment in East Asia for millennia. Nonetheless, the current methodologies for comprehensively assessing the quality of commercially available soybean paste through sensory evaluation or traditional instruments present significant challenges. In this study, contemporary detection techniques and machine learning methodologies were employed to quantitatively characterize and evaluate the overall quality of soybean paste. Sensory evaluations were conducted on 33 varieties of commercial soybean paste using three types of sensors: a colorimeter, an electronic nose (E-nose), and an electronic tongue (E-tongue) for detection purposes. Subsequently, machine learning models, including Support Vector Regression (SVR), Random Forest (RF), extreme Gradient Boosting (XGBoost), Bayesian Ridge Regression (BRR), Ridge Regression (RR), k-nearest neighbors (KNN), and Artificial Neural Network (ANN) were developed based on the sensory evaluation data to characterize and assess the overall quality of the soybean paste. The findings from both sensory evaluations and sensor detection indicated notable differences between the various soybean pastes. Soybean pastes can be distinguished using three sensors. The quantitative characterization model informed by the sensor data revealed that the SVR model exhibited the highest R² value of 0.9998 for the training set and 0.997 for the test set, which was close to the ideal value of 1. Additionally, the root mean square error (RMSE) for the test set was the lowest at 0.5359. These results suggest that SVR demonstrates superior performance in cross-validation and testing, aligning closely with human sensory perceptions, thereby establishing it as the most effective predictive model. This study underscores the potential of integrating sensor data with modern machine learning techniques to supplement traditional sensory evaluations for comprehensive characterization and assessment of soybean paste quality. The outcomes of this study offer significant insights and guidance for the advancement of the soybean paste industry and the enhancement of soybean paste quality.