A machine learning method for juice human sensory hedonic prediction using electronic sensory features

Abstract

This study proposed a method that combines fused electronic sensory analysis technology with artificial neural network to predict the human sensory hedonic of fruit juice. Quantitative descriptive analysis (QDA) and the scoring test method were utilized for human sensory evaluation. The first step involved modeling the fused e-sensory features with human sensory attributes, followed by establishing a fitting model of human sensory attributes and acceptance. The R² and RMSE values obtained were 0.77 and 0.42 (QDA method), and 0.63 and 0.63 (scoring test method). Finally, the relationship between the fusion e-sensory features and the human sensory hedonic was established. Model-1 achieved an R² of 0.95 and an RMSE of 0.04, while model-2 achieved an R² value of 0.88 and an RMSE value of 0.21. This study demonstrates the potential of fusing e-sensory technologies to replace human senses, which may lead to the development of devices with simultaneous multiple senses.

Graphical abstract

Highlights

• •Fusion e-sensory features with human sensory attribute modeling.
• •To build a model for fitting human sensory attribute scores to sensory hedonic.
• •ANN was used to predict human sensory hedonic based on fusion e-sensory features.
• •The potential of fused e-sensory technology to replace human senses.

1. Introduction

Consumers in the food market are increasingly pursuing personalization and diversification, which brings new challenges to fruit juice producer. Understanding consumer hedonic and making products that consumers like equals occupying the corresponding market share (Bi et al., 2022). Accurate prediction of consumer hedonic is vital for juice producers to attract new consumers and maintain their brand loyalty. Therefore, the development of a new model to rapidly assess food sensory quality or predict consumer hedonic is a useful goal.

Traditional juice sensory quality evaluation methods often rely on human sensory evaluation, which is an important means of product development and food research. The methods of food human sensory evaluation can be divided into difference test, scale and category test, descriptive analysis test, and emotion test according to their purpose. Among these, the scale category test method is one of the commonly used sensory evaluation methods, which includes the ranking test, category test, evaluation method, and fuzzy mathematical method. On the other hand, description analysis method has been widely used in food industry for decades since revealing the importance of nutrition and a balanced diet during World War II. It includes two steps: training sensory group members panelists to identify and quantify product characteristics (Gabrieli et al., 2022). However, traditional sensory methods have some limitations such as being subjective, costly, and having low reproducibility. They are time-consuming and expensive, especially when large numbers of panelists are required (Ross, 2021). Furthermore, sensory evaluation tends to vary with the physical condition of the evaluator, emotional changes, and influences from the external environment, making it a low objectivity method. Thus, there is a demand for a quick, unbiased, and inexpensive alternative to sensory evaluation methods.

Multi-sensor technologies have attracted research interest as alternative techniques to obtain accurate and comprehensive food sensory analysis results. Electronic nose (e-nose), electronic tongue (e-tongue) and electronic eye (e-eye) equipped with gas sensors, liquid sensors, and color sensors respectively, are widely used in the food field. These technologies mainly mimic the human olfactory, taste and visual systems, providing rapid detection and global information of the sample rather than specific ingredients. These sensors allow the detection of one or more compounds present in complex samples (Buratti et al., 2018). This approach is low-cost, time-saving, simple operation, and fast analysis (Minxu et al., 2019). Furthermore, the use of electronic sensor fusion techniques can provide more accurate knowledge of products than the use of a single sensor (Wang and Zhenbo, 2015; Zhi et al., 2017). Understanding consumer hedonic for sensory qualities of fruit juice is essential for manufacturers to process their products effectively. However, the electronic sensory (e-sensory) evaluation results can not directly reflect consumer hedonic but only indicate the product quality indirectly.

A data-fusion method for multi-sensors heterogeneous data was utilized to predict consumer hedonic as the main object in this paper. Several studies have focused on exploring the correlation between sensory attributes of food products and consumer hedonic. Zhi et al. (2016) used partial least squares regression to establish the relationship between product characteristics and overall preferences, thereby determining consumer preferences for flavored milk in different regions. Similarly, Yu et al. (2017) developed a partial least squares-artificial neural network hybrid model based on human sensory evaluation to predict the consumer acceptance scores for ready-to-drink green tea beverages. Castada et al. (2019) used principal component analysis to find a positive relationship between five sensory attributes and consumer preferences for Swiss cheese. In another study, Bi et al. (2020) used human sensory attributes to predict yogurt preferences using support vector machine approach. Recently, G.A.Kon et al. (Koné et al., 2022) used partial least squares regression to determine the relationship between chicken samples, their sensory attributes and consumer preferences. These studies applied a series of traditional statistical methods to try to predict consumer preferences based on sensory attributes. However, when dealing with large amounts of high-dimensional data, these conventional methods may not provide sufficient accuracy, scalability, and generalizability (Jiménez-Carvelo et al., 2019). Moreover, the error in the prediction results may be due to potential problems, such as individual differences in evaluators or poor data quality in sensory experiments. Machine learning can solve the above problems and can be used to map the multi-dimensional output of the sensor array to sensory perception, regardless of the selectivity of a single sensor (Ishihara et al., 2005). Artificial neural network (ANN) is an important algorithm in machine learning, based on biological neural networks (Kassanuk and Phasinam, 2021). ANN possess strong self-learning abilities and adaptability, which can effectively handle any linear and nonlinear functions in data and complex relationships among various parameters. As a robust task prediction tool, ANN can learn from a vast amount of data and even be applied to unlearned data. Therefore, ANN model will be applied to establish a link between multiple electronic sensory data and human sensory evaluation acceptance value in this paper.

In order to solve the current problems of time-consuming, labor-intensive, and limited scalability in food evaluation, this study aimed to objectively evaluate food quality from a multi-dimensional perspective and quickly predict human hedonic. To achieve this, the focus of this study was to combine the artificial neural network algorithm with fusion e-sensory technology of fruit juice with human senses (Fig. 1). The specific objectives of this study were: (1) to conduct e-sensory and two different human sensory assessment methods of market-purchased and self-made fruit juice, respectively; (2) to utilize artificial neural network technology to establish a model of fusion e-sensory features and human sensory attribute scores of fruit juice, and then to fit the human sensory attribute scores and human sensory hedonic; and (3) to establish the link between e-sensory and human sensory hedonic association for the purpose of predicting human sensory hedonic value using e-sensory features data of fruit juice.

Fig. 1.

Schematic diagram of the modeling process.

2. Material and methods

2.1. Sample preparation

A total of 287 samples were used in this study, consisting of 26 samples purchased from Tmall Supermarket comprising 12 different brands and 15 different flavors, and 261 self-made samples made from 9 different processes and 29 different formulas. The sea buckthorn fruit was sourced from Lvliang City, Shanxi Province, while the passion fruit was purchased from Dehong Autonomous Prefecture, Yunnan Province. Sucrose, citric acid, xanthan gum, and tea polyphenols were obtained from Shanghai Xinminrong Food Co., Ltd., Shengfa Biotechnology Co., Ltd., and Henan Wanbang Industrial Co., Ltd., Hangzhou Pulimeidi Biotechnology Co., Ltd., respectively. The self-made fruit juices were prepared using varying proportions of components, including sea buckthorn (10%–35%), passion fruit (10%–35%), sucrose (6%–14%), citric acid (0.09%–0.17%), and xanthan gum (0.14%–0.22%). A total of 30 kinds of sea buckthorn passion fruit juice beverages (“fruit juice”) were prepared using different proportions based on the formula presented in Supplementary Table S1 (Ren et al., 2023). Nine main process parameters were adopted to the preparation of fruit juice, as outlined in Supplementary Table S2.

2.2. Characterization of fruit juice by multiple electronic sensory technologies

The e-sensory analysis of fruit juice was performed by e-tongue, e-nose, colorimeter, and rheometer. Each sample of fruit juice was measured three times to ensure accuracy of results. The specific e-tongue device was the SA402B, which consists of a sensor array, signal acquisition system, and pattern recognition system. The sensor array is comprised of 6 lipid membrane sensors: AAE, CA0, CT0, C00, AE1, and GL1, along with 3 reference electrodes (Ag/AgCl). These sensors are designed to detect umami, salty, sour, bitter, astringent, and sweet flavors, respectively. A variety of fruit juices were placed in the e-tongue special tasting cup up to the scale line (about 35 mL). Then the tasting cup was inserted into the e-tongue instrument, and different procedures were employed to measure sweetness, sourness, bitterness, astringency, umami, and saltiness. Sweetness was measured using the GL1_test procedure in five parallel measurements, while Sample_Measurement procedure was used for the remaining tastes, which were measured four times in parallel. The resulting values of the e-tongue were converted into taste values utilizing the taste analysis application software. A reference solution was utilized as a control for numerical correction. The Foodstuff Evaluation + GL1. ece was used for taste value conversion in order to obtain data related to the e-tongue analysis of fruit juice. The formula for calculating the taste value was as follows.

$$R=V_s-V_r$$ (1)

$$CPA={V_r}^{\prime }-V_r$$ (2)

where R is the taste signal value. Change of membrane potential caused by adsorption (CPA) is an aftertaste signal value. Vr and Vs are the membrane potential values of the activated sensor sequentially immersed in the reference solution and the sample solution, respectively. Vr' is the re-measured membrane potential value of the sensor immersed in a new reference solution.

The e-nose device utilized in this study is the PEN3 (Airsense Analytics, GMBH, Schwerin, Germany), which consists of a dynamic headspace sampling module, gas detection units, and a data acquisition module. The PEN3 e-nose employs ten different metal oxide sensors as part of its sensor array, named W1C, W5S, W3C, W6S, W5C, W1S, W1W, W2S, W2W, and W3S. The general working procedures of e-nose include: initialization of the sensor, sample measurement and data acquisition, result output, and sensor cleaning. The sample was determined at room temperature 25 °C. The e-nose was preheated and clean air was passed inside before use. To ensure that the e-nose sensor could detect the odor of the sample, the prepared sample was sealed in advance, and the probe was inserted quickly during measurement. The time interval between each sample was 1 s, the cleaning time was 60 s, the zeroing time was 10 s, the determination time for each sample was 60 s, the carrier gas flow rate was 300 mL min⁻¹, and the sample injection flow rate was 300 mL min⁻¹.

The color of fruit juice was measured by colorimeter. Distilled water was used as a blank sample for correction. The samples color was measured by D65 standard light source and 10° plane measurement at room temperature of 25 °C. The sample color was expressed according to the CIE color system, also known as the L* a* b* color system. After measurement, four characteristics were obtained to characterize the color of fruit juice, namely L*, a*, b*, △E*ab, where L* represents brightness, a* represents red and green color, b* represents yellow and blue, △E*ab is the total color difference of fruit juice color, calculated by the following formula.

$$△\mathrm{E}*\mathrm{a}\mathrm{b}=\sqrt{{\left(△L^{*}\right)}^2+{\left(△a^{*}\right)}^2+{\left(△b^{*}\right)}^2}$$ (3)

The rheological properties of the fruit juice were determined using a Harker RS600 rheometer (Thermo Scientific, US), Mechato and Siche (2020). To measure the sample, it was added to a stainless-steel parallel plate with a diameter of 40 mm. The plate spacing was set to 1 mm, the shear rate was 0.01–300 s⁻¹, and the shear time was 120 s. The apparent viscosity of the sample and its variation of shear stress with shear rate were measured at room temperature of 25 °C. The shear stress (r) is analyzed according to the following formula.

$$\sigma =K{y}^n$$ (4)

where σ is shear stress (Pa). K is Power Law consistency index (Pa.sⁿ). y is shear rate (s⁻¹). n is flow behavior index (non-dimensional).

2.3. Quantitative descriptive analysis and the scoring test for fruit juice by trained panel

The sensory evaluation in this study was performed according to the National Standard of China (GB/T 10220-2012, 2012). There are two main steps, selection and training of evaluation panelists, and sensory evaluation. The selection of panelists referred to GB/T 16291.1-2012 “General guidelines for the selection, training, and management of Evaluation Personnel, Part 1” (GB/T 16291.1-2012, 2012). A sensory analysis team composed of twenty graduate students from Northeastern Agricultural University with an average age of 24–26 years was established through selection, training, and assessment. There were two groups of 10 people each, one group using quantitative descriptive analysis (QDA) and the other group using the rating method to assess the human sensory attributes and hedonic value of the juices. Each evaluator had reached the level of the preferred evaluator in the standard. The screening and training process of evaluator was carried out in a standard sensory analysis laboratory using standard methods.

The human sensory attributes and hedonic scores of fruit juice was performed using the linear scale method in quantitative descriptive analysis (Sánchez et al., 2020) and the scoring test method in scale and category tests. Statistical methods were employed to analyze the resulting data. The quantitative descriptive analysis method involved marking a position on a 15 cm line that the panelist believes can represent the sensory properties of the fruit juice. The left end of the line represented 'none' or '0′, while the right end represented 'maximum' or 'strongest'. The sensory evaluation table of fruit juice had six lines for each sample with the score corresponding to each line segment was 10 points. A straightedge was used to measure the length of the underlined part, and the sensory evaluation score was obtained by dividing the measured value by the length of the corresponding line segment multiplied by 10 points. For the scoring test method, the numerical scales used ranged from 1 to 10, with increasing hedonic. The assessment samples were coded with a 3-digit random number and poured into tasting cups of approximately 30 mL. For each sample, the sensory evaluation was performed three times. The final sensory evaluation score for each sample was taken as the average of the scores scored by the 10 evaluators. Before each tasting, the evaluators rinsed their mouths with purified water.

2.4. Establishment of artificial neural network model for fruit juice

Before establishing the model, data preprocessing is needed. The e-sensory data were averaged and fused as the model input, while the human sensory scores were used as the output. All data were normalized during data preprocessing, the dataset was randomly divided into training set and test set by 8:2 (Yang et al., 2023). Previous research has shown that KS sampling is a better method than random sampling (Rajer-Kanduč et al., 2003). KS sampling is based on an algorithm that first finds the two most distant objects in the entire object set and forms a group. Next, it selects the objects with the maximum and minimum distance from the objects in the group and adds them to the group. In this study, random sampling and KS sampling methods were used to divide the training and test sets.

Fruit juice e-sensory and human sensory attributes model was developed. Artificial neural network model was established using 19 characteristics measured by e-nose, e-tongue, colorimeter, and rheometer as input, and color, aroma, viscosity, sweetness, and acidity of human sensory attribute score of fruit juices as output, respectively. Then the human sensory attributes scores of the fruit juice were modeled with the human sensory hedonic score using an artificial neural network. Finally, a fully connected fruit juice e-sensory artificial neural network model was developed using 19 e-sensory features as input and the human sensory hedonic score as output. These models were based on Tensorflow (2.0.0) and keras (2.3.1) in python (3.6.12). To improve the accuracy of the model, parameters such as the number of hidden layers, number of neurons, and training batch size of the artificial neural network model were optimized.

2.5. Evaluation of artificial neural network model

To evaluate the model, the coefficient of determination R-squared (R²) and the root mean square error (RMSE) were used as metrics (Wang et al., 2021). R² reflects the accuracy of the model in fitting the data. Mathematically, the denominator represents the degree of dispersion of the original data, while the numerator represents the error between the predicted data and the original data. Dividing the two eliminates the influence of the degree of dispersion of the original data.

The greater the R² value, the better the model fits the data. RMSE represents the expected value of the square of the error. It is calculated as the square root of the square sum of the deviation between the observed value and the true value and the ratio of the number of observations n. RMSE is used to measure the deviation between the observed value and the true value.

$$R^2=\frac{SSR}{SST}=1-\frac{SSE}{SST}$$ (5)

$$SST=\sum _{i=1}^n{\left(y_i-{\bar{y}}_i\right)}^2$$ (6)

$$SSR=\sum _{i=1}^n{\left({\hat{y}}_i-{\bar{y}}_i\right)}^2$$ (7)

$$SSE=\sum _{i=1}^n{\left(y_i-{\hat{y}}_i\right)}^2$$ (8)

$$RMSE=\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(y_i-{\hat{y}}_i\right)}^2}$$ (9)

The calculation equation of R² was shown in (5) - (8), and RMSE was calculated as shown in (9). In these equations, SSR is the square sum of the difference between the predicted data and the mean value of the original data, SST is the square sum of the difference between the original data and the mean value, and SSE is the square sum of the error of the corresponding point between the fitted data and the original data, ${\hat{y}}_i$ is the predicted value, $y_i$ is the actual value, ${\bar{y}}_i$ is the mean value of the original data. n is the number of samples in the test set.

2.6. External data set validation

A total of thirteen samples were tested, including four different brands of beverages purchased (Nongfushanquan nfc orange juice, Xueshan probiotic compound juice beverage (red orange flavor), Leyuan Yipin sea buckthorn mixed juice beverage, Huiyuan 100 % orange juice) and nine self-made juices with the same formula but different processes. E-sensory evaluation and human sensory evaluation were performed on 13 samples. The e-sensory data was used as the input in the optimal model to predict the hedonic value and compared with the human sensory results.

3. Results and discussion

3.1. Fruit juice electronic and human sensory data set

The growing demand for high-quality fruit juice, driven by consumers' focus on personalized quality, has prompted increased attention in sensory quality and consumer perception. E-sensory evaluation is a method that uses computer and sensor technology to simulate human taste sensations, enabling the evaluation of fruit juice quality. E-sensory and human sensory evaluation of juice quality are based on the same material basis, that is, the chemical composition contained in food (Morten et al., 2020). Human sensory evaluation relies more on individual sensory experience and subjective judgment, resulting in inconsistencies and misinterpretations due to human variations. In contrast, e-sensory evaluation offers the advantage of providing numerical quantification of both the chemical components and taste characteristics of food, enabling rapid and consistent quality evaluation of juice. By utilizing e-sensory evaluation, manufacturers can gain a more accurate understanding of consumers' expectations and preferences for products, facilitating the development of products that better meet market needs and improve customer satisfaction. Although there is a growing body of research on evaluating sensory flavor profiles using e--sensory characteristics (Gabrieli et al., 2022; Jiang et al., 2021), the number of publications pertaining to machine learning and fusing e--sensory studies with consumer hedonic is relatively lacking (Nunes et al., 2023). Hence, this paper aims to utilize the fused e-sensory characteristics of juice to predict human sensory evaluation.

To predict human hedonic score, twenty-four e-sensory characteristics were collected for each sample. These characteristics were obtained through various sampling methods, including e-nose (10 characteristics), e-tongue (9 characteristics), colorimeter (4 characteristics), and rheometer (1 characteristic). To show the sensory differences in juices between different formulas and processes, a subset of 13 samples were randomly selected from 287 samples. Fig. 2A presents an e-nose sensor radar plot, showcasing the unique aroma profiles of different juice formulas. The plot visually represents the variations in the measured e-nose characteristics among the selected samples, highlighting the distinct sensory attributes of each juice. Similarly, Fig. 2B displays an e-tongue sensor radar plot, emphasizing the differences in taste profiles between the juice samples. This radar plot provides a visual representation of the e-tongue characteristics and their variations. Fig. 2C employs a bar graph to depict the color differences among the juice samples. The graph illustrates the variations in colorimeter characteristics, allowing for a comparison of the visual appearance of the juices. Lastly, Fig. 2D exhibits a line graph showcasing the rheological properties of the juice samples. The graph represents the measured rheometer characteristic, providing insights into the flow behavior and texture of the juices. These visual representations of the sensory data highlight the significant differences observed among the juice samples. The variations in the measured e-sensory characteristics signify that each juice possesses a unique taste, aroma, color, and texture. This is critical for consumers and manufactures. For consumers, they can choose the product that best suits their taste. For manufacturers, they can improve the quality and taste of their products, improve their manufacturing processes, and optimize their product formulas by understanding the differences in the e-sensory characteristics of juice products to create distinctive brands of juice.

Fig. 2.

Map of electronic sensory characteristics for different formulas of fruit juice. (A) E-nose sensors radar map, (B) E-tongue sensor radar chart, (C) Color difference bar graph, (D) Line chart of rheological property. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

For the purpose of e-sensory prediction of human sensory, human sensory attribute scores and human sensory hedonic score were obtained based on the QDA method and the scoring test. The scoring table for each sample included six sensory attributes: color, aroma, viscosity, sweetness, acidity, and overall hedonic value. A total of twelve samples were randomly selected from these attributes to draw a radar plot displayed in Fig. 3A and B. Following human sensory evaluations of juices produced using different formulas and processes, the obtained attribute scores varied greatly, thus resulting in significantly different overall sensory evaluation scores for each sample. This means that human sensory evaluation can determine which formula and processing method can improve the taste and quality of juice products, thereby improving market competitiveness. In addition, by comparing sensory evaluation scores for different juice products, manufacturers can rank their products and identify the most advantageous formulas and processes. This allows for the development of a more scientific and rational production strategy, resulting in increased production efficiency and market share. Therefore, the difference of human sensory evaluation scores of juice products with varying formulas and processing methods is very important.

Fig. 3.

Radar chart of human sensory evaluation attributes of different formulas of fruit juice. (A) QDA method, (B) Scoring test method.

3.2. Fruit juice e-sensory features and human sensory attributes model

Data pre-processing is a crucial step in establishing a high-quality model. In this study, e-tongue, e-nose, colorimeter, and rheological were used to obtain e-sensory characteristics. For the e-tongue data, all measurements are outputted as absolute value with the reference solution as the standard. The e-tongue reference solution is composed of KCl and tartaric acid. KCl is used to simulate the salty taste, and tartaric acid is used to simulate the sour taste in the sample. The tasteless point of the measured sour taste was −13 and the tasteless point of the salty taste was −6. If the taste value of a sample is lower than the tasteless point, it indicates that the sample does not contain the taste, and vice versa. The juice e-tongue data were filtered accordingly, yielding valid features of sweetness, sourness, richness, and saltiness. In addition to the e-tongue data, ten effective features obtained by the e-nose experiment, 4 features obtained by the color difference experiment, and 1 feature obtained by the rheological experiment. Ultimately, 19 valid e-sensory features were obtained.

To explore the correlation between e-sensory and human sensory evaluation, this part involved establishing models of e-sensory characteristics and human sensory attributes of fruit juice. Specifically, e-sensory feature data were filtered as the input of the models, and the five sensory attribute scores obtained from the human sensory evaluation of the juice were used as the output, respectively. Among the models established using e-sensory and human sensory attributes obtained by the QDA method, the R² values for color, aroma, sweetness, sourness, and viscosity were 0.90, 0.81, 0.85, 0.80, and 0.20, respectively, and the RMSE values were 0.47, 0.54, 0.64, 0.40, 0.77, respectively. For the models established using e-sensory and human sensory attributes obtained by the scoring test method, the R² values were 0.81, 0.60, 0.73, 0.70, 0.09, respectively, and the RMSE were 0.69, 0.78, 0.88, 0.52, 0.93, respectively. The R² of the model established by the attributes other than viscosity was greater than 0.5, indicating a considerable correlation between the e-sensory and human sensory attributes of fruit juice. The model performance was shown in Supplementary Tables S3a and S3b. Supplementary Table S3a was the model results obtained by the QDA method for the human sensory evaluation of the juice, while Supplementary Table S3b was the model results obtained by scoring test method. The fitting plots of the model were shown in Fig. 4 (1) (QDA method) and (2) (scoring test method).

Fig. 4.

Fitted artificial neural network model of e-sensory and human sensory evaluation attributes. (1) Human sensory attribute scores obtained by QDA method, (2) Human sensory attribute scores obtained by the scoring test. (A) color, (B) aroma, (C) sweet, (D) sour, (E) viscosity. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

The establishment of fruit juice fusion e-sensory characteristics and human sensory attributes model can achieve the prediction of human sensory attribute score of fruit juice or other new products by using the e-sensory characteristics. The attribute values are used to describe and quantify the taste and quality of the juice. In this way, these attribute parameters can be quantitatively monitored and controlled during the production process to ensure that the taste and quality of the juice meet the requirements. And it also helps the juice industry to better grasp the quality and taste of products, thereby improving the competitiveness of enterprises. In summary, the establishment of a model for predicting human sensory attributes (sweetness and acidity) using electronic sensory data of juice is only the first step to better control product quality and taste.

3.3. Fruit juice human sensory attributes and hedonic model

In practice, it is not enough to establish a model to predict human sensory attributes. The human sensory hedonic of juice is also a crucial factor that must be considered. Therefore, we established a human sensory hedonic model in addition to the human sensory attribute prediction model to accurately predict the human hedonic of juice taste. By optimizing the model, the R² and RMSE of the model based on the human sensory attribute scores measured by QDA were 0.77 and 0.42, respectively. The R² and RMSE of the optimal model based on the human sensory attribute scores measured by the scoring test were 0.63 and 0.63, respectively. The fitting plot of e-sensory features and human sensory hedonic score model was shown in Fig. 5A (QDA method) and B (scoring test method).

Fig. 5.

Fitted artificial neural network model of human sensory evaluation attributes with human sensory hedonic score of fruit juice. (A) Human sensory scores obtained by QDA method, (B) Human sensory scores obtained by the scoring test.

By establishing a relationship model between the human sensory attributes of fruit juice and the hedonic, it is possible to determine which sensory attributes are most important to consumers. Thereby the product formula and processing can be optimized to better meet the consumer needs and preferences. In addition, the model can also help predict the market reaction of new products and provide suggestions for improving existing products. The accuracy of the fitted model was greater than 0.5, it was still much less than 1. This indicates that there is a correlation between human sensory attribute scores and acceptance of juice, but the latter is not entirely determined by the results of human sensory attribute scores. The high score on human sensory attributes does not necessarily mean that consumers will accept the product, but it can indicate certain advantageous aspects of the product. If a juice scores high in multiple dimensions of the human sensory attributes, it is likely that the juice will be closer to consumer expectations and therefore may be more popular.

3.4. Fused e-sensory features and human sensory hedonic model

The model performance established by the fusion of e-sensory characteristics and human sensory attribute scores of juices is relatively better than the model performance established by the human sensory attribute scores and hedonic of juices. Therefore, in order to gain a more comprehensive understanding of the human sensory hedonic of fruit juice in multiple dimensions, it is very crucial to develop a model of the fused e-sensory data of juice with the human overall hedonic. Multi-electronic sensory technology-based modeling allows for faster and more accurate measurement characteristics of juice samples, which greatly reduces the cost and time of sensory evaluation and improves the repeatability of results. Moreover, by modeling the e-sensory data of juices with the human sensory overall hedonic, we can not only analyze the influence of different factors on the juice taste, but also determine which factors are most important and how to optimize the formula to improve product quality. In addition, this modeling method can also help juice producers develop new products that meet consumer needs and provide more accurate data support in marketing and promotion. Therefore, it is of great significance to model the fused e-sensory data of juice and the human overall acceptance for the development of juice industry.

This part is to explore the juice human sensory hedonic from a multidimensional perspective using the fused e-sensory technique. We used the fused 19 e-sensory indicators instead of the human sensory attributes as the model input, and human sensory hedonic value as the target to establish an artificial neural network. The optimal model performance was obtained with five hidden layers and 1024, 512, 256, 128, and 64 neurons, with R² of 0.95 and RMSE of 0.04 for the QDA method. This model was named model-1. In contrast, the optimal performance for the scoring test method was obtained with three hidden layers and 64, 32, and 16 neurons, resulting in an R² value of 0.88 and RMSE of 0.21. This model was designated as model-2.

To improve the model performance, data normalization was performed. For data set partitioning, two sampling methods, random sampling and KS sampling, were implemented and their results were compared. As shown in Supplementary Tables S4a and 4b, the model performance obtained by KS sampling method is more stable and better than that obtained by random sampling method. The R² increased from 0.86 to 0.95 for the model built by using QDA and from 0.71 to 0.88 for the model built by using the scoring test method for sensory evaluation. Several studies have shown that the use of KS sampling in artificial neural networks can improve the performance of the model (Riahi-Madvar et al., 2021; Singh et al., 2019). To prevent the over-fitting phenomenon of artificial neural network, we used the Dropout function to optimize the model during the model building process (Srivastava et al., 2014). In the process of each training batch, half of the feature detectors in the network were ignored to improve the generalization ability of the model, thereby significantly reducing the over-fitting phenomenon. In addition, the process of optimizing the number of hidden layers and neurons of the artificial neural network was shown in Supplementary Tables S4a and 4b.

Currently, there are studies in the field of food researches that utilize sensory properties of food for modeling. Yu et al. (2018) used PLS-ANN regression model to predict the consumer preference scores of the green tea model system for the purpose of developing green tea flavors. Bi et al. (2020) used a support vector machine (SVM) for yogurt preference prediction based on sensory attributes. Although these studies aimed to use algorithms for the prediction of preference value of food products, the model and the sensory features used were different. In contrast, this study used the fusion e-sensory characteristics data as the basic data. Moreover, combined with the artificial neural network model with strong self-learning capability to predict human sensory acceptance.

Among the established models of fused e-sensory characteristics and human sensory attributes, human sensory attributes and hedonic value, and fused e-sensory characteristics and human sensory hedonic value, the fused e-sensory characteristics and human sensory hedonic value model had the highest R² value of 0.95 and the lowest RMSE of 0.04. This indicates that the use of multi-electronic sensory techniques is better than the use of human sensory methods to evaluate juice quality. These three models were developed to better understand the sensory characteristics of fruit juice and their hedonic by human, and to provide relevant reference and guidance for juice manufacturers. The fusion of e-sensory data and human sensory attributes model aims to transform the physicochemical indicators of juice (e.g., aroma, sweetness, and sourness, etc.) into human subjective perception of these indicators, such as good or bad taste, aroma intensity, sweetness, and sourness suitability, etc. Through this model, we can rapidly and accurately predict the human sensory attribute of a certain juice, and thus better control the production process of juice. The artificial neural network model between human sensory attributes and hedonic of juice aims to transform human subjective perception into a quantifiable indicator-human sensory hedonic. The human sensory hedonic of the juice was measured by human sensory evaluation of a certain number of subjects. Using this model, we can better understand human hedonic for a certain kind of juice and provide relevant guidance for manufacturers. The artificial neural network model of juice fusion e-sensory data and human sensory hedonic was a combined application of the first two models. It aims to transform the physicochemical indicators of fruit juice into human sensory hedonic. Through this model, we can gain insight into the influence of different indicators on the human sensory hedonic of fruit juice and further optimize the production process of fruit juice to improve product quality and market competitiveness. In conclusion, there is a close connection between these three models. From physicochemical indicators of juice to human subjective perception to the human sensory overall hedonic, it is a step-by-step process that needs to be achieved by multiple models. Each model has its own unique meaning and role, but together they constitute a complete research framework, providing valuable information and guidance for juice manufacturers. Of course, our model has strong generalization capability and can be applied to other new products.

3.5. Model application

To further evaluate the performance of fruit juice e-sensory and human sensory hedonic model, 13 samples were selected for external validation. The 19 e-sensory characteristics were used as the input for both model-1 and model-2. The results showed that R² and RMSE of model-1 was 0.90 and 0.69, respectively. And the R² and RMSE of model-2 was 0.86 and 0.66, respectively. Comparing the model prediction results with the human sensory evaluation results, the model prediction results are correct within the allowable error range, demonstrating the feasibility of the method in this study.

This study successfully achieved the objective of predicting human sensory hedonic of fruit juice products using fused e-sensory features. When applying the model in practice, we can use transfer learning to train other data. This involves inputting the e-sensory data of a product to train the model, which can then predict the human sensory hedonic score. To some extent, this study provides a powerful potential possible method for alternative sensory evaluation. Additionally, it accelerates the understanding of human sensory hedonic for new products, and helps to improve their quality. The next step will be to improve the transparency of the model by interpreting the optimal model established for the purpose of better analysis of sensory evaluation.

4. Conclusions

This study proposed a method combining machine learning algorithm with multiple intelligent sensory instruments to achieve human sensory hedonic prediction. The multiple e-sensory techniques enable comprehensive characterization of key attributes such as color, aroma, sweetness, and sourness of fruit juice. The hedonic score was positively correlated with the color, aroma, sweetness, sourness, and viscosity of the juice. For quantitative analysis, the ANN model based on the fused data of multiple e-sensory techniques could predict the scores of different human sensory attributes and hedonic. The QDA method, coupled with data partitioning using the KS sampling method, outperforms the model built by the method for human sensory assessment of fruit juice using the scoring test with a randomly partitioned data set. This study successfully established the relationship between fruit juice e-sensory evaluation and human sensory hedonic. In the future, the proposed method holds promise for the development of a website or platform for predicting human sensory hedonic score for fruit juices or other food products. Overall, our study demonstrates the potential and value of machine learning-assisted e-sensory analysis in rapidly predicting juice human sensory hedonic. It provides scientific support for juice evaluation based on multiple e-sensory techniques, thereby aiding in the improvement and optimization the fruit products. In subsequent study, more effort can be directed towards the development of rapid detection method and data fusion strategy based on multiple intelligent sensory technologies.

Data and code availability

Data and code will be made available on request.

CRediT authorship contribution statement

Huihui Yang: Investigation, Data curation, Formal analysis, Visualization, Writing – original draft, Writing – review & editing. Yutang Wang: Investigation, Resources, Formal analysis, Project administration, Methodology, Writing – review & editing. Jinyong Zhao: Writing – review & editing. Ping Li: Writing – review & editing. Long Li: Supervision, Writing – review & editing. Fengzhong Wang: Supervision, Project administration, Methodology, Writing – review & editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the Supplementary data to this article:

Data availability

Data will be made available on request.