A Spectroscopic Method and System for Detecting the Content of Aroma Active Compounds in Grapefruit Hydrosol and Floral Tea

Abstract

This invention relates to the field of materials analysis technology, specifically disclosing a method and system for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectroscopy. The method involves constructing a simulated matrix sample library and collecting its first component characteristic spectra and sensory attribute quantification values ​​to establish a baseline prediction model. Then, by collecting the second component characteristic spectra and sensory attribute quantification values ​​of real samples, an enhanced prediction model is obtained through feature distribution adaptation and transfer learning techniques. Finally, volatile component features are extracted from the sample to be tested, and the target component characteristic spectra are input into the enhanced prediction model to directly output the sensory attribute quantification values ​​as the detection result. This invention effectively solves the problem of aroma contribution evaluation distortion caused by matrix-removal analysis in traditional gas chromatography-olfactometry, achieving accurate mapping from chemical components to sensory attributes, and significantly improving detection efficiency while ensuring detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of materials analysis technology, specifically to a method and system for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectroscopy. Background Technology

[0002] In the production and quality control of pomelo hydrosol flower tea, accurately identifying and quantifying its key aroma active substances is a core step in ensuring product flavor quality. For a long time, the industry has relied on gas chromatography-olfactometry (GC-O) and its derivative methods as the "gold standard" for flavor analysis. This technique involves chromatographically separating volatile components in a complex sample, followed by individual smelling and identification by evaluators to screen for active compounds contributing to the overall aroma and calculate their aroma intensity (such as the FD factor). However, this mainstream approach is essentially based on a simplified analytical model, assuming that the aroma contribution of a single compound obtained through separation-olfactometry can be directly linearly superimposed and is equivalent to its sensory performance in the complete product.

[0003] The existing technology has the following shortcomings: In olfactory evaluation using GC-O technology, aroma molecules arrive directly from the evaluator's nasal cavity via an inert chromatographic carrier gas. This completely strips away the complex matrix (including aqueous phase, organic acids, sugars, phenols, etc.) present in natural grapefruit hydrosol tea. This "pure" evaluation environment ignores the matrix masking and enhancement effects commonly found in actual product systems. For example, trace amounts of organic acids in the matrix may enhance the freshness of certain ester aromas, while bitter substances such as naringin may suppress the perceived intensity of sweet aromas. In actual tasting, the release kinetics of aroma molecules, their interactions with other components, and competitive binding on sensory receptors collectively determine their final flavor profile. The FD factor provided by GC-O only represents the potential aroma intensity of a compound under "ideal conditions," not its actual performance in the "real battlefield" (i.e., the complete product). This directly leads to a significant gap between GC-O data-guided flavor control (such as artificial blending) and the harmonious and full aroma profile of natural products. Blended products generally exhibit the drawbacks of heavy "artificial flavor", harsh aroma, and lack of layering. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectroscopy, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A spectral-based method for detecting the aroma active substances in grapefruit hydrosol flower tea: S1: Constructing a matrix simulation system by systematically mixing aqueous medium, acidity regulator, sugar components and phenolic substances to establish a simulated matrix sample library with different component ratios. S2: Multi-source information collection is performed on the simulated matrix sample library, including obtaining the first component characteristic spectrum of each simulated matrix using volatile component extraction technology, and obtaining the corresponding sensory attribute quantification values ​​through sensory quantification analysis. S3: Establish a benchmark prediction model, using the first component feature spectrum as the input of the benchmark prediction model and the quantified value of the sensory attribute as the output of the benchmark prediction model, and train the benchmark prediction model through a deep neural network algorithm. S4: Multi-source information collection was carried out on real grapefruit hydrosol flower tea samples, including obtaining the second component characteristic spectrum of real samples by using volatile component extraction technology, and obtaining the sensory attribute quantitative values ​​of real samples through sensory quantitative analysis. S5: Based on the multi-source information of the benchmark prediction model and real samples, the feature distribution of the benchmark prediction model parameters is adapted to obtain the enhanced prediction model; S6: Extract the volatile components of the grapefruit hydrosol flower tea sample to be tested, obtain the target component feature spectrum, input the target component feature spectrum into the enhanced prediction model, and output the comprehensive sensory attribute quantification value of the sample to be tested as the final detection result of the aroma active substance content.

[0006] As a further aspect of the present invention: the construction of the matrix simulation system specifically includes: A basic aqueous medium is established, using tertiary purified water as the solvent base. The conductivity of the aqueous medium is controlled within a preset range through ion exchange and reverse osmosis treatment, and solvent deoxygenation treatment is carried out under inert gas protection. A complex component system was configured. According to the natural proportion range of each component in pomelo hydrosol flower tea, the organic acid component, sugar component and polyphenolic substance were accurately weighed separately. Gradient dilution and isothermal mixing technology were used to achieve stepwise mixing of each component under constant temperature conditions through a microfluidic device. A sample library variation system was constructed. Based on the orthogonal experimental design method, the ratio and concentration of each component in the simulated matrix were systematically changed. By introducing nonlinear concentration gradient changes, a simulated matrix sample sequence with continuous concentration changes was constructed. The precise component composition parameters of each sample were recorded.

[0007] As a further aspect of the present invention: the process of obtaining the first component feature spectrum is as follows: A dynamic headspace extraction device was used to capture volatile components. The simulated matrix sample was placed in a constant temperature oscillating reactor and captured by passing through a multi-stage cold trap at a preset flow rate under inert gas, thereby achieving low-temperature enrichment of volatile components. Programmed temperature rise analysis is implemented to transfer the captured volatile components for thermal desorption under strict control of the heating rate. The separation and release of components with different boiling point ranges are achieved through multi-stage temperature gradient control. The analyzed volatile components were introduced into a coupled chromatographic system and separated using a combination of chromatographic columns. The mass spectrometric characteristics of each component were obtained by a mass spectrometer detector. The obtained mass spectrometry data were processed by peak alignment, denoising and normalization to establish a first component characteristic spectrum containing retention time, mass spectrometry characteristic peaks and relative intensities.

[0008] As a further aspect of the present invention: the process of obtaining the sensory attribute quantification value is as follows: Establish a standardized sensory evaluation environment in a laboratory with constant temperature and humidity and purified air, using a sensory evaluation room with specific lighting conditions to eliminate the interference of environmental factors on sensory judgment. The organization selected sensory evaluators and conducted sensory calibration using standard samples to establish a sensory attribute evaluation system that includes fruit aroma intensity, floral aroma characteristics, sweetness level, and freshness. During the evaluation process, an electronic recording system is used to record the curves of the intensity of each sensory attribute changing over time in real time, thereby obtaining the dynamic response characteristics of the sensory attributes. The recorded sensory intensity data is normalized and checked for consistency to generate quantitative values ​​of sensory attributes.

[0009] As a further aspect of the present invention: the establishment of the benchmark prediction model specifically includes: The retention time, mass spectrometry characteristic peaks, and relative intensity data in the first component feature spectrum are normalized to construct a multidimensional feature vector; A deep neural network architecture is constructed, adopting a multi-branch network structure design, including feature extraction branch, association analysis branch, and weight allocation branch. Each branch realizes the interactive transmission of feature information through cross-layer connections. The network parameters are initialized using pre-training, the network weights are optimized using the backpropagation algorithm, and finally the hyperparameters of the benchmark prediction model are adjusted using a fine-tuning strategy. The model performance is evaluated using cross-validation, key feature parameters are selected through feature importance analysis, and the network structure is iteratively optimized until the preset accuracy requirements are met.

[0010] As a further aspect of the present invention: the process of obtaining the second component characteristic spectrum is as follows: Real grapefruit hydrosol flower tea samples were placed in a constant temperature shaker and dynamically headspaced under strictly controlled temperature conditions. The volatile components were enriched using a collection tube filled with a composite adsorbent. The enriched volatile components are thermally desorbed under programmed temperature rise conditions, and low-temperature focusing and rapid desorption of volatile components are achieved through a multi-stage cold trap focusing device. The analyzed volatile components are introduced into a chromatographic system equipped with a composite chromatographic column, and the efficient separation of different volatile components is achieved by using a gradient temperature program and precise flow control. Mass spectra of each separated component were obtained using a mass spectrometer, and a second component characteristic spectrum containing retention time index, characteristic ion fragments, and relative abundance information was established.

[0011] As a further aspect of the present invention: the method of adapting the feature distribution of the parameters of the benchmark prediction model to obtain an enhanced prediction model based on multi-source information from the benchmark prediction model and real samples specifically includes: By calculating the distribution differences between simulated matrix samples and real samples in the component feature spectrum space, a feature mapping function is established to project and transform the feature representation in the benchmark prediction model to the feature space of the real sample. While keeping the core architecture of the benchmark prediction model unchanged, a hierarchical progressive optimization strategy is adopted to adjust the parameter weights from the shallow to the deep layers of the network in turn, so that the benchmark prediction model gradually adapts to the feature distribution law of real samples. By calculating the difference in feature distribution between the model outputs in the simulated domain and the real domain, a domain adaptation loss function is established. The network parameters are then iteratively optimized based on the backpropagation algorithm to minimize the difference in feature distribution between the two domains.

[0012] As a further aspect of the present invention: the extraction of volatile component characteristics from the grapefruit hydrosol flower tea sample to be tested specifically includes: The sample to be tested is placed in a constant temperature oscillation device, and the volatile components are brought to equilibrium in the headspace by precisely controlling the oscillation frequency and temperature parameters. By combining ultrasonic-assisted extraction and microwave-induced release, volatile components with different boiling point ranges in the sample are activated simultaneously. The obtained target component characteristic spectrum is aligned and corrected with the first component characteristic spectrum and the second component characteristic spectrum to eliminate the detection error caused by instrument drift. The corrected feature spectrum is subjected to baseline correction and peak area normalization to establish a target component feature spectrum consistent with the input dimension of the benchmark prediction model.

[0013] As a further aspect of the present invention: the step of inputting the target component characteristic spectrum into the enhanced prediction model and outputting the comprehensive sensory attribute quantification value of the sample to be tested as the final detection result of the aroma active substance content specifically includes: The target component feature spectrum is fused with the first component feature spectrum and the second component feature spectrum at the feature level, and feature enhancement is achieved by calculating feature similarity weights. Based on the degree of matching between the target component feature spectrum and the historical feature spectrum, the contribution weight of each feature dimension in the enhanced prediction model is dynamically adjusted. The target component feature spectrum is subjected to a dual-domain alignment transformation using the reference feature space established by the first component feature spectrum and the real feature space constructed by the second component feature spectrum. By enhancing the multi-level neural network calculation of the prediction model, a comprehensive sensory attribute quantification value including fruit aroma intensity, floral aroma characteristics, sweetness, and freshness is output.

[0014] A spectral-based system for detecting the content of aroma active substances in grapefruit hydrosol and floral tea includes: A simulated matrix construction module is used to construct a matrix simulation system. By systematically mixing aqueous medium, acidity regulator, sugar components and phenolic substances, a simulated matrix sample library with different component ratios is established. A multi-source information acquisition module is used to acquire multi-source information from a simulated matrix sample library, including obtaining the first component characteristic spectrum of each simulated matrix using volatile component extraction technology, and obtaining the corresponding sensory attribute quantification values ​​through sensory quantification analysis. A benchmark model training module is used to establish a benchmark prediction model. The first component feature spectrum is used as the input of the benchmark prediction model, and the sensory attribute quantification value is used as the output of the benchmark prediction model. The benchmark prediction model is obtained by training through a deep neural network algorithm. The real sample adaptation module is used to collect multi-source information from real grapefruit hydrosol flower tea samples, including obtaining the second component characteristic spectrum of the real sample by using volatile component extraction technology, and obtaining the sensory attribute quantification value of the real sample through sensory quantification analysis. The migration enhancement and optimization module adapts the feature distribution of the parameters of the benchmark prediction model based on the multi-source information of the benchmark prediction model and the real sample to obtain the enhanced prediction model. The sensory attribute prediction module is used to extract the volatile component characteristics of the grapefruit hydrosol flower tea sample to be tested, obtain the target component feature spectrum, input the target component feature spectrum into the enhanced prediction model, and output the comprehensive sensory attribute quantification value of the sample to be tested as the final detection result of the aroma active substance content.

[0015] The beneficial effects of this invention are: (1) This invention constructs a simulated matrix sample library that can simulate the matrix characteristics of real products, and uses transfer learning technology to adapt the benchmark prediction model from the simulated domain to the real domain, enabling the model to learn and predict the complex interactions between matrix components and aroma active substances. This invention enables the final obtained sensory attribute quantification values ​​to truly reflect the actual sensory performance of aroma active substances in the complete product system, reducing the "artificial fragrance" problem caused by neglecting matrix effects in traditional methods, and providing a more reliable data foundation for product quality control, flavor reproduction, and new product development.

[0016] (2) This invention establishes a prediction model based on a deep neural network. After training the enhanced prediction model, the sample to be tested only needs to undergo rapid volatile component feature extraction once (such as dynamic headspace extraction combined with GC-MS analysis). The obtained target component feature spectrum can be input into the model, and the comprehensive sensory attribute quantification value can be directly output in a short time. This realizes a rapid and accurate mapping from "chemical composition" to "sensory perception", avoiding the cumbersome process of organizing professional sensory evaluation for each test. While ensuring the accuracy of the results, it reduces time and economic costs, and provides key technical support for rapid quality monitoring and product iteration on the production line. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the system of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, this invention is a spectral-based method for detecting the content of aroma active substances in grapefruit hydrosol flower tea, comprising the following steps: S1: Construct a matrix simulation system by systematically mixing aqueous medium, acidity regulator, sugar components and phenolic substances to establish a simulated matrix sample library with different component ratios; S2: Multi-source information collection is performed on the simulated matrix sample library, including obtaining the first component characteristic spectrum of each simulated matrix using volatile component extraction technology, and obtaining the corresponding sensory attribute quantification values ​​through sensory quantification analysis. S3: Establish a benchmark prediction model, using the first component feature spectrum as the input of the benchmark prediction model and the quantified value of the sensory attribute as the output of the benchmark prediction model, and train the benchmark prediction model through a deep neural network algorithm. S4: Multi-source information collection was carried out on real grapefruit hydrosol flower tea samples, including obtaining the second component characteristic spectrum of real samples by using volatile component extraction technology, and obtaining the sensory attribute quantitative values ​​of real samples through sensory quantitative analysis. S5: Based on the multi-source information of the benchmark prediction model and real samples, the feature distribution of the benchmark prediction model parameters is adapted to obtain the enhanced prediction model; S6: Extract the volatile components of the grapefruit hydrosol flower tea sample to be tested, obtain the target component feature spectrum, input the target component feature spectrum into the enhanced prediction model, and output the comprehensive sensory attribute quantification value of the sample to be tested as the final detection result of the aroma active substance content.

[0021] In S1, a matrix simulation system is constructed. By systematically mixing aqueous medium, acidity regulator, sugar components, and phenolic substances, a simulated matrix sample library with different component ratios is established, specifically including: Three-stage purified water was used as the solvent base. Ion exchange resin was used to remove cation and anion impurities from the water, followed by reverse osmosis to further remove dissolved solids. Throughout the process, a smart sensor continuously monitored changes in the conductivity of the aqueous medium to ensure it remained stable within a preset range. After purification, high-purity inert gas was bubbled through a sealed container for at least 30 minutes to effectively remove dissolved oxygen molecules. This process prevents oxidation reactions of components during subsequent experiments, ensuring the stability of the experimental system.

[0022] It should be noted that the core of the intelligent sensor in this purified water treatment process is a high-precision conductivity sensor, which directly contacts the aqueous medium to collect the conductivity signal caused by the change in ion concentration in the solution in real time.

[0023] Secondly, a composite component system was configured. Based on systematic analysis of natural grapefruit hydrosol flower tea samples, the natural proportions of organic acid components, sugar components, and polyphenols were determined. Each component was precisely weighed using a precision analytical balance. A gradient dilution method was employed to prepare high-concentration stock solutions for each component. Then, isothermal mixing technology was used to perform stepwise mixing under constant temperature conditions using a microfluidic device. The microfluidic device precisely controlled the flow rate and mixing ratio of each channel to ensure that different components achieved a uniform molecular-level dispersion during the mixing process. The entire mixing process was conducted in a constant temperature environment, with temperature fluctuations controlled within ±0.5 degrees Celsius.

[0024] Finally, a variation system for the sample library was constructed. Based on orthogonal experimental design, the effects of the concentration ratios of each component in the simulated matrix on the sensory properties of aroma active substances were systematically investigated. The orthogonal experimental design included four factors: organic acid concentration, sugar concentration, polyphenol concentration, and pH value, with five levels for each factor. By introducing a nonlinear concentration gradient, the concentrations of each component exhibited a continuous distribution characteristic within their natural range of variation. In practice, the baseline concentration levels of each component were first determined, and then the experimental sequence was arranged according to the orthogonal array, preparing simulated matrix samples with different ratios sequentially. After each sample was prepared, its component composition parameters, such as the concentration values ​​of organic acid, sugar, polyphenol, and pH value, were recorded in detail. The final constructed simulated matrix sample sequence contained 125 independent samples, each with a unique combination of component composition parameters.

[0025] During implementation, special attention should be paid to the following technical details: the conductivity control range of the basic aqueous medium should be determined according to actual needs; during the inert gas protection process, the gas flow rate should be kept stable to avoid solvent evaporation due to excessive flow rate; the channel size and flow rate parameters of the microfluidic device need to be optimized and adjusted according to the specific mixing system; when designing orthogonal experiments, the level settings of each factor should cover the actual possible concentration range to ensure that the sample library has sufficient representativeness.

[0026] The simulated matrix sample library constructed using the above methods can effectively simulate the matrix characteristics of real grapefruit hydrosol flower tea, providing a reliable experimental platform for subsequent research on the sensory characteristics of aroma active substances. This sample library contains various possible combinations of component ratios, comprehensively reflecting the changes in the sensory characteristics of aroma active substances under different formulation conditions, laying a solid foundation for establishing accurate predictive models.

[0027] In the implementation process, a corresponding quality control system needs to be established. This includes regularly calibrating the accuracy of the analytical balance, monitoring the working status of the microfluidic device, and verifying the rationality of the orthogonal experimental design. These quality control measures can ensure the reliability and reproducibility of experimental data, providing reliable technical support for subsequent research.

[0028] Furthermore, the stability of experimental conditions must be considered during the construction of the sample library. The laboratory temperature should be maintained between 20 and 25 degrees Celsius, and the relative humidity between 40% and 60%. All laboratory equipment must undergo a rigorous cleaning procedure before use to ensure no external contamination is introduced. These measures can minimize experimental errors and improve the reliability of experimental results.

[0029] In S2, multi-source information is collected from the simulated matrix sample library, including obtaining the first component characteristic spectrum of each simulated matrix using volatile component extraction technology, and obtaining the corresponding sensory attribute quantification values ​​through sensory quantification analysis, specifically including: In the implementation of the volatile component extraction technology, dynamic headspace extraction is performed first. The simulated matrix sample is placed in a isothermal oscillating reactor, with the temperature controlled at 40 degrees Celsius and the oscillation frequency at 200 times per minute. High-purity nitrogen is introduced into the closed system as a carrier gas, with a flow rate controlled at 50 ml per minute. Volatile components are carried by the carrier gas into a multi-stage cold trap collection system. The first-stage cold trap temperature is -30 degrees Celsius, mainly used to capture high-boiling-point components; the second-stage cold trap temperature is -80 degrees Celsius, mainly used to capture medium-boiling-point components; and the third-stage cold trap temperature is -120 degrees Celsius, mainly used to capture low-boiling-point components. The collection time for each sample is 30 minutes to ensure sufficient enrichment of various volatile components. After collection is completed, the cold trap system is immediately switched to a closed state to prevent sample contamination.

[0030] Next, a programmed temperature-controlled desorption process was implemented. The captured volatile components were thermally desorbed under strictly controlled conditions. The desorption process consisted of three stages: the first stage involved increasing the temperature from the initial temperature to 80°C at a rate of 10°C per minute, primarily releasing low-boiling-point components; the second stage involved increasing the temperature from 80°C to 150°C at a rate of 5°C per minute, primarily releasing medium-boiling-point components; and the third stage involved increasing the temperature from 150°C to 250°C at a rate of 8°C per minute, primarily releasing high-boiling-point components. Each temperature stage was held for 5 minutes to ensure complete release of components within the corresponding boiling point range. A carrier gas was continuously introduced during the desorption process to transfer the released volatile components to the subsequent analysis system.

[0031] Then, chromatographic separation and mass spectrometry detection were performed. The analyzed volatile components were introduced into a gas chromatography-mass spectrometry (GC-MS) system. Chromatographic separation employed a combined column configuration: the first column was a weakly polar column, 30 meters long, with an inner diameter of 0.25 mm and a film thickness of 0.25 μm; the second column was a moderately polar column, 60 meters long, with an inner diameter of 0.32 mm and a film thickness of 0.5 μm. The column oven temperature program was set as follows: initial temperature 40°C, held for 3 minutes; increased to 120°C at 5°C / min; then increased to 250°C at 10°C / min, held for 10 minutes. High-purity helium was used as the carrier gas at a flow rate of 1.2 mL / min. The mass spectrometer detector ion source temperature was set to 230°C, the interface temperature to 280°C, and the mass scan range to 35 to 450 atomic mass units.

[0032] Finally, mass spectrometry data processing and characteristic spectrum establishment were performed. Peak alignment was performed on the obtained mass spectrometry data, using retention time as the alignment benchmark, with a time window set to 0.1 minutes. Denoising was performed using wavelet transform to remove baseline drift and random noise. Normalization was performed based on the total ion chromatogram, converting the intensity of each characteristic peak into relative intensity values. The established first-component characteristic spectrum contains the following information: the top 10 most intense characteristic ions recorded by mass spectrometry peaks and their relative abundance, expressed as a percentage. All data was stored in a standardized format for easy subsequent analysis and processing.

[0033] In the implementation of sensory quantitative analysis, a standardized sensory evaluation environment was first established. The sensory evaluation laboratory was divided into two areas: a preparation room and an evaluation room. The evaluation room had internal dimensions of 4 meters long, 3 meters wide, and 2.5 meters high, with neutral gray paint on the walls. The ambient temperature was controlled at 22 degrees Celsius, the relative humidity at 55%, and the background noise level below 35 decibels. The air purification system used a three-stage filtration system to ensure that the concentration of particulate matter in the air was below 0.1 milligrams per cubic meter. Each evaluation room was equipped with an independent ventilation system, and the interval between each sample evaluation was 5 minutes to ensure complete air renewal.

[0034] Next, sensory evaluators were selected and trained. Selection criteria included age between 20 and 45, no smoking history, and no olfactory or gustatory disorders. Qualified personnel were selected through a standard olfactory test, which included the ability to identify and differentiate standard aroma compounds such as limonene, linalool, and geraniol acetate. The training lasted two weeks, with three hours of training per day. Training content included theoretical learning, standard sample identification, and intensity calibration. Evaluators were trained using standard aroma samples to identify and judge the intensity of four main sensory attributes: fruity intensity, floral characteristics, sweetness, and freshness. An assessment was conducted after training; only evaluators who passed all assessment items were eligible to participate in formal evaluations.

[0035] Then, dynamic sensory evaluation and recording were implemented. During the evaluation process, each evaluator used a separate evaluation room. Samples were assigned a three-digit random code, and 15 ml of sample was provided in a standard evaluation cup each time, with the sample temperature controlled at 25 degrees Celsius. Evaluators began evaluating immediately upon receiving the sample, using an electronic recording system to record the intensity of each sensory attribute in real time. The recording system had four intensity scale bars, corresponding to fruit aroma intensity, floral characteristics, sweetness, and freshness, respectively. Evaluators recorded the intensity changes in real time by sliding the scale bars, with a sampling frequency of once per second. Each sample was evaluated for 60 seconds. After evaluation, evaluators immediately rinsed their mouths with water, rested for 5 minutes, and then proceeded to evaluate the next sample.

[0036] Finally, sensory data standardization was performed. The recorded sensory intensity data were time-axis aligned, with the sample contact time as the zero point. Normalization correction employed a min-max standardization method, transforming the intensity value of each sensory attribute to a range of 0 to 100. Consistency testing used analysis of variance, calculating the coefficient of variation for each evaluator across different replicates and discarding data with a coefficient of variation greater than 15%. For each sample, the average intensity value for all valid evaluators at each time point was calculated, generating a sensory attribute quantification matrix. This matrix contains the intensity values ​​of the four sensory attributes at 60 time points, totaling 240 quantified values, serving as the sensory attribute quantification value for each sample.

[0037] In S3, a baseline prediction model is established, using the first component feature spectrum as the input and the quantified value of the sensory attribute as the output. The baseline prediction model is obtained through training using a deep neural network algorithm, specifically including: In the data preprocessing stage, the various data in the first component characteristic spectrum are first systematically processed. Retention time data are processed using a min-max normalization method, mapping the original retention time values ​​to a range of 0 to 1. Specifically, the actual retention time of each characteristic peak is recorded, the shortest and longest retention times among all samples are identified, and each retention time value is converted into a standardized value using a linear transformation formula. Mass spectrometry characteristic peak data are based on the first 10 characteristic ions and their relative abundances. The mass-to-charge ratio of each characteristic ion is segmented and encoded, dividing the mass-to-charge ratio range into 10 intervals, and then uniquely encoded according to the interval to which the ion belongs. Relative intensity data are processed using a logarithmic transformation followed by normalization. The original intensity values ​​are first logarithmically calculated to base 10, and then the resulting values ​​are mapped to a range of 0 to 1. After these data processing steps are completed, the processing results of the three types of data are concatenated in a fixed order to form a 120-dimensional multidimensional feature vector, with each vector corresponding to the complete component characteristic information of a sample.

[0038] The deep neural network architecture employs a multi-branch design. The feature extraction branch consists of three fully connected layers: the first layer contains 256 neurons, the second layer contains 128 neurons, and the third layer contains 64 neurons. Each layer uses a modified linear unit (MRU) as the activation function. The association analysis branch uses a bidirectional long short-term memory (LSTM) network structure with 32 memory units to capture the temporal relationships between features. The weight allocation branch consists of two fully connected layers: the first layer contains 64 neurons, and the second layer contains 32 neurons, using a soft maximum activation function to output feature weights. The outputs of the three branches are merged in a concatenation layer and then fused through two fully connected layers: the first fusion layer contains 128 neurons, and the second fusion layer contains 64 neurons. The network concludes with an output layer containing four neurons, each corresponding to one of the four sensory attribute quantification values. Cross-layer connections are established between the branches: the output of the first layer of the feature extraction branch is directly connected to the input of the second layer of the weight allocation branch, and the output of the last time step of the association analysis branch is directly connected to the input of the second fusion layer.

[0039] The model training process is implemented in three phases. In the pre-training phase, an adaptive moment estimation algorithm is used, with an initial learning rate of 0.001 and a batch size of 32. An early stopping strategy is employed during training; pre-training terminates when the validation set loss no longer decreases for 10 consecutive training epochs. In the backpropagation optimization phase, mean squared error is used as the loss function. The gradient of each layer is calculated using the chain rule, and gradient clipping is used to limit the gradient norm to within 1.0. Weight updates employ stochastic gradient descent with momentum, a momentum parameter of 0.9, and a 5% decay in the learning rate every 20 epochs. In the fine-tuning phase, a layered learning rate strategy is used, with a smaller learning rate of 0.0001 for the lower-level parameters and a larger learning rate of 0.001 for the upper-level parameters. A cyclic learning rate scheduling is used, with each cycle containing 50 training epochs.

[0040] Model optimization and validation employ a k-fold cross-validation method, randomly dividing the dataset into five mutually exclusive subsets. Each iteration uses four subsets as the training set and one subset as the validation set, repeating this process five times to ensure each subset is used as the validation set once. Feature importance analysis uses a permutation importance method, randomly shuffling the values ​​of individual features and observing the degree of model performance degradation; features with more significant performance degradation have higher importance. Network structure optimization is achieved through stepwise pruning, first removing the lowest 10% of feature connections, retraining the model, and evaluating its performance. This process is iterated until the model performance degradation exceeds 5%. The final model performance evaluation uses both the coefficient of determination (COD) and root mean square error (RMSE) metrics, requiring an average COD of at least 0.85 and an average RMSE of no more than 0.15 across all folds.

[0041] In S4, multi-source information was collected from real grapefruit hydrosol flower tea samples, including obtaining the second component characteristic spectrum of the real samples using volatile component extraction technology, and obtaining the sensory attribute quantification values ​​of the real samples through sensory quantitative analysis, specifically including: In the volatile component enrichment stage, a sample of real grapefruit hydrosol flower tea was placed in a constant-temperature shaker for dynamic headspace trapping. The specific procedures were as follows: 5 mL of sample was placed in a 20 mL headspace vial, and the vial opening was sealed with a PTFE / silicone septum. The headspace vial was placed in a constant-temperature shaker, with the temperature controlled at 40 ± 0.5 degrees Celsius and the shaking frequency set to 250 rpm. The nitrogen flow rate was precisely controlled at 50 mL / min using a mass flow controller, and the trapping time was 30 minutes. The sample temperature was kept constant during the trapping process to ensure the stable release of volatile components. After trapping, the trapping tube was purged with high-purity nitrogen for 10 minutes to remove residual moisture and carbon dioxide.

[0042] In the thermal desorption separation stage, programmed temperature ramping conditions were used for the analysis of volatile components. The collection tube was placed in the thermal desorption apparatus, with an initial temperature of 35°C, and the temperature was increased to 250°C at a rate of 10°C / min, held for 5 minutes. The desorbed volatile components entered a multi-stage cold trap focusing apparatus via a transfer line, with the transfer line temperature maintained at 200°C. The multi-stage cold traps were cooled with liquid nitrogen. The first stage cold trap was at -30°C for moisture removal; the second stage cold trap was at -50°C for initial focusing; and the third stage cold trap was at -150°C for deep focusing of the volatile components. After focusing, the temperature was rapidly increased to 250°C at a rate of 15°C / s to instantly analyze the volatile components and introduce them into the chromatographic system. The entire thermal desorption process was carried out under high-purity helium carrier gas at a flow rate of 2.0 mL / min.

[0043] During the chromatographic separation stage, a chromatographic system equipped with a composite column was used for analysis. The system consisted of two capillary columns connected in series: the first was a DB-5MS column (30 m × 0.25 mm × 0.25 μm), and the second was a DB-WAX column (30 m × 0.25 mm × 0.25 μm). The column oven temperature program was set as follows: initial temperature 40°C, held for 2 minutes; increased to 100°C at 3°C / min; increased to 180°C at 5°C / min; and finally increased to 250°C at 10°C / min, held for 10 minutes. High-purity helium was used as the carrier gas in constant flow mode at a flow rate of 1.2 mL / min. The injection port temperature was set to 250°C, using splitless injection mode with a split ratio of 10:1.

[0044] During the mass spectrometry detection and characteristic spectrum establishment phase, a high-resolution mass spectrometer detector was used for analysis. The mass spectrometry conditions were set as follows: ion source temperature 230°C, interface temperature 280°C, ionization mode electron impact ionization, electron energy 70 electron volts, mass scan range 35-450 atomic mass units, and scan rate 5 times per second. The detected mass spectrometry data underwent peak identification, peak area integration, and background subtraction. When establishing the second component characteristic spectrum, the retention time of each chromatographic peak was recorded to an accuracy of 0.01 minutes. The top 5 characteristic ion fragments and their relative abundances were selected, and the relative abundances were normalized with the base peak set to 100. All data was stored in a standardized format, including retention time index, characteristic ion mass-to-charge ratio, and relative abundance information, forming a complete second component characteristic spectrum database.

[0045] In S5, based on the multi-source information of the baseline prediction model and real samples, the parameters of the baseline prediction model are adapted to the feature distribution to obtain an enhanced prediction model, specifically including: In the feature mapping function establishment stage, it is first necessary to quantify the distribution differences between simulated matrix samples and real samples in the feature space. The specific operation process is as follows: Feature vectors from 100 samples are randomly selected from the simulated matrix samples, and feature vectors from 50 samples are selected from the real samples. The mean vector and covariance matrix of the two sets of feature vectors are calculated, and the difference between the two distributions is calculated using the Mahalanobis distance formula. Based on this difference, a linear transformation matrix is ​​constructed, with a dimension of 120×120, consistent with the dimension of the feature vectors. The transformation matrix is ​​optimized using the least squares method, aiming to minimize the Euclidean distance between the transformed simulated matrix feature vectors and the real sample feature vectors. The calculated transformation matrix is ​​applied to the input layer of the benchmark prediction model to achieve the projection transformation of the simulated domain feature representation to the real sample feature space. This projection transformation process is automatically executed during model forward propagation, ensuring that the input features maintain distribution consistency with the target domain.

[0046] In the hierarchical progressive optimization phase, a strategy of gradually adjusting network parameters in three levels is adopted. The first level optimizes the parameters of the first two layers of the network, including the first fully connected layer of the feature extraction branch and the input layer of the association analysis branch. The second level extends to the middle layers of the network, including the second layer of the feature extraction branch, the hidden layer of the association analysis branch, and the first layer of the weight allocation branch. The third level covers all the remaining parameters of the network, including the third layer of the feature extraction branch, the output layer of the association analysis branch, the second layer of the weight allocation branch, and all fusion layers. Each level of optimization uses stochastic gradient descent with momentum. At each level transition, the model's performance on a real sample validation set is verified to ensure stable improvement in model performance.

[0047] In the domain adaptation loss function establishment and optimization phase, two aspects of loss calculation need to be considered simultaneously. First, the prediction loss is calculated using the mean squared error function to determine the difference between the model's predicted value and the quantified value of the real sensory attribute. Second, the domain difference loss is measured by calculating the maximum mean difference between the feature distributions of the simulated domain and the real domain in the model's hidden layers. The maximum mean difference is calculated using a radial basis function kernel, with the kernel width parameter set to half the median of the feature dimensions. The final domain adaptation loss function is a weighted average of the prediction loss and the domain difference loss, with weights of 0.7 and 0.3, respectively. Based on this composite loss function, the backpropagation algorithm is used to calculate the gradients of the parameters in each layer. During gradient calculation, the gradient of the domain difference loss term is normalized to prevent gradient explosion. Parameter updates use an adaptive learning rate method, with an initial learning rate set to 0.001, dynamically adjusted according to the loss decrease. The entire optimization process continues until the feature distribution difference between the two domains falls below a preset threshold of 0.05, or the maximum number of iterations is reached. After each iteration, the model's coefficient of determination and root mean square error on the real sample test set are evaluated to ensure that the model performance meets the application requirements.

[0048] In S6, the volatile component characteristics of the grapefruit hydrosol flower tea sample to be tested are extracted to obtain the target component characteristic spectrum. The target component characteristic spectrum is input into the enhanced prediction model, and the comprehensive sensory attribute quantification value of the sample to be tested is output as the final detection result of the aroma active substance content, specifically including: In the volatile component extraction stage, a constant-temperature oscillation equilibration process was first performed. 5 mL of the grapefruit hydrosol flower tea sample was placed in a 20 mL headspace vial, and the vial opening was sealed with a PTFE / silicone septum. The vial was placed in a constant-temperature oscillator, with the temperature set at 45°C, the oscillation frequency at 300 rpm, and the equilibration time at 20 minutes. Precise temperature control ensured that the volatile components reached a stable equilibrium in the headspace. The temperature control accuracy was maintained within ±0.2°C, and the oscillation amplitude was set to 5 mm to ensure thorough mixing of the sample and uniform distribution of the volatile components.

[0049] Next, ultrasonic-assisted extraction and microwave-induced release were performed. The equilibrated sample vials were placed in the ultrasonic-microwave synergistic extraction device. The ultrasonic parameters were set as follows: frequency 40 kHz, power 200 W, and treatment time 5 min. The two excitation methods were alternated, with ultrasonic treatment followed by microwave treatment. This synergistic effect effectively activated volatile components in the sample with different boiling point ranges, especially those with strong aroma compounds. The sample temperature was monitored in real time during the process to ensure it did not exceed 50 degrees Celsius.

[0050] Then, characteristic peak alignment correction is performed. The obtained target component characteristic spectrum is compared with the pre-established first and second component characteristic spectra. The alignment process uses retention time as a reference, with a time window set to 0.2 minutes. For each characteristic peak, its retention time deviation from the corresponding peak in the reference characteristic spectrum is calculated, and a linear interpolation method is used for time axis correction. Simultaneously, mass spectrometry characteristics are compared to ensure correct peak matching. For systematic errors caused by instrument drift, compensation is provided by establishing a retention time correction curve, which is based on the retention time variation patterns of 10 internal standards.

[0051] Finally, baseline correction and peak area normalization were performed. Baseline correction employed asymmetric least squares, using iterative calculations to determine the optimal baseline position. First, the baseline position was estimated, then the fitting parameters were adjusted, and this process was repeated until the baseline converged. Peak area normalization used the total peak area as a baseline, converting the area of ​​each feature peak into a relative percentage. Simultaneously, data format conversion was performed to ensure that the generated target component feature spectrum was completely consistent with the input dimension of the baseline prediction model, forming a 120-dimensional feature vector. After all preprocessing was completed, the quality of the target component feature spectrum was assessed, including signal-to-noise ratio calculation and peak shape symmetry checks, to ensure data reliability.

[0052] In the feature enhancement stage, multi-source feature spectrum fusion is performed first. The target component feature spectrum is fused with the first and second component feature spectra at the feature level. The similarity weights between the target feature spectrum and each reference feature spectrum are calculated using a cosine similarity algorithm, calculating the similarity values ​​between the target feature spectrum and the first and second component feature spectra respectively. Weights are assigned based on the similarity calculation results, with weights ranging from 0 to 1, and the sum of the two weights being 1. The target feature spectrum is then weighted and fused with the two reference feature spectra separately to generate two enhanced feature spectra.

[0053] Next, dynamic weight allocation is implemented. Based on the matching degree between the target component feature spectrum and the historical feature spectrum, the contribution weights of each feature dimension in the enhanced prediction model are dynamically adjusted. The matching degree is evaluated by calculating the Euclidean distance between the target feature spectrum and each sample in the historical feature spectrum database. The 20 historical samples with the smallest distances are selected as references, and the importance ranking of each feature dimension in these samples is analyzed. Based on the importance ranking results, different weights are assigned to each feature dimension, with higher weights given to feature dimensions with higher importance. The weights are calculated using an exponential decay function, with the highest weight not exceeding 1.5 times the standard weight and the lowest weight not less than 0.7 times the standard weight.

[0054] Then, a dual-domain alignment transformation is performed. The target component feature spectrum is projected onto the reference feature space established using the first component feature spectrum and the true feature space constructed using the second component feature spectrum. First, the target feature spectrum is projected onto the reference feature space, and distribution alignment is achieved through a linear transformation matrix. Then, the target feature spectrum is projected onto the true feature space using the same alignment method. The two projection results retain their respective feature representations, forming dual-domain feature representations. These feature representations incorporate both the stability of the reference feature space and the adaptability of the true feature space.

[0055] Finally, the output is calculated using an enhanced prediction model. The processed dual-domain feature representation is input into the enhanced prediction model. The model first extracts features, processing different types of feature representations through three branch networks. Then, feature fusion is performed, concatenating the outputs of each branch. Finally, a fully connected layer is used for calculation, outputting quantified values ​​of four sensory attributes, corresponding to fruit aroma intensity, floral aroma characteristics, sweetness, and freshness. The output results are standardized to ensure that the quantified value of each attribute is between 0 and 100, ultimately serving as the detection result for the content of aroma active substances.

[0056] Please see Figure 2 As shown, the spectral-based system for detecting the content of aroma active substances in grapefruit hydrosol flower tea includes: A simulated matrix construction module is used to construct a matrix simulation system. By systematically mixing aqueous medium, acidity regulator, sugar components and phenolic substances, a simulated matrix sample library with different component ratios is established. A multi-source information acquisition module is used to acquire multi-source information from a simulated matrix sample library, including obtaining the first component characteristic spectrum of each simulated matrix using volatile component extraction technology, and obtaining the corresponding sensory attribute quantification values ​​through sensory quantification analysis. A benchmark model training module is used to establish a benchmark prediction model. The first component feature spectrum is used as the input of the benchmark prediction model, and the sensory attribute quantification value is used as the output of the benchmark prediction model. The benchmark prediction model is obtained by training through a deep neural network algorithm. The real sample adaptation module is used to collect multi-source information from real grapefruit hydrosol flower tea samples, including obtaining the second component characteristic spectrum of the real sample by using volatile component extraction technology, and obtaining the sensory attribute quantification value of the real sample through sensory quantification analysis. The migration enhancement and optimization module adapts the feature distribution of the parameters of the benchmark prediction model based on the multi-source information of the benchmark prediction model and the real sample to obtain the enhanced prediction model. The sensory attribute prediction module is used to extract the volatile component characteristics of the grapefruit hydrosol flower tea sample to be tested, obtain the target component feature spectrum, input the target component feature spectrum into the enhanced prediction model, and output the comprehensive sensory attribute quantification value of the sample to be tested as the final detection result of the aroma active substance content.

[0057] The working principle of this invention is as follows: By constructing a simulated matrix sample library and collecting its first component feature spectrum and sensory attribute quantification value, a benchmark prediction model based on deep neural network is established; then, the second component feature spectrum and sensory attribute quantification value of real samples are collected, and an enhanced prediction model is obtained through feature distribution adaptation and transfer learning optimization; finally, by extracting the volatile component features of the sample to be tested, the obtained target component feature spectrum is input into the enhanced prediction model, and the comprehensive sensory attribute quantification value is directly output as the detection result of the aroma active substance content, realizing accurate mapping and rapid detection from chemical component features to sensory attributes.

[0058] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectroscopy, characterized in that, Includes the following steps: S1: Construct a matrix simulation system by systematically mixing aqueous medium, acidity regulator, sugar components and phenolic substances to establish a simulated matrix sample library with different component ratios; S2: Multi-source information collection is performed on the simulated matrix sample library, including obtaining the first component characteristic spectrum of each simulated matrix using volatile component extraction technology, and obtaining the corresponding sensory attribute quantification values ​​through sensory quantification analysis. S3: Establish a baseline prediction model, using the first component feature spectrum as the input and the quantified value of the sensory attribute as the output. The baseline prediction model is obtained through training using a deep neural network algorithm, specifically including: The retention time, mass spectrometry characteristic peaks, and relative intensity data in the first component feature spectrum are normalized to construct a multidimensional feature vector; A deep neural network architecture is constructed, adopting a multi-branch network structure design, including feature extraction branch, association analysis branch, and weight allocation branch. Each branch realizes the interactive transmission of feature information through cross-layer connections. The network parameters are initialized using pre-training, the network weights are optimized using the backpropagation algorithm, and finally the hyperparameters of the benchmark prediction model are adjusted using a fine-tuning strategy. The model performance is evaluated using cross-validation, key feature parameters are selected through feature importance analysis, and the network structure is iteratively optimized until the preset accuracy requirements are met. S4: Multi-source information collection was carried out on real grapefruit hydrosol flower tea samples, including obtaining the second component characteristic spectrum of real samples by using volatile component extraction technology, and obtaining the sensory attribute quantitative values ​​of real samples through sensory quantitative analysis. S5: Based on the multi-source information of the benchmark prediction model and real samples, the parameters of the benchmark prediction model are adapted to the feature distribution to obtain the enhanced prediction model. Specifically, this includes: by calculating the distribution difference between the simulated matrix sample and the real sample in the component feature spectrum space, establishing a feature mapping function, and projecting the feature representation in the benchmark prediction model to the feature space of the real sample. While keeping the core architecture of the benchmark prediction model unchanged, a hierarchical progressive optimization strategy is adopted to adjust the parameter weights from the shallow to the deep layers of the network in turn, so that the benchmark prediction model gradually adapts to the feature distribution law of real samples. By calculating the difference in feature distribution output by the model in the simulated domain and the real domain, a domain adaptation loss function is established, and the network parameters are iteratively optimized based on the backpropagation algorithm to minimize the difference in feature distribution between the two domains. S6: Extract the volatile component characteristics of the grapefruit hydrosol flower tea sample to obtain the target component characteristic spectrum. Input the target component characteristic spectrum into the enhanced prediction model, and output the comprehensive sensory attribute quantification value of the sample as the final detection result of the aroma active substance content, specifically including: The target component feature spectrum is fused with the first component feature spectrum and the second component feature spectrum at the feature level, and feature enhancement is achieved by calculating feature similarity weights. Based on the degree of matching between the target component feature spectrum and the historical feature spectrum, the contribution weight of each feature dimension in the enhanced prediction model is dynamically adjusted. The target component feature spectrum is subjected to a dual-domain alignment transformation using the reference feature space established by the first component feature spectrum and the real feature space constructed by the second component feature spectrum. By enhancing the multi-level neural network calculation of the prediction model, a comprehensive sensory attribute quantification value including fruit aroma intensity, floral aroma characteristics, sweetness, and freshness is output.

2. The method for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectroscopy according to claim 1, characterized in that, The constructed matrix simulation system specifically includes: A basic aqueous medium is established, using tertiary purified water as the solvent base. The conductivity of the aqueous medium is controlled within a preset range through ion exchange and reverse osmosis treatment, and solvent deoxygenation treatment is carried out under inert gas protection. A complex component system was configured. According to the natural proportion range of each component in pomelo hydrosol flower tea, the organic acid component, sugar component and polyphenolic substance were accurately weighed separately. Gradient dilution and isothermal mixing technology were used to achieve stepwise mixing of each component under constant temperature conditions through a microfluidic device. A sample library variation system was constructed. Based on the orthogonal experimental design method, the ratio and concentration of each component in the simulated matrix were systematically changed. By introducing nonlinear concentration gradient changes, a simulated matrix sample sequence with continuous concentration changes was constructed. The precise component composition parameters of each sample were recorded.

3. The method for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectroscopy according to claim 1, characterized in that, The process of obtaining the first component feature spectrum is as follows: A dynamic headspace extraction device was used to capture volatile components. The simulated matrix sample was placed in a constant temperature oscillating reactor and captured by passing through a multi-stage cold trap at a preset flow rate under inert gas, thereby achieving low-temperature enrichment of volatile components. Programmed temperature rise analysis is implemented to transfer the captured volatile components for thermal desorption under strict control of the heating rate. The separation and release of components with different boiling point ranges are achieved through multi-stage temperature gradient control. The analyzed volatile components were introduced into a coupled chromatographic system and separated using a combination of chromatographic columns. The mass spectrometric characteristics of each component were obtained by a mass spectrometer detector. The obtained mass spectrometry data were processed by peak alignment, denoising and normalization to establish a first component characteristic spectrum containing retention time, mass spectrometry characteristic peaks and relative intensities.

4. The method for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectroscopy according to claim 1, characterized in that, The process for obtaining the quantified values ​​of the sensory attributes is as follows: Establish a standardized sensory evaluation environment in a laboratory with constant temperature and humidity and purified air, using a sensory evaluation room with specific lighting conditions to eliminate the interference of environmental factors on sensory judgment. The organization selected sensory evaluators and conducted sensory calibration using standard samples to establish a sensory attribute evaluation system that includes fruit aroma intensity, floral aroma characteristics, sweetness level, and freshness. During the evaluation process, an electronic recording system is used to record the curves of the intensity of each sensory attribute changing over time in real time, thereby obtaining the dynamic response characteristics of the sensory attributes. The recorded sensory intensity data is normalized and checked for consistency to generate quantitative values ​​of sensory attributes.

5. The method for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectroscopy according to claim 1, characterized in that, The process of obtaining the second component characteristic spectrum is as follows: Real grapefruit hydrosol flower tea samples were placed in a constant temperature shaker and dynamically headspaced under strictly controlled temperature conditions. The volatile components were enriched using a collection tube filled with a composite adsorbent. The enriched volatile components are thermally desorbed under programmed temperature rise conditions, and low-temperature focusing and rapid desorption of volatile components are achieved through a multi-stage cold trap focusing device. The analyzed volatile components are introduced into a chromatographic system equipped with a composite chromatographic column, and the efficient separation of different volatile components is achieved by using a gradient temperature program and precise flow control. Mass spectra of each separated component were obtained using a mass spectrometer, and a second component characteristic spectrum containing retention time index, characteristic ion fragments, and relative abundance information was established.

6. The method for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectroscopy according to claim 1, characterized in that, The extraction of volatile component characteristics from the grapefruit hydrosol flower tea sample to be tested specifically includes: The sample to be tested is placed in a constant temperature oscillation device, and the volatile components are brought to equilibrium in the headspace by precisely controlling the oscillation frequency and temperature parameters. By combining ultrasonic-assisted extraction and microwave-induced release, volatile components with different boiling point ranges in the sample are activated simultaneously. The obtained target component characteristic spectrum is aligned and corrected with the first component characteristic spectrum and the second component characteristic spectrum to eliminate the detection error caused by instrument drift. The corrected feature spectrum is subjected to baseline correction and peak area normalization to establish a target component feature spectrum consistent with the input dimension of the benchmark prediction model.

7. A spectral-based system for detecting the content of active aroma compounds in grapefruit hydrosol and floral tea, characterized in that, The method for detecting the content of aroma active substances in grapefruit hydrosol flower tea based on spectral analysis as described in any one of claims 1-6 includes: A simulated matrix construction module is used to construct a matrix simulation system. By systematically mixing aqueous medium, acidity regulator, sugar components and phenolic substances, a simulated matrix sample library with different component ratios is established. A multi-source information acquisition module is used to acquire multi-source information from a simulated matrix sample library, including obtaining the first component characteristic spectrum of each simulated matrix using volatile component extraction technology, and obtaining the corresponding sensory attribute quantification values ​​through sensory quantification analysis. A benchmark model training module is used to establish a benchmark prediction model. The first component feature spectrum is used as the input of the benchmark prediction model, and the sensory attribute quantification value is used as the output of the benchmark prediction model. The benchmark prediction model is obtained by training through a deep neural network algorithm. The real sample adaptation module is used to collect multi-source information from real grapefruit hydrosol flower tea samples, including obtaining the second component characteristic spectrum of the real sample by using volatile component extraction technology, and obtaining the sensory attribute quantification value of the real sample through sensory quantification analysis. The migration enhancement and optimization module adapts the feature distribution of the parameters of the benchmark prediction model based on the multi-source information of the benchmark prediction model and the real sample to obtain the enhanced prediction model. The sensory attribute prediction module is used to extract the volatile component characteristics of the grapefruit hydrosol flower tea sample to be tested, obtain the target component feature spectrum, input the target component feature spectrum into the enhanced prediction model, and output the comprehensive sensory attribute quantification value of the sample to be tested as the final detection result of the aroma active substance content.

Images