Method and system for integrated analysis of multiple batch fermentation data for curd product evaluation

Abstract

The application discloses a multi-batch fermentation data integrated analysis method and system for evaluation of Quella products, relates to the technical field of fermentation, and comprises the following steps: collecting environmental physicochemical factors, microbial abundance and flavor component data in different stages of the fermentation process of Quella products in multiple regions and multiple batches, constructing a standardized data matrix after pretreatment, then screening key flavor substances by analyzing the activity values of flavor components to establish a flavor evaluation basis; determining key microorganisms and environmental factors affecting fermentation quality and screening representative microbial combinations by combining correlation analysis of microbial communities and flavor characteristics; and establishing a data mapping relationship based on a neural network to realize prediction and evaluation of the quality of Quella products and regulation of the fermentation process. The application solves the technical problem of insufficient accuracy of Quella quality evaluation in the prior art, and achieves the technical effect of improving the accuracy and stability of Quella quality evaluation.

Description

Technical Field

[0001] This invention relates to the field of fermentation technology, and more specifically to a method and system for integrated analysis of multi-batch fermentation data for the evaluation of trehal products. Background Technology

[0002] Traditional Tibetan chutney (or chutney) is a distinctive fermented dairy product made primarily from yak milk using traditional natural fermentation techniques. Its quality is influenced by multiple factors, including the fermentation environment, microbial community structure, and flavor metabolism processes of the Tibetan plateau region. Due to the vast territory of Tibet, differences exist in altitude, climate, raw milk characteristics, traditional production processes, fermentation time, and fermentation environment across different regions. This results in significant regional variations in the microbial community composition, metabolic activity, and characteristic flavor compounds of chutney from different areas. Even within the same region of Tibet, different production batches may exhibit quality differences due to fluctuations in raw material conditions and the natural fermentation process, leading to a lack of stability in the flavor characteristics and quality of traditional Tibetan chutney. Existing chutney quality evaluation methods struggle to comprehensively analyze the environmental physicochemical factors, microbial communities, and flavor components of chutney from different regions and batches within Tibet. They also fail to accurately identify the key microorganisms and core environmental factors causing quality differences across regions, thus affecting the accuracy, consistency, and regional applicability of traditional Tibetan chutney quality evaluations. Summary of the Invention

[0003] This application provides a method and system for integrated analysis of multi-batch fermentation data for the evaluation of trellite products, which is intended to address the technical problem of insufficient accuracy in the evaluation of trellite quality in existing technologies.

[0004] In view of the above problems, this application provides a method and system for integrated analysis of multi-batch fermentation data for the evaluation of trell products.

[0005] A first aspect of this application provides a method for integrated analysis of multi-batch fermentation data for evaluating Tria products, the method comprising: Samples of traditional cheese from multiple regions and batches were collected at different stages of fermentation. Dynamic environmental physicochemical factors, relative microbial abundance, and flavor component content were measured for each batch. Data preprocessing was performed to construct a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix. The odor and taste activity values ​​of each flavor component in the standardized flavor substance matrix were calculated. Core flavor compounds with activity values ​​greater than or equal to a preset activity threshold were extracted, and a characteristic flavor fingerprint standard matrix was constructed as a benchmark for cheese evaluation. Using the characteristic flavor fingerprint standard matrix as the reconstruction target and the standardized microbial community matrix as the reconstruction base, multidimensional reconstruction correlation was performed under nonnegativity and normalized unit simplex constraints. The nonnegative linear reconstruction coefficients of each microbial genus in the reconstruction base with respect to the reconstruction target were calculated. The standardized environmental physicochemical matrix was used as an explanatory variable to evaluate the standard... Redundancy analysis was performed on the microbial community matrix to determine the core environmental control factors. Based on the ratio of the signal power of the reconstructed target in the characteristic flavor space to the corresponding reconstruction error power, decibel processing was performed to determine the synergistic fermentation capacity index of each subset of the microbial community to be evaluated. Combining the non-negative linear reconstruction coefficients, spatial diversity and independence screening of the microbial community subsets were conducted according to a balanced selection strategy, eliminating abnormal batches and redundant genera with the same function, and extracting the target marker microbial genus combination that maximizes the synergistic fermentation capacity index while minimizing the sample size. The abundance data of the target marker microbial genus combination was used as the input layer of a neural network, and the index components in the characteristic flavor fingerprint standard matrix were used as the output layer for nonlinear mapping network training to construct a quality prediction model. Based on the quality prediction model and the core environmental control factors, quality evaluation and fermentation closed-loop correction were performed on unknown batches of Tria.

[0006] A second aspect of this application provides a multi-batch fermentation data integration and analysis system for evaluating Tria products, the system comprising: The data acquisition module collects samples from multiple regions and batches of traditional chebula at different stages of fermentation, measuring the dynamic environmental physicochemical factors, relative abundance of microorganisms, and content of flavor components for each batch. Data preprocessing is performed to construct a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix. The calculation module calculates the odor and taste activity values ​​of each flavor component in the standardized flavor substance matrix, extracts core flavor compounds with activity values ​​greater than or equal to a preset activity threshold, and constructs a characteristic flavor fingerprint standard matrix as a benchmark for chebula evaluation. The reconstruction and correlation module uses the characteristic flavor fingerprint standard matrix as the reconstruction target and the standardized microbial community matrix as the reconstruction basis, performing multidimensional reconstruction and correlation under nonnegativity and normalized unit simplex constraints, calculating the nonnegative linear reconstruction coefficients of each microbial genus in the reconstruction basis with respect to the reconstruction target. The redundancy analysis module uses the standardized environmental physicochemical matrix as an explanatory variable to perform multidimensional reconstruction and correlation on the chebula. The standardized microbial community matrix undergoes redundancy analysis to determine core environmental control factors. A decibelization module is used to decibelize the signal power of the reconstructed target within the characteristic flavor space to the corresponding reconstruction error power, determining the synergistic fermentation capacity index of each subset of the microbial community to be evaluated. A screening module, combining the non-negative linear reconstruction coefficients, uses a balanced selection strategy to screen the microbial community subsets for spatial diversity and independence, eliminating abnormal batches and redundant genera with the same function, and extracting the target marker microbial genus combination that maximizes the synergistic fermentation capacity index while minimizing the sample size. A quality evaluation module uses the abundance data of the target marker microbial genus combination as the input layer of a neural network, and the index components in the characteristic flavor fingerprint standard matrix as the output layer to train a nonlinear mapping network, constructing a quality prediction model. Based on the quality prediction model and the core environmental control factors, it performs quality evaluation and fermentation closed-loop correction for unknown batches of Tria.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application collected samples of traditional cheese from multiple regions and batches at different stages of fermentation. Dynamic environmental physicochemical factors, relative microbial abundance, and flavor component content were measured for each batch. Data preprocessing was performed to construct a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix. The aroma and taste activity values ​​of each flavor component in the standardized flavor substance matrix were calculated. Core flavor compounds with activity values ​​greater than or equal to a preset activity threshold were extracted, and a characteristic flavor fingerprint standard matrix was constructed as a benchmark for cheese evaluation. Using the characteristic flavor fingerprint standard matrix as the reconstruction target and the standardized microbial community matrix as the reconstruction base, multidimensional reconstruction correlation was performed under nonnegativity and normalized unit simplex constraints. The nonnegative linear reconstruction coefficients of each microbial genus in the reconstruction base with respect to the reconstruction target were calculated. The standardized environmental physicochemical matrix was used as an explanatory variable to evaluate the standard... Redundancy analysis is performed on the standardized microbial community matrix to determine core environmental control factors. Based on the ratio of the signal power of the reconstructed target in the characteristic flavor space to the corresponding reconstruction error power, decibel processing is performed to determine the synergistic fermentation capacity index of each subset of the microbial community to be evaluated. Combining the non-negative linear reconstruction coefficients, spatial diversity and independence screening of the microbial community subsets are conducted according to a balanced selection strategy, eliminating abnormal batches and redundant genera with the same function, and extracting the target marker microbial genus combination that maximizes the synergistic fermentation capacity index while minimizing the sample size. The abundance data of the target marker microbial genus combination is used as the input layer of a neural network, and the index components in the standard matrix of the characteristic flavor fingerprint are used as the output layer for nonlinear mapping network training to construct a quality prediction model. Based on the quality prediction model and the core environmental control factors, quality evaluation and fermentation closed-loop correction are performed on unknown batches of cheese. This invention solves the technical problem of insufficient accuracy in cheese quality evaluation in the prior art by constructing a flavor fingerprint spectrum and correlating the analysis of microbial communities and fermentation characteristics, thereby improving the accuracy and stability of cheese quality evaluation. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of a method for integrating and analyzing multi-batch fermentation data for evaluating Tria products, provided in an embodiment of this application. Figure 2The training and validation error convergence curve of the quality prediction model in the multi-batch fermentation data integration and analysis method for evaluating Tria products provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of a multi-batch fermentation data integration and analysis system for evaluating Tria products provided in an embodiment of this application.

[0010] Figure labeling: Data acquisition module 11, calculation module 12, reconstruction and correlation module 13, redundancy analysis module 14, decibel processing module 15, screening module 16, quality evaluation module 17. Detailed Implementation

[0011] This application provides a method and system for integrated analysis of multi-batch fermentation data for the evaluation of truffle products. It addresses the technical problem of insufficient accuracy in the evaluation of truffle quality in existing technologies by constructing a flavor fingerprint spectrum and correlating the microbial community with fermentation characteristics, thereby improving the accuracy and stability of truffle quality evaluation.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, this application provides a method for integrated analysis of multi-batch fermentation data for the evaluation of trehal products, the method comprising: Step S100: Collect samples from multiple regions and batches of traditional curry at different stages of fermentation, and determine the dynamic environmental physicochemical factors, relative abundance of microorganisms, and content of flavor components for each batch of samples. Perform data preprocessing and construct a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix.

[0015] In this embodiment, samples from multiple regions and batches of traditional curry were collected at the early, middle, and mature fermentation stages. Dynamic environmental physicochemical factors, relative abundance of microorganisms, and flavor component content of each batch of samples were obtained. The dynamic environmental physicochemical factors included pH value, dynamic moisture content, types and absolute contents of organic acids, and free amino acid content. The relative abundance of microorganisms was obtained through microbial community detection. The flavor component content was detected by headspace solid-phase microextraction combined with gas chromatography-mass spectrometry to obtain flavor component data including short-chain fatty acid compounds, ester compounds, and active free amino acids.

[0016] Next, the relative abundance data of microorganisms in each batch of samples were screened by abundance threshold and detection frequency threshold. Zero-value filling and sparse filtering were performed on the low abundance and low frequency microbial data to obtain the relative abundance data of cleaned microorganisms. The self-scaling method was used to center the mean and normalize the standard deviation of the relative abundance of cleaned microorganisms, dynamic environmental physicochemical factors and flavor component content, respectively, to construct a standardized microbial community matrix, a standardized environmental physicochemical matrix and a standardized flavor substance matrix.

[0017] Furthermore, the method provided in the application embodiments also includes: The environmental physicochemical factors include real-time pH value measured by a pH meter, dynamic moisture content measured by drying method, types and absolute contents of organic acids determined by liquid chromatography, and free amino acid content determined by pre-column derivatization reversed-phase high-performance liquid chromatography-ultraviolet detection method.

[0018] In this embodiment, each batch of samples was homogenized and the real-time pH value was measured at the same detection temperature. Before measurement, the pH meter was calibrated at two or three points using a standard buffer solution covering the expected pH range. After calibration, the pH electrode was rinsed with deionized water and surface residual liquid was removed. The pH electrode was then inserted into each batch of homogenized samples, ensuring full contact between the sensitive part of the electrode and the sample. The real-time pH value was recorded after the reading stabilized. Each batch of samples underwent at least three parallel measurements. The arithmetic mean of the sum of the real-time pH values ​​from each measurement and the result was used as the real-time pH value of that batch of samples. Dynamic moisture content was determined using a drying method. First, the weighing dish was dried to constant weight and the mass of the weighing dish was recorded. Then, a predetermined amount of each batch of samples was added to the weighing dish, and the total mass of the weighing dish and each batch of samples before drying was recorded. Subsequently, constant-temperature drying was performed at a set temperature. After drying, the dish was placed in a desiccator to cool to room temperature and weighed. The drying, cooling, and weighing process was repeated until two consecutive weighing results met the constant weight requirement. The water loss mass is obtained by subtracting the total mass after drying from the total mass before drying. The mass of each batch of samples before drying is obtained by subtracting the mass of the weighing dish from the total mass before drying. The dynamic moisture content is obtained by dividing the water loss mass by the mass of each batch of samples before drying and multiplying by 100%.

[0019] The types and absolute contents of organic acids were determined by liquid chromatography (LC). Pre-determined amounts of each batch of sample were weighed, and a fixed volume of organic acid extraction solution was added. After homogenization, shaking, or ultrasonic extraction, the samples were centrifuged, and the supernatant was collected and filtered through a microporous membrane. Multiple concentration gradient standard solutions containing the target organic acid were prepared, and the standard solutions and filtered supernatants were analyzed sequentially under the same LC conditions. The retention times of each chromatographic peak were compared with the retention times of the corresponding organic acid standards to determine the type of organic acid. A standard curve was established with the concentration of the organic acid standard solution as the x-axis and the corresponding chromatographic peak area as the y-axis. Based on the chromatographic peak area of ​​the corresponding organic acid in the filtered supernatant and the slope and intercept of the standard curve, the concentration of the organic acid in the filtered supernatant was calculated. The absolute content of each organic acid was obtained by multiplying the concentration by the final volume of the organic acid extraction solution and the total dilution factor of the pretreatment process, and then dividing by the mass of each batch of sample weighed during extraction.

[0020] The content of free amino acids was determined by pre-column derivatization reversed-phase high-performance liquid chromatography-ultraviolet detection. Pre-determined amounts of each batch of samples were weighed, and a fixed volume of free amino acid extraction solution was added for homogenization and extraction. After protein removal, the samples were centrifuged, and the supernatant was collected and filtered through a microporous membrane. Pre-column derivatization reagent was added to the filtered supernatant at a fixed volume ratio, and the derivatization reaction was completed at the set temperature and reaction time. All free amino acid standard solutions were treated under the same pre-column derivatization conditions and detected under the same reversed-phase high-performance liquid chromatography separation conditions and ultraviolet detection wavelength. The types of free amino acids are determined based on the correspondence between the retention times of each chromatographic peak and the free amino acid standard. A standard curve is established using the concentration of the free amino acid standard solution and the corresponding chromatographic peak area. The concentration of each free amino acid is calculated based on the corresponding chromatographic peak area in each batch of samples. The concentration is then multiplied by the final volume of the free amino acid extraction solution and the total dilution factor of the pretreatment, derivatization, and injection processes, and divided by the mass of each batch of samples weighed during extraction to obtain the content of each free amino acid. When the total content of free amino acids is required, the contents of each detected free amino acid are added together.

[0021] Furthermore, the method provided in the application embodiments also includes: The content of the flavor components was detected using headspace solid-phase microextraction combined with gas chromatography-mass spectrometry. The core flavor compounds included short-chain fatty acid compounds, ester compounds, and active free amino acids.

[0022] In this embodiment, after homogenizing each batch of samples, equal amounts of each batch were weighed and placed in headspace vials of identical specifications. An equal amount of internal standard was added and the vials were sealed. Headspace equilibration was performed at the same temperature, oscillation rate, and equilibration time to ensure a stable distribution of volatile flavor components between the sample and the headspace gas phase. After headspace equilibration, the activated solid-phase microextraction fiber was exposed to the headspace gas phase, and flavor components were adsorbed and enriched at the same extraction temperature and time. After extraction, the solid-phase microextraction fiber was immediately inserted into the gas chromatograph inlet for thermal desorption at the set desorption temperature and time. The desorbed flavor components entered the capillary column with the carrier gas, and separation was achieved based on the differences in the distribution characteristics of each flavor component between the stationary and mobile phases. The separated flavor components sequentially entered the mass spectrometer detector, where corresponding characteristic ions and mass spectra were obtained through electron impact ionization. The retention time, retention index, and characteristic ions of each chromatographic peak were compared with the corresponding information in short-chain fatty acid compound standards, ester compound standards, and standard mass spectrometry libraries to determine the types of short-chain fatty acid compounds and ester compounds.

[0023] Standard solutions of short-chain fatty acids and esters at multiple concentration levels were prepared and analyzed under the same headspace equilibrium, solid-phase microextraction, thermal desorption, and gas chromatography-mass spectrometry (GC-MS) conditions as for each batch of samples. Standard curves were established by comparing the concentration of the corresponding compound in each standard solution with the ratio of its peak area to the peak area of ​​the internal standard, obtaining the slope and intercept of the standard curves. The ratio of the peak area of ​​the corresponding compound to the peak area of ​​the internal standard in each batch of samples was substituted into the corresponding standard curve. The intercept of the standard curve was subtracted from the ratio, and then divided by the slope of the standard curve to obtain the concentration of the corresponding compound. The concentration was multiplied by the total volume and dilution factor corresponding to the previous processing steps, and then divided by the amount of each batch of samples weighed during the analysis to obtain the flavor component content of the corresponding short-chain fatty acids and esters. Parallel analyses were performed on the same sample, and the results were summed and divided by the number of parallel analyses to obtain the average value of the corresponding flavor component content.

[0024] The content of active free amino acids was determined using the aforementioned pre-column derivatization reversed-phase high-performance liquid chromatography-ultraviolet detection method. The corresponding components were determined based on the retention time of the active free amino acid standards. A standard curve was established using the concentration of the active free amino acid standard solution and the corresponding chromatographic peak area. The concentration of active free amino acids was calculated by substituting the corresponding chromatographic peak area from each batch of samples into the standard curve. The concentration was then multiplied by the total volume and total dilution factor corresponding to the extraction process, and divided by the amount of each batch of samples weighed during detection to obtain the content of active free amino acids. The contents of short-chain fatty acids, esters, and active free amino acids from the same batch and fermentation stage were correspondingly grouped to form core flavor compound content data.

[0025] Furthermore, the method provided in the application embodiments, which involves performing data preprocessing to construct a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix, also includes: For data in each batch of samples where the relative abundance of microorganisms is lower than a preset abundance threshold and the detection frequency is lower than a preset frequency threshold, zero-value filling and sparse filtering are performed to obtain the relative abundance of cleaned microorganisms. The relative abundance of cleaned microorganisms, dynamic environmental physicochemical factors, and flavor component content are respectively processed by mean centering and standard deviation normalization using the self-scaling method to obtain a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix.

[0026] In this embodiment, the relative abundance of microorganisms in each batch of samples is first organized according to the same batch order and fermentation stage order, forming relative abundance data with each batch of samples as rows and each microbial genus as columns. The relative abundance of each microorganism is compared with a preset abundance threshold. When the relative abundance of a corresponding microorganism is lower than the preset abundance threshold, its relative abundance is assigned to zero, completing zero-value filling. After zero-value filling, the detection frequency of each microbial genus in all batches of samples is calculated. The detection frequency is calculated by dividing the number of batches of samples with a non-zero relative abundance of the corresponding microbial genus by the total number of batches of samples. The obtained detection frequency is compared with a preset frequency threshold. Sparse filtering is performed on microbial genera with detection frequencies lower than the preset frequency threshold, deleting the relative abundance of the corresponding microbial genus in all batches of samples, and retaining the data corresponding to the remaining microbial genera to obtain the cleaned microbial relative abundance.

[0027] Next, self-standardization was performed on the relative abundance of cleaning microorganisms, dynamic environmental physicochemical factors, and flavor component content. For each microbial genus in the relative abundance of cleaning microorganisms, the arithmetic mean and sample standard deviation of that genus were calculated across all batches of samples. The arithmetic mean was subtracted from the relative abundance of microorganisms in each batch of samples to center the mean, and then the centered value was divided by the sample standard deviation to normalize the standard deviation. For each dynamic environmental physicochemical factor and each flavor component, the arithmetic mean and sample standard deviation were calculated across all batches of samples in the same manner. The arithmetic mean of the corresponding dynamic environmental physicochemical factor or flavor component content in each batch of samples was subtracted from the corresponding content, and then divided by the corresponding sample standard deviation. The sample standard deviation was obtained by taking the square root of the sum of the squares of the differences between each value and the arithmetic mean, divided by the number of batches minus one. Dynamic environmental physicochemical factors, microbial genus, or flavor components with a standard deviation of zero were not subject to standard deviation normalization.

[0028] After self-scaling, the results were arranged in the same batch and fermentation stage order. A standardized environmental physicochemical matrix was constructed with each batch of samples as rows and dynamic environmental physicochemical factors as columns. A standardized microbial community matrix was constructed with each batch of samples as rows and the microbial genera retained after sparse filtration as columns. A standardized flavor substance matrix was constructed with each batch of samples as rows and flavor components as columns. Each row in the standardized environmental physicochemical matrix, standardized microbial community matrix, and standardized flavor substance matrix corresponds to samples from the same batch and at the same fermentation stage.

[0029] Step S200: Calculate the odor activity value and taste activity value of each flavor component in the standardized flavor matrix, extract the core flavor compounds with an activity value greater than or equal to the preset activity threshold, and construct a characteristic flavor fingerprint spectrum standard matrix as a benchmark for the Triller evaluation standard.

[0030] In this embodiment, the absolute concentrations of each flavor component are first extracted from a standardized flavor matrix, and the sensory olfactory and taste thresholds of each flavor component in the dairy product medium are retrieved. The absolute concentrations of each flavor component are divided by their corresponding sensory olfactory and taste thresholds to calculate the odor activity value and taste activity value of each flavor component. The odor activity value and taste activity value are compared with preset activity thresholds, and flavor components with odor activity values ​​and / or taste activity values ​​greater than or equal to the preset activity thresholds are extracted. The extracted flavor components are then identified as core flavor compounds.

[0031] Next, the proportion distribution relationship of each core flavor compound is calculated according to the content of each core flavor compound in different batches and different fermentation stages. The content of each core flavor compound is divided by the sum of the contents of all core flavor compounds in the corresponding sample to obtain the proportion value of each core flavor compound. Then, with each batch of samples as rows and each core flavor compound as columns, the corresponding proportion values ​​are written into a matrix and the proportion distribution relationship of the core flavor compounds is solidified to construct a characteristic flavor fingerprint spectrum standard matrix as a benchmark for the evaluation of the core flavor compounds.

[0032] Furthermore, in the method provided in the application embodiments, calculating the odor activity value and taste activity value of each flavor component in the standardized flavor matrix further includes: The sensory olfactory threshold and taste threshold of each flavor component in the standardized flavor matrix in the dairy product medium are retrieved; the ratio of the absolute concentration of each flavor component to the corresponding sensory olfactory threshold and taste threshold is calculated to obtain the corresponding odor activity value and taste activity value; the characteristic components with odor activity value and / or taste activity value greater than or equal to the preset activity threshold are extracted as the core flavor compounds, and the proportional distribution relationship of the core flavor compounds is solidified into the characteristic flavor fingerprint spectrum standard matrix.

[0033] In this embodiment, the absolute concentration of each flavor component is first obtained according to the correspondence between batches of samples, fermentation stages, and flavor components in the standardized flavor matrix. Then, the sensory olfactory threshold and taste threshold of the corresponding flavor component in the dairy product medium are retrieved. The sensory olfactory threshold and taste threshold are recorded using the same concentration unit as the absolute concentration of each flavor component, and are associated one-to-one with each flavor component name. For a flavor component with a sensory olfactory threshold, the absolute concentration of the flavor component is divided by the corresponding sensory olfactory threshold, and the resulting ratio is used as the odor activity value. For a flavor component with a taste threshold, the absolute concentration of the flavor component is divided by the corresponding taste threshold, and the resulting ratio is used as the taste activity value. When the same flavor component has both sensory olfactory and taste thresholds, the corresponding odor activity value and taste activity value are calculated and recorded separately.

[0034] Next, for each batch of samples and each fermentation stage, the odor activity value and taste activity value of each flavor component are compared with preset activity thresholds. When the odor activity value is greater than or equal to the preset activity threshold, or the taste activity value is greater than or equal to the preset activity threshold, the corresponding characteristic component is extracted as the core flavor compound; when both the odor activity value and the taste activity value are greater than or equal to the preset activity threshold, the corresponding characteristic component is also selected as the core flavor compound. This comparison method is used to iterate through each flavor component in the standardized flavor matrix to obtain the core flavor compounds for each batch of samples and each fermentation stage.

[0035] Finally, the absolute concentrations of each core flavor compound in the same batch and fermentation stage were extracted. The absolute concentrations of all core flavor compounds were summed to obtain the total absolute concentration of core flavor compounds in the corresponding batch and fermentation stage. Then, the absolute concentration of each core flavor compound was divided by the total absolute concentration of core flavor compounds to obtain the proportion of the corresponding core flavor compound. Using each batch of samples and each fermentation stage as rows and core flavor compounds as columns, the proportion of each core flavor compound was written into the corresponding position, so that the sum of the proportions of each core flavor compound in each row is 1. This forms the proportion distribution relationship of core flavor compounds, and the proportion distribution relationship is solidified into a standard matrix of characteristic flavor fingerprint spectrum.

[0036] Step S300: Using the standard matrix of the characteristic flavor fingerprint spectrum as the reconstruction target and the standardized microbial community matrix as the reconstruction basis, perform multidimensional reconstruction correlation under the constraints of nonnegativity and normalization of the unit simplex, and calculate the nonnegative linear reconstruction coefficients of each microbial genus in the reconstruction basis with respect to the reconstruction target.

[0037] In this embodiment, the characteristic flavor fingerprint spectrum standard matrix is ​​used as the reconstruction target, and the data columns corresponding to each microbial genera in the standardized microbial community matrix are used as the reconstruction basis. Reconstruction coefficients are assigned to each microbial genera, and the reconstruction result is formed by a linear combination of the reconstruction basis and the reconstruction coefficients. The difference between the reconstruction target and the reconstruction result is calculated, and the elements in the difference are squared and summed. The iteration objective is to minimize the sum of squared errors, continuously updating the reconstruction coefficients corresponding to each microbial genera until the sum of squared errors reaches the convergence condition.

[0038] During the error squared minimization iteration, a non-negativity constraint is applied to each reconstruction coefficient, limiting those less than zero to zero to maintain non-negativity. Then, the sum of all non-negative reconstruction coefficients is calculated, and each non-negative reconstruction coefficient is divided by this sum, ensuring that the normalized sum of all reconstruction coefficients equals 1. The reconstruction coefficient that satisfies the non-negativity and normalization constraints of the unit simplex and minimizes the error squared sum is determined as the non-negative linear reconstruction coefficient of each microbial genus in the reconstruction substrate for the reconstruction target.

[0039] Furthermore, in the method provided in the application embodiments, the characteristic flavor fingerprint spectrum standard matrix is ​​used as the reconstruction target, the standardized microbial community matrix is ​​used as the reconstruction basis, and multidimensional reconstruction correlation is performed under the constraints of nonnegativity and normalization of the unit simplex. The method also calculates the nonnegative linear reconstruction coefficients of each microbial genus in the reconstruction basis with respect to the reconstruction target, and further includes: The linear combination of the reconstruction target and the reconstruction basis is subjected to an iterative calculation to minimize the squared error. During the iterative calculation, a non-negative constraint is applied to each reconstruction coefficient in the linear combination to forcibly remove negative coefficients so that all coefficients meet the requirement of positive promotion of microbial fermentation metabolism. A normalization constraint with a sum of 1 is applied to all non-negative reconstruction coefficients to eliminate the interference of background noise on the gain of multidimensional fitting, thereby obtaining the non-negative linear reconstruction coefficients.

[0040] In this embodiment, the characteristic flavor fingerprint matrix is ​​first used as the reconstruction target, and the standardized microbial community matrix is ​​used as the reconstruction base. Rows of both are mapped one-to-one according to the same batch and fermentation stage. Assuming the standardized microbial community matrix includes multiple microbial genera and the characteristic flavor fingerprint matrix includes multiple index components, a corresponding reconstruction coefficient is set for each microbial genera and each index component, forming a reconstruction coefficient matrix. The initial reconstruction coefficients are weighted equally according to the number of microbial genera; that is, each initial reconstruction coefficient corresponding to each index component is set to the reciprocal of the number of microbial genera, ensuring that all initial reconstruction coefficients corresponding to the same index component are non-negative and sum to 1. The standardized microbial community matrix is ​​multiplied by the reconstruction coefficient matrix to obtain a linear combination of the reconstruction base. The linear combination is then subtracted element-by-element from the reconstruction target, and the sum of the squared differences is obtained to obtain the sum of squared errors. The reconstruction coefficients are updated round by round according to the direction of change of the sum of squared errors relative to each reconstruction coefficient. After each round of updates, the linear combination and the sum of squared errors are recalculated. When the updated sum of squared errors is greater than the sum of squared errors before the update, the iteration step size is reduced and the current round of updates is re-executed until the sum of squared errors no longer increases. In this way, the minimum squared error calculation is performed.

[0041] In each round of iterative calculation, a non-negativity constraint is applied to each updated reconstructed coefficient. Each reconstructed coefficient is checked for being less than zero, and those less than zero are set to zero. Negative coefficients are forcibly removed from the linear combination of the reconstructed basis, ensuring that all remaining reconstructed coefficients are greater than or equal to zero. After completing the non-negativity constraint, the linear combination of the reconstructed basis and the corresponding sum of squared errors are recalculated using the processed reconstructed coefficients. This recalculated sum of squared errors serves as the basis for the next round of iterations, ensuring that all coefficients meet the requirement of positively promoting microbial fermentation metabolism.

[0042] Finally, sum all non-negative reconstruction coefficients corresponding to each index component, and divide each non-negative reconstruction coefficient by the sum of its corresponding non-negative reconstruction coefficients to make the sum of all normalized non-negative reconstruction coefficients equal to 1. This applies a normalized unit simplex constraint to all non-negative reconstruction coefficients. If the sum of the reconstruction coefficients after non-negative constraint processing is zero, then each reconstruction coefficient is reset to the reciprocal of the number of microbial genera, and the iteration step size is reduced before continuing the iteration. In each iteration, the error square minimization update, non-negativity constraint, and normalization constraint with a sum of 1 are executed sequentially until the difference between the sums of squared errors of two adjacent iterations is less than a preset convergence threshold, or the preset maximum number of iterations is reached. The reconstruction coefficients that satisfy the non-negativity constraint, the normalization constraint with a sum of 1, and have the minimum corresponding error square sum at the end of the iteration are determined as the non-negative linear reconstruction coefficients of each microbial genera for the reconstruction target, in order to eliminate the interference of background noise on the gain of multidimensional fitting.

[0043] Step S400: Using the standardized environmental physicochemical matrix as an explanatory variable, perform redundancy analysis on the standardized microbial community matrix to determine the core environmental control factors.

[0044] In this embodiment, a standardized environmental physicochemical matrix is ​​used as an explanatory variable, and a standardized microbial community matrix is ​​used as a response variable. Correspondence is established according to the same batch and the same fermentation stage. Redundancy analysis is used to calculate the correlation degree and significance between each environmental physicochemical factor and the community abundance index. Based on the redundancy analysis results, environmental physicochemical factors that meet the preset correlation and significance conditions with the community abundance index are screened, and the screened environmental physicochemical factors are determined as core environmental control factors.

[0045] Furthermore, in the method provided in the application embodiments, the standardized environmental physicochemical matrix is ​​used as an explanatory variable to perform redundancy analysis on the standardized microbial community matrix to determine core environmental control factors, and the method further includes: Each environmental physicochemical factor in the standardized environmental physicochemical matrix and the microbial community abundance index in the standardized microbial community matrix are extracted respectively; the Pearson correlation coefficient and significance value between each environmental physicochemical factor and the microbial community abundance index are calculated; the environmental physicochemical factors with Pearson correlation coefficients greater than or equal to a preset correlation threshold and significance values ​​less than a preset significance threshold are extracted as the core environmental control factors.

[0046] In this embodiment, a one-to-one correspondence is established between the rows of the standardized environmental physicochemical matrix and the standardized microbial community matrix, following the same batch sample order and fermentation stage order. Each column of the standardized environmental physicochemical matrix is ​​read sequentially, and the real-time pH value, dynamic moisture content, absolute organic acid content, or absolute free amino acid content corresponding to each column is determined as the corresponding environmental physicochemical factor. The standardized values ​​of these environmental physicochemical factors are then extracted across all batch samples and each fermentation stage. Similarly, each column of the standardized microbial community matrix is ​​read sequentially, and the standardized relative abundance of the corresponding microbial genus in each column across all batch samples and each fermentation stage is determined as the community abundance index. For each environmental physicochemical factor and each community abundance index, a row-by-row correspondence is established according to the same batch samples and the same fermentation stage. Data with valid values ​​in both are retained, and the number of valid data points used in subsequent calculations is recorded.

[0047] For any environmental physicochemical factor and any bacterial abundance index, calculate the arithmetic mean of all valid values ​​for both the environmental physicochemical factor and the bacterial abundance index. Subtract the corresponding arithmetic mean from each environmental physicochemical factor value, and subtract the corresponding arithmetic mean from each bacterial abundance index value in the same row. Multiply the two differences obtained in the same row and sum all products. Squash and sum the differences for each environmental physicochemical factor and each bacterial abundance index, respectively. Take the square root of the product of the two sums of squares, and divide the sum of the product of the differences by the square root of the product of the two sums of squares to obtain the corresponding Pearson correlation coefficient. The significance value is calculated based on the Pearson correlation coefficient and the number of valid data points. The number of valid data points is subtracted by 2, the result is divided by 1, and the square of the Pearson correlation coefficient is subtracted. The square root of the resulting ratio is then multiplied by the absolute value of the Pearson correlation coefficient to obtain the t-test statistic. The number of valid data points is subtracted by 2 to determine the degrees of freedom. Based on the Student t-distribution corresponding to these degrees of freedom, the one-sided tail probability greater than or equal to the t-test statistic is calculated. This one-sided tail probability is then multiplied by 2 to obtain the two-sided probability, which is then determined as the significance value. This yields the Pearson correlation coefficients and significance values ​​corresponding to each environmental physicochemical factor and each bacterial abundance index.

[0048] Finally, the Pearson correlation coefficients between each environmental physicochemical factor and each bacterial abundance index were compared with preset correlation thresholds, and the corresponding significance values ​​were compared with preset significance thresholds. When the Pearson correlation coefficient was greater than or equal to the preset correlation threshold and the corresponding significance value was less than the preset significance threshold, the corresponding environmental physicochemical factor was extracted. When the same environmental physicochemical factor and multiple bacterial abundance indicators met the above conditions, the environmental physicochemical factor was retained once, and all its corresponding Pearson correlation coefficients and significance values ​​were recorded. After completing the item-by-item comparison of all environmental physicochemical factors and all bacterial abundance indicators, the extracted environmental physicochemical factors were determined as core environmental control factors.

[0049] Step S500: Based on the ratio of the signal power of the reconstructed target in the characteristic flavor space to the corresponding reconstruction error power, perform decibel processing to determine the synergistic fermentation ability index of each subset of microbial communities to be evaluated.

[0050] In this embodiment, when performing decibel processing based on the ratio of the signal power of the reconstructed target in the characteristic flavor space to the corresponding reconstruction error power, the microbial feature matrix corresponding to each subset of the microbial community to be evaluated is extracted from the standardized microbial community matrix, and the reconstruction fitting vector of the microbial feature matrix to the reconstructed target is solved under the constraints of nonnegativity and normalization of the unit simplex. The signal power in the characteristic flavor space is obtained by squaring and summing each element in the reconstructed target. At the same time, the difference between the reconstructed target and the reconstruction fitting vector is calculated, and the reconstruction error power is obtained by squaring and summing each difference. Finally, the signal power is divided by the reconstruction error power, and the obtained ratio is logarithmically transformed to base 10 and multiplied by 10 to obtain the microbial community synergistic fermentation ability index corresponding to each subset of the microbial community to be evaluated.

[0051] Furthermore, in the method provided in the application embodiments, the ratio of the signal power of the reconstructed target in the characteristic flavor space to the corresponding reconstruction error power is processed in decibels to determine the synergistic fermentation ability index of each subset of the microbial community to be evaluated, and the method further includes: Microbial feature matrices corresponding to any subset of microbial communities to be evaluated are extracted from the standardized microbial community matrix, and their reconstruction fitting vectors to the reconstruction target are solved according to the unit simplex constraint. The square norm of the reconstruction target is calculated as the signal power in the characteristic flavor space, and the square norm of the difference between the reconstruction target and the reconstruction fitting vector is calculated as the reconstruction error power. The ratio of the signal power to the reconstruction error power is calculated, and the ratio is logarithmically transformed and multiplied by 10 to obtain the synergistic fermentation ability index of each subset of microbial communities to be evaluated.

[0052] In this embodiment, the genera of microorganisms included in any subset of the microbial community to be evaluated are determined from the standardized microbial community matrix. Data columns corresponding to each genera are extracted according to their column positions in the standardized microbial community matrix, and these data columns are arranged according to the batch sample order and fermentation stage order to form a microbial feature matrix corresponding to the subset of the microbial community to be evaluated. Assuming the subset of the microbial community to be evaluated contains m genera of microorganisms, a reconstruction coefficient vector containing m reconstruction coefficients is established, ensuring that each reconstruction coefficient corresponds one-to-one with each data column in the microbial feature matrix. The initial value of each reconstruction coefficient is set to 1 / m, ensuring that each reconstruction coefficient is greater than or equal to 0 and the sum of all reconstruction coefficients equals 1. The microbial feature matrix is ​​multiplied by the reconstruction coefficient vector to obtain the current reconstruction fitting vector. The difference between the current reconstruction fitting vector and each corresponding component in the reconstruction target is calculated, and the squares of each difference are summed to obtain the current reconstruction error. Based on the direction of change of the current reconstruction error relative to each reconstruction coefficient, each reconstruction coefficient is updated according to a preset iteration step size, and the updated reconstruction coefficients are projected onto a unit simplex. The unit simplex constraint means that each reconstruction coefficient is greater than or equal to 0 and the sum of all reconstruction coefficients equals 1. When performing unit simplex projection, the updated reconstruction coefficients are arranged in descending order of value. A common adjustment amount is determined that makes the sum of the adjusted non-negative reconstruction coefficients equal to 1. The common adjustment amount is subtracted from each updated reconstruction coefficient, and results less than 0 are set to 0, resulting in a new round of reconstruction coefficients that satisfy the unit simplex constraint. The reconstruction fitting vector and reconstruction error are recalculated using the new round of reconstruction coefficients. The process of updating reconstruction coefficients, performing unit simplex projection, and calculating reconstruction error is repeated until the difference in reconstruction error between two adjacent iterations is less than a preset convergence threshold, or the number of iterations reaches a preset maximum number of iterations. The vector obtained by multiplying the final reconstruction coefficient vector by the microbial feature matrix is ​​determined as the reconstruction fitting vector of the subset of the microbial community to be evaluated for the reconstruction target.

[0053] Next, following the order of the flavor components in the reconstructed target, each flavor component is read sequentially, squared, and summed to obtain the square norm of the reconstructed target. This square norm is used as the signal power in the flavor space. Following the same order of flavor components, the corresponding fitted component in the reconstructed fitting vector is subtracted from each flavor component in the reconstructed target to obtain the reconstruction error for each flavor component. Each reconstruction error is squared, and the sum of the squares is used to obtain the square norm of the difference between the reconstructed target and the reconstructed fitting vector. This square norm of the difference is used as the reconstruction error power.

[0054] Finally, the signal power is divided by the reconstruction error power to obtain the ratio of signal power to reconstruction error power for the subset of microbial communities to be evaluated. This ratio is then logarithmically transformed to base 10, and the resulting logarithmic value is multiplied by 10 to obtain the synergistic fermentation capacity index of the microbial community subset to be evaluated. Following the same process of microbial feature matrix extraction, reconstruction coefficient initialization, unit simplex constraint solving, reconstruction fitting vector calculation, signal power calculation, reconstruction error power calculation, and logarithmic transformation, each subset of microbial communities to be evaluated is processed sequentially to obtain the synergistic fermentation capacity index corresponding to each subset.

[0055] Step S600: Based on the non-negative linear reconstruction coefficient, the spatial diversity and independence of the microbial community subset are screened according to the balanced selection strategy, abnormal batches and redundant genera with the same function are eliminated, and the target marker microbial genera combination that maximizes the synergistic fermentation ability index of the microbial community and minimizes the sample size is extracted.

[0056] In this embodiment, the non-negative linear reconstruction coefficients are combined to arrange the non-negative linear reconstruction coefficient vectors corresponding to each characteristic flavor component of each microbial genus. For any two microbial genera, the non-negative linear reconstruction coefficients of the two genera on the same characteristic flavor component are subtracted, and the sum and square root of the squared differences are obtained to obtain the Euclidean distance between the two genera. Microbial genera whose Euclidean distance is greater than or equal to a preset spatial distance threshold are retained in the same microbial community subset to complete spatial diversity screening. The standardized relative abundance of any two microbial genera is paired row by row according to the same batch and the same fermentation stage, and the Pearson correlation coefficient between the two is calculated. Microbial genera whose absolute value of the Pearson correlation coefficient is less than a preset independence threshold are retained in the same microbial community subset to complete independence screening.

[0057] For each batch, a reconstructed fitting vector is calculated using the corresponding microbial feature matrix and non-negative linear reconstruction coefficients. The components at the same position in the reconstructed target and the reconstructed fitting vector are subtracted respectively, and the squared differences are summed to obtain the reconstruction error power of that batch. The mean and standard deviation of the reconstruction error power of all batches are calculated. The reconstructed error power of each batch is subtracted from the mean and divided by the standard deviation to obtain the standardized deviation value of the corresponding batch. Batches with standardized deviation values ​​greater than a preset anomaly threshold are identified as abnormal batches, and the data corresponding to abnormal batches are deleted. For any two retained microbial genera, the cosine similarity of their non-negative linear reconstruction coefficient vectors is calculated, and the Pearson correlation coefficient between their standardized relative abundances is calculated. When the cosine similarity is greater than or equal to a preset functional similarity threshold and the absolute value of the Pearson correlation coefficient is greater than or equal to a preset redundancy threshold, the two are identified as functionally redundant genera. One of the microbial genera is deleted, and the synergistic fermentation capacity index of the microbial community is recalculated. The microbial genera with a larger decrease in the synergistic fermentation capacity index after deletion are retained, and the other microbial genera is removed.

[0058] After spatial diversity and independence screening, and the removal of abnormal batches and redundant genera with the same function, the corresponding microbial feature matrices of each microbial community subset are extracted. The reconstructed fitting vector under the unit simplex constraint is resolved, and the corresponding synergistic fermentation capacity index of the microbial community is calculated. At the same time, the number of microbial genera contained in each microbial community subset is counted, and the number of microbial genera is determined as the sample size. According to the balanced selection strategy, the microbial community subset with the maximum value of the synergistic fermentation capacity index is selected first. When the synergistic fermentation capacity indices of multiple microbial community subsets are the same or the difference does not exceed the preset allowable deviation, the microbial community subset with the smallest sample size is selected, and the microbial genera contained in this microbial community subset are determined as the target marker microbial genera combination.

[0059] Step S700: Using the abundance data of the target marker microbial genus combination as the input layer of the neural network, and using the index components in the characteristic flavor fingerprint spectrum standard matrix as the output layer, a nonlinear mapping network is trained to construct a quality prediction model. Based on the quality prediction model and the core environmental control factors, the unknown batch of cheese is evaluated for quality and fermentation closed-loop correction is performed.

[0060] In this embodiment, according to the correspondence between historical batches and fermentation stages, the abundance data of target marker microbial genera combinations in the middle of fermentation of each historical batch are extracted, and an input sample set is formed according to the fixed arrangement order of the target marker microbial genera. The abundance data of the target marker microbial genera combinations is used as the input layer of the neural network. Simultaneously, the indicator components corresponding to the same historical batch finished product stage in the characteristic flavor fingerprint spectrum standard matrix are extracted, and a supervision label set is formed according to the fixed arrangement order of the indicator components. The indicator components in the characteristic flavor fingerprint spectrum standard matrix are used as the output layer. A backpropagation neural network with a single hidden layer is used for training a nonlinear mapping network. The abundance data of the target marker microbial genus combination in the input layer is processed by the weights of the hidden layer and the activation function and then passed to the output layer to obtain the predicted values ​​of each index component. The sum of squares momentum error of the difference between the predicted index component and the supervision label is calculated. Based on the sum of squares momentum error, the weight adjustment is calculated from the output layer to the hidden layer. Combining the current weight adjustment and the previous weight adjustment, the weights of the hidden layer and the output layer are iteratively updated in multiple rounds until the sum of squares momentum error meets the preset convergence condition. The network structure, hidden layer weights and output layer weights formed after network convergence are determined as the quality prediction model.

[0061] After the quality prediction model is constructed, abundance data of the target marker microbial genus assemblage during the mid-fermentation stage of unknown batches of Tria are collected. This data is processed according to the microbial genus arrangement order and standardized parameters used during the training of the quality prediction model to form an input vector for the unknown batch. This input vector is then input into the quality prediction model. The quality prediction model performs a nonlinear mapping on the unknown batch input vector, outputting predicted values ​​for each indicator component at the finished product stage. Predicted indicator component vectors are formed according to the indicator component arrangement order in the characteristic flavor fingerprint standard matrix. The predicted values ​​of each indicator component are subtracted item by item from the corresponding standard values ​​in the characteristic flavor fingerprint standard matrix to obtain the difference between the predicted and standard values. The absolute value of each difference is compared with the corresponding preset allowable deviation. When the absolute value of each difference is not greater than the corresponding preset allowable deviation, the quality of the unknown batch of Tria is determined to meet the requirements. When the absolute value of at least one difference is greater than the corresponding preset allowable deviation, the corresponding indicator component, difference, and direction of the difference are recorded to obtain the quality evaluation result of the unknown batch of Tria.

[0062] When the absolute value of the difference between the predicted value and the corresponding standard value of at least one indicator component exceeds the preset allowable deviation, the target marker microbial genus associated with that indicator component is determined based on the non-negative linear reconstruction coefficient. Then, the core environmental control factors for adjustment are determined based on the Pearson correlation coefficient between the abundance index of the target marker microbial genus and each core environmental control factor. Measured values ​​of the core environmental control factors at the current fermentation stage of the unknown batch of Triazo are collected, compared with the corresponding preset control ranges, and the core environmental control factors are adjusted according to the positive or negative direction of the Pearson correlation coefficient. This ensures that the abundance change of the target marker microbial genus combination is conducive to reducing the difference between the predicted values ​​and the corresponding standard values ​​of each indicator component. After adjustment, the abundance data of the target marker microbial genus combination is re-collected and input into the quality prediction model to obtain the predicted values ​​of each indicator component again. The difference between each predicted value and the corresponding standard value is recalculated. The core environmental control factor adjustment, abundance data collection, quality prediction, and difference comparison are repeated until the absolute value of each difference is no greater than the corresponding preset allowable deviation, thus completing the quality evaluation and fermentation closed-loop correction of the unknown batch of Triazo.

[0063] Furthermore, in the method provided in the application embodiments, using the abundance data of the target marker microbial genus assemblage as the input layer of the neural network, and using the index components in the characteristic flavor fingerprint spectrum standard matrix as the output layer for nonlinear mapping network training to construct a quality prediction model, it further includes: Abundance data of target marker microbial genus combinations corresponding to the mid-fermentation stage of each historical batch are extracted as input sample set, and indicator components corresponding to the finished product stage are extracted as supervision label set. Using a backpropagation neural network with a single hidden layer, the training objective is to minimize the sum of squares and momentum errors between the predicted indicator components and the supervision labels. The weights of the hidden layer are iteratively updated multiple times until the network converges, generating the quality prediction model.

[0064] In this embodiment, a data correspondence is first established according to historical batch numbers and fermentation stages. Abundance data of target marker microbial genus combinations corresponding to the mid-fermentation stage of each historical batch are extracted. Following a fixed arrangement of the target marker microbial genera, the abundance data of each historical batch are arranged as an input sample, forming an input sample set from all historical batches. Simultaneously, indicator components of the finished product stage of the same historical batch are extracted. Following the indicator component arrangement order in the characteristic flavor fingerprint standard matrix, the indicator components of the same historical batch are arranged as a supervision label, forming a supervision label set from all historical batches. Each input sample is paired with its corresponding supervision label according to the historical batch number, ensuring that the abundance data of the target marker microbial genus combinations in each input sample corresponds to the indicator components of the finished product stage of the same historical batch. This yields paired training data for training the backpropagation neural network.

[0065] Next, as Figure 2As shown, the input sample set and its corresponding supervisory label set are divided into a training set and a validation set according to a preset ratio. The training set participates in the weight update of the backpropagation neural network, while the validation set is used to calculate the sum of squares and momentum errors of the network on data that did not participate in the weight update. Using a backpropagation neural network with a single hidden layer, the abundance data of the target marker microbial genus combinations in the training set are input into the input layer of the neural network. The index components from the corresponding supervisory label set are set as the target output of the output layer. The hidden layer weights between the input layer and the single hidden layer, the output layer weights between the single hidden layer and the output layer, and the corresponding biases are initialized. During training, the abundance data in each input sample is first multiplied by the corresponding hidden layer weight. The products pointing to the same single hidden layer node are summed and the corresponding bias is added. Then, the single hidden layer output is processed by an activation function to obtain the single hidden layer output. Subsequently, the single hidden layer output is multiplied by the corresponding output layer weight. The products pointing to the same output layer node are summed and the corresponding bias is added to obtain each predicted index component. Each predicted index component in the training set is subtracted from the supervision label corresponding to the same input sample. The resulting differences are squared and summed to obtain the sum of squared errors. Based on these sums of squared errors, the error gradient corresponding to the hidden layer weights is calculated backward from the output layer to the single hidden layer. The error gradient is multiplied by the learning rate to obtain the hidden layer weight adjustment for this round. The product of the previous round's hidden layer weight adjustment and the momentum coefficient is added to the current round's hidden layer weight adjustment to form the hidden layer weight update amount corresponding to the sum of squared momentum errors. The hidden layer weights are updated accordingly, and the output layer weights and corresponding biases are updated using the same error backpropagation method. After each iteration, the training set and validation set are input into the updated backpropagation neural network. The corresponding predicted index components are obtained through forward computation, and the sum of squared momentum errors of the training set and validation set are calculated separately. The sum of squared momentum errors of the training set and validation set are recorded sequentially according to the iteration rounds to form... Figure 2 The error convergence curves shown are illustrated, where the solid line represents the sum of squares momentum error of the training set, the dashed line represents the sum of squares momentum error of the validation set, and the dotted line represents the iteration number corresponding to when the network reaches the preset convergence condition. Using the updated hidden layer weights, output layer weights, and corresponding biases, the calculation of prediction index components, the sum of squares momentum error, error backpropagation, and weight updates are repeatedly performed to gradually reduce the sum of squares momentum errors of both the training and validation sets with increasing iterations. When the sum of squares momentum error of the training set is less than a preset convergence threshold, or the change in the sum of squares momentum error of the training set between two adjacent iterations is less than a preset change threshold, and the sum of squares momentum error of the validation set remains stable, the backpropagation neural network is considered converged. The converged network structure, hidden layer weights, output layer weights, and corresponding biases are saved to generate the quality prediction model.

[0066] In summary, the embodiments of this application have at least the following technical effects: This application collected samples of traditional cheese from multiple regions and batches at different stages of fermentation. Dynamic environmental physicochemical factors, relative microbial abundance, and flavor component content were measured for each batch. Data preprocessing was performed to construct a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix. The aroma and taste activity values ​​of each flavor component in the standardized flavor substance matrix were calculated. Core flavor compounds with activity values ​​greater than or equal to a preset activity threshold were extracted, and a characteristic flavor fingerprint standard matrix was constructed as a benchmark for cheese evaluation. Using the characteristic flavor fingerprint standard matrix as the reconstruction target and the standardized microbial community matrix as the reconstruction base, multidimensional reconstruction correlation was performed under nonnegativity and normalized unit simplex constraints. The nonnegative linear reconstruction coefficients of each microbial genus in the reconstruction base with respect to the reconstruction target were calculated. The standardized environmental physicochemical matrix was used as an explanatory variable to evaluate the standard... Redundancy analysis is performed on the standardized microbial community matrix to determine core environmental control factors. Based on the ratio of the signal power of the reconstructed target in the characteristic flavor space to the corresponding reconstruction error power, decibel processing is performed to determine the synergistic fermentation capacity index of each subset of the microbial community to be evaluated. Combining the non-negative linear reconstruction coefficients, spatial diversity and independence screening of the microbial community subsets are conducted according to a balanced selection strategy, eliminating abnormal batches and redundant genera with the same function, and extracting the target marker microbial genus combination that maximizes the synergistic fermentation capacity index while minimizing the sample size. The abundance data of the target marker microbial genus combination is used as the input layer of a neural network, and the index components in the standard matrix of the characteristic flavor fingerprint are used as the output layer for nonlinear mapping network training to construct a quality prediction model. Based on the quality prediction model and the core environmental control factors, quality evaluation and fermentation closed-loop correction are performed on unknown batches of cheese. This invention solves the technical problem of insufficient accuracy in cheese quality evaluation in the prior art by constructing a flavor fingerprint spectrum and correlating the analysis of microbial communities and fermentation characteristics, thereby improving the accuracy and stability of cheese quality evaluation.

[0067] Example 2, based on the same inventive concept as the multi-batch fermentation data integration and analysis method for evaluating trelline products in the foregoing examples, such as... Figure 3 As shown, this application provides a multi-batch fermentation data integration and analysis system for evaluating trehalose products. The system and method embodiments in this application are based on the same inventive concept. The system includes: The data acquisition module 11 is used to collect samples from multiple regions and batches of traditional cheese at different stages of fermentation, and to measure the dynamic environmental physicochemical factors, relative abundance of microorganisms, and content of flavor components for each batch of samples. Data preprocessing is performed to construct a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix. The calculation module 12 is used to calculate the odor and taste activity values ​​of each flavor component in the standardized flavor substance matrix, extract core flavor compounds with activity values ​​greater than or equal to a preset activity threshold, and construct a characteristic flavor fingerprint standard matrix as a benchmark for cheese evaluation. The reconstruction and correlation module 13 is used to use the characteristic flavor fingerprint standard matrix as the reconstruction target and the standardized microbial community matrix as the reconstruction basis, performing multidimensional reconstruction and correlation under nonnegativity and normalized unit simplex constraints, and calculating the nonnegative linear reconstruction coefficients of each microbial genus in the reconstruction basis for the reconstruction target. The redundancy analysis module 14 is used to use the standardized environmental physicochemical matrix as an explanatory variable. Redundancy analysis is performed on the standardized microbial community matrix to determine core environmental control factors. A decibelization module 15 is used to decibelize the signal power of the reconstructed target within the characteristic flavor space to the corresponding reconstruction error power, determining the synergistic fermentation capacity index of each subset of the microbial community to be evaluated. A screening module 16 is used to combine the non-negative linear reconstruction coefficients and, according to a balanced selection strategy, screen the microbial community subsets for spatial diversity and independence, eliminating abnormal batches and redundant genera with the same function, and extracting the target marker microbial genus combination that maximizes the synergistic fermentation capacity index while minimizing the sample size. A quality evaluation module 17 is used to train a nonlinear mapping network using the abundance data of the target marker microbial genus combination as the input layer of a neural network and the index components in the characteristic flavor fingerprint standard matrix as the output layer, constructing a quality prediction model. Based on the quality prediction model and the core environmental control factors, quality evaluation and fermentation closed-loop correction are performed on unknown batches of Tria.

[0068] Furthermore, the system is also used to implement the following functions: The environmental physicochemical factors include real-time pH value measured by a pH meter, dynamic moisture content measured by drying method, types and absolute contents of organic acids determined by liquid chromatography, and free amino acid content determined by pre-column derivatization reversed-phase high-performance liquid chromatography-ultraviolet detection method.

[0069] Furthermore, the system is also used to implement the following functions: The content of the flavor components was detected using headspace solid-phase microextraction combined with gas chromatography-mass spectrometry. The core flavor compounds included short-chain fatty acid compounds, ester compounds, and active free amino acids.

[0070] Furthermore, the system is also used to implement the following functions: For data in each batch of samples where the relative abundance of microorganisms is lower than a preset abundance threshold and the detection frequency is lower than a preset frequency threshold, zero-value filling and sparse filtering are performed to obtain the relative abundance of cleaned microorganisms. The relative abundance of cleaned microorganisms, dynamic environmental physicochemical factors, and flavor component content are respectively processed by mean centering and standard deviation normalization using the self-scaling method to obtain a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix.

[0071] Furthermore, the system is also used to implement the following functions: The sensory olfactory threshold and taste threshold of each flavor component in the standardized flavor matrix in the dairy product medium are retrieved; the ratio of the absolute concentration of each flavor component to the corresponding sensory olfactory threshold and taste threshold is calculated to obtain the corresponding odor activity value and taste activity value; the characteristic components with odor activity value and / or taste activity value greater than or equal to the preset activity threshold are extracted as the core flavor compounds, and the proportional distribution relationship of the core flavor compounds is solidified into the characteristic flavor fingerprint spectrum standard matrix.

[0072] Furthermore, the system is also used to implement the following functions: The linear combination of the reconstruction target and the reconstruction basis is subjected to an iterative calculation to minimize the squared error. During the iterative calculation, a non-negative constraint is applied to each reconstruction coefficient in the linear combination to forcibly remove negative coefficients so that all coefficients meet the requirement of positive promotion of microbial fermentation metabolism. A normalization constraint with a sum of 1 is applied to all non-negative reconstruction coefficients to eliminate the interference of background noise on the gain of multidimensional fitting, thereby obtaining the non-negative linear reconstruction coefficients.

[0073] Furthermore, the system is also used to implement the following functions: Each environmental physicochemical factor in the standardized environmental physicochemical matrix and the microbial community abundance index in the standardized microbial community matrix are extracted respectively; the Pearson correlation coefficient and significance value between each environmental physicochemical factor and the microbial community abundance index are calculated; the environmental physicochemical factors with Pearson correlation coefficients greater than or equal to a preset correlation threshold and significance values ​​less than a preset significance threshold are extracted as the core environmental control factors.

[0074] Furthermore, the system is also used to implement the following functions: Microbial feature matrices corresponding to any subset of microbial communities to be evaluated are extracted from the standardized microbial community matrix, and their reconstruction fitting vectors to the reconstruction target are solved according to the unit simplex constraint. The square norm of the reconstruction target is calculated as the signal power in the characteristic flavor space, and the square norm of the difference between the reconstruction target and the reconstruction fitting vector is calculated as the reconstruction error power. The ratio of the signal power to the reconstruction error power is calculated, and the ratio is logarithmically transformed and multiplied by 10 to obtain the synergistic fermentation ability index of each subset of microbial communities to be evaluated.

[0075] Furthermore, the system is also used to implement the following functions: Abundance data of target marker microbial genus combinations corresponding to the mid-fermentation stage of each historical batch are extracted as input sample set, and indicator components corresponding to the finished product stage are extracted as supervision label set. Using a backpropagation neural network with a single hidden layer, the training objective is to minimize the sum of squares and momentum errors between the predicted indicator components and the supervision labels. The weights of the hidden layer are iteratively updated multiple times until the network converges, generating the quality prediction model.

[0076] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for integrated analysis of multi-batch fermentation data for evaluating trehal products, characterized in that, The method includes: Samples of traditional cheesin from multiple regions and batches were collected at different stages of fermentation. The dynamic environmental physicochemical factors, relative abundance of microorganisms, and content of flavor components of each batch of samples were measured. Data preprocessing was performed to construct a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix. The aroma activity value and taste activity value of each flavor component in the standardized flavor matrix are calculated, and the core flavor compounds with activity values ​​greater than or equal to the preset activity threshold are extracted to construct a characteristic flavor fingerprint spectrum standard matrix as a benchmark for the evaluation standard of the Triller. Using the standard matrix of the characteristic flavor fingerprint spectrum as the reconstruction target and the standardized microbial community matrix as the reconstruction basis, multidimensional reconstruction correlation is performed under the constraints of non-negativity and normalized unit simplex, and the non-negative linear reconstruction coefficients of each microbial genus in the reconstruction basis with respect to the reconstruction target are calculated. Using the standardized environmental physicochemical matrix as an explanatory variable, redundancy analysis was performed on the standardized microbial community matrix to determine the core environmental control factors. The ratio of the signal power of the reconstructed target in the characteristic flavor space to the corresponding reconstruction error power is decibelized to determine the synergistic fermentation ability index of each subset of microbial communities to be evaluated. Based on the non-negative linear reconstruction coefficient, the spatial diversity and independence of the microbial community subset are screened according to the balanced selection strategy. Abnormal batches and redundant genera with the same function are eliminated, and the target marker microbial genera combination that maximizes the synergistic fermentation ability index of the microbial community and minimizes the sample size is extracted. The abundance data of the target marker microbial genus combination is used as the input layer of the neural network, and the index components in the characteristic flavor fingerprint spectrum standard matrix are used as the output layer for nonlinear mapping network training to construct a quality prediction model. Based on the quality prediction model and the core environmental control factors, the unknown batch of cheese is evaluated for quality and fermentation closed-loop correction is performed.

2. The method for integrated analysis of multi-batch fermentation data for evaluating Tria products as described in claim 1, characterized in that, The environmental physicochemical factors include real-time pH value measured by a pH meter, dynamic moisture content measured by drying method, types and absolute contents of organic acids determined by liquid chromatography, and free amino acid content determined by pre-column derivatization reversed-phase high-performance liquid chromatography-ultraviolet detection method.

3. The method for integrated analysis of multi-batch fermentation data for evaluating Tria products as described in claim 1, characterized in that, The content of the flavor components was detected using headspace solid-phase microextraction combined with gas chromatography-mass spectrometry. The core flavor compounds included short-chain fatty acid compounds, ester compounds, and active free amino acids.

4. The method for integrated analysis of multi-batch fermentation data for evaluating Tria products as described in claim 1, characterized in that, Perform data preprocessing to construct standardized environmental physicochemical matrices, standardized microbial community matrices, and standardized flavor compound matrices, including: For data in each batch of samples where the relative abundance of microorganisms is lower than a preset abundance threshold and the detection frequency is lower than a preset frequency threshold, zero-value filling and sparse filtering are performed to obtain the relative abundance of cleaned microorganisms. The relative abundance of the cleaning microorganisms, dynamic environmental physicochemical factors, and flavor component content were subjected to mean centering and standard deviation normalization using the self-scaling method to obtain a standardized environmental physicochemical matrix, a standardized microbial community matrix, and a standardized flavor substance matrix.

5. The method for integrated analysis of multi-batch fermentation data for evaluating Tria products as described in claim 1, characterized in that, Calculating the odor activity value and taste activity value of each flavor component in the standardized flavor matrix includes: The sensory olfactory and taste thresholds of each flavor component in the standardized flavor matrix were retrieved in dairy products. Calculate the ratio of the absolute concentration of each flavor component to the corresponding sensory thresholds of smell and taste to obtain the corresponding odor activity value and taste activity value; Characteristic components with odor activity values ​​and / or taste activity values ​​greater than or equal to a preset activity threshold are extracted as the core flavor compounds, and the proportional distribution relationship of the core flavor compounds is solidified into the characteristic flavor fingerprint spectrum standard matrix.

6. The method for integrated analysis of multi-batch fermentation data for evaluating Tria products as described in claim 1, characterized in that, Using the standard matrix of the characteristic flavor fingerprint as the reconstruction target and the standardized microbial community matrix as the reconstruction basis, multidimensional reconstruction correlation is performed under the constraints of nonnegativity and normalized unit simplex. The nonnegative linear reconstruction coefficients of each microbial genus in the reconstruction basis with respect to the reconstruction target are calculated, including: The linear combination of the reconstruction objective and the reconstruction basis is subjected to an iterative calculation to minimize the squared error. During the iterative calculation process, non-negativity constraints are applied to each reconstruction coefficient in the linear combination, and negative coefficients are forcibly eliminated to ensure that all coefficients meet the requirements of positive promotion of microbial fermentation metabolism. A normalization constraint with a sum of 1 is applied to all non-negative reconstruction coefficients to eliminate background noise interference with the gain of the multidimensional fitting, thereby obtaining the non-negative linear reconstruction coefficients.

7. The method for integrated analysis of multi-batch fermentation data for evaluating Tria products as described in claim 1, characterized in that, Using the standardized environmental physicochemical matrix as explanatory variables, redundancy analysis was performed on the standardized microbial community matrix to determine core environmental control factors, including: Each environmental physicochemical factor in the standardized environmental physicochemical matrix and the abundance index of the microbial community in the standardized microbial community matrix are extracted respectively. Calculate the Pearson correlation coefficient and significance value between each environmental physicochemical factor and the bacterial community abundance index; Environmental physicochemical factors whose Pearson correlation coefficient is greater than or equal to a preset correlation threshold and whose significance value is less than a preset significance threshold are extracted as the core environmental control factors.

8. The method for integrated analysis of multi-batch fermentation data for evaluating Tria products as described in claim 1, characterized in that, Based on the ratio of the signal power of the reconstructed target in the characteristic flavor space to the corresponding reconstruction error power, decibel processing is performed to determine the synergistic fermentation capacity index of each subset of the microbial community to be evaluated, including: Extract the microbial feature matrix corresponding to any subset of the microbial community to be evaluated from the standardized microbial community matrix, and solve its reconstruction fitting vector to the reconstruction target according to the unit simplex constraint; The square norm of the reconstruction target is calculated as the signal power in the feature flavor space, and the square norm of the difference between the reconstruction target and the reconstruction fitting vector is calculated as the reconstruction error power. Calculate the ratio of the signal power to the reconstruction error power, perform a logarithmic transformation on the ratio, and multiply by 10 to obtain the synergistic fermentation capacity index of each subset of bacterial communities to be evaluated.

9. The method for integrated analysis of multi-batch fermentation data for evaluating Tria products as described in claim 1, characterized in that, Using the abundance data of the target marker microbial genus assemblage as the input layer of a neural network, and using the index components in the standard matrix of the characteristic flavor fingerprint as the output layer, a nonlinear mapping network is trained to construct a quality prediction model, including: Abundance data of target marker microbial genus combinations corresponding to the mid-stage of fermentation in each historical batch were extracted as the input sample set, and indicator components corresponding to the finished product stage were extracted as the supervision label set. Using a backpropagation neural network with a single hidden layer, the training objective is to minimize the sum of squared momentum errors between the predicted index components and the supervision labels. The hidden layer weights are iteratively updated multiple times until the network converges, thus generating the quality prediction model.

10. A multi-batch fermentation data integration and analysis system for evaluating trehalose products, characterized in that, The system is used to perform the multi-batch fermentation data integration and analysis method for evaluating Tria products as described in any one of claims 1-9, the system comprising: The data acquisition module is used to collect samples of traditional cheesin from multiple regions and batches at different stages of fermentation, and to measure the dynamic environmental physicochemical factors, relative abundance of microorganisms and content of flavor components of each batch of samples. Data preprocessing is performed to construct a standardized environmental physicochemical matrix, a standardized microbial community matrix and a standardized flavor substance matrix. The calculation module is used to calculate the odor activity value and taste activity value of each flavor component in the standardized flavor matrix, extract the core flavor compounds with an activity value greater than or equal to a preset activity threshold, and construct a characteristic flavor fingerprint spectrum standard matrix as a benchmark for the Triller evaluation standard. The reconstruction association module is used to take the standard matrix of the characteristic flavor fingerprint spectrum as the reconstruction target and the standardized microbial community matrix as the reconstruction basis, and to perform multidimensional reconstruction association under the constraints of nonnegativity and normalized unit simplex, and to calculate the nonnegative linear reconstruction coefficients of each microbial genus in the reconstruction basis with respect to the reconstruction target. The redundancy analysis module is used to perform redundancy analysis on the standardized environmental physicochemical matrix as an explanatory variable, and to determine the core environmental control factors. The decibelization module is used to perform decibelization based on the ratio of the signal power of the reconstructed target in the characteristic flavor space to the corresponding reconstruction error power, and to determine the synergistic fermentation ability index of each subset of microbial communities to be evaluated. The screening module is used to combine the non-negative linear reconstruction coefficients and perform spatial diversity and independence screening on the microbial community subset according to the balanced selection strategy, remove abnormal batches and redundant genera with the same function, and extract the target marker microbial genera combination that maximizes the synergistic fermentation ability index of the microbial community and minimizes the sample size. The quality evaluation module is used to train a nonlinear mapping network by using the abundance data of the target marker microbial genus combination as the input layer of the neural network and the index components in the characteristic flavor fingerprint spectrum standard matrix as the output layer, to construct a quality prediction model, and to perform quality evaluation and fermentation closed-loop correction on unknown batches of cheese based on the quality prediction model and the core environmental control factors.

Images