Rice wine food fermentation intelligent regulation method based on electronic sensory information fusion

Abstract

The application discloses a rice wine food fermentation intelligent regulation method based on electronic sensory information fusion, and relates to the technical field of food fermentation. The method comprises the following steps: obtaining sample fermentation data set of target rice wine product, and analyzing the sample fermentation data set to reconstruct response factors; analyzing the sample fermentation data set based on the response factor reconstruction result, obtaining fusion vector norm of the target rice wine product, combining the fusion vector norm with the sample fermentation data set, and establishing a standard fermentation state vector set; collecting electronic sensory information in real time through an electronic nose sensor array, and converting the electronic sensory information into real-time fermentation state vectors through the fusion vector norm; traversing the standard fermentation state vector set, comparing and analyzing the real-time fermentation state vectors, and generating and executing fermentation intelligent regulation according to the comparison and analysis result and a preset multi-dimensional regulation decision matrix. The application realizes adaptive and accurate identification and intelligent closed-loop regulation of the rice wine fermentation process.

Description

Technical Field

[0001] This invention relates to the field of food fermentation technology, specifically to an intelligent control method for rice wine fermentation based on electronic sensory information fusion. Background Technology

[0002] With the continuous development of intelligent technology in the food industry, the control of rice wine fermentation process is gradually shifting from relying on manual experience to monitoring and regulation based on sensor data. Existing technologies mainly use electronic noses to collect volatile substance response signals during fermentation, combined with threshold alarms or fixed pattern recognition, to achieve preliminary monitoring of the fermentation status.

[0003] However, rice wine fermentation is a complex biochemical process involving multiple variables and nonlinear changes. Existing technologies use threshold alarm-based monitoring, which can only reflect the isolated state of a single substance and cannot characterize the dynamic evolution of the overall flavor pattern of fermentation. It is also prone to misjudgment due to sensor drift. Fixed pattern recognition suffers from problems such as pattern rigidity, lack of distinction in the importance of feature dimensions, and disconnect between monitoring and control, resulting in insufficient accuracy in identifying the rice wine fermentation process and making it difficult to meet the needs of refined and intelligent production. Summary of the Invention

[0004] This invention provides an intelligent control method for rice wine fermentation based on electronic sensory information fusion, aiming to solve the technical problem of insufficient accuracy in identifying the rice wine fermentation process in existing technologies.

[0005] In view of the above problems, the present invention provides a method for intelligent control of rice wine fermentation based on electronic sensory information fusion, comprising: Obtain a sample fermentation dataset of the target rice wine product, and parse the sample fermentation dataset to reconstruct response factors; Based on the analysis of the sample fermentation dataset according to the response factor reconstruction results, the fusion vector paradigm of the target rice wine product is obtained, and a standard fermentation state vector set is established by combining the fusion vector paradigm with the sample fermentation dataset. Electronic sensory information is collected in real time by an electronic nose sensor array and converted into a real-time fermentation state vector through the fusion vector paradigm. The standard fermentation state vector set is traversed and compared with the real-time fermentation state vector. Based on the comparison analysis results and the preset multidimensional control decision matrix, intelligent fermentation control is generated and executed.

[0006] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides an intelligent control method for rice wine fermentation based on electronic sensory information fusion. By reconstructing response factors and using a fusion vector paradigm, a standard vector set comprehensively characterizing the fermentation state is constructed, enabling precise modeling of the fermentation characteristics of the target rice wine product. Real-time information acquisition and vector transformation based on an electronic nose sensor array dynamically monitors multi-dimensional sensory changes during the fermentation process. Furthermore, through intelligent comparison and analysis of the real-time fermentation state vector and the standard vector set, combined with a multi-dimensional control decision matrix, optimized control commands are automatically generated and executed. This effectively overcomes the shortcomings of traditional methods, such as single-point monitoring and lagging control, improving the stability, product consistency, and overall controllability of the rice wine fermentation process. It achieves adaptive, precise identification and intelligent closed-loop control of the rice wine fermentation process. Attached Figure Description

[0007] 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.

[0008] Figure 1 This is a flowchart illustrating the intelligent control method for rice wine fermentation based on electronic sensory information fusion provided in an embodiment of the present invention. Detailed Implementation

[0009] This invention provides an intelligent control method for rice wine fermentation based on electronic sensory information fusion, which addresses the technical problem of insufficient accuracy in identifying the rice wine fermentation process in existing technologies.

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0011] It should be noted that the terms "comprising" and "having" are 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 these processes, methods, products, or devices.

[0012] Examples, such as Figure 1 As shown, this invention provides a method for intelligent control of rice wine fermentation based on electronic sensory information fusion, the method comprising: S100: Obtain the sample fermentation dataset of the target rice wine product, and parse the sample fermentation dataset to reconstruct the response factors.

[0013] In this embodiment of the invention, a sample fermentation dataset of the target rice wine product is obtained, and the sample fermentation dataset is analyzed to reconstruct response factors. Rice wine fermentation is a complex process driven by microbial metabolism, and the fermentation quality is determined by the types, concentrations, and proportions of various volatile substances. Traditional fermentation monitoring technologies only focus on isolated signals of single or a few volatile substances, without combining them with the final quality inspection results of the product, and cannot clearly define the correlation between the signals and qualified products; they also ignore the proportional relationships between volatile substances, and it is difficult to comprehensively characterize the complex features of the fermentation state based on a single substance signal, thus affecting the accuracy of subsequent fermentation stage identification and control. To solve the above problems, this step first needs to construct a complete sample fermentation dataset with sensory information and quality information to ensure that the data has correlation and validity; then, through response factor reconstruction, the single volatile substance signal is combined with the proportional signals between substances to enrich the feature dimensions and lay the foundation for the subsequent construction of an accurate fusion vector paradigm.

[0014] Step S100 in the method provided in this embodiment of the invention includes: The process of obtaining a sample fermentation dataset of the target rice wine product includes: Under standard fermentation process, electronic sensory information of samples at multiple fermentation time points is collected by an electronic nose sensor array, wherein the electronic sensory information is the response signal of the electronic nose sensor array to volatile substances; Based on the obtained electronic sensory information of the samples, the corresponding rice wine product quality inspection information is recorded synchronously and combined to form the sample fermentation dataset.

[0015] First, under a standard fermentation process, electronic sensor arrays are used to collect electronic sensory information from samples at multiple fermentation time points. This electronic sensory information refers to the response signals of the electronic nose sensor array to volatile substances. The standard fermentation process refers to a fixed set of fermentation parameters that has been validated through multiple production runs by the company, consistently producing samples that meet industry food safety standards and the company's internal quality control requirements. These parameters include raw material ratios, fermentation temperature, cycle time, and inoculated microbial strains. The electronic nose sensor array is an electronic detection device that mimics the biological olfactory system. It consists of multiple sensors with specific sensitivity to different volatile substances, capable of simultaneously detecting multiple volatile substances and outputting electrical signals. The volatile substance response signal refers to the quantifiable electrical signal generated when the electronic nose sensor comes into contact with a volatile substance, resulting from changes in its physical properties. The signal strength is positively correlated with the concentration of the corresponding substance. The rice wine fermentation process is started according to the standard fermentation process, and parameters such as raw material ratio, fermentation temperature, cycle and inoculated strain are specified. An electronic nose sensor array of preset specifications is selected, and the sensitive volatile substances corresponding to each sensor are identified. The sampling frequency is set and the total sampling time point is determined. The electronic nose is connected to the gas sampling port of the fermentation tank, and a fixed volume of headspace gas is extracted each time. The sensor outputs a response signal and stores it in the data acquisition system.

[0016] For example, taking a certain brand of glutinous rice wine as the target product: Standard fermentation process: 50kg glutinous rice, 60kg purified water, ratio 1:1.2, inoculated with 0.3kg brewing yeast + 0.1kg Rhizopus, fermentation temperature 28℃, cycle 21 days; Electronic nose configuration: 12-channel MOS sensor array, each sensor corresponds to the following sensitive volatile substances: S1-ethanol, S2-acetic acid, S3-ethyl lactate, S4-isoamyl alcohol, the rest correspond to esters / aldehydes; Acquisition parameters: Data was collected every 6 hours, for a total of 85 time points: 0h, 6h, 12h…504h. 50mL of headspace gas was extracted from the tank each time. The results showed that at 6h of fermentation, the response signals were S1=2.3mV, S2=1.1mV, S3=0.8mV, and S4=0.5mV; at 24h of fermentation, the response signals were S1=3.8mV, S2=1.5mV, S3=1.2mV, and S4=0.7mV. Finally, 85 sets of 12-dimensional electronic sensory information were obtained.

[0017] Secondly, based on the acquired electronic sensory information of the samples, the corresponding rice wine product quality inspection information is simultaneously recorded and merged to form the sample fermentation dataset. Rice wine product quality inspection information refers to the set of product quality indicators obtained through physicochemical testing, microbiological testing, and sensory evaluation; it serves as a quality anchor for judging whether the fermentation state is qualified. At each electronic sensory information collection time point, a fixed volume of fermentation liquid sample is simultaneously extracted; three types of quality inspection indicators are detected: physicochemical indicators include alcohol content, total acid, and reducing sugar; microbiological indicators include yeast count; and sensory indicators are scored by professional wine tasters. The electronic sensory information and quality inspection information at the same time point are matched one-to-one and merged in the format of collection time point + electronic sensory signal + quality inspection indicator to form the sample fermentation dataset.

[0018] For example, based on the 85 sets of 12-dimensional electronic sensory information obtained above, 50 mL of fermentation broth was extracted at each collection time point; quality control results: after 6 hours of fermentation, the alcohol content was 2.1% vol, total acid was 1.2 g / L, reducing sugar was 15.3 g / 100 mL, and yeast count was 1.2 × 10⁻⁶. CFU / mL, sensory score 72 points; after 24 hours of fermentation, alcohol content 4.3% vol, total acid 1.8 g / L, reducing sugar 10.5 g / 100 mL, yeast count 3.5 × 10⁻⁶. 6 CFU / mL, sensory score of 78 points; similarly, 85 sets of quality inspection results were obtained. The above quality inspection results were bound to the 12-dimensional electronic sensory signals at the corresponding time points to form 85 complete records. Each record contains a time point + 12-dimensional signal + 5 quality inspection indicators.

[0019] The process of reconstructing response factors by parsing the sample fermentation dataset includes: Analyze the sample fermentation dataset, extract the response signals of volatile substances that occur more frequently than a preset frequency, and output them as the first set of candidate response factors; The first set of candidate response factors is randomly combined in pairs, and the ratio of each random combination of factor pairs is defined as the second candidate response factor to obtain the second set of candidate response factors. The first set of candidate response factors and the second set of candidate response factors are merged and output as the response factor reconstruction result.

[0020] First, the sample fermentation dataset is analyzed to extract volatile substance response signals with a frequency greater than a preset frequency, which are output as the first candidate response factor set. The preset frequency refers to a screening threshold set based on the total number of time points in the sample, used to extract volatile substance signals that occur consistently during fermentation and have a significant impact on quality. The first candidate response factor set is a collection of response signals from a single, frequently occurring volatile substance, reflecting the dynamic characteristics of that single substance. The frequency of occurrence of electronic sensor signals in each dimension of the sample dataset is statistically analyzed: frequency = number of valid detection time points / total number of collection time points; signals with a frequency greater than the preset frequency are filtered; and the filtered signals are integrated to output the first candidate response factor set.

[0021] For example, the preset frequency setting is as follows: a total of 85 acquisition time points, and a preset frequency threshold of 85%; frequency statistics: S1 occurs 98% of the time, S2 occurs 95% of the time, S3 occurs 92% of the time, S4 occurs 88% of the time, and the frequency of the other 8 signals is less than 80%; the result output is: the first set of alternative response factors is: {ethanol response signal S1, acetic acid response signal S2, ethyl lactate response signal S3, isoamyl alcohol response signal S4}.

[0022] Next, the first set of candidate response factors is randomly combined pairwise, and the ratio of each random combination is defined as the second candidate response factor, thus obtaining the second set of candidate response factors. The second set of candidate response factors refers to the set of proportional signals formed by pairwise combinations of signals from the first set of candidate response factors, reflecting the concentration correlation between the two volatile substances. All signals in the first set of candidate response factors are randomly combined pairwise; the preceding / following signal of each combination is defined as the second candidate response factor corresponding to that combination, and the signal ratio indirectly represents the concentration ratio of the substances; all pairwise combinations are traversed and integrated to form the second set of candidate response factors.

[0023] For example, pairwise random combinations: The first set of candidate response factors consists of 4 signals, forming a total of 6 combinations: (S1,S2), (S1,S3), (S1,S4), (S2,S3), (S2,S4), (S3,S4); Proportional calculation: The corresponding second set of candidate response factors are R1=S1 / S2, R2=S1 / S3, R3=S1 / S4, R4=S2 / S3, R5=S2 / S4, R6=S3 / S4; Result output: The second set of candidate response factors is: {R1, R2, R3, R4, R5, R6}.

[0024] Finally, the first and second candidate response factor sets are merged and output as the response factor reconstruction result. The response factor reconstruction result is a comprehensive feature set after merging the first and second candidate response factor sets, which includes both single substance signals and substance ratio signals, thus expanding the feature dimensions. The first and second candidate response factor sets are directly merged; duplicate factors are removed to form a complete set of response factors, i.e., the response factor reconstruction result. For example, merging the first candidate response factor set {S1, S2, S3, S4} with the second candidate response factor set {R1, R2, R3, R4, R5, R6} results in the response factor reconstruction result {S1, S2, S3, S4, R1, R2, R3, R4, R5, R6}, a total of 10 response factors, covering both single substance features and ratio correlation features.

[0025] In this embodiment of the invention, by acquiring a complete sample fermentation dataset, which simultaneously includes electronic sensory information and quality inspection information, data support is provided for subsequently establishing the signal-quality correlation, solving the problem of traditional datasets lacking quality anchors. The response factor reconstruction results not only cover highly stable single volatile substance signals, i.e., the first candidate response factor set, but also supplement with proportional signals reflecting the correlation between substances, i.e., the second candidate response factor set. This breaks through the limitation of traditional techniques relying solely on single signal features and significantly enriches the representational dimensions of fermentation states. Furthermore, by using preset frequency screening and pairwise combinations, the effectiveness and correlation of response factors are ensured, laying a solid foundation for the subsequent construction of the fusion vector paradigm and the establishment of a standard fermentation state vector set, thus improving the robustness and accuracy of the overall method.

[0026] S200: Based on the analysis of the sample fermentation dataset according to the response factor reconstruction results, obtain the fusion vector paradigm of the target rice wine product, and combine the fusion vector paradigm with the sample fermentation dataset to establish a standard fermentation state vector set.

[0027] In this embodiment of the invention, the sample fermentation dataset is analyzed based on the results of response factor reconstruction to obtain the fusion vector paradigm of the target rice wine product. A standard fermentation state vector set is then established by combining the fusion vector paradigm with the sample fermentation dataset. S100 has obtained a comprehensive feature set containing single substance signals and proportional signals through response factor reconstruction, but this set still suffers from dimensional redundancy and indiscriminate feature importance: some response factors contribute little to the characterization of the fermentation state, and directly using them for subsequent analysis would increase computational complexity and introduce interference; different response factors have different weights in their impact on fermentation quality, and without differentiation, key features may be obscured. Furthermore, the sample fermentation dataset contains both qualified and unqualified samples, requiring the selection of standard data corresponding to high-quality products to construct a dynamic fermentation state reference benchmark. Therefore, this step first uses principal component analysis to screen high-contribution features, defines normalized weights, and constructs a dimensionality-reduced and weighted fusion vector paradigm to improve the accuracy of feature characterization; then, based on qualified sample data, a dynamic standard fermentation state vector set is constructed according to the fermentation stage to provide a scientific reference for real-time fermentation state comparison.

[0028] Step S200 in the method provided in this embodiment of the invention includes: Among them, the sample fermentation dataset is analyzed based on the results of response factor reconstruction to obtain the fusion vector paradigm of the target rice wine product, including: Principal component analysis was performed by combining the first set of candidate response factors with the sample fermentation dataset to obtain the first contribution of each first candidate response factor. Principal component analysis was performed by combining the second set of alternative response factors with the sample fermentation dataset to obtain the second contribution of each second alternative response factor. Based on a preset cumulative contribution threshold, the first candidate response factor set and the second candidate response factor set are subjected to cumulative analysis based on the first contribution and the second contribution, respectively. Obtain the top M first candidate response factors and the top N second candidate response factors that satisfy the cumulative contribution threshold, and concatenate their dimensions to define the fusion vector paradigm, where M and N are both positive integers greater than or equal to 2.

[0029] First, principal component analysis (PCA) is performed on the first set of candidate response factors and the sample fermentation dataset to obtain the first contribution of each candidate response factor. PCA is a data dimensionality reduction and feature extraction method that maps high-dimensional data to a low-dimensional space through linear transformation. Each dimension in the low-dimensional space is a linear combination of the original features, and the principal components are uncorrelated, thus preserving the effective information of the original data to the greatest extent. The first contribution refers to the proportion of the variance of the original data explained by each principal component in the first set of candidate response factors; the higher the contribution, the stronger the ability of that principal component to characterize the fermentation state. Feature data corresponding to the first set of candidate response factors is extracted from the sample fermentation dataset; the feature data is preprocessed by standardization to eliminate dimensional differences; the PCA algorithm is then used to calculate the variance contribution of each principal component, i.e., the first contribution.

[0030] For example, from 85 sample records, the response signals of S1, S2, S3, and S4 at each time point are extracted to form a feature matrix of 4 columns × 85 rows; preprocessing: Z-score standardization is performed on the feature matrix; principal component analysis results: four principal components PC1, PC2, PC3, and PC4 are output, with corresponding first contribution values ​​of PC1=65%, PC2=20%, PC3=10%, and PC4=5%, respectively.

[0031] Secondly, principal component analysis is performed on the second set of alternative response factors and the sample fermentation dataset to obtain the second contribution of each alternative response factor. The second contribution refers to the proportion of the variance explained by each principal component in the original data within the second set of alternative response factors, reflecting the principal component's ability to characterize the proportional correlation features of substances. Feature data corresponding to the second set of alternative response factors is extracted from the sample fermentation dataset; the proportional data is standardized and preprocessed; and principal component analysis is used to calculate and output the variance contribution of each principal component, i.e., the second contribution.

[0032] For example, the proportions of R1(S1 / S2), R2(S1 / S3), R3(S1 / S4), R4(S2 / S3), R5(S2 / S4), and R6(S3 / S4) in 85 samples are calculated to form a feature matrix with 6 columns × 85 rows; preprocessing: Z-score standardization is performed on the proportion matrix; principal component analysis results: six principal components PCa, PCb, PCc, PCd, PCe, and PCf are output, with corresponding second contribution rates of PCa=60%, PCb=28%, PCc=7%, PCd=3%, PCe=1%, and PCf=1%, respectively.

[0033] Furthermore, based on a preset cumulative contribution threshold, cumulative analysis is performed on the first candidate response factor set and the second candidate response factor set, respectively, based on the first contribution and the second contribution. The cumulative contribution threshold is a preset critical value used to screen principal components, ensuring that the screened principal components retain most of the effective information of the original data, achieving dimensionality reduction without losing key features. For example, if the preset cumulative contribution threshold is 85%, the factors contained in the first and second candidate sets are sorted from largest to smallest according to their respective contributions.

[0034] Subsequently, the top M first-candidate response factors and the top N second-candidate response factors that satisfy the cumulative contribution threshold are obtained, and their dimensions are concatenated to form the fusion vector paradigm, where M and N are both positive integers greater than or equal to 2. M / N refers to the number of high-contribution principal components selected from the first and second candidate response factor sets, respectively, i.e., the top M first-contribution principal components and the top N second-contribution principal components. The first-contribution factors are accumulated in descending order until the cumulative contribution reaches the threshold, and the number of principal components M at this point is recorded; the second-contribution factors are accumulated in descending order until the cumulative contribution reaches the threshold, and the number of principal components N at this point is recorded; the top M first-candidate response factor principal components and the top N second-candidate response factor principal components are selected. For example, the cumulative contribution threshold is 85%; the first contribution accumulation is: PC1 (65%) + PC2 (20%) = 85%, which reaches the threshold, so M = 2, and PC1 and PC2 are selected; the second contribution accumulation is: PCa (60%) + PCb (28%) = 88% > 85%, which reaches the threshold, so N = 2, and PCa and PCb are selected.

[0035] The analysis of the sample fermentation dataset based on the response factor reconstruction results to obtain the fusion vector paradigm of the target rice wine product also includes: Based on the first contribution and the second contribution, normalized weights are defined for the dimensions corresponding to the first M first candidate response factors and the first N second candidate response factors, respectively, to generate a set of dimension fusion weights; The dimensional fusion weight set is associated with and stored with the vector dimension formed by concatenating the first M first candidate response factors and the first N second candidate response factors, thus forming the fusion vector paradigm.

[0036] First, based on the first contribution and the second contribution, normalized weights are defined for the dimensions corresponding to the top M first-option response factors and the top N second-option response factors, respectively, generating a dimension fusion weight set. Normalized weights convert the contribution of each principal component into a value between 0 and 1, making the total weights equal to 1, used to quantify the importance of different principal components in the fusion vector. The dimension fusion weight set is a collection containing the weights of the top M first principal components and the weights of the top N second principal components, corresponding one-to-one with the dimensions of the fusion vector. The normalized weights of the top M first principal components are calculated as follows: weight of a single principal component = contribution of that principal component / cumulative contribution of the top M principal components; the normalized weights of the top N second principal components are calculated as follows: weight of a single principal component = contribution of that principal component / cumulative contribution of the top N principal components; the two types of weights are then integrated to generate the dimension fusion weight set.

[0037] For example, the weights of the first M=2 principal components are calculated as follows: PC1 weight = 65% / (65%+20%) ≈ 0.765; PC2 weight = 20% / (65%+20%) ≈ 0.235; the weights of the first N=2 principal components are calculated as follows: PCa weight = 60% / (60%+28%) ≈ 0.682; PCb weight = 28% / (60%+28%) ≈ 0.318; the dimensional fusion weight set is: {0.765 (PC1), 0.235 (PC2), 0.682 (PCa), 0.318 (PCb)}, and the total weight is 1.

[0038] Secondly, the dimensional fusion weight set is associated with and stored as a vector dimension formed by concatenating the first M first candidate response factors and the first N second candidate response factors, thus constituting the fusion vector paradigm. The fusion vector paradigm refers to a unified feature transformation standard that includes high-contribution principal component dimensions plus corresponding normalized weights, used to convert raw sensory information into a weighted low-dimensional fermentation-state vector. The first M first principal components and the first N second principal components are concatenated sequentially to form the basic dimensions of the fusion vector; the dimensional fusion weight set is associated one-to-one with the concatenated basic dimensions, i.e., each dimension corresponds to a normalized weight; the associated dimension-weight combinations are stored to constitute the fusion vector paradigm.

[0039] For example, PC1 and PC2 are concatenated with PCa and PCb, with the basic dimensions being [PC1, PC2, PCa, PCb]; the association weights are: PC1-0.765: the single substance feature dominated by ethanol, with the highest weight; PC2-0.235: the auxiliary single substance features such as acetic acid; PCa-0.682: the proportion feature dominated by alcohol-ester ratio; PCb-0.318: the auxiliary proportion feature of acid-ester ratio; the fusion vector paradigm is: F=[PC1(0.765), PC2(0.235), PCa(0.682), PCb(0.318)], which is used to convert the original signal into a 4-dimensional weighted vector later.

[0040] Specifically, by combining the fusion vector paradigm with the sample fermentation dataset, a standard fermentation state vector set is established, including: The sample fermentation dataset was analyzed to obtain the quality inspection information of the rice wine products; Based on the quality inspection information of the rice wine products, filter the sample fermentation data in the sample fermentation dataset that meet the rice wine product output standards to obtain the standard fermentation dataset; Using the fermentation stage as an index, the standard fermentation dataset is segmented temporally, and the temporal segmentation results are traversed and vectorized based on the fusion vector paradigm to obtain a fermentation state vector set. The fermentation state vector set includes multiple fermentation state vector clusters, and each fermentation state vector cluster corresponds one-to-one with a fermentation stage. Based on statistical analysis methods, a standard fermentation state vector corresponding to each fermentation state vector cluster is constructed and merged to output the standard fermentation state vector set.

[0041] The standard fermentation state vector set includes at least a standard saccharification fermentation state vector, a standard acid-producing fermentation state vector, and a standard termination fermentation state vector.

[0042] First, the sample fermentation dataset is parsed to obtain the quality inspection information for rice wine products. The quality inspection information for rice wine products is defined as in S100, which includes quality data containing physicochemical, microbiological, and sensory indicators; this information is used to screen qualified samples. The sample fermentation dataset constructed in S100 is read; the quality inspection information fields are extracted from each record; the quality inspection information is then categorized and organized by field to form a quality inspection information list. For example, the quality inspection data for each of the 85 records is extracted, such as the record for 72 hours of fermentation (end of saccharification and fermentation): alcohol content 5.2% vol, total acid 1.8 g / L, reducing sugar 8.5 g / 100 mL, yeast count 4.2 × 10⁻⁶. 6 CFU / mL, sensory score 86 points; fermentation record at 168h (end of acid-producing fermentation): alcohol content 10.3% vol, total acid 2.9 g / L, reducing sugar 4.1 g / 100 mL, yeast count 7.8 × 10⁻⁶. 6CFU / mL, sensory score 90 points; fermentation 504h (fermentation terminated) record: alcohol content 12.1% vol, total acid 3.2 g / L, reducing sugar 2.6 g / 100 mL, yeast count 2.3 × 10⁻⁶. 6 CFU / mL, sensory score of 96. A quality inspection information list of 85 time points and 5 quality inspection indicators was ultimately generated, providing a basis for selecting qualified samples.

[0043] Secondly, based on the quality inspection information of the rice wine products, sample fermentation data that meet the rice wine product output standards are selected from the sample fermentation dataset to obtain the standard fermentation dataset. The rice wine product output standards are the quality thresholds for qualified products set by the enterprise, which must comply with food safety standards and regulations, clearly defining the acceptable range for each core indicator. The standard fermentation dataset refers to the set of samples in the sample fermentation dataset whose quality inspection information fully meets the output standards. It is the basic data for constructing the standard fermentation state vector, retaining only data related to high-quality fermentation. The rice wine product output standards are clearly defined; each record in the sample fermentation dataset is traversed, and the quality inspection information is verified one by one to ensure compliance with the standards; all records that meet the standards are selected and merged to form the standard fermentation dataset.

[0044] For example, the output standards are set as follows: alcohol content 10-13% vol, total acid ≤ 3.5 g / L, reducing sugar ≥ 2.5 g / 100 mL, and sensory score ≥ 85 points. 85 records are checked one by one, and records that do not meet the standards are removed, such as records with a sensory score of 78 points < 85 points after 24 hours of fermentation, records with a total acid of 3.6 g / L > 3.5 g / L after 48 hours of fermentation, and records with an alcohol content of 8.7% vol < 10% vol after 96 hours of fermentation. Finally, 70 records that meet the standards are selected to form the standard fermentation dataset.

[0045] Furthermore, using fermentation stages as indexes, the standard fermentation dataset is temporally segmented, and the segmentation results are traversed and vectorized based on the fusion vector paradigm to obtain a fermentation state vector set. This set includes multiple fermentation state vector clusters, each corresponding one-to-one with a fermentation stage. The standard fermentation state vector set includes at least a standard saccharification fermentation state vector, a standard acid-producing fermentation state vector, and a standard termination fermentation state vector. Fermentation stages refer to the typical temporal stages of rice wine fermentation, determined according to microbial metabolic characteristics and industry-standard classifications. These stages include: saccharification fermentation stage (0-72h): starch is converted into fermentable sugars, and yeast begins to proliferate; acid-producing fermentation stage (72-168h): lactic acid bacteria and other microorganisms metabolize to produce organic acids, regulating flavor; and termination fermentation stage (168-504h): alcohol and ester production gradually level off, and quality stabilizes. Temporal segmentation refers to dividing the standard fermentation dataset into subsets corresponding to each stage, using fermentation stages as indexes and time ranges to ensure the temporal correlation of the data. A fermentation state vector cluster refers to the set of vectors formed after transforming all samples within the same fermentation stage using the fusion vector paradigm, reflecting the distribution range of qualified fermentation states at that stage. The fermentation state vector set is the collection of fermentation state vector clusters corresponding to all fermentation stages, with a one-to-one correspondence between each fermentation stage. Typical stages of rice wine fermentation and their corresponding time ranges are defined; the standard fermentation dataset is time-series segmented according to the time range to obtain subsets for each stage; using the fusion vector paradigm, samples in each subset are vectorized, principal component scores are extracted, and weighted vectors are generated by combining weights; all vectorization results from the same stage constitute a fermentation state vector cluster, and all clusters are merged into the fermentation state vector set.

[0046] For example, the fermentation stages are defined as follows: saccharification fermentation (0-72h), acid-producing fermentation (72-168h), and termination fermentation (168-504h). Temporal segmentation: In the standard fermentation dataset of 70 records, there are 15 records for 0-72h, 25 records for 72-168h, and 30 records for 168-504h, forming three independent subsets. Vectorization: Taking a sample from the acid-producing fermentation stage as an example, its S1-S4 signals are projected using PCA to obtain PC1=0.9 and PC2=0.4, and the R1-R6 ratios are projected using PCA to obtain PCa=0.6 and PCb=0.3. Substituting these values ​​into the fusion vector paradigm F, a weighted vector is generated: [0.9×0.765=0.6885, 0.4×0.235=0.094, 0.6×0.682=0.4092, 0.3×0.318=0.0954]. Vector cluster generation: All 25 samples in the acid-producing fermentation stage are transformed according to the above method to form an acid-producing fermentation state vector cluster containing 25 4-dimensional vectors; similarly, a saccharification fermentation state vector cluster containing 15 vectors and a termination fermentation state vector cluster containing 30 vectors are obtained. The final fermentation state vector set = {saccharification vector cluster, acid-producing vector cluster, termination vector cluster}.

[0047] Finally, based on statistical analysis methods, standard fermentation state vectors corresponding to each fermentation state vector cluster are constructed and merged to output the standard fermentation state vector set. The statistical analysis method employs the arithmetic mean method, calculating the mean of all vectors in a vector cluster along their corresponding dimensions to obtain the most representative vector for that stage, i.e., the standard fermentation state vector. The standard fermentation state vector refers to the typical characteristic vector of each fermentation stage, reflecting the core characteristics of a qualified fermentation state at that stage, and can serve as a benchmark for real-time comparison. The standard fermentation state vector set is a collection of standard fermentation state vectors for all fermentation stages, including at least the standard saccharification fermentation state vector, the standard acid-producing fermentation state vector, and the standard termination fermentation state vector. For each fermentation state vector cluster, the arithmetic mean of all vectors along each dimension is calculated. The mean of each dimension = the sum of all vector values ​​in that dimension / the number of vectors. The mean vector of each vector cluster is the standard fermentation state vector for that stage. The standard fermentation state vectors of all stages are merged to output the standard fermentation state vector set.

[0048] For example, the saccharification and fermentation state vector cluster is calculated as follows: the mean values ​​of each dimension of the 15 vectors are: PC1 weighted mean = 0.45, PC2 weighted mean = 0.12, PCa weighted mean = 0.25, PCb weighted mean = 0.07, resulting in the standard saccharification and fermentation state vector = [0.45, 0.12, 0.25, 0.07]; the acid-producing fermentation state vector cluster is calculated as follows: the mean values ​​of each dimension of the 25 vectors are: PC1 weighted mean = 0.78, PC2 weighted mean = 0.21, PCa weighted mean = 0.42, PCb weighted mean = 0.15, resulting in the standard acid-producing fermentation state vector = [0.45, 0.12, 0.25, 0.07]. [78, 0.21, 0.42, 0.15]; Calculation of the terminated fermentation state vector cluster: The mean values ​​of each dimension of the 30 vectors are: PC1 weighted mean = 0.89, PC2 weighted mean = 0.28, PCa weighted mean = 0.56, PCb weighted mean = 0.22, resulting in the standard terminated fermentation state vector = [0.89, 0.28, 0.56, 0.22]; Standard fermentation state vector set = {[0.45, 0.12, 0.25, 0.07], [0.78, 0.21, 0.42, 0.15], [0.89, 0.28, 0.56, 0.22]}.

[0049] In this embodiment of the invention, high-value features that meet the cumulative contribution threshold are screened through principal component analysis, and a fusion vector paradigm is constructed by defining normalized weights. This achieves precise dimensionality reduction in the characterization of fermentation state and distinguishes feature importance through weight quantification, avoiding the problems of dimensional redundancy and the obscuring of core features. At the same time, qualified samples are screened based on the production standards of rice wine products, and data is segmented according to the time sequence of the industry-standard fermentation stages. A standard fermentation state vector set is constructed using the mean method, forming a dynamic and objective reference benchmark covering key fermentation processes. Ultimately, this achieves precise characterization and standardized dynamic anchoring of the fermentation state, effectively solving the defects of feature dimensional redundancy, indiscriminate weighting, and static rigidity in traditional technologies. This lays a solid and reliable technical foundation for subsequent real-time fermentation state comparison analysis and adaptive control strategy generation, improving the refinement and robustness of the entire control method.

[0050] S300: Electronic sensory information is collected in real time through an electronic nose sensor array and converted into a real-time fermentation state vector through the fusion vector paradigm.

[0051] In this embodiment of the invention, electronic sensor information is collected in real time through an electronic nose sensor array and converted into a real-time fermentation state vector using the fusion vector paradigm. S200 has successfully constructed a fusion vector paradigm containing high-contribution principal component dimensions and normalized weights, as well as a standard fermentation state vector set covering key fermentation stages, providing a unified and accurate reference benchmark for real-time fermentation state comparison. However, the real-time fermentation process is dynamically evolving, requiring continuous capture of instantaneous changes in the fermentation state. Traditional real-time monitoring only collects raw electronic sensor signals without combining them with a preset fusion vector paradigm for targeted conversion, resulting in problems such as dimensional redundancy and feature misalignment in the raw signals. This makes it impossible to directly compare them effectively with the standard fermentation state vector set established by S200, thus affecting the accuracy of subsequent stage identification and control. Therefore, this step requires real-time information collection via the electronic nose, relying on the fusion vector paradigm of S200 to complete the closed-loop conversion from raw signal to raw dimensional values ​​to weighted real-time vector, achieving adaptation between the real-time fermentation state and the standard benchmark, laying the data foundation for the comparison analysis of S400.

[0052] Step S300 in the method provided in this embodiment of the invention includes: The electronic nose sensor array is used to collect real-time electronic sensory information of the target rice wine product. Based on the fusion vector paradigm, the original values ​​corresponding to each vector dimension in the fusion vector paradigm are calculated for the real-time electronic sensory information. The original numerical values ​​are weighted according to the dimension fusion weight set associated with the fusion vector paradigm to generate the real-time fermentation state vector.

[0053] First, the electronic nose sensor array is used to collect real-time electronic sensory information of the target rice wine product. Real-time electronic sensory information refers to the instantaneous response signal of the electronic nose sensor array to volatile substances in the fermentation environment during real-time fermentation, reflecting the immediate characteristics of the current fermentation state. The acquisition principle is the same as the sample electronic sensory information in S100, but the application scenario is real-time fermentation. The 12-channel MOS electronic nose sensor array used in S100 is retained to ensure that the sensor type, the correspondence of sensitive substances, and the sample data acquisition stage are completely consistent, avoiding errors introduced by equipment differences. The real-time acquisition frequency is set to be consistent with the S100 sample acquisition frequency, once every 6 hours, balancing data timeliness and equipment wear and tear. The electronic nose sensor array is sealed and connected to the gas sampling port at the top of the real-time fermentation tank. Each time data is collected, 50mL of headspace gas is extracted from the tank. After the sensor comes into contact with the gas, it outputs an instantaneous response signal, which is automatically stored in the real-time data acquisition system.

[0054] For example, in a real-time fermentation scenario: the target rice wine product is started with real-time fermentation according to the S100 standard fermentation process: glutinous rice:water = 1:1.2, inoculated with brewer's yeast + Rhizopus, fermentation temperature 28℃; the current fermentation time is 120h, in the acid-producing fermentation stage defined by S200. Data acquisition parameters: the sensor array still consists of 12 MOS sensors, including S1 (ethanol sensitive), S2 (acetic acid sensitive), S3 (ethyl lactate sensitive), and S4 (isoamyl alcohol sensitive), with a data acquisition frequency of once every 6 hours and a sampling volume of 50mL. Data collection results: After 120 hours of fermentation, the real-time electronic sensory information output by the electronic nose was as follows: S1=5.6mV, S2=2.3mV, S3=1.9mV, S4=1.1mV, S5=1.3mV, S6=1.0mV, S7=0.9mV, S8=1.2mV, S9=1.5mV, S10=0.8mV, S11=1.1mV, S12=1.4mV.

[0055] Secondly, based on the fusion vector paradigm, the original values ​​corresponding to each vector dimension in the fusion vector paradigm are calculated for the real-time electronic sensory information. The original values ​​of the vector dimensions refer to the unweighted original scores obtained by projecting the real-time electronic sensory information onto each principal component dimension of the S200 fusion vector paradigm. These are the basic data for generating the real-time fermentation state vector, and the calculation logic is consistent with the principal component scores when constructing the S200 paradigm. The PCA model used for principal component analysis in S200 is invoked, including standardized parameters, principal component loading matrices, etc., which are completely consistent with the sample data processing parameters. From the real-time electronic sensory information, the core feature data corresponding to the S200 fusion vector paradigm is extracted, namely the real-time signals corresponding to the first candidate response factor set {S1, S2, S3, S4} of S100, and the real-time proportion values ​​corresponding to the second candidate response factor set {R1-R6}. The extracted core feature data is input into the PCA model, and through the same projection calculation as S200, the original scores corresponding to each principal component dimension (PC1, PC2, PCa, PCb) in the fusion vector paradigm are obtained, that is, the original values ​​of each vector dimension.

[0056] For example, from the real-time signal of 120h fermentation, S1=5.6mV, S2=2.3mV, S3=1.9mV, and S4=1.1mV are extracted, and the real-time proportion values ​​of the second alternative response factor set are calculated: R1=S1 / S2≈2.43, R2=S1 / S3≈2.95, R3=S1 / S4≈5.09, R4=S2 / S3≈1.21, R5=S2 / S4≈2.09, and R6=S3 / S4≈1.73; PCA projection calculation: The above real-time signals of S1-S4 and the real-time proportion values ​​of R1-R6 are input into the PCA model of S200; After calculation, the original values ​​of each dimension of the fused vector paradigm are: PC1=1.2, PC2=0.5, PCa=0.7, and PCb=0.4.

[0057] Finally, based on the dimension fusion weight set stored in the fusion vector paradigm, the original values ​​are weighted to generate the real-time fermentation state vector. The real-time fermentation state vector is a weighted low-dimensional vector formed by multiplying the original values ​​of each vector dimension with the corresponding weights of the S200 dimension fusion weight set. Its dimensions are consistent with the S200 standard fermentation state vector and can be directly used for subsequent comparative analysis. The dimension fusion weight set generated by S200 is called; following the one-to-one correspondence principle, the original values ​​of each vector dimension are multiplied with the corresponding normalized weights to obtain the weighted values ​​of each dimension; following the dimensional order of the S200 fusion vector paradigm, all weighted values ​​are concatenated to generate the final real-time fermentation state vector.

[0058] For example, the dimension fusion weight set of S200 is directly used, with PC1 corresponding to a weight of 0.765, PC2 corresponding to a weight of 0.235, PCa corresponding to a weight of 0.682, and PCb corresponding to a weight of 0.318; weighted calculation: PC1 weighted value = 1.2 × 0.765 = 0.918; PC2 weighted value = 0.5 × 0.235 = 0.1175; PCa weighted value = 0.7 × 0.682 = 0.4774; PCb weighted value = 0.4 × 0.318 = 0.1272. The weighted values ​​are concatenated in dimensional order to obtain the real-time fermentation state vector after 120 hours of fermentation = [0.918, 0.1175, 0.4774, 0.1272].

[0059] In this embodiment of the invention, by using the electronic nose sensor array of S100, the consistency of real-time and sample electronic sensory information acquisition is ensured, avoiding systematic errors caused by equipment differences. Based on the fusion vector paradigm of S200, a precise conversion of the entire process—feature extraction, raw numerical calculation, and weighted processing—is completed, generating a real-time fermentation state vector with dimensions consistent with and features aligned with the standard fermentation state vector. This process not only highlights the importance of core features through weighted processing, avoiding the problems of redundant original signal dimensions and lack of feature prominence, but also achieves high compatibility between real-time data and the standard benchmark. This provides highly reliable and timely data support for rapid comparison and accurate deviation analysis of real-time and standard vectors in S400, effectively solving the technical defects of signal conversion lag and incompatibility with the standard benchmark in traditional real-time monitoring.

[0060] S400: Traverse the standard fermentation state vector set, compare and analyze it with the real-time fermentation state vector, and generate and execute intelligent fermentation control based on the comparison and analysis results and the preset multidimensional control decision matrix.

[0061] In this embodiment of the invention, the standard fermentation state vector set is traversed and compared with the real-time fermentation state vector. Based on the comparison analysis results and the preset multidimensional control decision matrix, intelligent fermentation control is generated and executed. S300 has generated a real-time fermentation state vector with the same dimensions as the standard fermentation state vector in S200, providing a suitable data carrier for fermentation state comparison. However, traditional comparison only judges whether the state is abnormal by simple distance calculation, without considering the correlation and dispersion of each vector dimension, resulting in distance calculation deviation and inaccurate stage identification. Even if anomalies are identified, there is a lack of a system mechanism to transform deviation characteristics into specific control measures, and a closed loop cannot be formed. To this end, this step uses two precise comparison paths: Mahalanobis distance method with dimension correction and efficient Euclidean distance method to achieve precise positioning and deviation quantification analysis of real-time fermentation stages. Combined with the preset multidimensional control decision matrix, the deviation information is transformed into executable control instructions, solving the problems of inaccurate traditional comparison and lack of closed-loop control, and finally achieving intelligent adaptive control of the fermentation process.

[0062] Step S400 in the method provided in this embodiment of the invention includes: The process of traversing the standard fermentation state vector set and comparing it with the real-time fermentation state vector includes: The Euclidean distance between each standard fermentation state vector and the real-time fermentation state vector is calculated by iterating through the data. Obtain the dimension-covariance matrix corresponding to the fusion vector paradigm, and perform dimension correction on the Euclidean distance based on the dimension-covariance matrix to obtain multiple Mahalanobis distances; The real-time fermentation stage corresponding to the real-time fermentation state vector is determined based on the standard fermentation state vector corresponding to the smallest of the multiple Mahalanobis distances. Calculate the deviation vector between the standard fermentation state vector and the real-time fermentation state vector corresponding to the real-time fermentation stage, and combine the deviation vector with the real-time fermentation stage to output the comparison analysis result.

[0063] First, the Euclidean distance between each standard fermentation state vector and the real-time fermentation state vector is calculated. Euclidean distance: measures the straight-line distance between corresponding points in two vector spaces, reflecting the degree of intuitive difference between vectors. The calculation formula is: Where n is the vector dimension, RV is the real-time fermentation state vector, and SV is the standard fermentation state vector. The system calls the real-time fermentation state vector RV generated by S300; it calls the standard fermentation state vector set SV established by S200; and it calculates the distance between RV and each SV according to the Euclidean distance formula, recording all distance results.

[0064] For example, using the real-time fermentation state vector after 120 hours of fermentation = [0.918, 0.1175, 0.4774, 0.1272], we calculate the following using the standard acid-producing fermentation state vector, the standard terminated fermentation state vector, and the standard saccharification fermentation state vector: We also calculate the following using the standard acid-producing fermentation state vector = [0.78, 0.21, 0.42, 0.15]: ; Calculate with the standard terminated fermentation state vector = [0.89, 0.28, 0.56, 0.22]: ; Calculate with the standard saccharification and fermentation state vector = [0.45, 0.12, 0.25, 0.07]: .

[0065] Secondly, the dimension-covariance matrix corresponding to the fused vector paradigm is obtained, and the Euclidean distance is dimensionally corrected based on the dimension-covariance matrix to obtain multiple Mahalanobis distances. The dimension-covariance matrix is ​​a matrix calculated based on the S200 standard fermentation state vector set, describing the correlation and dispersion between the dimensions of each vector, and is used to correct the problems of inconsistent dimensional scales and correlation interference in Euclidean distance. Mahalanobis distance refers to the distance corrected by the covariance matrix, eliminating redundancy and interference between dimensions, and more accurately reflecting the true differences between vectors. The calculation formula is: Where RV is the real-time fermentation state vector and SV is the standard fermentation state vector. This is the inverse of the covariance matrix. Based on all sample vectors from the S200 standard fermentation state vector set (e.g., 15 vectors from the standard saccharification fermentation state vector cluster, 25 vectors from the standard acid-producing fermentation state vector cluster, and 30 vectors from the standard termination fermentation state vector cluster), the covariance matrix is ​​calculated by dimension. Inverting the covariance matrix yields... Substitute the vector corresponding to each Euclidean distance into the Mahalanobis distance formula to calculate the corrected Mahalanobis distance. For example, the final Mahalanobis distance set is {0.482 (standard saccharification fermentation state vector), 0.156 (standard acid-producing fermentation state vector), 0.213 (standard termination fermentation state vector)}.

[0066] Further, based on the standard fermentation state vector corresponding to the smallest of the multiple Mahalanobis distances, the real-time fermentation stage corresponding to the real-time fermentation state vector is determined. The real-time fermentation stage refers to the standard fermentation stage corresponding to the real-time fermentation state vector, determined by the standard fermentation state vector corresponding to the smallest Mahalanobis distance, reflecting the actual position of the current fermentation process. The Mahalanobis distance set is traversed, and the distance with the smallest value is selected; the standard fermentation state vector corresponding to this smallest distance is determined; the fermentation stage to which this standard vector belongs is the real-time fermentation stage. For example, in the Mahalanobis distance set, 0.156 is the minimum value, corresponding to the standard acid-producing fermentation state vector; the fermentation stage corresponding to the standard acid-producing fermentation state vector is the acid-producing fermentation stage (72-168h) defined by S200; the real-time fermentation stage of 120h fermentation is determined to be the acid-producing fermentation stage.

[0067] Finally, the deviation vector between the standard fermentation state vector corresponding to the real-time fermentation stage and the real-time fermentation state vector is calculated, and the deviation vector is combined with the real-time fermentation stage to output the comparison analysis result. The deviation vector is the difference vector between the real-time fermentation state vector and the corresponding standard fermentation state vector. Deviation vector = real-time vector - current stage standard vector. The difference in each dimension reflects the degree of deviation of that feature dimension; a positive deviation indicates above the standard, and a negative deviation indicates below the standard. The comparison analysis result contains comprehensive information including the real-time fermentation stage and the deviation vector, and is the core input for generating the control strategy. The standard fermentation state vector corresponding to the real-time fermentation stage is extracted as the current stage standard vector; the difference between the real-time fermentation state vector and the current stage standard vector is calculated dimension by dimension to form the deviation vector; the real-time fermentation stage and the deviation vector are combined to output the comparison analysis result.

[0068] For example, the current real-time stage is the acid-producing fermentation stage, and the corresponding current stage standard vector = standard acid-producing fermentation state vector = [0.78, 0.21, 0.42, 0.15]; deviation vector = real-time fermentation state vector after 120 hours of fermentation - standard acid-producing fermentation state vector = [0.918-0.78, 0.1175-0.21, 0.4774-0.42, 0.1272-0.15] = [0.138, -0.0925, 0.0574, -0.0228]. Therefore, the comparative analysis result = {stage: acid-producing fermentation stage, deviation vector: [0.138, -0.0925, 0.0574, -0.0228]}, where the meaning of each dimension of deviation is as follows: the first principal component positive deviation indicates that the total intensity of a single substance is too high, the second principal component negative deviation indicates that the acetic acid content is too low; the first proportional principal component positive deviation indicates that the alcohol-ester ratio is too high; the second proportional principal component negative deviation indicates that the acid-ester ratio is too low.

[0069] The process of traversing the standard fermentation state vector set and comparing it with the real-time fermentation state vector includes: The Euclidean distance between each standard fermentation state vector and the real-time fermentation state vector is calculated by iterating through the data. The real-time fermentation stage corresponding to the real-time fermentation state vector is determined based on the standard fermentation state vector corresponding to the smallest of the multiple Euclidean distances. Calculate the deviation vector between the standard fermentation state vector and the real-time fermentation state vector corresponding to the real-time fermentation stage, and combine the deviation vector with the real-time fermentation stage to output the comparison analysis result.

[0070] First, the Euclidean distance between each standard fermentation state vector and the real-time fermentation state vector is calculated iteratively. The calculation process is the same as above, and the Euclidean distance result is {0.503 (standard saccharification fermentation state vector), 0.179 (standard acid-producing fermentation state vector), 0.207 (standard termination fermentation state vector)}.

[0071] Secondly, the real-time fermentation stage corresponding to the real-time fermentation state vector is determined based on the standard fermentation state vector corresponding to the smallest of the multiple Euclidean distances. For example, the smallest Euclidean distance of 0.179 corresponds to the standard acid-producing fermentation state vector, so the real-time fermentation stage is the acid-producing fermentation stage.

[0072] Further, the deviation vector between the standard fermentation state vector and the real-time fermentation state vector corresponding to the real-time fermentation stage is calculated, and the deviation vector is combined with the real-time fermentation stage to output the comparison analysis result. For example, the deviation vector = real-time fermentation state vector after 120 hours of fermentation - standard acid-producing fermentation state vector = [0.138, -0.0925, 0.0574, -0.0228], and the comparison analysis result = {stage: acid-producing fermentation stage, deviation vector: [0.138, -0.0925, 0.0574, -0.0228]}.

[0073] Finally, based on the comparative analysis results and the preset multidimensional control decision matrix, intelligent fermentation control is generated and executed.

[0074] First, a multidimensional control decision matrix is ​​pre-set. This matrix uses fermentation stage and deviation type as joint indices to map specific control measures. Each index combination within the matrix corresponds to an experimentally verified optimal quantitative control scheme, serving as the core bridge connecting deviation analysis and control execution. The pre-set multidimensional control decision matrix uses row indices representing the three core fermentation stages defined by S200: saccharification fermentation, acid production fermentation, and termination fermentation. Column indices represent the deviation types across four dimensions of the fusion vector, including positive deviation, negative deviation, and no significant deviation. Each matrix cell explicitly records the corresponding quantitative control measures, such as temperature adjustment range, stirring cycle and duration, and aeration frequency and duration.

[0075] Secondly, intelligent fermentation control is generated and executed. Control commands refer to the conversion of control measures matched by the decision matrix into quantitative parameter commands that the fermentation equipment can directly execute. These commands include specific operating parameters such as temperature setpoint, aeration duration, and stirring cycle, ensuring that the equipment can accurately respond to control requirements. The real-time fermentation stage and deviation types in each dimension are obtained from the comparative analysis results. Core deviation dimensions with absolute values ​​greater than preset thresholds are identified first. If multiple core deviations exist, multiple related control measures are integrated to form a combined scheme. Using the real-time fermentation stage + core deviation type as a joint index, the corresponding combined control measures are accurately matched in the multi-dimensional control decision matrix to ensure a high degree of fit between the measures and deviation characteristics. The combined control measures are broken down into specific equipment operating parameters, and the corresponding actuators and operating standards for each measure are clearly defined. The parameter instructions are sent to the corresponding actuators through industrial communication protocols. The actuators adjust their operating status in real time and record information such as control time, parameters before and after adjustment, and equipment operating status to the system database. The real-time fermentation state vector is continuously monitored at the acquisition frequency set by S300, and the comparative analysis process of S400 is executed cyclically. If the absolute values ​​of each dimension of the deviation vector are less than the preset thresholds, the current parameters are maintained. If core deviations still exist, control measures are rematched until the fermentation state meets the standards.

[0076] For example, based on the preset decision matrix, for the acid-producing fermentation stage, a negative deviation of the second principal component indicates a low level of acetic acid, corresponding to a temperature reduction of 0.3℃ and stirring for 5 minutes every 30 minutes. A negative deviation of the second proportional principal component indicates a low acid-ester ratio, corresponding to maintaining the aeration rate and extending the stirring time. The real-time fermentation stage is the acid-producing fermentation stage, and the core deviations are the negative deviations of the second principal component and the second proportional principal component. A combined control measure is obtained through a joint index query: the fermentation temperature is reduced from 28℃ to 27.7℃, the stirring frequency is adjusted to stirring for 5 minutes every 30 minutes, and the current aeration rate is maintained unchanged. Instruction breakdown: The PLC temperature control system setting is adjusted to 27.7℃; the stirring motor operating parameters are set to a 30-minute cycle, with 5 minutes of stirring per cycle; the electromagnetic vent valve maintains its current opening frequency; Execution and monitoring: The PLC temperature control system adjusts the temperature feedback through the temperature sensor to stabilize the tank temperature at 27.7℃, and the stirring motor runs according to the set cycle; after 1 hour, real-time electronic sensor information is collected and converted into a new real-time fermentation state vector, and comparative analysis is performed again. If the absolute value of the deviation dimension is less than 0.05, the parameters are maintained; otherwise, the optimization and control continue.

[0077] In this embodiment of the invention, two precise comparison paths are employed: Mahalanobis distance method, which balances accuracy, and Euclidean distance method, which balances efficiency. These methods achieve precise positioning and quantitative identification of deviation dimensions during the real-time fermentation stage, solving the problems of traditional comparisons neglecting dimensional correlation and resulting in large identification errors. Simultaneously, relying on a preset multi-dimensional control decision matrix, the abstract deviation vector is transformed into specific, executable equipment control commands, constructing an intelligent closed loop of comparison analysis, decision generation, and execution feedback. This step ensures the adaptability of comparisons across different scenarios and corrects fermentation deviations in real time through closed-loop control. It effectively addresses the shortcomings of traditional control methods that rely on human experience and lack systematicity, ultimately improving the stability of the rice wine fermentation process and the quality consistency of different batches, further reducing the cost of manual intervention.

[0078] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides an intelligent control method for rice wine fermentation based on electronic sensory information fusion. First, by integrating electronic sensory information with quality inspection data and reconstructing response factors, it overcomes the limitations of single-signal representation, enriches the dimensions of fermentation state characteristics, and lays a foundation for high-quality analysis. Next, it selects high-contribution features to construct a weighted fusion vector paradigm, establishing a phased dynamic standard fermentation state vector set, effectively solving problems of dimensional redundancy, inconsistent feature weights, and rigid static benchmarks. Then, through the adaptation and conversion between real-time signals and the fusion vector paradigm, it breaks down the barriers between real-time data and standard benchmarks. Finally, through dual-path vector comparison, it accurately identifies fermentation stages and deviations, and combines a multi-dimensional control decision matrix to transform abstract deviations into quantitative control commands, forming an intelligent closed loop. This effectively improves the accuracy and robustness of fermentation state identification, ensures consistency in product quality across different batches, reduces reliance on manual experience, and improves production efficiency, providing a reliable solution for the intelligent upgrading of the traditional rice wine fermentation industry.

[0079] It should be noted that the order of the above embodiments of the present invention 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. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0081] This specification and accompanying drawings are merely illustrative examples of the invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its scope. Therefore, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is intended to include these modifications and modifications.

Claims

1. A method for intelligent control of rice wine fermentation based on electronic sensory information fusion, characterized in that, include: Obtain a sample fermentation dataset of the target rice wine product, and parse the sample fermentation dataset to reconstruct response factors; Based on the analysis of the sample fermentation dataset according to the response factor reconstruction results, the fusion vector paradigm of the target rice wine product is obtained, and a standard fermentation state vector set is established by combining the fusion vector paradigm with the sample fermentation dataset. Electronic sensory information is collected in real time by an electronic nose sensor array and converted into a real-time fermentation state vector through the fusion vector paradigm. The standard fermentation state vector set is traversed and compared with the real-time fermentation state vector. Based on the comparison analysis results and the preset multidimensional control decision matrix, intelligent fermentation control is generated and executed. The process of reconstructing response factors by parsing the sample fermentation dataset includes: Analyze the sample fermentation dataset, extract the response signals of volatile substances that occur more frequently than a preset frequency, and output them as the first set of candidate response factors; The first set of candidate response factors is randomly combined in pairs, and the ratio of each random combination of factor pairs is defined as the second candidate response factor to obtain the second set of candidate response factors. The first set of candidate response factors and the second set of candidate response factors are merged and output as the response factor reconstruction result; Among them, the sample fermentation dataset is analyzed based on the results of response factor reconstruction to obtain the fusion vector paradigm of the target rice wine product, including: Principal component analysis was performed by combining the first set of candidate response factors with the sample fermentation dataset to obtain the first contribution of each first candidate response factor. Principal component analysis was performed by combining the second set of alternative response factors with the sample fermentation dataset to obtain the second contribution of each second alternative response factor. Based on a preset cumulative contribution threshold, the first candidate response factor set and the second candidate response factor set are subjected to cumulative analysis based on the first contribution and the second contribution, respectively. Obtain the top M first candidate response factors and the top N second candidate response factors that satisfy the cumulative contribution threshold, and concatenate their dimensions to define the fusion vector paradigm, where M and N are both positive integers greater than or equal to 2.

2. The intelligent control method for rice wine fermentation based on electronic sensory information fusion as described in claim 1, characterized in that, Obtain a sample fermentation dataset of the target rice wine product, including: Under standard fermentation process, electronic sensory information of samples at multiple fermentation time points is collected by an electronic nose sensor array, wherein the electronic sensory information is the response signal of the electronic nose sensor array to volatile substances; Based on the obtained electronic sensory information of the samples, the corresponding rice wine product quality inspection information is recorded synchronously and combined to form the sample fermentation dataset.

3. The intelligent control method for rice wine fermentation based on electronic sensory information fusion as described in claim 1, characterized in that, By combining the aforementioned fusion vector paradigm with the aforementioned sample fermentation dataset, a standard fermentation state vector set is established, including: The sample fermentation dataset was analyzed to obtain the quality inspection information of the rice wine products; Based on the quality inspection information of the rice wine products, filter the sample fermentation data in the sample fermentation dataset that meet the rice wine product output standards to obtain the standard fermentation dataset; Using the fermentation stage as an index, the standard fermentation dataset is segmented temporally, and the temporal segmentation results are traversed and vectorized based on the fusion vector paradigm to obtain a fermentation state vector set. The fermentation state vector set includes multiple fermentation state vector clusters, and each fermentation state vector cluster corresponds one-to-one with a fermentation stage. Based on statistical analysis methods, a standard fermentation state vector corresponding to each fermentation state vector cluster is constructed and merged to output the standard fermentation state vector set.

4. The intelligent control method for rice wine fermentation based on electronic sensory information fusion as described in claim 1, characterized in that, Traversing the standard fermentation state vector set and comparing it with the real-time fermentation state vector includes: The Euclidean distance between each standard fermentation state vector and the real-time fermentation state vector is calculated by iterating through the data. Obtain the dimension-covariance matrix corresponding to the fusion vector paradigm, and perform dimension correction on the Euclidean distance based on the dimension-covariance matrix to obtain multiple Mahalanobis distances; The real-time fermentation stage corresponding to the real-time fermentation state vector is determined based on the standard fermentation state vector corresponding to the smallest of the multiple Mahalanobis distances. Calculate the deviation vector between the standard fermentation state vector and the real-time fermentation state vector corresponding to the real-time fermentation stage, and combine the deviation vector with the real-time fermentation stage to output the comparison analysis result.

5. The intelligent control method for rice wine fermentation based on electronic sensory information fusion as described in claim 1, characterized in that, Based on the analysis of the sample fermentation dataset using the response factor reconstruction results, the fusion vector paradigm of the target rice wine product is obtained, and the following is also included: Based on the first contribution and the second contribution, normalized weights are defined for the dimensions corresponding to the first M first candidate response factors and the first N second candidate response factors, respectively, to generate a set of dimension fusion weights; The dimensional fusion weight set is associated with and stored with the vector dimension formed by concatenating the first M first candidate response factors and the first N second candidate response factors, thus forming the fusion vector paradigm.

6. The intelligent control method for rice wine fermentation based on electronic sensory information fusion as described in claim 5, characterized in that, Electronic sensory information is acquired in real time through an electronic nose sensor array and converted into a real-time fermentation state vector using the fusion vector paradigm, including: The electronic nose sensor array is used to collect real-time electronic sensory information of the target rice wine product. Based on the fusion vector paradigm, the original values ​​corresponding to each vector dimension in the fusion vector paradigm are calculated for the real-time electronic sensory information. The original numerical values ​​are weighted according to the dimension fusion weight set associated with the fusion vector paradigm to generate the real-time fermentation state vector.

7. The intelligent control method for rice wine fermentation based on electronic sensory information fusion as described in claim 6, characterized in that, Traversing the standard fermentation state vector set and comparing it with the real-time fermentation state vector includes: The Euclidean distance between each standard fermentation state vector and the real-time fermentation state vector is calculated by iterating through the data. The real-time fermentation stage corresponding to the real-time fermentation state vector is determined based on the standard fermentation state vector corresponding to the smallest of the multiple Euclidean distances. Calculate the deviation vector between the standard fermentation state vector and the real-time fermentation state vector corresponding to the real-time fermentation stage, and combine the deviation vector with the real-time fermentation stage to output the comparison analysis result.

8. The intelligent control method for rice wine fermentation based on electronic sensory information fusion as described in claim 1, characterized in that, The standard fermentation state vector set includes at least the standard saccharification fermentation state vector, the standard acid-producing fermentation state vector, and the standard termination fermentation state vector.

Images