High-yield ligustrazine data-driven modeling method and system in industrial vinegar fermentation process

Abstract

The invention discloses a high-yield ligustrazine data-driven modeling method and system in an industrial vinegar fermentation process, and the method comprises the steps: carrying out the preprocessing of original multi-dimensional time series data in the vinegar fermentation process, marking the technology type, and constructing a condition training data set; constructing a time sequence generative adversarial network model fused with process semantics, and enabling the model to learn time sequence data distribution characteristics corresponding to different processes through adversarial training with process types as conditions; inputting a target process type to generate synthetic fermentation data, and evaluating the quality of the synthetic fermentation data from two dimensions of statistical characteristics and time sequence dynamics; and optimizing the generation model according to an evaluation result to obtain a final data driving model. According to the method, through process semantic fusion and a dynamic optimization mechanism, the problems of large modeling deviation and insufficient generalization ability in small-sample, high-dimension and strong-nonlinearity scenes are effectively solved, and the prediction precision of the key substances in the fermentation process and the model reliability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of bio-fermentation engineering and artificial intelligence algorithm technology, and in particular to a data-driven modeling method and system for high-yield tetramethylpyrazine production in industrial vinegar fermentation process. Background Technology

[0002] As a typical nonlinear, multi-stage biochemical system, vinegar fermentation has long faced a triple challenge in its industrial scale-up and optimization: process complexity, monitoring limitations, and data scarcity. Specifically, the quantitative analysis of key flavor compounds (such as acetoin and tetramethylpyrazine) typically relies on destructive offline chromatography techniques, resulting in low-frequency time-series data acquisition and sparse sample points, making it difficult to fully characterize the dynamic behavior of fermentation. Simultaneously, strong nonlinear coupling relationships exist among multidimensional parameters such as temperature, alcohol content, and total acidity, collectively influencing microbial metabolic pathways and enzyme reaction kinetics. This makes traditional univariate control strategies difficult to implement effectively, placing higher demands on multivariate collaborative modeling. Furthermore, different production processes (such as solid-state fermentation and mechanically enhanced fermentation) exhibit significant differences in metabolic mechanisms, and the high cost of detection further limits the scale and diversity of modeling data, leading to a modeling dilemma of "small sample size, high dimensionality, and strong nonlinearity" for the vinegar fermentation system as a whole.

[0003] In existing technologies, various modeling methods have been attempted for fermentation process analysis, but significant limitations remain. Mechanistic models or metabolic network models based on differential equations often oversimplify microbial symbiosis and competition, leading to deviations from reality in the simulation of key flavor compound generation dynamics. Traditional statistical methods such as Principal Component Analysis (PCA) suffer from severe loss of key process information and insufficient model generalization ability when performing dimensionality reduction under small sample conditions. Statistical regression models (such as Partial Least Squares Regression, PLS) and conventional neural network models also struggle to effectively capture the complex intrinsic relationships between variables when faced with high-dimensional, sparse, and strongly coupled data structures due to insufficient sample size. Furthermore, although generative adversarial networks (GANs) have the potential to synthesize multivariate sequences, standard GAN models lack explicit constraints on temporal dynamics and biochemical laws, resulting in significant deficiencies in temporal consistency and physical interpretability, limiting their reliable application in accurate fermentation process modeling. Therefore, existing technologies have not yet provided a modeling scheme for vinegar fermentation systems that can accurately reproduce the dynamic coupling of multiple processes and variables based on a small number of samples. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the problems in the prior art of vinegar fermentation process, which are caused by data sparsity, strong nonlinear coupling and significant process differences, resulting in large small sample modeling bias and insufficient model generalization ability, as well as the lack of temporal dynamic constraints in standard generative adversarial networks, which cannot simultaneously guarantee the statistical accuracy and temporal rationality of the generated data.

[0005] To address the aforementioned technical problems, this invention provides a data-driven modeling method for high-yield tetramethylpyrazine production during industrial vinegar fermentation, comprising the following steps: S1: Preprocess the original multidimensional time series data of vinegar fermentation process and label the corresponding process type to obtain the conditional training dataset; S2: Construct a temporal generative adversarial network model that integrates process semantics, and train the temporal generative adversarial network model using the conditional training dataset. Through adversarial training with process type as the condition, the temporal generative adversarial network model learns the temporal data distribution characteristics corresponding to different process types, and obtains a trained generative model. S3: Input the target process type into the generation model to obtain synthetic fermentation data, and evaluate the synthetic fermentation data to obtain quality evaluation results; S4: Optimize the generated model based on the quality assessment results to obtain the final data-driven model.

[0006] In one embodiment of the present invention, in step S1, the method for preprocessing the original multidimensional time series data in the vinegar fermentation process and labeling the corresponding process type to obtain the conditional training dataset is as follows: the original multidimensional time series data is cleaned, scaled and normalized, and time series segmented, and each processed data subsequence is labeled with its corresponding process type to obtain the conditional training dataset.

[0007] In one embodiment of the present invention, the method for performing time-series segmentation on the original multidimensional time-series data is to use a sliding window with a preset window size and step size to segment the time-series data after data cleaning and scale normalization processing, generating multiple subsequences that overlap in time.

[0008] In one embodiment of the present invention, the method for constructing the temporal generative adversarial network model that integrates process semantics in step S2 is: to construct an adversarial network composed of a generator and a discriminator; The generator includes an encoder and a decoder. The encoder takes the real time-series data sequence extracted from the conditional training dataset as input, maps it into a low-dimensional latent space vector, concatenates the low-dimensional latent space vector with the process type label corresponding to the real time-series data sequence, and inputs it into the decoder to obtain the generated reconstructed time-series data sequence. The discriminator opposes the generator and is used to determine whether the input data sequence is from real data or a reconstructed time series data sequence generated by the generator, and at the same time to determine its process type.

[0009] In one embodiment of the present invention, step S2, the method for training the temporal generative adversarial network model using the conditional training dataset is as follows: the conditional training dataset is divided into a training set and a validation set, the model is trained on the training set using an initial learning rate, the learning rate is reduced and training continues after the training reaches a preset number of rounds, and the changes in the reconstruction loss on the validation set are monitored at the same time, and the training process is terminated when the reconstruction loss no longer decreases for several consecutive training rounds.

[0010] In one embodiment of the present invention, the method for evaluating the synthetic fermentation data and obtaining the quality evaluation result in step S3 is as follows: the consistency between the synthetic fermentation data and the real data is quantitatively evaluated from two dimensions: statistical characteristics and time-series dynamics, to obtain the quality evaluation result.

[0011] In one embodiment of the present invention, step S4, the method for optimizing the generative model based on the quality assessment result, is as follows: Based on the quality assessment results, when the assessment results of the statistical feature dimension show that the numerical deviation exceeds a preset threshold, the weight of the relevant features in the generator in the loss function is increased; When the evaluation result of the time-series dynamic dimension shows that the coupling distortion exceeds a preset threshold, the number of layers or units in the neural network of the generator or discriminator is increased.

[0012] This invention also provides a data-driven modeling system for high-yield tetramethylpyrazine production during industrial vinegar fermentation, comprising the following modules: The data preprocessing module is used to preprocess the raw multidimensional time-series data of vinegar fermentation process and label the corresponding process types to obtain the conditional training dataset. The model training module is used to construct a temporal generative adversarial network model that integrates process semantics. The temporal generative adversarial network model is trained using the conditional training dataset. Through adversarial training with process type as the condition, the temporal generative adversarial network model learns the temporal data distribution characteristics corresponding to different process types, and obtains a trained generative model. The data generation and evaluation module is used to input the target process type into the generation model, obtain synthetic fermentation data, evaluate the synthetic fermentation data, and obtain quality evaluation results. The model optimization module is used to optimize the generated model based on the quality assessment results to obtain the final data-driven model.

[0013] The present invention also provides an electronic device, which includes a processor, a memory, and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to realize the data-driven modeling method for high-yield ligustrazine in the industrial vinegar fermentation process.

[0014] The present invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions for causing a computer device to execute the data-driven modeling method for high-yield tetramethylpyrazine in the industrial vinegar fermentation process.

[0015] The technical solution of the present invention has the following advantages compared with the prior art: This invention proposes a data-driven modeling method and system for vinegar fermentation based on temporal generative adversarial networks (GANs), constructing a complete technical system from data preprocessing, model building, training optimization to generation and evaluation. By introducing conditional generative adversarial training with process semantics, the model can accurately learn the multivariate coupling dynamics under different processes. An adaptive optimization feedback mechanism based on multidimensional evaluation indicators is established to ensure that the generated data has both statistical accuracy and temporal rationality. This effectively overcomes the inherent problems of traditional modeling techniques, such as large prediction biases of key substances and insufficient model generalization ability in small-sample, high-dimensional, and strongly nonlinear scenarios, significantly improving the accuracy and reliability of digital modeling of the fermentation process. It provides an innovative technical solution for the quantitative analysis, optimization control, and intelligent upgrading of vinegar production processes. Attached Figure Description

[0016] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0017] Figure 1 This is a schematic flowchart of a data-driven modeling method for high-yield tetramethylpyrazine production in industrial vinegar fermentation provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the three-dimensional relationship between fermentation time, temperature, and tetramethylpyrazine under traditional processes in this embodiment of the invention. Figure 3 This is a schematic diagram of the three-dimensional relationship between fermentation time, temperature, and acetoin under traditional processes in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the three-dimensional relationship between fermentation time, total acid, and tetramethylpyrazine under traditional processes in this embodiment of the invention. Figure 5 This is a schematic diagram illustrating the three-dimensional relationship between fermentation time, total acid, and acetoin under traditional processes in this embodiment of the invention. Figure 6This is a schematic diagram illustrating the three-dimensional relationship between fermentation time, alcohol content, and tetramethylpyrazine under traditional processes in this embodiment of the invention. Figure 7 This is a schematic diagram illustrating the three-dimensional relationship between fermentation time, alcohol content, and acetoin under traditional processes in this embodiment of the invention. Figure 8 This is a schematic diagram of the three-dimensional relationship between fermentation time, temperature, and tetramethylpyrazine under mechanical processing in an embodiment of the present invention; Figure 9 This is a schematic diagram of the three-dimensional relationship between fermentation time, temperature, and acetoin under mechanical processing in an embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the three-dimensional relationship between fermentation time, total acid, and tetramethylpyrazine under mechanical processing in an embodiment of the present invention. Figure 11 This is a schematic diagram of the three-dimensional relationship between fermentation time, total acid, and acetoin under mechanical processing in an embodiment of the present invention; Figure 12 This is a schematic diagram illustrating the three-dimensional relationship between fermentation time, alcohol content, and tetramethylpyrazine under mechanical processing in an embodiment of the present invention; Figure 13 This is a schematic diagram illustrating the three-dimensional relationship between fermentation time, alcohol content, and acetoin under mechanical processing in an embodiment of the present invention. Figure 14 This is a schematic diagram comparing the actual value and the model-generated value of ligustrazine under the traditional process in an embodiment of the present invention; Figure 15 This is a schematic diagram comparing the actual value and the model-generated value of acetoin under the traditional process in an embodiment of the present invention; Figure 16 This is a schematic diagram comparing the actual value and the model-generated value of ligustrazine under mechanical processing in an embodiment of the present invention; Figure 17 This is a schematic diagram comparing the actual value and the model-generated value of acetylene under mechanical processing in an embodiment of the present invention; Figure 18 This is a schematic diagram of the structure of a data-driven modeling system for high-yield tetramethylpyrazine in the industrial vinegar fermentation process provided in this embodiment of the invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0019] Example 1: like Figure 1 As shown, the present invention provides a data-driven modeling method for high-yield tetramethylpyrazine production in industrial vinegar fermentation, comprising the following steps: S1: Preprocess the original multidimensional time series data of vinegar fermentation process and label the corresponding process type to obtain the conditional training dataset; S2: Construct a temporal generative adversarial network model that integrates process semantics, and train the temporal generative adversarial network model using the conditional training dataset. Through adversarial training with process type as the condition, the temporal generative adversarial network model learns the temporal data distribution characteristics corresponding to different process types, and obtains a trained generative model. S3: Input the target process type into the generation model to obtain synthetic fermentation data, and evaluate the synthetic fermentation data to obtain quality evaluation results; S4: Optimize the generated model based on the quality assessment results to obtain the final data-driven model.

[0020] The data-driven modeling method for vinegar fermentation in this invention constructs a complete technical framework: First, multi-dimensional time-series data during fermentation is preprocessed and labeled with process types to form a conditional training set; then, a time-series generative adversarial network (GAN) integrating process semantics is established, and the data distribution characteristics under different processes are learned through conditional adversarial training; subsequently, synthetic data is generated based on the target process and multi-dimensional quality assessment is performed; finally, the model is iteratively optimized based on the assessment results to form an adaptive modeling closed loop. This innovative approach, through process semantic fusion and dynamic optimization mechanisms, overcomes the modeling bottleneck of traditional methods in small-sample, strongly nonlinear coupling scenarios, significantly improving the accuracy of key process parameters and flavor compound predictions, and providing reliable model support for precise control and process optimization of the fermentation process.

[0021] In step S1 of this embodiment, the original multidimensional time series data of vinegar fermentation process is preprocessed and labeled with the corresponding process type to obtain a conditional training dataset. The preprocessing stage provides a foundation for subsequent modeling work through the data processing flow.

[0022] Specifically, the original multidimensional time-series data is cleaned, scaled, and segmented into time series, and each processed data subsequence is labeled with its corresponding process type to obtain the conditional training dataset.

[0023] The method for performing time-series segmentation on the original multidimensional time-series data is to use a sliding window with a preset window size and step size to segment the time-series data after data cleaning and scale normalization, generating multiple subsequences that overlap in time.

[0024] To ensure the complete reproducibility of the deep learning model training process and effectively control the impact of random factors on the experimental results, this embodiment adopts a multi-level random seed fixing mechanism.

[0025] Specifically, the initial state of the random number generator is first set. , making the first The result of the random sampling Can be derived from functions Confirmed, among which This is a pseudo-random number generation function. Based on this, the random seeds at three levels—NumPy, deep learning framework, and system level—are uniformly set: the NumPy random seed `np.random.seed(42)` is fixed to control random perturbations during data augmentation and interpolation; the TensorFlow / PyTorch random seed `tf.random.set_seed(42)` is fixed to control the calculation of statistics in model parameter initialization and batch normalization; and the system-level random seed `random.seed(42)` is fixed to control the randomness of data shuffling and batch partitioning.

[0026] Through the above systematic setup, random operations in each stage of model training can be fully reproduced, thereby significantly improving the stability and comparability of the experiment and effectively avoiding inconsistencies in training results caused by random fluctuations.

[0027] Regarding data integrity handling, a multi-level interpolation repair strategy was adopted to address the potential missing values ​​in fermentation process data collection. Interpolation functions were constructed based on valid data points, prioritizing computationally robust linear interpolation to ensure data continuity. Simultaneously, considering the dynamic characteristics of multidimensional variables, cubic interpolation was flexibly selected to adapt to different data distribution patterns. In special cases where valid data points are severely insufficient, first or last value padding was used to ensure the structural integrity of the sequence. This hierarchical interpolation strategy ensures data integrity while preserving the dynamic characteristics of the original data to the greatest extent possible.

[0028] In the feature standardization process, MinMaxScaler is used to normalize the processed data, linearly mapping all features to the [0,1] interval. This process completely eliminates the dimensional differences between parameters such as temperature (°C), alcohol content (% vol), and total acidity (g / 100mL), unifying the numerical range of different features and significantly improving the convergence efficiency of subsequent model training and the stability of parameter optimization.

[0029] The time series segmentation process uses preset sliding window parameters, with a window size of 10 and a step size of 1, to systematically segment the cleaned and normalized time series data. This transforms the continuous vinegar fermentation process into multiple subsequences that are highly overlapping in time. This not only significantly enhances the model's ability to capture local temporal dynamic features, but also provides a more representative input data structure for subsequent modeling.

[0030] Furthermore, each processed data subsequence is labeled with its corresponding process type, forming a conditional training dataset with explicit process semantics. This labeling work provides the necessary conditional information for subsequent training of generative adversarial networks that fuse process features, enabling the model to effectively learn the dynamic differences under different process conditions.

[0031] Through this series of rigorous preprocessing operations, not only were problems such as missing data, inconsistent dimensions, and insufficient continuity in the original data resolved, but a high-quality, standardized training dataset with clear process annotations was also constructed. This provided a reliable data foundation for the effective training and accurate modeling of subsequent temporal generative adversarial networks, significantly improving the stability and reproducibility of the entire modeling process.

[0032] Furthermore, in step S2, after completing data preprocessing and constructing a conditional training dataset, a temporal generative adversarial network model that integrates process semantics is constructed and trained. A generative model that can fully learn the distribution characteristics of time series data under different process types is designed. Through the adversarial training mechanism, high-precision modeling of small sample fermentation processes is achieved. Through a systematic training strategy, it can accurately learn the dynamic distribution characteristics of time series under different process conditions from small sample fermentation data, thereby breaking through the limitations of traditional modeling methods in data-scarce scenarios.

[0033] Specifically, this embodiment constructs a conditional adversarial network system consisting of a generator and a discriminator. The generator adopts an encoder-decoder architecture as its core. The encoder first extracts real multivariate time-series data sequences (including temperature, alcohol content, total acidity, etc.) from the conditional training dataset, and nonlinearly maps them into a low-dimensional latent space vector through a multi-layer neural network. This process not only achieves efficient data compression, but also captures and extracts the most representative dynamic features from the original high-dimensional sequences.

[0034] The input to the model is defined as follows: Let the input multivariate time series be... Where T is the time step (corresponding to fermentation days), and D is the variable dimension (including temperature, alcohol content, total acid, tetramethylpyrazine, acetoin, and derived characteristics); the process condition labels are... (0 = traditional process, 1 = mechanical process), obtained after embedding and expansion. The sequence length is matched by copying along the time dimension. The final encoder input is a fusion vector of features and process conditions: , The encoder uses and Structure, Bi-LSTM layer for fused input Perform bidirectional temporal feature extraction, and the output is: , Where H represents unidirectional Hidden layer dimension (set to 64 in this embodiment). and They are respectively The weights and bias parameters.

[0035] To avoid overfitting, Dropout regularization is applied to the Bi-LSTM output with a Dropout rate of 0.3. The regularized output is... Through a single layer Compress bidirectional features into a fixed-length latent vector. L is the dimension of the latent space (set to 32 in this embodiment), which serves as the input to the generator and decoder. .

[0036] Furthermore, the encoder training is optimized through backpropagation of the decoder's reconstruction loss, ensuring that the latent vector z retains key information from the original data. The reconstruction loss is defined as the mean squared error (MSE), and its calculation method is as follows: , in Let be the value of the reconstructed sequence output by the decoder at step t and dimension d.

[0037] Subsequently, the model innovatively introduces a RepeatVector layer. The RepeatVector layer copies the low-dimensional latent vector output by the encoder along the temporal dimension, generating a vector sequence with the exact same length as the original input sequence: , It significantly enhances the model's ability to model long-term temporal dependencies and solves the gradient vanishing or forgetting problem that traditional recurrent neural networks are prone to when processing long sequences.

[0038] Meanwhile, process type is embedded as a key conditional variable in the generation process: the process type label is expanded in dimension and then concatenated with the latent vector sequence expanded by the RepeatVector layer, thus forming a feature matrix that deeply integrates process semantics. This matrix is ​​then fed into the decoder, which is typically composed of time-sensitive units such as LSTM (Long Short-Term Memory) layers, and is responsible for reconstructing the feature representation that integrates process information into a synthetic time-series data sequence that conforms to the specific process dynamics.

[0039] The restorer combines LSTM layers, temporal attention mechanisms, and fully connected layer structures. The LSTM layers process the expanded latent vectors. Perform feature expansion along the time dimension, and the output is: , in, , The weights and bias parameters for the LSTM layer are set to H, which is set to 64 to match the encoder.

[0040] This explicit process semantic fusion mechanism enables the generator to accurately learn and reproduce the dynamic patterns unique to different processes, such as the typical rapid rise in high temperature and subsequent decay characteristics in traditional processes.

[0041] In this embodiment, the discriminator is designed with an architecture that opposes the generator, and it is given a dual discriminative mission. The discriminator needs to determine not only whether the input time-series data sequence originates from a real fermentation process or is synthesized by the generator, but also the corresponding process type. To enhance its ability to identify subtle process features, an additional fully connected layer is added to the discriminator's network structure to improve its nonlinear classification capabilities.

[0042] This adversarial environment with dual discrimination constitutes a more rigorous and refined training framework. It forces the generator to continuously optimize the data it generates, making it statistically close to the real data and highly consistent with the target process label in terms of process characteristics. This effectively solves the core problems of standard GAN models, such as lack of process-specific constraints and poor physical interpretability of generated data.

[0043] During model training, a refined training strategy was employed to optimize model performance and prevent overfitting. The conditional training dataset was proportionally divided into a training set for parameter updates and a validation set for monitoring generalization ability. An adaptive learning rate scheduling strategy was used during training, with specific parameters set as follows: batch size of 32, latent space dimension of 16, and LSTM hidden layer dimension of 64. A relatively high initial learning rate (1e-3) was used in the early stages of training to accelerate model parameter convergence and quickly capture the macroscopic distribution characteristics of the data. After 30 training iterations, an exponential decay strategy was initiated, reducing the learning rate by 5%, gradually transitioning the training to a fine-tuning phase. This helps the model converge to a better local optimum and effectively suppresses overfitting. Simultaneously, an early stopping mechanism (patience=8) was introduced to continuously monitor the reconstruction loss on the validation set. If the loss did not decrease for eight consecutive training iterations, training was automatically terminated, thus achieving an optimal balance between model performance and computational efficiency. After approximately 150 rounds of iterative training, the validation set loss tends to stabilize, indicating that the model has fully learned the time series data distribution characteristics of different process types and has reached a convergence state.

[0044] Based on the above, the temporal generative adversarial network model that integrates process semantics constructed in this embodiment, through the organic combination of encoder-decoder and conditional adversarial training, relies on the enhancement of temporal structure by the RepeatVector layer, the conditional generation guidance of process labels, and the dual constraints of the discriminator, and is further combined with the robust training strategy of adaptive learning rate scheduling and early stopping, together constitute a generative model that can accurately learn and reproduce the nonlinear dynamics of vinegar fermentation under the complex conditions of small samples and multiple processes.

[0045] Furthermore, after the model training is completed and convergence is achieved, in step S3, fermentation data is synthesized using the trained generative model and a systematic quality assessment is performed.

[0046] Specifically, a specific target process type (such as label "0" for traditional processes or label "1" for mechanical processes) is input into the generative model. By randomly sampling in the latent space, a number of latent vectors equal to the number in the real dataset are generated. These latent vectors, along with the target process labels, are input into the generator, where their decoder network reconstructs the corresponding synthetic fermentation time-series data. To ensure that the generated data conforms to the biophysical laws of the fermentation process, a non-negativity constraint is imposed on the output data, requiring all physical quantities such as concentration and temperature to be non-negative, thereby guaranteeing the physical rationality of the synthesized data.

[0047] The data quality assessment stage constructs a multi-dimensional quantitative evaluation system to accurately measure the consistency between synthetic and real data from two core dimensions: statistical characteristic consistency and temporal dynamic reproduction degree. In the statistical characteristic dimension, multiple indicators are used for comprehensive evaluation: mean squared error (MSE) is calculated to quantify the average squared deviation between predicted and actual values, and is highly sensitive to outliers; mean absolute error (MAE) is calculated to provide a more robust measure of average deviation; the coefficient of determination (R²) is calculated to assess the degree to which the generated data explains the variability of the real data; and Wasserstein distance is used to measure the similarity of the probability distribution between the generated and real data, assessing the consistency of their joint distribution. In the temporal dynamic dimension, the focus is on evaluating the reproduction degree of the coupling relationship between key parameters, such as the degree of agreement between the temperature-total acidity dynamic correlation curve and the actual fermentation process, and the accuracy of temporal dynamics is quantified through morphological comparison analysis.

[0048] , , The Wasserstein distance measures the distance between two probability distributions, reflecting the degree of matching between the generated data distribution and the real data distribution. The Wasserstein distance is calculated as follows: , in, The probability distribution of the real data. To generate the probability distribution of the data, For all cases that simultaneously satisfy the marginal distribution, and The joint distribution set, This indicates that the boundary has been removed.

[0049] In addition, depending on the specific evaluation needs, traditional distance measures such as Euclidean distance, absolute distance, and Mahalanobis distance, as well as information theory indicators such as KL divergence and JS divergence, can be used as supplementary evaluation methods to comprehensively quantify the similarity between generated data and real data in various feature dimensions from different perspectives.

[0050] Furthermore, in step S4, the generated model is optimized in a targeted manner. The optimization process adopts a differentiated strategy based on the specific performance of the evaluation results. Specifically, when the evaluation of the statistical feature dimension shows that the concentration deviation of a specific key substance (such as acetoin) exceeds a preset threshold of 8%, the model's learning accuracy in generating the key substance is enhanced by increasing the weight of the corresponding feature in the loss function in the generator. When the evaluation of the time-series dynamic dimension shows that the distortion of the coupling relationship between variables exceeds a preset threshold of 10%, the model capacity is expanded by increasing the number of layers or units of the neural network in the generator or discriminator, thereby enhancing its ability to fit complex nonlinear coupling relationships.

[0051] This refined optimization mechanism based on multi-indicator feedback enables the model to continuously optimize during the iteration process, significantly improving the quality of the generated data and the model's generalization ability. Ultimately, after multiple rounds of generation-evaluation-optimization iterations, a final data-driven model capable of accurately simulating the dynamics of vinegar fermentation under different processes is obtained.

[0052] Furthermore, in this embodiment, after data generation and evaluation are completed, the differences between the generated data and the real data are presented intuitively through visualization analysis, further verifying the effectiveness and accuracy of the temporal generative adversarial network model.

[0053] Specifically, using various visualization methods such as three-dimensional relationship diagrams and key substance comparison diagrams, a detailed comparative analysis of the fermentation process under traditional and mechanical processes is conducted.

[0054] In this invention, visualization analysis is achieved through a three-dimensional time-series correlation map and key feature comparison curves. The core mathematical processing is as follows: Three-dimensional map: A three-dimensional coordinate system is constructed using "fermentation time t (X-axis) - temperature T (Y-axis) - key substance concentration C (Z-axis)". Kernel density estimation (KDE) is used to fill the surface, and the formula is: , in The Gaussian kernel function, with h as the bandwidth, smooths the surface while preserving peak characteristics; The moving average (window size = 3) calculated for key features of real / generated data (such as acetaminophen) is as follows: , After noise removal, a comparison curve is plotted to visually quantify the differences in peak position and rise / fall rate.

[0055] like Figures 2 to 13 As shown, three-dimensional correlation maps of fermentation time and temperature, alcohol content, total acid and key flavor compounds (tetramethylpyrazine and acetoin) were plotted, clearly demonstrating the dynamic change trends and coupling relationships of each variable under different process conditions.

[0056] At the same time, such as Figures 14 to 17 As shown, a comparison curve between the measured values ​​and the model-generated values ​​of key substances (tetramethylpyrazine and acetoin) is provided. Through intuitive graphical comparison, the consistency between the generated data and the real data is quantitatively evaluated.

[0057] The three-dimensional relationship diagram reveals significant differences in the trends of key variables between the mechanical and traditional processes, directly impacting fermentation efficiency and product quality. Regarding temperature changes, the mechanical process exhibits a stable upward trend, gradually increasing from an initial 28℃ to 46℃, followed by a slow decline. In contrast, the traditional process rapidly climbs to 45℃ before experiencing a rapid decline in the middle stages, with more drastic temperature fluctuations, reaching a difference of up to 17℃, demonstrating significantly lower stability than the mechanical process. As for alcohol content changes, the mechanical process shows a continuous and gradual decline, decreasing from an initial 3.75% vol to 0.74% vol. The traditional process, however, declines rapidly, from 3.77% vol to 0.42% vol, a rate 2.3 times faster than the mechanical process, indicating more vigorous alcohol metabolism in the traditional process. Regarding total acid production, the mechanical process showed a steady upward trend, increasing from 0.72 g / 100 ml to 3.42 g / 100 ml; the traditional process, however, saw a rapid increase, ultimately reaching 3.86 g / 100 ml, with a peak value 12.9% higher than the mechanical process, indicating that the traditional process had a greater advantage in acid production capacity. As for the formation pattern of tetramethylpyrazine, the mechanical process produced only trace amounts of tetramethylpyrazine in the later stages, with a maximum of no more than 0.23 mg / L; the traditional process, on the other hand, saw rapid growth in the middle stages, increasing from 0 to 0.37 mg / L, with a peak content 61% higher than the mechanical process. The differences were even more pronounced in acetoin production: the mechanical process showed a continuous linear increase in acetoin, ultimately reaching 2933 mg / L; the traditional process first rapidly reached its peak (3666 mg / L), then quickly declined to 720 mg / L. Although the peak value was 25% higher than the mechanical process, the decline rate was faster, and the later accumulation was far lower than that of the mechanical process.

[0058] Through the above visualization analysis, it can be concluded that temperature, as a core regulatory factor, provides more favorable metabolic conditions for the large-scale synthesis of acetoin by optimizing temperature maintenance in the mechanical process, resulting in slow consumption of alcohol and moderate accumulation of total acid. In contrast, the traditional process, due to its poor temperature stability, is less effective than the mechanical process in terms of acetoin accumulation and tetramethylpyrazine conversion. The inverse relationship between alcohol content and total acid reflects the metabolic intensity and dynamic balance, which is more conducive to acetoin storage in the mechanical process. Ultimately, the mechanical process can significantly increase the content of functional components such as tetramethylpyrazine, strengthening the coupling chain of "temperature-regulated metabolism - alcohol-acid balance affecting precursor accumulation - driving key component conversion". The higher and more stable fermentation temperature (mechanical process) is conducive to the large-scale synthesis of acetoin. At the end of fermentation, the acetoin content of the mechanical process was 2933.33 mg / L, which is more than 4 times that of the traditional process (706.67 mg / L). In addition, the statistical information of the generated data was saved in an Excel file with attached charts for subsequent in-depth analysis and report generation. This results presentation system not only provides an intuitive visual representation, but also supports the interpretability of the temporal generative adversarial network model through quantitative data, providing a clear direction for the optimization and intelligent control of vinegar production processes.

[0059] Example 2: like Figure 18 As shown, based on the same inventive concept as in Embodiment 1, this invention also provides a data-driven modeling system for high-yield tetramethylpyrazine production in industrial vinegar fermentation processes, used to implement the steps of the data-driven modeling method for high-yield tetramethylpyrazine production in industrial vinegar fermentation processes described in Embodiment 1, including the following modules: The data preprocessing module is used to preprocess the raw multidimensional time-series data of vinegar fermentation process and label the corresponding process types to obtain the conditional training dataset. The model training module is used to construct a temporal generative adversarial network model that integrates process semantics. The temporal generative adversarial network model is trained using the conditional training dataset. Through adversarial training with process type as the condition, the temporal generative adversarial network model learns the temporal data distribution characteristics corresponding to different process types, and obtains a trained generative model. The data generation and evaluation module is used to input the target process type into the generation model, obtain synthetic fermentation data, evaluate the synthetic fermentation data, and obtain quality evaluation results. The model optimization module is used to optimize the generated model based on the quality assessment results to obtain the final data-driven model.

[0060] The data preprocessing module, model training module, and data generation and model optimization module of the high-yield tetramethylpyrazine data-driven modeling system for the industrial vinegar fermentation process proposed in this embodiment are respectively used to implement steps S1, S2, S3 and S4 in the high-yield tetramethylpyrazine data-driven modeling method for the industrial vinegar fermentation process in Embodiment 1. To avoid redundancy, they will not be described in detail here.

[0061] Example 3: The present invention also provides an electronic device, which includes a processor, a memory, and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the data-driven modeling method for high-yield ligustrazine in the industrial vinegar fermentation process described in Embodiment 1.

[0062] Example 4: The present invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions for causing a computer device to execute the data-driven modeling method for high-yield tetramethylpyrazine in the industrial vinegar fermentation process described in Embodiment 1.

[0063] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0064] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0067] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A data-driven modeling method for high-yield tetramethylpyrazine production during industrial vinegar fermentation, characterized in that, Includes the following steps: S1: Preprocess the original multidimensional time series data of vinegar fermentation process and label the corresponding process type to obtain the conditional training dataset; S2: Construct a temporal generative adversarial network model that integrates process semantics, and train the temporal generative adversarial network model using the conditional training dataset. Through adversarial training with process type as the condition, the temporal generative adversarial network model learns the temporal data distribution characteristics corresponding to different process types, and obtains a trained generative model. S3: Input the target process type into the generation model to obtain synthetic fermentation data, and evaluate the synthetic fermentation data to obtain quality evaluation results; S4: Optimize the generated model based on the quality assessment results to obtain the final data-driven model.

2. The data-driven modeling method for high-yield tetramethylpyrazine production in industrial vinegar fermentation process according to claim 1, characterized in that: In step S1, the method for preprocessing the original multidimensional time-series data in the vinegar fermentation process and labeling the corresponding process type to obtain the conditional training dataset is as follows: the original multidimensional time-series data is cleaned, scaled, and segmented into time series, and each processed data subsequence is labeled with its corresponding process type to obtain the conditional training dataset.

3. The data-driven modeling method for high-yield tetramethylpyrazine production in industrial vinegar fermentation process according to claim 2, characterized in that: The method for performing time-series segmentation on the original multidimensional time-series data is to use a sliding window with a preset window size and step size to segment the time-series data after data cleaning and scale normalization, generating multiple subsequences that overlap in time.

4. The data-driven modeling method for high-yield tetramethylpyrazine production in industrial vinegar fermentation process according to claim 1, characterized in that: In step S2, the method for constructing a temporal generative adversarial network model that integrates process semantics is to construct an adversarial network consisting of a generator and a discriminator. The generator includes an encoder and a decoder. The encoder takes the real time-series data sequence extracted from the conditional training dataset as input, maps it into a low-dimensional latent space vector, concatenates the low-dimensional latent space vector with the process type label corresponding to the real time-series data sequence, and inputs it into the decoder to obtain the generated reconstructed time-series data sequence. The discriminator opposes the generator and is used to determine whether the input data sequence is from real data or a reconstructed time series data sequence generated by the generator, and at the same time to determine its process type.

5. The data-driven modeling method for high-yield tetramethylpyrazine production in industrial vinegar fermentation process according to claim 1, characterized in that: In step S2, the method for training the temporal generative adversarial network model using the conditional training dataset is as follows: the conditional training dataset is divided into a training set and a validation set, the model is trained on the training set using an initial learning rate, the learning rate is reduced and training continues after the training reaches a preset number of rounds, and the changes in the reconstruction loss on the validation set are monitored. The training process is terminated when the reconstruction loss no longer decreases for several consecutive training rounds.

6. The data-driven modeling method for high-yield tetramethylpyrazine production in industrial vinegar fermentation process according to claim 1, characterized in that: In step S3, the method for evaluating the synthetic fermentation data and obtaining the quality evaluation result is as follows: the consistency between the synthetic fermentation data and the real data is quantitatively evaluated from two dimensions: statistical characteristics and time-series dynamics, to obtain the quality evaluation result.

7. The data-driven modeling method for high-yield tetramethylpyrazine production in industrial vinegar fermentation process according to claim 6, characterized in that: In step S4, the method for optimizing the generated model based on the quality assessment results is as follows: Based on the quality assessment results, when the assessment results of the statistical feature dimension show that the numerical deviation exceeds a preset threshold, the weight of the relevant features in the generator in the loss function is increased; When the evaluation result of the time-series dynamic dimension shows that the coupling distortion exceeds a preset threshold, the number of layers or units in the neural network of the generator or discriminator is increased.

8. A data-driven modeling system for high-yield tetramethylpyrazine production during industrial vinegar fermentation, characterized in that, Includes the following modules: The data preprocessing module is used to preprocess the raw multidimensional time-series data of vinegar fermentation process and label the corresponding process types to obtain the conditional training dataset. The model training module is used to construct a temporal generative adversarial network model that integrates process semantics. The temporal generative adversarial network model is trained using the conditional training dataset. Through adversarial training with process type as the condition, the temporal generative adversarial network model learns the temporal data distribution characteristics corresponding to different process types, and obtains a trained generative model. The data generation and evaluation module is used to input the target process type into the generation model, obtain synthetic fermentation data, evaluate the synthetic fermentation data, and obtain quality evaluation results. The model optimization module is used to optimize the generated model based on the quality assessment results to obtain the final data-driven model.

9. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the data-driven modeling method for high-yield ligustrazine in the industrial vinegar fermentation process as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, which includes several instructions for causing a computer device to execute the data-driven modeling method for high-yield ligustrazine production in the industrial vinegar fermentation process as described in any one of claims 1 to 7.

Images