Method and system for predicting and evaluating fermentation quality of baijiu

Abstract

The application discloses a kind of white liquor fermentation quality prediction evaluation method and system, comprising: the environmental parameter of collection wine fermentation process, obtains the evaluation parameter of fermentation quality;Fermentation quality evaluation model is constructed, and environmental parameter and evaluation parameter are regarded as sample data and input to fermentation quality evaluation model, the training of network model is carried out, and the trained evaluation model is obtained;The environmental parameter to be measured in wine fermentation process is input to the trained evaluation model, and the predicted evaluation parameter is output;Using the evaluation parameter of prediction carries out evaluation to fermentation quality, and obtains evaluation result.The present application can significantly reduce human interference, and quickly obtain evaluation parameter using network model, to realize reliable and effective evaluation to fermentation quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of liquor fermentation, in particular to a method and system for predicting and evaluating the quality of liquor fermentation. BACKGROUND

[0002] The most significant embodiment of the quality of fermented grains during the fermentation process is the quality of base liquor, and the liquor aroma is mainly derived from the types and contents of chemical substances such as alcohols, esters, acids, aldehydes and phenols produced by biochemical reactions during the fermentation process, which have an important influence on the quality of fermented grains during the fermentation process.

[0003] Currently, the quality of liquor fermentation is usually predicted and evaluated by chemical analysis methods. Traditional chemical analysis methods, such as gas chromatography, liquid chromatography and spectral analysis, have high requirements for equipment, complex operation, and require personnel with professional skills to operate, and usually require complex sample pretreatment steps, which not only increases the experimental time and consumption, but also may introduce human error. In addition, these methods can only analyze certain specific components in the sample, and it is difficult to comprehensively reflect the changes of various substances in the fermentation process, which limits the reliable evaluation of the fermentation quality.

[0004] Therefore, a new technology for predicting and evaluating the quality of liquor fermentation is needed, which can significantly reduce human interference, quickly obtain evaluation parameters using a network model, and realize reliable and effective evaluation of the fermentation quality. SUMMARY

[0005] Therefore, the purpose of the present application is to overcome the defects in the prior art, provide a method and system for predicting and evaluating the quality of liquor fermentation, which can significantly reduce human interference, quickly obtain evaluation parameters using a network model, and realize reliable and effective evaluation of the fermentation quality.

[0006] The method for predicting and evaluating the quality of liquor fermentation of the present application comprises:

[0007] Collecting environmental parameters of the fermentation process of fermented grains to obtain evaluation parameters of the fermentation quality;

[0008] Constructing a fermentation quality evaluation model, inputting the environmental parameters and evaluation parameters as sample data into the fermentation quality evaluation model, training the network model, and obtaining a trained evaluation model;

[0009] Inputting the environmental parameters to be measured during the fermentation process of fermented grains into the trained evaluation model to output predicted evaluation parameters;

[0010] Using the predicted evaluation parameters to evaluate the fermentation quality to obtain an evaluation result.

[0011] Further, the environmental parameters include temperature, humidity, carbon dioxide and pH value; and the evaluation parameters include total acid, total ester and total alcohol.

[0012] Further, the fermentation quality evaluation model comprises a graph convolution network, a time convolution network and an attention mechanism.

[0013] The graph convolution network captures the complex relationship between nodes by defining a convolution operation on the graph structure in a non-Euclidean space, and extracts spatial features; the graph convolution network takes environmental parameters as nodes in the graph, and the interaction between environmental parameters is represented by edges; the feature representation of each node is updated by aggregating the information of neighboring nodes;

[0014] The time convolution network utilizes the parallel computing capability of the convolutional neural network and introduces a convolution operation in the time dimension; the time convolution network further analyzes the spatial features output by the graph convolution network in the time dimension to extract time features, and combines the spatial features and the time features to provide insights in the time sequence for the final prediction;

[0015] The attention mechanism calculates the weight of the features after the spatial features and the time features are fused, and comprehensively considers the weighted information to enhance the sensitivity of the fermentation quality evaluation model to important features.

[0016] Further, the graph convolution network takes the graph G(X, A) as input, updates the node feature information matrix X through graph convolution operation, and obtains a new graph G(X k ,A) through a nonlinear activation function, and finally outputs the graph G(X', A); wherein the node features in X and X k have different dimensions, and the hidden layer node feature dimension is d.

[0017] Further, the time convolution network comprises one-dimensional full convolution, causal convolution, dilated convolution and residual connection.

[0018] The time convolution network generates an output sequence with the same length as the input sequence through one-dimensional full convolution to preserve the information of the input data; the time convolution network sets causal convolution so that the current convolution result only depends on the present and past inputs.

[0019] The dilated convolution increases the number of point intervals of the convolution kernel to make the time convolution network have a larger receptive field, so that the input information can be extracted faster.

[0020] The residual connection is used to train a deep network to enable the network to pass information across layers; the output of the residual connection includes input information and convolution operation output information.

[0021] A white wine fermentation quality prediction and evaluation system comprises a parameter acquisition unit, a model training unit and a quality evaluation unit.

[0022] The parameter acquisition unit is configured to acquire environmental parameters of the fermented grains during the fermentation process and obtain evaluation parameters of the fermentation quality.

[0023] The model training unit is configured to construct a fermentation quality evaluation model, input the environmental parameters and the evaluation parameters as sample data into the fermentation quality evaluation model, train the network model, and obtain a trained evaluation model.

[0024] The quality evaluation unit is configured to input the to-be-detected environmental parameters during the fermentation process of the fermented grains into the trained evaluation model, output predicted evaluation parameters, and evaluate the fermentation quality by using the predicted evaluation parameters to obtain an evaluation result.

[0025] Further, the environmental parameters include temperature, humidity, carbon dioxide and pH value; and the evaluation parameters include total acid, total ester and total alcohol.

[0026] Further, the fermentation quality evaluation model includes a graph convolution network, a time convolution network and an attention mechanism.

[0027] The graph convolution network defines a convolution operation on a graph structure in a non-Euclidean space, captures the complex relationship between nodes, and extracts spatial features; the graph convolution network takes the environmental parameters as nodes in the graph, and the interaction between the environmental parameters is represented by edges; the feature representation of each node is updated by aggregating the information of the neighbor nodes.

[0028] The time convolution network utilizes the parallel computing capability of the convolutional neural network and introduces a convolution operation in the time dimension; the time convolution network further analyzes the spatial features output by the graph convolution network in the time dimension, extracts time features, and combines the spatial features and the time features, to provide insights in the time sequence for the final prediction.

[0029] The attention mechanism calculates the weight of the features after the spatial features and the time features are fused, and comprehensively considers the weighted information, to enhance the sensitivity of the fermentation quality evaluation model to important features.

[0030] Further, the graph convolution network takes the graph G(X, A) as input, updates the node feature information matrix X through the graph convolution operation, and obtains a new graph G(X k ,A) through a nonlinear activation function, and finally outputs the graph G(X', A); wherein the node features in X and X k have different dimensions, and the hidden layer node feature dimension is d.

[0031] Further, the time convolution network includes one-dimensional full convolution, causal convolution, dilated convolution and residual connection.

[0032] The time convolution network generates an output sequence with the same length as the input sequence through one-dimensional full convolution to retain the information of the input data; the time convolution network sets causal convolution so that the current convolution result only depends on the present and past input;

[0033] The dilated convolution makes the time convolution network obtain a larger receptive field by increasing the interval number of the points of the convolution kernel, so that the input information is extracted faster;

[0034] The residual connection is used for training a deep network to make the network transmit information across layers; the output of the residual connection includes input information and output information of convolution operation.

[0035] The method and system for predicting and evaluating the fermentation quality of liquor disclosed by the application have the following beneficial effects: an improved GCN-TCN-Attention model is constructed by combining a graph convolution network (GCN), a time convolution network (TCN) and an attention mechanism, so that the model can more comprehensively understand and predict the complexity of spatio-temporal data, and capture complex spatial interaction and temporal dynamic change in the fermentation process; the soft attention mechanism is introduced, so that the improved model can more accurately identify and emphasize the features that have a greater impact on the prediction result, thereby improving the prediction accuracy and robustness of the model. The application significantly reduces human interference, and the improved model can quickly obtain evaluation parameters, thereby realizing reliable and effective evaluation of the fermentation quality. BRIEF DESCRIPTION OF DRAWINGS

[0036] The application will be further described below in combination with the drawings and embodiments:

[0037] Figure 1 The figure is a schematic diagram of the graph convolution network structure of the application;

[0038] Figure 2 The figure is a schematic diagram of the time convolution network structure of the application. DETAILED DESCRIPTION

[0039] The application will be further described below in combination with the drawings and embodiments:

[0040] The embodiment discloses a method for predicting and evaluating the fermentation quality of liquor, which includes the following steps:

[0041] Collecting environmental parameters of the fermentation process of the fermented grains to obtain evaluation parameters of the fermentation quality;

[0042] Constructing a fermentation quality evaluation model, inputting the environmental parameters and the evaluation parameters as sample data into the fermentation quality evaluation model, training the network model, and obtaining a trained evaluation model;

[0043] The environmental parameters to be detected in the fermentation process of the fermented grains are input into the trained evaluation model, and predicted evaluation parameters are output.

[0044] The fermentation quality is evaluated by using the predicted evaluation parameters, and an evaluation result is obtained.

[0045] In this embodiment, the temperature, humidity, carbon dioxide, pH value and other parameters of the fermentation process of the fermented grains are important characteristic indexes in the fermentation process of the fermented grains and are also environmental parameters in the fermentation process of the fermented grains; the environmental parameters include temperature, humidity, carbon dioxide and pH value.

[0046] The most significant embodiment of the fermentation process quality of the fermented grains is the quality of the base liquor, and the liquor aroma is mainly derived from the types and contents of alcohols, esters, acids, aldehydes and phenols produced by biochemical reactions in the fermentation process. Among them, acids affect the taste balance and flavor characteristics of the liquor, have the effect of antibacterial and quality preservation, and mainly represent organic acids such as caproic acid, acetic acid, lactic acid and butyric acid, which provide a material basis for the flavor and ester synthesis of the liquor. Ester is the main component of liquor aroma and has a great influence on flavor, and the main representatives are ethyl acetate, ethyl butyrate, ethyl caproate and ethyl lactate, which are the four major esters in liquor. Alcohol is the most influential, contributing to the overall aroma of the liquor, and the main representatives are ethanol and isopropyl alcohol. Therefore, the main components and contents of total acid, total ester and total alcohol in the fermented grains can be used as important indexes for evaluating the fermentation quality of the fermented grains; the evaluation parameters include total acid, total ester and total alcohol.

[0047] In this embodiment, an improved GCN-TCN-Attention model combining graph convolution network (GCN), time convolution network (TCN) and attention mechanism is constructed as a fermentation quality evaluation model.

[0048] Graph Convolutional Network (GCN) is a deep learning method based on graph structure data. Its core idea is to learn a mapping function to enable nodes in the graph to aggregate their own and neighbor node features to generate new node representations. Traditional convolutional neural network (CNN) is good at processing spatial features in Euclidean space data, while GCN defines convolution operations on graph structures in non-Euclidean space to capture complex relationships between nodes and extract spatial features.

[0049] In the fermentation process, temperature, humidity, pH value and carbon dioxide concentration can be regarded as nodes in the graph, and their interactions are represented by edges. The GCN model updates the feature representation of each node by aggregating the information of the neighbor nodes, thereby learning the role and influence of each parameter in the fermentation process.

[0050] The GCN structure is as follows:Figure 1 As shown, GCN takes a graph G(X,A) as input, performs graph convolution to update the node feature information matrix X, and then passes it through a nonlinear activation function to obtain a new graph G(X). k The final output is graph G(X',A). X and X' are the graphs in the graph G(X',A). k The node features in the middle layer have different dimensions, and the hidden layer node features have a dimension of d.

[0051] The node feature matrix of the (k+1)th layer can be obtained by the following formula:

[0052] ;

[0053] In the formula I is the identity matrix; D is the degree matrix of A. W k X is the weight matrix of the k-th layer; k Let k be the feature matrix of the nodes in the k-th layer; It is a non-linear activation function.

[0054] Temporal Convolutional Networks (TCNs) are deep learning frameworks for sequence modeling and prediction. Compared to traditional Recurrent Neural Networks (RNNs), TCNs leverage the parallel computing power of Convolutional Neural Networks (CNNs) and, by introducing convolutional operations in the temporal dimension, exhibit superior performance in handling long-term dependencies. TCNs were chosen for time series data processing primarily due to their efficiency and flexibility in capturing temporal dependencies and dynamic changes.

[0055] Data during fermentation is typically collected chronologically. TCNs, through the use of causal convolutional and dilated convolutional layers, can flexibly adjust the scope of historical information perception while maintaining computational efficiency. The structure of TCNs allows the model to consider not only parameter values ​​at the current time point but also trends and patterns over a past period, thus providing more accurate predictions. In practical applications, TCNs further analyze the spatial features of the GCN output in the temporal dimension, extracting temporal features and combining them with spatial features to provide time-series insights for the final prediction.

[0056] The architecture of TCN includes one-dimensional fully convolutional layers, causal convolutional layers, dilated convolutional layers, and residual connections. It boasts advantages such as parallel computing capabilities, flexible receptive fields, stable gradients, and low memory footprint. TCN is primarily composed of multiple stacked residual modules, each containing two layers of dilated causal convolutional layers to increase the direct mapping of the input and accelerate computation. The structure of TCN is as follows: Figure 2 As shown.

[0057] where TCN generates an output sequence with the same length as the input sequence by one-dimensional full convolution to preserve the information of the input data;

[0058] Causal convolution makes the convolution have a causal relationship, and the current convolution result only depends on the present and past inputs, thereby improving the prediction accuracy. For the data y t , the input value can only be at time t and before time t, i.e., x0…x t , which is not affected by x t+1 …x t .

[0059] Dilated convolution is to increase the number of intervals of the points of the convolution kernel, so that TCN can obtain a larger receptive field, and the input information can be extracted faster. Assuming that the one-dimensional input sequence is x, the one-dimensional convolution kernel f:{0,1,…,k-1}→R, and the output value of the dilated sequence of the s-th element is:

[0060] ;

[0061] In the formula: is a dilated operation; is a dilated coefficient; is the size of the convolution kernel; is the sequence element corresponding to the element in the convolution kernel. With the rise of the network layer, the receptive field is expanded by increasing the size of or .

[0062] Residual connection is an effective method for training deep networks, which can pass information across layers. Assuming that the input of the residual module is x and the output is y:

[0063] ;

[0064] In the formula: H is a series of convolution function operations; P is a weight function of the convolution value.

[0065] As can be seen from the formula, the output of the residual connection contains the input information and the output information of the convolution operation, which can effectively ensure the accuracy of the TCN network.

[0066] That is, the design of the TCN module combines residual connection and dilated convolution. Dilated convolution expands the receptive field of the model by increasing the interval of the convolution kernel, which can more effectively extract long-time dependent features; and the residual connection ensures the stability of the gradient in the deep network, avoiding the problem of gradient disappearance, so that the model can still maintain stable training effect while increasing complexity, further improving the performance of the model in processing long sequence data.

[0067] Attention Mechanism is an advanced computational method that effectively improves the prediction accuracy of a model by prioritizing features with higher priority in situations where data is limited or information overload. This mechanism is particularly suitable for fermentation process prediction, as parameters at different time periods or under different environmental conditions have different effects on the final product.

[0068] Attention Mechanism assigns the model the ability to identify and emphasize factors that significantly affect the prediction target by self-learning the weights of input features. This process not only enhances the model's sensitivity to key information, but also significantly improves the accuracy and robustness of the prediction. Attention Mechanism is divided into Hard Attention and Soft Attention. Hard Attention mechanism only focuses on one input information at any time, while Soft Attention mechanism calculates the weights for all input information and considers these weighted information comprehensively, inputting them into the neural network for training. In this invention, Soft Attention is used because it can capture the importance differences of input information more meticulously. The relevant formula is as follows:

[0069] ; ;

[0070] where, is the score of the feature vector, and the larger the score, the greater the attention, is the weight value of the th input feature in the attention mechanism, i.e., the ratio of the score of the feature vector to the total, the final vector obtained by summing and averaging all vectors.

[0071] The invention innovatively combines Graph Convolutional Network (GCN) and Temporal Convolutional Network (TCN), enabling the model to capture complex spatial interaction and temporal dynamic changes in the fermentation process. GCN is responsible for extracting spatial interaction features between parameters in the fermentation process, effectively handling complex relationships between parameters such as temperature, humidity, pH, and CO2. TCN accurately captures dynamic changes in the time dimension through causal convolution and dilated convolution, extracting time features. This combination enables the model to more comprehensively understand and predict the complexity of spatio-temporal data. Through the Soft Attention mechanism, the model can more accurately identify and emphasize features that have a greater impact on the prediction result, thereby improving the prediction accuracy and robustness of the model. This multi-level Attention application enables the model to more effectively capture key information when dealing with high-dimensional complex relationship data.

[0072] In this embodiment, the corresponding temperature, humidity, carbon dioxide, pH value and other environmental parameters can be obtained from the historical data records of the fermentation of the fermented grains for several times; correspondingly, the evaluation parameters such as total acid, total ester and total alcohol can be obtained from the historical data records of the fermentation of the fermented grains for several times; the obtained environmental parameters and evaluation parameters are used as sample data for model training, that is, the environmental parameters are used as the input of the GCN-TCN-Attention improved model, and the evaluation parameters are used as the output of the GCN-TCN-Attention improved model, and the improved model is trained. Among them, the GCN is used to extract spatial features, and the complex relationship between the temperature, humidity, PH, CO2 and other parameters in the fermentation process is processed; the TCN extracts time features in the time dimension; the Attention mechanism is applied in multiple levels in the model to enhance the capture ability of key information; the Adam algorithm is used as the optimizer of the model, the cycle training times are 1500, the initial learning rate is 0.01, the gradient threshold is 1, and the learning rate is adjusted after 750 training times, and the learning rate reduction factor is 0.4.

[0073] Through the above training and processing of the GCN-TCN-Attention improved model, a trained evaluation model is obtained; the environmental parameters to be measured in the actual fermented grain fermentation process are input into the trained evaluation model, and the predicted evaluation parameters are output; according to the existing evaluation standard of the distillery, the fermentation quality is evaluated by using the predicted evaluation parameters, and the evaluation result can be obtained. Among them, the distillery can set the standard value range of the predicted evaluation parameters such as total acid, total ester and total alcohol according to its own production process and product quality requirements. By comparing the predicted evaluation parameters with the standard values, it can be judged whether the fermentation quality meets the standard, so as to obtain the final evaluation result.

[0074] The present application also relates to a prediction and evaluation system for the fermentation quality of liquor, which corresponds to the evaluation method of the above embodiment and can be understood as a system for realizing the above evaluation method. The evaluation system comprises a parameter acquisition unit, a model training unit and a quality evaluation unit.

[0075] The parameter acquisition unit is used to acquire the environmental parameters of the fermented grain fermentation process and obtain the evaluation parameters of the fermentation quality.

[0076] The model training unit is used to construct a fermentation quality evaluation model, input the environmental parameters and evaluation parameters as sample data into the fermentation quality evaluation model, train the network model, and obtain a trained evaluation model.

[0077] The quality evaluation unit is used to input the environmental parameters to be measured in the fermented grain fermentation process into the trained evaluation model, output the predicted evaluation parameters, and evaluate the fermentation quality by using the predicted evaluation parameters to obtain the evaluation result.

[0078] In order to better understand the GCN-TCN-Attention improved model used in the fermentation quality prediction evaluation method and system of the present application, the following is further described:

[0079] The data of temperature, humidity, carbon dioxide and pH value in the fermentation process of certain fermented grains were collected as environmental parameters, which were input into the BP model, the LSTM model and the GCN-TCN-Attention improved model respectively, to obtain the predicted output of total acid, total ester and total alcohol, and the model regression prediction results shown in Table 1 were obtained.

[0080] Table 1

[0081]

[0082] From the above Table 1, it can be seen that the fitting degree of the GCN-TCN-Attention improved model of the present application in the prediction performance of acid, ester and alcohol is basically higher than that of the BP neural network regression model and the LSTM long short-term memory neural network regression model, which indicates that the regression accuracy of the GCN-TCN-Attention improved model is better than that of the BP neural network regression model and the LSTM long short-term memory neural network regression model. Since the average absolute error (MAE) and the root mean square error (RMSE) of the GCN-TCN-Attention improved model are basically smaller than those of the BP neural network regression model and the LSTM long short-term memory neural network regression model, it indicates that the stability of the GCN-TCN-Attention improved model is better than that of the BP neural network regression model and the LSTM long short-term memory neural network regression model.

[0083] Therefore, from the comprehensive data, the GCN-TCN-Attention improved model of the present application has obvious advantages.

[0084] Finally, it is pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the purpose and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for predicting and evaluating the quality of liquor fermentation, characterized in that: The method comprises the following steps: Collecting environmental parameters of the fermentation process of the fermented grains to obtain evaluation parameters of the fermentation quality; the environmental parameters include temperature, humidity, carbon dioxide, and pH value; the evaluation parameters include total acid, total ester, and total alcohol; Constructing a fermentation quality evaluation model, inputting the environmental parameters and the evaluation parameters as sample data into the fermentation quality evaluation model, training the network model, and obtaining the trained evaluation model; the fermentation quality evaluation model includes a graph convolution network, a time convolution network, and an attention mechanism; The graph convolution network defines a convolution operation on a graph structure in a non-Euclidean space, captures the complex relationship between nodes, and extracts spatial features; the graph convolution network takes the environmental parameters as nodes in the graph, and the interaction between the environmental parameters is represented by edges; the feature representation of each node is updated by aggregating the information of the neighbor nodes; The time convolution network utilizes the parallel computing capability of the convolutional neural network and introduces a convolution operation in the time dimension; the time convolution network further analyzes the spatial features output by the graph convolution network in the time dimension, extracts time features, and combines them with the spatial features to provide insights into the time series for the final prediction; The attention mechanism calculates the weight of the features after the fusion of the spatial features and the time features, and comprehensively considers the weighted information to enhance the sensitivity of the fermentation quality evaluation model to important features; Inputting the to-be-tested environmental parameters in the fermentation process of the fermented grains into the trained evaluation model to output the predicted evaluation parameters; Using the predicted evaluation parameters to evaluate the fermentation quality to obtain the evaluation result.

2. The method of claim 1, wherein the method is characterized by: The graph convolutional network takes a graph G(X,A) as input, performs graph convolution operations to update the node feature information matrix X, and then passes it through a nonlinear activation function to obtain a new graph G(X). k The final output is graph G(X',A); where X and X' are variables. k The node features in the middle layer have different dimensions, and the hidden layer node features have a dimension of d.

3. The method of claim 1, wherein the method is characterized by: The time convolution network includes one-dimensional full convolution, causal convolution, dilated convolution, and residual connection; The time convolution network generates an output sequence with the same length as the input sequence through one-dimensional full convolution to retain the information of the input data; the time convolution network sets causal convolution so that the current convolution result only depends on the present and past inputs; The dilated convolution increases the number of point intervals of the convolution kernel to make the time convolution network have a larger receptive field and extract input information faster; The residual connection is used to train a deep network to transfer information across layers; the output of the residual connection includes the input information and the output information of the convolution operation.

4. A system for predicting and evaluating the quality of liquor fermentation, characterized by: The method comprises the following steps: The parameter acquisition unit is configured to collect environmental parameters of the fermentation process of the fermented grains to obtain evaluation parameters of the fermentation quality; the environmental parameters include temperature, humidity, carbon dioxide, and pH value; the evaluation parameters include total acid, total ester, and total alcohol; The model training unit is configured to construct a fermentation quality evaluation model, input the environmental parameters and the evaluation parameters as sample data into the fermentation quality evaluation model, train the network model, and obtain the trained evaluation model; the fermentation quality evaluation model includes a graph convolution network, a time convolution network, and an attention mechanism; The graph convolution network defines a convolution operation on a graph structure in a non-Euclidean space, captures the complex relationship between nodes, and extracts spatial features; the graph convolution network takes the environmental parameters as nodes in the graph, and the interaction between the environmental parameters is represented by edges; the feature representation of each node is updated by aggregating the information of the neighbor nodes; The graph convolution network captures the complex relationship between nodes and extracts spatial features by defining a convolution operation on the graph structure in a non-Euclidean space; the graph convolution network takes environmental parameters as nodes in the graph, and the interaction between environmental parameters is represented by edges; the feature representation of each node is updated by aggregating the information of neighboring nodes; The time convolution network utilizes the parallel computing capability of the convolutional neural network and introduces a convolution operation in the time dimension; the time convolution network further analyzes the spatial features output by the graph convolution network in the time dimension to extract time features, and combines the spatial features and the time features to provide insights into the time sequence for the final prediction; The attention mechanism calculates the weight of the features after the fusion of the spatial features and the time features, and comprehensively considers the weighted information to enhance the sensitivity of the fermentation quality evaluation model to important features; The quality evaluation unit is configured to input the environmental parameters to be measured in the fermentation process of the fermented grains into the trained evaluation model and output predicted evaluation parameters; and evaluate the fermentation quality by using the predicted evaluation parameters to obtain an evaluation result.

5. The system for predicting and evaluating the quality of white wine fermentation according to claim 4, characterized in that: The graph convolution network takes a graph G(X, A) as input, updates a node feature information matrix X through a graph convolution operation, obtains a new graph G(X k ,A) through a nonlinear activation function, and finally outputs a graph G(X k The node features in X and X have different dimensions, and the hidden layer node feature dimension is d.

6. The system for predicting and evaluating the quality of white wine fermentation according to claim 4, characterized in that: The time convolution network includes one-dimensional full convolution, causal convolution, dilated convolution, and residual connection; The time convolution network generates an output sequence with the same length as the input sequence by one-dimensional full convolution to retain the information of the input data; the time convolution network sets causal convolution so that the current convolution result only depends on the present and past inputs; The dilated convolution increases the number of point intervals of the convolution kernel to make the time convolution network have a larger receptive field and extract input information faster; The residual connection is used to train a deep network to enable the network to pass information across layers; the output of the residual connection includes input information and output information of the convolution operation.

Images