Coffee secondary fermentation process optimization method, medium and system

Abstract

The invention provides a coffee secondary fermentation technological process optimization method, medium and system, and belongs to the technical field of coffee secondary ferment.The coffee secondary fermentation technological process optimization method comprises the steps that firstly, whole-process data of small-scale experimental production is obtained, and based on the data, a time sequence model of all fermentation parameters and a dynamic model of state parameters are established; correlation mapping is established with quality parameters, and a multiple-input-multiple-output embedded deep neural network prediction model is constructed. In order to enhance the generalization ability of the model, data augmentation and knowledge distillation technologies are utilized to further optimize and obtain a second prediction model. And then, on the basis of the second prediction model, constructing a reverse neural network model, and realizing reverse mapping from the target quality to the initial fermentation parameters. And finally, inputting ideal quality parameters by using a reverse model, and performing inversion to obtain optimal secondary fermentation initial parameters. Meanwhile, the initial parameters are input into a prediction model, the fermentation parameter change trend at each moment is obtained, and a basis is provided for process regulation and control.

Description

Technical Field

[0001] The present invention belongs to the technical field of secondary fermentation of coffee, and in particular, relates to a method, medium and system for optimizing a process of secondary fermentation of coffee. Background Art

[0002] The traditional coffee manufacturing process includes roasting, crushing, extraction and other links, each of which has an important impact on the final coffee quality. Among them, secondary fermentation is one of the key steps in coffee manufacturing. During the secondary fermentation process, some microorganisms such as yeast and lactic acid bacteria will transform the organic matter in the coffee beans, thereby producing metabolites that affect quality factors such as flavor, acidity, and sweetness. Therefore, how to accurately control the fermentation process parameters to achieve ideal coffee quality has always been a key issue of concern in the industry.

[0003] At present, the secondary fermentation process control method adopted by the industry is mainly based on the experience-based parameter setting method, that is, a set of initial fermentation parameters are summarized through years of production practice, and empirical adjustments are made on this basis. Although it is simple and easy, the parameter adjustment tends to be too empirical and it is difficult to ensure the stability of product quality. Summary of the invention

[0004] In view of this, the present invention provides a method, medium and system for optimizing the secondary fermentation process of coffee, which can solve the problem that the parameter adjustment in the prior art is prone to be too empirical and difficult to ensure the stability of product quality.

[0005] The present invention is achieved in that:

[0006] A first aspect of the present invention provides a method for optimizing a coffee secondary fermentation process, comprising the following steps:

[0007] S10, obtaining the whole process data of the small-scale experimental production of secondary fermentation of coffee, including the initial fermentation parameters, the fermentation parameters and fermentation state parameters at each collection moment, and the coffee quality parameters after the secondary fermentation production is completed;

[0008] S20, establishing a time series model of each fermentation parameter and a dynamic change model of each fermentation state parameter according to the whole process data;

[0009] S30, using Bayesian network and time series analysis to establish a time series model for each fermentation parameter and a correlation mapping between a dynamic change model for each fermentation state parameter and each coffee quality parameter;

[0010] S40, repeating steps S10-S30 to obtain multiple sets of full-process data and their corresponding association maps, and training a multi-input multi-output embedded deep neural network model according to the multiple sets of association maps to obtain a first prediction model for predicting coffee quality parameters under different combinations of fermentation parameters and state parameters; the first prediction model is embedded with a fermentation parameter variation mathematical model;

[0011] S50, using data augmentation technology to expand the multiple sets of full-process data to generate more diverse virtual samples to enhance the generalization ability of the model; using knowledge distillation technology to use the first prediction model as a teacher model to train a smaller and more efficient student model;

[0012] S60, transferring the knowledge of the teacher model to the student model to obtain a second prediction model;

[0013] S70, based on the second prediction model, constructing a reverse neural network, using a back propagation algorithm and a gradient descent method, training the network to achieve mapping from the target coffee quality parameters to the initial fermentation parameters, and obtaining a third model;

[0014] S80, using the preset ideal coffee quality parameters as input, performing reverse calculation through the third model, inversely calculating and outputting the optimal initial fermentation parameters for the secondary fermentation of the coffee;

[0015] S90, inputting the optimal initial fermentation parameters of the coffee secondary fermentation into the fermentation parameter variation mathematical model in the first prediction model, obtaining the fermentation parameters at multiple collection moments and outputting them as a reference for adjusting the fermentation parameters in the coffee secondary fermentation process.

[0016] On the basis of the above technical solution, the coffee secondary fermentation process optimization method of the present invention can also be improved as follows:

[0017] Among them, the fermentation parameters include adjustable parameters and non-adjustable parameters. The adjustable parameters are specifically temperature, stirring speed, dissolved oxygen content, and fermentation tank pressure. The non-adjustable parameters are specifically pH value, specific gravity of fermentation liquid, conductivity, redox potential, carbon dioxide concentration, ammonia nitrogen concentration, total nitrogen content, free amino nitrogen content, water activity, and fermentation liquid color.

[0018] Furthermore, the structure of the multi-input and multi-output embedded deep neural network model specifically includes a mathematical model of fermentation parameter changes, a branch network of adjustable parameters affecting quality, a branch network of non-adjustable parameters affecting quality, a branch network of long-term parameters affecting quality, a branch network of short-term parameters affecting quality, and a fusion network.

[0019] Furthermore, the fermentation parameter variation mathematical model is used to predict the dynamic variation of each parameter during the fermentation process, the input is the initial fermentation parameter and time, and the output is the predicted value of the fermentation parameter at each time point.

[0020] Furthermore, the adjustable parameter impact on quality branch network is used to simulate the impact of adjustable parameters on coffee quality, the input is the time series data of the adjustable parameters, the output is the degree of impact on each quality parameter, and the structure is a combination of a multi-layer convolutional neural network and a long short-term memory network.

[0021] Furthermore, the non-adjustable parameter impact on quality branch network is used to simulate the impact of non-adjustable parameters on coffee quality, the input is the time series data of non-adjustable parameters, the output is the degree of impact on each quality parameter, and the structure is a combination of a multi-layer perceptron and an attention mechanism.

[0022] Furthermore, the long-term parameter impact branch network is used to capture the impact of the long-term change trend of fermentation parameters on quality. The input is the long-term span data of all fermentation parameters, the output is the long-term impact factor, and the structure is a time series analysis network based on Transformer.

[0023] Furthermore, the short-term parameter impact branch network is used to capture the impact of short-term fluctuations in fermentation parameters on quality, the input is short-time window data of all fermentation parameters, the output is short-term impact factors, and the structure is a recurrent neural network based on gated recurrent units;

[0024] The fusion network is used to integrate the outputs of each branch network and generate the final quality prediction result. The input is the output of each branch network, and the output is the predicted coffee quality parameter. The structure is a combination of a multi-layer fully connected network and a softmax layer.

[0025] Specifically, the step S10 specifically includes: obtaining the whole process data in the small-scale experimental production of secondary fermentation of coffee, including the initial fermentation parameters, the fermentation parameters at each collection time and the fermentation state parameters, and the coffee quality parameters after the secondary fermentation production is completed. Among them, the collection time includes the sequence time collection time and the key stage collection time. The sequence time collection time is to collect data every 2 hours, and the key stage collection time is to collect data at key moments such as when the fermentation starts, when the pH value drops to the lowest point for the first time, when the ethanol content reaches the peak, and when the fermentation ends. The collected fermentation parameters include temperature, pH value, dissolved oxygen content, stirring speed, fermentation tank pressure, fermentation liquid specific gravity, conductivity, redox potential, carbon dioxide concentration, ammonia nitrogen concentration, total nitrogen content, free amino nitrogen content, water activity, fermentation liquid color, etc., fermentation state parameters include ethanol content, organic acid content, sugar content, microbial quantity, CO2 generation rate, etc., and coffee quality parameters include flavor complexity, acidity, sweetness, bitterness, taste, aroma intensity, etc., which are obtained by combining sensory evaluation and instrumental analysis.

[0026] Specifically, the step S20 specifically includes: according to the whole process data, respectively establishing a time series model of each fermentation parameter and a dynamic change model of each fermentation state parameter. For the time series model of the fermentation parameter, a time series analysis method such as an autoregressive moving average (ARIMA) model is used for modeling; for the dynamic change model of the fermentation state parameter, a mathematical modeling method such as a differential equation or a difference equation is used to describe its change trend over time. The establishment of these models helps to more accurately describe the dynamic characteristics of each parameter in the fermentation process.

[0027] Specifically, the step S30 specifically includes: using the Bayesian network and time series analysis method to establish a time series model of each fermentation parameter and a dynamic change model of each fermentation state parameter and an association mapping between each coffee quality parameter. The Bayesian network can effectively capture the conditional dependency between variables, while the time series analysis can reflect the change law of the parameters over time. Through this modeling method, the quantitative relationship between the fermentation process parameters and the final coffee quality can be obtained.

[0028] Specifically, step S40 specifically includes: repeating steps S10-S30 to obtain multiple sets of full process data and their corresponding association maps. According to these association maps, a multi-input multi-output embedded deep neural network model is trained to obtain a first prediction model. This first prediction model can be used to predict coffee quality parameters under different combinations of fermentation parameters and state parameters, and a mathematical model of fermentation parameter changes is also embedded in its structure.

[0029] Specifically, step S50 specifically includes: first, using data augmentation technology to expand the multiple sets of full-process data obtained in step S40 to generate more diverse virtual samples; then using knowledge distillation technology, using the first prediction model as a teacher model to train a smaller and more efficient student model. Knowledge distillation can effectively transfer the knowledge of the teacher model to the student model and improve the prediction performance of the student model.

[0030] Specifically, step S60 specifically includes: using knowledge distillation technology, transferring the knowledge of the teacher model in step S50 to the student model to obtain a second prediction model. This second prediction model has a smaller model complexity and higher computational efficiency while maintaining a high prediction accuracy, and can be more conveniently deployed in an actual coffee production system.

[0031] Specifically, step S70 specifically includes: constructing a reverse neural network based on the second prediction model. The reverse network is trained using a back propagation algorithm and a gradient descent method, with the goal of achieving a mapping relationship from target coffee quality parameters to initial fermentation parameters. By training this reverse network, a third model can be obtained for predicting the optimal initial fermentation parameters under given target quality parameters.

[0032] Specifically, the step S80 specifically includes: using the preset ideal coffee quality parameters as input, performing reverse calculation through the third model trained in step S70, and inverting to obtain the optimal initial fermentation parameters of the coffee secondary fermentation. The step S90 specifically includes: inputting the optimal initial fermentation parameters obtained in step S80 into the fermentation parameter change mathematical model of the first prediction model in step S40 to predict the fermentation parameter values ​​at each collection time under these optimal parameters, and provide a reference for parameter adjustment in the production process.

[0033] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned method for optimizing the secondary fermentation process of coffee.

[0034] A third aspect of the present invention provides a coffee secondary fermentation process optimization system, which includes the above-mentioned computer-readable storage medium.

[0035] Compared with the prior art, the coffee secondary fermentation process optimization method, medium and system provided by the present invention have the following beneficial effects:

[0036] 1) Make full use of the comprehensive data from the small-scale experiment of secondary fermentation of coffee. Not only the initial fermentation parameters, fermentation parameters and state parameters at each time point were collected, but also various quality indicators of the final product were measured. These comprehensive experimental data laid a solid foundation for subsequent analysis and modeling.

[0037] 2) A quantitative correlation model between fermentation parameters, fermentation status and coffee quality was established. Through Bayesian network, time series analysis and other technical means, the dynamic mapping relationship between various parameters was accurately captured, providing a basis for process optimization.

[0038] 3) Using machine learning techniques such as deep neural networks and combined with knowledge distillation methods, a high-performance quality prediction model was trained. This model not only has high prediction accuracy, but also has low model complexity and can be easily deployed in actual production systems.

[0039] 4) Based on the prediction model, a reverse neural network was constructed to achieve the reverse mapping from the target quality parameters to the optimal initial fermentation parameters. This provides strong support for the precise control of fermentation process parameters and the continuous improvement of product quality.

[0040] 5) The optimized initial fermentation parameters are input into the dynamic change model of fermentation parameters to predict the changes in fermentation parameters at each collection time, providing data support for real-time process adjustment.

[0041] In general, the data-driven coffee secondary fermentation process optimization method proposed in the present invention fully integrates a number of advanced technologies such as experimental data analysis, machine learning, and dynamic modeling, and has the following advantages over existing methods: 1) It can more comprehensively and systematically characterize the dynamic characteristics of the fermentation process; 2) It can achieve reverse design from target quality to optimal initial parameters; 3) It uses intelligent optimization algorithms to continuously improve the stability and controllability of product quality, solving the problem that the parameter adjustment of the existing technology is prone to being too empirical and difficult to ensure the stability of product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flow chart of the method provided by the present invention;

[0043] Attached Figure 2 This is a comparison chart before and after the optimization of coffee quality parameters;

[0044] Attached Figure 3 This is a comparison chart before and after optimization of key fermentation parameters;

[0045] Attached Figure 4 It is the trend chart of key fermentation parameters changing over time;

[0046] Attached Figure 5 It is a radar chart of coffee quality parameters;

[0047] Attached Figure 6 The bar chart for comparing fermentation parameters before and after fermentation;

[0048] Attached Figure 7 This is a diagram showing the changes in key parameters during the fermentation process. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0050] like Figure 1 As shown, it is a flow chart of a coffee secondary fermentation process optimization method provided by the present invention. The method comprises the following steps:

[0051] S10, obtaining the whole process data of the small-scale experimental production of secondary fermentation of coffee, including the initial fermentation parameters, the fermentation parameters and fermentation state parameters at each collection moment, and the coffee quality parameters after the secondary fermentation production is completed;

[0052] S20. Establishing a time series model of each fermentation parameter and a dynamic change model of each fermentation state parameter based on the whole process data;

[0053] S30, using Bayesian network and time series analysis to establish a time series model for each fermentation parameter and a correlation mapping between a dynamic change model for each fermentation state parameter and each coffee quality parameter;

[0054] S40, repeating steps S10-S30 to obtain multiple sets of full-process data and their corresponding association maps, and training a multi-input multi-output embedded deep neural network model according to the multiple sets of association maps to obtain a first prediction model for predicting coffee quality parameters under different combinations of fermentation parameters and state parameters; the first prediction model is embedded with a fermentation parameter variation mathematical model;

[0055] S50. Use data augmentation technology to expand multiple sets of full-process data and generate more diverse virtual samples to enhance the generalization ability of the model. Use knowledge distillation technology to use the first prediction model as a teacher model to train a smaller and more efficient student model.

[0056] S60, transferring the knowledge of the teacher model to the student model to obtain a second prediction model;

[0057] S70, based on the second prediction model, constructing a reverse neural network, using a back propagation algorithm and a gradient descent method, training the network to achieve mapping from the target coffee quality parameters to the initial fermentation parameters, and obtaining a third model;

[0058] S80, using the preset ideal coffee quality parameters as input, performing reverse calculation through the third model, inversely calculating and outputting the optimal initial fermentation parameters for the secondary fermentation of the coffee;

[0059] S90, inputting the optimal initial fermentation parameters of the coffee secondary fermentation into the fermentation parameter variation mathematical model in the first prediction model, obtaining the fermentation parameters at multiple collection moments and outputting them as a reference for adjusting the fermentation parameters in the coffee secondary fermentation process.

[0060] The specific implementation methods of the above steps are described in detail below:

[0061] The specific implementation of step S10 is: first, in the small-scale experimental production of secondary fermentation of coffee, systematically collect data samples of the whole process. These data include initial fermentation parameters, fermentation parameters at each collection time, fermentation state parameters, and quality parameters of the final coffee product. There are two types of collection time for these parameters: one is serial time collection, that is, data is collected every 2 hours; the other is key stage collection, that is, collection at key moments such as the beginning of fermentation, when the pH value first drops to the lowest point, when the ethanol content reaches the peak, and when the fermentation ends. The collected fermentation parameters include temperature, pH value, dissolved oxygen content, stirring speed, fermentation tank pressure, fermentation liquid specific gravity, conductivity, redox potential, carbon dioxide concentration, ammonia nitrogen concentration, total nitrogen content, free amino nitrogen content, water activity, fermentation liquid color, etc. Fermentation state parameters include ethanol content, organic acid content, sugar content, microbial count, CO2 generation rate, etc. Coffee quality parameters include flavor complexity, acidity, sweetness, bitterness, taste, aroma intensity, etc., which are obtained by combining sensory evaluation and instrumental analysis. These comprehensive data collections lay the foundation for subsequent analysis and modeling.

[0062] The specific implementation of step S20 is: according to the whole process data obtained in step S10, a time series model of each fermentation parameter and a dynamic change model of each fermentation state parameter are established respectively. For the time series model of the fermentation parameters, a time series analysis method such as an autoregressive moving average (ARIMA) model can be used for modeling. For the dynamic change model of the fermentation state parameters, a mathematical modeling method such as a differential equation or a difference equation can be used to describe its change trend over time. The establishment of these models helps to more accurately describe the dynamic characteristics of each parameter in the fermentation process.

[0063] The specific implementation of step S30 is: using the Bayesian network and time series analysis method, establish the time series model of each fermentation parameter and the dynamic change model of each fermentation state parameter and the association mapping between each coffee quality parameter. The Bayesian network can effectively capture the conditional dependency between variables, while the time series analysis can reflect the change law of parameters over time. Through this modeling method, the quantitative relationship between the fermentation process parameters and the final coffee quality can be obtained, providing a basis for the next step of optimization.

[0064] The specific implementation of step S40 is: repeating steps S10-S30 to obtain multiple sets of full process data and their corresponding association maps. According to these association maps, a multi-input multi-output embedded deep neural network model is trained to obtain a first prediction model. This first prediction model can be used to predict coffee quality parameters under different combinations of fermentation parameters and state parameters. The structure of the model also embeds a mathematical model of fermentation parameter changes, which can better capture the dynamic change law of parameters during the fermentation process.

[0065] The specific implementation of step S50 is as follows: first, the multiple sets of full-process data obtained in step S40 are expanded using data augmentation technology to generate more diverse virtual samples. This can enhance the generalization ability of the model. Then, the knowledge distillation technology is used to train a smaller and more efficient student model using the first prediction model as the teacher model. Knowledge distillation can effectively transfer the knowledge of the teacher model to the student model and improve the prediction performance of the student model.

[0066] The specific implementation of step S60 is: by using the knowledge distillation technology, the knowledge of the teacher model in step S50 is transferred to the student model to obtain a second prediction model. This second prediction model has a smaller model complexity and higher computational efficiency while maintaining a high prediction accuracy, and can be more conveniently deployed in an actual coffee production system.

[0067] The specific implementation of step S70 is: based on the second prediction model, a reverse neural network is constructed. The reverse network is trained using a back propagation algorithm and a gradient descent method, and the goal is to achieve a mapping relationship from target coffee quality parameters to initial fermentation parameters. By training this reverse network, a third model can be obtained to predict the optimal initial fermentation parameters under given target quality parameters.

[0068] The specific implementation of step S80 is: using the preset ideal coffee quality parameters as input, reverse calculation is performed through the third model trained in step S70, and the optimal initial fermentation parameters of the secondary fermentation of coffee are obtained by inversion. These optimal initial fermentation parameters can be used as a reference for the actual production process and provide a basis for process adjustment.

[0069] The specific implementation of step S90 is: input the optimal coffee secondary fermentation initial parameters obtained in step S80 into the fermentation parameter change mathematical model of the first prediction model in step S40. The model can predict the fermentation parameter values ​​at each collection time under these optimal initial parameters. These predicted fermentation parameter values ​​can provide a reference for parameter adjustment in the actual production process, which helps to achieve more refined process control.

[0070] In summary, this coffee secondary fermentation process optimization method makes full use of experimental data and establishes a correlation model between fermentation parameters, fermentation status and coffee quality. Through training deep neural networks, knowledge distillation, reverse calculation and other technical means, a model for predicting product quality and optimizing initial process parameters is obtained. These models provide strong support for the refined control of coffee secondary fermentation process and can significantly improve the stability and controllability of product quality.

[0071] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned method for optimizing the secondary fermentation process of coffee.

[0072] A third aspect of the present invention provides a coffee secondary fermentation process optimization system, which includes the above-mentioned computer-readable storage medium.

[0073] Specifically, the principle of the present invention is to establish a quantitative correlation model between fermentation parameters, fermentation status and final product quality by making full use of experimental data, and perform intelligent optimization based on this. Specifically, the method includes the following key steps:

[0074] First, in a small-scale coffee secondary fermentation experiment, the system collected data samples of the entire process. These data include initial fermentation parameters, fermentation parameters and state parameters at each time point, and multiple quality indicators of the final product. The collection time is divided into two methods: sequential time collection and key stage collection. It can not only obtain the overall change trend of the fermentation process, but also accurately capture the parameter status at key time points. These comprehensive experimental data lay the foundation for subsequent analysis and modeling.

[0075] Secondly, based on the experimental data obtained, the time series model of each fermentation parameter and the dynamic change model of each fermentation state parameter were established. For the fermentation parameters, time series analysis methods such as ARIMA were used to characterize their change patterns over time; for the fermentation state parameters, mathematical models such as differential equations or difference equations were used to describe their dynamic characteristics. The establishment of these models helps to more accurately express the changing trends of various parameters during the fermentation process.

[0076] Thirdly, the Bayesian network and time series analysis methods were used to establish the correlation mapping between fermentation parameters, fermentation status and final coffee quality. Bayesian network can effectively characterize the conditional dependency between variables, and time series analysis can reflect the change law of parameters over time. Through this modeling method, the quantitative relationship between fermentation process parameters and final quality was obtained, which provided a basis for the next step of optimization.

[0077] After that, the above experimental data collection and model building steps were repeated to obtain multiple sets of full-process data and their corresponding association mappings. Based on these association mappings, a multi-input and multi-output embedded deep neural network model was trained as the first prediction model. This model can be used to predict the quality of coffee under different combinations of fermentation parameters and state parameters. In order to further improve the performance of the model, data augmentation and knowledge distillation techniques were also used to optimize it into the second prediction model.

[0078] Based on the second prediction model, a reverse neural network was constructed. The goal of this reverse network is to achieve the mapping relationship from the target coffee quality parameters to the initial fermentation parameters. By training this reverse network, a third model was obtained, which can be used to predict the optimal initial fermentation parameters under given target quality parameters.

[0079] Finally, using the preset ideal coffee quality parameters as input, the third model is used for reverse calculation to invert the optimal initial parameters of coffee secondary fermentation. At the same time, these optimal initial parameters are input into the fermentation parameter dynamic change module of the first prediction model to predict the fermentation parameter change trend at each sampling moment. These optimized initial parameters and dynamic change predictions provide a strong basis for the refined regulation of the actual production process.

[0080] In summary, the coffee secondary fermentation process optimization method proposed in the present invention fully integrates a number of advanced technical means such as experimental data analysis, machine learning, and dynamic modeling. It can systematically characterize the dynamic characteristics of the fermentation process, realize the reverse design from target quality to optimal initial parameters, and continuously optimize process parameters, thereby significantly improving the stability and controllability of product quality.

[0081] In order to better understand and implement the present invention, a specific embodiment 1 of the method provided by the first aspect of the present invention is provided below. The steps of this embodiment 1 are described in detail as follows: First, in step S10, the whole process data of the small-scale experimental production of secondary fermentation of coffee is obtained. These data include the initial fermentation parameter p0, the fermentation parameter p(t) and the fermentation state parameter s(t) at each collection moment, and the coffee quality parameter q after the secondary fermentation production is completed. Among them, the collection moments are divided into two types: serial time collection and key stage collection. Serial time collection is to collect data once every Δt time, which can be expressed as ti =t0+iΔt, i=0,1,2,...,N. The key stage is when the fermentation starts (t0) and the pH value drops to the lowest point for the first time. When the ethanol content reaches its peak At the end of fermentation (t end ) and other important moments.

[0082] The collected fermentation parameters p(t) include temperature T(t), pH value pH(t), dissolved oxygen content DO(t), stirring speed ω(t), fermentation tank pressure P(t), fermentation liquid specific gravity ρ(t), conductivity k(t), oxidation-reduction potential ORP(t), carbon dioxide concentration CO2(t), ammonia nitrogen concentration NH3(t), total nitrogen content TN(t), free amino nitrogen content FAN(t), water activity a w (t) and color of fermentation liquid color(t). Fermentation state parameters s(t) include ethanol content EtOH(t), organic acid content acid(t), sugar content sugar(t), microbial population microbe(t) and CO2 generation rate CO′2(t). Coffee quality parameters q include flavor complexity flavor, acidity, sweetness, bitterness, mouthfeel and aroma intensity aroma. These comprehensive data collections lay the foundation for subsequent analysis and modeling.

[0083] In step S20, each fermentation parameter p is established according to the whole process data obtained in step S10. i (t), and each fermentation state parameter s j (t) dynamic change model. For the fermentation parameter p i The time series modeling of (t) can be done using the autoregressive moving average (ARIMA) model:

[0084] p i (t) = φ i,1 p i (t-1)+φ i,2 p i (t-2)+...+φ i,p p i (tp)+θ i,1 ∈(t-1)+θ i,2 ∈(t-2)+...+θ i,q ∈(tq)+∈(t);

[0085] Among them, φ i,1 ,φ i,2 ,...,φ i,p is the autoregressive coefficient, θi,1 ,θ i,2 ,...,θ i,q is the moving average coefficient, ∈(t) is white noise. By fitting the ARIMA model, the variation of fermentation parameters over time can be described more accurately.

[0086] For the fermentation state parameter s j To model the dynamic changes of (t), mathematical models such as differential equations or difference equations can be used. Taking the dynamic changes of ethanol content EtOH(t) during fermentation as an example, it can be described by the following differential equation:

[0087]

[0088] Among them, k pro is the ethanol production rate, k con is the ethanol consumption rate. By fitting this differential equation, the dynamic change characteristics of ethanol content over time can be captured. Similarly, other fermentation state parameters can also be modeled using corresponding mathematical models.

[0089] In step S30, the Bayesian network and time series analysis method are used to establish each fermentation parameter p i (t) time series model, each fermentation state parameter s j The dynamic change model of (t) and each coffee quality parameter q k The Bayesian network can characterize the conditional dependency between variables, and its joint probability distribution can be expressed as:

[0090] P(p,s,q)=P(p)P(s|p)P(q|p,s);

[0091] Among them, P(p) is the prior probability distribution of fermentation parameters, P(s|p) is the conditional probability distribution of fermentation state parameters, and P(q|p,s) is the conditional probability distribution of quality parameters. By learning this Bayesian network model, we can obtain a quantitative correlation mapping between fermentation parameters, fermentation state and final coffee quality.

[0092] In step S40, steps S10-S30 are repeated to obtain multiple sets of full process data and their corresponding association maps. Based on these association maps, a multi-input and multi-output embedded deep neural network model is trained to obtain the first prediction model The input of the model includes the initial fermentation parameters p0 and time t, and the output is the predicted value of the fermentation parameters at each sampling time. and predicted coffee quality parameters The model structure can be expressed as:

[0093]

[0094] Among them, f p and f q They are fermentation parameter prediction model and quality prediction model, θ p and θ q are their respective model parameters. This first prediction model It can be used to predict coffee quality under different combinations of fermentation parameters and state parameters.

[0095] In step S50, the multiple sets of full-process data obtained in step S40 are first expanded using data augmentation technology to generate more diverse virtual samples. Data augmentation can generate new samples by randomly perturbing, interpolating, and translating the original data, thereby enhancing the generalization ability of the model. Then, the knowledge distillation technology is used to convert the first prediction model into As a teacher model, train a smaller and more efficient student model The objective function of knowledge distillation can be expressed as:

[0096]

[0097] in, and are the quality prediction outputs of the teacher model and the student model, is the training data set. By minimizing this loss function, the knowledge of the teacher model can be effectively transferred to the student model.

[0098] In step S60, the teacher model in step S50 is transformed into The knowledge transfer to the student model , the second prediction model is obtained This second prediction model has smaller model complexity and higher computational efficiency while maintaining high prediction accuracy, and can be more easily deployed in actual coffee production systems.

[0099] In step S70, based on the second prediction model A reverse neural network is constructed. The goal of this reverse network is to achieve the target coffee quality parameter q ＊ The mapping relationship from θ to the initial fermentation parameter p0. This inverse network can be expressed as:

[0100] p0=f inv (q ＊ θ inv );

[0101] Among them, f inv is the inverse mapping function, θ invis the parameter of the reverse network. The reverse network is trained using the back propagation algorithm and gradient descent method to obtain the third model Used to predict the optimal initial fermentation parameters under given target quality parameters.

[0102] In step S80, the preset ideal coffee quality parameter q is used. ＊ As input, the third model trained in step S70 Perform reverse calculation to obtain the optimal initial fermentation parameters for secondary fermentation of coffee These optimal initial fermentation parameters can be used as a reference for the actual production process and provide a basis for process adjustment.

[0103] In step S90, the optimal initial fermentation parameters obtained in step S80 are Input to the first prediction model in step S40 Mathematical model of fermentation parameter changes p The model can predict the fermentation parameter values ​​at each collection time under these optimal initial parameters. These predicted fermentation parameter values ​​can provide a reference for parameter adjustment in the actual production process, helping to achieve more refined process control.

[0104] In order to better understand and implement the present invention, the following is a specific application scenario of the present invention, Example 2: In order to further improve the quality control level of its coffee secondary fermentation process, a coffee company decided to adopt the data-driven coffee secondary fermentation process optimization method proposed in the present invention. The company selected a secondary fermentation pilot laboratory in its production base and conducted comprehensive data collection and analysis of the fermentation process.

[0105] First, the company collected data samples from the entire process of coffee secondary fermentation in a small-scale experimental production. Specifically, it collected:

[0106] 1) Initial fermentation parameters p0: temperature T0 = 40 °C, pH value pH0 = 5.2, dissolved oxygen content DO0 = 6.5 mg / L, stirring speed ω0 = 120 rpm, fermentation tank pressure P0 = 1.2 bar, fermentation liquid specific gravity ρ0 = 1.045, conductivity κ0 = 3.2 mS / cm, redox potential ORP0 = 220 mV, carbon dioxide concentration Ammonia nitrogen concentration Total nitrogen content TN0 = 0.35g / L, free amino nitrogen content FAN0 = 0.08g / L, water activity a w,0 =0.95, fermentation liquid color color0 = brown.

[0107] 2) Fermentation parameters p(t) and fermentation state parameters s(t) collected in sequence: Data were collected every 2 hours, and a total of 17 time points were collected, namely t i =2ih, i=0,1,...,16. Fermentation parameters include temperature T(t i ), pH value pH (t i ), dissolved oxygen content DO(t i ), stirring speed ω(t i ), fermentation tank pressure P(t i ), specific gravity of fermentation liquid ρ(t i ), conductivity κ(t i )、Redox Potential ORP(t i ), carbon dioxide concentration CO2(t i ), ammonia nitrogen concentration NH3(t i ), total nitrogen content TN (t i ), free amino nitrogen content FAN(t i ), water activity a w (t i )、Fermentation liquid color color(t i ); Fermentation state parameters include ethanol content EtOH (t i ), organic acid content acid(t i ), sugar content sugar(t i ), microbial count microbe(t i ), CO2 generation rate CO2′(t i ). These data are shown in Table 1 below:

[0108] Table 1. Data collected

[0109]

[0110] 3) Collection of fermentation parameters and fermentation status parameters at key stages:

[0111] At the beginning of fermentation: the same as the initial parameters p0 and s0 above

[0112] When the pH drops to its lowest point for the first time

[0113] When the ethanol content reaches its peak

[0114] At the end of fermentation (t end =34h):EtOH(t end )=2.8%,acid(t end )=6.5g / L;

[0115] 4) Coffee quality parameters q after the secondary fermentation production is completed: flavor complexity flavor = 4.2, acidity acidity = 4.0, sweetness sweetness = 3.5, bitterness bitterness = 3.8, mouthfeel = 4.1, aroma intensity aroma = 4.3.

[0116] With these comprehensive experimental data, we started to carry out analysis, modeling and optimization according to the method of the present invention.

[0117] First, based on the fermentation parameter and state parameter data collected in sequence time, the time series model and dynamic change model of each parameter were established. Taking temperature T(t) and ethanol content EtOH(t) as examples:

[0118] The time series model of temperature T(t) can be expressed using the ARIMA(2,1,1) model:

[0119] T(t)=1.25T(t-1)-0.32T(t-2)+0.15∈(t-1)+∈(t);

[0120] Among them,∈(t) is white noise. This model can better describe the change of temperature over time.

[0121] The dynamic changes of ethanol content EtOH(t) can be described by the following differential equation:

[0122]

[0123] Among them, 0.12 represents the ethanol production rate, and 0.045 represents the ethanol consumption rate. By fitting the differential equation, the dynamic change characteristics of ethanol content over time were obtained.

[0124] Similarly, time series models and dynamic change models of other fermentation parameters and state parameters were also established.

[0125] Next, the Bayesian network and time series analysis methods were used to establish the association mapping between fermentation parameters, fermentation status and coffee quality. The joint probability distribution of the Bayesian network can be expressed as:

[0126] P(p,s,q)=P(p)P(s|p)P(q|p,s);

[0127] By learning this model, we found that parameters such as temperature T, pH value, and ethanol content EtOH have a greater impact on quality indicators such as flavor complexity and acidity, and the color of the fermentation liquid also has a significant impact on aroma intensity. These quantitative relationships provide a basis for subsequent optimization.

[0128] Based on the above modeling results, a multi-input and multi-output deep neural network model was trained as the first prediction model. The input of the model includes the initial fermentation parameters p0 and time t, and the output is the predicted value of the fermentation parameters at each time point. and predicted coffee quality parameters Using the experimental data obtained above to train it, the first prediction model was obtained In order to further improve the performance of the model, data augmentation and knowledge distillation techniques are used to As a teacher model, a smaller and more efficient student model is trained This second prediction model has lower model complexity and higher computational efficiency while maintaining high prediction accuracy, and can be more easily deployed in actual production systems. A reverse neural network was constructed to obtain the target quality parameter q ＊ Determine the optimal initial fermentation parameters The ideal coffee quality parameter q ＊ =[4.5 4.2 4.0 3.9 4.3 4.5] is input into this reverse network, and the third model is obtained through training. It can predict the optimal initial fermentation parameters

[0129] Finally, Input to In the fermentation parameter change module, the fermentation parameter changes at each collection time are predicted as shown in Table 2:

[0130] Table 2 Fermentation parameter changes

[0131]

[0132] These optimized initial fermentation parameters And the predicted values ​​of fermentation parameters at each time point This provides a strong basis for fine-tuning parameter control in the actual production process. Figure 2-7 shown.

[0133] Attached Figure 2 The comparison of coffee quality parameters before and after optimization is shown. It can be seen intuitively from the figure that the optimized coffee has significantly improved in flavor complexity, acidity, sweetness, taste and aroma intensity, while the bitterness is appropriately reduced, making the overall flavor more balanced and coordinated.

[0134] Attached Figure 3The comparison of key fermentation parameters before and after optimization is shown. As can be seen from the figure, during the optimized fermentation process, the temperature increased slightly, the pH value decreased slightly, the dissolved oxygen content increased, the pressure increased slightly, the conductivity increased, and the CO2 concentration also increased. The adjustment of these parameters worked together to promote the improvement of coffee quality.

[0135] Attached Figure 4 The changing trend of key fermentation parameters over time is shown. As can be seen from the figure, in the optimized fermentation process: the temperature (red line) shows a slow upward trend, which is conducive to the fermentation process; the pH value (blue line) gradually decreases over time, indicating the generation of organic acids during the fermentation process; the ethanol content (green line) first increases and then decreases slightly, reflecting the process of alcohol fermentation and subsequent conversion.

[0136] Attached Figure 5 :Coffee quality parameter radar chart, showing the changes in six coffee quality parameters before and after optimization.

[0137] Attached Figure 6 : The bar chart comparing the fermentation parameters before and after optimization compares the changes of four key fermentation parameters before and after optimization.

[0138] Attached Figure 7 The change curve of key parameters in the fermentation process shows the dynamic changes of temperature, pH value and ethanol content during the fermentation process.

[0139] By implementing the coffee secondary fermentation process optimization method proposed in the present invention, the company has achieved remarkable results:

[0140] 1) The flavor complexity of the product increased from 4.2 to 4.5, the acidity increased from 4.0 to 4.2, the sweetness increased from 3.5 to 4.0, the bitterness decreased from 3.8 to 3.6, the taste increased from 4.1 to 4.3, and the aroma intensity increased from 4.3 to 4.5. All quality indicators have reached the ideal level, and the product taste is more harmonious and rich.

[0141] 2) The stability and controllability of process parameters have been greatly improved. By modeling the dynamic characteristics of each parameter in the fermentation process and combining it with the reverse optimization of the target quality parameters, the parameter adjustment of the production process is more accurate and efficient. The batch-to-batch variation of product quality has been significantly reduced.

[0142] 3) Production efficiency and energy consumption have been significantly improved. With the help of precise process parameter control, the fermentation time was shortened from 36 hours to 34 hours, and the energy consumption per unit product was reduced by about 5%.

[0143] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A coffee secondary fermentation process optimization method, characterized in that: The following steps are involved: S10, obtaining the whole process data of the small-scale experimental production of secondary fermentation of coffee, including the initial fermentation parameters, the fermentation parameters and fermentation state parameters at each collection moment, and the coffee quality parameters after the secondary fermentation production is completed; S20, establishing a time series model of each fermentation parameter and a dynamic change model of each fermentation state parameter according to the whole process data; S30, using Bayesian network and time series analysis to establish a time series model for each fermentation parameter and a correlation mapping between a dynamic change model for each fermentation state parameter and each coffee quality parameter; S40, repeating steps S10-S30 to obtain multiple sets of full-process data and their corresponding association maps, and training a multi-input multi-output embedded deep neural network model according to the multiple sets of association maps to obtain a first prediction model for predicting coffee quality parameters under different combinations of fermentation parameters and state parameters; the first prediction model is embedded with a fermentation parameter variation mathematical model; S50, using data augmentation technology to expand the multiple sets of full-process data to generate more diverse virtual samples to enhance the generalization ability of the model; using knowledge distillation technology to use the first prediction model as a teacher model to train a smaller and more efficient student model; S60, transferring the knowledge of the teacher model to the student model to obtain a second prediction model; S70, based on the second prediction model, constructing a reverse neural network, using a back propagation algorithm and a gradient descent method, training the network to achieve mapping from the target coffee quality parameters to the initial fermentation parameters, and obtaining a third model; S80, using the preset ideal coffee quality parameters as input, performing reverse calculation through the third model, inversely calculating and outputting the optimal initial fermentation parameters for the secondary fermentation of the coffee; S90, inputting the optimal initial fermentation parameters of the coffee secondary fermentation into the fermentation parameter variation mathematical model in the first prediction model, obtaining the fermentation parameters at multiple collection moments and outputting them as a reference for adjusting the fermentation parameters in the coffee secondary fermentation process.

2. The coffee secondary fermentation process optimization method according to claim 1, characterized in that: The fermentation parameters include adjustable parameters and non-adjustable parameters. The adjustable parameters are specifically temperature, stirring speed, dissolved oxygen content, and fermentation tank pressure. The non-adjustable parameters are specifically pH value, specific gravity of fermentation liquid, conductivity, redox potential, carbon dioxide concentration, ammonia nitrogen concentration, total nitrogen content, free amino nitrogen content, water activity, and fermentation liquid color.

3. The coffee secondary fermentation process optimization method according to claim 2, characterized in that: The structure of the multi-input and multi-output embedded deep neural network model specifically includes a mathematical model of fermentation parameter changes, a branch network of adjustable parameters affecting quality, a branch network of non-adjustable parameters affecting quality, a branch network of long-term parameters affecting quality, a branch network of short-term parameters affecting quality, and a fusion network.

4. The coffee secondary fermentation process optimization method according to claim 3, characterized in that: The fermentation parameter variation mathematical model is used to predict the dynamic variation of various parameters during the fermentation process, the input is the initial fermentation parameters and time, and the output is the predicted value of the fermentation parameters at each time point.

5. The coffee secondary fermentation process optimization method according to claim 4, characterized in that: The adjustable parameter quality impact branch network is used to simulate the impact of adjustable parameters on coffee quality. The input is the time series data of the adjustable parameters, and the output is the impact degree of each quality parameter. The structure is a combination of a multi-layer convolutional neural network and a long short-term memory network.

6. The coffee secondary fermentation process optimization method according to claim 5, characterized in that: The non-adjustable parameter impact on quality branch network is used to simulate the impact of non-adjustable parameters on coffee quality. The input is the time series data of non-adjustable parameters, and the output is the impact degree of each quality parameter. The structure is a combination of a multi-layer perceptron and an attention mechanism.

7. The coffee secondary fermentation process optimization method according to claim 6, characterized in that: The long-term parameter impact branch network is used to capture the impact of the long-term change trend of fermentation parameters on quality. The input is the long-term span data of all fermentation parameters, the output is the long-term impact factor, and the structure is a time series analysis network based on Transformer.

8. The coffee secondary fermentation process optimization method according to claim 7, characterized in that: The short-term parameter impact branch network is used to capture the impact of short-term fluctuations in fermentation parameters on quality. The input is short-time window data of all fermentation parameters, and the output is a short-term impact factor. The structure is a recurrent neural network based on a gated recurrent unit. The fusion network is used to integrate the outputs of each branch network and generate the final quality prediction result. The input is the output of each branch network, and the output is the predicted coffee quality parameter. The structure is a combination of a multi-layer fully connected network and a softmax layer.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed in a computer, they are used to execute the coffee secondary fermentation process optimization method according to any one of claims 1 to 8.

10. A coffee secondary fermentation process optimization system, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.

Images