Saccharomyces cerevisiae ethanol fermentation process intelligent feedback feeding control method

Abstract

The application discloses a kind of saccharomyces cerevisiae ethanol fermentation process intelligent feedback feeding control methods, including sampling spectrum, obtains sampling spectrum;The sampling spectrum is input into two-dimensional convolutional neural network based on spectrum amplification;Using mean square error as loss function, using Adam optimizer updates network weight when training raman concentration prediction model, obtains optimal solution, updates raman concentration prediction model.This method can learn, adjust through raman concentration prediction model, improve classification accuracy, the glucose concentration, glycerol concentration and ethanol concentration in saccharomyces cerevisiae ethanol fermentation process in bioreactor are predicted in real time online.Compared with conventional detection method, detection is fast and has little effect on fermentation process, reduces the risk of contamination, and can realize online monitoring such as ethanol and other volatile substance concentration, also can online monitoring glucose and glycerol and other non-volatile substance concentration, make up the deficiency in detection principle.

Description

Technical Field

[0001] This invention relates to the field of feeding control technology for ethanol fermentation, and particularly to an intelligent feedback feeding control method for the ethanol fermentation process of brewer's yeast. Background Technology

[0002] Industrial ethanol fermentation typically uses high-sugar batch fermentation due to its advantages such as high yield, high equipment utilization, low energy consumption, and low production cost. Glucose is the primary carbon and energy source for cells, consumed during cell culture and then converted into biomass and products. However, maintaining different glucose concentrations during culture alters the cellular metabolic state. Therefore, glucose concentration is usually a critical process control parameter in fermentation. In batch fermentation, high initial sugar concentrations (~30% (w / v)) can easily create a high osmotic pressure environment, leading to the generation of reactive oxygen species (ROS) and the accumulation of byproducts such as glycerol, thereby affecting the normal physiological metabolism of yeast and its ethanol fermentation capacity. Fed-batch fermentation can effectively reduce the inhibition caused by high sugar levels. Instead of adding the substrate to the fermentation broth all at once, the rate of sugar addition is controlled to match the yeast's metabolic rate, ensuring that the sugar concentration remains at a constant optimal level.

[0003] In existing technologies, in the fed-batch glucose method using online capacitance monitoring, the optimal bacterial growth during fermentation is achieved when the initial glucose concentration is 90–110 g / L. Furthermore, controlling the glucose concentration below 100 g / L through fed-batch fermentation increases ethanol yield by 15.4% compared to batch fermentation. However, this method lacks detailed exploration of controlling process sugar concentration, relies on manual control without online feedback, and suffers from a large control error range. Since residual sugar concentration detection during fermentation typically involves manual sampling and traditional offline analysis, frequent sampling not only poses a risk of contamination but also disrupts cell metabolism due to disturbances in the culture medium. Additionally, offline analysis is time-consuming, requiring sample pretreatment and analysis using instruments such as high-performance liquid chromatography (HPLC), making process feedback difficult to implement. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies do not conduct a more detailed investigation of sugar concentration control during the process, and process control is also done manually without online feedback control, resulting in a large control error range. Since the detection of residual sugar concentration during fermentation typically relies on manual sampling and traditional offline analysis, frequent sampling not only poses a risk of contamination but also affects cell metabolism due to disturbances in the culture medium. Furthermore, offline analysis is very time-consuming, requiring sample pretreatment and analysis using instruments such as high-performance liquid chromatography (HPLC), making process feedback difficult to implement.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: an intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast, comprising: sampling the spectrum to obtain a sampled spectrum; inputting the sampled spectrum into a two-dimensional convolutional neural network based on spectral amplification; using mean squared error as a loss function, using the Adam optimizer to update the network weights during the training of the Raman concentration prediction model to obtain the optimal solution, and updating the Raman concentration prediction model; collecting the current spectral data in the fermenter, inputting the current spectral data into the updated Raman concentration prediction model, outputting a glucose concentration value, and supplementing with a high-concentration glucose aqueous solution according to the glucose concentration value.

[0006] As a preferred embodiment of the intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast described in this invention, the method involves continuously sampling the spectrum to obtain multiple spectra.

[0007] As a preferred embodiment of the intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast described in this invention, the multiple spectra are spliced ​​together to form a sampling spectrum.

[0008] As a preferred embodiment of the intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast described in this invention, the two-dimensional convolutional neural network based on spectral amplification includes a convolutional layer, a BN layer, and a fully connected layer.

[0009] As a preferred embodiment of the intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast described in this invention, the convolutional layer includes a first convolutional layer, a second convolutional layer, and a third convolutional layer;

[0010] The first convolutional layer has a 5×5 kernel, 8 kernels, and an input of 6×5 × 11 × 8.

[0011] The second convolutional layer has a 3×3 kernel, 16 kernels, and an input of 4×509×16.

[0012] The third convolutional layer has a 2×2 kernel, a total of 32 kernels, and an input of 3×508×32.

[0013] As a preferred embodiment of the intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast described in this invention, the first convolutional layer, the second convolutional layer, and the third convolutional layer are activated using the ReLU activation function.

[0014] As a preferred embodiment of the intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast described in this invention, the BN layer includes a first BN layer, a second BN layer, and a third BN layer.

[0015] The first convolutional layer is connected to the first batch normalization layer;

[0016] The second convolutional layer is connected to the second batch normalization layer;

[0017] The third convolutional layer is connected to the third batch normalization (BN) layer.

[0018] As a preferred embodiment of the intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast described in this invention, the method employs mean squared error as the loss function and uses the Adam optimizer to update the network weights during the training of the Raman concentration prediction model, thereby obtaining the optimal solution of the Raman concentration prediction model, including:

[0019] Calculate the gradient g of the current parameters t Its mathematical expression is:

[0020]

[0021] Among them, f t (θ) represents the loss function at time step t, updating the estimates of the first and second moments, and its mathematical expression is:

[0022] m t =β1·m t-1 +(1-β1)·g t

[0023] v t =β2·v t-1 +(1-β2)·g t 2

[0024] The mathematical expressions for estimating the first and second moments after bias correction are as follows:

[0025]

[0026] The parameters are adjusted using the bias-corrected estimates of the first and second moments to more accurately estimate the descent direction of the parameters. The mathematical expression for this is:

[0027]

[0028] Where t represents the current iteration number, θ represents the parameters of the Raman concentration prediction model, α represents the learning rate, and m t v represents the estimate of the first-order moment. t This represents the estimate of the second moment.

[0029] As a preferred embodiment of the intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast described in this invention, the method includes: collecting current spectral data in the fermenter, inputting the current spectral data into an updated Raman concentration prediction model, outputting a glucose concentration value, and if the glucose concentration value is lower than a preset glucose concentration threshold, supplementing with a high-concentration glucose aqueous solution until the glucose concentration value reaches the glucose concentration threshold.

[0030] The beneficial effects of this invention are as follows: This method, through a Raman concentration prediction model, can learn and adjust to improve classification accuracy, enabling real-time online prediction of glucose, glycerol, and ethanol concentrations during the ethanol fermentation process of *Saccharomyces cerevisiae* in a bioreactor. Compared to conventional detection methods, this method is rapid and has almost no impact on the fermentation process, reducing the risk of contamination. Furthermore, it can achieve online monitoring of volatile substances such as ethanol, as well as non-volatile substances such as glucose and glycerol, thus overcoming the shortcomings of conventional detection principles. Attached Figure Description

[0031] Figure 1 This is a basic flowchart illustrating a method for intelligent feedback feeding control in the ethanol fermentation process of brewing yeast, provided as an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram showing the changes in glucose concentration in the first fed-batch fermentation process and the control batch of a smart feedback feeding control method for the ethanol fermentation process of brewing yeast, provided in an embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram of the glucose concentration change in a second-stage feed-fed batch fermentation process, provided by an embodiment of the intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast according to an embodiment of the present invention.

[0034] Figure 4 This is a schematic diagram illustrating the changes in glucose concentration during four batches of fermentation in industrial-scale production of a smart feedback feeding control method for the ethanol fermentation process of brewing yeast, as provided in an embodiment of the present invention.

[0035] Figure 5 This is a schematic diagram of the glucose concentration change in the third batch fermentation process of a smart feedback feeding control method for the ethanol fermentation process of brewing yeast, provided in an embodiment of the present invention.

[0036] Figure 6 This is a schematic diagram of the Raman online signal before pretreatment in an intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast, provided as an embodiment of the present invention.

[0037] Figure 7This is a schematic diagram of the preprocessed Raman online signal of a smart feedback feeding control method for the ethanol fermentation process of brewing yeast, provided in an embodiment of the present invention.

[0038] Figure 8 This is a schematic diagram of the real-time Raman prediction of the PLSR algorithm Raman concentration prediction model provided in an embodiment of the present invention during the control process of a glucose concentration of 30 g / L.

[0039] Figure 9 This is a schematic diagram of real-time Raman prediction using the SA-CNN algorithm Raman concentration prediction model provided in an embodiment of the present invention during a controlled concentration of 20 g / L.

[0040] Figure 10 This is a schematic diagram of the real-time Raman prediction of ethanol, glycerol and glucose concentrations using the SA-CNN algorithm Raman concentration prediction model provided in an embodiment of the present invention during the control process at a concentration of 20 g / L.

[0041] Figure 11 This is a schematic diagram comparing the cell dry weight of a sugar-supplemented batch and a control batch in an intelligent feedback feeding control method for the ethanol fermentation process of Saccharomyces cerevisiae, as provided in an embodiment of the present invention.

[0042] Figure 12 This is a schematic diagram comparing the ethanol concentration during the fermentation process of a sugar-supplemented batch and a control batch, according to an embodiment of the present invention, which provides an intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast.

[0043] Figure 13 This diagram illustrates a comparison of glycerol concentrations during the fermentation process of a sugar-supplemented batch and a control batch, provided by an embodiment of the present invention for an intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast. Detailed Implementation

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0045] Example 1

[0046] Reference Figure 1 As an embodiment of the present invention, a smart feedback feeding control method for the ethanol fermentation process of brewing yeast is provided, comprising:

[0047] S1: Sample the spectrum to obtain the sampled spectrum, including:

[0048] The spectrum is continuously sampled to obtain multiple spectra.

[0049] The multiple spectra are stitched together to form a sampled spectrum. A spectral amplification layer then stitches the continuous temporal sampled spectra according to their corresponding Raman shifts into a two-dimensional spectrum, which serves as the input to a two-dimensional convolutional neural network.

[0050] Raman spectroscopy can simultaneously measure multiple substances.

[0051] Raman spectroscopy, as a PAT (Process Analytical Technology) tool, relies on a stoichiometric Raman concentration prediction model that correlates Raman signals with parameters. Based on the constructed mathematical Raman concentration prediction model, Raman spectroscopy can be used directly for real-time monitoring of the fermentation process, replacing the tedious manual sampling and measurement. The application of Raman spectroscopy in PAT depends on using a calibrated Raman concentration prediction model that correlates spectral signals with analytical measurements.

[0052] S2: Inputting the sampled spectrum into a two-dimensional convolutional neural network (SA-CNN) based on spectral amplification includes:

[0053] Two-dimensional convolutional neural networks based on spectral amplification include convolutional layers, batch normalization (BN) layers, and fully connected layers.

[0054] Before being passed to the fully connected layer, the feature maps first need to be flattened into a one-dimensional vector. If there are multiple feature maps, they are unfolded side-by-side and concatenated into a long vector. This is because the neurons in the fully connected layer are connected to every element in the input data, thus requiring a one-dimensional data format. The flattened feature vector serves as the input to the fully connected layer.

[0055] The convolutional layer includes a first convolutional layer, a second convolutional layer, and a third convolutional layer;

[0056] The first convolutional layer has a 5×5 kernel, 8 kernels, and an input of 6×5 × 11 × 8.

[0057] The second convolutional layer has a 3×3 kernel, 16 kernels, and an input of 4×509×16.

[0058] The third convolutional layer has a 2×2 kernel, a total of 32 kernels, and an input of 3×508×32.

[0059] The first, second, and third convolutional layers are activated using the ReLU (Rectified linear unit) activation function.

[0060] The BN layer includes a first BN layer, a second BN layer, and a third BN layer;

[0061] The first convolutional layer is connected to the first batch normalization layer;

[0062] The second convolutional layer is connected to the second batch normalization layer;

[0063] The third convolutional layer is connected to the third batch normalization (BN) layer.

[0064] The first convolutional layer uses a large 5×5 kernel to capture a wider range of features. In the initial stages of the network, a larger kernel helps capture fundamental features and patterns in the input data. As the number of layers increases, the kernel size gradually decreases. This design helps reduce the number of parameters while preserving important feature information, allowing the network to focus more on detailed features and reducing the risk of overfitting. The activation function for each convolutional layer is ReLU (Rectified Linear Unit). ReLU retains many advantages of linear Raman concentration prediction models (such as ease of optimization) while introducing necessary nonlinearity. Furthermore, since the derivative of ReLU in the positive part is constant, it helps mitigate the vanishing gradient problem. Each convolutional layer is followed by a BN (Batch Normalization) layer. This design reduces the shifting problem of internal covariates, accelerating the convergence of the Raman concentration prediction model and enhancing its stability. Mean Squared Error (MSE) is used as the loss function. The Adam (Adaptive Moment Estimation) optimizer is used to efficiently update network weights when training a Raman concentration prediction model. By combining momentum and adaptive learning rate (RMSprop), the Adam optimizer helps to converge to the optimal solution quickly.

[0065] S3: Using mean squared error as the loss function, the Adam optimizer is used to update the network weights during the training of the Raman concentration prediction model to obtain the optimal solution. The update of the Raman concentration prediction model includes:

[0066] The Adam (adaptive moment estimation) optimizer is a widely used gradient descent optimization algorithm that combines the concepts of momentum and RMSprop (Root Mean Square Propagation) to achieve a fast and stable training process. The Adam optimizer's computational process and steps include:

[0067] Initialization parameters:

[0068] α: Learning rate (Recommended default value: 0.001)

[0069] β1, β2: Exponential decay rates estimated by the first and second moments (recommended default values: 0.9 and 0.999, respectively)

[0070] ∈: Used to prevent division by zero (recommended default value: 10-8)

[0071] θ: Parameter to be optimized

[0072] m0: First-order moment vector (initialized to 0)

[0073] v0: Second-order moment vector (initialized to 0)

[0074] t: Time step (initialized to 0)

[0075] For each training step:

[0076] Calculate the gradient g of the current parameters t Its mathematical expression is:

[0077]

[0078] Among them, f t (θ) represents the loss function at time step t, updating the estimates of the first and second moments, and its mathematical expression is:

[0079] m t =β1·m t-1 +(1-β1)·g t

[0080] v t =β2·v t-1 +(1-β2)·g t 2

[0081] The mathematical expressions for estimating the first and second moments after bias correction are as follows:

[0082]

[0083] The parameters are adjusted using the bias-corrected estimates of the first and second moments to more accurately estimate the descent direction of the parameters. The mathematical expression for this is:

[0084]

[0085] Where t represents the current iteration number, θ represents the parameters of the Raman concentration prediction model, α represents the learning rate, and m t v represents the estimate of the first-order moment. t This represents the estimate of the second moment, thus allowing for faster adaptation to the training process.

[0086] S4: Collect the current spectral data in the fermenter, input the current spectral data into the updated Raman concentration prediction model, output the glucose concentration value, and replenish the high-concentration glucose aqueous solution according to the glucose concentration value, including:

[0087] Collect the current spectral data in the fermenter, input the current spectral data into the updated Raman concentration prediction model, and output the glucose concentration value. If the glucose concentration value is lower than the preset glucose concentration threshold, add high-concentration glucose aqueous solution until the glucose concentration value reaches the glucose concentration threshold.

[0088] Feeding control is performed based on the average of the previous three consecutive glucose concentration output values ​​to increase process control stability. A glucose concentration value is set on the computer. When the concentration falls below the set value (within the range of -1 g / L), the computer uses a PID (Proportional Integral Differential) algorithm to control a peristaltic pump (Lange pump BT100-2J) to slowly add a high-concentration glucose solution (approximately 90%) into the fermenter until the set value is reached, at which point feeding stops. Each feeding cycle lasts 30 seconds, and the prediction interval is 15 seconds, resulting in a feeding control cycle of 45 seconds. Due to the lag and fluctuation of the predicted value, the glucose concentration may sometimes deviate significantly from the target value. The PID algorithm adjusts the pump speed based on the degree of deviation, with the pump speed generally ranging from 0.1 ml / min to 2.4 ml / min. When the deviation is small, the pump speed is slower to prevent excessive glucose solution from being added and causing the concentration to exceed the target value significantly. When the deviation is large, the pump speed is faster to quickly bring the glucose concentration in the fermenter to the target value. The average of three consecutive glucose concentration predictions is used for feeding control, and a PID algorithm is introduced to adjust the pump speed according to the degree of deviation from the target value, making the feedback feeding control more intelligent.

[0089] This method employs an optimized deep learning algorithm to predict glucose, glycerol, and ethanol concentrations during the ethanol fermentation process of *Saccharomyces cerevisiae* in a bioreactor, achieving better accuracy and generalization. Furthermore, the experiment involves implementing intelligent feedback feeding control by predicting the glucose concentration in the fermenter in real time, and comparing and analyzing process parameters under different feeding modes to find the optimal strategy for controlling residual sugar concentration and explore the reasons for these parameter changes. The application of this method not only improves prediction accuracy but also demonstrates good results in practical control tasks, which is of great significance for optimizing fermentation process control.

[0090] This method enables real-time online prediction of glucose, glycerol, and ethanol concentrations during the ethanol fermentation of Saccharomyces cerevisiae in a bioreactor. Compared to conventional detection methods, it is rapid and has almost no impact on the fermentation process, reducing the risk of contamination. Furthermore, while existing technologies utilize sensors such as electronic olfaction to measure volatile substances like ethanol during fermentation, their application is primarily limited by the volatility of the substances. The online Raman spectroscopy technology developed in this invention can not only monitor the concentrations of volatile substances like ethanol online, but also the concentrations of non-volatile substances like glucose and glycerol, overcoming the limitations of the detection principle.

[0091] The Raman concentration prediction model constructed using the SA-CNN algorithm in this method can achieve good predictions for both batch fermentation and fed-batch fermentation. Compared with traditional methods such as PLSR, the Raman concentration prediction model has greatly improved in terms of prediction accuracy, range, and practicality.

[0092] This method achieves intelligent feed control during the fermentation process, maintaining the glucose concentration at a precise and constant target concentration. Compared to offline measurement and manual control using instruments such as HPLC and biochemical analyzers, it significantly reduces control errors and enables real-time continuous control of the fermentation process.

[0093] By controlling the glucose concentration during the process, compared to the significant fluctuations in sugar concentration encountered in industrial production, the ethanol production rate was improved, the fermentation time was shortened, and the yield of the byproduct glycerol was reduced, providing an important basis for optimizing the industrial-scale ethanol production process. By controlling the initial sugar concentration at 100 g / L to ensure vigorous cell growth, and by controlling different constant sugar concentrations during the process, comparative analysis revealed that controlling the residual sugar concentration at 30 g / L resulted in the highest final ethanol yield and the lowest glycerol yield. Furthermore, controlling the residual sugar concentration at 80 g / L resulted in the highest ethanol production rate.

[0094] Example 2

[0095] Reference Figures 2-7 This is another embodiment of the present invention. Unlike the first embodiment, this embodiment provides an experimental verification of an intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast. To make the above-mentioned objectives, features and innovations of the present invention more apparent and understandable, the specific implementation methods of the present invention will be described in more detail below with reference to specific implementation cases.

[0096] The fermentation experimental procedures and conditions for a 5L fermenter in this method are as follows:

[0097] 1. Activation of microbial strains

[0098] In a clean bench, take 1 ml of glycerol-freezing brewer's yeast (CEN.PK 113-7D) culture, dissolve it in 9 mL of sterile water, and after mixing, dilute it to 10⁵. Take 40 μL and spread it on a plate (usually three plates). Incubate it in a 30℃ incubator (about 48 h) until a single colony grows.

[0099] 2. Seed culture

[0100] Seed culture was performed using 500mL shake flasks with a volume of 100mL. A single colony (as uniform in size as possible) was picked from the plate medium and inoculated into the shake flask. The culture conditions were 220rpm and 30℃. Samples were taken every 2 hours, diluted to the appropriate concentration, and the optical density (OD) was measured. After approximately 15-16 hours of culture, the OD600 of the seed culture was measured. At this point, the cells were in the late logarithmic growth stage, with an OD600 of approximately 6-8, which was suitable for inoculation into fermentation.

[0101] 3. Inoculate into a 5L fermenter for cultivation.

[0102] After the fermenter has been sterilized and cooled, inoculation should be carried out. It is important to insert the Raman electrode in advance and sterilize it together with the fermenter to prevent contamination caused by subsequent electrode insertion. The alcohol flame inoculation method is used during inoculation. That is, place the inoculation loop at the inoculation port, stuff the inoculation loop with cotton wool containing 75% alcohol, ignite it and start inoculation. Pour in 9 shake flasks of seed liquid and nitrogen source respectively. The inoculation volume is 30%, the culture volume is 3L, the fermentation temperature is 30℃, the rotation speed is 150rpm, the pressure is 0.05MPa, and anaerobic fermentation culture is carried out under nitrogen.

[0103] Calculation of cell dry weight during fermentation:

[0104] Samples were taken every 2 hours, and the dry weight of each sample was measured three times and averaged. The total sample volume was 6 ml. Three sheets of glass fiber filter paper were used for each sample. The glass fiber filter paper was dried in a microwave oven for 4 minutes, weighed, and placed in numbered covered petri dishes for later use. Approximately 6 ml of sample was taken from a 10 ml centrifuge tube through the sampling port. 2 ml of fermentation broth was drawn using a 5 ml pipette, and the accurate weight was recorded. The sample was filtered using pre-weighed glass fiber filter paper. The first sample was directly filtered, and the beaker and filter paper were washed twice with deionized water. Finally, the three sheets of filter paper were placed in labeled petri dishes. The three petri dishes were placed in a microwave oven for 4 minutes. After drying, they were weighed and the cell concentration was calculated.

[0105] Offline data determination of glucose, glycerol, and ethanol concentrations during fermentation:

[0106] Every so often, 2 ml of fermentation broth was taken from the fermenter and centrifuged at 6500 rpm for 5 min. The supernatant was then filtered through a 0.22 μm aqueous filter into two liquid chromatography vials. Mobile phase preparation (5 mM dilute sulfuric acid): 270 μL of concentrated sulfuric acid was diluted to 1 L to prepare 5 M dilute sulfuric acid. Then, 1 mL of 5 M dilute sulfuric acid was diluted to 1 L to prepare the 5 mM dilute sulfuric acid mobile phase. The solution was then filtered using a circulating water vacuum pump, followed by sonication for 30 min before use. Detector: A differential refractive index detector was used. Detection column: Aminex HPX-87H. Mobile phase flow rate: 0.6 mL / min. Detection temperature: 50℃. Injection volume: 15 μL.

[0107] Detection using Raman spectroscopy:

[0108] The Raman electrode was inserted into the fermenter and sterilized along with the fermenter. After sterilization, the fermenter was wrapped with aluminum foil to maintain a dark environment inside. Process data from inside the fermenter was collected using Raman analyzer software.

[0109] Raman analyzer online control intelligent feedback feeding operation steps:

[0110] The Raman analyzer collects spectral data at fixed intervals, transmitting the data to a computer-based Raman concentration prediction model, which outputs a glucose concentration. Due to the fluctuations in predicted glucose concentration, feed control is based on the average of the previous three consecutive glucose concentration predictions, increasing process control stability. A glucose concentration value is set on the computer. When the concentration falls below this value, the computer uses a PID algorithm to control a peristaltic pump, slowly adding a high-concentration glucose solution into the fermenter until the set value is reached, at which point feed control stops. Due to the lag and fluctuations in predicted values, the glucose concentration may sometimes deviate significantly from the target value. The PID algorithm adjusts the pump speed based on the degree of deviation. When the deviation is small, the pump speed is slower to prevent excessive glucose solution from being added and exceeding the target value. When the deviation is large, the pump speed is faster to quickly bring the glucose concentration in the fermenter to the target value. This design, using the average of three consecutive glucose concentration predictions for feed control and incorporating a PID algorithm to adjust the pump speed based on the degree of deviation, makes the feedback feed control more intelligent.

[0111] Reference Figure 2 The batch fermentation process with an initial glucose concentration of 285 g / L was used as the control group, and fermentation ended when the glucose concentration dropped to below 1 g / L.

[0112] The first type of fed-batch fermentation process:

[0113] like Figure 2In this method, the initial glucose concentration for the first fed-batch fermentation process is set to 100 g / L, with maintenance glucose concentrations of 20 g / L, 30 g / L, 40 g / L, 50 g / L, 65 g / L, and 80 g / L. When fermentation begins, the initial glucose concentration is 100 g / L. As fermentation progresses, the glucose concentration gradually decreases until it reaches the set value of 30 g / L. The Raman concentration prediction model controls the feeding based on the prediction results. One cycle is predicted to last 15 seconds, and feeding is performed once per cycle. The feed solution is a 925 g / L glucose aqueous solution with a volume of 600 ml. After all the glucose aqueous solution has been added, fermentation continues until the glucose concentration drops below 1 g / L, at which point fermentation is considered complete and stopped. The fermentation processes for the remaining maintenance processes with different glucose concentrations are described above.

[0114] The second type of supplemental fermentation process:

[0115] like Figure 3 In this method, the second fed-batch fermentation process is set with an initial glucose concentration of 100 g / L. When the glucose concentration drops to around 10 g / L, it is replenished to 80 g / L, for a total of two replenishments. When fermentation begins, the initial glucose concentration is 100 g / L. As fermentation progresses, the glucose concentration gradually decreases until it reaches around 10 g / L. The Raman concentration prediction model then controls the peristaltic pump to replenish the glucose to around 80 g / L once, stopping the replenishment. The Raman concentration prediction model has a prediction cycle of 15 seconds. As fermentation continues, when the concentration drops again to around 10 g / L, the remaining glucose solution is added to the fermenter. Therefore, the sugar solution is replenished twice during fermentation. The replenishment solution is a 925 g / L glucose solution with a volume of 600 ml. After all the glucose solution has been added, fermentation continues until the glucose concentration drops below 1 g / L, at which point fermentation is considered complete and stopped.

[0116] The third type of supplemental batch fermentation process:

[0117] Reference Figure 4 To simulate the changes in glucose concentration during actual industrial-scale production, we used the third type of fed-batch fermentation process to model the changes in glucose concentration during actual industrial production. Therefore, as... Figure 5 As shown, in the third type of fed-batch fermentation process, the glucose concentration is randomly set to vary below 100 g / L. The Raman concentration prediction model will control the feeding based on the prediction results. The predicted cycle time is 15 seconds. Feeding is performed once within one cycle. The feed solution is a 925 g / L glucose aqueous solution with a volume of 600 ml. After all the glucose aqueous solution is added, fermentation continues until the glucose concentration drops below 1 g / L, at which point fermentation is considered to be complete and fermentation is stopped.

[0118] It is worth noting that the total amount of glucose used in the control batch and the fed batch is the same, and all three fed batch fermentation processes are based on online monitoring and intelligent feedback feeding control using a Raman concentration prediction model.

[0119] Raman data analysis:

[0120] Figure 6 The online Raman signals at different time points during batch fermentation are shown, with each spectrum spaced 10 hours apart. Details of the signal acquisition process are provided in Section 2.2. Each spectrum ranges from 179 cm⁻¹ to 2484 cm⁻¹. Figure 6 The Raman signals clearly revealed information about the main components in the fermentation broth. Glucose characteristic peaks were mainly located at 911 cm⁻¹, 1060 cm⁻¹, 1125 cm⁻¹, and 1360 cm⁻¹, with an additional characteristic peak observed at 511 cm⁻¹ in the low wavenumber range. Ethanol showed distinct bands at 877 cm⁻¹, 1046 cm⁻¹, and 1455 cm⁻¹. Due to its low concentration, the byproduct glycerol did not show a clear characteristic peak band directly in the spectrum; however, its concentration information can be indirectly reflected by the content and proportion of other components.

[0121] Raman spectroscopy signal acquisition is susceptible to various factors, such as changes in excitation light intensity, stability noise of the CCD detector, fluorescence background of fluorescent substances in the sample, and scattering effects caused by the characteristics of the sample itself. Therefore, spectral preprocessing is essential. Figure 7 The image shows the spectrum after preprocessing. In this experiment, the preprocessing steps for the Raman signal included using a sliding median filter based on variance judgment to eliminate cosmic spikes, using Savitzky-Golay filtering to eliminate noise (window width of 7, fitting order of 2), using first-order difference to eliminate background fluorescence interference, improving the signal-to-noise ratio of Raman characteristic peaks, and analyzing and extracting the spectral bands containing the main Raman characteristic peaks, ranging from 386 cm⁻¹ to 1735 cm⁻¹.

[0122] Example 3

[0123] Reference Figures 8-13 This is another embodiment of the present invention. Unlike the first embodiment, this embodiment uses a traditional technical solution to compare and test the method of the present invention, and compares the test results with scientific demonstration methods to verify the real effect of the method.

[0124] Algorithm Comparison:

[0125] To achieve precise feed control, an accurate Raman concentration prediction model is crucial. In this experiment, we first constructed a regression Raman concentration prediction model for glucose concentration using only the Partial Least Squares Regression (PLSR) algorithm and a dataset from the batch feed process, to guide the feed control of the fermentation process. During control, the glucose concentration output from the Raman concentration prediction model is processed by a Kalman filter in the time domain to output the final predicted glucose concentration value. When necessary, we also perform dynamic calibration using a biochemical analyzer to ensure the accuracy of the prediction results.

[0126] Reference Figure 8 This paper demonstrates the real-time Raman spectral prediction using the PLSR algorithm Raman concentration prediction model during a 30 g / L glucose concentration control process. The PLSR algorithm successfully captured the dynamic changes in glucose concentration, reflecting the fluctuations in concentration during fermentation. Its root mean square error (RMSE) compared to HPLC results over six batches was 23.58 g / L. While the PLSR Raman concentration prediction model shows some effectiveness in predicting glucose concentration, it has limitations when handling complex or multi-batch data.

[0127] Meanwhile, collecting high-quality and abundant data is crucial for establishing an efficient Raman concentration prediction model. Especially during fermentation, the long cycle and time-consuming standard sampling methods mean that extensive sampling can impact smaller fermentation systems. Therefore, we first considered data augmentation to improve algorithm performance. We adopted a machine learning pseudo-label data augmentation method for individual fermentation processes. Before this, the data was cleaned to remove outliers, and then a separate regression Raman concentration prediction model was built for each completed fermentation process using the PLSR algorithm. This Raman concentration prediction model was used to predict the glucose concentration in all Raman spectra during fermentation, and Kalman filtering was used to reduce prediction variance. This was based on our prior knowledge that the concentration changes of each component in the fermentation broth are gradual over short periods during batch fermentation. Next, we relabeled the spectra using these pseudo-labels, creating three datasets: a batch fermentation process dataset (Batch(Remark)) (2 batches * 10,000 samples), a fed-batch process dataset (Fed-batch(Remark)) (2 batches * 10,000 samples), and a continuous control process dataset (Continuous(Remark)) (6 batches * 10,000 samples). When the datasets were augmented using labels (including HPLC calibration and pseudo-labels), their order of magnitude increased to 100 times that of the original datasets. The introduction of pseudo-labels is particularly important because it provides additional, though not entirely accurate, but valuable labeling information, which helps enhance the training and generalization capabilities of the Raman concentration prediction model. Furthermore, this method reduces reliance on expensive and time-consuming HPLC experiments, thereby improving data generation efficiency while maintaining data quality.

[0128] Subsequently, we improved the glucose concentration prediction and feedback Raman concentration prediction models using the relabeled batch feeding process dataset (Fed-batch(Remark)) and the continuous control process dataset (Continuous(Remark)) combined with a convolutional neural network (CNN) algorithm. Control experiments were conducted at six different control concentrations to evaluate the performance of the improved Raman concentration prediction model.

[0129] After modifying the algorithm, the variance of the Raman concentration prediction model was further reduced, allowing for faster and more accurate control. The control strategy was changed from adjusting the pump rate every 3 minutes to every 15 seconds. As shown in Table 1, the re-labeled batch feeding process dataset (Fed-batch(Remark)) and continuous control process dataset (Continuous(Remark)) were added to the PLSR and SA-CNN algorithms for training. Comparing the prediction results of the two algorithms revealed that the prediction error of the PLSR algorithm was 10.53 g / L, while the prediction error of the SA-CNN algorithm was reduced to 3.41 g / L, demonstrating a significant improvement in prediction performance. Figure 9 As shown, the improved Raman concentration prediction model demonstrated excellent predictive ability during the control of glucose concentration at 20 g / L. The SA-CNN Raman concentration prediction model effectively handled complex data structures and nonlinear relationships, and significantly reduced sampling errors by splicing continuous spectra, enabling it to achieve accurate predictions even at low concentrations. Figure 10 As shown, after achieving accurate prediction of glucose concentration, online prediction of ethanol and glycerol concentrations was performed. It can be seen that the prediction performance is good, with prediction accuracy of over 90%.

[0130] Table 1: Root mean square error of PLSR and SA-CNN algorithms combined with different training datasets on different control processes.

[0131]

[0132] RMSE is calculated in g / L; (only samples within 1 minute before and after the HPLC sampling time are calculated).

[0133] Reference Figure 11 The control batch was a batch fermentation with an initial glucose concentration of 285 g / L. In the experimental batches, the first type of feed batch had an initial glucose concentration of 100 g / L, and the maintenance glucose concentrations were 20 g / L, 30 g / L, 40 g / L, 50 g / L, 65 g / L, and 80 g / L, respectively. The second type of feed batch had an initial glucose concentration of 100 g / L, and when the glucose concentration dropped to about 10 g / L, it was replenished to 80 g / L, for a total of two replenishments. The third type of feed batch had the glucose concentration randomly varied below 100 g / L, so that the glucose concentration during fermentation was in a state of large fluctuation.

[0134] like Figure 11As shown, by comparing the cell growth during the fermentation process, it can be concluded that the cell dry weight of all three fed-batch fermentation processes was higher than that of the control batch. This is because the inhibitory effect of excessively high initial sugar concentration on the Saccharomyces cerevisiae cells was alleviated, resulting in improved specific growth rate and cell yield, ultimately increasing biomass. The highest cell concentrations of the first fed-batch fermentation process (80 g / L, 65 g / L, 50 g / L, 40 g / L, 30 g / L, 20 g / L) were 1.16, 1.26, 1.23, 1.17, 1.27, and 1.26 times that of the control batch, respectively. The highest cell concentrations of the second and third fed-batch fermentation processes were 1.23 and 1.25 times that of the control batch, respectively. It can be seen that the cell growth state was best when the sugar concentration was maintained at 30 g / L, which was 13.12 g / L, 1.27 times that of the control batch.

[0135] By comparing fermentation times, it was found that the sugar supplementation strategy not only increased the cell dry weight concentration during the process but also effectively shortened the fermentation time. The fermentation time for the first fed-batch fermentation process (80 g / L, 65 g / L, 50 g / L, 40 g / L, 30 g / L, 20 g / L) was shortened from 48 h in the control batch to 37 h, 37.5 h, 38 h, 39 h, 39 h, and 40.5 h, respectively. The fermentation times for the second and third fed-batch fermentation processes were shortened from 48 h in the control batch to 37 h and 41 h, respectively. The shortest fermentation time was achieved by maintaining a sugar concentration of 80 g / L and the second fed-batch fermentation process, while the longest fermentation times were achieved by maintaining a sugar concentration of 20 g / L and the third fed-batch fermentation process. This indicates that as the maintenance sugar concentration decreases, the fermentation time increases. This may be because overall cell activity decreases in the later stages of fermentation, and it is more difficult to maintain cell viability and the number of viable cells at excessively low sugar concentrations, thus resulting in a longer fermentation time.

[0136] Reference Figure 12 and Figure 13 Because 600ml of sugar solution was added during the fermentation of the fed batches for dilution, and some ethanol evaporated during fermentation, the ethanol and glycerol concentrations were standardized for easier comparison between different batches. Therefore, the final ethanol concentration was calculated as HPLC-measured ethanol concentration * 1.2 + evaporated ethanol concentration, where 1.2 is the dilution factor (3.6L / 3L = 1.2), and the evaporated ethanol concentration was calculated using Henry's Law. The final glycerol concentration was calculated as HPLC-measured glycerol concentration * dilution factor 1.2.

[0137] like Figure 12As shown, the ethanol yields of the first batch fermentation process (80 g / L, 65 g / L, 50 g / L, 40 g / L, 30 g / L, 20 g / L) were 137.25 g / L, 137.46 g / L, 138.41 g / L, 140.18 g / L, 140.68 g / L, and 140.61 g / L, respectively. Compared with the control batch (125.71 g / L), the ethanol yields increased by 9.18%, 9.35%, 10.1%, 11.51%, 11.91%, and 11.85%, respectively. The ethanol yields of the second and third fed batches were 135.46 g / L and 140.03 g / L, respectively, which increased by 7.76% and 11.39% compared with the control batch. The first fed-batch fermentation showed that maintaining a sugar concentration between 30 g / L and 80 g / L resulted in higher ethanol yields. However, maintaining a concentration of 20 g / L did not significantly increase ethanol yield compared to 30 g / L, although the fermentation time was longer and the ethanol production rate slower. This may be because at lower sugar concentrations (<20 g / L), carbon limitation during the exponential growth phase prevents full utilization of the sugar source, thus affecting cell activity. Comparing the final ethanol yields of the three fed-batch fermentation processes, the highest yield (140.68 g / L) was achieved with a maintenance concentration of 30 g / L. Correspondingly, the ethanol production rate of the first fed-batch fermentation process (80 g / L, 65 g / L, 50 g / L, 40 g / L, 30 g / L, 20 g / L) increased from 2.62 g / L / h in the control batch to 3.71 g / L / h, 3.67 g / L / h, 3.64 g / L / h, 3.59 g / L / h, 3.61 g / L / h, and 3.47 g / L / h, respectively, representing increases of 41.6%, 40.1%, 38.9%, 37%, 37.8%, and 32.4%. The second and third fed-batch fermentation processes increased to 3.66 g / L / h and 3.42 g / L / h, respectively, representing increases of 39.7% and 30.5%. Comparing the ethanol production rates of the three fed-batch fermentation processes, it can be seen that the highest ethanol production rate (3.71 g / L / h) is achieved when the maintenance process concentration is 80 g / L. As the maintenance sugar concentration decreases, the ethanol production rate declines, indicating that the higher the maintenance sugar concentration (20 g / L-80 g / L), the higher the ethanol production rate. The third fed-batch fermentation process simulates the glucose concentration changes in actual industrial production. Maintaining a process concentration of 80 g / L significantly increases the ethanol production rate, by 8.5% compared to the third fed-batch fermentation process. From the perspective of actual industrial production efficiency, maintaining a process concentration of 80 g / L is the best. However, product yield is also extremely important in actual production. The final ethanol yield at a maintenance process concentration of 30 g / L is higher than that of other fed-batch processes. Therefore, from the perspective of product yield, maintaining a process concentration of 30 g / L is the best.

[0138] In addition, the concentration of glycerol, a byproduct of the sugar-supplemented batches, also decreased significantly, such as... Figure 13 As shown, the byproduct glycerol concentrations in the first fed-batch fermentation process (80 g / L, 65 g / L, 50 g / L, 40 g / L, 30 g / L, 20 g / L) were 7.53 g / L, 7.25 g / L, 6.88 g / L, 6.87 g / L, 6.72 g / L, and 6.74 g / L, respectively, representing 73.6%, 70.9%, 67.3%, 67.2%, 65.7%, and 65.9% of the control group's glycerol concentration (10.23 g / L). The byproduct glycerol concentrations in the second and third fed-batch fermentations were 7.2 g / L and 6.97 g / L, respectively, representing 70.4% and 68.1% of the control group's (10.23 g / L). This shows that in the first type of fed-batch fermentation, the higher the glucose concentration maintained during the process, the higher the concentration of the byproduct glycerol. However, starting from 20 g / L, it was not lower than 30 g / L, indicating that fermentation conditions with excessively low sugar concentrations are not conducive to the cell's utilization of carbon sources to produce ethanol. Therefore, the ethanol-glycerol ratios for the first fed-batch fermentation process (80 g / L, 65 g / L, 50 g / L, 40 g / L, 30 g / L, 20 g / L) were 18.24, 18.96, 20.11, 20.4, 20.93, and 20.87, respectively, representing increases of 48.4%, 54.3%, 63.6%, 65.9%, 70.3%, and 69.8% compared to the control batch (12.29). The ethanol-glycerol ratios for the second and third fed-batch fermentations were 18.8 and 20.09, respectively, representing increases of 52.9% and 63.47% compared to the control batch (12.29). Based on the results of the ethanol-glycerol ratio in the three fed-batch fermentation processes, the first fed-batch fermentation process maintained the highest ethanol-glycerol ratio of 30 g / L, which means that maintaining a low glucose concentration during fermentation helps to reduce the concentration of glycerol by-products. In other words, a low glucose concentration during the production period helps more carbon source flow to ethanol.

[0139] Compared with commonly used traditional machine learning methods (such as partial least squares regression, PLSR), this invention employs an optimized deep learning algorithm to predict glucose, glycerol, and ethanol concentrations during the ethanol fermentation process of *Saccharomyces cerevisiae* in a bioreactor, exhibiting better accuracy and generalization. Furthermore, the experiment involves implementing intelligent feedback feeding control by predicting glucose, glycerol, and ethanol concentrations in the fermenter in real time, and comparing and analyzing process parameters under different feeding modes to find the optimal strategy for controlling residual sugar concentration and explore the reasons for these parameter changes. The application of this method not only improves prediction accuracy but also demonstrates good results in practical control tasks, which is of great significance for optimizing fermentation process control.

[0140] It should be recognized that embodiments of the present invention can be implemented or carried out by a combination of computer hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium. The method can be implemented using standard programming techniques—including a non-transitory computer-readable storage medium configured with a computer program, wherein such a storage medium causes the computer to operate in a specific and predefined manner—according to the methods and drawings described in the specific embodiments. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if desired, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit (ASIC).

[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent feedback feeding control in the ethanol fermentation process of brewing yeast, characterized in that, include: The spectrum is sampled to obtain the sampled spectrum; The sampled spectrum is input into a two-dimensional convolutional neural network based on spectral amplification; The mean squared error is used as the loss function, and the Adam optimizer is used to update the network weights when training the Raman concentration prediction model to obtain the optimal solution and update the Raman concentration prediction model. Collect the current spectral data in the fermenter, input the current spectral data into the updated Raman concentration prediction model, output the glucose concentration value, and replenish the high-concentration glucose aqueous solution according to the glucose concentration value; The method employs mean squared error as the loss function and uses the Adam optimizer to update network weights during Raman concentration prediction model training. The optimal solution for the Raman concentration prediction model includes: Calculate the gradient of the current parameters Its mathematical expression is: ； in, Indicates at time step The loss function is used to update the estimates of the first and second moments, and its mathematical expression is: ； ； in, , The exponential decay rate is estimated for the first and second moments; The mathematical expressions for estimating the first and second moments after bias correction are as follows: ； The parameters are adjusted using the bias-corrected estimates of the first and second moments to more accurately estimate the descent direction of the parameters. The mathematical expression for this is: ； in, Indicates the current iteration number. This represents the parameters of the Raman concentration prediction model. Indicates the learning rate. This represents an estimate of the first-order moment. This represents the estimate of the second moment.

2. The intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast as described in claim 1, characterized in that: The spectrum is continuously sampled to obtain multiple spectra.

3. The intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast as described in claim 2, characterized in that: The multiple spectra are spliced ​​together to form a sampling spectrum.

4. The intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast as described in claim 3, characterized in that: Two-dimensional convolutional neural networks based on spectral amplification include convolutional layers, batch normalization (BN) layers, and fully connected layers.

5. The intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast as described in claim 4, characterized in that: The convolutional layer includes a first convolutional layer, a second convolutional layer, and a third convolutional layer; The first convolutional layer has a 5×5 kernel, 8 kernels, and an input of 6×5 × 11 × 8. The second convolutional layer has a 3×3 kernel, 16 kernels, and an input of 4×509×16. The third convolutional layer has a 2×2 kernel, a total of 32 kernels, and an input of 3×508×32.

6. The intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast as described in claim 5, characterized in that: The first, second, and third convolutional layers are activated using the ReLU activation function.

7. The intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast as described in claim 6, characterized in that: The BN layer includes a first BN layer, a second BN layer, and a third BN layer; The first convolutional layer is connected to the first batch normalization layer; The second convolutional layer is connected to the second batch normalization layer; The third convolutional layer is connected to the third batch normalization (BN) layer.

8. The intelligent feedback feeding control method for the ethanol fermentation process of brewing yeast as described in claim 7, characterized in that: Collect the current spectral data in the fermenter, input the current spectral data into the updated Raman concentration prediction model, and output the glucose concentration value. If the glucose concentration value is lower than the preset glucose concentration threshold, add high-concentration glucose aqueous solution until the glucose concentration value reaches the glucose concentration threshold.

Images