Method and system for predicting enzyme concentration in microbial fermentation process

Abstract

The present invention provides a method and system for predicting enzyme concentrations in a microbial fermentation process. The method comprises: obtaining online data for historical tank batches and a current tank batch to be predicted, as well as offline data for the tank batch corresponding to the online data; performing statistical analysis on the offline data and designing classification criteria to classify the historical tank batches; establishing a hybrid prediction training database for the online and offline data according to the classified tank batch categories; training each tank batch data category using an artificial neural network prediction model and an extreme gradient boosting prediction model based on the hybrid prediction training database; performing nonlinear fusion of the trained artificial neural network prediction model and extreme gradient boosting prediction model based on the extreme gradient boosting algorithm to achieve hybrid prediction; and performing a rolling update on the hybrid prediction training database each time a tank batch fermentation ends. The present invention is used to predict enzyme concentrations in a microbial fermentation process in advance, achieving higher-precision predictions.

Description

Technical Field

[0001] The present invention relates to the field of biotechnology, in particular to a method and system for predicting the concentration of biological enzymes in a microbial fermentation process. Background Art

[0002] In actual microbial fermentation production processes, accurate and timely prediction of key state variables (such as cell concentration, substrate concentration, and enzyme concentration) is crucial. Some fermentation state variables are difficult to measure online, requiring sampling of the fermentation system and offline analysis to obtain relevant data. However, sampling poses a risk of contamination to the fermentation system, and the time lag associated with offline sampling and analysis hinders real-time production operations during the fermentation process. While relevant sensors are available, these are often very expensive and have poor reliability, making them unsuitable for large-scale industrial production processes. Therefore, prediction techniques for microbial fermentation process state variables have attracted considerable research interest. Computer technology can generate a large amount of discrete measurement data from fermentation processes. Data-driven process modeling can provide state prediction and fault diagnosis capabilities for fermentation processes, making it a hot topic in fermentation process research.

[0003] Chinese patent application number 201310661816.4, filed on December 9, 2013, describes a method for online prediction of biofermentation yield based on a Bayesian combination neural network. This patent first categorizes all historical batches into superior, medium, and inferior batches. A neural network predictor is established for each type of fermentation batch. The three neural networks are then trained separately. The output values ​​of the three neural networks are calculated based on the latest input vector of the batch to be predicted. The yield forecast for the fermentation process is derived by weighting the output values ​​of the three neural networks. This method does not classify the batches to be predicted, and the output values ​​of the three trained neural networks are not always applicable to that batch. For example, when the batch to be predicted is superior, the neural network predictor corresponding to the superior batch can provide a good forecast of the batch's yield. However, the neural network predictors corresponding to the medium and inferior batches lack superior features in the training data, resulting in less reliable forecasts of the batch's yield. Summary of the Invention

[0004] In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for predicting the concentration of biological enzymes in a microbial fermentation process.

[0005] According to one aspect of the present invention, a method for predicting the concentration of an enzyme in a microbial fermentation process is provided, the method comprising:

[0006] Obtaining online data of historical tank batches and current tank batches to be forecasted, as well as offline data of tank batches corresponding to the online data;

[0007] Performing statistical analysis on the offline data and designing classification criteria to classify historical tank batches;

[0008] For the online data and the offline data, respectively, a hybrid forecast training database is established according to the classified tank batch categories;

[0009] Based on the hybrid forecast training database, each type of tank batch data is trained using the artificial neural network forecast model and the extreme gradient boosting forecast model respectively;

[0010] Based on the extreme gradient boosting algorithm, the trained artificial neural network forecast model and the extreme gradient boosting forecast model are nonlinearly fused to achieve hybrid forecasting;

[0011] Whenever a tank batch fermentation is completed, the mixed prediction training database is rolled over and updated.

[0012] Furthermore, the online data of historical tank batches and the current tank batch to be predicted, as well as the offline data of the tank batch corresponding to the online data, are obtained, wherein: the online data includes fermentation liquid temperature T, fermentation liquid pH, fermentation liquid volume V, fermentation tank pressure P, ventilation volume flow F, dissolved oxygen DO and stirring blade speed ω; the offline data includes offline recorded biological enzyme concentration.

[0013] Furthermore, the statistical analysis is performed on the offline data and a classification standard is designed, wherein: according to the classification function J c,i (t) values ​​are classified and the classification function J is calculated c,i The formula for the value of (t) is:

[0014]

[0015] Where, J c,i (t) represents the classification function value of the i-th tank batch at time t, T W Indicates the window width that needs to be considered in order to achieve classification, J i (t) represents the enzyme concentration of the i-th tank batch at time t.

[0016] Furthermore, the classification of historical tank batches includes: defining J c,ave (t) is the mean of the classification function at time t, σ(t) is the standard deviation of the classification function at time t, α is the confidence coefficient, and the classification standard is designed as follows:

[0017] J c,i (t)<J c,ave (t)-α·σ(t), for low-yield tank batches;

[0018] J c,ave (t)-α·σ(t)≤J c,i(t)≤J c,ave (t)+α·σ(t), is the average tank batch;

[0019] J c,i (t)>J c,ave (t)+α·σ(t), for high-yield tank batches.

[0020] Furthermore, a hybrid forecast training database is established for the online data and the offline data respectively, wherein: the data in the hybrid forecast training database is evenly distributed, and the data sets are concentrated from production data of the same period of time.

[0021] Furthermore, each type of tank batch data is trained using an artificial neural network prediction model and an extreme gradient boosting prediction model, respectively, wherein: the artificial neural network prediction model is based on the collected fermentation production process state data, establishes a feature model between input and output data, and simulates the actual fermentation process.

[0022] Furthermore, each type of tank batch data is trained using an artificial neural network prediction model and an extreme gradient boosting prediction model, respectively, wherein: the input variables of the extreme gradient boosting prediction model include detectable variables affecting the concentration of the biological enzyme, and the output variables include the advance predicted biological enzyme concentration. During the training process of the extreme gradient boosting prediction model, new classification trees are continuously added to the current model to improve the model prediction accuracy.

[0023] Furthermore, the nonlinear fusion of the trained artificial neural network prediction model and the extreme gradient boosting prediction model based on the extreme gradient boosting algorithm includes:

[0024] The prediction result of a single model is used as input, that is, x f =[U A (t u +c),U X (t u +c)], where U A (t u +c) represents the prediction results of the artificial neural network prediction model, U X (t u +c) represents the prediction result of the extreme gradient boosting prediction model, c represents the lead prediction time; the actual enzyme concentration is used as the output, that is, y f =[U(t u +c)], using the self-learning ability of the model to optimize the weight of a single model in the hybrid forecasting model.

[0025] Furthermore, each time a tank batch fermentation is completed, the mixed forecast training database is rolled updated, including:

[0026] If the tank batch is an abnormal tank batch, the tank batch selected by the hybrid forecast training database remains unchanged;

[0027] If the enzyme concentration curve of the late batch is similar to that of an earlier batch within the confidence region, the earlier batch is replaced by the later batch;

[0028] If the enzyme concentration curve in the middle and late stages of a tank batch is not similar to that in the confidence region of any historical tank batch, it is directly added to the hybrid prediction training database.

[0029] According to another aspect of the present invention, a system for predicting the concentration of an enzyme in a microbial fermentation process is provided, the system comprising:

[0030] Acquisition module: acquires online data of historical tank batches and current tank batches to be forecasted, as well as offline data of tank batches corresponding to the online data;

[0031] Classification module: performing statistical analysis on the offline data and designing classification standards to classify historical tank batches;

[0032] Database establishment module: establishing a hybrid forecast training database for the online data and the offline data according to the classified tank batch categories;

[0033] Training and testing module: Based on the hybrid forecast training database, each type of tank batch data is trained using an artificial neural network forecast model and an extreme gradient boosting forecast model;

[0034] Hybrid forecast module: Based on the extreme gradient boosting algorithm, the trained artificial neural network forecast model and the extreme gradient boosting forecast model are nonlinearly fused to achieve hybrid forecasting;

[0035] Rolling update module: Whenever a tank batch fermentation is completed, the mixed prediction training database is rolled updated.

[0036] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0037] 1. The present invention first determines the category of the batch to be predicted and then performs model prediction using artificial neural network and extreme gradient boosting respectively. Then, a hybrid prediction of the two models is performed based on the extreme gradient boosting algorithm. Thus, the weight of a single model in the hybrid prediction model can be optimized by the self-learning ability of the model, which can achieve a higher-precision prediction of the enzyme concentration.

[0038] 2. The present invention does not require additional measurement points or other equipment, only the addition of a software computing module to the existing control system, resulting in low implementation costs. The present invention can be applied to fermentation production sites and has great potential in guiding online monitoring and scheduling optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0040] Figure 1 Schematic diagram of a method for predicting the concentration of biological enzymes in a microbial fermentation process according to one embodiment of the present invention;

[0041] Figure 2 This is the classification result of a batch of 272 cans in one embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the principle of an artificial neural network predictor in one embodiment of the present invention;

[0043] Figure 4 Schematic diagram of the training process of the extreme gradient boosting prediction model in one embodiment of the present invention;

[0044] Figure 5 is the average relative error distribution of 232 average enzyme activity tank batch advance mixing prediction 1 hour in one embodiment of the present invention;

[0045] Figure 6 The average relative error distribution of the 232 average enzyme activity tank batch advance mixing prediction 2h in one embodiment of the present invention;

[0046] Figure 7 3h average relative error distribution of 232 average enzyme activity tank batch advance mixing predictions in one embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several variations and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0048] The present invention provides a method for predicting the concentration of enzymes in a microbial fermentation process. Figure 1 , the method comprising:

[0049] Step 1: Obtain online data for historical tank batches and the current tank batch to be forecasted, as well as offline data for the tank batches corresponding to the online data. Historical tank batches are tank batches that have completed fermentation and have complete data; the current tank batch to be forecasted is a tank batch that is currently fermenting; and the tank batches corresponding to the online data are the tank batches for which data is already available for the historical tank batches and the current tank batch to be forecasted.

[0050] Step 2: Perform statistical analysis based on offline data and design classification criteria to classify historical tank batches;

[0051] Step 3: For online data and offline data, establish a hybrid prediction training database according to the classified tank batch categories;

[0052] Step 4: Based on the hybrid forecast training database, each type of tank batch data is trained using the artificial neural network forecast model and the extreme gradient boosting forecast model respectively;

[0053] Step 5: Based on the extreme gradient boosting algorithm, the trained artificial neural network prediction model and the extreme gradient boosting prediction model are nonlinearly fused to obtain a hybrid prediction model, and the hybrid prediction is realized using the model;

[0054] Step 6: Whenever a batch fermentation is completed, the mixed prediction training database is updated in a rolling manner.

[0055] In some embodiments, in step 1, historical tank batch online data of a given tank number and batch and known online data of a current tank batch to be predicted are read from a factory distributed control system database, and offline data of the tank batch corresponding to the read online data are collected, wherein: the online data includes fermentation liquid temperature T, fermentation liquid pH, fermentation liquid volume V, fermentation tank pressure P, ventilation volume flow F, dissolved oxygen DO, and stirring blade speed ω; the offline data includes offline recorded biological enzyme concentration.

[0056] In some embodiments, in step 2, statistical analysis is performed based on the offline data and a classification standard is designed, wherein: according to the classification function J c,i (t) values ​​are classified and the classification function J is calculated c,i The formula for the value of (t) is:

[0057]

[0058] Where, J c,i (t) represents the classification function value of the i-th tank batch at time t, T W Indicates the window width that needs to be considered in order to achieve classification, J i (t) represents the enzyme concentration of the i-th tank batch at time t.

[0059] Further, the tank batch is classified, including: defining J c,ave (t) is the mean of the classification function at time t, σ(t) is the standard deviation of the classification function at time t, α is the confidence coefficient, and the classification standard is designed as: J c,i (t)<J c,ave (t)-α·σ(t), for low-yield tank batches; J c,ave (t)-α·σ(t)≤J c,i (t)≤J c,ave(t)+α·σ(t), is the average tank batch; J c,i (t)>J c,ave (t)+α·σ(t), is a high-yield batch. Therefore, the batches are divided into three categories.

[0060] In some embodiments, in step three, the collected online and offline data are divided into three categories according to a designed classification standard. Training databases are then established for each of the classified tank batches. The hybrid forecast training database includes both online and offline data, with the online data serving as input for model training and the offline data serving as output. The establishment of the hybrid forecast training database should adhere to the following principles: the database size should be moderate, the data in the database should be evenly distributed, the data should represent the operating conditions of a normal tank batch, and the data should be concentrated on production data from a consistent period, such as one to three months, to minimize gradual variations.

[0061] In some embodiments, in step 4, the training and testing of the artificial neural network prediction model is based on an artificial neural network algorithm. The essence of the artificial neural network algorithm is to simulate nonlinear systems using artificial neurons, which theoretically has the ability to approximate arbitrary functions and their derivatives of various orders. Using the artificial neural network prediction model to predict fermentation process state variables does not require a precise physical mathematical model of the fermentation process, that is, it does not require prior knowledge of the process. Therefore, for each type of tank batch data, an artificial neural network prediction model and an extreme gradient boosting prediction model are trained separately. In particular, the artificial neural network prediction model uses the collected fermentation production process state data as a basis to establish a characteristic model between input and output data, simulate the actual fermentation process, and guide production operations. When modeling the fermentation process and predicting state variables online, some data that can be detected online in real time and offline analysis data are used to predict the state variable values ​​at a certain point in the future. In this prediction process, the artificial neural network prediction model acts as a model that approximates the actual physical fermentation process, and parameters such as the internal neuron connection weights are the specific expression of the model. The input variables of the neural network state variable predictor generally include some production variables that are easy to detect online in real time, such as fermentation liquid pH, temperature, dissolved oxygen, oxygen concentration and carbon dioxide concentration in exhaust gas, etc. The output variables can be some biological variables that are difficult to detect online in real time, such as bacterial concentration, enzyme concentration and substrate concentration.

[0062] In some embodiments, in step 4, the training and testing of the extreme gradient boosting prediction model are based on the extreme gradient boosting algorithm. The extreme gradient boosting algorithm belongs to the Boosting algorithm and is an efficient implementation of the gradient boosting decision tree. It has the advantages of accuracy and good interpretability in prediction problems and is suitable for efficient parallel computing. For each type of tank batch data, an artificial neural network prediction model and an extreme gradient boosting prediction model are used for training respectively, wherein: the input variables of the extreme gradient boosting prediction model include detectable variables that affect the concentration of the bio-enzyme, and the output variables include the advance predicted bio-enzyme concentration. During the training process of the extreme gradient boosting prediction model, new classification trees are continuously added to the current model to improve the model prediction accuracy.

[0063] Specifically, select online detection data as input data set A input , and collect the corresponding offline test data as the output data set A output Input dataset A input It mainly includes detectable variables that affect product synthesis, such as fermentation liquid temperature T, fermentation liquid pH, fermentation liquid volume V, fermentation tank pressure P, ventilation volume flow F, dissolved oxygen DO and stirring blade speed ω, etc. The output data set A output Mainly refers to the offline recorded biological enzyme concentration.

[0064] Let the training set input data be x j (j=1,2,...,N), the training set output data is y j (j=1,2,...,N), the prediction result obtained using the extreme gradient boosting algorithm is have:

[0065]

[0066] Where k is the number of the tree, M is the total number of trees in the algorithm, and f k (x j ) is x j The weight of the kth tree.

[0067] The objective function of the extreme gradient boosting forecasting model can be expressed as:

[0068]

[0069] Where N is the number of samples, is the training error function of the jth sample, Ω(f k ) is the regularization function of the kth tree. To minimize the objective function, both the error function and the regularization function should be small.

[0070] During the training process of the extreme gradient boosting prediction model, new classification trees are continuously added to the current model to improve the prediction accuracy of the model. In the case of a prediction model consisting of m-1 trees, when the mth tree is added to the model, the following is obtained:

[0071]

[0072] Where, For the first m-1 trees, the sample x j The prediction result, f m (x j ) is x j The weight of the mth tree.

[0073] Using Taylor's second-order expansion to optimize the objective function Obj, we can get:

[0074]

[0075] Where p j ,q j are the first-order and second-order partial derivatives of the training error function, respectively.

[0076] Expand Ω(f m ), define φ l is the data set belonging to the decision tree node l, δ l is the weight of leaf node l. For a tree with a fixed structure, the value of the objective function is only related to δ l So δ l By taking the derivative, we can get the optimal weight of each leaf node and the optimal value of the objective function:

[0077]

[0078]

[0079] Where, T L is the total number of leaf nodes, and λ and γ are weighting factors.

[0080] After the decision tree structure is determined, the weight of each leaf node can be determined. The results of new additions to the tree are obtained by the exact greedy algorithm, that is, each split can minimize the objective function until the objective function has no room for reduction or reaches a pre-set maximum depth.

[0081] The training and testing of the artificial neural network prediction model and the extreme gradient boosting prediction model are independent of each other. In step 5, in order to improve the prediction accuracy, considering the complex nonlinear relationship between the two models and the actual enzyme concentration, a hybrid prediction of the two models is performed based on the extreme gradient boosting algorithm. Specifically, based on the extreme gradient boosting algorithm, the trained artificial neural network prediction model and the extreme gradient boosting prediction model are nonlinearly fused, including: taking the prediction result of a single model as input, that is, x f =[U A (t u +c),U X (t u +c)], where U A (t u +c) represents the prediction results of the artificial neural network prediction model, U X (t u +c) represents the prediction result of the extreme gradient boosting prediction model, c represents the lead prediction time; the actual enzyme concentration is used as the output, that is, y f =[U(t u +c)], the self-learning ability of the model is used to optimize the weight of a single model in the hybrid prediction model, thereby achieving a more accurate prediction of the concentration of biological enzymes.

[0082] In some embodiments, in step 6, since fermentation tank-related detection equipment may wear out to a certain extent over time, and climate change may also have a certain impact on fermentation production, the training database of the forecast model needs to be rolled over. Specifically, each time a batch of fermentation is completed, the mixed forecast training database is rolled over and updated, including:

[0083] If the tank batch is an abnormal tank batch, that is, a tank batch contaminated with bacteria or a tank batch with a bio-enzyme concentration significantly lower than the lowest bio-enzyme concentration curve in the historical tank batches, the tank batch cannot be added to the database composed of tank batches under normal working conditions, but should be discarded. The tank batch selected in the mixed prediction training database will remain the same;

[0084] If the enzyme concentration curve in the middle and late stages of a batch is similar to that of an earlier batch within the confidence region, the middle and late stages of the batch refer to the middle and late stages of the entire fermentation process of the batch, and the earlier batch refers to the historical batch in the training database, indicating that the data of the two batches have similar effects on model training and can therefore be substituted for each other. The newer data can be selected and used to replace the earlier batch.

[0085] If the enzyme concentration curve in the middle and late stages of a tank batch is not similar to that in the confidence region of any historical tank batch, but the tank batch is not an abnormal tank batch that needs to be discarded and can be used to enrich the database diversity, it will be directly added to the hybrid prediction training database.

[0086] In the above embodiment, the data in the database is made consistent with the operating conditions of the last 1 to 3 months through rolling updates, which can minimize the impact of device losses and climate change and improve the accuracy of model forecasting.

[0087] The present invention will be described in more detail below with reference to an example of prediction of enzyme activity of products produced by industrial fermentation of xylanase.

[0088] Xylanase is mainly used in brewing and feed industries. It can decompose the cell walls and β-glucan of raw materials in brewing or feed industries, reduce the viscosity of materials in brewing, promote the release of effective substances, reduce non-starch polysaccharides in feed grains, and promote the absorption and utilization of nutrients. Xylanase is a secondary metabolite of Streptomyces and can be obtained by intermittent fermentation of Streptomyces. The flow diagram of the method for predicting the concentration of biological enzymes in the microbial fermentation process in this embodiment is shown in FIG. Figure 1 , specifically including the following steps:

[0089] S1. Read the historical online data of the tank batch with a given tank number and batch and the known online data of the tank batch to be predicted from the factory distributed control system database, and collect the offline data of the tank batch corresponding to the read online data. The online data mainly includes the fermentation liquid temperature T, the fermentation liquid pH, the fermentation liquid volume V, the fermentation tank pressure P, the ventilation volume flow F, the dissolved oxygen DO and the stirring blade speed ω, etc. The offline data refers to the product enzyme activity recorded offline (in a specific embodiment, the biological enzyme concentration is characterized by the product enzyme activity). As the fermentation production proceeds, the above online data will be adjusted accordingly. For example, for 80m 3 In the middle and late stages of the xylanase fermentation tank, the temperature is generally controlled at around 34°C and the rotation speed is maintained at 200 rpm.

[0090] S2, perform statistical analysis based on the offline data obtained in S1, and calculate the classification function J using the following formula c,i The value of (t):

[0091]

[0092] Where, J c,i (t) represents the classification function value of the i-th tank batch at time t, T W Indicates the window width that needs to be considered in order to achieve classification, J i (t) represents the enzyme activity of the product of the i-th tank batch at time t.

[0093] Definition J c,ave (t) is the mean of the classification function at time t, σ(t) is the standard deviation of the classification function at time t, α is the confidence coefficient, and the classification standard design is shown in Table 1:

[0094] Table 1 Tank batch classification standards

[0095] condition category <![CDATA[J c,i (t)<J c,ave (t)-a·s(t)]]> Low enzyme activity tank batch <![CDATA[J c,ave (t)-α·σ(t)≤J c,i (t)≤J c,ave (t)+α·σ(t)]]> Average enzyme activity tank batch <![CDATA[J c,i (t)>J c,ave (t)+α·σ(t)]]> High enzyme activity tank batch

[0096] α is usually between 1.04 and 1.65, corresponding to 85% to 95% confidence limits, and here it is 1.28, corresponding to 90% confidence limit. For example, the classification results of 272 batches of xylanase production plants are shown in Figure 2 The values ​​of fermentation culture time and enzyme activity in the figure are normalized.

[0097] S3. Divide the collected online and offline data into three categories according to the classification criteria designed in S2, and establish training databases for each category. Furthermore, the establishment of the training database should adhere to the following principles: (a) The database size should be moderate; (b) The database data should be evenly distributed; (c) The database data should represent the operating conditions of a normal tank batch; and (d) The database data should be concentrated on production data from the same period (1 to 3 months), excluding any gradual changes as much as possible. Specifically, the training database contains online and offline data from 272 tank batches from a xylanase production plant over a 3-month period.

[0098] S4. Use artificial neural network algorithm to train and test the model for each type of tank batch data.

[0099] The principle diagram of artificial neural network predictor can be found in Figure 3 When modeling the fermentation process and predicting state variables online, some online real-time detectable data and offline analysis data are used to predict the state variable values ​​at a certain moment in the future. In this prediction process, the artificial neural network prediction model acts as a model that approximates the real physical fermentation process. The internal neuron connection weights and other parameters are the specific expression of the model. The input variables of the neural network state variable predictor include detectable variables that have been tested to affect the enzyme activity of the product, specifically the current moment t u , current enzyme activity U(t u ), the discrete fermentation liquid pH (t) from the current moment to the previous 8 hours, the discrete dissolved oxygen DO (t) from the current moment to the previous 8 hours, that is, the model input is x u =[t u ,U(t u ),pH(t),DO(t)] T ,t=t u -8,t u -7,...,t u The output variable specifically refers to the product enzyme activity of 1h, 2h and 3h ahead, that is, the model output is y u =[U(t u +1),U(t u +2),U(t u +3)] TTaking the batch of 232 average enzyme activity tanks as an example, traversal training and testing were carried out, and the data of batch 231 were selected as the training set and the data of batch 1 were selected as the test set. After completing parameter tuning, a trained artificial neural network prediction model was obtained.

[0100] S5. Use the extreme gradient boosting algorithm to train and test the model for each type of tank batch data.

[0101] Input variables include detectable variables that have been tested to affect the enzyme activity of the product, specifically the current time t u , current enzyme activity U(t u ), the discrete fermentation liquid pH (t) from the current moment to the previous 8 hours, the discrete dissolved oxygen DO (t) from the current moment to the previous 8 hours, that is, the model input is x u =[t u ,U(t u ),pH(t),DO(t)] T ,t=t u -8,t u -7,...,t u The output variable specifically refers to the product enzyme activity of 1h, 2h and 3h ahead, that is, the model output is y u =[U(t u +1),U(t u +2),U(t u +3)] T .

[0102] The training process of the extreme gradient boosting forecast model is shown in Figure 4 , remember the training set input data as x j (j=1,2,...,N), the training set output data is y j (j=1,2,...,N), the prediction result obtained using the extreme gradient boosting algorithm is have:

[0103]

[0104] Where k is the number of the tree, M is the total number of trees in the algorithm, and f k (x j ) is x j The weight of the kth tree.

[0105] The objective function of the extreme gradient boosting forecasting model can be expressed as:

[0106]

[0107] Where N is the number of samples, is the training error function of the jth sample, Ω(f k) is the regularization function of the kth tree. To minimize the objective function, both the error function and the regularization function should be small.

[0108] During the training process of the extreme gradient boosting prediction model, new classification trees are continuously added to the current model to improve the prediction accuracy of the model. In the case of a prediction model consisting of m-1 trees, when the mth tree is added to the model, the following is obtained:

[0109]

[0110] Where, For the first m-1 trees, the sample x j The prediction result, f m (x j ) is x j The weight of the mth tree.

[0111] Using Taylor's second-order expansion to optimize the objective function Obj, we can get:

[0112]

[0113] Where p j ,q j are the first-order and second-order partial derivatives of the training error function, respectively.

[0114] Expand Ω(f m ), define φ l is the data set belonging to the decision tree node l, δ l is the weight of leaf node l. For a tree with a fixed structure, the value of the objective function is only related to δ l So δ l By taking the derivative, we can get the optimal weight of each leaf node and the optimal value of the objective function:

[0115]

[0116]

[0117] Where, T L is the total number of leaf nodes, and λ and γ are weighting factors.

[0118] Once the decision tree structure is determined, the weights for each leaf node can be determined. Newly added tree nodes are obtained using an exact greedy algorithm, where each split minimizes the objective function until no further reduction is achieved or the pre-set maximum depth is reached. Using 232 batches of average enzyme activity as an example, training and testing were performed, selecting data from batches 231 as the training set and data from batch 1 as the test set. Parameters were then optimized to obtain a trained extreme gradient boosting prediction model.

[0119] S6. In order to improve the prediction accuracy, considering the complex nonlinear relationship between the two models in S4 and S5 and the actual product enzyme activity, a hybrid prediction of the two models is performed based on the extreme gradient boosting algorithm.

[0120] The prediction result of a single model is used as input, that is, x f =[U A (t u +c),U X (t u +c)], where U A (t u +c) represents the enzyme activity prediction result of the artificial neural network prediction model, U X (t u +c) represents the enzyme activity prediction result of the extreme gradient boosting prediction model, c represents the lead prediction time, c = 1, 2, 3. The actual product enzyme activity is output, that is, y f =[U(t u +c)], leveraging the model's self-learning capabilities to optimize the weight of a single model within a hybrid prediction model, achieving more accurate predictions of product enzyme activity. Using 232 batches of average enzyme activity as an example, a traversal training and testing process was performed, selecting data from batches 231 as the training set and data from batch 1 as the test set. After parameter tuning, a trained hybrid prediction model based on extreme gradient boosting (i.e., an enzyme activity hybrid prediction model) was obtained.

[0121] In order to evaluate and verify the forecast effect, the relative error is defined as:

[0122]

[0123] Where U p (r) represents the predicted value of enzyme activity of the product at the rth sampling moment of a single tank batch, U m (r) represents the measured value of product enzyme activity at the rth sampling moment of a single tank batch.

[0124] In order to estimate the accuracy of the model, the mean relative error is introduced:

[0125]

[0126] Where R represents the number of enzyme activity sampling points for a single tank batch. The average relative error distribution of the average enzyme activity tank batches for 1 hour, 2 hours, and 3 hours of advance mixing is shown in Figure 5-Figure 7 ,After calculation, the average values ​​of e for 232 batches of 1h, 2h, and 3h advance ,forecasts are 1.12%, 1.71%, and 2.13%, respectively, indicating that ,the hybrid forecast model has high accuracy and robustness.

[0127] S7. Because fermentation tank-related detection equipment will wear out to a certain extent over time, and climate change will also have a certain impact on fermentation production, the training database of the hybrid forecasting model needs to be rolled over. Each time a batch of fermentation ends, the training database is rolled over as follows: (a) If the batch is an abnormal batch, the batch selected in the database remains unchanged; (b) If the product enzyme activity curve of the batch in the middle and late stages is similar to that of an earlier batch within the confidence region, the earlier batch is replaced with the batch; (c) If the product enzyme activity curve of the batch in the middle and late stages is not similar to that of any historical batch within the confidence region, the batch is directly added to the training database.

[0128] In the above embodiment of the present invention, the category of the batch to be predicted is first determined, and then the artificial neural network and extreme gradient lifting model predictions are performed separately. Then, a hybrid prediction of the two models is performed based on the extreme gradient lifting algorithm. Thus, the weight of the single model in the hybrid prediction model can be optimized by the self-learning ability of the model, and a more accurate product concentration prediction can be achieved. The verification results show that the embodiment of the present invention does not require the addition of additional measuring points or other equipment, and only requires the addition of a software calculation module to the existing control system, which has a low implementation cost. The embodiment of the present invention can be applied to fermentation production sites and has good application potential in guiding online monitoring and scheduling optimization.

[0129] Based on the same inventive concept as above, another embodiment of the present invention provides a system for predicting the concentration of an enzyme in a microbial fermentation process, the system comprising:

[0130] Acquisition module: acquires online data of historical tank batches and current tank batches to be forecasted, as well as offline data of tank batches corresponding to the online data;

[0131] Classification module: Statistical analysis is performed on offline data and classification criteria are designed to classify historical tank batches;

[0132] Database establishment module: According to the classified tank batch categories, a hybrid forecast training database is established for online data and offline data respectively;

[0133] Training and testing module: Based on the hybrid forecast training database, each type of tank batch data is trained using the artificial neural network forecast model and the extreme gradient boosting forecast model;

[0134] Hybrid forecast module: Based on the extreme gradient boosting algorithm, the artificial neural network forecast model and the extreme gradient boosting forecast model are nonlinearly integrated to achieve hybrid forecasting;

[0135] Rolling update module: Whenever a tank batch fermentation is completed, the mixed prediction training database is rolled updated.

[0136] The technologies implemented by each module in the embodiment of the system for predicting the concentration of biological enzymes in the microbial fermentation process of the present invention can adopt the technical features corresponding to each step in the method for predicting the concentration of biological enzymes in the microbial fermentation process, which will not be repeated here.

[0137] Based on the same inventive concept, another embodiment of the present invention also provides an electronic terminal, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it is used to execute the method for predicting the concentration of biological enzymes in the microbial fermentation process in the above embodiment.

[0138] In the above embodiment, the memory is used to store programs; the memory may include volatile memory (English: volatile memory), such as random-access memory (English: random-access memory, abbreviated: RAM), such as static random-access memory (English: static random-access memory, abbreviated: SRAM), double data rate synchronous dynamic random access memory (English: Double Data Rate Synchronous Dynamic Random Access Memory, abbreviated: DDR SDRAM), etc.; the memory may also include non-volatile memory (English: non-volatile memory), such as flash memory (English: flash memory). The memory is used to store computer programs (such as applications, functional modules, etc. that implement the above method), computer instructions, etc., and the above computer programs, computer instructions, etc. can be partitioned and stored in one or more memories. And the above computer programs, computer instructions, data, etc. can be called by the processor.

[0139] The aforementioned computer programs, computer instructions, etc. may be partitioned and stored in one or more memories, and the aforementioned computer programs, computer instructions, data, etc. may be called by a processor.

[0140] The processor is configured to execute the computer program stored in the memory to implement the various steps of the method involved in the above embodiment. For details, please refer to the relevant description in the above method embodiment.

[0141] The processor and memory can be independent structures or integrated structures. When the processor and memory are independent structures, the memory and processor can be coupled via a bus.

[0142] Based on the same inventive concept, another embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to execute the method for predicting the concentration of biological enzymes in the microbial fermentation process in the above embodiment.

[0143] Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transfer of computer programs from one location to another. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer. An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and storage medium can be located in an ASIC. Alternatively, the ASIC can be located in a user device. Of course, the processor and storage medium can also exist as discrete components in a communication device.

[0144] In the above embodiment of the present invention, a hybrid prediction training database is established according to the tank batch category. The category of the batch to be predicted is first determined, and then the artificial neural network and extreme gradient boosting model predictions are performed respectively. Then, a hybrid prediction of the two models is performed based on the extreme gradient boosting algorithm. The self-learning ability of the model is used to optimize the weight of the single model in the hybrid prediction model, thereby achieving a more accurate prediction of the product concentration.

[0145] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various modifications or variations within the scope of the claims without affecting the essence of the present invention. The above preferred features may be used in any combination as long as they do not conflict with each other.

Claims

1. A method for predicting the concentration of biological enzymes in a microbial fermentation process, characterized in that: include: Obtaining online data of historical tank batches and current tank batches to be forecasted, as well as offline data of tank batches corresponding to the online data; Performing statistical analysis on the offline data and designing classification criteria to classify historical tank batches; For the online data and the offline data, respectively, a hybrid forecast training database is established according to the classified tank batch categories; Based on the hybrid forecast training database, each type of tank batch data is trained using an artificial neural network forecast model and an extreme gradient boosting forecast model respectively; Based on the extreme gradient boosting algorithm, the trained artificial neural network forecast model and the extreme gradient boosting forecast model are nonlinearly fused to achieve hybrid forecasting; Whenever a tank batch fermentation is completed, the mixed prediction training database is rolled over and updated.

2. The method for predicting the concentration of biological enzymes in a microbial fermentation process according to claim 1, characterized in that: The online data of historical tank batches and the current tank batch to be predicted are obtained, as well as the offline data of the tank batch corresponding to the online data, wherein: the online data includes fermentation liquid temperature T, fermentation liquid pH, fermentation liquid volume V, fermentation tank pressure P, ventilation volume flow F, dissolved oxygen DO and stirring blade speed ω; the offline data includes offline recorded biological enzyme concentration.

3. The method for predicting the concentration of biological enzymes in a microbial fermentation process according to claim 1, wherein: The statistical analysis is performed on the offline data and the classification criteria are designed, wherein: according to the classification function J c,i (t) values ​​are classified and the classification function J is calculated c,i The formula for the value of (t) is: Where, J c,i (t) represents the classification function value of the i-th tank batch at time t, T W Indicates the window width that needs to be considered in order to achieve classification, J i (t) represents the enzyme concentration of the i-th tank batch at time t.

4. The method for predicting the concentration of biological enzymes in a microbial fermentation process according to claim 3, characterized in that: The classification of historical tank batches includes: defining J c,ave (t) is the mean of the classification function at time t, σ(t) is the standard deviation of the classification function at time t, α is the confidence coefficient, and the classification standard is designed as follows: J c,i (t)<J c,ave (t)-α·σ(t), for low-yield tank batches; J c,ave (t)-α·σ(t)≤J c,i (t)≤J c,ave (t)+α·σ(t), is the average tank batch; J c,i (t)>J c,ave (t)+α·σ(t), for high-yield tank batches.

5. The method for predicting the concentration of biological enzymes in a microbial fermentation process according to claim 1, characterized in that: A hybrid forecast training database is established for the online data and the offline data respectively, wherein: the data in the hybrid forecast training database is evenly distributed, and the data sets are mainly derived from production data of the same period of time.

6. The method for predicting the concentration of biological enzymes in a microbial fermentation process according to claim 1, characterized in that: Each type of tank batch data is trained using an artificial neural network prediction model and an extreme gradient boosting prediction model, respectively, wherein: the artificial neural network prediction model is based on the collected fermentation production process state data, establishes a feature model between input and output data, and simulates the actual fermentation process.

7. The method for predicting the concentration of biological enzymes in a microbial fermentation process according to claim 1, characterized in that: Each type of tank batch data is trained using an artificial neural network prediction model and an extreme gradient boosting prediction model, respectively, wherein: the input variables of the extreme gradient boosting prediction model include detectable variables that affect the concentration of the biological enzyme, and the output variables include the advance predicted biological enzyme concentration. During the training process of the extreme gradient boosting prediction model, new classification trees are continuously added to the current model to improve the model prediction accuracy.

8. The method for predicting the concentration of biological enzymes in a microbial fermentation process according to claim 1, characterized in that: The nonlinear fusion of the trained artificial neural network prediction model and the extreme gradient boosting prediction model based on the extreme gradient boosting algorithm includes: The prediction result of a single model is used as input, that is, x f =[U A (t u +c),U X (t u +c)], where U A (t u +c) represents the prediction results of the artificial neural network prediction model, U X (t u +c) represents the prediction result of the extreme gradient boosting prediction model, c represents the lead prediction time; the actual enzyme concentration is used as the output, that is, y f =[U(t u +c)], using the self-learning ability of the model to optimize the weight of a single model in the hybrid forecasting model.

9. The method for predicting the concentration of biological enzymes in a microbial fermentation process according to claim 1, characterized in that: Whenever a tank batch fermentation is completed, the mixed forecast training database is rolled updated, including: If the tank batch is an abnormal tank batch, the tank batch selected by the hybrid forecast training database remains unchanged; If the enzyme concentration curve of the late batch is similar to that of an earlier batch within the confidence region, the earlier batch is replaced by the later batch; If the enzyme concentration curve in the middle and late stages of a tank batch is not similar to that in the confidence region of any historical tank batch, it is directly added to the hybrid prediction training database.

10. A prediction system for enzyme concentration in a microbial fermentation process, characterized in that: include: Acquisition module: acquires online data of historical tank batches and current tank batches to be forecasted, as well as offline data of tank batches corresponding to the online data; Classification module: performing statistical analysis on the offline data and designing classification standards to classify historical tank batches; Database establishment module: establishing a hybrid forecast training database for the online data and the offline data according to the classified tank batch categories; Training and testing module: Based on the hybrid forecast training database, each type of tank batch data is trained using an artificial neural network forecast model and an extreme gradient boosting forecast model; Hybrid forecast module: Based on the extreme gradient boosting algorithm, the trained artificial neural network forecast model and the extreme gradient boosting forecast model are nonlinearly fused to achieve hybrid forecasting; Rolling update module: Whenever a tank batch fermentation is completed, the mixed prediction training database is rolled updated.

Images