Soft-sensing method and controller for key parameters in fermentation process of marine alkaline protease

Abstract

The application discloses a soft measurement method and a controller for key parameters in a marine alkaline protease fermentation process, which is based on intelligent calculation of a hardware platform and computer system software, and realizes online estimation of a target by online acquisition of process data through measuring instruments. Real-time detection values of the key parameters are obtained by using the method, solving the problem of lag in manual offline sampling detection. A soft measurement model is built by using a bidirectional long short-term memory network, which not only combines historical and future information in the fermentation process, but also effectively retains or eliminates key information and bad information in the data, so that the prediction result can better reflect the time sequence of the data. A local model is established by using an instant learning strategy to replace a traditional global model, and a real-time prediction model is established. In addition, the multi-head self-attention mechanism is used to capture the correlation between data, solving the problem of reduced internal correlation of data when selecting local data by instant learning.

Description

Technical Field

[0001] This invention belongs to the field of soft measurement technology, and specifically designs an online estimation method to solve the problem of three key biological parameters—cell concentration, substrate concentration, and relative enzyme activity—that are difficult to detect online in real time using physical sensors during the fermentation production of marine alkaline protease. Background Technology

[0002] Marine alkaline protease exhibits high stability in industrial production, and its activity does not decrease with the continuation of fermentation. However, the internal mechanism of fermentation is quite complex, characterized by strong coupling, nonlinearity, and time-varying properties. Therefore, controlling fermentation conditions to improve protease yield and quality is a key research focus. Key parameters in the fermentation process, such as cell concentration, substrate concentration, and relative enzyme activity, cannot be directly measured in real-time using physical sensors like temperature, pressure, and humidity. Instead, they require manual extraction and offline measurement. Offline detection suffers from long latency and significant measurement errors, which severely hinders the overall design and optimization of fermentation process control, significantly reducing production efficiency, product quality, and yield. Therefore, real-time monitoring of key parameters in the marine alkaline protease fermentation process is of great importance.

[0003] The fermentation process of proteases is characterized by multiple operating conditions, time-varying nature, and nonlinearity. Traditional machine learning modeling methods and single global modeling strategies cannot simultaneously address these issues, resulting in prediction models with low accuracy and poor generalization performance. However, Bidirectional Long Short-Term Memory (BiLSTM) networks in deep learning can not only capture historical and future information during fermentation but also effectively retain or remove key and unwanted information from the data, allowing prediction results to better reflect the temporal nature of the data. To address the multi-operating-condition characteristic, a Just-In-Time (JITL) strategy is used to build multiple local models to replace the traditional global model. A real-time prediction model is established through ensemble processing. Furthermore, the advantage of Multi-Head Self-Attention (MHSA) in capturing the correlations between data points is utilized, solving the problem of reduced internal correlations when JITL selects local data.

[0004] Therefore, for the complex system of marine alkaline protease fermentation, this invention provides an integrated soft measurement method for key parameters of marine alkaline protease fermentation based on a combination of bidirectional long short-term memory network (BiLSTM), multi-head self-attention mechanism (MHSA), and just-in-time learning (JITL). Summary of the Invention

[0005] In marine alkaline protease fermentation, offline measurement of key parameters (such as cell concentration) using manual detection methods is prone to contamination of the fermentation environment by other microorganisms and results exhibit significant lag. To address these issues, this invention provides an integrated soft measurement method for key parameters in marine alkaline protease fermentation, combining Bidirectional Long Short-Term Memory (BiLSTM), Multi-Head Self-Attention (MHSA), and Just-In-Time (JITL) learning. First, conventional physical sensors are used to acquire values ​​for physical parameters such as temperature, pressure, and humidity during the fermentation process. Then, the soft measurement method is used to achieve real-time online prediction of the key parameters. The specific process includes the following:

[0006] Step 1: Fermentation data classification. Collect auxiliary variables and key parameter data from several batches of historical fermenters with the same fermentation environment. Define a part of the data as the training dataset and the other part as the test dataset. The training dataset is used to train BiLSTM, and the test dataset is used to test the prediction accuracy and generalization ability of the model.

[0007] Step 2: Establish a local sample set. Use Wasserstein distance to select the top few similar samples with the highest similarity to the input data to be predicted from the historical samples, and calculate the similarity weight. Finally, select an appropriate number of historical data around each similar sample as the local sample dataset.

[0008] Step 3: Select the principal components. Calculate the cumulative variance contribution rate ζ of the auxiliary variables in each local sample dataset using principal component analysis. Select the auxiliary variables with ζ ≥ 90% as the input principal components.

[0009] Step 4: Improve data correlation: Feed the input principal component data from Step 3 into the constructed multi-head self-attention mechanism network.

[0010] Step 5: Build the BiLSTM model and debug the model parameters;

[0011] Step 6: Establishment of the real-time prediction model for key parameters: When the sample data to be predicted arrives, multiple local sample datasets are established and fed into the built BiLSTM model to obtain the predicted values ​​of key parameters. Multiple predicted values ​​are integrated according to similarity weights to calculate the final real-time predicted values ​​of key parameters.

[0012] Furthermore, the training dataset mentioned in step 1 accounts for 80% of the total historical dataset, with the remaining 20% ​​serving as the test dataset. Auxiliary variables are selected from external variables that can be directly measured by physical sensors and are closely related to the fermentation process, including: fermentation broth pH, ​​fermenter temperature T, air flow rate Q, dissolved oxygen DO, carbon dioxide concentration CO2, internal pressure Pa, ventilation rate η, glucose replenishment rate α, ammonia replenishment rate β, and stirring speed V. Key parameters are the cell concentration X, substrate concentration S, and protease activity P in the fermentation broth.

[0013] Furthermore, the specific process of step 2 is as follows:

[0014] Step 2.1: Calculate the sample data P to be predicted a ∈R n With each sample data P in the historical database b∈{1,2,...,j} ∈R n The Wasserstein distance between the data points is used as the similarity value; the smaller the distance, the greater the similarity value.

[0015] Step 2.2: Arrange the similarity values ​​in descending order {d1, d2, ..., dk} according to the number of historical working conditions k. k} Calculate the weight coefficient using the similarity value:

[0016]

[0017] Step 2.3: Around each selected similar sample P h∈{1,2,...,k} Select m historical samples as the local sample dataset {P m1 ,P m2 ,...,P mk}

[0018] Furthermore, in step 3, the auxiliary variables are transformed into input principal components using principal component analysis, and the components with a cumulative variance contribution rate ζ≥90% are used as input principal components.

[0019] Furthermore, the process of improving data correlation by utilizing the constructed multi-head self-attention mechanism network in step 4 is as follows:

[0020] Step 4.1: Initialize and set the weight matrices for different spaces And input X i Multiplying by the weight matrices of different spaces respectively yields q. (h,i) k (h,i) v (h,i) .

[0021] Step 4.2: Obtain q (h,i) respectively with k (h,1),k (h,2) ...k (h,N) Perform vector dot product operation and get the result a (h,1) ,a (h,2) ...a (h,N) The input is processed by the Softmax layer to obtain attention weight values ​​a' between (0,1). (h,1) ,a' (h,2) ...a' (h,N) .

[0022] Step 4.3: Compare the obtained weight values ​​with the corresponding v values. (h,1) ,v (h,2) ...v (h,N) Multiply the results of multiplying different spaces, then add the results of multiplying them in different spaces, and after a linear transformation, obtain the result that is the same as the input X. i The corresponding output b i .

[0023] Furthermore, the specific process of step 6 is as follows:

[0024] Step 6.1, when the input vector X of the sample to be tested... i Upon arrival, multiple local sample datasets and similarity weight coefficients δ are obtained through calculations in step 2. l .

[0025] Step 6.2: Feed each local sample dataset into the constructed BiLSTM model and obtain multiple predicted values ​​y. s .

[0026] Step 6.3: Calculate the real-time predicted values ​​of key parameters based on the predicted values ​​and similarity weighting coefficients.

[0027]

[0028] The beneficial effects of this invention are:

[0029] 1. This invention utilizes physical sensors to directly measure the values ​​of auxiliary variables, as well as a small number of offline measurements of key parameters, to establish an integrated JITL-MHSA-BiLSTM soft measurement model for key parameters during marine alkaline protease fermentation, thus solving the problem of difficulty in real-time online detection of key parameters during fermentation.

[0030] 2. By using a soft measurement model to obtain real-time detection values ​​of key parameters, the problem of data lag caused by manual offline sampling and detection is solved, thereby making the fermentation process more controllable and improving the quality and yield of marine alkaline protease.

[0031] 3. This invention utilizes a bidirectional long short-term memory network (BiLSTM) to build a basic soft measurement model. This model not only combines historical and future information during the fermentation process but also effectively retains or removes key and harmful information from the data, enabling the prediction results to better reflect the temporal sequence of the data.

[0032] 4. This invention addresses the multi-condition nature of the fermentation process by using a real-time learning strategy to establish a local model instead of the traditional global model, thereby creating a real-time prediction model. At the same time, it utilizes the advantage of multi-head self-attention mechanism to capture the correlation between data, solving the problem of reduced internal correlation of data when real-time learning selects local data. Attached Figure Description

[0033] Figure 1 A flowchart illustrating the fermentation process of marine alkaline protease and the setup of measuring instruments;

[0034] Figure 2 A flowchart for establishing a soft measurement model for key parameters in the marine alkaline protease fermentation process;

[0035] Figure 3 Diagram of the basic prediction model for the MHSA-BiLSTM network;

[0036] like Figure 1 The system includes: 1. biological fermentation tube, 2. air purification device, 3. water supply system, 4. gas flow sensor, 5. speed sensor, 6. CO2 detector, 7. pressure gauge, 8. resistance temperature detector, 9. dissolved oxygen electrode, 10. pH electrode, 11. peristaltic pump, 12. centrifuge, 13. intelligent controller, and 14. host computer.

[0037] Figure 1 The meanings of the symbols used are explained below:

[0038] Fermentation broth pH, ​​fermenter temperature T, air flow rate Q, dissolved oxygen DO, exhaust gas concentration CO2, tank pressure Pa, ventilation rate η, stirring speed V, glucose replenishment acceleration rate α, ammonia replenishment acceleration rate β. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and specific implementation examples of marine alkaline protease.

[0040] like Figure 2 As shown, the specific implementation process of this invention is as follows:

[0041] 1. Data Collection and Selection of Auxiliary Variables

[0042] Variables that can be directly measured by physical sensors and are closely related to the fermentation process of marine alkaline protease were selected as auxiliary variables for the soft sensor model.

[0043] (1) Acquisition of fermentation process data

[0044] The selected strain was YS-80, a high-yield, low-temperature alkaline protease strain. During fermentation, the tank temperature was controlled at 28±0.5℃, the motor stirring speed at 250 r / min, the tank pressure at 0.04 MPa, the aeration rate at 0.4 L / min, the pH at 9.5, and the dissolved oxygen at 45%–75%. The fermentation cycle was 48 hours, and data from 10 batches were collected. Figure 1 The fermentation control system shown includes a bioreactor 1, an air purification device 2, and a water supply system 3. The intelligent controller 13 is electrically connected to a gas flow sensor 4, a speed sensor 5, a CO2 detector 6, a pressure gauge 7, a resistance temperature detector 8, a dissolved oxygen electrode 9, a pH electrode 10, a peristaltic pump 11, and a host computer 14.

[0045] During the fermentation process, air flow rate Q is collected by gas flow sensor 4, motor stirring speed V is collected by speed sensor 5, fermenter temperature T is collected by thermal resistor 8, pH of fermentation broth pH is collected by pH electrode 10, internal pressure Pa is collected by pressure gauge 7, ventilation volume η is collected by gas flow sensor 4, carbon dioxide concentration CO2 is collected by CO2 detector 6, dissolved oxygen concentration is collected by dissolved oxygen electrode 9, glucose replenishment rate α and ammonia replenishment rate β are collected by peristaltic pump 11, the fermentation broth is separated by centrifuge 12, and finally the offline analysis value of cell concentration is obtained by offline detection and stored in host computer 14 by intelligent controller 13.

[0046] (2) Selecting auxiliary variables

[0047] The auxiliary variables selected are: pH of fermentation broth, T of fermenter, Q of air flow rate, DO of dissolved oxygen, CO2 concentration of exhaust gas, Pa of internal pressure, η of ventilation, V of stirring speed, α of glucose replenishment acceleration rate, and β of ammonia replenishment acceleration rate.

[0048] 2. Classification of fermentation data

[0049] Auxiliary variables and key parameter data of several batches of historical fermenters with the same fermentation environment were collected. A portion of the data was defined as the training dataset and the other portion as the test dataset. The training dataset was used to train BiLSTM, and the test dataset was used as the test sample to test the prediction accuracy and generalization ability of the real-time model. The test dataset accounted for 80% of the total fermentation dataset collected.

[0050] 3. Establish a local sample dataset

[0051] (1) Calculate the similarity value

[0052] Calculate the sample data P to be predicted a ∈Rn With each sample data P in the test dataset b∈{1,2,...,j} ∈R n The Wasserstein distance between them. Where P a P b ∈R n Where n and j represent the auxiliary variable and the number of historical sample data, respectively, the calculation expression is:

[0053]

[0054] In the formula, inf represents the infimum, (P a ,P b ) represents all P a and P b Let γ be the set of all joint distributions of the marginal distributions. For each possible joint distribution γ, a real sample x and a target generated sample y are sampled. The lower bound of the expected value of the distance ||xy|| between the samples is the Wasserstein distance. This distance is denoted as the similarity value between the data; the smaller the distance, the greater the similarity value.

[0055] (2) Calculate similarity weights

[0056] Arrange the similarity values ​​in descending order based on the number of test datasets {d1, d2, ..., d8}, and calculate the weight coefficients using the similarity values:

[0057]

[0058] (3) Establish a local sample dataset

[0059] Around each selected similar sample P h∈{1,2,...,8} We selected 1000 historical samples as local sample datasets, resulting in a total of eight local sample datasets. Each dataset has a dimension of n×m, where n is the number of auxiliary variables and m is 1000. The value of m can be chosen according to your requirements.

[0060] 4. Determine the input principal components of the local sample dataset.

[0061] Principal component analysis was used to analyze eight local sample datasets P. n×m Principal component analysis was performed sequentially.

[0062] Taking the first local sample dataset as an example, the correlation of auxiliary variables is assigned to new variables. This process not only reduces the number of auxiliary variables, but also retains the main features of the auxiliary variables.

[0063] First, consider the sample set X = {X1, X2, ..., X...} m All features of} are centralized, X i={x1,x2,...x n Let} represent the i-th feature vector of n samples, and μ represent the average value. The calculation process is as follows:

[0064]

[0065] {x1,x2,...,x n}={x1-μ,x2-μ,...,x n -μ}

[0066] Suppose the new coordinates after principal component analysis are W = {w1, w2, ... w} m The objective function is obtained using the maximum projection variance method:

[0067]

[0068] Solving the objective function using the Lagrange multiplier method yields:

[0069] XX T W = λW

[0070] λ is the covariance matrix XX T The eigenvalues ​​are ordered as follows: λ1≥λ2≥...≥λ n Calculate the variance contribution rate:

[0071]

[0072] The number of principal components is the number of eigenvalues ​​corresponding to a cumulative variance contribution rate greater than 90%.

[0073] The cumulative variance contribution rate of the principal components in the local sample dataset is shown in Table 1.

[0074] Table 1 Cumulative Variance Contribution Rate of Principal Components

[0075]

[0076]

[0077] As shown in Table 1, the cumulative variance contribution rate of the first 6 principal components already accounts for 91%, so the first 6 principal components are selected as input principal components.

[0078] 5. Improve data correlation

[0079] The local sample dataset {(x) processed by 4 i ,y i |i=1,2,...,8},x i ∈R 6 For 6-dimensional input data, y iThe output of the R-dimensional sample is fed into Figure 3 The multi-head self-attention mechanism network built in the middle is operated as follows:

[0080] X i Multiply by weight matrices of three randomly initialized spaces respectively. Get q (d,m) k (d,m) v (d,m) The calculation process is as follows:

[0081]

[0082] in d represents the number of input sample data, and d represents the number of spaces. It is a randomly generated m×n weight matrix, q (d,m) For the query vector, k (d,m) Let v be the key vector. (d,m) It is a value vector. Using the obtained q... (d,m) respectively with k (d,1) ,k (d,2) ...k (d,1000) Perform the vector dot product operation and output the result a (d ,1),a (d,2) ...a (d,1000) The input is processed by the Softmax layer to obtain attention weight values ​​a' between (0,1). (d,1) ,a' (d,2) ...a' (d,1000) The obtained weight values ​​are compared with the corresponding v values. (d,1) ,v (d,2) ...v (d,1000) Multiply, then add the results of multiplying different spaces, and after a linear transformation, obtain the result equal to the input X. i The corresponding output b i The calculation process can be represented as follows:

[0083]

[0084] 6. Build a BiLSTM prediction model

[0085] like Figure 3 As shown, the specific process of building the model includes:

[0086] Assume the BiLSTM network has three input states at time t: cell state C at time t-1. t-1 The hidden state h at time t-1 t-1 and the input vector X at time t t There are two output states: cell state C t and hidden state h t;

[0087] Step 6.1, calculate the state f of the forget gate. t The calculation process is as follows:

[0088] f t =σ(W xf x t +W hf h t-1 +b f )

[0089] In the formula, σ represents the activation function sigmoid; W xf W hf b represents the weight. f Represents the bias vector;

[0090] Step 6.2: Update the cell state using the input information. The calculation process is as follows:

[0091]

[0092] In the formula W xi W hi W xc W hc This represents the weights after initialization, b i ,b c Represents the corresponding bias vector, i t This represents the vector to be updated. This represents a candidate vector.

[0093] Step 6.3: Update the cell state with the new information and maintain it until the next time step. The cell state is continuously updated. The calculation process is as follows:

[0094]

[0095] In the formula, * represents dot product;

[0096] Step 6.4, the output state, as the last part, is mainly used to determine the value of the hidden state at the next time step and the output at the current time step. The calculation process is as follows:

[0097]

[0098] In the formula W xo W ho b represents the initial weight in the output state. o Represents the bias vector, O t This represents the vector to be output.

[0099] A BiLSTM model was built using the Deep Network Designer in MATLAB, and the model parameters are shown in Table 2.

[0100] Table 2

[0101] Number of hidden layer units 250 Drop rate 0.2 Maximum number of iterations 250 Minimum batch number 2 Initial learning rate 0.01 Optimize gradient descent method Adam Gradient threshold 1 Learning rate factor 0.2

[0102] Integration of key parameter prediction values

[0103] The predicted values ​​of key parameters {y1, y2, ..., y8} are obtained using the prediction model, where 8 is the number of local sample datasets. Each predicted value is then multiplied by its corresponding similarity weight coefficient δ. l Finally, the values ​​are added together to obtain the final integrated real-time prediction value Y. i .

[0104]

[0105] After the soft measurement model is established, it is implemented using embedded C language programming and embedded into the intelligent controller 13. When the input vector X of the batch to be predicted is... i+1 After the measurement instrument reads the data into the intelligent controller 13, the predicted values ​​of the key parameters are calculated using a soft measurement program, and the prediction results are sent to the host computer 14 for display.

[0106] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent methods or modifications that do not depart from the technology of the present invention should be included within the scope of protection of the present invention.

Claims

1. A soft-sensing method for key parameters in a fermentation process of marine alkaline protease, characterized in that, It comprises the following steps: Step 1: fermentation data classification, collecting auxiliary variables and key variable data of several batches of historical fermentation tank batches with the same fermentation environment, defining part of the data as a training data set and the other part as a test data set, using the training data set to train BiLSTM, and using the test data set to verify the prediction accuracy and generalization ability of the model; In step 1, the auxiliary variables include: fermentation broth pH, fermentation tank temperature T, air flow Q, dissolved oxygen DO, carbon dioxide concentration CO2, tank pressure Pa, ventilation amount η, glucose supplement rate α, ammonia water supplement rate β and stirring speed V; In step 1, the key variables are the cell concentration X, substrate concentration S and protease activity P in the fermentation broth; In step 1, the training data set accounts for 80% of the total historical data set, and the remaining 20% is the test data set; Step 2: Establish a local sample set, select the top few similar samples from the historical samples with the highest similarity to the input data to be predicted, calculate the similarity weight, and finally select an appropriate number of historical data around each similar sample as a local sample data set; Step 3: Input principal component selection, calculate the cumulative variance contribution rate ζ of the auxiliary variables in each local sample data set using principal component analysis, and select appropriate auxiliary variables as input principal components according to the ζ value; The specific process of step 3 includes: Step 3.1: First, center all features of the sample set X = {X1, X2,... X m}, X i = {x1, x2,... x n} represents the i-th feature vector of n samples, the calculation process is as follows: {x1, x2,..., x n} = {x1-μ, x2-μ,..., x n -μ} Step 3.2: Assume the new coordinates after principal component analysis is W = {w1, w2,... w m}, the objective function is obtained by the maximum projection variance method: Solve the objective function using the Lagrange multiplier method to get: XX T W = λW λ is the eigenvalue of the covariance matrix XX T , which is sequentially sorted: λ1≥ λ2≥... ≥ λ n , the variance contribution rate is calculated: Select the auxiliary variables with ζ≥90% as the input principal components; Step 4: Improve data correlation: input the input principal component data in step 3 into the multi-head self-attention mechanism network built; Step 5: Build a BiLSTM model and debug the model parameters; Step 6: Establishment of real-time prediction model of key variables: when the sample data to be predicted arrives, establish multiple local sample data sets and input them into the built BiLSTM model to obtain the predicted value of the key variables, integrate multiple predicted values according to the similarity weight, and calculate the final real-time prediction value of the key variables.

2. The soft-sensing method of marine alkaline protease fermentation process key parameters according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: Calculate the sample data P to be predicted a ∈R n Wasserstein distance between each sample data P in the historical database b∈{1,2,...,j} ∈R n and P, where P a , P b ∈R n , n and j represent auxiliary variables and the number of historical sample data respectively, and the calculation expression is: inf P a b ) denotes the set of all joint distributions with marginal distributions P a and P b . From each possible joint distribution γ, a real sample x and a target generated sample y are sampled. The lower bound of the expectation E of the distance ||x-y|| is the Wasserstein distance. The distance is denoted as the similarity value between data, and the smaller the distance is, the larger the similarity value is.​ Step 2.2: Arrange the similarity values from large to small according to the number of historical working conditions k {d1, d2,..., dk} k}, and calculate the weight coefficient using the similarity value: Step 2.3: Selecting each similar sample P around the selected P h∈{1,2,...,k} m historical samples as a local sample dataset {P m1 ,P m2 ,...,P mk}.

3. The soft-sensing method of marine alkaline protease fermentation process key parameters according to claim 1, characterized in that, The specific process of step 4 includes: Step 4.1: initialize the weight matrix pair (W h Q , W h K , W h V ) for different spaces, and multiply the input X i by the weight matrix pair of different spaces to obtain q (h,i) , k (h,i) , v (h,i) , respectively, where q (h,i) is a query vector, k (h,i) is a key vector, and v (h,i) is a value vector; Step 4.2: Perform vector dot product operation on the obtained q (h,i) and k (h,1) , respectively, and input the result a (h,2) to the Softmax layer to obtain the attention weight values a'(h,1), a'(h,2)...a'(h,N) between (0,1). (h,N) (h,1) (h,2) (h,N) ​​​​ Step 4.3: The obtained weight values are multiplied with the corresponding v (h,1) , (h,2) ...v (h,N) respectively, and the results of multiplication of different spaces are added, and the output b i corresponding to the input X i is obtained after linear transformation.

4. The soft-sensing method of marine alkaline protease fermentation process key parameters according to claim 1, characterized in that, The specific process of building the model in step 5 includes: Assume that the BiLSTM network has three input states at time t: cell state C t-1 , hidden layer state h t-1 , and input vector X t at time t, and two output states: C t and h t ; Step 5.1, Compute the state of the forget gate f t The computation is as follows: f t = σ(W xf x t + W hf h t-1 + b f ) where σ denotes the activation function sigmoid; W xf ,W hf denotes the weights, b f denotes the bias vector; Step 5.2, update the cell state with the input information, the calculation process is as follows: where W xi ,W hi ,W xc ,W hc denotes the initialized weight, b i ,b c denotes the corresponding bias vector, i t denotes the vector to be updated, denotes the candidate vector; Step 5.3, update the new information to the cell state and keep it to the next moment, where the cell state is always updated, the calculation process is as follows: Where * represents dot product; Step 5.4, output the state as the last part, mainly used to determine the value of the next moment hidden layer state and the output of the current moment, the calculation process is as follows: where W xo ,W ho denotes the initialized weights in the output state, b o denotes the bias vector, O t denotes the partial vector to be output; The BiLSTM model parameters are set as shown in the following table 。 5. The soft-sensing method of marine alkaline protease fermentation process key parameters according to claim 1, characterized in that, The specific process of step 6 is as follows: The prediction model is used to obtain prediction values {y1, y2,..., y8} of the key parameters, 8 being the number of local sample data sets, the prediction values are multiplied by corresponding similarity weight coefficients respectively, and finally added to obtain the final integrated real-time prediction value Y i :

6. A controller characterized by comprising: The controller can execute the soft-sensing method of the key variables of the marine alkaline protease fermentation process according to any one of claims 1-5.

Images