A fermentation process soft measurement modeling method based on domain adaptation graph network

Abstract

The application discloses a fermentation process soft measurement modeling method based on a domain adaptation graph network, and the method comprises the following steps: (1) data acquisition and analysis: obtaining different batch data through a penicillin simulation software; (2) input variable selection: selecting input variables having high correlation with output variables; (3) modeling training: constructing a process soft measurement model based on the DAGCN; and (4) model performance evaluation: evaluating the model performance based on target domain test data. The method can extract the relationship between process variables, further adapt the nonlinear difference between different batches based on a fine-tuning technology, and improve the use range of the model.

Description

Technical Field

[0001] This invention relates to the field of product quality prediction in fermentation processes, and in particular to a soft measurement modeling method for fermentation processes based on domain-adaptive graph networks. Background Technology

[0002] Penicillin fermentation is a typical biochemical reaction process, mainly referring to the metabolic activities of penicillin bacteria growing and synthesizing antibiotics under different environments. The fermentation process is highly time-varying and nonlinear. As a typical batch process, it exhibits multi-batch characteristics. However, due to the finite product lifecycle and the cost-effectiveness of investment, online measurement of product quality presents significant challenges. Data-driven soft measurement methods, due to their low cost and low latency, have gradually replaced expensive and time-consuming offline measurement methods and are widely used in the field of critical quality prediction for process products.

[0003] Deep learning-based data-driven methods, such as convolutional neural networks and autoencoders, have demonstrated good capabilities in historical data mining. Most applications assume that training and testing data have similar process characteristics. However, the batch process of penicillin fermentation often involves changing operating conditions or feedstock ratios to produce products of various specifications. Due to the nonlinear differences between processes, the predictive performance of models built for a specific process tends to decline across different processes. Furthermore, re-collecting labeled data to train models for new processes is time-consuming and labor-intensive, making it impossible to build reliable soft sensor models.

[0004] Domain-adaptive soft measurement methods based on transfer learning have been gradually applied to intermittent processes. These methods relax restrictions on process nonlinearity differences and the amount of labeled data, aiming to transfer data knowledge from related domains to assist in current domain modeling. However, existing research often has certain shortcomings: they ignore the relationships between variables, instead directly encapsulating them as abstract features, which to some extent limits the application of the models.

[0005] Graph neural networks, a branch of deep learning, can learn data with graph structures, which helps in understanding the relationships between different nodes. As a variant of traditional graph neural networks, graph convolutional networks (GCNs) demonstrate powerful representation capabilities by performing convolution operations on structured data. However, the application of GCNs in process industries is currently limited, especially in batch processes. Research on how to better adapt to the nonlinear differences between batches in batch processes and establish domain-adaptive models suitable for batch switching processes is lacking. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention proposes a Domain Adaptive Graph Convolution Network (DAGCN) method for soft measurement modeling of the penicillin fermentation process. DAGCN considers the topological structure between variables and the temporal characteristics of the process, aiming to expand the model's applicability by utilizing data information from related processes. The specific modeling process is as follows: First, the relevant process data is modeled based on GCN to capture the topological structure information of the data, where the relationship matrix of the variables is trained end-to-end. Temporal information from the data collection is incorporated to capture the temporal characteristics of the process. Fully connected layers learn the mapping relationship between process variables and output variables. Finally, the parameters of the fully connected layers are fine-tuned based on a small amount of data from the current batch, expanding the model's applicability. The proposed model demonstrates good performance during the intermittent process of penicillin fermentation.

[0007] The technical solution adopted by this invention to solve its technical problem is:

[0008] A soft sensor modeling method for fermentation processes based on domain-adaptive graph networks, the method comprising the following steps:

[0009] (1) Data acquisition and analysis

[0010] Data from different batches was obtained using penicillin simulation software.

[0011] (2) Input variable selection

[0012] Select input variables that are highly correlated with the output variables as model inputs.

[0013] (3) Modeling training

[0014] Construct a process soft measurement model based on DAGCN.

[0015] (4) Model performance evaluation

[0016] The model performance was evaluated based on test data from the target domain.

[0017] Furthermore, the process of step (1) is as follows:

[0018] Step 1.1: Simulate different reaction conditions of the penicillin fermentation process using the Pensim simulation software. The specific variable variables in the fermentation process are: time (t), aeration rate (AR), agitator power (AP), and substrate feed temperature (T). SF Substrate concentration (E) SDissolved oxygen concentration (E) DO ), Container Volume (V) V Carbon dioxide concentration (E) CO2 ), pH of the solution, and fermentation temperature (T). G The output variable is penicillin concentration (E), etc. P To reflect a real production process, 2% Gaussian noise was added to the process. This resulted in three different batch characteristics of penicillin fermentation processes, denoted as batches G1, G2, and G3.

[0019] Step 1.2: Divide the data from different batches into source domain and target domain data, where the source domain has a sufficient amount of labeled data, while the target domain has only a small amount of labeled data.

[0020] Step 1.3: Due to the differences in dimensions between different input variables, it is necessary to standardize the data. The specific formula is shown below:

[0021]

[0022] Where X is the original data without standardization, X' is the data after standardization, and μ and δ are the mean and standard deviation of the data, respectively.

[0023] Furthermore, the process of step (2) is as follows:

[0024] The interdependence between the source domain input variable x and the output variable y is measured based on the mutual information (MI) correlation between the two variables. The specific calculation is as follows:

[0025]

[0026] Where P(x) and P(y) are the marginal probability density functions of individual variables x and y, respectively, and I(x,y) and P(x,y) are the MI coefficient and joint probability density function of the two variables, respectively. Specifically, the larger the MI value, the higher the correlation between the variables.

[0027] Furthermore, the process of step (3) is as follows:

[0028] Step 3.1: Based on the MI ranking variable, select the main process variable as the model input. Train the GCN module using source domain data. The specific graph convolution operation can be represented by a variation as follows:

[0029]

[0030] GCN(X) Θ A) represents the graph convolution operation, D -1 / 2 AD -1 / 2 For the graph convolution kernel, D = ∑ j A j Let X be the degree matrix of adjacency matrix A. Θ The characteristic matrix is ​​represented by the symbol. Where H represents the batch size, C represents the number of channels, P represents the time step, and V represents the number of variables. I represents the identity matrix, conv(.) represents the convolution operation, σ(.) is the activation function, and W is the weight matrix.

[0031] To reduce the impact of feature location on the predicted variable, the features extracted by GCN are input into the Fully Connected Layers (FCL) module to learn the mapping relationship between the features and the output variable.

[0032] The root mean square error (RMSE) is used as the model loss. To learn an adjacency matrix A with better generalization performance, element-level entropy constraints are added to encourage structural discretization. The specific loss formula is as follows:

[0033]

[0034] L(A)=-Alog2(A)-(1-A)log2(1-A)

[0035] Where y j and and represent the true value and predicted value of the source domain prediction data, respectively, where N is the number of training samples in the source domain. L(.) represents the element-level entropy, and γ represents the regularization coefficient.

[0036] Step 3.2: Based on the frozen GCN module parameter information, fine-tune the FCL module parameters using a small amount of labeled data in the target domain to expand the applicability of the model.

[0037] Furthermore, the process of step (4) is as follows:

[0038] The model's performance is evaluated on the test set data in the target domain, specifically using RMSE and Mean Absolute Error (MAE) metrics.

[0039]

[0040]

[0041] Where y i and These are the true and predicted values ​​of the target domain prediction data, respectively. The test sample size of the target domain is U. The smaller the RMSE and MAE indices, the better the model performance.

[0042] The beneficial effects of this invention are mainly reflected in the following: This invention proposes a soft sensor modeling method for fermentation processes based on domain-adaptive graph networks. While capturing variable relationships and process temporal characteristics, it further expands the applicability of the model based on fine-tuning techniques. Attached Figure Description

[0043] Figure 1 This is a diagram showing the relationship between the input and output variables of this invention;

[0044] Figure 2 This is the domain adaptation graph network structure diagram of the present invention;

[0045] Figure 3 In the figure, Figure (a) and Figure (b) are the absolute prediction error graphs of the baseline model of the present invention and DAGCN in batch G2 to batch G1 and batch G2 to batch G3, respectively. Detailed Implementation

[0046] The present invention will now be further described with reference to the accompanying drawings.

[0047] Reference Figures 1-3 A soft sensor modeling method for fermentation processes based on domain-adaptive graph networks, the method comprising the following steps:

[0048] (1) Obtain the process dataset, the process is as follows:

[0049] Step 1.1: Set different penicillin fermentation reaction conditions using the simulation platform (Pensim) and add 2% Gaussian noise to simulate real data characteristics. The specific sampling time is twice per hour, and the fermentation time is 400 hours. By changing the range of input variables, 800 samples were collected for each batch, and process data for three batches, G1, G2, and G3, were collected in total.

[0050] Step 1.2: Define the source domain as batches with relatively sufficient labeled data, while the target domain is a batch with only a small amount of labeled data. Specifically, when batch G1 is used as the source domain data, 800 labeled samples are available. Batches G2 and G3 are used as the target domains, and the target domain training data is set with 5 labeled samples, with the remaining 795 samples used as test data.

[0051] Step 1.3: Standardize the data from different batches to reduce the impact of differences in data units on modeling.

[0052] The specific formula is shown below:

[0053]

[0054] Where X is the original data without standardization, X' is the data after standardization, and μ and δ are the mean and standard deviation of the data, respectively.

[0055] (2) Select and output input variables that are highly correlated, as follows:

[0056] Step 2.1: Calculate the MI coefficients between the input and output variables in the source domain data, such as... Figure 1 As shown.

[0057] Step 2.2: Select and E P E with high correlation S E DO V V E CO2 T G The six variables t are used as input variables.

[0058] (3) The modeling and training process is as follows:

[0059] Step 3.1: Based on the selected input variables, train the GCN and FCL modules using source domain data.

[0060] Step 3.2: Fix the parameters of the GCN module trained in the source domain, and fine-tune the FCL parameters based on a small amount of labeled data in the target domain. The specific structure diagram of DAGCN is shown below. Figure 2 As shown.

[0061] Table 1. Performance of the baseline model and DAGCN on the target domain test set.

[0062]

[0063]

[0064] (4) The model performance evaluation process is as follows:

[0065] The comparison method uses a source domain-based GCN trained to directly predict the target domain and a Domain Adaptation Extreme Learning Machine (DAELM) to train the model. Specifically, the model is trained using labeled data from the source domain and a small amount of labeled data from the target domain. The prediction results are shown in Table 1, and the absolute prediction errors are as follows: Figure 3As shown in (a) and (b), DAGCN achieved the lowest RMSE and MAE scores across all cases. Specifically, the GCN model, while learning the relationships between variables, further captured the temporal characteristics of process variables based on time information, resulting in predicted trends similar to the true values. However, due to nonlinear differences between batches, the performance of directly trained source domain models degrades across different domains. DAGCN, with the help of fine-tuning techniques, updated the parameters of the model's output layer based on a small amount of labeled data from the target domain, improving the model's applicability. Because DAELM's ability to extract temporal and data structure features is insufficient, it cannot effectively capture the characteristics of the penicillin growth process, resulting in larger prediction errors and inferior performance compared to DAGCN. Therefore, DAGCN achieved the best prediction performance across all cases.

[0066] The method of this invention employs a soft measurement modeling method for fermentation processes based on domain-adaptive graph networks, which improves the model's ability to capture variable relationships and its applicability across batches, and has universality and applicability.

[0067] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A soft sensor modeling method for fermentation processes based on domain-adaptive graph networks, characterized in that... The method includes the following steps: (1) Data acquisition and analysis: Data from different batches were obtained using penicillin simulation software; (2) Input variable selection: Select input variables that are highly correlated with the output variables as model inputs; (3) Modeling and training: Construct a process soft measurement model based on DAGCN; the specific process is as follows: Step 3.1: Use the process variables selected based on MI as input to the model, and train the GCN module with source domain data. The specific graph convolution operation can be represented by a variation as follows: ； GCN(X) Θ A) represents the graph convolution operation, D -1 / 2 AD -1 / 2 For the graph convolution kernel, D = ∑ j A j Let X be the degree matrix of adjacency matrix A. Θ Let X represent the characteristic matrix. Θ ∈ℝ H×C×P×V Where P represents the time step, C represents the number of channels, H represents the batch size; I represents the autocorrelation identity matrix, conv represents the convolution operation, σ is the activation function, and W is the weight matrix; To reduce the impact of feature location on the predicted variables, the features extracted by GCN are input into the fully connected layer module FCL to learn the mapping relationship between the variables and the output variables. The root mean squared error (RMSE) is used as the model loss. To learn an adjacency matrix A with better generalization performance, element-level entropy constraints are added to encourage structural discretization. The specific loss formula is as follows: ； ； Where y j and j These are the true value and the predicted value, respectively, where N is the number of training samples in the source domain; L(.) represents the element-level entropy, and γ represents the regularization coefficient; Step 3.2: Based on the frozen GCN module parameter information, fine-tune the FCL module parameters using a small amount of labeled data in the target domain to expand the applicability of the model; (4) Model performance evaluation: Evaluate model performance based on target domain test data.

2. The soft sensor modeling method for fermentation processes based on domain-adaptive graph networks as described in claim 1, characterized in that, The process of step (1) is as follows: Step 1.1: Obtain data from different batches during the fermentation process; Step 1.2: Divide the data into source domain and target domain data; Step 1.3: Standardize the data.

3. A soft sensor modeling method for fermentation processes based on domain-adaptive graph networks as described in claim 1 or 2, characterized in that, The process of step (2) is as follows: The mutual information (MI) correlation between two variables is used to measure the interdependence between the source domain input variable x and the output variable y. The specific calculation is as follows: ； Where P(x) and P(y) are the marginal probability density functions of individual variables x and y, and I(x, y) and P(x, y) are the MI coefficient and joint probability density function of the two variables, respectively; specifically, the larger the MI value, the higher the correlation between the variables.

4. A soft sensor modeling method for fermentation processes based on domain-adaptive graph networks as described in claim 1 or 2, characterized in that, The process of step (4) is as follows: The model's performance is evaluated on the test set data in the target domain, specifically using RMSE and mean absolute error (MAE) metrics. The smaller the RMSE and MAE metrics, the better the model's performance.

Images