Domain adaptation width learning industrial penicillin concentration prediction method based on parameter migration

Abstract

The invention discloses a domain adaptation width learning industrial penicillin concentration prediction method based on parameter migration, and the method is based on BLS, guarantees the calculation precision, and improves the calculation efficiency of a model. According to the method, learnable parameter conversion mapping is designed, the output parameters of the source domain and the target domain are projected into the same parameter space, and the data distribution difference under different domains is reduced to the maximum extent. A distribution alignment regular term of a source domain and a target domain is designed based on MMD, and the migration efficiency of data between the domains is improved. An output parameter regular term and a migration parameter regular term are constructed in an objective function, and overfitting and knowledge negative migration of the model are avoided. A model parameter optimization algorithm is designed, alternative optimization is carried out on output parameters and migration parameters, and optimal model parameters are calculated in a self-adaptive mode. The method is applied to soft measurement modeling of the industrial penicillin fermentation process with the multi-working-condition characteristic, and high-precision prediction of the penicillin concentration in the industrial penicillin fermentation process can be achieved.

Description

Technical Field

[0001] The invention relates to a parameter migration-based domain adaptation width learning industrial penicillin concentration prediction method, belonging to the technical field of soft measurement. Background Art

[0002] In modern industrial processes, measuring quality variables, key process variables, and economic indicators is crucial for building smart factories. However, due to complex process environments and measurement delays, most quality variables are difficult to detect online, such as penicillin content detection during industrial penicillin fermentation. The industrial penicillin fermentation process is a typical biochemical reaction process, describing the metabolic activities of penicillin bacteria as they grow and synthesize antibiotics under environmental conditions such as culture medium, pH, and temperature. In actual penicillin fermentation processes, the inability to effectively monitor key variables such as product concentration, bacterial concentration, and substrate concentration using online detection instruments poses challenges to its control and optimization. To effectively ensure product quality during penicillin fermentation, soft sensing technologies used in industrial processes are required to predict these key variables online.

[0003] Soft sensing technology, which uses easily measurable process variables to establish mathematical models and identify key quality variables, is an effective solution. Currently, data-driven soft sensing methods, which rely solely on process data to build predictive models, are independent of production experience and process knowledge and are widely used in the monitoring and optimization of complex industrial processes. Among them, data modeling methods centered around machine learning have garnered widespread attention due to their superior nonlinear modeling capabilities and generalization performance.

[0004] Broad Learning System (BLS), as an efficient machine learning method, is primarily used to process data with complex structures or high-dimensional features. BLS has a rich hidden layer space that encompasses both linear and nonlinear features of the input data, namely, feature nodes and enhancement nodes. This gives BLS powerful feature extraction capabilities and has been widely used in the field of soft sensing. However, traditional BLS focuses on modeling independent and identically distributed test and training sets, lacking the ability to adapt to multi-condition industrial processes. Therefore, it is particularly important to address the problem of unbalanced data distribution and explore and study how to enable soft sensing models to adapt to changes in working conditions and cope with data drift.

[0005] Transfer learning transfers knowledge learned in an old task (source domain) to a related but different task (target domain) to improve the performance of the new model. Domain adaptation methods based on parameter transfer, as a transfer learning method, map the data in the source and target domains into the same space, giving them the same distribution. This allows them to leverage the training data in the source domain to address the lack of training data in the target domain. The key to these methods is how to measure and reduce the distribution difference between the source and target domains.

[0006] To sum up, the existing soft measurement methods lack soft measurement models that can handle cross-operating condition problems with large distribution differences. Therefore, there is still room for further improvement in the prediction accuracy of penicillin concentration in the industrial penicillin fermentation process. Summary of the Invention

[0007] In order to solve the above problems, the present invention provides a domain adaptation width learning method for industrial penicillin concentration prediction based on parameter migration, and applies it to the soft measurement modeling of industrial penicillin fermentation process with multi-operating condition characteristics to achieve high-precision prediction of penicillin concentration in industrial penicillin fermentation process.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] In a first aspect, the present invention provides a method for predicting industrial penicillin concentration by domain adaptation width learning based on parameter migration, the method comprising:

[0010] Step 1: Obtain data in the bioreactor under different working conditions, and regard the samples under the working condition to be predicted as the target domain data {X t ,y t}, will be with X t Data with similar but different operating conditions are considered as source domain data {X s ,y s};

[0011] Step 2: Based on the width learning system and transfer learning method, a domain-adaptive width learning model DABLS-PT based on parameter transfer in cross-operating scenarios is constructed;

[0012] Step 3: Randomly initialize the parameters of the domain adaptation width learning model DABLS-PT;

[0013] Step 4: Use the width learning structure in the domain adaptation width learning model DABLS-PT to obtain the feature nodes and enhanced nodes of the source domain and target domain respectively;

[0014] Step 5: Merge the feature nodes and enhancement nodes of the source domain and target domain to obtain the hidden layer H s and H t ;

[0015] Step 6: Construct the objective function of the domain adaptation width learning model DABLS-PT and calculate the optimal migration parameter M and output parameter β of the domain adaptation width learning model DABLS-PT in an alternating optimization manner;

[0016] Step 7: Use the trained domain adaptation width learning model DABLS-PT to predict the industrial penicillin concentration under the new working conditions.

[0017] In one embodiment of the present invention, the domain adaptation width learning model DABLS-PT includes a source domain regressor, a target domain regressor, and a source domain-target domain conversion mapping; the step 4 specifically includes: extracting features from source domain and target domain samples to obtain feature nodes, and performing nonlinear transformation on the feature nodes to obtain enhanced nodes;

[0018] Utilizing the source-target domain conversion mapping to realize parameter association between the source domain regressor and the target domain regressor to achieve distribution alignment between the two domains, including:

[0019] Convert the source domain output parameters to the target domain, assuming that there is a linear relationship between the output parameters of the two domains:

[0020] β t =Mβ s (1)

[0021] in, are the source domain model parameters, are the target domain model parameters, is the conversion mapping parameter, n h Represents the number of hidden layer nodes; the transformation matrix M multiplies the source domain model parameter β s It represents the process of migrating the output parameters of the source domain to the target domain.

[0022] In one embodiment of the present invention, the objective function of the domain adaptation width learning model DABLS-PT in step 6 is:

[0023] J DABLS-PT =J e,s (X s ,y s ,β s )+J e,t (X t ,y t ,β t )+J d (X s ,β s ,X t ,β t )+J r (M,β s ) (2) Among them, J e,s (X s ,y s ,β s ) and J e,t (X t ,y t ,β t ) represent the loss of source domain and target domain respectively, J d (X s ,β s ,X t ,β t ) represents the distribution alignment regularization term between the source domain and the target domain, J r (M,β s ) is the output parameter of the domain adaptation width learning model DABLS-PT and the regularization term of the conversion mapping.

[0024] In one embodiment of the present invention, step 6 further includes source domain and target domain losses, and the optimization objectives of the source domain regressor and the target domain regressor are:

[0025]

[0026] Among them, P s and P t is the weight coefficient of the model for errors in different fields, represents the predicted value of the i-th source domain sample, The predicted value of the j-th target domain sample, n s Indicates the number of source domain data, n t Indicates the number of target domain data.

[0027] In one embodiment of the present invention, step 6 further includes a distribution alignment regularization term, which uses the maximum mean difference (MMD) criterion to measure the difference between the two distributions P and Q. In transfer learning, MMD is defined as:

[0028]

[0029] According to formula (4), the calculation formula of the distribution alignment regularization term of the source domain and target domain prediction output is:

[0030]

[0031] Wherein, φ(·) in formula (4) is the mapping function.

[0032] In one embodiment of the present invention, step 6 further includes output parameter and migration parameter regularization terms;

[0033] The domain adaptation width learning model DABLS-PT uses different output parameters β s and βt Regression modeling is performed on the source domain and target domain data; in order to prevent overfitting caused by the number of source domain samples being larger than the target domain, β s Regularization is performed; at the same time, to avoid negative migration, β is limited s and β t The domain adaptation width learning model DABLS-PT converts the source domain output parameters to the target domain through the conversion matrix M, and converts β s and β t Abbreviated as β and Mβ; the parameter regularization term of the domain adaptation width learning model DABLS-PT is designed as follows:

[0034]

[0035] Among them, P e and P f is the penalty coefficient, and ‖·‖1 is the L1 norm.

[0036] In one embodiment of the present invention, the overall optimization goal of the domain adaptation width learning model DABLS-PT is:

[0037]

[0038] Among them, the L1 norm is difficult to derive, so the marginal optimization theory is used to construct an alternative function to simplify it, and it is abbreviated as w: =β-Mβ, then there is The rest of formula (8) is expressed as Then formula (8) can be rewritten as

[0039] In one embodiment of the present invention, the optimization method of formula (8) is specifically as follows:

[0040] If formula (9) is selected as the replacement function, then -L(w) increases monotonically;

[0041]

[0042] in,

[0043] prove:

[0044]

[0045] When w=w *(t) When the equality is established, then -L(w)-Q(w|w *(t) ) takes the minimum value; further:

[0046]

[0047] So -L(w) is monotonically increasing, and the proof is complete;

[0048] ‖β-Mβ‖1 is used in optimization (β-Mβ) T Q(β-Mβ) is used as a substitute, and formula (8) is rewritten as:

[0049]

[0050] In one embodiment of the present invention, the auxiliary variables of the industrial penicillin fermentation process include cooling water flow, dissolved oxygen concentration, bulk volume, bulk weight, percentage of carbon dioxide in exhaust gas, percentage of oxygen in exhaust gas, and carbon absorption rate.

[0051] In a second aspect, the present invention provides application of the method in a fermentation process.

[0052] The beneficial effects of the present invention are:

[0053] When the operating conditions of an industrial penicillin fermentation process change, the data distributions of the new and old conditions become inconsistent, leading to mismatches in the soft-sensing model. Furthermore, the scarcity of samples for the new operating conditions makes it difficult to accurately establish a new soft-sensing model. To address this issue, the present invention provides a domain-adaptive wide-learning method for industrial penicillin concentration prediction based on parameter transfer to improve the model's adaptability across operating scenarios. Within the framework of a wide-learning system, the present invention designs a learnable source-target domain parameter transformation mapping to minimize distribution differences, align the predicted output distributions of the target and source domains, and achieve cross-domain shared knowledge transfer. The present invention constructs output parameter regularization terms, transfer parameter regularization terms, and a distribution alignment regularization term based on the maximum mean difference criterion to avoid negative knowledge transfer and overfitting in the cross-domain soft-sensing model. The present invention also proposes an alternating optimization algorithm for model parameters to achieve adaptive learning of output and transfer parameters. The performance of the proposed model is validated using industrial penicillin fermentation and three-phase flow industrial processes. Compared with other transfer soft-sensing methods, the results demonstrate that the proposed model achieves higher prediction accuracy for cross-domain soft-sensing problems, providing a powerful technical support for the optimization and control of practical industrial processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0055] Figure 1 It is a structural diagram of the DABLS-PT model of the present invention.

[0056] Figure 2 It is a schematic structural diagram of the bioreactor used in the present invention.

[0057] Figure 3 It is a three-dimensional data distribution diagram of different working conditions of the present invention.

[0058] Figure 4(a) shows the prediction effect of each model on penicillin concentration under working condition 1→2.

[0059] Figure 4(b) shows the prediction effect of each model on penicillin concentration under working conditions 1→3.

[0060] Figure 5(a) is a scatter plot of the predicted penicillin concentrations of each model under working condition 1→2.

[0061] Figure 5(b) is a scatter plot of the predicted penicillin concentrations of each model under working conditions 1→3.

[0062] Figure 6(a) is a box plot of the error in predicting penicillin concentration for each model under working condition 1→2.

[0063] Figure 6(b) is a box plot of the error in predicting penicillin concentration for each model under working condition 1→3. DETAILED DESCRIPTION

[0064] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0065] Example 1

[0066] This embodiment provides a method for predicting industrial penicillin concentration using domain adaptation width learning based on parameter migration. The method is used to predict the concentration of penicillin in an industrial penicillin fermentation process under new operating conditions. The method includes:

[0067] Step 1: Obtain data in the bioreactor under different working conditions, and regard the samples under the working condition to be predicted as the target domain data {X t ,y t}, will be with X t Data with similar but different operating conditions are considered as source domain data {X s ,y s};

[0068] Step 2: Based on the width learning system and transfer learning method, a domain-adaptive width learning model DABLS-PT based on parameter transfer in cross-operating scenarios is constructed;

[0069] Step 3: Randomly initialize the parameters of the domain adaptation width learning model DABLS-PT;

[0070] Step 4: Use the width learning structure in the domain adaptation width learning model DABLS-PT to obtain the feature nodes and enhanced nodes of the source domain and target domain respectively;

[0071] Step 5: Merge the feature nodes and enhancement nodes of the source domain and target domain to obtain the hidden layer H s and H t ;

[0072] Step 6: Construct the objective function of the domain adaptation width learning model DABLS-PT and calculate the optimal migration parameter M and output parameter β of the domain adaptation width learning model DABLS-PT in an alternating optimization manner;

[0073] Step 7: Use the trained domain adaptation width learning model DABLS-PT to predict the industrial penicillin concentration under the new working conditions.

[0074] Example 2

[0075] This example provides a method for predicting industrial penicillin concentration using domain-adaptive breadth learning (DABLS-PT) based on parameter migration in cross-operating scenarios. This method constructs a domain-adaptive breadth learning model (DABLS-PT) and then uses this model to predict penicillin concentration during industrial penicillin fermentation. The model is first introduced as follows:

[0076] Given modeling data: source domain data {X s ,y s}, target domain data {X t ,y t}, where the target domain data is insufficient to train the new soft sensor model. The Domain Adaptive Broad Learning System with Parameter Transferring (DABLS-PT) model structure is as follows: Figure 1 As shown in Figure 2, it consists of three parts: a source domain regressor, a target domain regressor, and a source-target domain conversion mapping. First, feature extraction is performed on source and target domain samples to obtain feature nodes. Then, a nonlinear transformation is performed on the feature nodes to obtain enhanced nodes, which are then combined as a hidden layer. Finally, the source-target domain conversion mapping is used to associate the parameters of the source and target domain regressors, thereby achieving distribution alignment between the two domains.

[0077] DABLS-PT aims to transfer knowledge gained from training in past working conditions to modeling tasks in new working conditions. To complete the target domain modeling task, the transfer target is the BLS regressor parameters in the source domain. To improve model performance and enable it to better adapt to the data distribution of the new working condition, DABLS-PT, based on the idea of ​​parameter transfer, designs a trainable parameter conversion mapping to convert the output parameters of the source domain to the target domain. This assumes a linear relationship between the output parameters of the two domains:

[0078] β t =Mβs (1)

[0079] in, are the source domain model parameters, are the target domain model parameters, is the conversion mapping parameter, n h Represents the number of hidden layer nodes. The transformation matrix M multiplied by the source domain model parameter β s This represents the process of migrating the output parameters of the source domain to the target domain. This allows the source and target domain regressors to achieve knowledge transfer by sharing parameter space, allowing the soft sensing model to better utilize source domain knowledge and improve learning performance in the target domain.

[0080] To overcome the data distribution differences between the source and target domains, a distribution alignment regularization term is constructed. To prevent model overfitting and negative transfer, regularization terms for output parameters and transformation mapping are designed. Therefore, the optimization objective function of the DABLS-PT model is:

[0081] J DABLS-PT =J e,s (X s ,y s ,β s )+J e,t (X t ,y t ,β t )+J d (X s ,β s ,X t ,β t )+J r (M,β s ) (2)

[0082] Among them, J e,s (X s ,y s ,β s ) and J e,t (X t ,y t ,β t ) represent the loss of source domain and target domain respectively, J d (X s ,β s ,X t ,β t ) represents the distribution alignment regularization term between the source domain and the target domain, J r (M,β s ) is the output parameter of the DABLS-PT model and the regularization term of the transformation mapping.

[0083] The specific definition of the optimization objective function of the DABLS-PT model is as follows:

[0084] (1) Source domain and target domain losses

[0085] As a soft sensor model, DABLS-PT first considers the prediction accuracy of the target domain task. However, due to the extremely small number of target domain samples, it is necessary to use source domain data to assist the model learning. Therefore, the optimization objectives of the source domain and target domain regressors are:

[0086]

[0087] Among them, P s and P t is the weight coefficient of the model for errors in different fields, represents the predicted value of the i-th source domain sample, The predicted value of the j-th target domain sample, n s Indicates the number of source domain data, n t Indicates the number of target domain data.

[0088] (2) Distribution alignment regularization term

[0089] The core goal of transfer learning is to reduce the data distribution difference between the source domain and the target domain. The basis for this is to reasonably measure the distribution difference between different domains. To this end, this paper introduces the Maximum Mean Discrepancy (MMD) criterion to measure the difference between two distributions P and Q. In transfer learning, MMD is defined as:

[0090]

[0091] Where φ(·) is the mapping function.

[0092] According to the MMD criterion, the distribution difference between the source domain and the target domain prediction data can be effectively quantified, and then this difference can be reduced to achieve high-performance cross-domain soft sensing modeling. According to formula (4), the calculation formula for the distribution alignment regularization term of the source domain and target domain prediction output is:

[0093]

[0094] (3) Output parameter and migration parameter regularization

[0095] The DABLS-PT model uses different output parameters β s and β t Regression modeling is performed on the source domain and target domain data. In order to prevent overfitting caused by the number of source domain samples being much larger than the target domain, β s Regularization is performed; at the same time, to avoid negative migration, it is necessary to limit β s and β tThe DABLS-PT model transforms the source domain output parameters into the target domain through the transformation matrix M, as shown in formula (1). For the convenience of representation, the present invention will hereinafter refer to β s and β t Abbreviated as β and Mβ. The parameter regularization term of the DABLS-PT model is designed as follows:

[0096]

[0097] Among them, P e and P f is the penalty coefficient, and ‖·‖1 is the L1 norm.

[0098] Combining the above three loss functions, the overall optimization goal of the proposed DABLS-PT model is:

[0099]

[0100] Among them, the L1 norm is difficult to derive, so the marginal optimization theory is used to construct an alternative function to simplify it, and it is abbreviated as w: =β-Mβ, then there is The rest of formula (8) is expressed as Then formula (8) can be rewritten as The optimization method is given by Proposition 1.

[0101] Proposition 1: If equation (9) is selected as the replacement function, then -L(w) is monotonically increasing.

[0102]

[0103] in,

[0104] prove:

[0105]

[0106] When w=w *(t) When the equality is established, then -L(w)-Q(w|w *(t) ) takes the minimum value. Further:

[0107]

[0108] Therefore, -L(w) is monotonically increasing, and Proposition 1 is proved.

[0109] Therefore, ‖β-Mβ‖1 can be replaced by (β-Mβ) in the optimization. T Q(β-Mβ) is used as a substitute, and formula (8) can be rewritten as:

[0110]

[0111] In order to verify the prediction effect of the DABLS-PT model proposed in this invention in the process of industrial penicillin fermentation, Figure 2 The bioreactor shown in the figure is used as an example for the experiment. Figure 2 As shown in Figure 1, it describes the metabolic activities of penicillin bacteria as they grow and synthesize antibiotics under environmental conditions such as culture medium, pH, and temperature. In actual penicillin fermentation processes, the inability to effectively monitor key variables such as product concentration, cell concentration, and substrate concentration using online instrumentation poses challenges in control and optimization. To effectively ensure product quality during penicillin fermentation, soft sensing techniques used in industrial processes are employed to predict these key variables online. To establish the soft sensing model, penicillin concentration was selected as the key variable for prediction, and seven related variables were selected as auxiliary input variables, as shown in Table 1. By setting different initial reaction conditions for the penicillin fermentation process, data from different operating conditions was generated. Each batch of fermentation lasted between 168 and 232 hours. In this experiment, data from three different operating conditions were collected: 1,130 data sets for operating condition 1, 1,150 data sets for operating condition 2, and 968 data sets for operating condition 3. Figure 3 The data distribution difference plot is drawn by selecting the first three-dimensional data. The data is the normalized result. Considering that all three operating conditions can be used as source conditions, in case 1, operating condition 1 is used as the source condition, and operating conditions 2 and 3 are used as target conditions; in case 2, operating condition 2 is used as the source condition, and operating conditions 1 and 3 are used as target conditions; in case 3, operating condition 3 is used as the source condition, and operating conditions 1 and 2 are used as target conditions.

[0112] Table 1 Input variables in industrial penicillin process

[0113]

[0114] For a more comprehensive comparison and analysis, the prediction results and prediction errors of the BLS(S), BLS(T), BLS(S+T), DAELM, DAPT, DAMRRWNN, and the proposed DABLS-PT model are visualized (using Case 1 as an example), as shown in Figures 4(a), 4(b), 5(a), 5(b), 6(a), and 6(b). Table 2 also provides the evaluation metrics for penicillin concentration prediction for each model. Overall, regardless of the data selection used in the training set, the basic BLS model exhibits very poor tracking error between the predicted values ​​and the true curve under the new operating condition. The model even fails when predicting Condition 3. Table 2 also shows that the traditional BLS model exhibits significant deviations in various evaluation metrics and significantly lower prediction performance when predicting conditions with significant data distribution differences. This is due to the different data distributions between the new and old operating conditions, and BLS is unable to achieve a balanced distribution between the two domains. In addition, although the predicted values ​​of the models DAELM, DAPT, and DAMRRWNN can track the true values ​​with the addition of the relevant domain adaptation algorithm, it can be clearly seen from Figures 5(a), 5(b), 6(a), and 6(b) that these three models still have significant prediction errors. The MAE indicator in the table more intuitively reflects that these three models are significantly inferior to the algorithm proposed in this invention. The DABLS-PT model proposed in this invention can achieve parameter migration between the source domain and the target domain, and better achieve distribution alignment between the two domains. Whether from the prediction result graph, error comparison graph, or evaluation index comparison table, it reflects good prediction performance.

[0115] Table 2 Comparison of the prediction evaluation indicators of penicillin concentration by each model

[0116]

[0117] In summary, the present invention provides a domain adaptation width learning industrial penicillin concentration prediction method based on parameter migration. The method is based on BLS and ensures calculation accuracy while improving the calculation efficiency of the model. The present invention designs a learnable parameter conversion mapping to project the output parameters of the source domain and target domain into the same parameter space, minimizing the data distribution differences under different domains. The present invention designs distribution alignment regularization terms for the source domain and target domain based on MMD, thereby improving the migration efficiency of data between domains. Output parameter regularization terms and migration parameter regularization terms are constructed in the objective function to avoid overfitting of the model and negative knowledge transfer. The present invention designs a model parameter optimization algorithm to alternately optimize the output parameters and migration parameters and adaptively calculate the optimal model parameters. The present invention verifies the performance of the proposed model through the industrial penicillin fermentation process and the three-phase flow industrial process. The results show that compared with other migration soft measurement methods, the proposed model has higher prediction accuracy when dealing with cross-domain soft measurement problems, providing a strong technical guarantee for the optimization control of actual industrial processes.

[0118] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

[0119] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting industrial penicillin concentration by domain adaptation width learning based on parameter migration, characterized in that: The method comprises: Step 1: Obtain data in the bioreactor under different working conditions, and regard the samples under the working condition to be predicted as the target domain data {X t ,y t }, will be with X t Data with similar but different operating conditions are considered as source domain data {X s ,y s }; Step 2: Based on the width learning system and transfer learning method, a domain-adaptive width learning model DABLS-PT based on parameter transfer in cross-operating scenarios is constructed; Step 3: Randomly initialize the parameters of the domain adaptation width learning model DABLS-PT; Step 4: Use the width learning structure in the domain adaptation width learning model DABLS-PT to obtain the feature nodes and enhanced nodes of the source domain and target domain respectively; Step 5: Merge the feature nodes and enhancement nodes of the source domain and target domain to obtain the hidden layer H s and H t ; Step 6: Construct the objective function of the domain adaptation width learning model DABLS-PT and calculate the optimal migration parameter M and output parameter β of the domain adaptation width learning model DABLS-PT in an alternating optimization manner; Step 7: Use the trained domain adaptation width learning model DABLS-PT to predict the industrial penicillin concentration under the new working conditions.

2. The method according to claim 1, characterized in that The domain adaptation width learning model DABLS-PT includes a source domain regressor, a target domain regressor, and a source domain-target domain conversion mapping; the step 4 specifically includes: extracting features from source domain and target domain samples to obtain feature nodes, and performing nonlinear transformation on the feature nodes to obtain enhanced nodes; Utilizing the source-target domain conversion mapping to realize parameter association between the source domain regressor and the target domain regressor to achieve distribution alignment between the two domains, including: Convert the source domain output parameters to the target domain, assuming that there is a linear relationship between the output parameters of the two domains: β t ＝Mβ s (1) in, are the source domain model parameters, are the target domain model parameters, is the conversion mapping parameter, n h Represents the number of hidden layer nodes; the transformation matrix M multiplies the source domain model parameter β s It represents the process of migrating the output parameters of the source domain to the target domain.

3. The method according to claim 2, characterized in that The objective function of the domain adaptation width learning model DABLS-PT in step 6 is: J DABLS-PT =J e,s (X s ,y s ,b s )+J e,t (X t ,y t ,b t )+J d (X s ,b s ,X t ,b t )+J r (M,b s ) (2) Among them, J e,s (X s ,y s ,β s ) and J e,t (X t ,y t ,β t ) represent the loss of source domain and target domain respectively, J d (X s ,β s ,X t ,β t ) represents the distribution alignment regularization term between the source domain and the target domain, J r (M,β s ) is the output parameter of the domain adaptation width learning model DABLS-PT and the regularization term of the conversion mapping.

4. The method according to claim 3, characterized in that Step 6 also includes source domain and target domain losses. The optimization objectives of the source domain regressor and the target domain regressor are: Among them, P s and P t is the weight coefficient of the model for errors in different fields, represents the predicted value of the i-th source domain sample, The predicted value of the j-th target domain sample, n s Indicates the number of source domain data, n t Indicates the number of target domain data.

5. The method according to claim 4, characterized in that Step 6 also includes a distribution alignment regularization term, which uses the maximum mean difference (MMD) criterion to measure the difference between the two distributions P and Q. In transfer learning, MMD is defined as: According to formula (4), the calculation formula of the distribution alignment regularization term of the source domain and target domain prediction output is: Wherein, φ(·) in formula (4) is the mapping function.

6. The method according to claim 5, characterized in that Said step 6 also includes output parameter and migration parameter regularization terms; The domain adaptation width learning model DABLS-PT uses different output parameters β s and β t Perform regression modeling on source and target domain data; In order to prevent overfitting caused by the number of source domain samples being larger than the target domain, β s Regularization is performed; at the same time, to avoid negative migration, β is limited s and β t The domain adaptation width learning model DABLS-PT converts the source domain output parameters to the target domain through the conversion matrix M, and converts β s and β t Abbreviated as β and Mβ; the parameter regularization term of the domain adaptation width learning model DABLS-PT is designed as follows: Among them, P e and P f is the penalty coefficient, and ‖·‖1 is the L1 norm.

7. The method according to claim 6, characterized in that The overall optimization goal of the domain adaptation width learning model DABLS-PT is: Among them, the L1 norm is difficult to derive, so the marginal optimization theory is used to construct an alternative function to simplify it, and it is simply written as w: =β-Mβ, then there is The rest of formula (8) is expressed as Then formula (8) can be rewritten as 8. The method according to claim 7, characterized in that The optimization method of formula (8) is specifically as follows: If formula (9) is selected as the replacement function, then -L(w) increases monotonically; in, prove: When w=w *(t) When the equality is established, then -L(w)-Q(w|w *(t) ) takes the minimum value; further: So -L(w) is monotonically increasing, and the proof is complete; ‖β-Mβ‖1 is used in optimization (β-Mβ) T Q(β-Mβ) is used as a substitute, and formula (8) is rewritten as:

9. The method according to claim 1, characterized in that Auxiliary variables of the industrial penicillin fermentation process include cooling water flow, dissolved oxygen concentration, bulk volume, bulk weight, percentage of carbon dioxide in exhaust gas, percentage of oxygen in exhaust gas, and carbon absorption rate.

10. Use of the method according to any one of claims 1 to 9 in a fermentation process.

Images