Empirically-derived microbial manufacturing process cell concentration sensing methods and systems

Abstract

The application discloses a kind of microbial manufacturing process bacteria concentration perception method and system based on experience transmission, method includes: according to the historical data of microbial manufacturing process, construct bacteria concentration prediction general model, obtains its model parameters based on source system historical data training and establishes target system model structure;By introducing latent variable quantification model difference between source system and target system, a unified experience transmission probability framework is constructed;Based on the framework, the iterative algorithm combining expectation and gradient descent is used to estimate the target system parameters until convergence;The converged parameters are substituted into the general model to construct an online soft sensor, and process auxiliary variable data is collected in real time and the bacteria concentration perception value is output.The application effectively suppresses negative transfer by explicitly modeling the differences between systems, and in the industrial scenario where data is scarce and distribution is inconsistent, it realizes robust, accurate and real-time perception of bacteria concentration.

Description

Technical Field

[0001] This invention relates to the field of industrial process status monitoring and measurement technology, and in particular to a method and system for sensing cell concentration in a microbial manufacturing process based on experience transfer. Background Technology

[0002] As a crucial bioreaction process, the real-time and accurate sensing of cell concentration in microbial manufacturing is essential for ensuring stable process operation, improving product yield, and achieving intelligent regulation. However, existing cell concentration monitoring technologies still face many challenges in practical applications. Currently, two main methods are employed: one is laboratory detection methods based on offline sampling, which, while highly accurate, suffers from significant lag and cannot support real-time closed-loop optimization and control; the other is data-driven soft measurement modeling methods, which can achieve online inference, but their performance heavily relies on a large number of high-quality, uniformly distributed training samples and is extremely sensitive to non-ideal data characteristics such as process noise, missing data, and batch-to-batch fluctuations, resulting in limited generalization ability in real-world industrial scenarios.

[0003] In recent years, to mitigate the impact of data scarcity or inconsistent distribution on modeling accuracy, transfer learning and experience transfer techniques have been gradually introduced into the field of microbial process modeling. These techniques aim to utilize prior knowledge from historical batches, similar systems, or laboratory-scale data to assist in the efficient modeling of the target system. The source system typically refers to a model system built based on historical or similar process data, while the target system refers to the actual industrial process system currently being monitored. The source domain corresponds to the data distribution represented by the source system (e.g., historically stable batches), and the target domain corresponds to the data distribution of the target system (e.g., currently operating batches). However, existing transfer methods still face significant challenges when applied to microbial manufacturing processes. Most methods have high computational complexity, rely on large-scale matrix operations or iterative optimization, and are difficult to adapt to high-dimensional, dynamic real-time industrial scenarios. More importantly, these methods are sensitive to differences in data distribution between the source and target systems. When the operating conditions or characteristics of the source and target domains differ significantly, forcibly transferring knowledge may cause the model to learn incorrect correlated or irrelevant features in the target task, leading to a decline in performance rather than an improvement. This phenomenon is known as "negative transfer." Negative transfer not only fails to improve the model accuracy of the target system by leveraging historical experience, but may also degrade the perception results due to knowledge misguidance, thus severely limiting the reliable application of such technologies in complex and ever-changing industrial scenarios. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the problems of existing technologies that rely on a large amount of labeled data, are sensitive to differences in data distribution, have high computational complexity and are prone to negative transfer, resulting in low accuracy of bacterial concentration sensing, poor real-time performance and weak generalization ability under limited sample and non-ideal data conditions.

[0005] To address the aforementioned technical problems, this invention provides a method for sensing cell concentration in a microbial manufacturing process based on experience transfer, comprising the following steps: S1: Preprocess historical data of the microbial manufacturing process, construct a general model for predicting cell concentration, train the model parameters of the source system based on historical data of the source system, and establish the model structure of the target system; S2: Based on the general model for predicting bacterial cell concentration, the model differences between the source system and the target system are quantified, and a unified empirical transfer probability framework is constructed by introducing latent variables; S3: Based on the aforementioned empirical transfer probability framework, estimate the target system parameters through iterative calculation until the target system model parameters converge. S4: Substitute the converged target system model parameters into the general model to construct an online soft sensor; use the online soft sensor to collect process auxiliary variable data of the target system in real time, and calculate and output the real-time sensing value of bacterial cell concentration based on the collected data.

[0006] In one embodiment of the present invention, in step S1, the general model for predicting bacterial cell concentration is a linear dynamic model, specifically: Source system model: , Target system model: , in, and These are the cell concentration output vectors for the source system and the target system, respectively. and This is the auxiliary variable information matrix for the corresponding system. and Let be the vector of model parameters to be estimated. and Let be independent and identically distributed Gaussian noise vectors, with covariance matrices of respectively and .

[0007] In one embodiment of the present invention, in step S2, based on the general model for predicting bacterial cell concentration, the method for quantifying the model differences between the source system and the target system by introducing latent variables to construct a unified empirical transfer probability framework is as follows: The source system model parameters obtained from the training of the general model for predicting bacterial cell concentration are used. As prior knowledge, a latent variable is defined. Characterizing the source system model parameters and the target system model parameters to be estimated The differences between them; reconstructing the observation model of the source system into one based on the parameters of the target system. With latent variables The common form of expression is represented as: ; By associating the source system and the target system in the same probabilistic model, we obtain an empirical transfer probabilistic framework that can be used for knowledge transfer and joint parameter estimation.

[0008] In one embodiment of the present invention, the latent variable It is assumed that the prior distribution follows a Gaussian distribution with a mean of zero, i.e. ,in This is the preset covariance matrix.

[0009] In one embodiment of the present invention, step S3, based on the empirical transfer probability framework, estimates the target system parameters through iterative calculation until the target system model parameters converge, as follows: The sample transfer identification algorithm, which combines expectation and gradient descent, is executed, alternating between the following two steps in each iteration: Perform the expectation calculation step: based on the current estimated values ​​of the target system parameters and all observed data, calculate the posterior probability distribution of the latent variables, and construct the expectation function of the complete data log-likelihood with respect to the posterior distribution, denoted as the Q function; The gradient descent step is performed to calculate the gradient of the Q function with respect to the target system parameters. The target system parameters are updated along the gradient direction of the Q function with respect to the target system parameters. The step size used during the update is dynamically adjusted according to the spectral radius criterion to ensure the convergence of the iteration process. Repeat the above expected value calculation and gradient descent steps until the change in the target system parameters is less than the preset convergence threshold. At this point, the target system model parameters are considered to have converged.

[0010] In one embodiment of the present invention, the step size The adjustment must meet the following conditions: ,in Map the gradient to the Jacobian matrix at the current iteration point. Spectral radius.

[0011] In one embodiment of the present invention, the Q function is defined as the expectation of the complete data log-likelihood with respect to the posterior distribution of the latent variables: , in This is the observation dataset.

[0012] In one embodiment of the present invention, the method further includes adaptive model updating, specifically: monitoring the prediction error of the online soft sensor output results, and when the average value of the prediction error within a preset sliding window continuously exceeds a preset threshold, correcting the target system model parameters using the latest collected data.

[0013] This invention also provides a cell concentration sensing system for microbial manufacturing processes based on experience transfer, comprising the following modules: The model building module is used to preprocess historical data of the microbial manufacturing process, build a general model for predicting cell concentration, train the model parameters of the source system based on historical data of the source system, and establish the model structure of the target system. The difference quantification and empirical modeling module is used to quantify the model differences between the source system and the target system based on the general model for predicting bacterial cell concentration, and to construct a unified empirical transfer probability framework by introducing latent variables. The parameter iterative estimation module is used to estimate the target system parameters through iterative calculation based on the empirical transfer probability framework until the target system model parameters converge. The sensing and output module is used to substitute the converged target system model parameters into the general model to construct an online soft sensor; and to use the online soft sensor to collect process auxiliary variable data of the target system in real time, and to calculate and output the real-time sensing value of bacterial cell concentration based on the collected data.

[0014] The present invention also provides a computer storage medium storing a computer software product, the computer software product including a plurality of instructions for causing a computer device to execute the experience-based microbial manufacturing process cell concentration sensing method.

[0015] The technical solution of the present invention has the following advantages over the prior art: The microbial cell concentration sensing method based on experience transfer described in this invention introduces latent variables for explicit quantification and compensates for model differences between the source and target systems, constructing an interpretable experience transfer probability framework. This effectively suppresses negative transfer problems caused by inconsistent data distribution, significantly improving the robustness and accuracy of sensing in limited sample and high-noise environments. A lightweight iterative strategy combining expectation-gradient descent is employed to replace complex high-dimensional matrix operations in traditional methods, greatly reducing computational burden and meeting the real-time requirements of industrial processes. Simultaneously, prior knowledge from historical data is fully utilized, reducing dependence on labeled data of the target system and lowering modeling costs and time. An adaptive step size adjustment mechanism based on the spectral radius criterion ensures stable convergence of the parameter estimation process. Furthermore, the system in this invention possesses online monitoring and adaptive update capabilities, continuously tracking process changes. Thus, in real-world industrial scenarios with scarce data and fluctuating operating conditions, it achieves efficient, accurate, and stable sensing of cell concentration, providing reliable technical support for intelligent monitoring and optimized control of microbial manufacturing processes. Attached Figure Description

[0016] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0017] Figure 1 This is a schematic flowchart of the cell concentration sensing method for microbial manufacturing process based on experience transfer provided in this embodiment of the invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0019] Example 1: like Figure 1 As shown, this invention provides a method for sensing cell concentration in a microbial manufacturing process based on experience transfer, comprising the following steps: S1: Preprocess historical data of the microbial manufacturing process, construct a general model for predicting cell concentration, train the model parameters of the source system based on historical data of the source system, and establish the model structure of the target system; S2: Based on the general model for predicting bacterial cell concentration, the model differences between the source system and the target system are quantified, and a unified empirical transfer probability framework is constructed by introducing latent variables; S3: Based on the aforementioned empirical transfer probability framework, estimate the target system parameters through iterative calculation until the target system model parameters converge. S4: Substitute the converged target system model parameters into the general model to construct an online soft sensor.

[0020] This invention proposes a microbial cell concentration sensing method based on experience transfer in microbial manufacturing processes. This method preprocesses historical data to construct a general model for predicting cell concentration, trains the model parameters based on source system data, and then establishes the model structure for the target system. Furthermore, by quantifying the model differences between the source and target systems, latent variables are introduced to construct a unified experience transfer probability framework, and iterative calculations are used to estimate the target system parameters until convergence. Finally, the converged parameters are used to construct an online soft sensor, enabling real-time cell concentration sensing based on real-time auxiliary variable data. This approach effectively utilizes historical experience, significantly reducing reliance on large amounts of labeled data from the target system. Even in industrial scenarios with scarce or inconsistent data distribution, it achieves efficient and stable online cell concentration estimation, improving the real-time nature of sensing and the model's generalization ability, providing reliable support for precise monitoring and optimized control of microbial manufacturing processes.

[0021] In this embodiment, in step S1, historical data of the microbial manufacturing process is preprocessed to construct a general model for predicting cell concentration. Based on the historical data of the source system, the model parameters of the source system are obtained through training, and the model structure of the target system is established.

[0022] Specifically, in this embodiment, for the target process (such as fermentation), key measurable process variables are selected as auxiliary variables. These variables typically include fermentation temperature. pH value, dissolved oxygen concentration substrate concentration These factors directly affect bacterial growth and metabolism, and form the information basis for constructing soft measurement models.

[0023] The collected historical data from the source system, such as data from historically stable batches or laboratory-scale experiments, needs to undergo preprocessing operations such as outlier removal, missing value imputation, and standardization. These operations can eliminate differences in data units, improve data quality, and provide a reliable data foundation for subsequent model training.

[0024] Based on the preprocessed data, a general model structure for predicting bacterial cell concentration is constructed.

[0025] In this embodiment, a linear dynamic model is selected as the basic framework. This model has a clear structure, strong parameter interpretability, and can effectively capture the dynamic relationship between auxiliary variables and cell concentration output, while also taking into account computational efficiency, making it suitable for the real-time requirements of industrial scenarios. The linear dynamic model provides mathematical definitions for both the source system and the target system.

[0026] The source system model is defined as follows: , The target system model is defined as follows: , in, and These are the cell concentration output vectors for the source system and the target system, respectively. and This is the auxiliary variable information matrix for the corresponding system, where each row corresponds to the observed value of the auxiliary variable at a sampling time. The vector of model parameters to be estimated reflects the weights of the process variables on the cell concentration. and A zero-mean Gaussian white noise vector is used to characterize modeling error and process random disturbances, and their covariance matrices are respectively... and .

[0027] The source and target systems use the same model structure, reflecting the common foundation between processes; however, differences in parameters and noise characteristics are allowed. The allowance for variation reflects the specificity of processes across different batches, scales, or operating conditions. This design enables the universal model to possess both structural uniformity and flexibility to adapt to changes between systems.

[0028] Furthermore, the model parameters are obtained by training based on historical data from the source system. This is achieved using preprocessed source system data. Using the least squares parameter identification method, the estimated values ​​of the source system parameters can be calculated: .

[0029] This step aims to extract stable, reusable process experience knowledge from historical data, and to use parameters... It is encapsulated in the form of prior information for subsequent experience transfer.

[0030] At the same time, a model with the exact same structure as the target system is established, but its parameters are different. The initial state is unknown and needs to be estimated using a subsequently constructed empirical transfer probability framework.

[0031] This completes the construction of a general prediction model, the extraction of experience from the source system, and the establishment of the target system model structure, providing a unified model foundation and structured prior support for the next step of quantifying differences between systems and realizing knowledge transfer.

[0032] Furthermore, in step S2, based on the general model for predicting bacterial cell concentration, the model differences between the source system and the target system are quantified, and a unified probabilistic framework for experience transfer is constructed by introducing latent variables. Step S2 aims to explicitly model and quantify the model differences between the source system and the target system, thereby transforming the experience transfer problem into a structured probabilistic inference task and providing a unified and interpretable mathematical framework for subsequent parameter estimation.

[0033] Specifically, in this embodiment, a latent variable is introduced. Specifically used to characterize source system model parameters With the target system model parameters to be estimated The differences between them.

[0034] Latent variables The core idea is to explicitly model the parameter offset between the source system and the target system, rather than treating it as unexplainable noise or error.

[0035] Based on the source system parameters obtained through training As prior knowledge, the observation model of the source system is reconstructed to match the parameters of the target system. Connect them. The reconstructed source system model is represented as: .

[0036] This reconstruction indicates that the observed data of the source system Not only can it be derived from the parameters of the source system itself Explanation can also be derived from the target system parameters. Subtract one difference item To explain. When When the source system and target system parameters are completely identical, experience can be directly transferred without adjustment; when In this case, the parameter adjustments required for knowledge transfer are quantified.

[0037] In order to perform rigorous probabilistic inference within the empirical transitive probability framework, it is necessary to study the latent variables. Assign a prior distribution. In this embodiment, assume the latent variable... It follows a Gaussian prior distribution with a mean of zero, that is: , in It is a pre-defined covariance matrix that reflects prior knowledge of the magnitude of differences between systems. Setting the mean to zero indicates that, in the absence of data for the target system, the expected value of the difference between the source and target systems can be initially assumed to be zero, i.e., a tendency to believe that historical experience is transferable; while the covariance matrix... This controls the degree of uncertainty regarding this prior belief, and its value can be set based on domain knowledge or fluctuations between historical batches.

[0038] By introducing latent variables and its prior distribution, source system model With the target system model They are incorporated into a single, unified probabilistic model. The probabilistic model incorporates observations from the source system. Observation of the target system Target system parameters and differential latent variables These elements, when connected, form a complete Bayesian network. This unified empirical transitive probability framework successfully transforms the problem of using source system knowledge to assist in target system modeling into a problem based on observed data. Under the condition, for the target parameter and latent variables The problem of probabilistic inference for joint posterior estimation.

[0039] The construction of the empirical transitivity framework has the following important functions and effects: it utilizes latent variables... The explicit separation and quantification of system-specific differences enable the model to clearly distinguish shareable common knowledge (reflected in the model structure). and some parameters (in the middle) and the specific parts that need adjustment (reflected in) This provides a mechanistic guarantee for suppressing negative migration caused by distribution mismatch. Secondly, the Bayesian probability-based modeling approach naturally integrates prior knowledge (source system parameters) with current observations (target system data), achieving an organic combination of experience and data. Finally, the empirical transitivity framework lays a solid theoretical foundation for the subsequent design of efficient and stable parameter estimation algorithms. Iterative algorithms such as expectation-gradient descent can be used to perform rigorous and efficient inference calculations within this empirical transitivity framework.

[0040] Further, in step S3, after obtaining the empirical transitive probability framework, the target system parameters are estimated through iterative calculation until the target system model parameters converge. Step S3 is based on the constructed model containing latent variables. A unified probabilistic model is designed to utilize observational data from the source system. and partial observation data of the target system The optimal parameters of the target system are estimated through an iterative algorithm. .

[0041] In this embodiment, a sample migration identification algorithm (EGD-STI) combining Expectation and Gradient Descent is used. This algorithm alternates between two steps in each iteration: the Expectation step (E step) and the Gradient Descent step (GD step) until the parameters converge.

[0042] The core task of the expected step (E-step) is based on the current step number. The target system parameter estimate in the next iteration and all observation data Inferring latent variables The posterior probability distribution. In the model assumptions, noise... , and latent variable priors All are Gaussian distributed, therefore latent variables The posterior distribution of also follows a Gaussian distribution, i.e. Its posterior mean With covariance It can be obtained analytically through Bayes' theorem: , .

[0043] This calculation process essentially utilizes data from the source system. And current understanding of the target parameters To correct for differences between systems The estimate. This reflects the most likely parameter offset between the source and target systems under the current parameter settings. This quantifies the uncertainty of the estimate. Subsequently, based on this post-aperential distribution, a log-likelihood function for the complete data is constructed. Regarding the expectation of this posterior distribution, i.e., the Q function: .

[0044] The Q function combines the goodness of fit of the target system data (first term) with the degree of fit of the source system's empirical knowledge after considering the uncertainty of differences (second term), providing a clear optimization objective for subsequent parameter updates.

[0045] The task of the gradient descent step (GD step) is to replace the analytical maximization (M step) in the traditional expectation maximization (EM) algorithm, by moving along the Q function with respect to the objective parameters. The gradient direction is used to update the parameter estimates. The Q-function is calculated with respect to... gradient: .

[0046] The gradient consists of two parts: the first part Derived from the target system data, the parameters are adjusted to fit the current observations; Part Two Derived from the source system experience, its function is to shift the target parameters to the prior source parameters after difference correction (i.e., (In some form) directional traction. This structure enables an automatic trade-off between experience and current data.

[0047] The parameter update formula is: , in For the first The step size of the next iteration.

[0048] Step length The adaptive adjustment mechanism is crucial for ensuring the convergence and stability of the algorithm. In this embodiment, the step size is dynamically adjusted based on the spectral radius criterion. The Jacobian matrix of the gradient mapping is defined as follows: ,in To ensure iterative convergence, the step size must meet the following conditions: , in, It is a matrix spectral radius, It is a matrix The largest eigenvalue. In practical calculations, it can be estimated using numerical methods such as the power iteration method. The algorithm uses the spectral radius or the largest eigenvalue to set a conservative step size that satisfies the convergence condition. This mechanism avoids the oscillation or slow convergence problems that may be caused by a fixed step size, enabling the algorithm to quickly and smoothly approach the optimal solution.

[0049] Repeat the E-step and GD-step above, calculating the parameter update amount after each iteration. When the update amount of this parameter is less than the preset convergence threshold... (For example When the algorithm converges, the final estimated values ​​of the target system parameters are obtained. .

[0050] If noise covariance , If it is unknown, it can be embedded as a hyperparameter to be estimated into the aforementioned EM framework for synchronous updating.

[0051] Furthermore, in step S4, after obtaining the converged target system model parameters, these parameters are substituted into the general cell concentration prediction model to construct an online soft sensor suitable for the target system. This soft sensor, based on real-time collected auxiliary variable data from the fermentation process, such as temperature, pH, dissolved oxygen concentration, and substrate concentration, forms an auxiliary variable information vector for the current moment, and then calculates and outputs the real-time sensing value of the cell concentration.

[0052] To achieve continuous adaptation to dynamic changes in the process, this embodiment also includes a model adaptive update mechanism, which defines the average prediction error within the sliding window. .like Continuously exceeding the threshold This triggers a model update: using recent data as new samples, steps S2 to S4 are repeated partially or entirely to adjust the parameters. Online adaptive correction is used to track the slow time-varying characteristics of the process. The target system model parameters are re-estimated or incrementally updated using the latest acquired data, so that the model can track the slow time-varying operating conditions and maintain perception accuracy and robustness.

[0053] This allows for the effective use of historical experience and a reduction in reliance on labeled data of the target system during the initial modeling phase, while also enabling stable performance through online adaptation during long-term operation.

[0054] Through the above steps, this invention constructs a complete technology chain from probabilistic modeling, difference quantification, efficient iterative estimation to online application. Its core lies in explicitly modeling and conveying system differences through latent variables, and utilizing the EGD-STI algorithm to achieve an effective balance between computational complexity and estimation accuracy. This enables robust, accurate, and real-time sensing of bacterial concentration in industrial scenarios characterized by scarce and inconsistent data distribution.

[0055] Example 2: Based on the same inventive concept as Embodiment 1, the present invention also provides a cell concentration sensing system for a microbial manufacturing process based on experience transfer, used to implement the steps of the cell concentration sensing method for a microbial manufacturing process based on experience transfer described in Embodiment 1, including the following modules: The model building module is used to preprocess historical data of the microbial manufacturing process, build a general model for predicting cell concentration, train the model parameters of the source system based on historical data of the source system, and establish the model structure of the target system. The difference quantification and empirical modeling module is used to quantify the model differences between the source system and the target system based on the general model for predicting bacterial cell concentration, and to construct a unified empirical transfer probability framework by introducing latent variables. The parameter iterative estimation module is used to estimate the target system parameters through iterative calculation based on the empirical transfer probability framework until the target system model parameters converge. The sensing and output module is used to substitute the converged target system model parameters into the general model to construct an online soft sensor; and to use the online soft sensor to collect process auxiliary variable data of the target system in real time, and to calculate and output the real-time sensing value of bacterial cell concentration based on the collected data.

[0056] The model construction module, difference quantification and experience modeling module, parameter iterative estimation module, and sensing and result output module of the microbial manufacturing process cell concentration sensing system based on experience transfer proposed in this embodiment are respectively used to implement steps S1, S2, S3 and S4 in the microbial manufacturing process cell concentration sensing method based on experience transfer in Embodiment 1. To avoid redundancy, they will not be described in detail here.

[0057] Example 3: The present invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions for causing a computer device to execute the experience-based microbial manufacturing process cell concentration sensing method described in Embodiment 1.

[0058] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for sensing cell concentration in a microbial manufacturing process based on experience transfer, characterized in that, Includes the following steps: S1: Preprocess historical data of the microbial manufacturing process, construct a general model for predicting cell concentration, train the model parameters of the source system based on historical data of the source system, and establish the model structure of the target system; S2: Based on the general model for predicting bacterial cell concentration, the model differences between the source system and the target system are quantified, and a unified empirical transfer probability framework is constructed by introducing latent variables; S3: Based on the aforementioned empirical transfer probability framework, estimate the target system parameters through iterative calculation until the target system model parameters converge. S4: Substitute the converged target system model parameters into the general model to construct an online soft sensor; The online soft sensor is used to collect process auxiliary variable data of the target system in real time, and the real-time sensing value of bacterial cell concentration is calculated and output based on the collected data.

2. The microbial cell concentration sensing method based on experience transfer in the microbial manufacturing process according to claim 1, characterized in that: In step S1, the general model for predicting bacterial cell concentration is a linear dynamic model, specifically: Source system model: , Target system model: , in, and These are the cell concentration output vectors for the source system and the target system, respectively. and This is the auxiliary variable information matrix for the corresponding system. and Let be the vector of model parameters to be estimated. and Let be independent and identically distributed Gaussian noise vectors, with covariance matrices of respectively and .

3. The microbial cell concentration sensing method based on experience transfer in the microbial manufacturing process according to claim 1, characterized in that: In step S2, based on the general model for predicting bacterial cell concentration, the method for quantifying the model differences between the source system and the target system and constructing a unified empirical transfer probability framework by introducing latent variables is as follows: The source system model parameters obtained from the training of the general model for predicting bacterial cell concentration are used. As prior knowledge, a latent variable is defined. Characterizing the source system model parameters and the target system model parameters to be estimated The differences between them; reconstructing the observation model of the source system into one based on the parameters of the target system. With latent variables The common form of expression is represented as: ； By associating the source system and the target system in the same probabilistic model, we obtain an empirical transfer probabilistic framework that can be used for knowledge transfer and joint parameter estimation.

4. The microbial cell concentration sensing method based on experience transfer in the microbial manufacturing process according to claim 3, characterized in that: The latent variables It is assumed that the prior distribution follows a Gaussian distribution with a mean of zero, i.e. ,in This is the preset covariance matrix.

5. The microbial cell concentration sensing method based on experience transfer in the microbial manufacturing process according to claim 1, characterized in that: In step S3, the method for estimating the target system parameters through iterative calculation based on the empirical transfer probability framework until the target system model parameters converge is as follows: The sample transfer identification algorithm, which combines expectation and gradient descent, is executed, alternating between the following two steps in each iteration: Perform the expectation calculation step: based on the current estimated values ​​of the target system parameters and all observed data, calculate the posterior probability distribution of the latent variables, and construct the expectation function of the complete data log-likelihood with respect to the posterior distribution, denoted as the Q function; The gradient descent step is performed to calculate the gradient of the Q function with respect to the target system parameters. The target system parameters are updated along the gradient direction of the Q function with respect to the target system parameters. The step size used during the update is dynamically adjusted according to the spectral radius criterion to ensure the convergence of the iteration process. Repeat the above expected value calculation and gradient descent steps until the change in the target system parameters is less than the preset convergence threshold. At this point, the target system model parameters are considered to have converged.

6. The microbial cell concentration sensing method based on experience transfer in the microbial manufacturing process according to claim 5, characterized in that: The step size The adjustment must meet the following conditions: ,in The gradient mapping Jacobian matrix for the current iteration point Spectral radius.

7. The microbial cell concentration sensing method based on experience transfer in the microbial manufacturing process according to claim 5, characterized in that: The Q-function is defined as the expectation of the complete data log-likelihood with respect to the posterior distribution of the latent variables: ， in This is the observation dataset.

8. The microbial cell concentration sensing method based on experience transfer in the microbial manufacturing process according to claim 1, characterized in that: It also includes adaptive model updates, specifically: monitoring the prediction error of the online soft sensor output results, and when the average value of the prediction error within a preset sliding window continuously exceeds a preset threshold, using the latest collected data to correct the target system model parameters.

9. A cell concentration sensing system for a microbial manufacturing process based on experience transfer, characterized in that, Includes the following modules: The model building module is used to preprocess historical data of the microbial manufacturing process, build a general model for predicting cell concentration, train the model parameters of the source system based on historical data of the source system, and establish the model structure of the target system. The difference quantification and empirical modeling module is used to quantify the model differences between the source system and the target system based on the general model for predicting bacterial cell concentration, and to construct a unified empirical transfer probability framework by introducing latent variables. The parameter iterative estimation module is used to estimate the target system parameters through iterative calculation based on the empirical transfer probability framework until the target system model parameters converge. The sensing and result output module is used to substitute the converged target system model parameters into the general model to construct an online soft sensor. The online soft sensor is used to collect process auxiliary variable data of the target system in real time, and the real-time sensing value of bacterial cell concentration is calculated and output based on the collected data.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, the computer software product including several instructions for causing a computer device to execute the cell concentration sensing method for microbial manufacturing process based on experience transfer as described in any one of claims 1 to 8.

Images