Method for obtaining protein purification operation parameters and related device

Abstract

The application discloses a method for obtaining protein purification operation parameters and a related device, and relates to the technical field of biological manufacturing process control. The method comprises the following steps: obtaining directly measurable process parameters of a protein purification system detected by a sensor; processing the directly measurable process parameters by using a hybrid prediction model of a fusion purification process mechanism model and a data-driven residual compensation model and a state estimator to obtain non-directly measurable state variables of the protein purification system; and based on the state variables, performing rolling optimization within a model predictive control framework, adjusting operation parameter constraint conditions, and generating an optimal operation parameter sequence of the protein purification system. The application can realize high-precision real-time estimation of non-directly measurable state variables, real-time adaptive optimization of operation parameters in a dynamic environment with multivariable strong coupling, and maintenance of high consistency of key quality attributes.

Description

Technical Field

[0001] This application relates to the field of biomanufacturing process control technology, specifically to a method and related apparatus for obtaining protein purification operation parameters. Background Technology

[0002] Protein purification is a core step in biopharmaceutical manufacturing, and its process control level directly determines the quality attributes, production costs, and production efficiency of the final drug. In the current biomanufacturing industry, sensor-feedback-based automated control systems are typically used to ensure stable operation of the purification process. Existing mainstream technologies primarily rely on proportional-integral-derivative (PID) controllers to adjust basic process parameters such as temperature, pressure, and flow rate, or combine them with preset rule logic (such as fixed threshold triggering) to control key operational steps such as chromatography elution and membrane filtration switching. Furthermore, some advanced facilities are beginning to explore the introduction of mechanistic models or simple data-driven models, using offline calculations or static simulations to assist in formulating process parameters, aiming to achieve a basic description and monitoring of material balance and mass transfer kinetics during purification.

[0003] However, in complex actual protein purification production scenarios, due to fluctuations in raw material properties, the time-dependent degradation of purification media performance, and interference from environmental factors, existing control technologies struggle to acquire, in real-time and accurately, key state variables that cannot be directly measured, such as column concentration distribution and membrane fouling levels. This lack of internal state awareness prevents control systems from proactively adjusting before process deviations occur, often only responding passively after significant fluctuations in product quality. For a long time, achieving real-time adaptive optimization of operating parameters in a dynamic environment characterized by strong coupling of multiple variables, nonlinearity, and numerous unmeasurable states—in order to maintain high consistency in key quality attributes and reduce batch failure risk—has been a persistent technical challenge in this field. Summary of the Invention

[0004] In view of the above problems, this application provides a method and related apparatus for obtaining protein purification operating parameters. This method addresses the challenges of obtaining real-time, accurate, and non-measurable key state variables such as column concentration distribution and membrane fouling levels in complex actual protein purification production scenarios due to fluctuations in raw material properties, the time-dependent degradation of purification medium performance, and environmental interference. This results in the control system's inability to proactively adjust before process deviations occur, often only responding passively after significant fluctuations in product quality. Furthermore, it struggles to achieve real-time adaptive optimization of operating parameters in dynamic environments with strong multivariate coupling, nonlinearity, and numerous unmeasurable states, thereby maintaining high consistency in key quality attributes and reducing batch failure risk. The specific solution is as follows:

[0005] The first aspect of this application provides a method for obtaining protein purification operation parameters, including:

[0006] Obtain directly measurable process parameters of the protein purification system obtained from sensor detection;

[0007] By using a hybrid prediction model that combines a purification process mechanism model and a data-driven residual compensation model, along with a state estimator, the directly measurable process parameters are processed to obtain the non-directly measurable state variables of the protein purification system.

[0008] Based on the state variables, rolling optimization is performed within the model predictive control framework to adjust the constraints of the operating parameters and generate the optimal operating parameter sequence for the protein purification system.

[0009] In one possible implementation, the output of the hybrid prediction model is the sum of the predicted value of the purification process mechanism model and the output value of the data-driven residual compensation model; the state variables that cannot be directly measured include at least one of the following: dynamic binding capacity of the purification medium, intracolumn protein concentration distribution, or membrane fouling degree.

[0010] The step of performing rolling optimization within the model predictive control framework based on the state variables and adjusting the constraints of the operating parameters includes: when the state variables that cannot be directly measured represent the degradation of the performance of the purification medium, adjusting the constraint boundaries of the feed load, switching time, or operating pressure in the objective function or constraints of the model predictive control framework.

[0011] In one possible implementation, it further includes: determining that the performance of the purification medium has degraded when the dynamic binding capacity of the purification medium is detected to decrease with the usage period;

[0012] The adjustment of the constraint boundaries of feed load, switching time, or operating pressure in the objective function or constraints of the model predictive control framework includes:

[0013] When the dynamic binding capacity drops to a first preset ratio of the initial capacity, the upper limit of the feed load constraint is lowered to the load value corresponding to the first preset ratio, and the switching time of the purification system is adjusted.

[0014] When the dynamic combined capacity drops to a second preset ratio of the initial capacity, a media replacement warning signal is generated; wherein the second preset ratio is less than the first preset ratio.

[0015] In one possible implementation, the hybrid prediction model combining the fusion purification process mechanism model and the data-driven residual compensation model satisfies the following relationship:

[0016]

[0017] in, This is the output of the hybrid prediction model at the current time. This is a mechanism model for the purification process. For the state variables that cannot be directly measured, These are the operating parameters of the protein purification system. These are the model parameters for the purification process mechanism model. For the data-driven residual compensation model, The model parameters of the data-driven residual compensation model are... This is an input feature vector constructed based on the directly measurable process parameters.

[0018] In one possible implementation, the input feature vector This includes: the directly measurable process parameters and operating parameters of the protein purification system at the current moment, as well as the historical residual sequence of the directly measurable process parameters;

[0019] When the protein purification system is a chromatography system, the purification process mechanism model includes the mass conservation equation based on the axial dispersion model and the Langmuir adsorption isotherm; when the protein purification system is a membrane filtration system, the purification process mechanism model includes the flux model based on Darcy's law and the filter cake contamination model.

[0020] In one possible implementation, the process of obtaining the indirectly measurable state variables of the protein purification system using a hybrid prediction model and a state estimator includes:

[0021] The following state estimation process is performed by the state estimator, wherein the state estimator employs an unscented Kalman filter or an extended Kalman filter algorithm:

[0022] The state estimator calculates the predicted system state value and the prediction error covariance matrix at the current moment based on the state transition function of the hybrid prediction model.

[0023] The state estimator calculates the Kalman gain matrix, which is obtained based on the prediction error covariance matrix and in combination with the observation Jacobian matrix or the unscented transformation.

[0024] Using the Kalman gain matrix, the directly measurable process parameters are used as observations to correct the predicted system state values, resulting in updated system state estimates. The non-measurable state variables of the protein purification system are then obtained from the updated system state estimates.

[0025] In one possible implementation, the method further includes:

[0026] Calculate the prediction residual between the predicted measured value corresponding to the directly measurable process parameter and the directly measurable process parameter, wherein the predicted measured value is calculated using the indirectly measurable state variable;

[0027] An anomaly alarm is triggered when the predicted residual exceeds a dynamic threshold determined based on the moving mean and standard deviation, or when the statistic of the multivariate statistical process control exceeds a preset control limit.

[0028] In response to the anomaly alarm, the root cause variable leading to the anomaly is determined by contribution graph analysis or by performing a fault hypothesis simulation on the hybrid prediction model.

[0029] A tiered response strategy is executed based on the root dependent variable. The tiered response strategy includes at least one of the following: alarm logging, tightening the constraints of the model predictive control framework, switching to a safe operating mode, or requesting manual intervention.

[0030] A second aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the method for obtaining protein purification operation parameters as described in the first aspect or any implementation thereof.

[0031] A third aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0032] The memory is used to store computer programs;

[0033] The processor is used to execute the computer program so that the electronic device can implement the method for obtaining protein purification operation parameters as described in the first aspect or any implementation thereof.

[0034] A fourth aspect of this application provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the method for obtaining protein purification operation parameters described in the first aspect or any implementation thereof.

[0035] Using the above technical solution, this application provides a method and related apparatus for obtaining protein purification operation parameters. This solution obtains directly measurable process parameters of the protein purification system detected by sensors, and uses a hybrid prediction model that integrates the purification process mechanism model and the data-driven residual compensation model, along with a state estimator, to process the directly measurable process parameters, thereby obtaining the non-directly measurable state variables of the protein purification system. Then, based on the state variables, rolling optimization is performed within the model predictive control framework to adjust the constraints of the operation parameters and generate the optimal operation parameter sequence of the protein purification system. By constructing a hybrid predictive model driven by both mechanism and data, the physicochemical laws of the purification process can be accurately described and model residuals can be compensated. By using a state estimator to integrate measured data with model predictions, high-precision real-time estimation of unmeasurable state variables such as intracolumn protein concentration distribution, membrane fouling level, and resin dynamic binding capacity can be achieved. On this basis, rolling optimization using a model predictive control framework can dynamically adjust the constraints of operating parameters based on the estimated real-time state. This effectively solves the problem that in complex protein purification production scenarios, due to fluctuations in raw material properties, time-dependent degradation of purification medium performance, and interference from environmental factors, existing control technologies struggle to obtain real-time and accurate key unmeasurable state variables such as intracolumn concentration distribution and membrane fouling level. This results in the control system being unable to make proactive adjustments before process deviations occur, and can only respond passively after significant fluctuations in product quality. It is difficult to achieve real-time adaptive optimization of operating parameters in a dynamic environment with strong coupling of multiple variables, nonlinearity, and a large number of unmeasurable states to maintain high consistency of key quality attributes and reduce the risk of batch failure. Therefore, it avoids process runaway and batch failures caused by lack of state awareness, and improves the stability, reliability, and product consistency of the protein purification process. Attached Figure Description

[0036] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0037] Figure 1 A flowchart illustrating a method for obtaining protein purification operation parameters provided in this application;

[0038] Figure 2 The flowchart illustrates a method for obtaining protein purification operation parameters provided in this application, which utilizes a state estimator to perform a state estimation process to obtain state variables that cannot be directly measured.

[0039] Figure 3 A flowchart illustrating another method for obtaining protein purification operation parameters provided in this application;

[0040] Figure 4 This is a schematic diagram of an electronic device structure provided in this application. Detailed Implementation

[0041] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0042] As will be known to those skilled in the art, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0043] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0044] like Figure 1 The diagram shown is a flowchart of a method for obtaining protein purification operation parameters according to an embodiment of this application. The method includes:

[0045] Step S101: Obtain the directly measurable process parameters of the protein purification system detected by the sensor.

[0046] Measurable process parameters refer to process variables that characterize the system's operating status and are directly acquired by online monitoring instruments installed on the protein purification system. Examples of directly measurable process parameters include ultraviolet absorbance (UV280) measured by an online ultraviolet spectrometer, pressure drop across a chromatography column or membrane module measured by a pressure sensor, buffer conductivity measured by a conductivity meter, liquid flow rate measured by a flow meter, and pH value measured by a pH meter.

[0047] This step is typically performed by the data acquisition module deployed in the control system. The data acquisition module reads the current measurements from each sensor at a preset sampling period via industrial fieldbus (such as OPC UA, Modbus, etc.) or analog input channels, and performs filtering and unit conversion processing to form a structured real-time data sequence. This real-time data sequence is then synchronously transmitted to the subsequent hybrid prediction model and state estimator as input conditions for state estimation and prediction calculations.

[0048] Step S102: Using a hybrid prediction model combining the fusion purification process mechanism model and the data-driven residual compensation model, along with a state estimator, the directly measurable process parameters are processed to obtain the indirectly measurable state variables of the protein purification system.

[0049] A hybrid prediction model, consisting of a purification process mechanism model and a data-driven residual compensation model, is used to describe the dynamic behavior of a protein purification system. The purification process mechanism model, based on physicochemical principles (such as the law of conservation of mass, adsorption kinetics, and Darcy's law), describes the main evolutionary trends of the system. The data-driven residual compensation model captures complex nonlinearities, perturbations, or model mismatches not covered by the mechanism model. The state estimator, in the presence of noise and uncertainty, fuses model predictions with measured data to infer the true state of the system.

[0050] State variables that cannot be directly measured can refer to internal variables that cannot be directly read in real time by existing sensors, but are crucial for evaluating purification effectiveness and system performance. For example, such variables include the dynamic binding capacity of purification media (such as chromatography resins and filter membranes), the protein concentration distribution (axial or radial distribution) within the chromatography column or flow channel, and the degree of membrane fouling (such as filter cake resistance and pore blockage rate).

[0051] In the specific processing, the hybrid prediction model predicts the system output for the next moment based on historical states and current operating parameters. A state estimator (e.g., an Extended Kalman Filter (EKF) or an Unscented Kalman Filter (UKF)) uses the hybrid prediction model as a state transition predictor to calculate the prior state estimate for the current moment. Subsequently, the state estimator calculates the Kalman gain, reflecting the confidence weight between the model prediction and the sensor measurements, and uses this Kalman gain to correct the prior state estimate using the directly measurable process parameters obtained in step S101 as observations, thereby obtaining the posterior system state estimate. The non-directly measurable state variables of the protein purification system are extracted from the updated system state estimate. This fusion processing effectively overcomes sensor noise and model errors, providing high-precision key state information for subsequent optimized control.

[0052] In this embodiment, the hybrid prediction model, which integrates the purification process mechanism model and the data-driven residual compensation model, together with a state estimator implemented using algorithms such as Unscented Kalman Filter (UKF) or Extended Kalman Filter (EKF), constitutes a digital twin model of the physical protein purification system in virtual space. This digital twin model interacts with a sensor network deployed on the purification equipment (such as an online ultraviolet spectrometer and pressure sensors) at high frequency, enabling it to reflect the dynamic behavior of the physical system in real time. It not only utilizes the mechanism model to ensure the physical interpretability and robustness of the predictions, but also corrects for model bias and sensor noise through the data-driven model and state estimator, thereby achieving high-precision real-time estimation of indirect state variables such as intracolumn protein concentration distribution, resin dynamic binding capacity, and membrane fouling degree. This dynamic, high-fidelity digital twin model provides an accurate and reliable virtual simulation environment for subsequent rolling optimization and generation of optimal operating parameter sequences within the model predictive control framework, forming the core foundation for intelligent real-time control of the protein purification process.

[0053] Step S103: Based on the state variables, perform rolling optimization within the model predictive control framework, adjust the constraints of the operating parameters, and generate the optimal operating parameter sequence for the protein purification system.

[0054] Model predictive control (MMC) is an advanced model-based control strategy that uses a system model at each sampling time to predict the system dynamics over a certain time domain and solves the optimization problem to determine the optimal control inputs for the present and future. Rolling optimization refers to implementing only the first control action in the optimal operating parameter sequence and then re-predicting and optimizing based on new measurement data at the next sampling time, thus achieving closed-loop feedback control. Operating parameter constraints are physical or process safety limitations imposed on adjustable operating variables, such as upper and lower limits of feed flow rate, buffer switching time range, and maximum operating pressure. The optimal operating parameter sequence refers to a series of operating parameter values ​​that optimize a preset objective function (such as maximizing yield, maximizing purity, or minimizing production time) while satisfying all constraints.

[0055] In this step, the non-measurable state variables obtained in step S102 (such as the current resin binding capacity and column concentration distribution) are first used as the initial state of the predictive model in the model predictive control framework. Next, the prediction and control time domains are set within the model predictive control framework, and the objective function and constraints are defined. Based on a hybrid predictive model, the system performs multi-step predictive simulations under given operating parameter constraints, using optimization algorithms (such as sequential quadratic programming, interior-point methods, or linear programming) to find the future sequence of operating parameters that optimizes the objective function. Specifically, when the system state changes (e.g., resin aging leading to a decrease in binding capacity), the model predictive control framework dynamically adjusts the operating parameter constraints (e.g., appropriately reducing the feed load limit) to adapt to the real-time characteristics of the system. Finally, the generated optimal operating parameter sequence is output to the control system to guide the operation of the purification equipment. Through rolling optimization based on real-time state estimation, the protein purification process can always operate under optimal conditions, effectively coping with raw material fluctuations and equipment performance changes, thereby improving product quality consistency and production efficiency.

[0056] The aforementioned technologies work synergistically to achieve intelligent closed-loop control of the protein purification process. Measurable process parameters provided by sensors serve as observational inputs for state estimation. The hybrid predictive model, combined with the state estimator, enables high-precision online reconstruction of state variables that cannot be directly measured, solving the problem of unmeasurable information in key processes. Furthermore, the model predictive control framework performs rolling optimization based on accurate state information, dynamically adjusting operational parameter constraints to generate an optimal sequence of operational parameters adapted to the system's real-time characteristics. This collaborative mechanism allows the control system to proactively adapt to disturbances such as raw material fluctuations and media aging, thereby ensuring the stability of the protein purification process and the consistency of product quality.

[0057] In one optional embodiment, the output of the hybrid prediction model is the sum of the predicted value of the purification process mechanism model and the output value of the data-driven residual compensation model; the non-measurable state variables include at least one of the dynamic binding capacity of the purification medium, the protein concentration distribution within the column, or the degree of membrane fouling. Step S103, based on the state variables, performs rolling optimization within the model predictive control framework, and adjusts the specific process of adjusting the constraints of the operating parameters. This may include: when the non-measurable state variables characterize the degradation of the purification medium performance, adjusting the constraint boundaries of the feed load, switching time, or operating pressure in the objective function or constraints of the model predictive control framework.

[0058] This application employs a parallel fusion architecture to construct a hybrid prediction model, combining a purification process mechanism model based on physical laws with a data-driven residual compensation model trained on historical data. The output of the hybrid prediction model is not the prediction result of a single model, but rather the sum of the linear superposition of the predicted values ​​from the purification process mechanism model and the output values ​​from the data-driven residual compensation model. The purification process mechanism model is typically established based on mass conservation equations, mass transfer kinetic equations, and adsorption / desorption equilibrium equations to describe the spatiotemporal evolution of protein concentration, impurity concentration, and product quality attributes during purification, providing a basic predictive trend consistent with physical properties. However, due to perturbations in actual industrial environments that are difficult to model accurately, such as raw material differences, equipment aging, or minor environmental fluctuations, residual deviations often exist between the predicted values ​​of the purification process mechanism model and the actual sensor measurements. The data-driven residual compensation model is configured to capture and predict residual biases that cannot be explained by these non-physical mechanism models. Its construction process can be as follows: collect residual sequences between the purification process mechanism model's predicted values ​​and the sensor's measured values ​​from historical operational data; use a temporal neural network (e.g., Long Short-Term Memory network LSTM or Transformer) as its architecture, taking the historical residual sequence, current operating parameters, and state variables as inputs, to train the network to predict the residual values ​​at the next time step; update the network weights using a backpropagation algorithm until the model's prediction error converges on the validation set. By adding the predicted values ​​from the purification process mechanism model to the output values ​​of the data-driven residual compensation model, the hybrid prediction model retains the interpretability and generalization ability of the physical mechanism while using data-driven correction of model biases, thus outputting high-precision system prediction results.

[0059] By combining a hybrid predictive model with a state estimator (e.g., Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF), it is possible to acquire in real-time the state variables of a protein purification system that are not directly measurable. These unmeasurable state variables refer to internal parameters that cannot be directly read by existing online sensors (such as UV spectrometers, conductivity meters, pressure sensors, etc.) but are crucial for monitoring and controlling the purification process. The dynamic binding capacity (DBC) of the purification medium is a key indicator characterizing the adsorption capacity of chromatographic media (such as resins). Its value determines the total amount of protein that the medium can bind, and this parameter tends to decrease with increasing usage cycles and chemical cleaning losses. Intra-column protein concentration distribution refers to the spatial distribution of protein concentration in the liquid and solid phases along the axial or radial direction of the chromatography column. It reflects the movement of the adsorption band and the mass transfer effect, directly affecting the shape and resolution of the elution peak. Membrane fouling refers to the increase in membrane resistance during membrane filtration due to gel layer formation, pore blockage, or adsorption deposition, typically characterized by the membrane flux decay rate or the transmembrane pressure (TMP) rise rate. The estimation process for the aforementioned intangible state variables can be as follows: The state estimator utilizes externally observable data such as outlet concentration, pressure, and conductivity measured by sensors, combined with prior knowledge from the purification process mechanism model, to continuously refine the estimated value of the internal state vector through a recursive algorithm. For the dynamic binding capacity of the purification medium, the state estimator can online inversely calculate the current effective binding capacity based on the shape change of the breakthrough curve. For the degree of membrane fouling, the state estimator can, based on Darcy's law, use measured data of transmembrane pressure and flux to calculate the membrane fouling resistance in real time. By accurately estimating these intangible state variables, a deep understanding of the internal operating state of the purification system can be gained, providing accurate decision-making basis for subsequent optimization control.

[0060] This application introduces an adaptive constraint adjustment mechanism based on the health status of the purification medium within a model predictive control framework. The model predictive control framework performs rolling optimization within each control cycle, aiming to generate the optimal sequence of operating parameters for a future period. In traditional control strategies, the constraint boundaries of operating parameters are usually fixed. However, as the purification medium (such as the resin in a chromatography column or a filter membrane) is used more frequently, its physical properties degrade. Continuing to use the initial constraint boundaries can easily lead to product quality defects (such as resin breakthrough, decreased yield) or equipment damage. By monitoring the aforementioned intangible state variables, the performance status of the purification medium is evaluated in real time. When the state variable characterizes a degradation in the performance of the purification medium—for example, when the dynamic binding capacity of the purification medium calculated by the state estimator drops below a preset percentage of the initial value, or when the calculated membrane fouling resistance exceeds a preset threshold—the constraint adjustment logic within the model predictive control framework is triggered.

[0061] Specifically, strategies for adjusting the constraints of operating parameters include: for feed load, when the dynamic binding capacity of the purification medium decays, to prevent protein penetration, the upper limit of the feed load in the constraints of the model predictive control framework is reduced, i.e., the loading volume or mass per cycle is reduced, so that the loading amount matches the current decayed binding capacity; for switching time, in batch operations such as continuous capture chromatography (PCC), when the adsorption band movement speed slows down due to the deterioration of medium performance, the constraint range of switching time is extended to ensure that the target protein is fully captured before column switching; for operating pressure, in membrane filtration, when the degree of membrane fouling causes a sharp increase in resistance, to maintain flux and avoid excessive pressure damage to the membrane material, the constraints of operating pressure are tightened or relaxed, for example, limiting the maximum transmembrane pressure (TMP) to protect the membrane module, or maintaining relative flux stability by adjusting the operating pressure. These adjustments to the constraints are achieved by modifying the inequality constraints in the model predictive control framework optimization problem. This mechanism, based on state feedback, forces the optimization algorithm to search for the optimal sequence of operating parameters within a new feasible domain. Through this dynamic constraint adjustment mechanism, the performance degradation of the purification medium can be adaptively managed, fully utilizing the residual lifetime of the medium without sacrificing product quality, extending the medium's service life, and ensuring the safety and stability of the entire purification process.

[0062] This application constructs a closed-loop intelligent control system by combining the high-precision output of a hybrid predictive model, real-time sensing of state variables that cannot be directly measured, and an adaptive constraint adjustment strategy based on media performance degradation. This system can dynamically adapt to the time-varying characteristics of the equipment state, transforming internal variables characterizing the health state of the purification medium into executable control constraint changes. Thus, within the model predictive control framework, it achieves refined and robust adjustment of operating parameters, solving the technical problem that traditional fixed-constraint control easily leads to yield loss or equipment damage in equipment aging scenarios.

[0063] In one possible implementation, the method further includes determining that the performance of the purification medium has degraded when the dynamic binding capacity of the purification medium is detected to decrease over the usage period.

[0064] In the objective function or constraints of the aforementioned model predictive control framework, adjusting the constraint boundaries of feed load, switching time, or operating pressure can specifically include: when the dynamic binding capacity drops to a first preset proportion of the initial capacity, lowering the upper limit of the feed load constraint to the load value corresponding to the first preset proportion, and adjusting the switching time of the purification system; when the dynamic binding capacity drops to a second preset proportion of the initial capacity, generating a media replacement warning signal, wherein the second preset proportion is less than the first preset proportion.

[0065] The dynamic binding capacity of a purification medium refers to the maximum amount of target protein that the medium can bind under specific process conditions, and it is a core indicator characterizing the medium's performance. With increasing usage cycles, the dynamic binding capacity of the medium typically declines due to ligand shedding, pore blockage, or structural denaturation. By monitoring the change trend of the dynamic binding capacity in real time and analyzing its correlation with the usage cycle, early signs of performance degradation can be identified. This monitoring method differs from direct detection relying solely on the breakthrough curve; instead, it is based on real-time state estimates obtained from the aforementioned state estimator, enabling the early detection of microscopic performance degradation that is not yet clearly manifested in the effluent concentration. The logic for determining performance degradation can be, for example, comparing the current dynamic binding capacity with historical benchmark values, or fitting the slope of the degradation curve. When the degradation rate exceeds a preset range or the cumulative degradation reaches a certain level, the medium's performance is confirmed to have entered the degradation phase. This process provides a precise triggering basis for subsequent graded control strategies.

[0066] The first preset ratio serves as a parameter defining the initial warning threshold for media performance, triggering a protective control strategy. This strategy aims to maintain product quality by adjusting operating parameters when the media still possesses working capacity but its performance has deteriorated. When the monitored dynamic binding capacity drops to the first preset ratio of the initial capacity, it signifies a substantial reduction in the media's sample loading capacity. Maintaining the original feed load would significantly increase the risk of penetration. In this case, the upper limit of the feed load constraint is forcibly lowered by modifying the constraints within the model predictive control framework. The lowered load value matches the current level of dynamic binding capacity. For example, if the first preset ratio is 85%, the upper limit of the feed load constraint is set to 85% of the initial load, ensuring that the sample loading amount does not exceed the media's current actual carrying capacity, thereby avoiding protein loss and purity degradation due to overload.

[0067] A reduction in feed load directly affects the time required to reach the preset capture volume. To ensure the efficiency of the cyclic operation, the switching time needs to be calculated and updated synchronously. For systems such as continuous capture chromatography, the switching time adjustment logic can be recalculated based on the ratio of the current feed load to the effective binding capacity of the medium, to optimize the cycle time or extend the sample loading time to compensate for insufficient sample loading capacity. This coordinated adjustment ensures that the system can maintain efficient operation within safe operating boundaries in the early stages of medium performance degradation, extending the batch production cycle.

[0068] The second preset ratio serves as a parameter defining the critical threshold for media performance limits. A value less than the first preset ratio indicates that the media performance is nearing failure and requires replacement or regeneration. When the dynamic binding capacity further decreases to the second preset ratio, it indicates that the media's residual binding capacity can no longer meet the basic requirements for stable production. Continuing to maintain production by reducing the load will severely impact economic efficiency or pose uncontrollable quality risks. At this point, a media replacement warning signal is generated. This signal can be, for example, a visual alarm sent to the operating terminal or a maintenance work order written into the production planning system, prompting the replacement of the purification media at an appropriate time (e.g., after the current batch ends).

[0069] A tiered response mechanism is established by setting a second preset ratio, which is less than the first preset ratio. The first preset ratio triggers adaptive parameter adjustments, primarily to address natural performance degradation and maximize the remaining value of the medium. The second preset ratio triggers preventative maintenance warnings, mainly to prevent sudden failures and provide a sufficient preparation window. This tiered control logic combines the real-time nature of state monitoring with the predictive nature of predictive control, enabling the protein purification process to automatically and smoothly transition to a safe state when faced with the common disturbance of purification medium performance degradation. This avoids batch losses due to medium failure and maximizes the utilization of the medium's lifespan.

[0070] In the aforementioned technical process, dynamic real-time capacity monitoring provides the foundation for state awareness; the hierarchical threshold system composed of the first and second preset ratios provides a quantitative standard for decision-making; and constraint adjustment and early warning generation within the model predictive control framework provide specific execution means. By monitoring the decay trend and dynamically reducing the load or triggering early warnings, a shift from passively responding to faults to actively managing the media lifecycle is achieved. This ensures that the protein purification system remains in a controlled and optimal operating state during the performance degradation of the purification media, effectively reducing production risks caused by media aging.

[0071] In another embodiment, the hybrid prediction model that integrates the purification process mechanism model and the data-driven residual compensation model satisfies the following relationship:

[0072]

[0073] in, This is the output of the hybrid prediction model at the current time. This is a mechanism model for the purification process. For the state variables that cannot be directly measured, These are the operating parameters of the protein purification system. These are the model parameters for the purification process mechanism model. For the data-driven residual compensation model, The model parameters of the data-driven residual compensation model are... This is an input feature vector constructed based on the directly measurable process parameters.

[0074] The output of the hybrid prediction model at the current time This output characterizes the comprehensive state response of the protein purification system at the current moment, encompassing directly measurable process parameters (such as UV absorbance, conductivity, and pressure) as well as key quality attributes that cannot be directly measured. This output serves as the core basis for subsequent control decisions, and its accuracy directly determines the optimization effect of model predictive control (MPC).

[0075] Purification process mechanism model This forms the framework of the hybrid prediction model, which is essentially a mathematical description based on physicochemical laws (such as the law of conservation of mass, mass transfer kinetics equations, and adsorption isotherm equations). It is responsible for capturing the fundamental evolutionary laws with clear physical meaning, whether linear or nonlinear, during the purification process. The inputs include state variables that cannot be directly measured. (e.g., intracolumn protein concentration distribution, resin binding capacity, etc.) and operating parameters (e.g., flow rate, pH value, salt concentration gradient, etc.), model parameters These correspond to constants that characterize physicochemical properties (such as axial diffusion coefficient, mass transfer coefficient, equilibrium constant, etc.). These parameters are usually obtained through offline experimental determination or theoretical calculation and have strong physical interpretability.

[0076] Data-driven residual compensation model It is used to correct the deviation between the predicted values ​​of the mechanistic model and the actual observed values. In actual industrial production, due to raw material fluctuations, equipment aging, sensor noise, and the simplification assumptions of the mechanistic model itself, a single mechanistic model may have prediction errors. By learning residual patterns from historical data, compensation is provided for these nonlinear and time-varying characteristics that are difficult to model accurately. In this embodiment, For example, temporal neural network architectures such as Long Short-Term Memory (LSTM) or Transformer can be used to capture long-term dependencies in time series data. Their model parameters... Optimize and adjust through online or offline training processes.

[0077] State variables that cannot be directly measured These can refer to process variables that cannot be directly obtained through conventional sensors but have a decisive impact on product quality and process stability. These variables are usually estimated in real time by state estimators (such as Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF)) based on sensor measurements and mechanistic models. Accurate estimation is crucial for understanding the internal dynamics of the system (such as in-column adsorption saturation and membrane fouling), and is the foundation for achieving closed-loop control and anomaly diagnosis.

[0078] Operating parameters of protein purification system It is a physical quantity that the controller can adjust, and it is a direct input that affects the system state and product output. During the tomography process, Including but not limited to feed flow rate, buffer switching time, elution gradient slope, etc.; in the membrane filtration process, These include transmembrane pressure (TMP), feed rate, and tangential velocity. These parameters constitute the optimization variables in model predictive control, and can be adjusted appropriately. It can enable proactive intervention in the technological process.

[0079] Model parameters of the purification process mechanism model These parameters are key factors determining the accuracy of mechanistic models. They reflect the inherent properties of the purification medium (such as chromatography resins and filtration membranes) and the characteristics of reaction kinetics. For example, The maximum adsorption capacity of the resin (can be included) ), adsorption equilibrium constant ( ), membrane inherent resistance ( ), etc. In the early stages of model building, The model was determined through fitting small-scale experimental data; in actual operation, when the model prediction error continues to increase, it can also trigger adjustments. Online calibration is performed to accommodate performance degradation or batch-to-batch differences in the purification medium.

[0080] Model parameters of data-driven residual compensation model The structure and weight distribution of the neural network are defined. For neural network classes... , It includes the weight matrices, bias vectors, and activation function parameters for each layer. These parameters do not have intuitive physical meaning but are obtained through training by minimizing the prediction loss function (such as mean squared error, MSE). Under the incremental learning mechanism, It will be continuously updated as new data accumulates, thereby improving the model's adaptability to complex working conditions.

[0081] Input feature vectors constructed based on directly measurable process parameters It is the input data source for the data-driven residual compensation model. It is not a simple accumulation of raw sensor data, but a structured data sequence after feature engineering. The construction process can be as follows: First, high-frequency raw data (such as UV280 signals at 10 times per second) is collected from the sensor network; second, the raw data is cleaned, denoised, and normalized; then, statistical features (such as moving averages, variances, and rates of change) or time series segments that reflect the trend of process changes are extracted; finally, the processed feature data is assembled into a high-dimensional vector. . The quality of construction directly affects The learning effect of residual patterns is usually determined by combining domain knowledge to select the key features most sensitive to process deviations.

[0082] The hybrid prediction model constructed using the aforementioned relationships combines the strong extrapolation capability of the mechanistic model with the high fitting accuracy of the data-driven model. This gray-box modeling approach ensures the model's robustness under unfamiliar operating conditions while fully utilizing the massive amounts of data accumulated during production to correct deviations, thereby achieving high-fidelity simulation of the protein purification system and providing reliable technical support for real-time optimization control.

[0083] In one optional embodiment, this application also provides specific methods for constructing input feature vectors and different system mechanism models.

[0084] Step 1: Construct the input feature vector .

[0085] Step 2: When the protein purification system is a chromatography system, construct a purification process mechanism model based on the mass conservation equation of the axial dispersion model and the Langmuir adsorption isotherm.

[0086] Step 3: When the protein purification system is a membrane filtration system, construct a purification process mechanism model based on Darcy's law flux model and filter cake contamination model.

[0087] Input feature vector This includes the directly measurable process parameters and operating parameters of the protein purification system at the current moment, as well as the historical residual sequence of the directly measurable process parameters. This vector serves as the basic data structure for constructing a data-driven residual compensation model, combining the raw data collected by sensors with the system's control history to provide the model with a complete informational context that reflects the current operating conditions and past deviation trends.

[0088] When the protein purification system is a chromatography system, the purification process mechanism model includes the mass conservation equation based on the axial dispersion model and the Langmuir adsorption isotherm; when the protein purification system is a membrane filtration system, the purification process mechanism model includes the flux model based on Darcy's law and the filter cake contamination model.

[0089] The directly measurable process parameters of the protein purification system at the current moment constitute a snapshot of the system's current operating state. These include protein absorbance values ​​detected by an online UV spectrometer, solution conductivity detected by a conductivity meter, system pressure detected by a pressure sensor, fluid flow rate detected by a flow meter, and temperature detected by a temperature sensor. Operating parameters can refer to input variables set and executed by the control system, such as pump flow rate in a chromatography system, the rate of change of salt concentration in gradient elution, or switching time commands, and transmembrane pressure (TMP) and feed flow rate setpoints in a membrane filtration system. The historical residual sequence of directly measurable process parameters can refer to the error sequence between sensor measurements and mechanistic model predictions within a previously set time window. This sequence contains unmodeled dynamic information that the mechanistic model cannot explain, such as minor environmental disturbances or sensor drift.

[0090] The input feature vector is formed by aligning the three types of data according to their timestamps and concatenating them into a high-dimensional vector. It can capture the multidimensional correlation characteristics of the system in time and space, thus providing accurate input basis for data-driven residual compensation models.

[0091] A chromatography system is a unit operation that utilizes a stationary phase medium to perform differential adsorption and separation of target proteins and impurities. The purification process mechanism model aims to describe the migration and distribution of substances within the chromatography column from a physical and chemical perspective.

[0092] The mass conservation equation based on the axial dispersion model describes the convection, diffusion, and adsorption dynamics of proteins in the liquid phase as they move axially along the mobile phase in the chromatography column. This equation comprehensively considers fluid velocity, axial diffusion coefficient, and mass transfer rate between the liquid and solid phases, reflecting the continuous nature of protein concentration changes with time and spatial location. Langmuir adsorption isotherms describe the adsorption equilibrium behavior of protein molecules on the solid medium surface, defining the nonlinear relationship between the solid surface binding concentration and the liquid free concentration, specifically characterizing the saturated adsorption characteristics when the number of adsorption sites is limited.

[0093] By combining the mass conservation equation based on the axial dispersion model with the Langmuir adsorption isotherm, the constructed mechanistic model can accurately simulate the protein breakthrough curve, elution peak shape, and loading dynamics during chromatography, providing physically interpretable basic predictions for the mixed prediction model. .

[0094] Membrane filtration systems utilize semi-permeable membranes to achieve feed concentration or buffer replacement under pressure. In this scenario, the core of the purification process mechanism model lies in describing the variation of permeate flux and the accumulation process of membrane fouling.

[0095] A flux model based on Darcy's law establishes a linear relationship between transmembrane pressure and permeate flux. This model treats the filtration process as fluid flow through a porous medium, where fluid flux is directly proportional to the driving pressure and inversely proportional to the total filtration resistance, which is composed of inherent membrane resistance, cake layer resistance, and membrane pore blockage resistance. The cake layer fouling model describes the dynamic process of trapped proteins or impurities accumulating on the membrane surface to form a cake layer. It elucidates the mechanism by which cake layer resistance increases with filtration volume or time, and the direct impact of cake layer thickness on flux decay.

[0096] By combining the flux model based on Darcy's law with the cake fouling model, it is possible to predict the flux decay trend and transmembrane pressure changes in the membrane filtration process in real time, thereby effectively supporting the accurate estimation of the membrane filtration status by the hybrid prediction model.

[0097] The synergistic effect of the above-mentioned technical features allows for the customized construction of mechanistic models based on the physical characteristics of different purification systems. By using feature vectors containing rich historical deviation information to train data-driven models, the fused hybrid prediction model can retain the robustness and extrapolation ability of the physical mechanism while correcting mechanistic deviations through data-driven methods, thus achieving high-precision prediction of the entire protein purification process.

[0098] In one alternative embodiment, such as Figure 2 The diagram shows a flowchart illustrating a method for obtaining protein purification operation parameters according to an embodiment of this application, in which a state estimator is used to perform a state estimation process to obtain state variables that cannot be directly measured. The method further includes:

[0099] Step S201: The state estimator calculates the predicted system state value and the prediction error covariance matrix at the current moment based on the state transition function of the hybrid prediction model.

[0100] In the protein purification process, to address the issues of sparse sensor measurement data and inherent errors in the mechanistic model, a state estimator is introduced as the core module for data fusion. This state estimator is configured to use either the Unscented Kalman Filter (UKF) algorithm or the Extended Kalman Filter (EKF) algorithm, aiming to optimally estimate the true state of the system under nonlinear and non-Gaussian noise environments.

[0101] The calculation of the system state prediction depends on the system state estimate from the previous time step and the currently applied operating parameters. The state transition function describes the physical laws governing the evolution of the system state over time, and its specific form is determined by the hybrid prediction model. For example, for a chromatography process involving an axial dispersion model, the state transition function is a discretized set of partial differential equations characterizing the axial diffusion and convection of protein concentration within the column; for a membrane filtration process involving a cake fouling model, it is a kinetic equation representing the changes in membrane flux and fouling resistance with operating pressure. By solving this state transition function, the prior estimate before the introduction of measured data correction at the current time step is obtained, i.e., the system state prediction.

[0102] Simultaneously, to quantify the uncertainty of the prediction process, the prediction error covariance matrix needs to be calculated. This matrix represents a measure of confidence in the predicted system state, and its evolution is driven by the process noise covariance matrix. In the unscented Kalman filter algorithm, the prediction error covariance matrix is ​​obtained statistically by propagating the Sigma points generated by the unscented transformation through the state transition function; while in the extended Kalman filter algorithm, the covariance distribution after the state transition is approximated by linearizing the state transition function and using the Jacobian matrix. This process enables dynamic tracking of the system state uncertainty, providing the necessary statistical basis for the subsequent calculation of the Kalman gain.

[0103] Step S202: The state estimator calculates the Kalman gain matrix, which is based on the prediction error covariance matrix and is calculated in combination with the observation Jacobian matrix or the unscented transformation.

[0104] After completing the prior prediction of the current state, the preparation work for the update phase is crucial, and the core of this phase lies in calculating the Kalman gain matrix. The Kalman gain matrix determines the weight distribution between the measured data acquired by the sensor and the model predictions in subsequent correction steps; essentially, it measures the reliability of the observed information relative to the prior prediction information.

[0105] The process of calculating the Kalman gain matrix fully utilizes the prediction error covariance matrix obtained in the preceding steps and the characteristics of the system's observation model. If the extended Kalman filter algorithm is used, the observation Jacobian matrix needs to be calculated. This matrix characterizes the local linear mapping relationship between the system state variables and directly measurable process parameters, such as the gradient relationship between in-column protein concentration and out-of-column UV absorbance. Combining the prediction error covariance matrix, the observation Jacobian matrix, and the measurement noise covariance matrix, the Kalman gain matrix is ​​obtained through matrix operations. If the unscented Kalman filter algorithm is used, there is no need to calculate the complex Jacobian matrix. Instead, the unscented transformation is used again to map the Sigma points in the prediction stage to the measurement space through the observation function. By statistically analyzing the interaction between the covariance of these propagation points and the measurement noise covariance, the Kalman gain matrix is ​​calculated.

[0106] The Kalman gain matrix is ​​a time-varying matrix that dynamically changes with the system state and measurement accuracy. When the prediction error covariance is large (i.e., insufficient confidence in the model's prediction) or the measurement noise is low, the Kalman gain matrix increases, meaning the state estimator assigns higher weight to the sensor's measured data. Conversely, when the model prediction is more accurate or the measurement is severely affected by noise, the Kalman gain matrix decreases, and the state estimator relies more on the model's prediction results. This adaptive weight adjustment mechanism effectively resists noise interference and is key to achieving high-precision state estimation.

[0107] Step S203: Using the Kalman gain matrix, the directly measurable process parameters are used as observations to correct the predicted system state values, resulting in updated system state estimates. The non-measurable state variables of the protein purification system are then obtained from the updated system state estimates.

[0108] After calculating the Kalman gain matrix, a final correction operation is performed to update the system state predictions with a posterior estimate that more closely approximates the actual physical state. A high-precision sensor network deployed on the purification equipment collects directly measurable process parameters in real time as observations. These directly measurable process parameters include absorbance detected by an online ultraviolet spectrometer, ion intensity detected by a conductivity meter, transmembrane pressure difference detected by a pressure sensor, and fluid velocity detected by a flow meter.

[0109] The correction process involves calculating the residual between the observed values ​​and the predicted observed values ​​derived from the system state predictions. This residual reflects the deviation between the model predictions and the actual measurements. Then, the Kalman gain matrix is ​​multiplied by this residual to obtain a state correction. This state correction is then added to the system state prediction obtained in the preceding steps to obtain the updated system state estimate. This mathematical operation is essentially a weighted correction to the predicted state, ensuring that the state estimate, while satisfying physical mechanism constraints, approximates the system state revealed by the actual sensor data to the greatest extent possible.

[0110] The updated system state estimate is a high-dimensional vector containing multi-dimensional variables. This includes smoothed estimates of directly measurable process parameters, but more importantly, it includes indirect state variables of the protein purification system. These indirect state variables are key internal parameters that physical sensors cannot directly access, such as the protein concentration distribution along the axial direction within the chromatography column, the current dynamic binding capacity (DBC) of the purification medium, the degree of membrane fouling during membrane filtration, and the thickness of the filter cake. Through the Kalman filter-based state estimation process described above, external, directly measurable process parameters can be mapped to internal state information, thus providing insight into the purification process's "black box" and offering accurate and reliable state feedback for subsequent model predictive control.

[0111] One possible implementation is, such as Figure 3 As shown, the method provided in this application embodiment further includes:

[0112] Step S104: Calculate the prediction residual between the predicted measured value corresponding to the directly measurable process parameter and the directly measurable process parameter, wherein the predicted measured value is calculated using the indirectly measurable state variable.

[0113] Using the system state variables obtained in the preceding steps, the intangible state variables are mapped to the measurable output space of the sensors through state observation equations, thereby obtaining predicted measurements. These predicted measurements represent the theoretical prediction of the physical system output by the digital twin model based on the current state estimate. Specifically, the state variables are substituted into a preset observation matrix or nonlinear observation function to calculate the theoretical readings of sensors such as ultraviolet absorbance, conductivity, pressure, and flow rate at the corresponding time. Subsequently, the predicted measurements are compared point-by-point with the actual sensor data collected at the same time, and the deviation between the two is calculated; this deviation is the prediction residual.

[0114] Prediction residuals not only reflect numerical differences but are also key indicators characterizing the accuracy of model predictions and the consistency of system operation. Their physical nature reveals the degree of mismatch between the theoretical model and the physical entity at the current moment. This mismatch may stem from model parameter drift, unmeasured disturbance inputs, or sensor malfunctions. In this embodiment, the prediction residuals are constructed as a multi-dimensional vector, with each dimension corresponding to a sensor's measurement channel. By calculating this residual sequence in real time, subtle abnormal state changes in the system can be captured, providing a quantitative basis for subsequent anomaly detection.

[0115] Step S105: When the predicted residual exceeds the dynamic threshold determined based on the moving mean and standard deviation, or when the statistic of the multivariate statistical process control exceeds the preset control limit, an abnormal alarm is triggered.

[0116] Based on the generated predicted residuals, a dual-track monitoring mechanism is employed to statistically analyze the system status and determine whether the system is in an abnormal state. This mechanism combines univariate residual monitoring with multivariate statistical monitoring to comprehensively cover the characteristics of different types of faults. For univariate residual monitoring, the system dynamically calculates the moving mean and standard deviation of the residuals and constructs a dynamic threshold that adaptively adjusts with operating conditions based on these two statistics. When the predicted residual in any dimension exceeds this dynamic threshold, it is considered that a single point of anomaly has occurred.

[0117] The mathematical expression of this process is as follows: First, calculate the dynamic threshold of the residual. The formula is:

[0118]

[0119] In the formula, This represents the moving average of the predicted residuals. This represents the moving standard deviation of the predicted residuals. This is a preset sensitivity coefficient (e.g., set to 2 or 3 to adjust the sensitivity of the alarm). When predicting residuals... satisfy When this occurs, an exception detection is triggered.

[0120] For multivariate statistical process control, dimensionality reduction techniques such as Principal Component Analysis (PCA) are used to extract latent features under normal operating conditions, and Hotelling is calculated. The sum of squared prediction error (SPE) is used to determine the system's operating status as deviating from the normal range when either of these two statistics exceeds a preset control limit (usually determined based on quantiles of historical data distributions, such as the 99% confidence limit), triggering an anomaly alarm. By combining univariate and multivariate monitoring logic, it is possible to effectively identify abrupt faults and gradual performance drift, ensuring robust anomaly detection.

[0121] Step S106: In response to an anomaly alarm, determine the root cause variable leading to the anomaly by performing a fault hypothesis simulation on a hybrid prediction model through contribution graph analysis or by performing a fault hypothesis simulation.

[0122] Once an anomaly alarm is triggered, a root cause diagnostic procedure is immediately initiated to locate the specific physical source or operational variable causing the anomaly from multidimensional sensor data, i.e., the root cause variable. The diagnostic process mainly involves two paths.

[0123] The first approach is contribution plot analysis. This method decomposes the aforementioned multivariate statistics (such as SPE) and calculates the contribution rate of each sensor variable to the overall statistical error. In practice, for each measurement variable, the ratio of its projected distance in the residual space to the overall residual modulus is calculated. The variable with the largest ratio is the one that contributes the most to the current abnormal operating condition and is thus identified as a potential root cause variable.

[0124] The second approach involves performing fault hypothesis simulations on a hybrid predictive model. This method presupposes a series of possible fault hypotheses, such as sudden changes in feed concentration, a decrease in mass transfer coefficient, or sensor zero-point drift. In a digital twin environment, each fault hypothesis is sequentially introduced into the hybrid predictive model, simulating its output response under current operating conditions, and calculating the degree of match between the simulated output and the actual measured values. The variable or parameter corresponding to the fault hypothesis with the highest degree of match is identified as the root cause variable leading to the anomaly. By combining contribution plot analysis with fault hypothesis simulation, both rapid diagnosis and mechanism verification are achieved, significantly improving the accuracy of root cause localization.

[0125] Step S107: Execute a graded response strategy based on the root dependent variable. The graded response strategy includes at least one of alarm logging, tightening the constraints of the model predictive control framework, switching to a safe operating mode, or requesting manual intervention.

[0126] Once the root cause variable is identified, a tiered response strategy is automatically matched and executed from a pre-defined strategy library based on its severity, scope of impact, and reversibility. This strategy follows a principle of gradual escalation from mild to severe, aiming to maintain production continuity as much as possible while ensuring system safety.

[0127] Specifically, when the root cause variable represents a minor disturbance or temporary fluctuation, a Level 1 response is executed, which involves only recording an alarm and storing the abnormal event, its occurrence time, and root cause information in a log database for subsequent trend analysis and auditing. When the root cause variable indicates that the process parameters are approaching the safety boundary but still have room for adjustment, a Level 2 response is executed, which involves tightening the constraints of the model predictive control framework. For example, this might involve narrowing the allowable range of the manipulated variables or increasing the weight of quality indicators in the objective function, forcing the control system to take more conservative actions to suppress the spread of the anomaly.

[0128] When the root cause indicates that the system has entered a hazardous area or that the performance of major equipment has significantly deteriorated, a Level 3 response is executed, automatically switching to a safe operating mode. In this mode, the control logic shifts from pursuing optimal performance to maintaining system stability, such as automatically reducing load, shutting down high-risk sub-equipment, or switching to a backup control loop. If the root cause exceeds the system's processing capacity or involves significant safety risks, a Level 4 response is executed, requesting manual intervention, sending the highest priority alarm to the control panel, suspending automatic control functions, and awaiting operator confirmation or executing an emergency shutdown. This differentiated, tiered response mechanism achieves a shift from passive response to proactive defense, effectively reducing the risk of abnormal operating conditions escalating into production accidents.

[0129] The synergistic effect of these steps, through calculating predicted residuals and utilizing dynamic thresholds and statistics for real-time monitoring, enables the sensitive detection of minute abnormal fluctuations during protein purification. A diagnostic logic combining contribution graph analysis and fault hypothesis simulation allows for rapid localization and mechanism verification of the root causes of anomalies. A graded response strategy based on the severity of the root cause provides the system with a complete chain of handling mechanisms, from log recording to safe shutdown. This closed-loop control architecture of monitoring-diagnosis-response enables the protein purification system to maintain a high degree of autonomy and safety in the face of raw material fluctuations, equipment aging, and environmental disturbances, significantly improving the robustness and intelligence of the production process.

[0130] This application also provides an electronic device in its embodiments. (See reference...) Figure 4 The diagram illustrates a structural schematic of an electronic device suitable for implementing the method for obtaining protein purification operation parameters in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 4 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0131] like Figure 4As shown, the electronic device may include a processor (e.g., a central processing unit, a graphics processing unit, etc.) 401 and a memory 402 connected to the processor 401, wherein:

[0132] Memory 402 is used to store computer programs;

[0133] The processor 401 is used to execute a computer program to enable the electronic device to implement any of the methods for obtaining protein purification operation parameters provided in the embodiments of this application.

[0134] This application also provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the protein purification operation parameters obtained in this application.

[0135] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the protein purification operation parameters obtained in this application.

[0136] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0137] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0138] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0139] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A method for obtaining protein purification operation parameters, characterized in that, include: Obtain directly measurable process parameters of the protein purification system obtained from sensor detection; By using a hybrid prediction model that combines a purification process mechanism model and a data-driven residual compensation model, along with a state estimator, the directly measurable process parameters are processed to obtain the non-directly measurable state variables of the protein purification system. Based on the state variables, rolling optimization is performed within the model predictive control framework to adjust the constraints of the operating parameters and generate the optimal operating parameter sequence for the protein purification system.

2. The method according to claim 1, characterized in that, The output of the hybrid prediction model is the sum of the predicted value of the purification process mechanism model and the output value of the data-driven residual compensation model; the state variables that cannot be directly measured include at least one of the following: dynamic binding capacity of the purification medium, protein concentration distribution in the column, or membrane fouling degree. The step of performing rolling optimization within the model predictive control framework based on the state variables and adjusting the constraints of the operating parameters includes: when the state variables that cannot be directly measured represent the degradation of the performance of the purification medium, adjusting the constraint boundaries of the feed load, switching time, or operating pressure in the objective function or constraints of the model predictive control framework.

3. The method according to claim 2, characterized in that, Also includes: When the dynamic binding capacity of the purification medium is detected to decrease over the usage period, it is determined that the performance of the purification medium has degraded. The adjustment of the constraint boundaries of feed load, switching time, or operating pressure in the objective function or constraints of the model predictive control framework includes: When the dynamic binding capacity drops to a first preset ratio of the initial capacity, the upper limit of the feed load constraint is lowered to the load value corresponding to the first preset ratio, and the switching time of the purification system is adjusted. When the dynamic combined capacity drops to a second preset ratio of the initial capacity, a media replacement warning signal is generated; wherein the second preset ratio is less than the first preset ratio.

4. The method according to claim 2, characterized in that, The hybrid prediction model combining the fusion purification process mechanism model and the data-driven residual compensation model satisfies the following relationship: in, This is the output of the hybrid prediction model at the current time. This is a mechanism model for the purification process. For the state variables that cannot be directly measured, These are the operating parameters of the protein purification system. These are the model parameters for the purification process mechanism model. For the data-driven residual compensation model, The model parameters of the data-driven residual compensation model are... This is an input feature vector constructed based on the directly measurable process parameters.

5. The method according to claim 4, characterized in that, The input feature vector This includes: the directly measurable process parameters and operating parameters of the protein purification system at the current moment, as well as the historical residual sequence of the directly measurable process parameters; When the protein purification system is a chromatography system, the purification process mechanism model includes the mass conservation equation based on the axial dispersion model and the Langmuir adsorption isotherm; when the protein purification system is a membrane filtration system, the purification process mechanism model includes the flux model based on Darcy's law and the filter cake contamination model.

6. The method according to claim 1, characterized in that, The process of obtaining the indirectly measurable state variables of the protein purification system using a hybrid prediction model and a state estimator includes: The following state estimation process is performed by the state estimator, wherein the state estimator employs an unscented Kalman filter or an extended Kalman filter algorithm: The state estimator calculates the predicted system state value and the prediction error covariance matrix at the current moment based on the state transition function of the hybrid prediction model. The state estimator calculates the Kalman gain matrix, which is obtained based on the prediction error covariance matrix and in combination with the observation Jacobian matrix or the unscented transformation. Using the Kalman gain matrix, the directly measurable process parameters are used as observations to correct the predicted system state values, resulting in updated system state estimates. The non-measurable state variables of the protein purification system are then obtained from the updated system state estimates.

7. The method according to claim 1, characterized in that, The method further includes: Calculate the prediction residual between the predicted measured value corresponding to the directly measurable process parameter and the directly measurable process parameter, wherein the predicted measured value is calculated using the indirectly measurable state variable; An anomaly alarm is triggered when the predicted residual exceeds a dynamic threshold determined based on the moving mean and standard deviation, or when the statistic of the multivariate statistical process control exceeds a preset control limit. In response to the anomaly alarm, the root cause variable leading to the anomaly is determined by contribution graph analysis or by performing a fault hypothesis simulation on the hybrid prediction model. A tiered response strategy is executed based on the root dependent variable. The tiered response strategy includes at least one of the following: alarm logging, tightening the constraints of the model predictive control framework, switching to a safe operating mode, or requesting manual intervention.

8. A computer program product, characterized in that, Includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to perform the method for obtaining protein purification operation parameters as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the method for obtaining protein purification operation parameters as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the method for obtaining protein purification operation parameters as described in any one of claims 1 to 7.

Images